From d0c614ade336ba6c644c545f50291ebb3ecd09a2 Mon Sep 17 00:00:00 2001 From: Aurelio Vivas Date: Wed, 9 Sep 2026 21:15:43 +0000 Subject: [PATCH 01/13] Vendor minimal PyTorch RecordFunction headers for pytorch backend Copies the transitive header closure of ATen/record_function.h (605 files) from Aurora's `module load frameworks` PyTorch install (2.10.0a0), following the same vendoring pattern hip/ze/mpi use for their real API headers. Verified standalone compile against these headers with the POC tracer. Co-Authored-By: Claude Sonnet 5 --- .../pytorch/include/ATen/MethodOperators.h | 449 ++ backends/pytorch/include/ATen/StorageUtils.h | 54 + backends/pytorch/include/ATen/core/ATen_fwd.h | 51 + .../include/ATen/core/CheckMemoryFormat.h | 29 + .../ATen/core/DeprecatedTypeProperties.h | 144 + .../core/DeprecatedTypePropertiesRegistry.h | 38 + backends/pytorch/include/ATen/core/Dict.h | 401 ++ backends/pytorch/include/ATen/core/Dict_inl.h | 213 + .../pytorch/include/ATen/core/DimVector.h | 18 + backends/pytorch/include/ATen/core/Dimname.h | 53 + .../pytorch/include/ATen/core/Generator.h | 194 + backends/pytorch/include/ATen/core/IListRef.h | 638 ++ .../pytorch/include/ATen/core/IListRef_inl.h | 208 + backends/pytorch/include/ATen/core/List.h | 496 ++ backends/pytorch/include/ATen/core/List_inl.h | 358 + .../pytorch/include/ATen/core/NamedTensor.h | 148 + .../pytorch/include/ATen/core/QuantizerBase.h | 89 + backends/pytorch/include/ATen/core/Tensor.h | 103 + .../include/ATen/core/TensorAccessor.h | 66 + .../pytorch/include/ATen/core/TensorBase.h | 1098 ++++ .../pytorch/include/ATen/core/TensorBody.h | 5799 +++++++++++++++++ backends/pytorch/include/ATen/core/blob.h | 209 + .../pytorch/include/ATen/core/custom_class.h | 33 + .../pytorch/include/ATen/core/dynamic_type.h | 252 + .../pytorch/include/ATen/core/functional.h | 59 + backends/pytorch/include/ATen/core/ivalue.h | 1642 +++++ .../pytorch/include/ATen/core/ivalue_inl.h | 2578 ++++++++ .../pytorch/include/ATen/core/ivalue_to.h | 41 + backends/pytorch/include/ATen/core/jit_type.h | 2435 +++++++ .../pytorch/include/ATen/core/jit_type_base.h | 602 ++ .../pytorch/include/ATen/core/operator_name.h | 103 + .../include/ATen/core/qualified_name.h | 166 + .../include/ATen/core/rref_interface.h | 46 + backends/pytorch/include/ATen/core/symbol.h | 152 + .../pytorch/include/ATen/core/type_factory.h | 113 + backends/pytorch/include/ATen/core/type_ptr.h | 59 + .../include/ATen/ops/_addmm_activation_ops.h | 45 + .../ops/_autocast_to_full_precision_ops.h | 34 + .../ops/_autocast_to_reduced_precision_ops.h | 34 + .../pytorch/include/ATen/ops/_backward_ops.h | 34 + .../pytorch/include/ATen/ops/_coalesced_ops.h | 56 + backends/pytorch/include/ATen/ops/_conj_ops.h | 34 + .../include/ATen/ops/_conj_physical_ops.h | 45 + backends/pytorch/include/ATen/ops/_dimI_ops.h | 34 + backends/pytorch/include/ATen/ops/_dimV_ops.h | 34 + .../pytorch/include/ATen/ops/_fw_primal_ops.h | 34 + .../pytorch/include/ATen/ops/_indices_ops.h | 34 + .../include/ATen/ops/_is_all_true_ops.h | 34 + .../include/ATen/ops/_is_any_true_ops.h | 34 + .../include/ATen/ops/_is_zerotensor_ops.h | 34 + .../include/ATen/ops/_lazy_clone_ops.h | 34 + .../pytorch/include/ATen/ops/_neg_view_ops.h | 34 + .../ATen/ops/_nested_tensor_size_ops.h | 45 + .../ops/_nested_tensor_storage_offsets_ops.h | 45 + .../ATen/ops/_nested_tensor_strides_ops.h | 45 + backends/pytorch/include/ATen/ops/_nnz_ops.h | 34 + .../include/ATen/ops/_reshape_alias_ops.h | 34 + .../ATen/ops/_sparse_mask_projection_ops.h | 45 + .../pytorch/include/ATen/ops/_to_dense_ops.h | 45 + .../include/ATen/ops/_to_sparse_bsc_ops.h | 45 + .../include/ATen/ops/_to_sparse_bsr_ops.h | 45 + .../include/ATen/ops/_to_sparse_csc_ops.h | 45 + .../include/ATen/ops/_to_sparse_csr_ops.h | 45 + .../pytorch/include/ATen/ops/_to_sparse_ops.h | 67 + .../pytorch/include/ATen/ops/_values_ops.h | 34 + .../pytorch/include/ATen/ops/_version_ops.h | 34 + backends/pytorch/include/ATen/ops/abs_ops.h | 56 + .../pytorch/include/ATen/ops/absolute_ops.h | 56 + backends/pytorch/include/ATen/ops/acos_ops.h | 56 + backends/pytorch/include/ATen/ops/acosh_ops.h | 56 + backends/pytorch/include/ATen/ops/add_ops.h | 89 + .../pytorch/include/ATen/ops/addbmm_ops.h | 56 + .../pytorch/include/ATen/ops/addcdiv_ops.h | 56 + .../pytorch/include/ATen/ops/addcmul_ops.h | 56 + backends/pytorch/include/ATen/ops/addmm_ops.h | 78 + backends/pytorch/include/ATen/ops/addmv_ops.h | 56 + backends/pytorch/include/ATen/ops/addr_ops.h | 56 + .../pytorch/include/ATen/ops/adjoint_ops.h | 34 + backends/pytorch/include/ATen/ops/alias_ops.h | 34 + .../pytorch/include/ATen/ops/align_as_ops.h | 34 + .../pytorch/include/ATen/ops/align_to_ops.h | 45 + backends/pytorch/include/ATen/ops/all_ops.h | 111 + .../pytorch/include/ATen/ops/allclose_ops.h | 34 + backends/pytorch/include/ATen/ops/amax_ops.h | 45 + backends/pytorch/include/ATen/ops/amin_ops.h | 45 + .../pytorch/include/ATen/ops/aminmax_ops.h | 45 + backends/pytorch/include/ATen/ops/and_ops.h | 67 + backends/pytorch/include/ATen/ops/angle_ops.h | 45 + backends/pytorch/include/ATen/ops/any_ops.h | 111 + .../pytorch/include/ATen/ops/arccos_ops.h | 56 + .../pytorch/include/ATen/ops/arccosh_ops.h | 56 + .../pytorch/include/ATen/ops/arcsin_ops.h | 56 + .../pytorch/include/ATen/ops/arcsinh_ops.h | 56 + .../pytorch/include/ATen/ops/arctan2_ops.h | 56 + .../pytorch/include/ATen/ops/arctan_ops.h | 56 + .../pytorch/include/ATen/ops/arctanh_ops.h | 56 + .../pytorch/include/ATen/ops/argmax_ops.h | 45 + .../pytorch/include/ATen/ops/argmin_ops.h | 45 + .../pytorch/include/ATen/ops/argsort_ops.h | 67 + .../pytorch/include/ATen/ops/argwhere_ops.h | 34 + .../pytorch/include/ATen/ops/as_strided_ops.h | 45 + .../include/ATen/ops/as_strided_scatter_ops.h | 45 + backends/pytorch/include/ATen/ops/asin_ops.h | 56 + backends/pytorch/include/ATen/ops/asinh_ops.h | 56 + backends/pytorch/include/ATen/ops/atan2_ops.h | 56 + backends/pytorch/include/ATen/ops/atan_ops.h | 56 + backends/pytorch/include/ATen/ops/atanh_ops.h | 56 + .../pytorch/include/ATen/ops/baddbmm_ops.h | 78 + .../pytorch/include/ATen/ops/bernoulli_ops.h | 111 + .../pytorch/include/ATen/ops/bincount_ops.h | 45 + .../include/ATen/ops/bitwise_and_ops.h | 111 + .../include/ATen/ops/bitwise_left_shift_ops.h | 111 + .../include/ATen/ops/bitwise_not_ops.h | 56 + .../pytorch/include/ATen/ops/bitwise_or_ops.h | 111 + .../ATen/ops/bitwise_right_shift_ops.h | 111 + .../include/ATen/ops/bitwise_xor_ops.h | 111 + backends/pytorch/include/ATen/ops/bmm_ops.h | 67 + .../include/ATen/ops/broadcast_to_ops.h | 34 + .../pytorch/include/ATen/ops/cauchy_ops.h | 56 + .../include/ATen/ops/ccol_indices_ops.h | 34 + backends/pytorch/include/ATen/ops/ceil_ops.h | 56 + backends/pytorch/include/ATen/ops/chalf_ops.h | 34 + .../include/ATen/ops/cholesky_inverse_ops.h | 45 + .../pytorch/include/ATen/ops/cholesky_ops.h | 45 + .../include/ATen/ops/cholesky_solve_ops.h | 45 + backends/pytorch/include/ATen/ops/chunk_ops.h | 34 + .../pytorch/include/ATen/ops/clamp_max_ops.h | 89 + .../pytorch/include/ATen/ops/clamp_min_ops.h | 89 + backends/pytorch/include/ATen/ops/clamp_ops.h | 89 + backends/pytorch/include/ATen/ops/clip_ops.h | 89 + backends/pytorch/include/ATen/ops/clone_ops.h | 45 + .../pytorch/include/ATen/ops/coalesce_ops.h | 34 + .../include/ATen/ops/col_indices_ops.h | 34 + backends/pytorch/include/ATen/ops/conj_ops.h | 34 + .../include/ATen/ops/conj_physical_ops.h | 56 + .../pytorch/include/ATen/ops/contiguous_ops.h | 34 + backends/pytorch/include/ATen/ops/copy_ops.h | 56 + .../pytorch/include/ATen/ops/copysign_ops.h | 89 + .../pytorch/include/ATen/ops/corrcoef_ops.h | 34 + backends/pytorch/include/ATen/ops/cos_ops.h | 56 + backends/pytorch/include/ATen/ops/cosh_ops.h | 56 + .../include/ATen/ops/count_nonzero_ops.h | 67 + backends/pytorch/include/ATen/ops/cov_ops.h | 34 + backends/pytorch/include/ATen/ops/cross_ops.h | 45 + .../include/ATen/ops/crow_indices_ops.h | 34 + .../pytorch/include/ATen/ops/cummax_ops.h | 67 + .../pytorch/include/ATen/ops/cummin_ops.h | 67 + .../pytorch/include/ATen/ops/cumprod_ops.h | 89 + .../pytorch/include/ATen/ops/cumsum_ops.h | 89 + backends/pytorch/include/ATen/ops/data_ops.h | 34 + .../pytorch/include/ATen/ops/deg2rad_ops.h | 56 + .../pytorch/include/ATen/ops/dense_dim_ops.h | 34 + .../pytorch/include/ATen/ops/dequantize_ops.h | 67 + backends/pytorch/include/ATen/ops/det_ops.h | 34 + .../pytorch/include/ATen/ops/detach_ops.h | 45 + .../pytorch/include/ATen/ops/diag_embed_ops.h | 45 + backends/pytorch/include/ATen/ops/diag_ops.h | 45 + .../pytorch/include/ATen/ops/diagflat_ops.h | 34 + .../pytorch/include/ATen/ops/diagonal_ops.h | 45 + .../include/ATen/ops/diagonal_scatter_ops.h | 45 + backends/pytorch/include/ATen/ops/diff_ops.h | 45 + .../pytorch/include/ATen/ops/digamma_ops.h | 56 + backends/pytorch/include/ATen/ops/dist_ops.h | 45 + backends/pytorch/include/ATen/ops/div_ops.h | 155 + .../pytorch/include/ATen/ops/divide_ops.h | 133 + backends/pytorch/include/ATen/ops/dot_ops.h | 45 + .../pytorch/include/ATen/ops/dsplit_ops.h | 45 + backends/pytorch/include/ATen/ops/eq_ops.h | 89 + backends/pytorch/include/ATen/ops/equal_ops.h | 34 + backends/pytorch/include/ATen/ops/erf_ops.h | 56 + backends/pytorch/include/ATen/ops/erfc_ops.h | 56 + .../pytorch/include/ATen/ops/erfinv_ops.h | 56 + backends/pytorch/include/ATen/ops/exp2_ops.h | 56 + backends/pytorch/include/ATen/ops/exp_ops.h | 56 + .../pytorch/include/ATen/ops/expand_as_ops.h | 34 + .../pytorch/include/ATen/ops/expand_ops.h | 34 + backends/pytorch/include/ATen/ops/expm1_ops.h | 56 + .../include/ATen/ops/exponential_ops.h | 56 + .../include/ATen/ops/fill_diagonal_ops.h | 34 + backends/pytorch/include/ATen/ops/fill_ops.h | 89 + backends/pytorch/include/ATen/ops/fix_ops.h | 56 + .../pytorch/include/ATen/ops/flatten_ops.h | 67 + backends/pytorch/include/ATen/ops/flip_ops.h | 45 + .../pytorch/include/ATen/ops/fliplr_ops.h | 34 + .../pytorch/include/ATen/ops/flipud_ops.h | 34 + .../include/ATen/ops/float_power_ops.h | 111 + .../include/ATen/ops/floor_divide_ops.h | 89 + backends/pytorch/include/ATen/ops/floor_ops.h | 56 + backends/pytorch/include/ATen/ops/fmax_ops.h | 45 + backends/pytorch/include/ATen/ops/fmin_ops.h | 45 + backends/pytorch/include/ATen/ops/fmod_ops.h | 89 + backends/pytorch/include/ATen/ops/frac_ops.h | 56 + backends/pytorch/include/ATen/ops/frexp_ops.h | 45 + .../pytorch/include/ATen/ops/gather_ops.h | 67 + backends/pytorch/include/ATen/ops/gcd_ops.h | 56 + backends/pytorch/include/ATen/ops/ge_ops.h | 89 + .../pytorch/include/ATen/ops/geometric_ops.h | 56 + backends/pytorch/include/ATen/ops/geqrf_ops.h | 45 + backends/pytorch/include/ATen/ops/ger_ops.h | 45 + .../include/ATen/ops/greater_equal_ops.h | 89 + .../pytorch/include/ATen/ops/greater_ops.h | 89 + backends/pytorch/include/ATen/ops/gt_ops.h | 89 + .../ATen/ops/hardshrink_backward_ops.h | 45 + .../pytorch/include/ATen/ops/hardshrink_ops.h | 45 + .../include/ATen/ops/hash_tensor_ops.h | 45 + .../pytorch/include/ATen/ops/heaviside_ops.h | 56 + backends/pytorch/include/ATen/ops/histc_ops.h | 45 + .../pytorch/include/ATen/ops/histogram_ops.h | 67 + .../pytorch/include/ATen/ops/hsplit_ops.h | 45 + backends/pytorch/include/ATen/ops/hypot_ops.h | 56 + backends/pytorch/include/ATen/ops/i0_ops.h | 56 + .../pytorch/include/ATen/ops/igamma_ops.h | 56 + .../pytorch/include/ATen/ops/igammac_ops.h | 56 + .../pytorch/include/ATen/ops/index_add_ops.h | 67 + .../pytorch/include/ATen/ops/index_copy_ops.h | 78 + .../pytorch/include/ATen/ops/index_fill_ops.h | 133 + backends/pytorch/include/ATen/ops/index_ops.h | 45 + .../pytorch/include/ATen/ops/index_put_ops.h | 56 + .../include/ATen/ops/index_reduce_ops.h | 56 + .../include/ATen/ops/index_select_ops.h | 67 + .../pytorch/include/ATen/ops/indices_ops.h | 34 + backends/pytorch/include/ATen/ops/inner_ops.h | 45 + .../pytorch/include/ATen/ops/int_repr_ops.h | 45 + .../pytorch/include/ATen/ops/inverse_ops.h | 45 + .../include/ATen/ops/is_coalesced_ops.h | 34 + .../pytorch/include/ATen/ops/is_complex_ops.h | 34 + .../pytorch/include/ATen/ops/is_conj_ops.h | 34 + .../include/ATen/ops/is_distributed_ops.h | 34 + .../include/ATen/ops/is_floating_point_ops.h | 34 + .../include/ATen/ops/is_inference_ops.h | 34 + .../pytorch/include/ATen/ops/is_leaf_ops.h | 34 + .../pytorch/include/ATen/ops/is_neg_ops.h | 34 + .../pytorch/include/ATen/ops/is_nonzero_ops.h | 34 + .../pytorch/include/ATen/ops/is_pinned_ops.h | 34 + .../include/ATen/ops/is_same_size_ops.h | 34 + .../pytorch/include/ATen/ops/is_set_to_ops.h | 34 + .../pytorch/include/ATen/ops/is_signed_ops.h | 34 + .../pytorch/include/ATen/ops/isclose_ops.h | 34 + .../pytorch/include/ATen/ops/isfinite_ops.h | 34 + backends/pytorch/include/ATen/ops/isinf_ops.h | 45 + backends/pytorch/include/ATen/ops/isnan_ops.h | 45 + .../pytorch/include/ATen/ops/isneginf_ops.h | 45 + .../pytorch/include/ATen/ops/isposinf_ops.h | 45 + .../pytorch/include/ATen/ops/isreal_ops.h | 34 + backends/pytorch/include/ATen/ops/istft_ops.h | 34 + backends/pytorch/include/ATen/ops/item_ops.h | 34 + backends/pytorch/include/ATen/ops/kron_ops.h | 45 + .../pytorch/include/ATen/ops/kthvalue_ops.h | 67 + backends/pytorch/include/ATen/ops/lcm_ops.h | 56 + backends/pytorch/include/ATen/ops/ldexp_ops.h | 56 + backends/pytorch/include/ATen/ops/le_ops.h | 89 + backends/pytorch/include/ATen/ops/lerp_ops.h | 89 + .../pytorch/include/ATen/ops/less_equal_ops.h | 89 + backends/pytorch/include/ATen/ops/less_ops.h | 89 + .../pytorch/include/ATen/ops/lgamma_ops.h | 56 + backends/pytorch/include/ATen/ops/log10_ops.h | 56 + backends/pytorch/include/ATen/ops/log1p_ops.h | 56 + backends/pytorch/include/ATen/ops/log2_ops.h | 56 + .../pytorch/include/ATen/ops/log_normal_ops.h | 56 + backends/pytorch/include/ATen/ops/log_ops.h | 56 + .../include/ATen/ops/log_softmax_ops.h | 56 + .../pytorch/include/ATen/ops/logaddexp2_ops.h | 45 + .../pytorch/include/ATen/ops/logaddexp_ops.h | 45 + .../include/ATen/ops/logcumsumexp_ops.h | 67 + .../pytorch/include/ATen/ops/logdet_ops.h | 34 + .../include/ATen/ops/logical_and_ops.h | 56 + .../include/ATen/ops/logical_not_ops.h | 56 + .../pytorch/include/ATen/ops/logical_or_ops.h | 56 + .../include/ATen/ops/logical_xor_ops.h | 56 + backends/pytorch/include/ATen/ops/logit_ops.h | 56 + .../pytorch/include/ATen/ops/logsumexp_ops.h | 67 + .../pytorch/include/ATen/ops/lshift_ops.h | 89 + backends/pytorch/include/ATen/ops/lt_ops.h | 89 + .../pytorch/include/ATen/ops/lu_solve_ops.h | 45 + backends/pytorch/include/ATen/ops/mH_ops.h | 34 + backends/pytorch/include/ATen/ops/mT_ops.h | 34 + .../include/ATen/ops/masked_fill_ops.h | 89 + .../include/ATen/ops/masked_scatter_ops.h | 56 + .../include/ATen/ops/masked_select_ops.h | 45 + .../pytorch/include/ATen/ops/matmul_ops.h | 45 + .../pytorch/include/ATen/ops/matrix_H_ops.h | 34 + .../pytorch/include/ATen/ops/matrix_exp_ops.h | 34 + .../include/ATen/ops/matrix_power_ops.h | 45 + backends/pytorch/include/ATen/ops/max_ops.h | 111 + .../pytorch/include/ATen/ops/maximum_ops.h | 45 + backends/pytorch/include/ATen/ops/mean_ops.h | 89 + .../pytorch/include/ATen/ops/median_ops.h | 89 + backends/pytorch/include/ATen/ops/min_ops.h | 111 + .../pytorch/include/ATen/ops/minimum_ops.h | 45 + backends/pytorch/include/ATen/ops/mm_ops.h | 67 + backends/pytorch/include/ATen/ops/mode_ops.h | 67 + .../pytorch/include/ATen/ops/moveaxis_ops.h | 45 + .../pytorch/include/ATen/ops/movedim_ops.h | 45 + backends/pytorch/include/ATen/ops/msort_ops.h | 45 + backends/pytorch/include/ATen/ops/mul_ops.h | 89 + .../include/ATen/ops/multinomial_ops.h | 45 + .../pytorch/include/ATen/ops/multiply_ops.h | 78 + backends/pytorch/include/ATen/ops/mv_ops.h | 45 + .../pytorch/include/ATen/ops/mvlgamma_ops.h | 56 + .../pytorch/include/ATen/ops/nan_to_num_ops.h | 56 + .../pytorch/include/ATen/ops/nanmean_ops.h | 45 + .../pytorch/include/ATen/ops/nanmedian_ops.h | 89 + .../include/ATen/ops/nanquantile_ops.h | 67 + .../pytorch/include/ATen/ops/nansum_ops.h | 45 + .../include/ATen/ops/narrow_copy_ops.h | 45 + .../pytorch/include/ATen/ops/narrow_ops.h | 45 + backends/pytorch/include/ATen/ops/ne_ops.h | 89 + backends/pytorch/include/ATen/ops/neg_ops.h | 56 + .../pytorch/include/ATen/ops/negative_ops.h | 56 + .../pytorch/include/ATen/ops/new_empty_ops.h | 45 + .../include/ATen/ops/new_empty_strided_ops.h | 45 + .../pytorch/include/ATen/ops/new_full_ops.h | 45 + .../pytorch/include/ATen/ops/new_ones_ops.h | 45 + .../pytorch/include/ATen/ops/new_zeros_ops.h | 45 + .../pytorch/include/ATen/ops/nextafter_ops.h | 56 + .../include/ATen/ops/nonzero_numpy_ops.h | 34 + .../pytorch/include/ATen/ops/nonzero_ops.h | 45 + .../include/ATen/ops/nonzero_static_ops.h | 45 + backends/pytorch/include/ATen/ops/norm_ops.h | 155 + .../pytorch/include/ATen/ops/normal_ops.h | 144 + .../pytorch/include/ATen/ops/not_equal_ops.h | 89 + .../pytorch/include/ATen/ops/numpy_T_ops.h | 34 + backends/pytorch/include/ATen/ops/or_ops.h | 67 + backends/pytorch/include/ATen/ops/orgqr_ops.h | 45 + backends/pytorch/include/ATen/ops/ormqr_ops.h | 45 + backends/pytorch/include/ATen/ops/outer_ops.h | 45 + .../pytorch/include/ATen/ops/output_nr_ops.h | 34 + .../pytorch/include/ATen/ops/permute_ops.h | 34 + .../pytorch/include/ATen/ops/pin_memory_ops.h | 34 + .../pytorch/include/ATen/ops/pinverse_ops.h | 34 + .../pytorch/include/ATen/ops/polygamma_ops.h | 56 + .../pytorch/include/ATen/ops/positive_ops.h | 34 + backends/pytorch/include/ATen/ops/pow_ops.h | 111 + backends/pytorch/include/ATen/ops/prelu_ops.h | 34 + backends/pytorch/include/ATen/ops/prod_ops.h | 89 + backends/pytorch/include/ATen/ops/put_ops.h | 56 + .../include/ATen/ops/q_per_channel_axis_ops.h | 34 + .../ATen/ops/q_per_channel_scales_ops.h | 45 + .../ATen/ops/q_per_channel_zero_points_ops.h | 45 + .../pytorch/include/ATen/ops/q_scale_ops.h | 34 + .../include/ATen/ops/q_zero_point_ops.h | 34 + backends/pytorch/include/ATen/ops/qr_ops.h | 45 + .../pytorch/include/ATen/ops/qscheme_ops.h | 34 + .../pytorch/include/ATen/ops/quantile_ops.h | 67 + .../pytorch/include/ATen/ops/rad2deg_ops.h | 56 + .../pytorch/include/ATen/ops/random_ops.h | 122 + backends/pytorch/include/ATen/ops/ravel_ops.h | 34 + .../pytorch/include/ATen/ops/reciprocal_ops.h | 56 + .../include/ATen/ops/record_stream_ops.h | 34 + .../include/ATen/ops/refine_names_ops.h | 34 + backends/pytorch/include/ATen/ops/relu_ops.h | 56 + .../pytorch/include/ATen/ops/remainder_ops.h | 111 + .../pytorch/include/ATen/ops/rename_ops.h | 45 + .../pytorch/include/ATen/ops/renorm_ops.h | 56 + .../include/ATen/ops/repeat_interleave_ops.h | 67 + .../pytorch/include/ATen/ops/repeat_ops.h | 45 + .../include/ATen/ops/requires_grad_ops.h | 34 + .../pytorch/include/ATen/ops/reshape_as_ops.h | 34 + .../pytorch/include/ATen/ops/reshape_ops.h | 34 + .../pytorch/include/ATen/ops/resize_as_ops.h | 56 + .../include/ATen/ops/resize_as_sparse_ops.h | 56 + .../pytorch/include/ATen/ops/resize_ops.h | 56 + .../include/ATen/ops/resolve_conj_ops.h | 34 + .../include/ATen/ops/resolve_neg_ops.h | 34 + .../include/ATen/ops/retain_grad_ops.h | 34 + .../include/ATen/ops/retains_grad_ops.h | 34 + backends/pytorch/include/ATen/ops/roll_ops.h | 45 + backends/pytorch/include/ATen/ops/rot90_ops.h | 45 + backends/pytorch/include/ATen/ops/round_ops.h | 89 + .../include/ATen/ops/row_indices_ops.h | 34 + .../pytorch/include/ATen/ops/rshift_ops.h | 89 + backends/pytorch/include/ATen/ops/rsqrt_ops.h | 56 + .../include/ATen/ops/scatter_add_ops.h | 67 + .../pytorch/include/ATen/ops/scatter_ops.h | 177 + .../include/ATen/ops/scatter_reduce_ops.h | 56 + .../pytorch/include/ATen/ops/select_ops.h | 45 + .../include/ATen/ops/select_scatter_ops.h | 45 + .../pytorch/include/ATen/ops/set_data_ops.h | 34 + backends/pytorch/include/ATen/ops/set_ops.h | 166 + backends/pytorch/include/ATen/ops/sgn_ops.h | 56 + .../pytorch/include/ATen/ops/sigmoid_ops.h | 56 + backends/pytorch/include/ATen/ops/sign_ops.h | 56 + .../pytorch/include/ATen/ops/signbit_ops.h | 45 + backends/pytorch/include/ATen/ops/sin_ops.h | 56 + backends/pytorch/include/ATen/ops/sinc_ops.h | 56 + backends/pytorch/include/ATen/ops/sinh_ops.h | 56 + backends/pytorch/include/ATen/ops/size_ops.h | 45 + .../include/ATen/ops/slice_inverse_ops.h | 34 + backends/pytorch/include/ATen/ops/slice_ops.h | 34 + .../include/ATen/ops/slice_scatter_ops.h | 45 + .../pytorch/include/ATen/ops/slogdet_ops.h | 45 + backends/pytorch/include/ATen/ops/smm_ops.h | 34 + .../pytorch/include/ATen/ops/softmax_ops.h | 56 + backends/pytorch/include/ATen/ops/sort_ops.h | 111 + .../pytorch/include/ATen/ops/sparse_dim_ops.h | 34 + .../include/ATen/ops/sparse_mask_ops.h | 45 + .../ATen/ops/sparse_resize_and_clear_ops.h | 56 + .../include/ATen/ops/sparse_resize_ops.h | 56 + backends/pytorch/include/ATen/ops/split_ops.h | 45 + .../include/ATen/ops/split_with_sizes_ops.h | 34 + backends/pytorch/include/ATen/ops/sqrt_ops.h | 56 + .../pytorch/include/ATen/ops/square_ops.h | 56 + .../pytorch/include/ATen/ops/squeeze_ops.h | 111 + .../pytorch/include/ATen/ops/sspaddmm_ops.h | 45 + backends/pytorch/include/ATen/ops/std_ops.h | 122 + backends/pytorch/include/ATen/ops/stft_ops.h | 45 + .../pytorch/include/ATen/ops/stride_ops.h | 45 + backends/pytorch/include/ATen/ops/sub_ops.h | 89 + .../pytorch/include/ATen/ops/subtract_ops.h | 78 + backends/pytorch/include/ATen/ops/sum_ops.h | 89 + .../include/ATen/ops/sum_to_size_ops.h | 34 + backends/pytorch/include/ATen/ops/svd_ops.h | 45 + .../pytorch/include/ATen/ops/swapaxes_ops.h | 45 + .../pytorch/include/ATen/ops/swapdims_ops.h | 45 + backends/pytorch/include/ATen/ops/t_ops.h | 45 + .../include/ATen/ops/take_along_dim_ops.h | 45 + backends/pytorch/include/ATen/ops/take_ops.h | 45 + backends/pytorch/include/ATen/ops/tan_ops.h | 56 + backends/pytorch/include/ATen/ops/tanh_ops.h | 56 + .../include/ATen/ops/tensor_split_ops.h | 56 + backends/pytorch/include/ATen/ops/tile_ops.h | 34 + .../pytorch/include/ATen/ops/to_dense_ops.h | 34 + .../pytorch/include/ATen/ops/to_mkldnn_ops.h | 45 + backends/pytorch/include/ATen/ops/to_ops.h | 67 + .../include/ATen/ops/to_padded_tensor_ops.h | 45 + .../include/ATen/ops/to_sparse_bsc_ops.h | 34 + .../include/ATen/ops/to_sparse_bsr_ops.h | 34 + .../include/ATen/ops/to_sparse_csc_ops.h | 34 + .../include/ATen/ops/to_sparse_csr_ops.h | 34 + .../pytorch/include/ATen/ops/to_sparse_ops.h | 45 + backends/pytorch/include/ATen/ops/topk_ops.h | 45 + backends/pytorch/include/ATen/ops/trace_ops.h | 45 + .../pytorch/include/ATen/ops/transpose_ops.h | 56 + .../include/ATen/ops/triangular_solve_ops.h | 45 + backends/pytorch/include/ATen/ops/tril_ops.h | 56 + backends/pytorch/include/ATen/ops/triu_ops.h | 56 + .../include/ATen/ops/true_divide_ops.h | 78 + backends/pytorch/include/ATen/ops/trunc_ops.h | 56 + .../pytorch/include/ATen/ops/type_as_ops.h | 34 + .../pytorch/include/ATen/ops/unbind_ops.h | 45 + .../pytorch/include/ATen/ops/unflatten_ops.h | 45 + .../pytorch/include/ATen/ops/unfold_ops.h | 34 + .../pytorch/include/ATen/ops/uniform_ops.h | 56 + .../include/ATen/ops/unsafe_chunk_ops.h | 34 + .../include/ATen/ops/unsafe_split_ops.h | 45 + .../ATen/ops/unsafe_split_with_sizes_ops.h | 45 + .../pytorch/include/ATen/ops/unsqueeze_ops.h | 45 + .../pytorch/include/ATen/ops/values_ops.h | 34 + backends/pytorch/include/ATen/ops/var_ops.h | 122 + backends/pytorch/include/ATen/ops/vdot_ops.h | 45 + .../pytorch/include/ATen/ops/view_as_ops.h | 34 + backends/pytorch/include/ATen/ops/view_ops.h | 45 + .../pytorch/include/ATen/ops/vsplit_ops.h | 45 + backends/pytorch/include/ATen/ops/where_ops.h | 89 + backends/pytorch/include/ATen/ops/xlogy_ops.h | 111 + backends/pytorch/include/ATen/ops/xor_ops.h | 67 + backends/pytorch/include/ATen/ops/zero_ops.h | 56 + .../pytorch/include/ATen/record_function.h | 799 +++ backends/pytorch/include/c10/core/Allocator.h | 455 ++ .../pytorch/include/c10/core/AutogradState.h | 90 + backends/pytorch/include/c10/core/Backend.h | 414 ++ .../pytorch/include/c10/core/DefaultDtype.h | 20 + backends/pytorch/include/c10/core/Device.h | 221 + .../include/c10/core/DeviceCapability.h | 81 + .../pytorch/include/c10/core/DeviceGuard.h | 207 + .../pytorch/include/c10/core/DeviceType.h | 35 + .../pytorch/include/c10/core/DispatchKey.h | 750 +++ .../pytorch/include/c10/core/DispatchKeySet.h | 977 +++ backends/pytorch/include/c10/core/Event.h | 142 + .../pytorch/include/c10/core/GeneratorImpl.h | 116 + .../pytorch/include/c10/core/InferenceMode.h | 96 + backends/pytorch/include/c10/core/Layout.h | 67 + .../pytorch/include/c10/core/MemoryFormat.h | 268 + .../pytorch/include/c10/core/OptionalRef.h | 36 + backends/pytorch/include/c10/core/QScheme.h | 60 + .../pytorch/include/c10/core/SafePyObject.h | 125 + backends/pytorch/include/c10/core/Scalar.h | 471 ++ .../pytorch/include/c10/core/ScalarType.h | 285 + .../include/c10/core/ScalarTypeToTypeMeta.h | 62 + backends/pytorch/include/c10/core/Storage.h | 293 + .../pytorch/include/c10/core/StorageImpl.h | 398 ++ backends/pytorch/include/c10/core/Stream.h | 182 + .../pytorch/include/c10/core/StreamGuard.h | 178 + backends/pytorch/include/c10/core/SymBool.h | 184 + backends/pytorch/include/c10/core/SymFloat.h | 123 + backends/pytorch/include/c10/core/SymInt.h | 586 ++ .../pytorch/include/c10/core/SymIntArrayRef.h | 113 + .../pytorch/include/c10/core/SymNodeImpl.h | 261 + .../include/c10/core/SymbolicShapeMeta.h | 234 + .../pytorch/include/c10/core/TensorImpl.h | 3333 ++++++++++ .../pytorch/include/c10/core/TensorOptions.h | 791 +++ .../include/c10/core/UndefinedTensorImpl.h | 54 + .../pytorch/include/c10/core/WrapDimMinimal.h | 53 + backends/pytorch/include/c10/core/alignment.h | 35 + backends/pytorch/include/c10/core/impl/COW.h | 37 + .../include/c10/core/impl/COWDeleter.h | 71 + .../c10/core/impl/DeviceGuardImplInterface.h | 417 ++ .../c10/core/impl/HermeticPyObjectTLS.h | 67 + .../include/c10/core/impl/InlineDeviceGuard.h | 438 ++ .../include/c10/core/impl/InlineEvent.h | 152 + .../include/c10/core/impl/InlineStreamGuard.h | 265 + .../c10/core/impl/LocalDispatchKeySet.h | 174 + .../include/c10/core/impl/PyInterpreter.h | 257 + .../c10/core/impl/PyInterpreterHooks.h | 45 + .../include/c10/core/impl/PyObjectSlot.h | 70 + .../include/c10/core/impl/SizesAndStrides.h | 336 + .../c10/core/impl/TorchDispatchModeTLS.h | 72 + .../include/c10/core/impl/VirtualGuardImpl.h | 117 + backends/pytorch/include/c10/macros/Export.h | 6 + backends/pytorch/include/c10/macros/Macros.h | 6 + backends/pytorch/include/c10/util/AlignOf.h | 181 + backends/pytorch/include/c10/util/ArrayRef.h | 326 + backends/pytorch/include/c10/util/BFloat16.h | 6 + backends/pytorch/include/c10/util/Backtrace.h | 36 + backends/pytorch/include/c10/util/C++17.h | 75 + .../pytorch/include/c10/util/ConstexprCrc.h | 137 + .../pytorch/include/c10/util/Deprecated.h | 7 + backends/pytorch/include/c10/util/DimVector.h | 22 + backends/pytorch/include/c10/util/Exception.h | 875 +++ .../include/c10/util/ExclusivelyOwned.h | 145 + .../c10/util/ExclusivelyOwnedTensorTraits.h | 80 + backends/pytorch/include/c10/util/Flags.h | 247 + .../include/c10/util/Float4_e2m1fn_x2.h | 6 + .../pytorch/include/c10/util/Float8_e4m3fn.h | 6 + .../include/c10/util/Float8_e4m3fnuz.h | 6 + .../pytorch/include/c10/util/Float8_e5m2.h | 6 + .../include/c10/util/Float8_e5m2fnuz.h | 6 + .../pytorch/include/c10/util/Float8_e8m0fnu.h | 6 + .../pytorch/include/c10/util/FunctionRef.h | 80 + backends/pytorch/include/c10/util/Half.h | 13 + backends/pytorch/include/c10/util/IdWrapper.h | 82 + backends/pytorch/include/c10/util/Lazy.h | 125 + backends/pytorch/include/c10/util/Logging.h | 378 ++ .../pytorch/include/c10/util/MaybeOwned.h | 242 + .../include/c10/util/Metaprogramming.h | 6 + backends/pytorch/include/c10/util/Optional.h | 65 + .../include/c10/util/OptionalArrayRef.h | 242 + backends/pytorch/include/c10/util/Registry.h | 334 + .../pytorch/include/c10/util/SmallVector.h | 1472 +++++ .../pytorch/include/c10/util/StringUtil.h | 267 + .../include/c10/util/ThreadLocalDebugInfo.h | 90 + backends/pytorch/include/c10/util/Type.h | 35 + backends/pytorch/include/c10/util/TypeCast.h | 215 + backends/pytorch/include/c10/util/TypeIndex.h | 132 + backends/pytorch/include/c10/util/TypeList.h | 6 + .../include/c10/util/TypeSafeSignMath.h | 6 + .../pytorch/include/c10/util/TypeTraits.h | 6 + .../pytorch/include/c10/util/UniqueVoidPtr.h | 145 + .../pytorch/include/c10/util/accumulate.h | 129 + backends/pytorch/include/c10/util/bit_cast.h | 6 + backends/pytorch/include/c10/util/bits.h | 6 + backends/pytorch/include/c10/util/complex.h | 83 + .../pytorch/include/c10/util/complex_math.h | 411 ++ .../pytorch/include/c10/util/complex_utils.h | 51 + backends/pytorch/include/c10/util/hash.h | 384 ++ .../pytorch/include/c10/util/intrusive_ptr.h | 1278 ++++ backends/pytorch/include/c10/util/irange.h | 128 + .../pytorch/include/c10/util/llvmMathExtras.h | 910 +++ .../pytorch/include/c10/util/logging_common.h | 79 + .../c10/util/logging_is_not_google_glog.h | 186 + .../c10/util/order_preserving_flat_hash_map.h | 2222 +++++++ backends/pytorch/include/c10/util/overflows.h | 105 + .../pytorch/include/c10/util/python_stub.h | 9 + backends/pytorch/include/c10/util/qint32.h | 6 + backends/pytorch/include/c10/util/qint8.h | 6 + backends/pytorch/include/c10/util/quint2x4.h | 6 + backends/pytorch/include/c10/util/quint4x2.h | 6 + backends/pytorch/include/c10/util/quint8.h | 6 + .../pytorch/include/c10/util/safe_numerics.h | 119 + .../pytorch/include/c10/util/string_utils.h | 27 + .../pytorch/include/c10/util/string_view.h | 648 ++ backends/pytorch/include/c10/util/typeid.h | 720 ++ .../torch/headeronly/core/DeviceType.h | 128 + .../include/torch/headeronly/core/Layout.h | 44 + .../torch/headeronly/core/MemoryFormat.h | 46 + .../torch/headeronly/core/ScalarType.h | 381 ++ .../torch/headeronly/core/TensorAccessor.h | 462 ++ .../include/torch/headeronly/macros/Export.h | 153 + .../include/torch/headeronly/macros/Macros.h | 694 ++ .../torch/headeronly/macros/cmake_macros.h | 14 + .../include/torch/headeronly/util/BFloat16.h | 480 ++ .../torch/headeronly/util/Deprecated.h | 102 + .../include/torch/headeronly/util/Exception.h | 83 + .../torch/headeronly/util/Float4_e2m1fn_x2.h | 47 + .../torch/headeronly/util/Float8_e4m3fn.h | 531 ++ .../torch/headeronly/util/Float8_e4m3fnuz.h | 444 ++ .../torch/headeronly/util/Float8_e5m2.h | 458 ++ .../torch/headeronly/util/Float8_e5m2fnuz.h | 448 ++ .../torch/headeronly/util/Float8_e8m0fnu.h | 226 + .../torch/headeronly/util/Float8_fnuz_cvt.h | 69 + .../include/torch/headeronly/util/Half.h | 788 +++ .../headeronly/util/HeaderOnlyArrayRef.h | 248 + .../torch/headeronly/util/Metaprogramming.h | 237 + .../include/torch/headeronly/util/TypeList.h | 548 ++ .../torch/headeronly/util/TypeSafeSignMath.h | 148 + .../torch/headeronly/util/TypeTraits.h | 164 + .../include/torch/headeronly/util/bit_cast.h | 50 + .../include/torch/headeronly/util/bits.h | 71 + .../include/torch/headeronly/util/complex.h | 616 ++ .../headeronly/util/floating_point_utils.h | 38 + .../include/torch/headeronly/util/qint32.h | 22 + .../include/torch/headeronly/util/qint8.h | 24 + .../include/torch/headeronly/util/quint2x4.h | 23 + .../include/torch/headeronly/util/quint4x2.h | 23 + .../include/torch/headeronly/util/quint8.h | 22 + 605 files changed, 78930 insertions(+) create mode 100644 backends/pytorch/include/ATen/MethodOperators.h create mode 100644 backends/pytorch/include/ATen/StorageUtils.h create mode 100644 backends/pytorch/include/ATen/core/ATen_fwd.h create mode 100644 backends/pytorch/include/ATen/core/CheckMemoryFormat.h create mode 100644 backends/pytorch/include/ATen/core/DeprecatedTypeProperties.h create mode 100644 backends/pytorch/include/ATen/core/DeprecatedTypePropertiesRegistry.h create mode 100644 backends/pytorch/include/ATen/core/Dict.h create mode 100644 backends/pytorch/include/ATen/core/Dict_inl.h create mode 100644 backends/pytorch/include/ATen/core/DimVector.h create mode 100644 backends/pytorch/include/ATen/core/Dimname.h create mode 100644 backends/pytorch/include/ATen/core/Generator.h create mode 100644 backends/pytorch/include/ATen/core/IListRef.h create mode 100644 backends/pytorch/include/ATen/core/IListRef_inl.h create mode 100644 backends/pytorch/include/ATen/core/List.h create mode 100644 backends/pytorch/include/ATen/core/List_inl.h create mode 100644 backends/pytorch/include/ATen/core/NamedTensor.h create mode 100644 backends/pytorch/include/ATen/core/QuantizerBase.h create mode 100644 backends/pytorch/include/ATen/core/Tensor.h create mode 100644 backends/pytorch/include/ATen/core/TensorAccessor.h create mode 100644 backends/pytorch/include/ATen/core/TensorBase.h create mode 100644 backends/pytorch/include/ATen/core/TensorBody.h create mode 100644 backends/pytorch/include/ATen/core/blob.h create mode 100644 backends/pytorch/include/ATen/core/custom_class.h create mode 100644 backends/pytorch/include/ATen/core/dynamic_type.h create mode 100644 backends/pytorch/include/ATen/core/functional.h create mode 100644 backends/pytorch/include/ATen/core/ivalue.h create mode 100644 backends/pytorch/include/ATen/core/ivalue_inl.h create mode 100644 backends/pytorch/include/ATen/core/ivalue_to.h create mode 100644 backends/pytorch/include/ATen/core/jit_type.h create mode 100644 backends/pytorch/include/ATen/core/jit_type_base.h create mode 100644 backends/pytorch/include/ATen/core/operator_name.h create mode 100644 backends/pytorch/include/ATen/core/qualified_name.h create mode 100644 backends/pytorch/include/ATen/core/rref_interface.h create mode 100644 backends/pytorch/include/ATen/core/symbol.h create mode 100644 backends/pytorch/include/ATen/core/type_factory.h create mode 100644 backends/pytorch/include/ATen/core/type_ptr.h create mode 100644 backends/pytorch/include/ATen/ops/_addmm_activation_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_autocast_to_full_precision_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_autocast_to_reduced_precision_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_backward_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_coalesced_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_conj_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_conj_physical_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_dimI_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_dimV_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_fw_primal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_indices_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_is_all_true_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_is_any_true_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_is_zerotensor_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_lazy_clone_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_neg_view_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_nested_tensor_size_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_nested_tensor_storage_offsets_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_nested_tensor_strides_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_nnz_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_reshape_alias_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_sparse_mask_projection_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_to_dense_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_to_sparse_bsc_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_to_sparse_bsr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_to_sparse_csc_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_to_sparse_csr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_to_sparse_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_values_ops.h create mode 100644 backends/pytorch/include/ATen/ops/_version_ops.h create mode 100644 backends/pytorch/include/ATen/ops/abs_ops.h create mode 100644 backends/pytorch/include/ATen/ops/absolute_ops.h create mode 100644 backends/pytorch/include/ATen/ops/acos_ops.h create mode 100644 backends/pytorch/include/ATen/ops/acosh_ops.h create mode 100644 backends/pytorch/include/ATen/ops/add_ops.h create mode 100644 backends/pytorch/include/ATen/ops/addbmm_ops.h create mode 100644 backends/pytorch/include/ATen/ops/addcdiv_ops.h create mode 100644 backends/pytorch/include/ATen/ops/addcmul_ops.h create mode 100644 backends/pytorch/include/ATen/ops/addmm_ops.h create mode 100644 backends/pytorch/include/ATen/ops/addmv_ops.h create mode 100644 backends/pytorch/include/ATen/ops/addr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/adjoint_ops.h create mode 100644 backends/pytorch/include/ATen/ops/alias_ops.h create mode 100644 backends/pytorch/include/ATen/ops/align_as_ops.h create mode 100644 backends/pytorch/include/ATen/ops/align_to_ops.h create mode 100644 backends/pytorch/include/ATen/ops/all_ops.h create mode 100644 backends/pytorch/include/ATen/ops/allclose_ops.h create mode 100644 backends/pytorch/include/ATen/ops/amax_ops.h create mode 100644 backends/pytorch/include/ATen/ops/amin_ops.h create mode 100644 backends/pytorch/include/ATen/ops/aminmax_ops.h create mode 100644 backends/pytorch/include/ATen/ops/and_ops.h create mode 100644 backends/pytorch/include/ATen/ops/angle_ops.h create mode 100644 backends/pytorch/include/ATen/ops/any_ops.h create mode 100644 backends/pytorch/include/ATen/ops/arccos_ops.h create mode 100644 backends/pytorch/include/ATen/ops/arccosh_ops.h create mode 100644 backends/pytorch/include/ATen/ops/arcsin_ops.h create mode 100644 backends/pytorch/include/ATen/ops/arcsinh_ops.h create mode 100644 backends/pytorch/include/ATen/ops/arctan2_ops.h create mode 100644 backends/pytorch/include/ATen/ops/arctan_ops.h create mode 100644 backends/pytorch/include/ATen/ops/arctanh_ops.h create mode 100644 backends/pytorch/include/ATen/ops/argmax_ops.h create mode 100644 backends/pytorch/include/ATen/ops/argmin_ops.h create mode 100644 backends/pytorch/include/ATen/ops/argsort_ops.h create mode 100644 backends/pytorch/include/ATen/ops/argwhere_ops.h create mode 100644 backends/pytorch/include/ATen/ops/as_strided_ops.h create mode 100644 backends/pytorch/include/ATen/ops/as_strided_scatter_ops.h create mode 100644 backends/pytorch/include/ATen/ops/asin_ops.h create mode 100644 backends/pytorch/include/ATen/ops/asinh_ops.h create mode 100644 backends/pytorch/include/ATen/ops/atan2_ops.h create mode 100644 backends/pytorch/include/ATen/ops/atan_ops.h create mode 100644 backends/pytorch/include/ATen/ops/atanh_ops.h create mode 100644 backends/pytorch/include/ATen/ops/baddbmm_ops.h create mode 100644 backends/pytorch/include/ATen/ops/bernoulli_ops.h create mode 100644 backends/pytorch/include/ATen/ops/bincount_ops.h create mode 100644 backends/pytorch/include/ATen/ops/bitwise_and_ops.h create mode 100644 backends/pytorch/include/ATen/ops/bitwise_left_shift_ops.h create mode 100644 backends/pytorch/include/ATen/ops/bitwise_not_ops.h create mode 100644 backends/pytorch/include/ATen/ops/bitwise_or_ops.h create mode 100644 backends/pytorch/include/ATen/ops/bitwise_right_shift_ops.h create mode 100644 backends/pytorch/include/ATen/ops/bitwise_xor_ops.h create mode 100644 backends/pytorch/include/ATen/ops/bmm_ops.h create mode 100644 backends/pytorch/include/ATen/ops/broadcast_to_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cauchy_ops.h create mode 100644 backends/pytorch/include/ATen/ops/ccol_indices_ops.h create mode 100644 backends/pytorch/include/ATen/ops/ceil_ops.h create mode 100644 backends/pytorch/include/ATen/ops/chalf_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cholesky_inverse_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cholesky_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cholesky_solve_ops.h create mode 100644 backends/pytorch/include/ATen/ops/chunk_ops.h create mode 100644 backends/pytorch/include/ATen/ops/clamp_max_ops.h create mode 100644 backends/pytorch/include/ATen/ops/clamp_min_ops.h create mode 100644 backends/pytorch/include/ATen/ops/clamp_ops.h create mode 100644 backends/pytorch/include/ATen/ops/clip_ops.h create mode 100644 backends/pytorch/include/ATen/ops/clone_ops.h create mode 100644 backends/pytorch/include/ATen/ops/coalesce_ops.h create mode 100644 backends/pytorch/include/ATen/ops/col_indices_ops.h create mode 100644 backends/pytorch/include/ATen/ops/conj_ops.h create mode 100644 backends/pytorch/include/ATen/ops/conj_physical_ops.h create mode 100644 backends/pytorch/include/ATen/ops/contiguous_ops.h create mode 100644 backends/pytorch/include/ATen/ops/copy_ops.h create mode 100644 backends/pytorch/include/ATen/ops/copysign_ops.h create mode 100644 backends/pytorch/include/ATen/ops/corrcoef_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cos_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cosh_ops.h create mode 100644 backends/pytorch/include/ATen/ops/count_nonzero_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cov_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cross_ops.h create mode 100644 backends/pytorch/include/ATen/ops/crow_indices_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cummax_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cummin_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cumprod_ops.h create mode 100644 backends/pytorch/include/ATen/ops/cumsum_ops.h create mode 100644 backends/pytorch/include/ATen/ops/data_ops.h create mode 100644 backends/pytorch/include/ATen/ops/deg2rad_ops.h create mode 100644 backends/pytorch/include/ATen/ops/dense_dim_ops.h create mode 100644 backends/pytorch/include/ATen/ops/dequantize_ops.h create mode 100644 backends/pytorch/include/ATen/ops/det_ops.h create mode 100644 backends/pytorch/include/ATen/ops/detach_ops.h create mode 100644 backends/pytorch/include/ATen/ops/diag_embed_ops.h create mode 100644 backends/pytorch/include/ATen/ops/diag_ops.h create mode 100644 backends/pytorch/include/ATen/ops/diagflat_ops.h create mode 100644 backends/pytorch/include/ATen/ops/diagonal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/diagonal_scatter_ops.h create mode 100644 backends/pytorch/include/ATen/ops/diff_ops.h create mode 100644 backends/pytorch/include/ATen/ops/digamma_ops.h create mode 100644 backends/pytorch/include/ATen/ops/dist_ops.h create mode 100644 backends/pytorch/include/ATen/ops/div_ops.h create mode 100644 backends/pytorch/include/ATen/ops/divide_ops.h create mode 100644 backends/pytorch/include/ATen/ops/dot_ops.h create mode 100644 backends/pytorch/include/ATen/ops/dsplit_ops.h create mode 100644 backends/pytorch/include/ATen/ops/eq_ops.h create mode 100644 backends/pytorch/include/ATen/ops/equal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/erf_ops.h create mode 100644 backends/pytorch/include/ATen/ops/erfc_ops.h create mode 100644 backends/pytorch/include/ATen/ops/erfinv_ops.h create mode 100644 backends/pytorch/include/ATen/ops/exp2_ops.h create mode 100644 backends/pytorch/include/ATen/ops/exp_ops.h create mode 100644 backends/pytorch/include/ATen/ops/expand_as_ops.h create mode 100644 backends/pytorch/include/ATen/ops/expand_ops.h create mode 100644 backends/pytorch/include/ATen/ops/expm1_ops.h create mode 100644 backends/pytorch/include/ATen/ops/exponential_ops.h create mode 100644 backends/pytorch/include/ATen/ops/fill_diagonal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/fill_ops.h create mode 100644 backends/pytorch/include/ATen/ops/fix_ops.h create mode 100644 backends/pytorch/include/ATen/ops/flatten_ops.h create mode 100644 backends/pytorch/include/ATen/ops/flip_ops.h create mode 100644 backends/pytorch/include/ATen/ops/fliplr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/flipud_ops.h create mode 100644 backends/pytorch/include/ATen/ops/float_power_ops.h create mode 100644 backends/pytorch/include/ATen/ops/floor_divide_ops.h create mode 100644 backends/pytorch/include/ATen/ops/floor_ops.h create mode 100644 backends/pytorch/include/ATen/ops/fmax_ops.h create mode 100644 backends/pytorch/include/ATen/ops/fmin_ops.h create mode 100644 backends/pytorch/include/ATen/ops/fmod_ops.h create mode 100644 backends/pytorch/include/ATen/ops/frac_ops.h create mode 100644 backends/pytorch/include/ATen/ops/frexp_ops.h create mode 100644 backends/pytorch/include/ATen/ops/gather_ops.h create mode 100644 backends/pytorch/include/ATen/ops/gcd_ops.h create mode 100644 backends/pytorch/include/ATen/ops/ge_ops.h create mode 100644 backends/pytorch/include/ATen/ops/geometric_ops.h create mode 100644 backends/pytorch/include/ATen/ops/geqrf_ops.h create mode 100644 backends/pytorch/include/ATen/ops/ger_ops.h create mode 100644 backends/pytorch/include/ATen/ops/greater_equal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/greater_ops.h create mode 100644 backends/pytorch/include/ATen/ops/gt_ops.h create mode 100644 backends/pytorch/include/ATen/ops/hardshrink_backward_ops.h create mode 100644 backends/pytorch/include/ATen/ops/hardshrink_ops.h create mode 100644 backends/pytorch/include/ATen/ops/hash_tensor_ops.h create mode 100644 backends/pytorch/include/ATen/ops/heaviside_ops.h create mode 100644 backends/pytorch/include/ATen/ops/histc_ops.h create mode 100644 backends/pytorch/include/ATen/ops/histogram_ops.h create mode 100644 backends/pytorch/include/ATen/ops/hsplit_ops.h create mode 100644 backends/pytorch/include/ATen/ops/hypot_ops.h create mode 100644 backends/pytorch/include/ATen/ops/i0_ops.h create mode 100644 backends/pytorch/include/ATen/ops/igamma_ops.h create mode 100644 backends/pytorch/include/ATen/ops/igammac_ops.h create mode 100644 backends/pytorch/include/ATen/ops/index_add_ops.h create mode 100644 backends/pytorch/include/ATen/ops/index_copy_ops.h create mode 100644 backends/pytorch/include/ATen/ops/index_fill_ops.h create mode 100644 backends/pytorch/include/ATen/ops/index_ops.h create mode 100644 backends/pytorch/include/ATen/ops/index_put_ops.h create mode 100644 backends/pytorch/include/ATen/ops/index_reduce_ops.h create mode 100644 backends/pytorch/include/ATen/ops/index_select_ops.h create mode 100644 backends/pytorch/include/ATen/ops/indices_ops.h create mode 100644 backends/pytorch/include/ATen/ops/inner_ops.h create mode 100644 backends/pytorch/include/ATen/ops/int_repr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/inverse_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_coalesced_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_complex_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_conj_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_distributed_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_floating_point_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_inference_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_leaf_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_neg_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_nonzero_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_pinned_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_same_size_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_set_to_ops.h create mode 100644 backends/pytorch/include/ATen/ops/is_signed_ops.h create mode 100644 backends/pytorch/include/ATen/ops/isclose_ops.h create mode 100644 backends/pytorch/include/ATen/ops/isfinite_ops.h create mode 100644 backends/pytorch/include/ATen/ops/isinf_ops.h create mode 100644 backends/pytorch/include/ATen/ops/isnan_ops.h create mode 100644 backends/pytorch/include/ATen/ops/isneginf_ops.h create mode 100644 backends/pytorch/include/ATen/ops/isposinf_ops.h create mode 100644 backends/pytorch/include/ATen/ops/isreal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/istft_ops.h create mode 100644 backends/pytorch/include/ATen/ops/item_ops.h create mode 100644 backends/pytorch/include/ATen/ops/kron_ops.h create mode 100644 backends/pytorch/include/ATen/ops/kthvalue_ops.h create mode 100644 backends/pytorch/include/ATen/ops/lcm_ops.h create mode 100644 backends/pytorch/include/ATen/ops/ldexp_ops.h create mode 100644 backends/pytorch/include/ATen/ops/le_ops.h create mode 100644 backends/pytorch/include/ATen/ops/lerp_ops.h create mode 100644 backends/pytorch/include/ATen/ops/less_equal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/less_ops.h create mode 100644 backends/pytorch/include/ATen/ops/lgamma_ops.h create mode 100644 backends/pytorch/include/ATen/ops/log10_ops.h create mode 100644 backends/pytorch/include/ATen/ops/log1p_ops.h create mode 100644 backends/pytorch/include/ATen/ops/log2_ops.h create mode 100644 backends/pytorch/include/ATen/ops/log_normal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/log_ops.h create mode 100644 backends/pytorch/include/ATen/ops/log_softmax_ops.h create mode 100644 backends/pytorch/include/ATen/ops/logaddexp2_ops.h create mode 100644 backends/pytorch/include/ATen/ops/logaddexp_ops.h create mode 100644 backends/pytorch/include/ATen/ops/logcumsumexp_ops.h create mode 100644 backends/pytorch/include/ATen/ops/logdet_ops.h create mode 100644 backends/pytorch/include/ATen/ops/logical_and_ops.h create mode 100644 backends/pytorch/include/ATen/ops/logical_not_ops.h create mode 100644 backends/pytorch/include/ATen/ops/logical_or_ops.h create mode 100644 backends/pytorch/include/ATen/ops/logical_xor_ops.h create mode 100644 backends/pytorch/include/ATen/ops/logit_ops.h create mode 100644 backends/pytorch/include/ATen/ops/logsumexp_ops.h create mode 100644 backends/pytorch/include/ATen/ops/lshift_ops.h create mode 100644 backends/pytorch/include/ATen/ops/lt_ops.h create mode 100644 backends/pytorch/include/ATen/ops/lu_solve_ops.h create mode 100644 backends/pytorch/include/ATen/ops/mH_ops.h create mode 100644 backends/pytorch/include/ATen/ops/mT_ops.h create mode 100644 backends/pytorch/include/ATen/ops/masked_fill_ops.h create mode 100644 backends/pytorch/include/ATen/ops/masked_scatter_ops.h create mode 100644 backends/pytorch/include/ATen/ops/masked_select_ops.h create mode 100644 backends/pytorch/include/ATen/ops/matmul_ops.h create mode 100644 backends/pytorch/include/ATen/ops/matrix_H_ops.h create mode 100644 backends/pytorch/include/ATen/ops/matrix_exp_ops.h create mode 100644 backends/pytorch/include/ATen/ops/matrix_power_ops.h create mode 100644 backends/pytorch/include/ATen/ops/max_ops.h create mode 100644 backends/pytorch/include/ATen/ops/maximum_ops.h create mode 100644 backends/pytorch/include/ATen/ops/mean_ops.h create mode 100644 backends/pytorch/include/ATen/ops/median_ops.h create mode 100644 backends/pytorch/include/ATen/ops/min_ops.h create mode 100644 backends/pytorch/include/ATen/ops/minimum_ops.h create mode 100644 backends/pytorch/include/ATen/ops/mm_ops.h create mode 100644 backends/pytorch/include/ATen/ops/mode_ops.h create mode 100644 backends/pytorch/include/ATen/ops/moveaxis_ops.h create mode 100644 backends/pytorch/include/ATen/ops/movedim_ops.h create mode 100644 backends/pytorch/include/ATen/ops/msort_ops.h create mode 100644 backends/pytorch/include/ATen/ops/mul_ops.h create mode 100644 backends/pytorch/include/ATen/ops/multinomial_ops.h create mode 100644 backends/pytorch/include/ATen/ops/multiply_ops.h create mode 100644 backends/pytorch/include/ATen/ops/mv_ops.h create mode 100644 backends/pytorch/include/ATen/ops/mvlgamma_ops.h create mode 100644 backends/pytorch/include/ATen/ops/nan_to_num_ops.h create mode 100644 backends/pytorch/include/ATen/ops/nanmean_ops.h create mode 100644 backends/pytorch/include/ATen/ops/nanmedian_ops.h create mode 100644 backends/pytorch/include/ATen/ops/nanquantile_ops.h create mode 100644 backends/pytorch/include/ATen/ops/nansum_ops.h create mode 100644 backends/pytorch/include/ATen/ops/narrow_copy_ops.h create mode 100644 backends/pytorch/include/ATen/ops/narrow_ops.h create mode 100644 backends/pytorch/include/ATen/ops/ne_ops.h create mode 100644 backends/pytorch/include/ATen/ops/neg_ops.h create mode 100644 backends/pytorch/include/ATen/ops/negative_ops.h create mode 100644 backends/pytorch/include/ATen/ops/new_empty_ops.h create mode 100644 backends/pytorch/include/ATen/ops/new_empty_strided_ops.h create mode 100644 backends/pytorch/include/ATen/ops/new_full_ops.h create mode 100644 backends/pytorch/include/ATen/ops/new_ones_ops.h create mode 100644 backends/pytorch/include/ATen/ops/new_zeros_ops.h create mode 100644 backends/pytorch/include/ATen/ops/nextafter_ops.h create mode 100644 backends/pytorch/include/ATen/ops/nonzero_numpy_ops.h create mode 100644 backends/pytorch/include/ATen/ops/nonzero_ops.h create mode 100644 backends/pytorch/include/ATen/ops/nonzero_static_ops.h create mode 100644 backends/pytorch/include/ATen/ops/norm_ops.h create mode 100644 backends/pytorch/include/ATen/ops/normal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/not_equal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/numpy_T_ops.h create mode 100644 backends/pytorch/include/ATen/ops/or_ops.h create mode 100644 backends/pytorch/include/ATen/ops/orgqr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/ormqr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/outer_ops.h create mode 100644 backends/pytorch/include/ATen/ops/output_nr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/permute_ops.h create mode 100644 backends/pytorch/include/ATen/ops/pin_memory_ops.h create mode 100644 backends/pytorch/include/ATen/ops/pinverse_ops.h create mode 100644 backends/pytorch/include/ATen/ops/polygamma_ops.h create mode 100644 backends/pytorch/include/ATen/ops/positive_ops.h create mode 100644 backends/pytorch/include/ATen/ops/pow_ops.h create mode 100644 backends/pytorch/include/ATen/ops/prelu_ops.h create mode 100644 backends/pytorch/include/ATen/ops/prod_ops.h create mode 100644 backends/pytorch/include/ATen/ops/put_ops.h create mode 100644 backends/pytorch/include/ATen/ops/q_per_channel_axis_ops.h create mode 100644 backends/pytorch/include/ATen/ops/q_per_channel_scales_ops.h create mode 100644 backends/pytorch/include/ATen/ops/q_per_channel_zero_points_ops.h create mode 100644 backends/pytorch/include/ATen/ops/q_scale_ops.h create mode 100644 backends/pytorch/include/ATen/ops/q_zero_point_ops.h create mode 100644 backends/pytorch/include/ATen/ops/qr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/qscheme_ops.h create mode 100644 backends/pytorch/include/ATen/ops/quantile_ops.h create mode 100644 backends/pytorch/include/ATen/ops/rad2deg_ops.h create mode 100644 backends/pytorch/include/ATen/ops/random_ops.h create mode 100644 backends/pytorch/include/ATen/ops/ravel_ops.h create mode 100644 backends/pytorch/include/ATen/ops/reciprocal_ops.h create mode 100644 backends/pytorch/include/ATen/ops/record_stream_ops.h create mode 100644 backends/pytorch/include/ATen/ops/refine_names_ops.h create mode 100644 backends/pytorch/include/ATen/ops/relu_ops.h create mode 100644 backends/pytorch/include/ATen/ops/remainder_ops.h create mode 100644 backends/pytorch/include/ATen/ops/rename_ops.h create mode 100644 backends/pytorch/include/ATen/ops/renorm_ops.h create mode 100644 backends/pytorch/include/ATen/ops/repeat_interleave_ops.h create mode 100644 backends/pytorch/include/ATen/ops/repeat_ops.h create mode 100644 backends/pytorch/include/ATen/ops/requires_grad_ops.h create mode 100644 backends/pytorch/include/ATen/ops/reshape_as_ops.h create mode 100644 backends/pytorch/include/ATen/ops/reshape_ops.h create mode 100644 backends/pytorch/include/ATen/ops/resize_as_ops.h create mode 100644 backends/pytorch/include/ATen/ops/resize_as_sparse_ops.h create mode 100644 backends/pytorch/include/ATen/ops/resize_ops.h create mode 100644 backends/pytorch/include/ATen/ops/resolve_conj_ops.h create mode 100644 backends/pytorch/include/ATen/ops/resolve_neg_ops.h create mode 100644 backends/pytorch/include/ATen/ops/retain_grad_ops.h create mode 100644 backends/pytorch/include/ATen/ops/retains_grad_ops.h create mode 100644 backends/pytorch/include/ATen/ops/roll_ops.h create mode 100644 backends/pytorch/include/ATen/ops/rot90_ops.h create mode 100644 backends/pytorch/include/ATen/ops/round_ops.h create mode 100644 backends/pytorch/include/ATen/ops/row_indices_ops.h create mode 100644 backends/pytorch/include/ATen/ops/rshift_ops.h create mode 100644 backends/pytorch/include/ATen/ops/rsqrt_ops.h create mode 100644 backends/pytorch/include/ATen/ops/scatter_add_ops.h create mode 100644 backends/pytorch/include/ATen/ops/scatter_ops.h create mode 100644 backends/pytorch/include/ATen/ops/scatter_reduce_ops.h create mode 100644 backends/pytorch/include/ATen/ops/select_ops.h create mode 100644 backends/pytorch/include/ATen/ops/select_scatter_ops.h create mode 100644 backends/pytorch/include/ATen/ops/set_data_ops.h create mode 100644 backends/pytorch/include/ATen/ops/set_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sgn_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sigmoid_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sign_ops.h create mode 100644 backends/pytorch/include/ATen/ops/signbit_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sin_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sinc_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sinh_ops.h create mode 100644 backends/pytorch/include/ATen/ops/size_ops.h create mode 100644 backends/pytorch/include/ATen/ops/slice_inverse_ops.h create mode 100644 backends/pytorch/include/ATen/ops/slice_ops.h create mode 100644 backends/pytorch/include/ATen/ops/slice_scatter_ops.h create mode 100644 backends/pytorch/include/ATen/ops/slogdet_ops.h create mode 100644 backends/pytorch/include/ATen/ops/smm_ops.h create mode 100644 backends/pytorch/include/ATen/ops/softmax_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sort_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sparse_dim_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sparse_mask_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sparse_resize_and_clear_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sparse_resize_ops.h create mode 100644 backends/pytorch/include/ATen/ops/split_ops.h create mode 100644 backends/pytorch/include/ATen/ops/split_with_sizes_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sqrt_ops.h create mode 100644 backends/pytorch/include/ATen/ops/square_ops.h create mode 100644 backends/pytorch/include/ATen/ops/squeeze_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sspaddmm_ops.h create mode 100644 backends/pytorch/include/ATen/ops/std_ops.h create mode 100644 backends/pytorch/include/ATen/ops/stft_ops.h create mode 100644 backends/pytorch/include/ATen/ops/stride_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sub_ops.h create mode 100644 backends/pytorch/include/ATen/ops/subtract_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sum_ops.h create mode 100644 backends/pytorch/include/ATen/ops/sum_to_size_ops.h create mode 100644 backends/pytorch/include/ATen/ops/svd_ops.h create mode 100644 backends/pytorch/include/ATen/ops/swapaxes_ops.h create mode 100644 backends/pytorch/include/ATen/ops/swapdims_ops.h create mode 100644 backends/pytorch/include/ATen/ops/t_ops.h create mode 100644 backends/pytorch/include/ATen/ops/take_along_dim_ops.h create mode 100644 backends/pytorch/include/ATen/ops/take_ops.h create mode 100644 backends/pytorch/include/ATen/ops/tan_ops.h create mode 100644 backends/pytorch/include/ATen/ops/tanh_ops.h create mode 100644 backends/pytorch/include/ATen/ops/tensor_split_ops.h create mode 100644 backends/pytorch/include/ATen/ops/tile_ops.h create mode 100644 backends/pytorch/include/ATen/ops/to_dense_ops.h create mode 100644 backends/pytorch/include/ATen/ops/to_mkldnn_ops.h create mode 100644 backends/pytorch/include/ATen/ops/to_ops.h create mode 100644 backends/pytorch/include/ATen/ops/to_padded_tensor_ops.h create mode 100644 backends/pytorch/include/ATen/ops/to_sparse_bsc_ops.h create mode 100644 backends/pytorch/include/ATen/ops/to_sparse_bsr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/to_sparse_csc_ops.h create mode 100644 backends/pytorch/include/ATen/ops/to_sparse_csr_ops.h create mode 100644 backends/pytorch/include/ATen/ops/to_sparse_ops.h create mode 100644 backends/pytorch/include/ATen/ops/topk_ops.h create mode 100644 backends/pytorch/include/ATen/ops/trace_ops.h create mode 100644 backends/pytorch/include/ATen/ops/transpose_ops.h create mode 100644 backends/pytorch/include/ATen/ops/triangular_solve_ops.h create mode 100644 backends/pytorch/include/ATen/ops/tril_ops.h create mode 100644 backends/pytorch/include/ATen/ops/triu_ops.h create mode 100644 backends/pytorch/include/ATen/ops/true_divide_ops.h create mode 100644 backends/pytorch/include/ATen/ops/trunc_ops.h create mode 100644 backends/pytorch/include/ATen/ops/type_as_ops.h create mode 100644 backends/pytorch/include/ATen/ops/unbind_ops.h create mode 100644 backends/pytorch/include/ATen/ops/unflatten_ops.h create mode 100644 backends/pytorch/include/ATen/ops/unfold_ops.h create mode 100644 backends/pytorch/include/ATen/ops/uniform_ops.h create mode 100644 backends/pytorch/include/ATen/ops/unsafe_chunk_ops.h create mode 100644 backends/pytorch/include/ATen/ops/unsafe_split_ops.h create mode 100644 backends/pytorch/include/ATen/ops/unsafe_split_with_sizes_ops.h create mode 100644 backends/pytorch/include/ATen/ops/unsqueeze_ops.h create mode 100644 backends/pytorch/include/ATen/ops/values_ops.h create mode 100644 backends/pytorch/include/ATen/ops/var_ops.h create mode 100644 backends/pytorch/include/ATen/ops/vdot_ops.h create mode 100644 backends/pytorch/include/ATen/ops/view_as_ops.h create mode 100644 backends/pytorch/include/ATen/ops/view_ops.h create mode 100644 backends/pytorch/include/ATen/ops/vsplit_ops.h create mode 100644 backends/pytorch/include/ATen/ops/where_ops.h create mode 100644 backends/pytorch/include/ATen/ops/xlogy_ops.h create mode 100644 backends/pytorch/include/ATen/ops/xor_ops.h create mode 100644 backends/pytorch/include/ATen/ops/zero_ops.h create mode 100644 backends/pytorch/include/ATen/record_function.h create mode 100644 backends/pytorch/include/c10/core/Allocator.h create mode 100644 backends/pytorch/include/c10/core/AutogradState.h create mode 100644 backends/pytorch/include/c10/core/Backend.h create mode 100644 backends/pytorch/include/c10/core/DefaultDtype.h create mode 100644 backends/pytorch/include/c10/core/Device.h create mode 100644 backends/pytorch/include/c10/core/DeviceCapability.h create mode 100644 backends/pytorch/include/c10/core/DeviceGuard.h create mode 100644 backends/pytorch/include/c10/core/DeviceType.h create mode 100644 backends/pytorch/include/c10/core/DispatchKey.h create mode 100644 backends/pytorch/include/c10/core/DispatchKeySet.h create mode 100644 backends/pytorch/include/c10/core/Event.h create mode 100644 backends/pytorch/include/c10/core/GeneratorImpl.h create mode 100644 backends/pytorch/include/c10/core/InferenceMode.h create mode 100644 backends/pytorch/include/c10/core/Layout.h create mode 100644 backends/pytorch/include/c10/core/MemoryFormat.h create mode 100644 backends/pytorch/include/c10/core/OptionalRef.h create mode 100644 backends/pytorch/include/c10/core/QScheme.h create mode 100644 backends/pytorch/include/c10/core/SafePyObject.h create mode 100644 backends/pytorch/include/c10/core/Scalar.h create mode 100644 backends/pytorch/include/c10/core/ScalarType.h create mode 100644 backends/pytorch/include/c10/core/ScalarTypeToTypeMeta.h create mode 100644 backends/pytorch/include/c10/core/Storage.h create mode 100644 backends/pytorch/include/c10/core/StorageImpl.h create mode 100644 backends/pytorch/include/c10/core/Stream.h create mode 100644 backends/pytorch/include/c10/core/StreamGuard.h create mode 100644 backends/pytorch/include/c10/core/SymBool.h create mode 100644 backends/pytorch/include/c10/core/SymFloat.h create mode 100644 backends/pytorch/include/c10/core/SymInt.h create mode 100644 backends/pytorch/include/c10/core/SymIntArrayRef.h create mode 100644 backends/pytorch/include/c10/core/SymNodeImpl.h create mode 100644 backends/pytorch/include/c10/core/SymbolicShapeMeta.h create mode 100644 backends/pytorch/include/c10/core/TensorImpl.h create mode 100644 backends/pytorch/include/c10/core/TensorOptions.h create mode 100644 backends/pytorch/include/c10/core/UndefinedTensorImpl.h create mode 100644 backends/pytorch/include/c10/core/WrapDimMinimal.h create mode 100644 backends/pytorch/include/c10/core/alignment.h create mode 100644 backends/pytorch/include/c10/core/impl/COW.h create mode 100644 backends/pytorch/include/c10/core/impl/COWDeleter.h create mode 100644 backends/pytorch/include/c10/core/impl/DeviceGuardImplInterface.h create mode 100644 backends/pytorch/include/c10/core/impl/HermeticPyObjectTLS.h create mode 100644 backends/pytorch/include/c10/core/impl/InlineDeviceGuard.h create mode 100644 backends/pytorch/include/c10/core/impl/InlineEvent.h create mode 100644 backends/pytorch/include/c10/core/impl/InlineStreamGuard.h create mode 100644 backends/pytorch/include/c10/core/impl/LocalDispatchKeySet.h create mode 100644 backends/pytorch/include/c10/core/impl/PyInterpreter.h create mode 100644 backends/pytorch/include/c10/core/impl/PyInterpreterHooks.h create mode 100644 backends/pytorch/include/c10/core/impl/PyObjectSlot.h create mode 100644 backends/pytorch/include/c10/core/impl/SizesAndStrides.h create mode 100644 backends/pytorch/include/c10/core/impl/TorchDispatchModeTLS.h create mode 100644 backends/pytorch/include/c10/core/impl/VirtualGuardImpl.h create mode 100644 backends/pytorch/include/c10/macros/Export.h create mode 100644 backends/pytorch/include/c10/macros/Macros.h create mode 100644 backends/pytorch/include/c10/util/AlignOf.h create mode 100644 backends/pytorch/include/c10/util/ArrayRef.h create mode 100644 backends/pytorch/include/c10/util/BFloat16.h create mode 100644 backends/pytorch/include/c10/util/Backtrace.h create mode 100644 backends/pytorch/include/c10/util/C++17.h create mode 100644 backends/pytorch/include/c10/util/ConstexprCrc.h create mode 100644 backends/pytorch/include/c10/util/Deprecated.h create mode 100644 backends/pytorch/include/c10/util/DimVector.h create mode 100644 backends/pytorch/include/c10/util/Exception.h create mode 100644 backends/pytorch/include/c10/util/ExclusivelyOwned.h create mode 100644 backends/pytorch/include/c10/util/ExclusivelyOwnedTensorTraits.h create mode 100644 backends/pytorch/include/c10/util/Flags.h create mode 100644 backends/pytorch/include/c10/util/Float4_e2m1fn_x2.h create mode 100644 backends/pytorch/include/c10/util/Float8_e4m3fn.h create mode 100644 backends/pytorch/include/c10/util/Float8_e4m3fnuz.h create mode 100644 backends/pytorch/include/c10/util/Float8_e5m2.h create mode 100644 backends/pytorch/include/c10/util/Float8_e5m2fnuz.h create mode 100644 backends/pytorch/include/c10/util/Float8_e8m0fnu.h create mode 100644 backends/pytorch/include/c10/util/FunctionRef.h create mode 100644 backends/pytorch/include/c10/util/Half.h create mode 100644 backends/pytorch/include/c10/util/IdWrapper.h create mode 100644 backends/pytorch/include/c10/util/Lazy.h create mode 100644 backends/pytorch/include/c10/util/Logging.h create mode 100644 backends/pytorch/include/c10/util/MaybeOwned.h create mode 100644 backends/pytorch/include/c10/util/Metaprogramming.h create mode 100644 backends/pytorch/include/c10/util/Optional.h create mode 100644 backends/pytorch/include/c10/util/OptionalArrayRef.h create mode 100644 backends/pytorch/include/c10/util/Registry.h create mode 100644 backends/pytorch/include/c10/util/SmallVector.h create mode 100644 backends/pytorch/include/c10/util/StringUtil.h create mode 100644 backends/pytorch/include/c10/util/ThreadLocalDebugInfo.h create mode 100644 backends/pytorch/include/c10/util/Type.h create mode 100644 backends/pytorch/include/c10/util/TypeCast.h create mode 100644 backends/pytorch/include/c10/util/TypeIndex.h create mode 100644 backends/pytorch/include/c10/util/TypeList.h create mode 100644 backends/pytorch/include/c10/util/TypeSafeSignMath.h create mode 100644 backends/pytorch/include/c10/util/TypeTraits.h create mode 100644 backends/pytorch/include/c10/util/UniqueVoidPtr.h create mode 100644 backends/pytorch/include/c10/util/accumulate.h create mode 100644 backends/pytorch/include/c10/util/bit_cast.h create mode 100644 backends/pytorch/include/c10/util/bits.h create mode 100644 backends/pytorch/include/c10/util/complex.h create mode 100644 backends/pytorch/include/c10/util/complex_math.h create mode 100644 backends/pytorch/include/c10/util/complex_utils.h create mode 100644 backends/pytorch/include/c10/util/hash.h create mode 100644 backends/pytorch/include/c10/util/intrusive_ptr.h create mode 100644 backends/pytorch/include/c10/util/irange.h create mode 100644 backends/pytorch/include/c10/util/llvmMathExtras.h create mode 100644 backends/pytorch/include/c10/util/logging_common.h create mode 100644 backends/pytorch/include/c10/util/logging_is_not_google_glog.h create mode 100644 backends/pytorch/include/c10/util/order_preserving_flat_hash_map.h create mode 100644 backends/pytorch/include/c10/util/overflows.h create mode 100644 backends/pytorch/include/c10/util/python_stub.h create mode 100644 backends/pytorch/include/c10/util/qint32.h create mode 100644 backends/pytorch/include/c10/util/qint8.h create mode 100644 backends/pytorch/include/c10/util/quint2x4.h create mode 100644 backends/pytorch/include/c10/util/quint4x2.h create mode 100644 backends/pytorch/include/c10/util/quint8.h create mode 100644 backends/pytorch/include/c10/util/safe_numerics.h create mode 100644 backends/pytorch/include/c10/util/string_utils.h create mode 100644 backends/pytorch/include/c10/util/string_view.h create mode 100644 backends/pytorch/include/c10/util/typeid.h create mode 100644 backends/pytorch/include/torch/headeronly/core/DeviceType.h create mode 100644 backends/pytorch/include/torch/headeronly/core/Layout.h create mode 100644 backends/pytorch/include/torch/headeronly/core/MemoryFormat.h create mode 100644 backends/pytorch/include/torch/headeronly/core/ScalarType.h create mode 100644 backends/pytorch/include/torch/headeronly/core/TensorAccessor.h create mode 100644 backends/pytorch/include/torch/headeronly/macros/Export.h create mode 100644 backends/pytorch/include/torch/headeronly/macros/Macros.h create mode 100644 backends/pytorch/include/torch/headeronly/macros/cmake_macros.h create mode 100644 backends/pytorch/include/torch/headeronly/util/BFloat16.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Deprecated.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Exception.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Float4_e2m1fn_x2.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Float8_e4m3fn.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Float8_e4m3fnuz.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Float8_e5m2.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Float8_e5m2fnuz.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Float8_e8m0fnu.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Float8_fnuz_cvt.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Half.h create mode 100644 backends/pytorch/include/torch/headeronly/util/HeaderOnlyArrayRef.h create mode 100644 backends/pytorch/include/torch/headeronly/util/Metaprogramming.h create mode 100644 backends/pytorch/include/torch/headeronly/util/TypeList.h create mode 100644 backends/pytorch/include/torch/headeronly/util/TypeSafeSignMath.h create mode 100644 backends/pytorch/include/torch/headeronly/util/TypeTraits.h create mode 100644 backends/pytorch/include/torch/headeronly/util/bit_cast.h create mode 100644 backends/pytorch/include/torch/headeronly/util/bits.h create mode 100644 backends/pytorch/include/torch/headeronly/util/complex.h create mode 100644 backends/pytorch/include/torch/headeronly/util/floating_point_utils.h create mode 100644 backends/pytorch/include/torch/headeronly/util/qint32.h create mode 100644 backends/pytorch/include/torch/headeronly/util/qint8.h create mode 100644 backends/pytorch/include/torch/headeronly/util/quint2x4.h create mode 100644 backends/pytorch/include/torch/headeronly/util/quint4x2.h create mode 100644 backends/pytorch/include/torch/headeronly/util/quint8.h diff --git a/backends/pytorch/include/ATen/MethodOperators.h b/backends/pytorch/include/ATen/MethodOperators.h new file mode 100644 index 000000000..fd1f397d4 --- /dev/null +++ b/backends/pytorch/include/ATen/MethodOperators.h @@ -0,0 +1,449 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from MethodOperators.h + +#ifdef TORCH_ASSERT_NO_OPERATORS +#error This change adds a dependency on native_functions.yaml, \ + meaning the file will need to be re-compiled every time an operator \ + is changed or added. Consider if your change would be better placed in \ + another file, or if a more specific header might achieve the same goal. \ + See NOTE: [Tensor vs. TensorBase] +#endif + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace at { +namespace _ops { + +} // namespace _ops +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/StorageUtils.h b/backends/pytorch/include/ATen/StorageUtils.h new file mode 100644 index 000000000..9647f6411 --- /dev/null +++ b/backends/pytorch/include/ATen/StorageUtils.h @@ -0,0 +1,54 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +namespace at { + +class TensorBase; + +// Here we define a series of utils to create/manipulate ATen backed +// c10 storage implementations. + +/** + * Create a new shared memory storage impl managed by file descriptor + * + * @param size size in bytes + */ +C10_EXPORT c10::intrusive_ptr new_shm_fd_storage(size_t size); + +/** + * Copy src to dst + * Caller must guarantee the validness of the storage objects + * during the entire copy process, esp. when it's async. + * + * This can probably live in c10 namespace later if needed, + * but for now keep it in at to keep implementation simple. + * + * @param dst dst tensor + * @param src src tensor + * @param non_blocking (default false) whether this operation blocks caller + */ +C10_EXPORT void storage_copy( + c10::Storage& dst, + const c10::Storage& src, + bool non_blocking = false); + +/** + * In place change the storage to shm based. + * + * This is only applicable to CPU tensors not already shared. + * Otherwise, it's a no op to mirror the THP tensor behavior: + * https://pytorch.org/docs/stable/generated/torch.Tensor.share_memory_.html + * + * @param t a tensor + */ +C10_EXPORT void share_memory_(TensorBase& t); + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/ATen_fwd.h b/backends/pytorch/include/ATen/core/ATen_fwd.h new file mode 100644 index 000000000..68d4b7e2e --- /dev/null +++ b/backends/pytorch/include/ATen/core/ATen_fwd.h @@ -0,0 +1,51 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once +#include + +// Forward declarations of core ATen types used in dispatch functions +namespace c10 { + +template +class List; +template +class IListRef; +class Stream; +class Scalar; +class SymInt; +class SymIntList; +struct Storage; +struct TensorOptions; +template +class ArrayRef; +template +class OptionalArrayRef; + +} // namespace c10 + +namespace at { + +class Tensor; +class OptionalTensorRef; +struct Dimname; +struct Generator; +using TensorList = c10::ArrayRef; +using ITensorListRef = c10::IListRef; +using IOptTensorListRef = c10::IListRef; +using DimnameList = c10::ArrayRef; +using IntArrayRef = c10::ArrayRef; +using OptionalIntArrayRef = c10::OptionalArrayRef; +using OptionalSymIntArrayRef = c10::OptionalArrayRef; + +using c10::Stream; +using c10::Storage; +using c10::QScheme; +using c10::Scalar; +using c10::SymInt; +using c10::SymIntList; +using c10::TensorOptions; + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/CheckMemoryFormat.h b/backends/pytorch/include/ATen/core/CheckMemoryFormat.h new file mode 100644 index 000000000..c02296ff5 --- /dev/null +++ b/backends/pytorch/include/ATen/core/CheckMemoryFormat.h @@ -0,0 +1,29 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +namespace c10::impl { + +inline std::optional +check_tensor_options_and_extract_memory_format( + const TensorOptions& options, + std::optional memory_format) { + TORCH_CHECK( + options.requires_grad_opt() != true, + "Operators taking TensorOptions cannot take a TensorOptions with " + "options.requires_grad set as true. This isn't implemented yet."); + TORCH_CHECK( + !(options.has_memory_format() && memory_format.has_value()), + "Cannot set memory_format both in TensorOptions and explicit argument; please delete " + "the redundant setter."); + if (memory_format.has_value()) { + return memory_format; + } else { + return options.memory_format_opt(); + } +} + +} // namespace impl namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/DeprecatedTypeProperties.h b/backends/pytorch/include/ATen/core/DeprecatedTypeProperties.h new file mode 100644 index 000000000..2ea6e095b --- /dev/null +++ b/backends/pytorch/include/ATen/core/DeprecatedTypeProperties.h @@ -0,0 +1,144 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include + + +namespace at { + +class Tensor; + +// This class specifies a Backend and a ScalarType. Currently, it primarily +// serves as a replacement return value for Tensor::type(). Previously, +// Tensor::type() returned Type&, but we are changing Type to not be +// dtype-specific. +class TORCH_API DeprecatedTypeProperties { + public: + DeprecatedTypeProperties(Backend backend, ScalarType scalar_type) + : backend_(backend), scalar_type_(scalar_type) {} + + Backend backend() const { + return backend_; + } + + Layout layout() const { + return layout_from_backend(backend_); + } + + bool is_sparse() const { + return layout_from_backend(backend()) == kSparse; + } + + bool is_sparse_csr() const { + return layout_from_backend(backend()) == kSparseCsr; + } + + c10::DeviceType device_type() const { + return backendToDeviceType(backend_); + } + + bool is_cuda() const { + return backendToDeviceType(backend_) == kCUDA; + } + + ScalarType scalarType() const { + return scalar_type_; + } + + caffe2::TypeMeta typeMeta() const { + return scalarTypeToTypeMeta(scalar_type_); + } + + bool operator==(const DeprecatedTypeProperties& other) const { + return backend_ == other.backend() && scalar_type_ == other.scalarType(); + } + + bool operator!=(const DeprecatedTypeProperties& other) const { + return !(*this == other); + } + + std::string toString() const { + std::string base_str; + if (backend_ == Backend::Undefined || scalar_type_ == ScalarType::Undefined) { + base_str = "UndefinedType"; + } else { + base_str = std::string(at::toString(backend_)) + at::toString(scalar_type_) + "Type"; + } + return base_str; + } + + DeprecatedTypeProperties & toBackend(Backend b) const { + return globalDeprecatedTypePropertiesRegistry().getDeprecatedTypeProperties( + b, scalar_type_); + } + + DeprecatedTypeProperties & toScalarType(ScalarType s) const { + return globalDeprecatedTypePropertiesRegistry().getDeprecatedTypeProperties( + backend_, s); + } + + DeprecatedTypeProperties & cpu() const { + return toBackend(Backend::CPU); + } + + DeprecatedTypeProperties & cuda() const { + return toBackend(Backend::CUDA); + } + + DeprecatedTypeProperties & hip() const { + return toBackend(Backend::HIP); + } + + DeprecatedTypeProperties & privateUser1() const { + return toBackend(Backend::PrivateUse1); + } + + /// Constructs the `TensorOptions` from a type and a `device_index`. + TensorOptions options(int16_t device_index = -1) const { + return TensorOptions().dtype(typeMeta()) + .device(device_type(), static_cast(device_index)) + .layout(layout()); + } + + /// Constructs the `TensorOptions` from a type and a Device. Asserts that + /// the device type matches the device type of the type. + TensorOptions options(std::optional device_opt) const { + if (!device_opt.has_value()) { + return options(-1); + } else { + Device device = device_opt.value(); + AT_ASSERT(device.type() == device_type()); + return options(device.index()); + } + } + + operator TensorOptions() const { + return options(); + } + + int64_t id() const { + return static_cast(backend()) * + static_cast(ScalarType::NumOptions) + + static_cast(scalarType()); + } + + Tensor unsafeTensorFromTH(void * th_pointer, bool retain) const; + Storage unsafeStorageFromTH(void * th_pointer, bool retain) const; + Tensor copy(const Tensor & src, bool non_blocking=false, std::optional to_device={}) const; + + private: + Backend backend_; + ScalarType scalar_type_; +}; + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/DeprecatedTypePropertiesRegistry.h b/backends/pytorch/include/ATen/core/DeprecatedTypePropertiesRegistry.h new file mode 100644 index 000000000..0119d4c13 --- /dev/null +++ b/backends/pytorch/include/ATen/core/DeprecatedTypePropertiesRegistry.h @@ -0,0 +1,38 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// In order to preserve bc, we make DeprecatedTypeProperties instances unique +// just like they are for Type. + +#include +#include +#include + +namespace at { + +class DeprecatedTypeProperties; + +struct TORCH_API DeprecatedTypePropertiesDeleter { + void operator()(DeprecatedTypeProperties * ptr); +}; + +class TORCH_API DeprecatedTypePropertiesRegistry { + public: + DeprecatedTypePropertiesRegistry(); + + DeprecatedTypeProperties& getDeprecatedTypeProperties(Backend p, ScalarType s) const; + +private: + // NOLINTNEXTLINE(*c-array*) + std::unique_ptr registry + [static_cast(Backend::NumOptions)] + [static_cast(ScalarType::NumOptions)]; +}; + +TORCH_API DeprecatedTypePropertiesRegistry& globalDeprecatedTypePropertiesRegistry(); + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/Dict.h b/backends/pytorch/include/ATen/core/Dict.h new file mode 100644 index 000000000..35b829519 --- /dev/null +++ b/backends/pytorch/include/ATen/core/Dict.h @@ -0,0 +1,401 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { +struct IValue; +template class Dict; +struct Type; + +namespace impl { + +using valid_dict_key_types = guts::typelist::typelist< + int64_t, + std::string, + double, + c10::complex, + bool, + at::Tensor +>; +} + +namespace detail { + +struct DictKeyHash { + size_t operator()(const IValue& ivalue) const; +}; + +struct DictKeyEqualTo { + bool operator()(const IValue& lhs, const IValue& rhs) const; +}; + +struct DictImpl final : public c10::intrusive_ptr_target { + using dict_map_type = ska_ordered::order_preserving_flat_hash_map; + struct DictElementTypes final { + TypePtr keyType; + TypePtr valueType; + }; + + explicit DictImpl(dict_map_type dict_, DictElementTypes elementTypes_) + : dict(std::move(dict_)) + , elementTypes(std::move(elementTypes_)) {} + dict_map_type dict; + + DictElementTypes elementTypes; + + intrusive_ptr copy() const; + friend TORCH_API bool operator==(const DictImpl& lhs, const DictImpl& rhs); +}; + +} + +namespace impl { +template class DictIterator; + +/** + * A reference to an entry in the Dict. + * Use the `key()` and `value()` methods to read the element. + */ +template +class DictEntryRef final { +public: + explicit DictEntryRef(Iterator iterator) + : iterator_(std::move(iterator)) {} + + decltype(auto) key() const { + return iterator_->first.template to(); + } + + decltype(auto) value() const { + return iterator_->second.template to(); + } + + template + void setValue(Value_&& value) const { + static_assert(std::is_constructible_v, "Wrong type for the value argument of setValue()"); + iterator_->second = Value(std::forward(value)); + } + ~DictEntryRef() = default; + +private: + // allow copying and moving, but only our friends (i.e. the Dict class) can do + // it. Copying/moving this reference wrapper would be too ambiguous to allow it + // in the public API. + DictEntryRef(const DictEntryRef&) = default; + DictEntryRef& operator=(const DictEntryRef&) = default; + DictEntryRef(DictEntryRef&&) noexcept = default; + DictEntryRef& operator=(DictEntryRef&& rhs) & noexcept = default; + + Iterator iterator_; + friend class DictIterator; + friend class Dict; +}; + +// this wraps map_type::iterator to make sure user code can't rely +// on it being the type of the underlying map. +template +class DictIterator final { +public: + // C++17 friendly std::iterator implementation + using iterator_category = std::forward_iterator_tag; + using value_type = DictEntryRef; + using difference_type = std::ptrdiff_t; + using pointer = value_type*; + using reference = value_type&; + + explicit DictIterator() = default; + ~DictIterator() = default; + + DictIterator(const DictIterator& rhs): entryRef_(rhs.entryRef_) {} + DictIterator(DictIterator&& rhs) noexcept: entryRef_(std::move(rhs.entryRef_)) {} + DictIterator& operator=(const DictIterator& rhs) = default; + DictIterator& operator=(DictIterator&& rhs) noexcept { + entryRef_ = std::move(rhs.entryRef_); + return *this; + } + + DictIterator& operator++() { + ++entryRef_.iterator_; + return *this; + } + + DictIterator operator++(int) { + DictIterator copy(*this); + ++*this; + return copy; + } + + const DictEntryRef& operator*() const { + return entryRef_; + } + + const DictEntryRef* operator->() const { + return &entryRef_; + } + + friend difference_type operator-(const DictIterator& lhs, const DictIterator& rhs) { + return lhs.entryRef_.iterator_ - rhs.entryRef_.iterator_; + } + +private: + explicit DictIterator(Iterator iterator): entryRef_(std::move(iterator)) {} + + const Iterator& get_iterator_() const { + return entryRef_.iterator_; + } + + friend bool operator==(const DictIterator& lhs, const DictIterator& rhs) { + return lhs.get_iterator_() == rhs.get_iterator_(); + } + + friend bool operator!=(const DictIterator& lhs, const DictIterator& rhs) { + return lhs.get_iterator_() != rhs.get_iterator_(); + } + + friend bool operator<(const DictIterator& lhs, const DictIterator& rhs) { + return lhs.get_iterator_() < rhs.get_iterator_(); + } + + friend bool operator<=(const DictIterator& lhs, const DictIterator& rhs) { + return lhs.get_iterator_() <= rhs.get_iterator_(); + } + + friend bool operator>(const DictIterator& lhs, const DictIterator& rhs) { + return lhs.get_iterator_() > rhs.get_iterator_(); + } + + friend bool operator>=(const DictIterator& lhs, const DictIterator& rhs) { + return lhs.get_iterator_() >= rhs.get_iterator_(); + } + + DictEntryRef entryRef_; + + friend class DictIterator; + friend class Dict; +}; + +template Dict toTypedDict(Dict dict); +template Dict toGenericDict(Dict dict); +} + +/** + * An object of this class stores a map from Key to Value. + * + * This is a pointer type. After a copy, both Dicts + * will share the same storage: + * + * > Dict a; + * > Dict b = a; + * > b.insert(3, "three"); + * > ASSERT("three" == a.at(3)); + * + * We use this class in the PyTorch kernel API because that + * allows us to do optimizations and switch out the underlying + * map implementation without breaking backwards compatibility + * for the kernel API. + */ +template +// NOLINTNEXTLINE(cppcoreguidelines-special-member-functions) +class Dict final { +private: + static_assert((std::is_same_v && std::is_same_v) || guts::typelist::contains::value, "Invalid Key type for Dict. We only support int64_t, double, bool, and string."); + + // impl_ stores the underlying map as a ska_ordered::order_preserving_flat_hash_map. + // We intentionally don't offer conversion from/to + // order_preserving_flat_hash_map, return references to it or something like that, + // because such operations would get expensive if we switch out + // the actual map implementation. + // This is an intrusive_ptr because Dict is a pointer type. + // Invariant: This will never be a nullptr, there will always be a valid + // DictImpl. + c10::intrusive_ptr impl_; + + explicit Dict(c10::intrusive_ptr&& impl); + friend struct IValue; + template friend Dict impl::toTypedDict(Dict); + template friend Dict impl::toGenericDict(Dict); + +public: + using key_type = Key; + using mapped_type = Value; + using size_type = typename detail::DictImpl::dict_map_type::size_type; + using iterator = impl::DictIterator; + + /** + * Creates an empty dict. + */ + explicit Dict(); + + /** + * Create a generic dict with runtime type information. + * This only works for c10::impl::GenericDict and is not part of the public API + * but only supposed to be used internally by PyTorch. + */ + explicit Dict(TypePtr keyType, TypePtr valueType); + + ~Dict() = default; + + Dict(const Dict&) = default; + Dict& operator=(const Dict&) = default; + + /** + * Create a new Dict pointing to a deep copy of the same data. + * The Dict returned is a new dict with separate storage. + * Changes in it are not reflected in the original dict or vice versa. + */ + Dict copy() const; + + /** + * Returns an iterator to the first element of the container. + * If the container is empty, the returned iterator will be equal to end(). + */ + iterator begin() const; + + /** + * Returns an iterator to the element following the last element of the container. + * This element acts as a placeholder; attempting to access it results in undefined behavior. + */ + iterator end() const; + + /** + * Checks if the container has no elements. + */ + bool empty() const; + + /** + * Returns the number of elements in the container. + */ + size_type size() const; + + /** + * Erases all elements from the container. After this call, size() returns zero. + * Invalidates any references, pointers, or iterators referring to contained elements. May also invalidate past-the-end iterators. + */ + void clear() const; + + /** + * Inserts element(s) into the container, if the container doesn't already contain an element with an equivalent key. + * May invalidate any references, pointers, or iterators referring to contained elements. + * + * @return A pair consisting of an iterator to the inserted element (or to the element that prevented the insertion) and a bool denoting whether the insertion took place. + */ + template + std::pair insert(Key_&& key, Value_&& value) const; + + /** + * If an element with the given key already exists, it is overwritten with the given value. + * Otherwise, a new element with the given key and value are inserted. + * May invalidate any references, pointers, or iterators referring to contained elements. + * + * @return The bool component is true if the insertion took place and false if the assignment took place. The iterator component is pointing at the element that was inserted or updated. + */ + template + std::pair insert_or_assign(Key_&& key, Value_&& value) const; + + /** + * Removes the element pointed to by iter. + * May invalidate any references, pointers, or iterators referring to contained elements. + * The iterator iter must be valid and dereferenceable. Thus the end() iterator (which is valid, but is not dereferenceable) cannot be used as a value for iter. + */ + void erase(iterator iter) const; + + /** + * Removes the element with the given key, if it exists. + * May invalidate any references, pointers, or iterators referring to contained elements. + * + * @return The number of elements removed. This is either '1' if an element with the key existed, or '0' if it didn't. + */ + [[nodiscard]] size_t erase(const Key& key) const; + + /** + * Returns the mapped value of the element with key equivalent to key. + * If no such element exists, an exception of type std::out_of_range is thrown. + */ + Value at(const Key& key) const; + + /** + * Finds an element with key equivalent to key. + * + * @return Iterator to an element with key equivalent to key. + * If no such element is found, past-the-end (see end()) iterator is returned. + */ + iterator find(const Key& key) const; + + /** + * Checks if there is an element with key equivalent to key in the container. + * + * @return true if there is such an element, otherwise false. + */ + bool contains(const Key& key) const; + + /** + * Increase the capacity so that at least count elements can be stored without + * having to reallocate or rehash. + */ + void reserve(size_type count) const; + + /** + * Value equality comparison. This function implements Python-like semantics for + * equality: two dicts with the same identity (e.g. same pointer) trivially + * compare equal, otherwise each element is compared for equality. + */ + template + friend bool operator==( + const Dict& lhs, + const Dict& rhs); + template + friend bool operator!=( + const Dict& lhs, + const Dict& rhs); + + /** + * Identity comparison. Returns true if and only if `rhs` represents the same + * Dict object as `this`. + */ + bool is(const Dict& rhs) const; + + // private API for now because the return type will change to TypePtr + // instead of std::optional once types are mandatory. + TypePtr keyType() const; + TypePtr valueType() const; + + // [unsafe set type] + // These functions mutate the tagged type of this dictionary in place. + // There is no checking that the members of the dictionary are instances + // of the new types, nor is there a check that other IValues which + // hold references to this dictionary have the right static type. + // This functionality is used only in the unpickler, where at + // creation type the real type of the dictionary is unknown, but + // then later recovered from the static type information of the + // unpickled object. + void unsafeSetKeyType(TypePtr t); + void unsafeSetValueType(TypePtr t); +}; + +namespace impl { +// GenericDict is how IValue stores dicts. It is, however, not part of the +// public API. Kernels should use Dicts with concrete Key, Value types instead +// (maybe except for some internal prim ops). +using GenericDict = Dict; + +} +} + +namespace torch { + template using Dict = c10::Dict; +} + +#include // IWYU pragma: keep + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/Dict_inl.h b/backends/pytorch/include/ATen/core/Dict_inl.h new file mode 100644 index 000000000..d80e69ae6 --- /dev/null +++ b/backends/pytorch/include/ATen/core/Dict_inl.h @@ -0,0 +1,213 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace c10 { +namespace detail { +inline bool DictKeyEqualTo::operator()(const IValue& lhs, const IValue& rhs) const { + if (lhs.isTensor() && rhs.isTensor()) { + // for tensors, we compare only by identity (following how it's done in Python). + return lhs.is(rhs); + } + // Otherwise, we first compare by identity for efficiency, then by value (see: + // [container equality]) + return _fastEqualsForContainer(lhs, rhs); +} +} + +template decltype(auto) getTypePtr(); +std::string toString(const Type& type); + +namespace impl { + +template +Dict toTypedDict(GenericDict dict) { + TORCH_INTERNAL_ASSERT(*getTypePtr() == *dict.impl_->elementTypes.keyType, "Tried to cast a Dict<", toString(*dict.impl_->elementTypes.keyType), ", ", toString(*dict.impl_->elementTypes.valueType) ,"> to a Dict<", toString(*getTypePtr()), ", ", toString(*getTypePtr()), ">. Key types mismatch."); + TORCH_INTERNAL_ASSERT(*getTypePtr() == *dict.impl_->elementTypes.valueType, "Tried to cast a Dict<", toString(*dict.impl_->elementTypes.keyType), ", ", toString(*dict.impl_->elementTypes.valueType) ,"> to a Dict<", toString(*getTypePtr()), ", ", toString(*getTypePtr()), ">. Value types mismatch."); + + return Dict(std::move(dict.impl_)); +} + +template +GenericDict toGenericDict(Dict dict) { + return GenericDict(std::move(dict.impl_)); +} +} + +namespace detail { + +inline size_t DictKeyHash::operator()(const IValue& ivalue) const { + if (ivalue.isInt()) { + return std::hash()(ivalue.toInt()); + } else if (ivalue.isString()) { + return std::hash()(ivalue.toStringView()); + } else if (ivalue.isDouble()) { + return std::hash()(ivalue.toDouble()); + } else if (ivalue.isComplexDouble()) { + return c10::hash>()(ivalue.toComplexDouble()); + } else if (ivalue.isBool()) { + return std::hash()(ivalue.toBool()); + } else if (ivalue.isTensor()) { + return std::hash()(ivalue.toTensor().unsafeGetTensorImpl()); + } else if (ivalue.isDevice()) { + return std::hash()(ivalue.toDevice()); + } else { + TORCH_CHECK(false, "Can't hash IValues with tag '", ivalue.tagKind(), "'"); + } +} + +inline intrusive_ptr DictImpl::copy() const { + return make_intrusive(dict, elementTypes); +} + +} + +template +Dict::Dict() + :Dict(make_intrusive( + detail::DictImpl::dict_map_type(), + detail::DictImpl::DictElementTypes{getTypePtr(), getTypePtr()})) { + static_assert(!std::is_same_v, "This constructor is not valid for Dict. Please use c10::impl::GenericDict(keyType, valueType) instead."); + static_assert(!std::is_same_v, "This constructor is not valid for Dict<_, IValue>. Please use c10::impl::GenericDict(keyType, valueType) instead."); +} + +template +Dict::Dict(TypePtr keyType, TypePtr valueType) +: Dict(make_intrusive( + detail::DictImpl::dict_map_type(), + detail::DictImpl::DictElementTypes {std::move(keyType), std::move(valueType)})) { + static_assert(std::is_same_v, "This constructor is only valid for c10::impl::GenericDict."); + static_assert(std::is_same_v, "This constructor is only valid for c10::impl::GenericDict."); +} + +template +Dict::Dict(c10::intrusive_ptr&& impl): impl_(std::move(impl)) {} + +template +Dict Dict::copy() const { + return Dict(impl_->copy()); +} + +template +typename Dict::iterator Dict::begin() const { + return iterator{impl_->dict.begin()}; +} + +template +typename Dict::iterator Dict::end() const { + return iterator{impl_->dict.end()}; +} + +template +bool Dict::empty() const { + return impl_->dict.empty(); +} + +template +typename Dict::size_type Dict::size() const { + return impl_->dict.size(); +} + +template +void Dict::clear() const { + impl_->dict.clear(); +} + +template +template +std::pair::iterator, bool> Dict::insert(Key_&& key, Value_&& value) const { + static_assert(std::is_constructible_v, "Wrong type for the key argument of Dict::insert"); + static_assert(std::is_constructible_v, "Wrong type for the value argument of Dict::insert"); + auto inserted = impl_->dict.emplace( + Key(std::forward(key)), + Value(std::forward(value))); + return {iterator{inserted.first}, inserted.second}; +} + +template +template +std::pair::iterator, bool> Dict::insert_or_assign(Key_&& key, Value_&& value) const { + static_assert(std::is_constructible_v, "Wrong type for the key argument of Dict::insert_or_assign"); + static_assert(std::is_constructible_v, "Wrong type for the value argument of Dict::insert_or_assign"); + auto inserted = impl_->dict.insert_or_assign( + Key(std::forward(key)), + Value(std::forward(value))); + return {iterator{inserted.first}, inserted.second}; +} + +template +void Dict::erase(iterator iter) const { + impl_->dict.erase(iter.entryRef_.iterator_); +} + +template +[[nodiscard]] size_t Dict::erase(const Key& key) const { + return impl_->dict.erase(key); +} + +template +Value Dict::at(const Key& key) const { + return impl_->dict.at(key).template to(); +} + +template +typename Dict::iterator Dict::find(const Key& key) const { + return iterator{impl_->dict.find(key)}; +} + +template +bool Dict::contains(const Key& key) const { + return end() != find(key); +} + +template +void Dict::reserve(size_type count) const { + impl_->dict.reserve(count); +} + +template +TypePtr Dict::keyType() const { + return impl_->elementTypes.keyType; +} + +template +TypePtr Dict::valueType() const { + return impl_->elementTypes.valueType; +} +template +void Dict::unsafeSetKeyType(TypePtr t) { + impl_->elementTypes.keyType = std::move(t); +} + +template +void Dict::unsafeSetValueType(TypePtr t) { + impl_->elementTypes.valueType = std::move(t); +} + +template +bool operator==(const Dict& lhs, const Dict& rhs) { + // Dicts with the same identity trivially compare equal. + if (lhs.impl_ == rhs.impl_) { + return true; + } + + // Otherwise compare the values + return *lhs.impl_ == *rhs.impl_; +} + +template +bool operator!=(const Dict& lhs, const Dict& rhs) { + return !(lhs == rhs); +} + +template +bool Dict::is(const Dict& rhs) const { + return this->impl_ == rhs.impl_; +} +} + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/DimVector.h b/backends/pytorch/include/ATen/core/DimVector.h new file mode 100644 index 000000000..39f8179ba --- /dev/null +++ b/backends/pytorch/include/ATen/core/DimVector.h @@ -0,0 +1,18 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once +#include + +namespace at { + +// Redeclaring 'DimVector' type and size inside 'at' namespace. +// This is done to avoid modifying every use into their 'c10' +// equivalent. + +using c10::kDimVectorStaticSize; +using c10::DimVector; + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/Dimname.h b/backends/pytorch/include/ATen/core/Dimname.h new file mode 100644 index 000000000..eb8a83403 --- /dev/null +++ b/backends/pytorch/include/ATen/core/Dimname.h @@ -0,0 +1,53 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include + +namespace at { + +enum class NameType: uint8_t { BASIC, WILDCARD }; + +struct TORCH_API Dimname { + static Dimname fromSymbol(Symbol name); + static Dimname wildcard(); + static bool isValidName(const std::string& name); + + NameType type() const { return type_; } + Symbol symbol() const { return name_; } + + bool isBasic() const { return type_ == NameType::BASIC; } + bool isWildcard() const { return type_ == NameType::WILDCARD; } + + bool matches(Dimname other) const; + std::optional unify(Dimname other) const; + + private: + Dimname(Symbol name) + : name_(name), type_(NameType::BASIC) {} + Dimname(Symbol name, NameType type) + : name_(name), type_(type) {} + + Symbol name_; + NameType type_; +}; + +using DimnameList = c10::ArrayRef; + +TORCH_API std::ostream& operator<<(std::ostream& out, const Dimname& dimname); + +inline bool operator==(const Dimname& lhs, const Dimname& rhs) { + return lhs.symbol() == rhs.symbol(); +} + +inline bool operator!=(const Dimname& lhs, const Dimname& rhs) { + return !(lhs == rhs); +} + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/Generator.h b/backends/pytorch/include/ATen/core/Generator.h new file mode 100644 index 000000000..0f312f4e2 --- /dev/null +++ b/backends/pytorch/include/ATen/core/Generator.h @@ -0,0 +1,194 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include + +#include +#include +#include +#include + +// For the record I don't think this is a correct pimpl idiom. +// Including Impl header in interface header defeats the purpose +// because you can't change Impl private members without forcing +// everything that included the interface to rebuild. +// Impl should be forward-declared in the interface header instead. +#include + +/** + * Note [Generator] + * ~~~~~~~~~~~~~~~~ + * A Pseudo Random Number Generator (PRNG) is an engine that uses an algorithm to + * generate a seemingly random sequence of numbers, that may be later be used in creating + * a random distribution. Such an engine almost always maintains a state and requires a + * seed to start off the creation of random numbers. Often times, users have + * found it beneficial to be able to explicitly create, retain, and destroy + * PRNG states and also be able to have control over the seed value. + * + * A Generator in ATen gives users the ability to read, write and modify a PRNG engine. + * For instance, it does so by letting users seed a PRNG engine, fork the state of the + * engine, etc. + * + * By default, there is one generator per device, and a device's generator is + * lazily created. A user can use the torch.Generator() api to create their own generator. + */ + +/** + * Note [Acquire lock when using random generators] + * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + * Generator and its derived classes are NOT thread-safe. Please note that most of the + * places where we have inserted locking for generators are historically based, and we + * haven't actually checked that everything is truly thread safe (and it probably isn't). + * Please use the public mutex_ when using any methods from these classes, except for the + * read-only methods. You can learn about the usage by looking into the unittests + * (aten/src/ATen/cpu_generator_test.cpp) and other places where we have used lock_guard. + * + * TODO: Look into changing the threading semantics of Generators in ATen (e.g., making + * them non-thread safe and instead making the generator state splittable, to accommodate + * forks into other threads). + */ + +namespace at { + +class Tensor; + +struct TORCH_API Generator { + Generator() = default; + + explicit Generator(c10::intrusive_ptr gen_impl) + : impl_(std::move(gen_impl)) { + TORCH_CHECK(impl_.get(), "GeneratorImpl with nullptr is not supported"); + } + + bool operator==(const Generator& rhs) const { + return this->impl_ == rhs.impl_; + } + + bool operator!=(const Generator& rhs) const { + return !((*this) == rhs); + } + + bool defined() const { + return static_cast(impl_); + } + + c10::GeneratorImpl* unsafeGetGeneratorImpl() const { + return impl_.get(); + } + + c10::GeneratorImpl* unsafeReleaseGeneratorImpl() { + return impl_.release(); + } + + const c10::intrusive_ptr& getIntrusivePtr() const { + return impl_; + } + + void set_current_seed(uint64_t seed) { impl_->set_current_seed(seed); } + // Sets the offset of Generator state to the desired offset. This is currently + // supported for only Philox based Generators, i.e., CUDA and MPS. + void set_offset(uint64_t offset) { impl_->set_offset(offset); } + + // Returns the offset of Generator state. This is currently supported for only + // Philox based Generators, i.e., CUDA and MPS. + uint64_t get_offset() const { return impl_->get_offset(); } + + uint64_t current_seed() const { return impl_->current_seed(); } + + uint64_t seed() { return impl_->seed(); } + + // Implementation not inlined to prevent cycle reference between + // `ATen/core/Generator.h` and `ATen/core/Tensor.h` + void set_state(const at::Tensor& new_state); + + at::Tensor get_state() const; + + void graphsafe_set_state(const Generator& new_state); + + Generator graphsafe_get_state() const; + + std::mutex& mutex() { + return impl_->mutex_; + } + + DispatchKeySet key_set() const { + return impl_->key_set(); + } + + Device device() const { return impl_->device(); } + + inline void set_pyobj(PyObject* pyobj) const noexcept { + impl_->set_pyobj(pyobj); + } + + inline PyObject* pyobj() const noexcept { + return impl_->pyobj(); + } + + template + T* get() const { return static_cast(impl_.get()); } + + Generator clone() const { + return Generator(impl_->clone()); + } + + private: + c10::intrusive_ptr impl_; +}; + +template +Generator make_generator(Args&&... args) { + return Generator(c10::make_intrusive(std::forward(args)...)); +} + +/** + * Utility function to static cast input Generator* to + * the backend generator type (CPU/CUDAGeneratorImpl etc.) + */ +template +inline T * check_generator(std::optional gen) { + TORCH_CHECK(gen.has_value(), "Expected Generator but received nullopt"); + TORCH_CHECK(gen->defined(), "Generator with undefined implementation is not allowed"); + TORCH_CHECK(T::device_type() == gen->device().type(), "Expected a '", T::device_type(), "' device type for generator but found '", gen->device().type(), "'"); + return gen->get(); +} + +/** + * Utility function used in tensor implementations, which + * supplies the default generator to tensors, if an input generator + * is not supplied. The input Generator* is also static casted to + * the backend generator type (CPU/CUDAGeneratorImpl etc.) + */ +template +inline T* get_generator_or_default(const std::optional& gen, const Generator& default_gen) { + return gen.has_value() && gen->defined() ? check_generator(gen) : check_generator(default_gen); +} + +namespace detail { + +/** + * Helper function for checking the validity of new random generator + * state. Right now following conditions are checked: + * + * - The new state tensor must be a torch.ByteTensor + * - Data of the new state tensor must be contiguous + */ +inline void check_rng_state(const c10::TensorImpl& new_state) { + TORCH_CHECK_TYPE( + new_state.layout() == kStrided && new_state.device().type() == kCPU && new_state.dtype() == kByte, + "RNG state must be a torch.ByteTensor" + ); + + TORCH_CHECK(new_state.is_contiguous(), "RNG state must be contiguous"); +} + +} // namespace detail + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/IListRef.h b/backends/pytorch/include/ATen/core/IListRef.h new file mode 100644 index 000000000..694907d44 --- /dev/null +++ b/backends/pytorch/include/ATen/core/IListRef.h @@ -0,0 +1,638 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +#include +#include +#include +#include + +/* + * [Note: IListRef] + * Wrapper around different API containers (e.g. boxed and unboxed). + * + * What is it? + * =========== + * It is a tagged union of both boxed and unboxed API containers. + * Working implementations: + * + * - `IListRef` + * - `IListRef` + * + * Note that `IListRef` is a view type. Meaning that it won't own the + * tensors it holds. It's intended to be used only as argument parameters. + * Specifically, where these 2 worlds overlap. + * + * What is this for? + * ================= + * Historically, PyTorch has maintained 2 different APIs: the unboxed + * (called from C++ API and Python eager mode) and boxed APIs (called + * from the TorchScript JIT, mobile interpreter, and boxed fallbacks). + * + * Calling unboxed kernels from the boxed "world" and vice-versa may + * result in non-negligible overhead. Lists are one of those types: + * + * - Boxed world: `c10::List` + * - Unboxed world: `c10::ArrayRef` + * + * In this context, `c10::IListRef` solves this problem by wrapping those + * 2 container types, so that we don't need to convert from one to + * the other. + * + * (see https://github.com/pytorch/pytorch/issues/66328) + * + * What does it do? + * ================ + * This container wraps around the different tagged containers + * (currently, only boxed and unboxed), without incurring in extra + * overhead for converting from one to another. It does so while + * exposing usual container methods, which dispatch to corresponding + * implementations. + * + * While it works with different container types, it introduces + * overhead for repeatedly calling member functions (since those will + * get dispatched, again). Therefore, you should only use it to iterate + * through the list up to one time. If you need to do more complex things, + * call `materialize()` first. + * + * Adding support for a new Tag + * ============================ + * Suppose we want to add a new tag: `Chest`. Here are the steps + * we would have to go through: + * + * 1. Add a line for it in the macro `TORCH_ILISTREF_FORALL_TAGS`. + * + * #define TORCH_ILISTREF_FORALL_TAGS(_, ...) \ + * ... + * _(Chest, ##__VA_ARGS__) + * + * 2. Add type aliases, union members, and constructors. + * + * template + * class IListRef { + * ... + * using chest_type = + * typename detail::IListRefTagImpl::list_type; + * ... + * IListRef(...) : tag_(IListRefTag::Chest) { + * ... + * } + * ... + * union Payload { + * ... + * chest_type chest; + * ... + * }; + * ... + * }; + * + * 3. Add a default implementation for it (in 'IListRef_inl.h'). It's + * preferable to make the default implementation work for `T = Tensor` + * (both `Unboxed` and `Boxed` do it). + * + * template + * class IListRefTagImplBase { + * public: + * using elem_type = ListElemT; + * using list_type = ChestContainer; + * + * static const list_type& unwrap(const IListRef& ilist) { ... } + * + * static typename list_type::const_iterator& unwrap( + * IListRefIterator& it) { ... } + * + * static const typename list_type::const_iterator& unwrap( + * const IListRefIterator& it) { ... } + * + * static IListRefConstRef iterator_get( + * const typename list_type::const_iterator& it) { ... } + * } + * + * 4. Add an specialization for each of the already supported types. + * Finally, for consistency, add them to the tracking list. + * (see [Note: IListRefTagImpl Specializations]) + * + * template <> + * class IListRefTagImpl + * : public IListRefTagImplBase {}; + * + * Adding support for a new Type + * ============================= + * Suppose we want to add support for a new type: `Matrix`. + * Here are the steps we would have to go through: + * + * 1. Add an specialization for each of the existing tags. + * For consistency, add them to the tracking list. + * (see [Note: IListRefTagImpl Specializations]) + * + * template <> + * class IListRefTagImpl + * : public IListRefTagImplBase {}; + * + * template <> + * class IListRefTagImpl + * : public IListRefTagImplBase {}; + * + * Common Problems + * =============== + * 1. One of `IListRef(Iterator)` methods are failing to compile. + * + * That may be happening because the container type you added + * is not compatible with the code written for that method. If + * that's true, then you might have to transform that code into + * a static method call (see `List::operator[]` method). + * + * 2. Can't make `IListRefIterator::operator*` return a const-reference. + * + * First, keep in mind that we assume that boxed containers will + * have to deal with `IValue` (e.g. `c10::List`). In this context, + * what may be happening is that `IValue` doesn't store internally + * your type `T`. Instead, it constructs a type new `T` every time + * you try to get `T` for it (see `IListRef`). + */ + +namespace c10 { +template +class IListRef; + +/* + * Applies arbitrary macros to each `IListRefTag`. + */ +#define TORCH_ILISTREF_FORALL_TAGS(_, ...) \ + _(Unboxed, ##__VA_ARGS__) \ + _(Boxed, ##__VA_ARGS__) \ + _(Materialized, ##__VA_ARGS__) + +/* + * Defines a "switch-case" for `TAG`. Inside, it executes `BODY`, + * while bringing to scope: + * + * - `ImplT`: the implementation class for `TAG` + * - `this_`: the result of unwrapping `this` + */ +#define TORCH_ILISTREF_UNWRAP_CASE(TAG, BODY) \ + case c10::IListRefTag::TAG: { \ + using ImplT = c10::detail::IListRefTagImpl; \ + auto& this_ = ImplT::unwrap(*this); \ + BODY \ + } break; + +/* + * Dispatches the unwrap call, depending on `TAG`, followed by + * the execution of `BODY`. It aborts if `TAG` is not a `IListRefTag`. + * + * This macro is useful because it allows us to handle different + * types (that correspond to different tags) to be implemented + * only once. We can do it even when the implementation of the + * different tags aren't syntactically the same, by dispatching + * it to a function (e.g. `ImplT::(this_)`). + */ +#define TORCH_ILISTREF_UNWRAP(TAG, BODY) \ + C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum") \ + switch (TAG) { \ + TORCH_ILISTREF_FORALL_TAGS(TORCH_ILISTREF_UNWRAP_CASE, BODY) \ + break; \ + default: \ + TORCH_INTERNAL_ASSERT(false, "invalid IListRef tag."); \ + } \ + C10_DIAGNOSTIC_POP() + +enum class IListRefTag { +#define DEFINE_TAG(tag, ...) tag, + TORCH_ILISTREF_FORALL_TAGS(DEFINE_TAG) +#undef DEFINE_TAG + None +}; + +namespace detail { +/* + * Type alias that specifies whether we return a reference or a copy of `T`. + * + * What is this for? + * ================= + * Since values in the boxed world are represented by an `IValue`, we also + * depend on whether it can be converted to a const-reference (`Tensor`) or + * has to create a new copy of `T` (`OptionalTensorRef`). + */ +template +using IListRefConstRef = typename ivalue_to_const_ref_overload_return::type; + +/* + * Interface that implements key functions for each `IListRefTag` type. + * + * What is this for? + * ================= + * Given an `IListRef(Iterator)`, some methods have to be implemented + * differently for each `TAG`. Therefore, the methods inside this class + * are used as dispatch targets for the different `IListRefTag` values. + * + * You should create an specialization of this class for each possible + * combination of `IListRefTag` type (except `None`) and element types + * (e.g. `Tensor`). + * + * What does it do? + * ================ + * 1. defines static methods to be used as dispatch targets by both + * `IListRef` and `IListRefIterator` (see the implementation of + * `IListRefTagImplBase`). + * + * 2. defines the `elem_type` and `list_type` aliases that will be + * used in the definition of `IListRef`. In general, we should do + * so by inheriting from `IListRefTagImplBase`. + * + * [Note: IListRefTagImpl Specialization] + * ====================================== + * For `IListRef(Iterator)`: + * - + * - + * - + * + * For `IListRef(Iterator)`: + * - + * - + * - + */ +template +class IListRefTagImpl {}; + +/* + * Base implementation of `IListRefTagImpl` methods. + * + * What is this for? + * ================= + * This should make adding specializations for new types easier. For + * example, one should be able to add a new type just by making its + * `IListRefTagImpl` specialization inherit from `IListRefTagImplBase`. + * + * You should create a partial specialization for this class only if + * you introduce a new `IListRefTag`. The idea being that there is one + * default implementation for each possible value of `IListRefTag`. + * + * What does it do? + * ================ + * 1. defines `elem_type` as an alias to `ListElemT`. + * + * 1. defines `list_type` as an alias to the default container type + * that will hold a collection of `elem_type`. The idea being that + * all types tagged as `TAG` will have `list_type` as its container, + * with different `elem_type`. + * + * 3. defines the default implementation for each of the methods that + * are supposed to be defined on `IListRefTagImpl` specializations. + * + * 4. inheriting from `IListRefTagImplBase` also means + * that the payload of the type `IListRef` will be of type `list_type` + * when it is tagged as `TAG`. + */ +template +class IListRefTagImplBase {}; + +/* + * Materialized container for `IListRef`. + * + * What is this for? + * ================= + * Container that groups `T` references together. This exchanges the + * overhead of every method call from `IListRef` for a dynamic allocation. + * + * You should use this container instead of `IListRef` if: + * + * - You are going to iterate the list more than once + * - You need to repeatedly access arbitrary elements (using `operator[]`) + * What does it do? + + * ================ + * Removes the reference (&) from the type, and wraps it into a + * `std::reference_wrapper`. If `IListRefConstRef` is not a + * reference type, then it's left unchanged. + */ +template +using _MaterializedIListRefElem = std::conditional_t< + std::is_reference_v, + typename std::reference_wrapper>, + T>; + +template +using MaterializedIListRefElem = _MaterializedIListRefElem>; + +template +using MaterializedIListRef = std::vector>; + +} // namespace detail + +/* + * Iterator for `IListRef`. + * + * What is it? + * =========== + * Currently, a `std::bidirectional_iterator` that wraps the iterator + * types defined for each of the `IListRefTag`. + * + * One should be able to use it, as if it were the unwrapped + * iterators themselves. + + * What does it do? + * ================ + * Similarly to `IListRef`, this is a wrapper class. Specifically, it + * wraps each container's `const_iterator` type alias. So, for example, + * given that the container for `IListRefTag::Boxed` is `c10::List`, this + * iterator will wrap a `c10::List::const_iterator`. + * + * [Note: MSVC Iterator Debug] + * =========================== + * MSVC `vector::iterator` implementation (used in the boxed variant) + * makes it so this union's destructor, copy-constructor (assignment), and + * move-constructor (assignment) are implicitly deleted. + * + * Therefore, we need to explicitly define them as needed. Follows a list + * of places where these are needed and their reason: + * + * - `Payload` destructor: + * it is deleted only if the macro `_ITERATOR_DEBUG_LEVEL` is set to 2. + * + * - `IListRefIterator` destructor: + * same as above. However, we need to explicitly call the variant + * destructor explicitly. + * + * - `IListRefIterator` copy-constructor: + * it is deleted only if the macro `_ITERATOR_DEBUG_LEVEL` is different + * than 0. + */ +template +class IListRefIterator { + private: +#define DEFINE_FRIEND_CLASS(TAG, ...) \ + friend class detail::IListRefTagImpl; \ + friend class detail::IListRefTagImplBase< \ + IListRefTag::TAG, \ + T, \ + typename detail::IListRefTagImpl::elem_type>; + TORCH_ILISTREF_FORALL_TAGS(DEFINE_FRIEND_CLASS) +#undef DEFINE_FRIEND_CLASS + + public: + // C++17 friendly std::iterator implementation + using iterator_category = std::bidirectional_iterator_tag; + using value_type = T; + using difference_type = std::ptrdiff_t; + using pointer = T*; + using reference = T&; + + using unboxed_iterator_type = typename detail:: + IListRefTagImpl::list_type::const_iterator; + using boxed_iterator_type = typename detail:: + IListRefTagImpl::list_type::const_iterator; + using materialized_iterator_type = + typename detail::MaterializedIListRef::const_iterator; + + IListRefIterator() : tag_(IListRefTag::None) {} + +#if defined(_MSC_VER) && _ITERATOR_DEBUG_LEVEL != 0 + // See [Note: MSVC Iterator Debug] + IListRefIterator(const IListRefIterator& iterator) + : tag_(iterator.tag_) { + switch (tag_) { + case IListRefTag::Boxed: + payload_.boxed_iterator = iterator.payload_.boxed_iterator; + break; + case IListRefTag::Unboxed: + payload_.unboxed_iterator = iterator.payload_.unboxed_iterator; + break; + case IListRefTag::Materialized: + payload_.materialized_iterator = iterator.payload_.materialized_iterator; + break; + default: + TORCH_INTERNAL_ASSERT(false, "invalid IListRef tag."); + } + } +#endif + +#if defined(_MSC_VER) && _ITERATOR_DEBUG_LEVEL == 2 + // See [Note: MSVC Iterator Debug] + ~IListRefIterator() noexcept(false) { + switch (tag_) { + case IListRefTag::Boxed: + payload_.boxed_iterator.~boxed_iterator_type(); + break; + case IListRefTag::Unboxed: + payload_.unboxed_iterator.~unboxed_iterator_type(); + break; + case IListRefTag::Materialized: + payload_.materialized_iterator.~materialized_iterator_type(); + break; + default: + TORCH_INTERNAL_ASSERT(false, "invalid IListRef tag."); + } + } +#endif + + IListRefIterator(boxed_iterator_type boxed) : tag_(IListRefTag::Boxed) { + payload_.boxed_iterator = boxed; + } + + IListRefIterator(unboxed_iterator_type unboxed) : tag_(IListRefTag::Unboxed) { + payload_.unboxed_iterator = unboxed; + } + + IListRefIterator(materialized_iterator_type materialized) : tag_(IListRefTag::Materialized) { + payload_.materialized_iterator = materialized; + } + + detail::IListRefConstRef operator*() const { + TORCH_ILISTREF_UNWRAP(tag_, { return ImplT::iterator_get(this_); }); + } + + IListRefIterator& operator++() { + TORCH_ILISTREF_UNWRAP(tag_, { ++this_; }); + return *this; + } + + IListRefIterator operator++(int) { + auto old = *this; + TORCH_ILISTREF_UNWRAP(tag_, { ++this_; }); + return old; + } + + IListRefIterator& operator--() { + TORCH_ILISTREF_UNWRAP(tag_, { --this_; }); + return *this; + } + + IListRefIterator operator--(int) { + auto old = *this; + TORCH_ILISTREF_UNWRAP(tag_, { --this_; }); + return old; + } + + bool operator==(const IListRefIterator& rhs) const { + if (tag_ != rhs.tag_) { + return false; + } + TORCH_ILISTREF_UNWRAP(tag_, { + auto& rhs_it = ImplT::unwrap(rhs); + return this_ == rhs_it; + }); + } + + bool operator!=(const IListRefIterator& rhs) const { + return !(*this == rhs); + } + + private: + union Payload { + boxed_iterator_type boxed_iterator; + unboxed_iterator_type unboxed_iterator; + materialized_iterator_type materialized_iterator; + void* _init_ptr; + Payload() : _init_ptr(nullptr) {} +#if defined(_MSC_VER) + // See [Note: MSVC Iterator Debug] + ~Payload() {} +#endif + }; + + Payload payload_; + IListRefTag tag_; +}; + +/* + * See [Note: IListRef] + */ +template +class IListRef { + private: +#define DEFINE_FRIEND_CLASS(TAG, ...) \ + friend class detail::IListRefTagImpl; \ + friend class detail::IListRefTagImplBase< \ + IListRefTag::TAG, \ + T, \ + typename detail::IListRefTagImpl::elem_type>; + TORCH_ILISTREF_FORALL_TAGS(DEFINE_FRIEND_CLASS) +#undef DEFINE_FRIEND_CLASS + + public: + using unboxed_type = + typename detail::IListRefTagImpl::list_type; + using boxed_type = + typename detail::IListRefTagImpl::list_type; + using materialized_type = + typename detail::MaterializedIListRef; + + using iterator = IListRefIterator; + using const_iterator = IListRefIterator; + using reverse_iterator = std::reverse_iterator; + using value_type = typename iterator::value_type; + + IListRef() : tag_(IListRefTag::None) {} + + IListRef(const boxed_type& boxed) : tag_(IListRefTag::Boxed) { + payload_.boxed = &boxed; + } + + IListRef(const unboxed_type& unboxed) : tag_(IListRefTag::Unboxed) { + payload_.unboxed = unboxed; + } + + IListRef(const std::initializer_list& list) : tag_(IListRefTag::Unboxed) { + payload_.unboxed = at::ArrayRef(list); + } + + template < + typename... UnboxedConstructorArgs, + typename = std::enable_if_t< + std::is_constructible_v>> + IListRef(UnboxedConstructorArgs&&... args) : tag_(IListRefTag::Unboxed) { + payload_.unboxed = unboxed_type(std::forward(args)...); + } + + IListRef(const materialized_type& materialized) : tag_(IListRefTag::Materialized) { + payload_.materialized = &materialized; + } + + size_t size() const { + TORCH_ILISTREF_UNWRAP(tag_, { return this_.size(); }); + } + + bool empty() const { + return size() == 0; + } + + iterator begin() const { + TORCH_ILISTREF_UNWRAP(tag_, { return this_.begin(); }); + } + + iterator end() const { + TORCH_ILISTREF_UNWRAP(tag_, { return this_.end(); }); + } + + detail::IListRefConstRef front() const { + TORCH_ILISTREF_UNWRAP(tag_, { return ImplT::front(this_); }); + } + + /* + * Materializes the `IListRef` into a `std::vector`. + * + * This should be used when one wishes to either: + * + * - iterate over the list more than once: each `IListRefIterator` + * member function call has to go through a switch, introducing + * non-negligible overhead + * + * - randomly access an arbitrary element using `operator[]`: + * same reason as above + */ + detail::MaterializedIListRef materialize() const { + if (isMaterialized()) { + return toMaterialized(); + } + + detail::MaterializedIListRef materialized; + materialized.reserve(size()); + for (const auto& t : *this) { + materialized.emplace_back(t); + } + return materialized; + } + +#define DEFINE_CHECK(TAG, ...) \ + bool is##TAG() const { \ + return tag_ == IListRefTag::TAG; \ + } + TORCH_ILISTREF_FORALL_TAGS(DEFINE_CHECK) +#undef DEFINE_CHECK + + bool isNone() const { + return tag_ == IListRefTag::None; + } + +#define DEFINE_CASTING(TAG, ...) \ + const typename detail::IListRefTagImpl::list_type& \ + to##TAG() const { \ + TORCH_INTERNAL_ASSERT(is##TAG()); \ + return detail::IListRefTagImpl::unwrap(*this); \ + } + TORCH_ILISTREF_FORALL_TAGS(DEFINE_CASTING) +#undef DEFINE_CASTING + + private: + union Payload { + const boxed_type* boxed; + unboxed_type unboxed; + const materialized_type* materialized; + Payload() : boxed(nullptr) {} + }; + + Payload payload_; + IListRefTag tag_; +}; + +} // namespace c10 + +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/IListRef_inl.h b/backends/pytorch/include/ATen/core/IListRef_inl.h new file mode 100644 index 000000000..98bf27276 --- /dev/null +++ b/backends/pytorch/include/ATen/core/IListRef_inl.h @@ -0,0 +1,208 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace at { +class Tensor; +class OptionalTensorRef; +} + + +namespace c10::detail { + +/* + * Specializations of `IListRefTagImplBase` that implement the default + * implementation for `IListRefTag::Unboxed`. + */ +template +class IListRefTagImplBase { + public: + using elem_type = ListElemT; + using list_type = ArrayRef; + + /* + * These `unwrap` static methods unwraps the inner containers out + * of `IListRef` (and `IListRefIterator`). They are required when + * the macro `TORCH_ILISTREF_UNWRAP` is called. + */ + static const list_type& unwrap(const IListRef& ilist) { + return ilist.payload_.unboxed; + } + + static typename list_type::const_iterator& unwrap(IListRefIterator& it) { + return it.payload_.unboxed_iterator; + } + + static const typename list_type::const_iterator& unwrap( + const IListRefIterator& it) { + return it.payload_.unboxed_iterator; + } + + /* + * We have these function (besides the `unwrap`s above) because the + * implementation for both `IListRef::operator[]` and `IListRefIterator::operator*` + * weren't syntactically equal for the existing tags at the time + * (`Unboxed` and `Boxed`). + */ + static IListRefConstRef front(const list_type& lst) { + return lst.front(); + } + + static IListRefConstRef iterator_get( + const typename list_type::const_iterator& it) { + return *it; + } +}; + +/* + * Specializations of `IListRefTagImplBase` that implement the default + * implementation for `IListRefTag::Boxed`. + */ +template +class IListRefTagImplBase { + public: + using elem_type = ListElemT; + using list_type = List; + + static const list_type& unwrap(const IListRef& ilist) { + return *ilist.payload_.boxed; + } + + static typename list_type::const_iterator& unwrap(IListRefIterator& it) { + return it.payload_.boxed_iterator; + } + + static const typename list_type::const_iterator& unwrap( + const IListRefIterator& it) { + return it.payload_.boxed_iterator; + } + + static IListRefConstRef front(const list_type& lst) { + return lst[0]; + } + + static IListRefConstRef iterator_get( + const typename list_type::const_iterator& it) { + return (*it).get().toTensor(); + } +}; + +/* + * Specializations of `IListRefTagImplBase` that implement the default + * implementation for `IListRefTag::Materialized`. + */ +template +class IListRefTagImplBase> { + public: + using elem_type = MaterializedIListRefElem; + using list_type = MaterializedIListRef; + + static const list_type& unwrap(const IListRef& ilist) { + return *ilist.payload_.materialized; + } + + static typename list_type::const_iterator& unwrap(IListRefIterator& it) { + return it.payload_.materialized_iterator; + } + + static const typename list_type::const_iterator& unwrap( + const IListRefIterator& it) { + return it.payload_.materialized_iterator; + } + + static IListRefConstRef front(const list_type& lst) { + return lst[0]; + } + + static IListRefConstRef iterator_get( + const typename list_type::const_iterator& it) { + return *it; + } +}; + +/* + * [Note: ITensorListRef] + * Specializations necessary for `IListRef` type. + * + * Since the default implementations are usually done with supporting + * `Tensor` in mind, we only have to inherit from the base implementations. + */ +template <> +class IListRefTagImpl + : public IListRefTagImplBase {}; + +template <> +class IListRefTagImpl + : public IListRefTagImplBase {}; + +template <> +class IListRefTagImpl + : public IListRefTagImplBase< + IListRefTag::Materialized, + at::Tensor, + MaterializedIListRefElem> {}; + +/* + * [Note: IOptTensorListRef] + * Specializations necessary for `IListRef` type. + * + * We can't get an `at::OptionalTensorRef` directly from an instance of + * `List>` (the type that corresponds to the boxed world). + * + * So, the default implementation won't help us. Thus, we have to implement + * this method ourselves. + */ +template <> +class IListRefTagImpl + : public IListRefTagImplBase {}; + +template <> +class IListRefTagImpl + : public IListRefTagImplBase> { + + public: + /* + * Given an instance of the types corresponding to the `Boxed` tag, we override + * the default implementation, so that we can return a `at::OptionalTensorRef`. + */ + static IListRefConstRef iterator_get( + const typename list_type::const_iterator& it) { + C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wdangling-reference") + const auto& ivalue = (*it).get(); + C10_DIAGNOSTIC_POP() + if (!ivalue.isNone()) { + const auto& tensor = ivalue.toTensor(); + return (tensor.defined()) ? tensor : at::OptionalTensorRef{}; + } + return {}; + } +}; + +template <> +class IListRefTagImpl + : public IListRefTagImplBase< + IListRefTag::Materialized, + at::OptionalTensorRef, + MaterializedIListRefElem> {}; + +} // namespace c10::detail + + +namespace at { + +// [Note: ITensorListRef] +using ITensorListRef = c10::IListRef; +using ITensorListRefIterator = c10::IListRefIterator; +using MaterializedITensorListRef = c10::detail::MaterializedIListRef; +// [Note: IOptTensorListRef] +using IOptTensorListRef = c10::IListRef; +using IOptTensorListRefIterator = c10::IListRefIterator; +using MaterializedIOptTensorListRef = c10::detail::MaterializedIListRef; + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/List.h b/backends/pytorch/include/ATen/core/List.h new file mode 100644 index 000000000..f109430c5 --- /dev/null +++ b/backends/pytorch/include/ATen/core/List.h @@ -0,0 +1,496 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace at { +class Tensor; +} +namespace c10 { +struct IValue; +template class List; +struct Type; + +namespace detail { + +struct ListImpl final : public c10::intrusive_ptr_target { + using list_type = std::vector; + + explicit TORCH_API ListImpl(list_type list_, TypePtr elementType_); + + list_type list; + + TypePtr elementType; + + intrusive_ptr copy() const { + return make_intrusive(list, elementType); + } + friend TORCH_API bool operator==(const ListImpl& lhs, const ListImpl& rhs); +}; +} + +namespace impl { + +template class ListIterator; + +template class ListElementReference; + +template +void swap(ListElementReference&& lhs, ListElementReference&& rhs) noexcept; + +template +bool operator==(const ListElementReference& lhs, const T& rhs); + +template +bool operator==(const T& lhs, const ListElementReference& rhs); + +template +struct ListElementConstReferenceTraits { + // In the general case, we use IValue::to(). + using const_reference = typename c10::detail::ivalue_to_const_ref_overload_return::type; +}; + +// There is no to() overload for std::optional. +template<> +struct ListElementConstReferenceTraits> { + using const_reference = std::optional>; +}; + +template +class ListElementReference final { +public: + operator std::conditional_t< + std::is_reference_v::type>, + const T&, + T>() const; + + ListElementReference& operator=(T&& new_value) &&; + + ListElementReference& operator=(const T& new_value) &&; + + // assigning another ref to this assigns the underlying value + ListElementReference& operator=(ListElementReference&& rhs) && noexcept; + + const IValue& get() const& { + return *iterator_; + } + + friend void swap(ListElementReference&& lhs, ListElementReference&& rhs) noexcept; + + ListElementReference(const ListElementReference&) = delete; + ListElementReference& operator=(const ListElementReference&) = delete; + ~ListElementReference() = default; + +private: + ListElementReference(Iterator iter) + : iterator_(iter) {} + + // allow moving, but only our friends (i.e. the List class) can move us + ListElementReference(ListElementReference&&) noexcept = default; + ListElementReference& operator=(ListElementReference&& rhs) & noexcept { + iterator_ = std::move(rhs.iterator_); + return *this; + } + + friend class List; + friend class ListIterator; + + Iterator iterator_; +}; + +// this wraps vector::iterator to make sure user code can't rely +// on it being the type of the underlying vector. +template +class ListIterator final { + public: + // C++17 friendly std::iterator implementation + using iterator_category = std::random_access_iterator_tag; + using value_type = T; + using difference_type = std::ptrdiff_t; + using pointer = T*; + using reference = ListElementReference; + + explicit ListIterator() = default; + ~ListIterator() = default; + + ListIterator(const ListIterator&) = default; + ListIterator(ListIterator&&) noexcept = default; + ListIterator& operator=(const ListIterator&) = default; + ListIterator& operator=(ListIterator&&) noexcept = default; + + ListIterator& operator++() { + ++iterator_; + return *this; + } + + ListIterator operator++(int) { + ListIterator copy(*this); + ++*this; + return copy; + } + + ListIterator& operator--() { + --iterator_; + return *this; + } + + ListIterator operator--(int) { + ListIterator copy(*this); + --*this; + return copy; + } + + ListIterator& operator+=(typename List::size_type offset) { + iterator_ += offset; + return *this; + } + + ListIterator& operator-=(typename List::size_type offset) { + iterator_ -= offset; + return *this; + } + + ListIterator operator+(typename List::size_type offset) const { + return ListIterator{iterator_ + offset}; + } + + ListIterator operator-(typename List::size_type offset) const { + return ListIterator{iterator_ - offset}; + } + + friend difference_type operator-(const ListIterator& lhs, const ListIterator& rhs) { + return lhs.iterator_ - rhs.iterator_; + } + + ListElementReference operator*() const { + return {iterator_}; + } + + ListElementReference operator[](typename List::size_type offset) const { + return {iterator_ + offset}; + } + +private: + explicit ListIterator(Iterator iterator): iterator_(std::move(iterator)) {} + + Iterator iterator_; + + friend bool operator==(const ListIterator& lhs, const ListIterator& rhs) { + return lhs.iterator_ == rhs.iterator_; + } + + friend bool operator!=(const ListIterator& lhs, const ListIterator& rhs) { + return !(lhs == rhs); + } + + friend bool operator<(const ListIterator& lhs, const ListIterator& rhs) { + return lhs.iterator_ < rhs.iterator_; + } + + friend bool operator<=(const ListIterator& lhs, const ListIterator& rhs) { + return lhs.iterator_ <= rhs.iterator_; + } + + friend bool operator>(const ListIterator& lhs, const ListIterator& rhs) { + return lhs.iterator_ > rhs.iterator_; + } + + friend bool operator>=(const ListIterator& lhs, const ListIterator& rhs) { + return lhs.iterator_ >= rhs.iterator_; + } + + friend class ListIterator; + friend class List; +}; + +template List toTypedList(List list); +template List toList(List&& list); +template List toList(const List& list); +const IValue* ptr_to_first_element(const List& list); +} + +/** + * An object of this class stores a list of values of type T. + * + * This is a pointer type. After a copy, both Lists + * will share the same storage: + * + * > List a; + * > List b = a; + * > b.push_back("three"); + * > ASSERT("three" == a.get(0)); + * + * We use this class in the PyTorch kernel API instead of + * std::vector, because that allows us to do optimizations + * and switch out the underlying list implementation without + * breaking backwards compatibility for the kernel API. + */ +template +// NOLINTNEXTLINE(cppcoreguidelines-special-member-functions) +class List final { +private: + // This is an intrusive_ptr because List is a pointer type. + // Invariant: This will never be a nullptr, there will always be a valid + // ListImpl. + c10::intrusive_ptr impl_; + + using internal_reference_type = impl::ListElementReference; + using internal_const_reference_type = typename impl::ListElementConstReferenceTraits::const_reference; + +public: + using value_type = T; + using size_type = typename c10::detail::ListImpl::list_type::size_type; + using iterator = impl::ListIterator; + using const_iterator = impl::ListIterator; + using reverse_iterator = impl::ListIterator; + + /** + * Constructs an empty list. + */ + explicit List(); + + /** + * Constructs a list with some initial values. + * Example: + * List a({2, 3, 4}); + */ + List(std::initializer_list initial_values); + explicit List(ArrayRef initial_values); + + /** + * Create a generic list with runtime type information. + * This only works for c10::impl::GenericList and is not part of the public API + * but only supposed to be used internally by PyTorch. + */ + explicit List(TypePtr elementType); + + List(const List&) = default; + List& operator=(const List&) = default; + ~List() = default; + + /** + * Create a new List pointing to a deep copy of the same data. + * The List returned is a new list with separate storage. + * Changes in it are not reflected in the original list or vice versa. + */ + List copy() const; + + /** + * Returns the element at specified location pos, with bounds checking. + * If pos is not within the range of the container, an exception of type std::out_of_range is thrown. + */ + internal_const_reference_type get(size_type pos) const; + + /** + * Moves out the element at the specified location pos and returns it, with bounds checking. + * If pos is not within the range of the container, an exception of type std::out_of_range is thrown. + * The list contains an invalid element at position pos afterwards. Any operations + * on it before re-setting it are invalid. + */ + value_type extract(size_type pos) const; + + /** + * Returns a reference to the element at specified location pos, with bounds checking. + * If pos is not within the range of the container, an exception of type std::out_of_range is thrown. + * + * You cannot store the reference, but you can read it and assign new values to it: + * + * List list = ...; + * list[2] = 5; + * int64_t v = list[1]; + */ + internal_const_reference_type operator[](size_type pos) const; + + internal_reference_type operator[](size_type pos); + + /** + * Assigns a new value to the element at location pos. + */ + void set(size_type pos, const value_type& value) const; + + /** + * Assigns a new value to the element at location pos. + */ + void set(size_type pos, value_type&& value) const; + + /** + * Returns an iterator to the first element of the container. + * If the container is empty, the returned iterator will be equal to end(). + */ + iterator begin() const; + + /** + * Returns an iterator to the element following the last element of the container. + * This element acts as a placeholder; attempting to access it results in undefined behavior. + */ + iterator end() const; + + /** + * Checks if the container has no elements. + */ + bool empty() const; + + /** + * Returns the number of elements in the container + */ + size_type size() const; + + /** + * Increase the capacity of the vector to a value that's greater or equal to new_cap. + */ + void reserve(size_type new_cap) const; + + /** + * Erases all elements from the container. After this call, size() returns zero. + * Invalidates any references, pointers, or iterators referring to contained elements. Any past-the-end iterators are also invalidated. + */ + void clear() const; + + /** + * Inserts value before pos. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + iterator insert(iterator pos, const T& value) const; + + /** + * Inserts value before pos. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + iterator insert(iterator pos, T&& value) const; + + /** + * Inserts a new element into the container directly before pos. + * The new element is constructed with the given arguments. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + template + iterator emplace(iterator pos, Args&&... value) const; + + /** + * Appends the given element value to the end of the container. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + void push_back(const T& value) const; + + /** + * Appends the given element value to the end of the container. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + void push_back(T&& value) const; + + /** + * Appends the given list to the end of the container. Uses at most one memory allocation. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + void append(List lst) const; + + /** + * Appends the given element value to the end of the container. + * The new element is constructed with the given arguments. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + template + void emplace_back(Args&&... args) const; + + /** + * Removes the element at pos. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + iterator erase(iterator pos) const; + + /** + * Removes the elements in the range [first, last). + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + iterator erase(iterator first, iterator last) const; + + /** + * Removes the last element of the container. + * Calling pop_back on an empty container is undefined. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + void pop_back() const; + + /** + * Resizes the container to contain count elements. + * If the current size is less than count, additional default-inserted elements are appended. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + void resize(size_type count) const; + + /** + * Resizes the container to contain count elements. + * If the current size is less than count, additional copies of value are appended. + * May invalidate any references, pointers, or iterators referring to contained elements. Any past-the-end iterators may also be invalidated. + */ + void resize(size_type count, const T& value) const; + + /** + * Value equality comparison. This function implements Python-like semantics for + * equality: two lists with the same identity (e.g. same pointer) trivially + * compare equal, otherwise each element is compared for equality. + */ + template + friend bool operator==(const List& lhs, const List& rhs); + + template + friend bool operator!=(const List& lhs, const List& rhs); + + /** + * Identity comparison. Returns true if and only if `rhs` represents the same + * List object as `this`. + */ + bool is(const List& rhs) const; + + std::vector vec() const; + + /** + * Returns the number of Lists currently pointing to this same list. + * If this is the only instance pointing to this list, returns 1. + */ + // TODO Test use_count + size_t use_count() const; + + TypePtr elementType() const; + + // See [unsafe set type] for why this exists. + void unsafeSetElementType(TypePtr t); + +private: + explicit List(c10::intrusive_ptr&& elements); + explicit List(const c10::intrusive_ptr& elements); + friend struct IValue; + template friend List impl::toTypedList(List); + template friend List impl::toList(List&&); + template friend List impl::toList(const List&); + friend const IValue* impl::ptr_to_first_element(const List& list); +}; + +namespace impl { +// GenericList is how IValue stores lists. It is, however, not part of the +// public API. Kernels should use Lists with concrete types instead +// (maybe except for some internal prim ops). +using GenericList = List; + +} +} + +namespace torch { + template using List = c10::List; +} + +#include // IWYU pragma: keep + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/List_inl.h b/backends/pytorch/include/ATen/core/List_inl.h new file mode 100644 index 000000000..55c1e24c2 --- /dev/null +++ b/backends/pytorch/include/ATen/core/List_inl.h @@ -0,0 +1,358 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace c10 { + +template decltype(auto) getTypePtr(); +std::string toString(const Type& type); + +template +List::List(c10::intrusive_ptr&& elements) +: impl_(std::move(elements)) {} + +template +List::List(const c10::intrusive_ptr& elements) +: impl_(elements) {} + +template +List::List() +: List(make_intrusive( + typename c10::detail::ListImpl::list_type(), + getTypePtr())) { + static_assert(!std::is_same_v, "This constructor is not valid for List. Please use c10::impl::GenericList(elementType) instead."); +} + +template +List::List(ArrayRef values) +: List(make_intrusive( + typename c10::detail::ListImpl::list_type(), + getTypePtr())) { + static_assert(!std::is_same_v, "This constructor is not valid for List. Please use c10::impl::GenericList(elementType)."); + impl_->list.reserve(values.size()); + for (const T& element : values) { + impl_->list.push_back(element); + } +} + +template +List::List(std::initializer_list initial_values) +: List(ArrayRef(initial_values)) { + static_assert(!std::is_same_v, "This constructor is not valid for List. Please use c10::impl::GenericList(elementType)."); +} + +template +List::List(TypePtr elementType) +: List(make_intrusive( + typename c10::detail::ListImpl::list_type(), + std::move(elementType))) { + static_assert(std::is_same_v || std::is_same_v>, + "This constructor is only valid for c10::impl::GenericList or List."); +} + +namespace impl { +template +List toTypedList(impl::GenericList list) { + // If there's other instances of the list (i.e. list.use_count() > 1), then we have to be invariant + // because upcasting would allow people to add types into the new list that would break the old list. + // However, if there aren't any other instances of this list (i.e. list.use_count() == 1), then we can + // allow upcasting. This can be a perf improvement since we can cast List to List> + // without having to copy it. This is also used to provide backwards compatibility with some old models + // that serialized the index arguments to aten::index, aten::index_put, aten::index_put_ and aten::index_put_impl_ + // as List before we changed that argument to be List>. When deserializing, we + // have list.use_count() == 1 and can deserialize the List directly as List>. + TORCH_CHECK(*list.impl_->elementType == *getTypePtr() + || (list.use_count() == 1 && list.impl_->elementType->isSubtypeOf(*getTypePtr())) + , "Tried to cast a List<", toString(*list.impl_->elementType), "> to a List<", toString(*getTypePtr()), ">. Types mismatch."); + return List(std::move(list.impl_)); +} + +template +impl::GenericList toList(List&& list) { + return GenericList(std::move(list.impl_)); +} +template +impl::GenericList toList(const List& list) { + return GenericList(list.impl_); +} +} + +template +List List::copy() const { + return List(impl_->copy()); +} + +namespace detail { + template + T list_element_to(T element) { + return element; + } + template + T list_element_to(const IValue& element) { + return element.template to(); + } + template + T list_element_to(IValue&& element) { + return std::move(element).template to(); + } + template + struct ListElementFrom { + static IValue from(const T& element) { + return element; + } + static IValue from(T&& element) { + return std::move(element); + } + }; + template<> + struct ListElementFrom { + static const IValue& from(const IValue& element) { + return element; + } + static IValue&& from(IValue&& element) { + return std::move(element); + } + }; +} + +namespace impl { + +template +ListElementReference::operator std::conditional_t< + std::is_reference_v::type>, + const T&, + T>() const { + return iterator_->template to(); +} + +template +ListElementReference& ListElementReference::operator=(T&& new_value) && { + *iterator_ = c10::detail::ListElementFrom::from(std::move(new_value)); + return *this; +} + +template +ListElementReference& ListElementReference::operator=(const T& new_value) && { + *iterator_ = c10::detail::ListElementFrom::from(new_value); + return *this; +} + +template +ListElementReference& ListElementReference::operator=(ListElementReference&& rhs) && noexcept { + *iterator_ = *rhs.iterator_; + return *this; +} + +template +void swap(ListElementReference&& lhs, ListElementReference&& rhs) noexcept { + std::swap(*lhs.iterator_, *rhs.iterator_); +} + +template +bool operator==(const ListElementReference& lhs, const T& rhs) { + const T& lhs_tmp = lhs; + return lhs_tmp == rhs; +} + +template +inline bool operator==(const T& lhs, const ListElementReference& rhs) { + return rhs == lhs; +} + +template +inline typename ListElementConstReferenceTraits::const_reference +list_element_to_const_ref(const IValue& element) { + return element.template to(); +} + +template<> +inline typename ListElementConstReferenceTraits>::const_reference +list_element_to_const_ref>(const IValue& element) { + return element.toOptionalStringRef(); +} + +} // namespace impl + +template +void List::set(size_type pos, const value_type& value) const { + impl_->list.at(pos) = c10::detail::ListElementFrom::from(value); +} + +template +void List::set(size_type pos, value_type&& value) const { + impl_->list.at(pos) = c10::detail::ListElementFrom::from(std::move(value)); +} + +template +typename List::internal_const_reference_type List::get(size_type pos) const { + return operator[](pos); +} + +template +typename List::internal_const_reference_type List::operator[](size_type pos) const { + return c10::impl::list_element_to_const_ref(impl_->list.at(pos)); +} + +template +typename List::internal_reference_type List::operator[](size_type pos) { + static_cast(impl_->list.at(pos)); // Throw the exception if it is out of range. + return {impl_->list.begin() + static_castlist)::difference_type>(pos)}; +} + +template +typename List::value_type List::extract(size_type pos) const { + auto& elem = impl_->list.at(pos); + auto result = c10::detail::list_element_to(std::move(elem)); + // Reset the list element to a T() instead of None to keep it correctly typed + elem = c10::detail::ListElementFrom::from(T{}); + return result; +} + +template +typename List::iterator List::begin() const { + return iterator(impl_->list.begin()); +} + +template +typename List::iterator List::end() const { + return iterator(impl_->list.end()); +} + +template +bool List::empty() const { + return impl_->list.empty(); +} + +template +typename List::size_type List::size() const { + return impl_->list.size(); +} + +template +void List::reserve(size_type new_cap) const { + impl_->list.reserve(new_cap); +} + +template +void List::clear() const { + impl_->list.clear(); +} + +template +typename List::iterator List::insert(iterator pos, const T& value) const { + return iterator { impl_->list.insert(pos.iterator_, c10::detail::ListElementFrom::from(value)) }; +} + +template +typename List::iterator List::insert(iterator pos, T&& value) const { + return iterator { impl_->list.insert(pos.iterator_, c10::detail::ListElementFrom::from(std::move(value))) }; +} + +template +template +typename List::iterator List::emplace(iterator pos, Args&&... value) const { + // TODO Use list_element_from? + return iterator { impl_->list.emplace(pos.iterator_, std::forward(value)...) }; +} + +template +void List::push_back(const T& value) const { + impl_->list.push_back(c10::detail::ListElementFrom::from(value)); +} + +template +void List::push_back(T&& value) const { + impl_->list.push_back(c10::detail::ListElementFrom::from(std::move(value))); +} + +template +void List::append(List b) const { + if (b.use_count() == 1) { + impl_->list.insert(impl_->list.end(), make_move_iterator(b.impl_->list.begin()), make_move_iterator(b.impl_->list.end())); + } else { + impl_->list.insert(impl_->list.end(), b.impl_->list.begin(), b.impl_->list.end()); + } +} + +template +template +void List::emplace_back(Args&&... args) const { + // TODO Use list_element_from? + impl_->list.push_back(T(std::forward(args)...)); +} + +template +typename List::iterator List::erase(iterator pos) const { + return iterator { impl_->list.erase(pos.iterator_) }; +} + +template +typename List::iterator List::erase(iterator first, iterator last) const { + return iterator { impl_->list.erase(first.iterator_, last.iterator_) }; +} + +template +void List::pop_back() const { + impl_->list.pop_back(); +} + +template +void List::resize(size_type count) const { + impl_->list.resize(count, T{}); +} + +template +void List::resize(size_type count, const T& value) const { + impl_->list.resize(count, value); +} + +template +bool operator==(const List& lhs, const List& rhs) { + // Lists with the same identity trivially compare equal. + if (lhs.impl_ == rhs.impl_) { + return true; + } + + // Otherwise, just compare values directly. + return *lhs.impl_ == *rhs.impl_; +} + +template +bool operator!=(const List& lhs, const List& rhs) { + return !(lhs == rhs); +} + +template +bool List::is(const List& rhs) const { + return this->impl_ == rhs.impl_; +} + +template +std::vector List::vec() const { + std::vector result(begin(), end()); + return result; +} + +template +size_t List::use_count() const { + return impl_.use_count(); +} + +template +TypePtr List::elementType() const { + return impl_->elementType; +} + +template +void List::unsafeSetElementType(TypePtr t) { + impl_->elementType = std::move(t); +} + +} + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/NamedTensor.h b/backends/pytorch/include/ATen/core/NamedTensor.h new file mode 100644 index 000000000..890373220 --- /dev/null +++ b/backends/pytorch/include/ATen/core/NamedTensor.h @@ -0,0 +1,148 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace at { + +class TensorBase; + +// XXX: This file exists because TensorImpl is in c10, but Dimname is in ATen. +// Due to the c10/ATen library split, TensorImpl cannot depend on Dimname, +// so we have a couple of workarounds. +// +// In the long term, we'll move Dimname to c10 and everything in this file +// can be refactored out. The main blocker for that is that "c10::Symbol" +// actually exists outside of c10 and needs to be moved in. + +// TensorImpl has a unique_ptr field. +// XXX: Ideally we would just put std::optional> into TensorImpl. +// +// This class has an important invariant: there must be at least ONE +// non-wildcard +struct TORCH_API NamedTensorMeta final : public c10::NamedTensorMetaInterface { + // This enum is to remind people that the invariant on constructors is that + // the list of dimnames must have at least one non-wildcard + enum HAS_NON_WILDCARD { + HasNonWildcard + }; + + explicit NamedTensorMeta(HAS_NON_WILDCARD /*unused*/, DimnameList names) + : names_(names.vec()) { + check_invariants(); + } + explicit NamedTensorMeta(HAS_NON_WILDCARD /*unused*/, std::vector&& names) + : names_(std::move(names)) { + check_invariants(); + } + + std::unique_ptr clone() const override { + return std::make_unique(HasNonWildcard, names_); + } + + DimnameList names() const { return names_; } + + // Used for an assertion in TensorImpl.h + int64_t slow_dim() const override { + return static_cast(names_.size()); + } + + void check_invariants() const { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + std::any_of(names_.begin(), names_.end(), [](const Dimname& n) { return !n.isWildcard(); })); + } + + void set_names(HAS_NON_WILDCARD /*unused*/, DimnameList new_names) { + TORCH_INTERNAL_ASSERT(new_names.size() == names_.size()); + std::copy(new_names.begin(), new_names.end(), names_.begin()); + check_invariants(); + } + + void set_names(HAS_NON_WILDCARD /*unused*/, std::vector&& new_names) { + TORCH_INTERNAL_ASSERT(new_names.size() == names_.size()); + names_ = std::move(new_names); + check_invariants(); + } + + // INVARIANT: at least one Dimname is non-WILDCARD + std::vector names_; +}; + +// When NamesMode is disabled, then all operations ignore tensors' names fields. +// Concretely speaking, all tensors are treated as having nullopt names. +struct TORCH_API NamesMode { + static bool is_enabled(); + static void set_enabled(bool enabled); +}; + + +// A RAII, thread local (!) guard that enables or disables names upon +// construction, and sets it back to the original value upon destruction. +struct TORCH_API NoNamesGuard { + NoNamesGuard() : prev_mode(NamesMode::is_enabled()) { + NamesMode::set_enabled(false); + } + NoNamesGuard(const NoNamesGuard&) = delete; + NoNamesGuard(NoNamesGuard&&) = delete; + NoNamesGuard& operator=(const NoNamesGuard&) = delete; + NoNamesGuard& operator=(NoNamesGuard&&) = delete; + ~NoNamesGuard() { + if (initialized) { + reset(); + } + } + void reset() { + TORCH_INTERNAL_ASSERT(initialized); + NamesMode::set_enabled(prev_mode); + } + private: + bool prev_mode; + bool initialized{true}; +}; + +void check_names_valid_for(const TensorBase& tensor, DimnameList names); +void check_names_valid_for(size_t tensor_dim, DimnameList names); + +// Sets the names of `tensor` to be `names`. +TORCH_API const TensorBase& internal_set_names_inplace(const TensorBase& tensor, std::optional names); +TORCH_API const TensorBase& internal_set_names_inplace(const TensorBase& tensor, std::vector&& names, bool validate_names); + +constexpr size_t kMaxNamedTensorDim = 64; + +DimnameList default_names(size_t len); + +namespace impl { + +// Some helper functions on TensorImpl. Useful for working with names in TH. +// XXX: Ideally these would exist as methods on TensorImpl +TORCH_API void internal_set_names_inplace(TensorImpl* impl, std::optional names, bool validate_names); +TORCH_API void internal_set_names_inplace(TensorImpl* impl, std::vector&& names, bool validate_names); + +void check_names_valid_for(TensorImpl* impl, DimnameList names); + +// Returns true if the tensor's names exist and are not all 'None'. +// Returns false if the tensor's names don't exist (were not allocated), +// or if all names are 'None'. +// We treat not-allocated-names the same as allocated names that are all 'None'. +TORCH_API bool has_names(const TensorImpl* impl); + +// Returns the names of the tensor's dimensions. +// Unnamed tensors are treated as having 'None' in all dimension; this method +// would return a DimnameList of all 'None's for an unnamed tensor. +TORCH_API DimnameList get_names(const TensorImpl* impl); + +// This is more of an implementation detail; one should use impl::get_names / +// Tensor::names() whenever possible because it provides a cleaner API. +// Returns the names of the tensor if they have been allocated; returns nullopt +// instead if the haven't been. The names of a tensor are not allocated if a +// tensor is constructed with names=None. +TORCH_API std::optional get_opt_names(const TensorImpl* impl); + +} // namespace impl + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/QuantizerBase.h b/backends/pytorch/include/ATen/core/QuantizerBase.h new file mode 100644 index 000000000..fbb5f92f0 --- /dev/null +++ b/backends/pytorch/include/ATen/core/QuantizerBase.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +namespace at { + +class Tensor; +struct QTensorImpl; +struct Quantizer; +using ConstQuantizerPtr = const c10::intrusive_ptr&; +using QuantizerPtr = c10::intrusive_ptr; + +/** + * Quantizer is the class for storing all the information + * that's necessary to perform quantize and dequantize + * operation. + * + * We might have different types of quantization schemes and this is + * the base class for all quantizers. + * + * QTensorImpl will hold a pointer to Quantizer so that we can support + * different quantization schemes on Tensor. + * + * For example, the most common quantization scheme, Affine Quantization, + * requires scale and zero_point as parameters, we'll store scale and zero_point + * inside the instance and we can use it to quantize a float Tensor or + * dequantize a quantized Tensor. + * + * When you add new types of leaf Quantizer class, please also + * make sure to add a corresponding QScheme enum since + * they should have one to one mapping. + * + * Note about intrusive_ptr: + * Quantized Tensor holds an intrusive_ptr to Quantizer, and multiple Tensor can + * share the same Quantizer. Quantizer should be immutable. + */ +struct TORCH_API Quantizer : public c10::intrusive_ptr_target { + // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members) + const ScalarType scalar_type_; + explicit Quantizer(ScalarType scalar_type) : scalar_type_(scalar_type) {} + ~Quantizer() override = default; + + // Copied from torch/csrc/jit/ir/scope.h + QuantizerPtr intrusive_from_this() { + c10::raw::intrusive_ptr::incref(this); // we are creating a new pointer + // from a raw `this` pointer + // so we need to bump the refcount + // to account for this ownership + return c10::intrusive_ptr::reclaim(this); + } + + /** + * Each concrete Quantizer type should have a unique QScheme type. + */ + virtual QScheme qscheme() const = 0; + + ScalarType scalar_type() const { + return scalar_type_; + } + + /** + * quantize a float Tensor into a quantized Tensor. + */ + virtual Tensor quantize(const Tensor& t) = 0; + + /** + * dequantize a quantized Tensor into a float Tensor. + */ + virtual Tensor dequantize(const Tensor& t) = 0; + + /** + * dequantize a quantized Tensor into a float Tensor, out= variant + */ + virtual Tensor& dequantize_out(Tensor& out, const Tensor& t) = 0; + + /** + * Compare against `other` for equality. + */ + virtual bool equalTo(QuantizerPtr other) const = 0; +}; + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/Tensor.h b/backends/pytorch/include/ATen/core/Tensor.h new file mode 100644 index 000000000..cc8558a2b --- /dev/null +++ b/backends/pytorch/include/ATen/core/Tensor.h @@ -0,0 +1,103 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace at { +// NOLINTNEXTLINE(cppcoreguidelines-special-member-functions) +class TORCH_API OptionalTensorRef { + public: + OptionalTensorRef() = default; + + ~OptionalTensorRef() { + ref_.unsafeReleaseTensorImpl(); + } + + OptionalTensorRef(const TensorBase& src) + : ref_(Tensor::unsafe_borrow_t{}, src) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(src.defined()); + } + + OptionalTensorRef(const OptionalTensorRef& rhs) + : ref_(Tensor::unsafe_borrow_t{}, rhs.ref_) {} + + OptionalTensorRef(OptionalTensorRef&& rhs) = default; + OptionalTensorRef& operator=(OptionalTensorRef rhs) { + std::swap(ref_, rhs.ref_); + return *this; + } + + bool has_value() const { + return ref_.defined(); + } + + const Tensor& getTensorRef() const & { + return ref_; + } + + const Tensor& operator*() const & { + return ref_; + } + + const Tensor* operator->() const & { + return &ref_; + } + + operator bool() const { + return ref_.defined(); + } + + private: + Tensor ref_; +}; + +// Use to convert a TensorBase (that may be undefined) to an at::Tensor +// without bumping refcount. +class TORCH_API TensorRef { + public: + ~TensorRef() { + ref_.unsafeReleaseTensorImpl(); + } + + TensorRef(const TensorBase& src) + : ref_(Tensor::unsafe_borrow_t{}, src) {} + TensorRef(TensorRef&& other) = default; + TensorRef(const TensorRef&) = default; + TensorRef& operator=(const TensorRef&) = default; + TensorRef& operator=(TensorRef&&) = default; + + const Tensor& operator*() const & { + return ref_; + } + private: + Tensor ref_; +}; + +template +auto Tensor::register_hook(T&& hook) const -> Tensor::hook_return_void_t { + // Return the grad argument in case of a hook with void return type to have an + // std::function with Tensor return type + static_assert(std::is_same_v, + "Expected hook to return void"); + return _register_hook([fn=std::forward(hook)](const TensorBase& grad_base) { + TensorRef grad(grad_base); + fn(*grad); + return Tensor(); + }); +} + +template +auto Tensor::register_hook(T&& hook) const -> Tensor::hook_return_var_t { + return _register_hook([fn=std::forward(hook)](const TensorBase& grad_base) { + TensorRef grad(grad_base); + Tensor ret = fn(*grad); + return TensorBase(std::move(ret)); + }); +} + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/TensorAccessor.h b/backends/pytorch/include/ATen/core/TensorAccessor.h new file mode 100644 index 000000000..4fc9c1e8a --- /dev/null +++ b/backends/pytorch/include/ATen/core/TensorAccessor.h @@ -0,0 +1,66 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace at { + +using torch::headeronly::DefaultPtrTraits; +#if defined(__CUDACC__) || defined(__HIPCC__) + using torch::headeronly::RestrictPtrTraits; +#endif + +template class PtrTraits = DefaultPtrTraits, typename index_t = int64_t> +using TensorAccessorBase = torch::headeronly::detail::TensorAccessorBase; + +template class PtrTraits = DefaultPtrTraits, typename index_t = int64_t> +using TensorAccessor = torch::headeronly::detail::TensorAccessor; + +namespace detail { + +template +struct IndexBoundsCheck { + IndexBoundsCheck(index_t i) { + TORCH_CHECK_INDEX( + 0 <= i && i < index_t{N}, + "Index ", + i, + " is not within bounds of a tensor of dimension ", + N); + } +}; +} // namespace detail + +template class PtrTraits = DefaultPtrTraits, typename index_t = int64_t> +using GenericPackedTensorAccessorBase = torch::headeronly::detail::GenericPackedTensorAccessorBase, T, N, PtrTraits, index_t>; + +template class PtrTraits = DefaultPtrTraits, typename index_t = int64_t> +using GenericPackedTensorAccessor = torch::headeronly::detail::GenericPackedTensorAccessor, detail::IndexBoundsCheck, T, N, PtrTraits, index_t>; + +// Can't put this directly into the macro function args because of commas +#define AT_X GenericPackedTensorAccessor + +// Old name for `GenericPackedTensorAccessor` +template class PtrTraits = DefaultPtrTraits, typename index_t = int64_t> +C10_DEFINE_DEPRECATED_USING(PackedTensorAccessor, AT_X) + +#undef AT_X + +template class PtrTraits = DefaultPtrTraits> +using PackedTensorAccessor32 = GenericPackedTensorAccessor; + +template class PtrTraits = DefaultPtrTraits> +using PackedTensorAccessor64 = GenericPackedTensorAccessor; +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/TensorBase.h b/backends/pytorch/include/ATen/core/TensorBase.h new file mode 100644 index 000000000..3244b247a --- /dev/null +++ b/backends/pytorch/include/ATen/core/TensorBase.h @@ -0,0 +1,1098 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// See https://github.com/pytorch/pytorch/issues/161660 +// This compile flag is intended to be passed in to CppExtensions that rely on +// the stable ABI via the `extra_compile_args` argument. This is a stopgap +// solution to ensure that non-stable libtorch APIs are not used in the extension. +// The long term solution is to have a torch_stable target that excludes headers +// that are not in torch/stable or torch/headeronly. +// See test/cpp_extensions/torch_stable_test_extension/setup.py for an example +// of how this is used. +#ifdef TORCH_STABLE_ONLY +#error \ + "TensorBase.h should not be included when TORCH_STABLE_ONLY compile flag is passed" +#endif + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include + +namespace c10 { +class Scalar; +} + +namespace torch::autograd { + +struct Node; + +} // namespace torch::autograd + +namespace at { + +class Tensor; +class TensorBase; + +// Convert Tensor to TensorBase without any need to include Tensor.h +TORCH_API const TensorBase& get_tensor_base(const Tensor& t); + +namespace impl { +inline bool variable_excluded_from_dispatch() { +#ifdef C10_MOBILE + // Please read the comment in `VariableFallbackKernel.cpp` about the background of this change. + return true; +#else + return c10::impl::tls_local_dispatch_key_set().excluded_.isSupersetOf(c10::autograd_dispatch_keyset); +#endif +} + +} + +// NOTE: [Tensor vs. TensorBase] +// +// Tensor, being the central data structure in PyTorch, gets used and +// its header included almost everywhere. Unfortunately this means +// every time an operator signature is updated or changed in +// native_functions.yaml, you (and every other PyTorch developer) need +// to recompile all of ATen and its dependencies. +// +// TensorBase aims to break up these header dependencies, and improve +// incremental build times for all PyTorch developers. TensorBase +// represents a reference counted handle to TensorImpl, exactly the +// same as Tensor. However, TensorBase doesn't have code generated +// methods in its API and thus no dependence on native_functions.yaml. +// +// Usage tips +// ---------- +// - You can `#define TORCH_ASSERT_NO_OPERATORS` at the top of a .cpp +// or .cu file to ensure it has no header dependencies on +// native_functions.yaml (direct or indirect). +// - Tensor inherits from TensorBase, so functions taking +// `const TensorBase &` are callable with Tensor as well. +// - TensorBase can be converted to Tensor with `Tensor(tensor_base)`, +// but this requires a reference-count bump. OptionalTensorRef, on +// the other hand, can materialize a `const Tensor &` without +// touching the reference-count. +class TORCH_API TensorBase { + public: + struct unsafe_borrow_t { explicit unsafe_borrow_t() = default; }; + + protected: + // Create a Tensor with a +0 reference count. Special care must be + // taken to avoid decrementing this reference count at destruction + // time. Intended to support MaybeOwnedTraits. + explicit TensorBase(unsafe_borrow_t /*unused*/, const TensorBase& rhs) + : impl_(c10::intrusive_ptr(rhs.impl_.get(), c10::raw::DontIncreaseRefcount{})) {} + friend MaybeOwnedTraits; + + public: + TensorBase() = default; + // This constructor should not be used by end users and is an implementation + // detail invoked by autogenerated code. + explicit TensorBase( + c10::intrusive_ptr tensor_impl) + : impl_(std::move(tensor_impl)) { + TORCH_CHECK(impl_.get(), "TensorImpl with nullptr is not supported"); + } + TensorBase(const TensorBase&) = default; + TensorBase(TensorBase&&) noexcept = default; + ~TensorBase() noexcept = default; + + public: + // Creates a new wrapper from TensorImpl. Intentionally a free method because + // it should be used with care. Checks necessary invariants + static TensorBase wrap_tensor_impl( + c10::intrusive_ptr tensor_impl) { + TensorBase r(std::move(tensor_impl)); + r.enforce_invariants(); + return r; + } + + int64_t dim() const { + return impl_->dim(); + } + int64_t storage_offset() const { + return impl_->storage_offset(); + } + + TensorBase contiguous(MemoryFormat memory_format=MemoryFormat::Contiguous) const { + if (is_contiguous_or_false(memory_format)) { + return *this; + } else { + return __dispatch_contiguous(memory_format); + } + } + + /// Should be used if *this can reasonably be expected to be contiguous and + /// performance is important. + /// Compared to contiguous, it saves a reference count + /// increment/decrement if *this is already contiguous, at the cost + /// in all cases of an extra pointer of stack usage, an extra branch + /// to access, and an extra branch at destruction time. + c10::MaybeOwned expect_contiguous( + MemoryFormat memory_format=MemoryFormat::Contiguous) const &; + + // Use .contiguous() instead. Trying to borrow from a prvalue + // will only lead to trouble and dangling references. + c10::MaybeOwned expect_contiguous( + MemoryFormat memory_format=MemoryFormat::Contiguous) && = delete; + + const TensorBase& fill_(const c10::Scalar& scalar) const; + const TensorBase& zero_() const; + + TensorBase to(at::TensorOptions options={}, bool non_blocking=false, bool copy=false, std::optional memory_format=std::nullopt) const; + + bool is_complex() const { + return at::isComplexType(this->scalar_type()); + } + + bool is_floating_point() const { + return at::isFloatingType(this->scalar_type()); + } + + bool is_signed() const { + return at::isSignedType(this->scalar_type()); + } + + c10::SymInt sym_size(int64_t dim) const { + return impl_->sym_size(dim); + } + + c10::SymInt sym_stride(int64_t dim) const { + const auto sizes = this->sym_strides(); + const auto ndim = static_cast(sizes.size()); + // false is passed to maybe_wrap_dim so behavior is identical to array access (but with wrapping) + return sizes[c10::maybe_wrap_dim(dim, ndim, /*wrap_scalar=*/false)]; + + } + + int64_t size(int64_t dim) const { + return impl_->size(dim); + } + + int64_t stride(int64_t dim) const { + const auto strides = this->strides(); + const auto ndim = static_cast(strides.size()); + // false is passed to maybe_wrap_dim so behavior is identical to array access (but with wrapping) + return strides[c10::maybe_wrap_dim(dim, ndim, /*wrap_scalar=*/false)]; + } + + TensorImpl * unsafeGetTensorImpl() const { + return impl_.get(); + } + TensorImpl * unsafeReleaseTensorImpl() { + return impl_.release(); + } + const c10::intrusive_ptr& getIntrusivePtr() const { + return impl_; + } + + c10::intrusive_ptr unsafeReleaseIntrusivePtr() { + return std::move(impl_); + } + + bool defined() const { + return impl_; + } + + void reset() { + impl_.reset(); + } + +#if defined (_MSC_VER) + TensorBase& operator=(const TensorBase& x) & { + impl_ = x.impl_; + return *this; + }; + TensorBase& operator=(TensorBase&& x) & noexcept { + impl_ = std::move(x.impl_); + return *this; + } +#else + TensorBase& operator=(const TensorBase& x) & = default; + TensorBase& operator=(TensorBase&& x) & noexcept = default; +#endif + + // Ban assignment to rvalues, since at::Tensor (weirdly) performs a deep copy here + TensorBase& operator=(const TensorBase&) && = delete; + TensorBase& operator=(TensorBase&&) && noexcept = delete; + + bool is_same(const TensorBase& other) const noexcept { + return impl_ == other.impl_; + } + size_t use_count() const noexcept { + return impl_.use_count(); + } + size_t weak_use_count() const noexcept { + return impl_.weak_use_count(); + } + bool is_uniquely_owned() const noexcept { + return impl_.is_uniquely_owned(); + } + + std::string toString() const; + + IntArrayRef sizes() const { + return impl_->sizes(); + } + c10::SymIntArrayRef sym_sizes() const { + return impl_->sym_sizes(); + } + c10::SymIntArrayRef sym_strides() const { + return impl_->sym_strides(); + } + IntArrayRef strides() const { + return impl_->strides(); + } + // See impl::get_opt_names in ATen/NamedTensor.h for docs. + std::optional opt_names() const { + return impl::get_opt_names(unsafeGetTensorImpl()); + } + // See impl::get_names in ATen/NamedTensor.h for docs. + DimnameList names() const { + return impl::get_names(unsafeGetTensorImpl()); + } + int64_t ndimension() const { + return dim(); + } + + bool is_contiguous(at::MemoryFormat memory_format=at::MemoryFormat::Contiguous) const { + return impl_->is_contiguous(memory_format); + } + + // Like is_contiguous, but more dynamic shape-friendly. May return a symbolic representation of + // contiguity instead of SymTrue SymFalse, when results are data-dependent. + c10::SymBool sym_is_contiguous(at::MemoryFormat memory_format=at::MemoryFormat::Contiguous) const { + if (impl_->has_symbolic_sizes_strides()) { + return impl_->sym_is_contiguous(memory_format); + } + return impl_->is_contiguous(memory_format); + } + + // Like is_contiguous, but more dynamic shape-friendly. Can returns + // false instead of throwing data-dependent errors for tensors with unbacked + // sizes or strides. + bool is_contiguous_or_false(at::MemoryFormat memory_format=at::MemoryFormat::Contiguous) const { + if (impl_->has_symbolic_sizes_strides()) { + return impl_->sym_is_contiguous(memory_format).guard_or_false(__FILE__, __LINE__); + } + return impl_->is_contiguous(memory_format); + } + + bool is_non_overlapping_and_dense() const { + return impl_->is_non_overlapping_and_dense(); + } + + at::MemoryFormat suggest_memory_format( + bool channels_last_strides_exact_match = false) const { + // Setting channels_last_strides_exact_match to true forces function to + // check 0,1 - sized dimension strides. + if (layout() == at::kStrided) { + if (impl_->is_strides_like_channels_last()) { + if (!channels_last_strides_exact_match || + get_channels_last_strides_2d(sizes()) == strides()) { + return at::MemoryFormat::ChannelsLast; + } + } + else if (impl_->is_strides_like_channels_last_3d()) { + if (!channels_last_strides_exact_match || + get_channels_last_strides_3d(sizes()) == strides()) { + return at::MemoryFormat::ChannelsLast3d; + } + } + } + return at::MemoryFormat::Contiguous; + } + + // Total bytes consumed by the "view" of elements of the array. Does not + // include size of metadata. The number reported here does not necessarily + // correspond to the true physical memory consumed by a tensor; instead, + // it reports the memory the tensor would take *if* it were contiguous. + // Defined to be numel() * itemsize() + size_t nbytes() const { + TORCH_CHECK(layout () != at::kSparse, + "nbytes is not defined for sparse tensors. If you want the size of the constituent " \ + "tensors, add the nbytes of the indices and values. If you want the size of the " \ + "equivalent dense tensor, multiply numel() by element_size()"); + return impl_->numel() * impl_->itemsize(); + } + + c10::SymInt sym_nbytes() const { + TORCH_CHECK(layout () != at::kSparse, + "nbytes is not defined for sparse tensors. If you want the size of the constituent " \ + "tensors, add the nbytes of the indices and values. If you want the size of the " \ + "equivalent dense tensor, multiply numel() by element_size()"); + return impl_->sym_numel() * impl_->itemsize(); + } + + int64_t numel() const { + return impl_->numel(); + } + + c10::SymInt sym_numel() const { + return impl_->sym_numel(); + } + + c10::SymInt sym_storage_offset() const { + return impl_->sym_storage_offset(); + } + + // Length of one array element in bytes. This is the traditional + // Numpy naming. + size_t itemsize() const { + return impl_->itemsize(); + } + + // Same as itemsize(). This is the PyTorch naming. + int64_t element_size() const { + return static_cast(impl_->itemsize()); + } + + DispatchKeySet key_set() const { + return impl_->key_set(); + } + ScalarType scalar_type() const { + return typeMetaToScalarType(impl_->dtype()); + } + bool has_storage() const { + return defined() && impl_->has_storage(); + } + const Storage& storage() const { + return impl_->storage(); + } + bool is_alias_of(const at::TensorBase& other) const{ + return impl_->storage().is_alias_of(other.storage()); + } + + // Move the storage backend to shm based + // to enable memory sharing across processes. + // + // NB1: the ideal behavior of this API still requires further discussion + // but for now we are inclined to keep it consistent with existing THP behavior + // https://github.com/pytorch/pytorch/blob/4dca9bde0552afc67b5b74f4a0696fe6055709c4/torch/storage.py#L196-L212 + // so we don't assert on anything here and rely on caller knowing + // what it's doing. + // + // NB2: this currently provides Linux fd based shm support only + // to simplify the storage lifetime management logic in ATen + // and similarly for now we are not adding support for file system based + // shm support like in THP due to additional GC manager support needed + // to prevent leaks. + // As such, calling this from non supported systems (e.g. Windows) would fail. + void share_memory_() { + at::share_memory_(*this); + } + + inline bool _is_zerotensor() const { + return impl_->_is_zerotensor(); + } + + inline void _set_zero(bool zero) const { + impl_->_set_zero(zero); + } + + inline bool is_conj() const { + return impl_->is_conj(); + } + + // sets the conjugate bit of a tensor. + // NOTE: Conjugate bit is supposed to be a read-only field. Only change this, if you are sure + // that's what you want. Changing this might lead to incorrect behavior since conjugation is + // a lazy operation and we rely on this bit to determine if a conjugation needs to be materialized. + inline void _set_conj(bool conjugate) const { + impl_->_set_conj(conjugate); + } + + inline bool is_neg() const { + return impl_->is_neg(); + } + + // sets the negative bit of a tensor. + // NOTE: Negative bit is supposed to be a read-only field. Only change this, if you are sure + // that's what you want. Changing this might lead to incorrect behavior since we rely on this + // bit to determine if a negation needs to be materialized. + inline void _set_neg(bool negative) const { + impl_->_set_neg(negative); + } + + /// Returns a `Tensor`'s layout. + Layout layout() const { + return impl_->layout(); + } + + /// Returns a `Tensor`'s dtype (`TypeMeta`). + caffe2::TypeMeta dtype() const { + return impl_->dtype(); + } + + /// Returns a `Tensor`'s device. + inline Device device() const { + return impl_->device(); + } + + /// Returns a `Tensor`'s device index. + DeviceIndex get_device() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->get_device(); + } + + /// Returns if a `Tensor` has CPU backend. + bool is_cpu() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_cpu(); + } + + /// Returns if a `Tensor` has CUDA backend. + bool is_cuda() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_cuda(); + } + + /// Returns if a `Tensor` has IPU backend. + bool is_ipu() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_ipu(); + } + + /// Returns if a `Tensor` has XPU backend. + bool is_xpu() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_xpu(); + } + + /// Returns if a `Tensor` has XLA backend. + bool is_xla() const { + return impl_->is_xla(); + } + + /// Returns if a `Tensor` has MTIA backend. + bool is_mtia() const { + return impl_->is_mtia(); + } + + /// Returns if a `Tensor` has HPU backend. + bool is_hpu() const { + return impl_->is_hpu(); + } + + /// Returns if a `Tensor` has Lazy backend. + bool is_lazy() const { + return impl_->is_lazy(); + } + + /// Returns if a `Tensor` has HIP backend. + bool is_hip() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_hip(); + } + + /// Returns if a `Tensor` has VE backend. + bool is_ve() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_ve(); + } + + /// Returns if a `Tensor` has PrivateUse1 backend. + bool is_privateuseone() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_privateuseone(); + } + + /// Returns if a `Tensor` has sparse backend. + bool is_sparse() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_sparse(); + } + + /// Returns is a `Tensor` has a sparse CSR backend. + bool is_sparse_csr() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_sparse_csr(); + } + + /// Returns if a `Tensor` is mkldnn tensor. + bool is_mkldnn() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_mkldnn(); + } + + /// Returns if a `Tensor` is mps tensor. + bool is_mps() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_mps(); + } + + /// Returns if a `Tensor` is maia tensor. + bool is_maia() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_maia(); + } + + /// Returns if a `Tensor` is vulkan tensor. + bool is_vulkan() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_vulkan(); + } + + /// Returns if a `Tensor` is metal tensor. + bool is_metal() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_metal(); + } + + /// Returns if a `Tensor` has quantized backend. + bool is_quantized() const { + // NB: this is not a native function to avoid dispatching overhead. + return impl_->is_quantized(); + } + + /// Returns if a `Tensor` is a meta tensor. Meta tensors can + /// also have other designations. + bool is_meta() const { + return impl_->is_meta(); + } + + /// Returns if a `Tensor` is an inference tensor. + bool is_inference() const { + return impl_->is_inference(); + } + + // Returns if a `Tensor` is a NestedTensor. + bool is_nested() const { + return impl_->is_nested(); + } + + /// If a tensor is a quantized tensor, returns its quantizer + /// TODO: it's not in native_functions.yaml yet as it's not exposed to python + QuantizerPtr quantizer() const; + + /// Returns if a `Tensor` has any dimension names + bool has_names() const { + // If a user is using unnamed tensors, then we can short-circuit right here. + // Otherwise, impl::has_names attempts to retrieve names. + if (!impl_->has_named_tensor_meta()) { + return false; + } + return impl::has_names(unsafeGetTensorImpl()); + } + + /// Returns a `Tensor`'s dimension names data structure + const NamedTensorMeta* get_named_tensor_meta() const { + return static_cast(impl_->named_tensor_meta()); + } + + NamedTensorMeta* get_named_tensor_meta() { + return static_cast(impl_->named_tensor_meta()); + } + + /// Returns the `TensorOptions` corresponding to this `Tensor`. Defined in + /// TensorOptions.h. + TensorOptions options() const { + return TensorOptions().dtype(dtype()) + .device(device()) + .layout(layout()); + } + + const void* const_data_ptr() const { + return this->unsafeGetTensorImpl()->data(); + } + + void* mutable_data_ptr() const { + return this->unsafeGetTensorImpl()->mutable_data(); + } + + // TODO(#97856) Make this return a const pointer. This currently + // returns a non-const pointer because of the large + // number of clients that we still want to audit before + // migrating to mutable_data_ptr(). + void* data_ptr() const { + return mutable_data_ptr(); + } + + template , int> = 0> + const T* const_data_ptr() const; + + template , int> = 0> + const std::remove_const_t* const_data_ptr() const; + + template + T* mutable_data_ptr() const; + + // Legacy interface during the migration to indicate that a callsite + // has not been audited for mutability. + // + // Do not add new uses of this, use const_data_ptr() if possible, + // mutable_data_ptr() otherwise. + // + // TODO(#97856) Make this return a const pointer. This is currently + // const because of the vast number of clients that + // rely on this. + template + T* data_ptr() const; + + // Purposely not defined here to avoid inlining + void print() const; + + // Return a `TensorAccessor` for CPU `Tensor`s. You have to specify scalar type and + // dimension. + template + TensorAccessor accessor() const& { + static_assert(N > 0, "accessor is used for indexing tensor, for scalars use *data_ptr()"); + TORCH_CHECK(dim() == N, "TensorAccessor expected ", N, " dims but tensor has ", dim()); + T* ptr = nullptr; + if constexpr (std::is_const_v) { + ptr = const_data_ptr(); + } else { + ptr = mutable_data_ptr(); + } + return TensorAccessor(ptr,sizes().data(),strides().data()); + } + template + TensorAccessor accessor() && = delete; + + // Return a `GenericPackedTensorAccessor` for CUDA `Tensor`s. You have to specify scalar type and + // dimension. You can optionally specify RestrictPtrTraits as a template parameter to + // cast the data pointer to a __restrict__ pointer. + // In order to use this, your CUDA kernel has to take a corresponding GenericPackedTensorAccessor + // as an argument. + template class PtrTraits = DefaultPtrTraits, typename index_t = int64_t> + GenericPackedTensorAccessor generic_packed_accessor() const& { + static_assert(N > 0, "accessor is used for indexing tensor, for scalars use *data_ptr()"); + TORCH_CHECK(dim() == N, "TensorAccessor expected ", N, " dims but tensor has ", dim()); + T* ptr = nullptr; + if constexpr (std::is_const_v) { + ptr = const_data_ptr(); + } else { + ptr = mutable_data_ptr(); + } + return GenericPackedTensorAccessor(static_cast::PtrType>(ptr),sizes().data(),strides().data()); + } + template class PtrTraits = DefaultPtrTraits, typename index_t = int64_t> + GenericPackedTensorAccessor generic_packed_accessor() && = delete; + + template class PtrTraits = DefaultPtrTraits> + PackedTensorAccessor32 packed_accessor32() const& { + TORCH_CHECK( + impl_->numel() <= + static_cast(std::numeric_limits::max()), + "numel needs to be smaller than int32_t max; otherwise, please use packed_accessor64"); + return generic_packed_accessor(); + } + template class PtrTraits = DefaultPtrTraits> + PackedTensorAccessor32 packed_accessor32() && = delete; + + template class PtrTraits = DefaultPtrTraits> + PackedTensorAccessor64 packed_accessor64() const& { + return generic_packed_accessor(); + } + template class PtrTraits = DefaultPtrTraits> + PackedTensorAccessor64 packed_accessor64() && = delete; + + // ~~~~~ Autograd API ~~~~~ + + /// \fn bool is_leaf() const; + /// + /// All Tensors that have `requires_grad()` which is ``false`` will be leaf Tensors by convention. + /// + /// For Tensors that have `requires_grad()` which is ``true``, they will be leaf Tensors if they were + /// created by the user. This means that they are not the result of an operation and so + /// `grad_fn()` is `nullptr`. + /// + /// Only leaf Tensors will have their `grad()` populated during a call to `backward()`. + /// To get `grad()` populated for non-leaf Tensors, you can use `retain_grad()`. + /// + /// Example: + /// @code + /// auto a = torch::rand(10, torch::requires_grad()); + /// std::cout << a.is_leaf() << std::endl; // prints `true` + /// + /// auto b = torch::rand(10, torch::requires_grad()).to(torch::kCUDA); + /// std::cout << b.is_leaf() << std::endl; // prints `false` + /// // b was created by the operation that cast a cpu Tensor into a cuda Tensor + /// + /// auto c = torch::rand(10, torch::requires_grad()) + 2; + /// std::cout << c.is_leaf() << std::endl; // prints `false` + /// // c was created by the addition operation + /// + /// auto d = torch::rand(10).cuda(); + /// std::cout << d.is_leaf() << std::endl; // prints `true` + /// // d does not require gradients and so has no operation creating it (that is tracked by the autograd engine) + /// + /// auto e = torch::rand(10).cuda().requires_grad_(); + /// std::cout << e.is_leaf() << std::endl; // prints `true` + /// // e requires gradients and has no operations creating it + /// + /// auto f = torch::rand(10, torch::device(torch::kCUDA).requires_grad(true)); + /// std::cout << f.is_leaf() << std::endl; // prints `true` + /// // f requires grad, has no operation creating it + /// @endcode + + /// \fn void backward(const Tensor & gradient={}, std::optional retain_graph=std::nullopt, bool create_graph=false, std::optional inputs=std::nullopt) const; + /// + /// Computes the gradient of current tensor with respect to graph leaves. + /// + /// The graph is differentiated using the chain rule. If the tensor is + /// non-scalar (i.e. its data has more than one element) and requires + /// gradient, the function additionally requires specifying ``gradient``. + /// It should be a tensor of matching type and location, that contains + /// the gradient of the differentiated function w.r.t. this Tensor. + /// + /// This function accumulates gradients in the leaves - you might need to + /// zero them before calling it. + /// + /// \param gradient Gradient w.r.t. the + /// tensor. If it is a tensor, it will be automatically converted + /// to a Tensor that does not require grad unless ``create_graph`` is True. + /// None values can be specified for scalar Tensors or ones that + /// don't require grad. If a None value would be acceptable then + /// this argument is optional. + /// \param retain_graph If ``false``, the graph used to compute + /// the grads will be freed. Note that in nearly all cases setting + /// this option to True is not needed and often can be worked around + /// in a much more efficient way. Defaults to the value of + /// ``create_graph``. + /// \param create_graph If ``true``, graph of the derivative will + /// be constructed, allowing to compute higher order derivative + /// products. Defaults to ``false``. + /// \param inputs Inputs w.r.t. which the gradient will be accumulated into + /// ``at::Tensor::grad``. All other Tensors will be ignored. If not + /// provided, the gradient is accumulated into all the leaf Tensors + /// that were used to compute the current tensor. + /// When inputs are provided and a given input is not a leaf, + /// the current implementation will call its grad_fn (even though it is not strictly needed to get this gradients). + /// It is an implementation detail on which the user should not rely. + /// See https://github.com/pytorch/pytorch/pull/60521#issuecomment-867061780 for more details. + + /// \fn Tensor detach() const; + /// + /// Returns a new Tensor, detached from the current graph. + /// The result will never require gradient. + + /// \fn Tensor & detach_() const; + /// + /// Detaches the Tensor from the graph that created it, making it a leaf. + /// Views cannot be detached in-place. + + /// \fn void retain_grad() const; + /// + /// Enables this Tensor to have their :attr:`grad` populated during + /// :func:`backward`. This is a no-op for leaf tensors. + + /// \fn bool retains_grad() const; + /// + /// Is ``true`` if this Tensor is non-leaf and its :attr:`grad` is enabled to be + /// populated during :func:`backward`, ``false`` otherwise. + + const TensorBase& set_requires_grad(bool requires_grad) const { + impl_->set_requires_grad(requires_grad); + return *this; + } + bool requires_grad() const { + return impl_->requires_grad(); + } + + // The Forward AD API functions below are low level and are not to be used by end + // users who should use the API provided in torch/csrc/autograd.h + + /// This function returns the forward gradient for this Tensor at the given level. + const Tensor& _fw_grad(uint64_t level) const { + return impl_->_fw_grad(level, *this); + } + + /// This function can be used to set the value of the forward grad. + /// Note that the given new_grad might not be used directly if it has different + /// metadata (size/stride/storage offset) compared to this Tensor. In that case, + /// new_grad content will be copied into a new Tensor + void _set_fw_grad(const TensorBase& new_grad, uint64_t level, bool is_inplace_op) const { + impl_->_set_fw_grad(new_grad, *this, level, is_inplace_op); + } + + /// NOTE: This is similar to the legacy `.data()` function on `Variable`, and is intended + /// to be used from functions that need to access the `Variable`'s equivalent `Tensor` + /// (i.e. `Tensor` that shares the same storage and tensor metadata with the `Variable`). + /// + /// One notable difference with the legacy `.data()` function is that changes to the + /// returned `Tensor`'s tensor metadata (e.g. sizes / strides / storage / storage_offset) + /// will not update the original `Variable`, due to the fact that this function + /// shallow-copies the `Variable`'s underlying TensorImpl. + at::TensorBase tensor_data() const; + + /// NOTE: `var.variable_data()` in C++ has the same semantics as `tensor.data` + /// in Python, which create a new `Variable` that shares the same storage and + /// tensor metadata with the original `Variable`, but with a completely new + /// autograd history. + /// + /// NOTE: If we change the tensor metadata (e.g. sizes / strides / + /// storage / storage_offset) of a variable created from `var.variable_data()`, those + /// changes will not update the original variable `var`. In `.variable_data()`, we set + /// `allow_tensor_metadata_change_` to false to make such changes explicitly illegal, + /// in order to prevent users from changing metadata of `var.variable_data()` + /// and expecting the original variable `var` to also be updated. + at::TensorBase variable_data() const; + + // Gradient Node and Edges + //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + /// Gets the gradient function of the `Variable`. If this is a leaf variable, + /// the pointer returned will be null. + /// + /// For View Variables: + /// Gets the up-to-date grad_fn. If the shared data or base was modified, we + /// re-create the grad_fn to express the up-to-date view relationship between + /// this and the base Variable. + const std::shared_ptr& grad_fn() const; + + // Hooks + //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + template + using hook_return_void_t = std::enable_if_t>, unsigned>; + template + using hook_return_var_t = std::enable_if_t, TensorBase>, unsigned>; + + /// Registers a backward hook. + /// + /// The hook will be called every time a gradient with respect to the Tensor is computed. + /// The hook should have one of the following signature: + /// ``` + /// hook(TensorBase grad) -> TensorBase + /// ``` + /// ``` + /// hook(TensorBase grad) -> void + /// ``` + /// The hook should not modify its argument, but it can optionally return a new gradient + /// which will be used in place of `grad`. + /// + /// This function returns the index of the hook in the list which can be used to remove hook. + /// + /// Example: + /// @code + /// auto v = torch::tensor({0., 0., 0.}, torch::requires_grad()); + /// auto h = v.register_hook([](torch::Tensor grad){ return grad * 2; }); // double the gradient + /// v.backward(torch::tensor({1., 2., 3.})); + /// // This prints: + /// // ``` + /// // 2 + /// // 4 + /// // 6 + /// // [ CPUFloatType{3} ] + /// // ``` + /// std::cout << v.grad() << std::endl; + /// v.remove_hook(h); // removes the hook + /// @endcode + template + hook_return_void_t register_hook(T&& hook) const; + template + hook_return_var_t register_hook(T&& hook) const; + +protected: + unsigned _register_hook(std::function hook) const; + +public: + + /// Remove hook at given position + void remove_hook(unsigned pos) const; + + // Variable methods + //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + bool is_leaf() const; + + int64_t output_nr() const; + + void set_data(const TensorBase & new_data) const; + + TensorBase data() const; + + int64_t _version() const; + + void retain_grad() const; + + bool retains_grad() const; + + const TensorBase& requires_grad_(bool _requires_grad=true) const; + + std::optional grad_dtype() const; + + void set_grad_dtype(const std::optional& grad_dtype) const; + + // View Variables + //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + /// Returns true if this `Variable` is a view of another `Variable`. + bool is_view() const; + + /// Returns the `Variable` that this `Variable` is a view of. If this + /// `Variable` is not a view, throw a `std::runtime_error`. + const TensorBase& _base() const; + + // Miscellaneous + //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + const std::string& name() const; + +protected: + void enforce_invariants(); + c10::intrusive_ptr impl_; + +private: + TensorBase __dispatch_contiguous(c10::MemoryFormat /*memory_format*/) const; +}; + +inline DeviceIndex get_device(const TensorBase& self) { + return self.get_device(); +} + +template +auto TensorBase::register_hook(T&& hook) const -> TensorBase::hook_return_void_t { + // Return the grad argument in case of a hook with void return type to have an + // std::function with Tensor return type + static_assert(std::is_same_v, + "Expected hook to return void"); + return _register_hook([fn=std::forward(hook)](const TensorBase& grad) { + fn(grad); + return TensorBase(); + }); +} + +template +auto TensorBase::register_hook(T&& hook) const -> TensorBase::hook_return_var_t { + return _register_hook(std::forward(hook)); +} + +namespace detail { +// Helper creator for Tensor class which doesn't requires the users to pass +// in an intrusive_ptr instead it just converts the argument passed to +// requested intrusive_ptr type. +template +TensorBase make_tensor_base(Args&&... args) { + return TensorBase(c10::make_intrusive(std::forward(args)...)); +} + +} // namespace detail + +inline DispatchKey legacyExtractDispatchKey(const TensorBase& t) { + return legacyExtractDispatchKey(t.key_set()); +} + +} // namespace at + +namespace c10 { +template <> +struct MaybeOwnedTraits { + using owned_type = at::TensorBase; + using borrow_type = at::TensorBase; + + static borrow_type createBorrow(const owned_type& from) { + // NOTE: this can be implemented without the special + // unsafe_borrow_t Tensor constructor as + // + // return borrow_type(c10::intrusive_ptr::reclaim(from.unsafeGetTensorImpl())); + // + // but that hurts inlining due to the nullptr check in the + // Tensor(c10::intrusive_ptr<...>) constructor. We already know + // that from.impl_ isn't null because from is a valid Tensor, so + // we needn't do the check again. (using __builtin_assume can + // avoid this, but wouldn't be portable to MSVC.) + return borrow_type(borrow_type::unsafe_borrow_t{}, from); + } + + static void assignBorrow(borrow_type& lhs, const borrow_type& rhs) { + lhs.unsafeReleaseTensorImpl(); + // See above note: this can be implemented with public API + // similarly to createBorrow(), but that would hurt inlining. + lhs = borrow_type(borrow_type::unsafe_borrow_t{}, rhs); + } + + static void destroyBorrow(borrow_type& toDestroy) { + toDestroy.unsafeReleaseTensorImpl(); // "leak" it, but it was already +0. + } + + static const owned_type& referenceFromBorrow(const borrow_type& borrow) { + return borrow; + } + + static const owned_type* pointerFromBorrow(const borrow_type& borrow) { + return &borrow; + } + + static bool debugBorrowIsValid(const borrow_type& /*borrow*/) { + return true; + } +}; + +template <> +struct ExclusivelyOwnedTraits : public c10::ExclusivelyOwnedTensorTraits {}; +} // namespace c10 + +namespace at { + +inline c10::MaybeOwned borrow_from_optional_tensor( + const std::optional& opt) { + return opt.has_value() + ? c10::MaybeOwned::borrowed(*opt) + : c10::MaybeOwned::owned(std::in_place); +} + +inline c10::MaybeOwned TensorBase::expect_contiguous(MemoryFormat memory_format) const & { + if (is_contiguous(memory_format)) { + return c10::MaybeOwned::borrowed(*this); + } else { + return c10::MaybeOwned::owned(__dispatch_contiguous(memory_format)); + } +} + +namespace symint { + +template +using enable_if_symint = std::enable_if_t>; +template +using enable_if_int = std::enable_if_t>; + +template > +c10::SymIntArrayRef sizes(const TensorBase& t) { return t.sym_sizes(); } +template > +IntArrayRef sizes(const TensorBase& t) { return t.sizes(); } + +template > +c10::SymInt size(const TensorBase& t, int64_t dim) { return t.sym_size(dim); } +template > +int64_t size(const TensorBase& t, int64_t dim) { return t.size(dim); } + +template > +c10::SymIntArrayRef strides(const TensorBase& t) { return t.sym_strides(); } +template > +IntArrayRef strides(const TensorBase& t) { return t.strides(); } + +template > +c10::SymInt numel(const TensorBase& t) { return t.sym_numel(); } +template > +int64_t numel(const TensorBase& t) { return t.numel(); } + +} // namespace symint + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/TensorBody.h b/backends/pytorch/include/ATen/core/TensorBody.h new file mode 100644 index 000000000..f67ce6fbf --- /dev/null +++ b/backends/pytorch/include/ATen/core/TensorBody.h @@ -0,0 +1,5799 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#ifdef TORCH_ASSERT_NO_OPERATORS +#error This change adds a dependency on native_functions.yaml, \ + meaning the file will need to be re-compiled every time an operator \ + is changed or added. Consider if your change would be better placed in \ + another file, or if a more specific header might achieve the same goal. \ + See NOTE: [Tensor vs. TensorBase] +#endif + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + + +#include + +namespace c10{ +template class List; +template class IListRef; +} +namespace at { +struct Generator; +struct Type; +class DeprecatedTypeProperties; +class Tensor; +} // namespace at +namespace at { +namespace indexing { +struct TensorIndex; +} // namespace indexing +} // namespace at + +namespace torch { namespace autograd { + +struct Node; + +}} // namespace torch::autograd + +namespace at { + +class OptionalTensorRef; +class TensorRef; +class Tensor; +using TensorList = ArrayRef; +using ITensorList = c10::IListRef; + +using Stream = c10::Stream; + +// Tensor is a "generic" object holding a pointer to the underlying TensorImpl object, which +// has an embedded reference count. In this way, Tensor is similar to boost::intrusive_ptr. +// +// For example: +// +// void func(Tensor a) { +// Tensor b = a; +// ... +// } +// +// In this example, when we say Tensor b = a, we are creating a new object that points to the +// same underlying TensorImpl, and bumps its reference count. When b goes out of scope, the +// destructor decrements the reference count by calling release() on the TensorImpl it points to. +// The existing constructors, operator overloads, etc. take care to implement the correct semantics. +// +// Note that Tensor can also be NULL, i.e. it is not associated with any underlying TensorImpl, and +// special care must be taken to handle this. +class TORCH_API Tensor: public TensorBase { + protected: + // Create a Tensor with a +0 reference count. Special care must be + // taken to avoid decrementing this reference count at destruction + // time. Intended to support MaybeOwnedTraits. + explicit Tensor(unsafe_borrow_t, const TensorBase& rhs): TensorBase(unsafe_borrow_t{}, rhs) {} + friend MaybeOwnedTraits; + friend OptionalTensorRef; + friend TensorRef; + + public: + Tensor() = default; + // This constructor should not be used by end users and is an implementation + // detail invoked by autogenerated code. + explicit Tensor( + c10::intrusive_ptr tensor_impl) + : TensorBase(std::move(tensor_impl)) {} + Tensor(const Tensor &tensor) = default; + Tensor(Tensor &&tensor) = default; + + // Implicitly move-constructible from TensorBase, but must be explicit to increase refcount + explicit Tensor(const TensorBase &base): TensorBase(base) {} + /*implicit*/ Tensor(TensorBase &&base): TensorBase(std::move(base)) {} + + // Creates a new wrapper from TensorImpl. Intentionally a free method because + // it should be used with care. Checks necessary invariants + static Tensor wrap_tensor_impl( + c10::intrusive_ptr tensor_impl) { + return TensorBase::wrap_tensor_impl(std::move(tensor_impl)); + } + + Tensor contiguous(MemoryFormat memory_format=MemoryFormat::Contiguous) const { + return TensorBase::contiguous(memory_format); + } + + Tensor conj() const { + if (!this->is_complex()) { + return *this; + } + + C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum") + switch (this->layout()) { + case at::kSparse: + case at::kSparseCsr: + case at::kSparseCsc: + case at::kSparseBsr: + case at::kSparseBsc: + return this->conj_physical(); + default: + return this->_conj(); + } + C10_DIAGNOSTIC_POP() + } + + // Aliased by Dimname overloads, so need explicit using + using TensorBase::size; + using TensorBase::sym_size; + using TensorBase::stride; + + /// Should be used if *this can reasonably be expected to be contiguous and + /// performance is important. + /// Compared to contiguous, it saves a reference count + /// increment/decrement if *this is already contiguous, at the cost + /// in all cases of an extra pointer of stack usage, an extra branch + /// to access, and an extra branch at destruction time. + c10::MaybeOwned expect_contiguous(MemoryFormat memory_format=MemoryFormat::Contiguous) const &; + + // Use .contiguous() instead. Trying to borrow from a prvalue Tensor + // will only lead to trouble and dangling references. + c10::MaybeOwned expect_contiguous(MemoryFormat memory_format=MemoryFormat::Contiguous) && = delete; + + // The following overloads are very intriguing. Consider the following + // program: + // + // x[1] = 3; + // + // We would expect that the first entry of x is written to 3. But how can we + // actually achieve this? x[1] evaluates to a tensor... + // + // The answer is, using a ref-qualifier. x[1] is an rvalue, which cannot be + // (profitably) assigned to in the traditional sense, so we overload + // assignment to mean, "Actually, copy 3 into the tensor data." This is done + // with an rvalue-reference ref-qualified overload (the methods with && at the + // end of their type.) + // + // There's one more fly in the ointment: We also want + // + // Tensor x = y; + // + // to work, and we want it NOT to copy. So we need a traditional operator= + // overload. But we MUST specify a mutable lvalue ref-qualifier, to + // disambiguate the traditional overload from the rvalue-reference + // ref-qualified overload. Otherwise, it will be ambiguous, because + // a non ref-qualified method is eligible for all situations. + + // Unfortunately, we have to write these constructors out manually + // to work around an MSVC bug: + // error C2580: 'at::Tensor &at::Tensor::operator =(const at::Tensor &) &': + // multiple versions of a defaulted special member functions are not allowed + // Tensor& operator=(const Tensor&) & = default; + // Tensor& operator=(Tensor&&) & = default; + + // Also MSVC will wrongly issue the following warning with the aforementioned fix + // warning C4522: 'at::Tensor': multiple assignment operators specified + // Let's just skip the warning. + // + // TODO: temporarily disabled + + Tensor& operator=(const TensorBase& x) & noexcept { + impl_ = x.getIntrusivePtr(); + return *this; + } + Tensor& operator=(TensorBase&& x) & noexcept { + impl_ = x.unsafeReleaseIntrusivePtr(); + return *this; + } + + Tensor& operator=(const Tensor &x) & noexcept { + return operator=(static_cast(x)); + } + Tensor& operator=(Tensor &&x) & noexcept { + return operator=(static_cast(x)); + } + + Tensor& operator=(const Scalar &v) && { + return fill_(v); + } + Tensor& operator=(const Tensor &rhs) && { + return copy_(rhs); + } + Tensor& operator=(Tensor&& rhs) && { + return copy_(rhs); + } + + C10_DEPRECATED_MESSAGE("Tensor.type() is deprecated. Instead use Tensor.options(), which in many cases (e.g. in a constructor) is a drop-in replacement. If you were using data from type(), that is now available from Tensor itself, so instead of tensor.type().scalar_type(), use tensor.scalar_type() instead and instead of tensor.type().backend() use tensor.device().") + DeprecatedTypeProperties & type() const { + return globalDeprecatedTypePropertiesRegistry().getDeprecatedTypeProperties( + dispatchKeyToBackend(legacyExtractDispatchKey(key_set())), + scalar_type()); + } + + Tensor toType(ScalarType t) const { + return to(options().dtype(t), /*non_blocking*/ false, /*copy*/ false); + } + + // TODO: Deprecate me + Tensor toBackend(Backend b) const { + return to(options().device(backendToDeviceType(b)).layout(layout_from_backend(b)), /*non_blocking*/ false, /*copy*/ false); + } + + C10_DEPRECATED_MESSAGE("Tensor.is_variable() is deprecated; everything is a variable now. (If you want to assert that variable has been appropriately handled already, use at::impl::variable_excluded_from_dispatch())") + bool is_variable() const noexcept { + return !at::impl::variable_excluded_from_dispatch(); + } + + template + C10_DEPRECATED_MESSAGE("Tensor.data() is deprecated. Please use Tensor.data_ptr() instead.") + T * data() const { + return data_ptr(); + } + + template + T item() const; + + template class PtrTraits = DefaultPtrTraits, typename index_t = int64_t> + C10_DEPRECATED_MESSAGE("packed_accessor is deprecated, use packed_accessor32 or packed_accessor64 instead") + GenericPackedTensorAccessor packed_accessor() const & { + return generic_packed_accessor(); + } + template class PtrTraits = DefaultPtrTraits, typename index_t = int64_t> + C10_DEPRECATED_MESSAGE("packed_accessor is deprecated, use packed_accessor32 or packed_accessor64 instead") + GenericPackedTensorAccessor packed_accessor() && = delete; + + Tensor operator~() const { + return bitwise_not(); + } + Tensor operator-() const { + return neg(); + } + Tensor& operator+=(const Tensor & other) { + return add_(other); + } + Tensor& operator+=(const Scalar & other) { + return add_(other); + } + Tensor& operator-=(const Tensor & other) { + return sub_(other); + } + Tensor& operator-=(const Scalar & other) { + return sub_(other); + } + Tensor& operator*=(const Tensor & other) { + return mul_(other); + } + Tensor& operator*=(const Scalar & other) { + return mul_(other); + } + Tensor& operator/=(const Tensor & other) { + return div_(other); + } + Tensor& operator/=(const Scalar & other) { + return div_(other); + } + Tensor& operator&=(const Tensor & other) { + return bitwise_and_(other); + } + Tensor& operator|=(const Tensor & other) { + return bitwise_or_(other); + } + Tensor& operator^=(const Tensor & other) { + return bitwise_xor_(other); + } + Tensor operator[](const Scalar & index) const { + if (!index.isIntegral(false)) { + TORCH_CHECK_INDEX(false, "Can only index tensors with integral scalars"); + } + return this->operator[](index.toLong()); + } + Tensor operator[](const Tensor & index) const { + // These properties are checked in the Scalar constructor, but we already + // check them here to provide more useful diagnostics for the user. + if (!index.defined()) { + TORCH_CHECK_INDEX(false, "Can only index with tensors that are defined"); + } + if (index.dim() != 0) { + TORCH_CHECK_INDEX(false, + "Can only index with tensors that are scalars (zero-dim)"); + } + // The Scalar(Tensor) constructor is explicit, so we need to call it. + return this->operator[](index.item()); + } + Tensor operator[](int64_t index) const { + return select(0, index); + } + + Tensor index(ArrayRef indices) const; + Tensor index(std::initializer_list indices) const; + + Tensor & index_put_(ArrayRef indices, Tensor const & rhs); + Tensor & index_put_(ArrayRef indices, const Scalar& v); + Tensor & index_put_(std::initializer_list indices, Tensor const & rhs); + Tensor & index_put_(std::initializer_list indices, const Scalar& v); + + Tensor cpu() const { + return to(options().device(c10::DeviceType::CPU), /*non_blocking*/ false, /*copy*/ false); + } + + // TODO: The Python version also accepts arguments + Tensor cuda() const { + return to(options().device(c10::DeviceType::CUDA), /*non_blocking*/ false, /*copy*/ false); + } + + Tensor hip() const { + return to(options().device(c10::DeviceType::HIP), /*non_blocking*/ false, /*copy*/ false); + } + + Tensor ve() const { + return to(options().device(c10::DeviceType::VE), /*non_blocking*/ false, /*copy*/ false); + } + + Tensor vulkan() const { + return to(options().device(c10::DeviceType::Vulkan), /*non_blocking*/ false, /*copy*/ false); + } + + Tensor metal() const { + return to(options().device(c10::DeviceType::Metal), /*non_blocking*/ false, /*copy*/ false); + } + + Tensor meta() const { + return to(options().device(c10::DeviceType::Meta), /*non_blocking*/ false, /*copy*/ false); + } + + // ~~~~~ Autograd API ~~~~~ + + /// \fn bool is_leaf() const; + /// + /// All Tensors that have `requires_grad()` which is ``false`` will be leaf Tensors by convention. + /// + /// For Tensors that have `requires_grad()` which is ``true``, they will be leaf Tensors if they were + /// created by the user. This means that they are not the result of an operation and so + /// `grad_fn()` is `nullptr`. + /// + /// Only leaf Tensors will have their `grad()` populated during a call to `backward()`. + /// To get `grad()` populated for non-leaf Tensors, you can use `retain_grad()`. + /// + /// Example: + /// @code + /// auto a = torch::rand(10, torch::requires_grad()); + /// std::cout << a.is_leaf() << std::endl; // prints `true` + /// + /// auto b = torch::rand(10, torch::requires_grad()).to(torch::kCUDA); + /// std::cout << b.is_leaf() << std::endl; // prints `false` + /// // b was created by the operation that cast a cpu Tensor into a cuda Tensor + /// + /// auto c = torch::rand(10, torch::requires_grad()) + 2; + /// std::cout << c.is_leaf() << std::endl; // prints `false` + /// // c was created by the addition operation + /// + /// auto d = torch::rand(10).cuda(); + /// std::cout << d.is_leaf() << std::endl; // prints `true` + /// // d does not require gradients and so has no operation creating it (that is tracked by the autograd engine) + /// + /// auto e = torch::rand(10).cuda().requires_grad_(); + /// std::cout << e.is_leaf() << std::endl; // prints `true` + /// // e requires gradients and has no operations creating it + /// + /// auto f = torch::rand(10, torch::device(torch::kCUDA).requires_grad(true)); + /// std::cout << f.is_leaf() << std::endl; // prints `true` + /// // f requires grad, has no operation creating it + /// @endcode + + /// \fn void backward(const Tensor & gradient={}, std::optional retain_graph=std::nullopt, bool create_graph=false, std::optional inputs=std::nullopt) const; + /// + /// Computes the gradient of current tensor with respect to graph leaves. + /// + /// The graph is differentiated using the chain rule. If the tensor is + /// non-scalar (i.e. its data has more than one element) and requires + /// gradient, the function additionally requires specifying ``gradient``. + /// It should be a tensor of matching type and location, that contains + /// the gradient of the differentiated function w.r.t. this Tensor. + /// + /// This function accumulates gradients in the leaves - you might need to + /// zero them before calling it. + /// + /// \param gradient Gradient w.r.t. the + /// tensor. If it is a tensor, it will be automatically converted + /// to a Tensor that does not require grad unless ``create_graph`` is True. + /// None values can be specified for scalar Tensors or ones that + /// don't require grad. If a None value would be acceptable then + /// this argument is optional. + /// \param retain_graph If ``false``, the graph used to compute + /// the grads will be freed. Note that in nearly all cases setting + /// this option to True is not needed and often can be worked around + /// in a much more efficient way. Defaults to the value of + /// ``create_graph``. + /// \param create_graph If ``true``, graph of the derivative will + /// be constructed, allowing to compute higher order derivative + /// products. Defaults to ``false``. + /// \param inputs Inputs w.r.t. which the gradient will be accumulated into + /// ``at::Tensor::grad``. All other Tensors will be ignored. If not + /// provided, the gradient is accumulated into all the leaf Tensors + /// that were used to compute the current tensor. + /// When inputs are provided and a given input is not a leaf, + /// the current implementation will call its grad_fn (even though it is not strictly needed to get this gradients). + /// It is an implementation detail on which the user should not rely. + /// See https://github.com/pytorch/pytorch/pull/60521#issuecomment-867061780 for more details. + void backward(const Tensor & gradient={}, std::optional retain_graph=std::nullopt, bool create_graph=false, std::optional inputs=std::nullopt) const { + // NB: Adding this wrapper to _backward here because we'd like our + // 'backwards' api to accept the 'inputs' argument optionally. Since code gen + // currently does not support optional of TensorList our approach is to replace + // backward in native_functions.yaml with _backward and call it here instead. + if (inputs.has_value()) { + TORCH_CHECK(inputs.value().size() > 0, "'inputs' argument to backward cannot be empty") + this->_backward(inputs.value(), gradient, retain_graph, create_graph); + } else { + this->_backward({}, gradient, retain_graph, create_graph); + } + } + + /// \fn Tensor detach() const; + /// + /// Returns a new Tensor, detached from the current graph. + /// The result will never require gradient. + + /// \fn Tensor & detach_() const; + /// + /// Detaches the Tensor from the graph that created it, making it a leaf. + /// Views cannot be detached in-place. + + /// \fn void retain_grad() const; + /// + /// Enables this Tensor to have their :attr:`grad` populated during + /// :func:`backward`. This is a no-op for leaf tensors. + + /// \fn bool retains_grad() const; + /// + /// Is ``true`` if this Tensor is non-leaf and its :attr:`grad` is enabled to be + /// populated during :func:`backward`, ``false`` otherwise. + + const Tensor& set_requires_grad(bool requires_grad) const { + TensorBase::set_requires_grad(requires_grad); + return *this; + } + + /// Return a mutable reference to the gradient. This is conventionally + /// used as `t.grad() = x` to set a gradient to a completely new tensor. + /// Note that this function work with a non-const Tensor and is not + /// thread safe. + Tensor& mutable_grad() const { + return impl_->mutable_grad(); + } + + /// This function returns an undefined tensor by default and returns a defined tensor + /// the first time a call to `backward()` computes gradients for this Tensor. + /// The attribute will then contain the gradients computed and future calls + /// to `backward()` will accumulate (add) gradients into it. + const Tensor& grad() const { + const Tensor& maybe_grad = impl_->grad(); + if (!is_leaf() && !retains_grad() && !maybe_grad.defined()) { + TORCH_WARN( + "The .grad attribute of a Tensor that is not a leaf Tensor is being accessed. Its .grad " + "attribute won't be populated during autograd.backward(). If you indeed want the .grad " + "field to be populated for a non-leaf Tensor, use .retain_grad() on the non-leaf Tensor. " + "If you access the non-leaf Tensor by mistake, make sure you access the leaf Tensor " + "instead. See github.com/pytorch/pytorch/pull/30531 for more information."); + } + return maybe_grad; + } + + // The Forward AD API functions below are low level and are not to be used by end + // users who should use the API provided in torch/csrc/autograd.h + + /// This function returns the forward gradient for this Tensor at the given level. + const Tensor& _fw_grad(uint64_t level) const { + return impl_->_fw_grad(level, *this); + } + + /// This function can be used to set the value of the forward grad. + /// Note that the given new_grad might not be used directly if it has different + /// metadata (size/stride/storage offset) compared to this Tensor. In that case, + /// new_grad content will be copied into a new Tensor + void _set_fw_grad(const TensorBase& new_grad, uint64_t level, bool is_inplace_op) const { + impl_->_set_fw_grad(new_grad, *this, level, is_inplace_op); + } + + + // STOP. Thinking of adding a method here, which only makes use + // of other ATen methods? Define it in native_functions.yaml. + + //example + //Tensor * add(Tensor & b); + void __dispatch__backward(at::TensorList inputs, const ::std::optional & gradient={}, ::std::optional retain_graph=::std::nullopt, bool create_graph=false) const; + void __dispatch_set_data(const at::Tensor & new_data) const; + at::Tensor __dispatch_data() const; + bool __dispatch_is_leaf() const; + int64_t __dispatch_output_nr() const; + int64_t __dispatch__version() const; + at::Tensor & __dispatch_requires_grad_(bool requires_grad=true) const; + void __dispatch_retain_grad() const; + bool __dispatch_retains_grad() const; + at::Tensor _fw_primal(int64_t level) const; + at::Tensor & rename_(::std::optional names) const; + at::Tensor rename(::std::optional names) const; + at::Tensor align_to(at::DimnameList names) const; + at::Tensor align_to(at::DimnameList order, int64_t ellipsis_idx) const; + at::Tensor align_as(const at::Tensor & other) const; + at::Tensor refine_names(at::DimnameList names) const; + at::Tensor abs() const; + at::Tensor & abs_() const; + at::Tensor absolute() const; + at::Tensor & absolute_() const; + at::Tensor angle() const; + at::Tensor sgn() const; + at::Tensor & sgn_() const; + at::Tensor chalf(::std::optional memory_format=::std::nullopt) const; + at::Tensor _conj() const; + at::Tensor __dispatch_conj() const; + at::Tensor _conj_physical() const; + at::Tensor conj_physical() const; + at::Tensor & conj_physical_() const; + at::Tensor resolve_conj() const; + at::Tensor resolve_neg() const; + at::Tensor _neg_view() const; + at::Tensor acos() const; + at::Tensor & acos_() const; + at::Tensor arccos() const; + at::Tensor & arccos_() const; + at::Tensor add(const at::Tensor & other, const at::Scalar & alpha=1) const; + at::Tensor & add_(const at::Tensor & other, const at::Scalar & alpha=1) const; + at::Tensor add(const at::Scalar & other, const at::Scalar & alpha=1) const; + at::Tensor & add_(const at::Scalar & other, const at::Scalar & alpha=1) const; + at::Tensor addmv(const at::Tensor & mat, const at::Tensor & vec, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor & addmv_(const at::Tensor & mat, const at::Tensor & vec, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor addr(const at::Tensor & vec1, const at::Tensor & vec2, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor & addr_(const at::Tensor & vec1, const at::Tensor & vec2, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor _is_all_true() const; + at::Tensor _is_any_true() const; + at::Tensor all(int64_t dim, bool keepdim=false) const; + at::Tensor all(at::OptionalIntArrayRef dim, bool keepdim=false) const; + at::Tensor all(at::Dimname dim, bool keepdim=false) const; + bool allclose(const at::Tensor & other, double rtol=1e-05, double atol=1e-08, bool equal_nan=false) const; + at::Tensor any(int64_t dim, bool keepdim=false) const; + at::Tensor any(at::OptionalIntArrayRef dim, bool keepdim=false) const; + at::Tensor any(at::Dimname dim, bool keepdim=false) const; + at::Tensor argmax(::std::optional dim=::std::nullopt, bool keepdim=false) const; + at::Tensor argmin(::std::optional dim=::std::nullopt, bool keepdim=false) const; + at::Tensor acosh() const; + at::Tensor & acosh_() const; + at::Tensor arccosh() const; + at::Tensor & arccosh_() const; + at::Tensor asinh() const; + at::Tensor & asinh_() const; + at::Tensor arcsinh() const; + at::Tensor & arcsinh_() const; + at::Tensor atanh() const; + at::Tensor & atanh_() const; + at::Tensor arctanh() const; + at::Tensor & arctanh_() const; + at::Tensor as_strided(at::IntArrayRef size, at::IntArrayRef stride, ::std::optional storage_offset=::std::nullopt) const; + at::Tensor as_strided_symint(c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset=::std::nullopt) const; + const at::Tensor & as_strided_(at::IntArrayRef size, at::IntArrayRef stride, ::std::optional storage_offset=::std::nullopt) const; + const at::Tensor & as_strided__symint(c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset=::std::nullopt) const; + at::Tensor asin() const; + at::Tensor & asin_() const; + at::Tensor arcsin() const; + at::Tensor & arcsin_() const; + at::Tensor atan() const; + at::Tensor & atan_() const; + at::Tensor arctan() const; + at::Tensor & arctan_() const; + at::Tensor baddbmm(const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor & baddbmm_(const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor bernoulli(::std::optional generator=::std::nullopt) const; + at::Tensor & bernoulli_(const at::Tensor & p, ::std::optional generator=::std::nullopt) const; + at::Tensor & bernoulli_(double p=0.5, ::std::optional generator=::std::nullopt) const; + at::Tensor bernoulli(double p, ::std::optional generator=::std::nullopt) const; + at::Tensor bincount(const ::std::optional & weights={}, int64_t minlength=0) const; + at::Tensor bincount_symint(const ::std::optional & weights={}, c10::SymInt minlength=0) const; + at::Tensor bitwise_not() const; + at::Tensor & bitwise_not_() const; + at::Tensor copysign(const at::Tensor & other) const; + at::Tensor & copysign_(const at::Tensor & other) const; + at::Tensor copysign(const at::Scalar & other) const; + at::Tensor & copysign_(const at::Scalar & other) const; + at::Tensor _lazy_clone() const; + at::Tensor logical_not() const; + at::Tensor & logical_not_() const; + at::Tensor logical_xor(const at::Tensor & other) const; + at::Tensor & logical_xor_(const at::Tensor & other) const; + at::Tensor logical_and(const at::Tensor & other) const; + at::Tensor & logical_and_(const at::Tensor & other) const; + at::Tensor logical_or(const at::Tensor & other) const; + at::Tensor & logical_or_(const at::Tensor & other) const; + at::Tensor bmm(const at::Tensor & mat2) const; + at::Tensor broadcast_to(at::IntArrayRef size) const; + at::Tensor broadcast_to_symint(c10::SymIntArrayRef size) const; + at::Tensor ceil() const; + at::Tensor & ceil_() const; + ::std::vector unsafe_chunk(int64_t chunks, int64_t dim=0) const; + ::std::vector chunk(int64_t chunks, int64_t dim=0) const; + ::std::vector tensor_split(int64_t sections, int64_t dim=0) const; + ::std::vector tensor_split_symint(c10::SymInt sections, int64_t dim=0) const; + ::std::vector tensor_split(at::IntArrayRef indices, int64_t dim=0) const; + ::std::vector tensor_split_symint(c10::SymIntArrayRef indices, int64_t dim=0) const; + ::std::vector tensor_split(const at::Tensor & tensor_indices_or_sections, int64_t dim=0) const; + at::Tensor clamp(const ::std::optional & min, const ::std::optional & max=::std::nullopt) const; + at::Tensor clamp(const ::std::optional & min={}, const ::std::optional & max={}) const; + at::Tensor & clamp_(const ::std::optional & min, const ::std::optional & max=::std::nullopt) const; + at::Tensor & clamp_(const ::std::optional & min={}, const ::std::optional & max={}) const; + at::Tensor clamp_max(const at::Scalar & max) const; + at::Tensor clamp_max(const at::Tensor & max) const; + at::Tensor & clamp_max_(const at::Scalar & max) const; + at::Tensor & clamp_max_(const at::Tensor & max) const; + at::Tensor clamp_min(const at::Scalar & min) const; + at::Tensor clamp_min(const at::Tensor & min) const; + at::Tensor & clamp_min_(const at::Scalar & min) const; + at::Tensor & clamp_min_(const at::Tensor & min) const; + at::Tensor clip(const ::std::optional & min, const ::std::optional & max=::std::nullopt) const; + at::Tensor clip(const ::std::optional & min={}, const ::std::optional & max={}) const; + at::Tensor & clip_(const ::std::optional & min, const ::std::optional & max=::std::nullopt) const; + at::Tensor & clip_(const ::std::optional & min={}, const ::std::optional & max={}) const; + at::Tensor __dispatch_contiguous(at::MemoryFormat memory_format=c10::MemoryFormat::Contiguous) const; + at::Tensor & copy_(const at::Tensor & src, bool non_blocking=false) const; + at::Tensor cos() const; + at::Tensor & cos_() const; + at::Tensor cosh() const; + at::Tensor & cosh_() const; + at::Tensor count_nonzero(at::IntArrayRef dim) const; + at::Tensor count_nonzero(::std::optional dim=::std::nullopt) const; + at::Tensor cov(int64_t correction=1, const ::std::optional & fweights={}, const ::std::optional & aweights={}) const; + at::Tensor corrcoef() const; + ::std::tuple cummax(int64_t dim) const; + ::std::tuple cummax(at::Dimname dim) const; + ::std::tuple cummin(int64_t dim) const; + ::std::tuple cummin(at::Dimname dim) const; + at::Tensor cumprod(int64_t dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor & cumprod_(int64_t dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor cumprod(at::Dimname dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor & cumprod_(at::Dimname dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor cumsum(int64_t dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor & cumsum_(int64_t dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor cumsum(at::Dimname dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor & cumsum_(at::Dimname dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor diag_embed(int64_t offset=0, int64_t dim1=-2, int64_t dim2=-1) const; + at::Tensor diagflat(int64_t offset=0) const; + at::Tensor diagonal(int64_t offset=0, int64_t dim1=0, int64_t dim2=1) const; + at::Tensor diagonal(at::Dimname outdim, at::Dimname dim1, at::Dimname dim2, int64_t offset=0) const; + at::Tensor & fill_diagonal_(const at::Scalar & fill_value, bool wrap=false) const; + at::Tensor diff(int64_t n=1, int64_t dim=-1, const ::std::optional & prepend={}, const ::std::optional & append={}) const; + at::Tensor div(const at::Tensor & other) const; + at::Tensor & div_(const at::Tensor & other) const; + at::Tensor div(const at::Tensor & other, ::std::optional rounding_mode) const; + at::Tensor & div_(const at::Tensor & other, ::std::optional rounding_mode) const; + at::Tensor div(const at::Scalar & other) const; + at::Tensor & div_(const at::Scalar & other) const; + at::Tensor div(const at::Scalar & other, ::std::optional rounding_mode) const; + at::Tensor & div_(const at::Scalar & other, ::std::optional rounding_mode) const; + at::Tensor divide(const at::Tensor & other) const; + at::Tensor & divide_(const at::Tensor & other) const; + at::Tensor divide(const at::Scalar & other) const; + at::Tensor & divide_(const at::Scalar & other) const; + at::Tensor divide(const at::Tensor & other, ::std::optional rounding_mode) const; + at::Tensor & divide_(const at::Tensor & other, ::std::optional rounding_mode) const; + at::Tensor divide(const at::Scalar & other, ::std::optional rounding_mode) const; + at::Tensor & divide_(const at::Scalar & other, ::std::optional rounding_mode) const; + at::Tensor true_divide(const at::Tensor & other) const; + at::Tensor & true_divide_(const at::Tensor & other) const; + at::Tensor true_divide(const at::Scalar & other) const; + at::Tensor & true_divide_(const at::Scalar & other) const; + at::Tensor dot(const at::Tensor & tensor) const; + at::Tensor vdot(const at::Tensor & other) const; + at::Tensor new_empty(at::IntArrayRef size, at::TensorOptions options={}) const; + at::Tensor new_empty(at::IntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const; + at::Tensor new_empty_symint(c10::SymIntArrayRef size, at::TensorOptions options={}) const; + at::Tensor new_empty_symint(c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const; + at::Tensor new_empty_strided(at::IntArrayRef size, at::IntArrayRef stride, at::TensorOptions options={}) const; + at::Tensor new_empty_strided(at::IntArrayRef size, at::IntArrayRef stride, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const; + at::Tensor new_empty_strided_symint(c10::SymIntArrayRef size, c10::SymIntArrayRef stride, at::TensorOptions options={}) const; + at::Tensor new_empty_strided_symint(c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const; + at::Tensor new_full(at::IntArrayRef size, const at::Scalar & fill_value, at::TensorOptions options={}) const; + at::Tensor new_full(at::IntArrayRef size, const at::Scalar & fill_value, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const; + at::Tensor new_full_symint(c10::SymIntArrayRef size, const at::Scalar & fill_value, at::TensorOptions options={}) const; + at::Tensor new_full_symint(c10::SymIntArrayRef size, const at::Scalar & fill_value, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const; + at::Tensor new_zeros(at::IntArrayRef size, at::TensorOptions options={}) const; + at::Tensor new_zeros(at::IntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const; + at::Tensor new_zeros_symint(c10::SymIntArrayRef size, at::TensorOptions options={}) const; + at::Tensor new_zeros_symint(c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const; + at::Tensor new_ones(at::IntArrayRef size, at::TensorOptions options={}) const; + at::Tensor new_ones(at::IntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const; + at::Tensor new_ones_symint(c10::SymIntArrayRef size, at::TensorOptions options={}) const; + at::Tensor new_ones_symint(c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const; + const at::Tensor & resize_(at::IntArrayRef size, ::std::optional memory_format=::std::nullopt) const; + const at::Tensor & resize__symint(c10::SymIntArrayRef size, ::std::optional memory_format=::std::nullopt) const; + at::Tensor erf() const; + at::Tensor & erf_() const; + at::Tensor erfc() const; + at::Tensor & erfc_() const; + at::Tensor exp() const; + at::Tensor & exp_() const; + at::Tensor exp2() const; + at::Tensor & exp2_() const; + at::Tensor expm1() const; + at::Tensor & expm1_() const; + at::Tensor expand(at::IntArrayRef size, bool implicit=false) const; + at::Tensor expand_symint(c10::SymIntArrayRef size, bool implicit=false) const; + at::Tensor expand_as(const at::Tensor & other) const; + at::Tensor flatten(int64_t start_dim=0, int64_t end_dim=-1) const; + at::Tensor flatten(int64_t start_dim, int64_t end_dim, at::Dimname out_dim) const; + at::Tensor flatten(at::Dimname start_dim, at::Dimname end_dim, at::Dimname out_dim) const; + at::Tensor flatten(at::DimnameList dims, at::Dimname out_dim) const; + at::Tensor unflatten(int64_t dim, at::IntArrayRef sizes) const; + at::Tensor unflatten_symint(int64_t dim, c10::SymIntArrayRef sizes) const; + at::Tensor unflatten(at::Dimname dim, at::IntArrayRef sizes, at::DimnameList names) const; + at::Tensor unflatten_symint(at::Dimname dim, c10::SymIntArrayRef sizes, at::DimnameList names) const; + at::Tensor & fill_(const at::Scalar & value) const; + at::Tensor & fill_(const at::Tensor & value) const; + at::Tensor floor() const; + at::Tensor & floor_() const; + at::Tensor floor_divide(const at::Tensor & other) const; + at::Tensor & floor_divide_(const at::Tensor & other) const; + at::Tensor floor_divide(const at::Scalar & other) const; + at::Tensor & floor_divide_(const at::Scalar & other) const; + at::Tensor frac() const; + at::Tensor & frac_() const; + at::Tensor gcd(const at::Tensor & other) const; + at::Tensor & gcd_(const at::Tensor & other) const; + at::Tensor lcm(const at::Tensor & other) const; + at::Tensor & lcm_(const at::Tensor & other) const; + at::Tensor index(const c10::List<::std::optional> & indices) const; + at::Tensor & index_copy_(int64_t dim, const at::Tensor & index, const at::Tensor & source) const; + at::Tensor index_copy(int64_t dim, const at::Tensor & index, const at::Tensor & source) const; + at::Tensor & index_copy_(at::Dimname dim, const at::Tensor & index, const at::Tensor & source) const; + at::Tensor index_copy(at::Dimname dim, const at::Tensor & index, const at::Tensor & source) const; + at::Tensor & index_put_(const c10::List<::std::optional> & indices, const at::Tensor & values, bool accumulate=false) const; + at::Tensor index_put(const c10::List<::std::optional> & indices, const at::Tensor & values, bool accumulate=false) const; + at::Tensor isclose(const at::Tensor & other, double rtol=1e-05, double atol=1e-08, bool equal_nan=false) const; + at::Tensor isnan() const; + bool is_distributed() const; + bool __dispatch_is_floating_point() const; + bool __dispatch_is_complex() const; + bool __dispatch_is_conj() const; + bool __dispatch__is_zerotensor() const; + bool __dispatch_is_neg() const; + at::Tensor isreal() const; + bool is_nonzero() const; + bool is_same_size(const at::Tensor & other) const; + bool __dispatch_is_signed() const; + bool __dispatch_is_inference() const; + at::Tensor kron(const at::Tensor & other) const; + ::std::tuple kthvalue(int64_t k, int64_t dim=-1, bool keepdim=false) const; + ::std::tuple kthvalue_symint(c10::SymInt k, int64_t dim=-1, bool keepdim=false) const; + ::std::tuple kthvalue(int64_t k, at::Dimname dim, bool keepdim=false) const; + ::std::tuple kthvalue_symint(c10::SymInt k, at::Dimname dim, bool keepdim=false) const; + at::Tensor nan_to_num(::std::optional nan=::std::nullopt, ::std::optional posinf=::std::nullopt, ::std::optional neginf=::std::nullopt) const; + at::Tensor & nan_to_num_(::std::optional nan=::std::nullopt, ::std::optional posinf=::std::nullopt, ::std::optional neginf=::std::nullopt) const; + at::Tensor ldexp(const at::Tensor & other) const; + at::Tensor & ldexp_(const at::Tensor & other) const; + at::Tensor log() const; + at::Tensor & log_() const; + at::Tensor log10() const; + at::Tensor & log10_() const; + at::Tensor log1p() const; + at::Tensor & log1p_() const; + at::Tensor log2() const; + at::Tensor & log2_() const; + at::Tensor logaddexp(const at::Tensor & other) const; + at::Tensor logaddexp2(const at::Tensor & other) const; + at::Tensor xlogy(const at::Tensor & other) const; + at::Tensor xlogy(const at::Scalar & other) const; + at::Tensor & xlogy_(const at::Tensor & other) const; + at::Tensor & xlogy_(const at::Scalar & other) const; + at::Tensor log_softmax(int64_t dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor log_softmax(at::Dimname dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor logcumsumexp(int64_t dim) const; + at::Tensor logcumsumexp(at::Dimname dim) const; + at::Tensor logsumexp(at::IntArrayRef dim, bool keepdim=false) const; + at::Tensor logsumexp(at::DimnameList dim, bool keepdim=false) const; + at::Tensor matmul(const at::Tensor & other) const; + at::Tensor matrix_power(int64_t n) const; + at::Tensor matrix_exp() const; + ::std::tuple aminmax(::std::optional dim=::std::nullopt, bool keepdim=false) const; + ::std::tuple max(int64_t dim, bool keepdim=false) const; + ::std::tuple max(at::Dimname dim, bool keepdim=false) const; + at::Tensor amax(at::IntArrayRef dim={}, bool keepdim=false) const; + at::Tensor mean(::std::optional dtype=::std::nullopt) const; + at::Tensor mean(at::OptionalIntArrayRef dim, bool keepdim=false, ::std::optional dtype=::std::nullopt) const; + at::Tensor mean(at::DimnameList dim, bool keepdim=false, ::std::optional dtype=::std::nullopt) const; + at::Tensor nanmean(at::OptionalIntArrayRef dim=::std::nullopt, bool keepdim=false, ::std::optional dtype=::std::nullopt) const; + at::Tensor median() const; + ::std::tuple median(int64_t dim, bool keepdim=false) const; + ::std::tuple median(at::Dimname dim, bool keepdim=false) const; + at::Tensor nanmedian() const; + ::std::tuple nanmedian(int64_t dim, bool keepdim=false) const; + ::std::tuple nanmedian(at::Dimname dim, bool keepdim=false) const; + ::std::tuple min(int64_t dim, bool keepdim=false) const; + ::std::tuple min(at::Dimname dim, bool keepdim=false) const; + at::Tensor amin(at::IntArrayRef dim={}, bool keepdim=false) const; + at::Tensor mm(const at::Tensor & mat2) const; + ::std::tuple mode(int64_t dim=-1, bool keepdim=false) const; + ::std::tuple mode(at::Dimname dim, bool keepdim=false) const; + at::Tensor mul(const at::Tensor & other) const; + at::Tensor & mul_(const at::Tensor & other) const; + at::Tensor mul(const at::Scalar & other) const; + at::Tensor & mul_(const at::Scalar & other) const; + at::Tensor multiply(const at::Tensor & other) const; + at::Tensor & multiply_(const at::Tensor & other) const; + at::Tensor multiply(const at::Scalar & other) const; + at::Tensor & multiply_(const at::Scalar & other) const; + at::Tensor mv(const at::Tensor & vec) const; + at::Tensor mvlgamma(int64_t p) const; + at::Tensor & mvlgamma_(int64_t p) const; + at::Tensor narrow_copy(int64_t dim, int64_t start, int64_t length) const; + at::Tensor narrow_copy_symint(int64_t dim, c10::SymInt start, c10::SymInt length) const; + at::Tensor narrow(int64_t dim, int64_t start, int64_t length) const; + at::Tensor narrow_symint(int64_t dim, c10::SymInt start, c10::SymInt length) const; + at::Tensor narrow(int64_t dim, const at::Tensor & start, int64_t length) const; + at::Tensor narrow_symint(int64_t dim, const at::Tensor & start, c10::SymInt length) const; + at::Tensor permute(at::IntArrayRef dims) const; + at::Tensor movedim(at::IntArrayRef source, at::IntArrayRef destination) const; + at::Tensor movedim(int64_t source, int64_t destination) const; + at::Tensor moveaxis(at::IntArrayRef source, at::IntArrayRef destination) const; + at::Tensor moveaxis(int64_t source, int64_t destination) const; + at::Tensor numpy_T() const; + at::Tensor matrix_H() const; + at::Tensor mT() const; + at::Tensor mH() const; + at::Tensor adjoint() const; + bool is_pinned(::std::optional device=::std::nullopt) const; + at::Tensor pin_memory(::std::optional device=::std::nullopt) const; + at::Tensor pinverse(double rcond=1e-15) const; + at::Tensor rad2deg() const; + at::Tensor & rad2deg_() const; + at::Tensor deg2rad() const; + at::Tensor & deg2rad_() const; + at::Tensor ravel() const; + at::Tensor reciprocal() const; + at::Tensor & reciprocal_() const; + at::Tensor neg() const; + at::Tensor & neg_() const; + at::Tensor negative() const; + at::Tensor & negative_() const; + at::Tensor repeat(at::IntArrayRef repeats) const; + at::Tensor repeat_symint(c10::SymIntArrayRef repeats) const; + at::Tensor repeat_interleave(const at::Tensor & repeats, ::std::optional dim=::std::nullopt, ::std::optional output_size=::std::nullopt) const; + at::Tensor repeat_interleave_symint(const at::Tensor & repeats, ::std::optional dim=::std::nullopt, ::std::optional output_size=::std::nullopt) const; + at::Tensor repeat_interleave(int64_t repeats, ::std::optional dim=::std::nullopt, ::std::optional output_size=::std::nullopt) const; + at::Tensor repeat_interleave_symint(c10::SymInt repeats, ::std::optional dim=::std::nullopt, ::std::optional output_size=::std::nullopt) const; + at::Tensor reshape(at::IntArrayRef shape) const; + at::Tensor reshape_symint(c10::SymIntArrayRef shape) const; + at::Tensor _reshape_alias(at::IntArrayRef size, at::IntArrayRef stride) const; + at::Tensor _reshape_alias_symint(c10::SymIntArrayRef size, c10::SymIntArrayRef stride) const; + at::Tensor reshape_as(const at::Tensor & other) const; + at::Tensor round() const; + at::Tensor & round_() const; + at::Tensor round(int64_t decimals) const; + at::Tensor & round_(int64_t decimals) const; + at::Tensor relu() const; + at::Tensor & relu_() const; + at::Tensor prelu(const at::Tensor & weight) const; + at::Tensor hardshrink(const at::Scalar & lambd=0.5) const; + at::Tensor hardshrink_backward(const at::Tensor & grad_out, const at::Scalar & lambd) const; + at::Tensor rsqrt() const; + at::Tensor & rsqrt_() const; + at::Tensor select(at::Dimname dim, int64_t index) const; + at::Tensor select(int64_t dim, int64_t index) const; + at::Tensor select_symint(int64_t dim, c10::SymInt index) const; + at::Tensor sigmoid() const; + at::Tensor & sigmoid_() const; + at::Tensor logit(::std::optional eps=::std::nullopt) const; + at::Tensor & logit_(::std::optional eps=::std::nullopt) const; + at::Tensor sin() const; + at::Tensor & sin_() const; + at::Tensor sinc() const; + at::Tensor & sinc_() const; + at::Tensor sinh() const; + at::Tensor & sinh_() const; + at::Tensor detach() const; + at::Tensor & detach_() const; + int64_t size(at::Dimname dim) const; + at::Tensor slice(int64_t dim=0, ::std::optional start=::std::nullopt, ::std::optional end=::std::nullopt, int64_t step=1) const; + at::Tensor slice_symint(int64_t dim=0, ::std::optional start=::std::nullopt, ::std::optional end=::std::nullopt, c10::SymInt step=1) const; + at::Tensor slice_inverse(const at::Tensor & src, int64_t dim=0, ::std::optional start=::std::nullopt, ::std::optional end=::std::nullopt, int64_t step=1) const; + at::Tensor slice_inverse_symint(const at::Tensor & src, int64_t dim=0, ::std::optional start=::std::nullopt, ::std::optional end=::std::nullopt, c10::SymInt step=1) const; + at::Tensor slice_scatter(const at::Tensor & src, int64_t dim=0, ::std::optional start=::std::nullopt, ::std::optional end=::std::nullopt, int64_t step=1) const; + at::Tensor slice_scatter_symint(const at::Tensor & src, int64_t dim=0, ::std::optional start=::std::nullopt, ::std::optional end=::std::nullopt, c10::SymInt step=1) const; + at::Tensor select_scatter(const at::Tensor & src, int64_t dim, int64_t index) const; + at::Tensor select_scatter_symint(const at::Tensor & src, int64_t dim, c10::SymInt index) const; + at::Tensor diagonal_scatter(const at::Tensor & src, int64_t offset=0, int64_t dim1=0, int64_t dim2=1) const; + at::Tensor as_strided_scatter(const at::Tensor & src, at::IntArrayRef size, at::IntArrayRef stride, ::std::optional storage_offset=::std::nullopt) const; + at::Tensor as_strided_scatter_symint(const at::Tensor & src, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset=::std::nullopt) const; + at::Tensor smm(const at::Tensor & mat2) const; + at::Tensor softmax(int64_t dim, ::std::optional dtype=::std::nullopt) const; + at::Tensor softmax(at::Dimname dim, ::std::optional dtype=::std::nullopt) const; + ::std::vector unsafe_split(int64_t split_size, int64_t dim=0) const; + ::std::vector unsafe_split_symint(c10::SymInt split_size, int64_t dim=0) const; + ::std::vector split(int64_t split_size, int64_t dim=0) const; + ::std::vector split_symint(c10::SymInt split_size, int64_t dim=0) const; + ::std::vector split(at::IntArrayRef split_size, int64_t dim=0) const; + ::std::vector split_symint(c10::SymIntArrayRef split_size, int64_t dim=0) const; + ::std::vector unsafe_split_with_sizes(at::IntArrayRef split_sizes, int64_t dim=0) const; + ::std::vector unsafe_split_with_sizes_symint(c10::SymIntArrayRef split_sizes, int64_t dim=0) const; + ::std::vector split_with_sizes(at::IntArrayRef split_sizes, int64_t dim=0) const; + ::std::vector split_with_sizes_symint(c10::SymIntArrayRef split_sizes, int64_t dim=0) const; + ::std::vector hsplit(int64_t sections) const; + ::std::vector hsplit(at::IntArrayRef indices) const; + ::std::vector vsplit(int64_t sections) const; + ::std::vector vsplit(at::IntArrayRef indices) const; + ::std::vector dsplit(int64_t sections) const; + ::std::vector dsplit(at::IntArrayRef indices) const; + at::Tensor squeeze() const; + at::Tensor squeeze(int64_t dim) const; + at::Tensor squeeze(at::Dimname dim) const; + at::Tensor squeeze(at::IntArrayRef dim) const; + at::Tensor & squeeze_() const; + at::Tensor & squeeze_(int64_t dim) const; + at::Tensor & squeeze_(at::IntArrayRef dim) const; + at::Tensor & squeeze_(at::Dimname dim) const; + at::Tensor sspaddmm(const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor stft(int64_t n_fft, ::std::optional hop_length, ::std::optional win_length, const ::std::optional & window, bool normalized, ::std::optional onesided=::std::nullopt, ::std::optional return_complex=::std::nullopt, ::std::optional align_to_window=::std::nullopt) const; + at::Tensor stft(int64_t n_fft, ::std::optional hop_length=::std::nullopt, ::std::optional win_length=::std::nullopt, const ::std::optional & window={}, bool center=true, c10::string_view pad_mode="reflect", bool normalized=false, ::std::optional onesided=::std::nullopt, ::std::optional return_complex=::std::nullopt, ::std::optional align_to_window=::std::nullopt) const; + at::Tensor istft(int64_t n_fft, ::std::optional hop_length=::std::nullopt, ::std::optional win_length=::std::nullopt, const ::std::optional & window={}, bool center=true, bool normalized=false, ::std::optional onesided=::std::nullopt, ::std::optional length=::std::nullopt, bool return_complex=false) const; + int64_t stride(at::Dimname dim) const; + at::Tensor sum(::std::optional dtype=::std::nullopt) const; + at::Tensor sum(at::OptionalIntArrayRef dim, bool keepdim=false, ::std::optional dtype=::std::nullopt) const; + at::Tensor sum(at::DimnameList dim, bool keepdim=false, ::std::optional dtype=::std::nullopt) const; + at::Tensor nansum(at::OptionalIntArrayRef dim=::std::nullopt, bool keepdim=false, ::std::optional dtype=::std::nullopt) const; + at::Tensor hash_tensor(at::IntArrayRef dim={}, bool keepdim=false, int64_t mode=0) const; + at::Tensor sum_to_size(at::IntArrayRef size) const; + at::Tensor sum_to_size_symint(c10::SymIntArrayRef size) const; + at::Tensor sqrt() const; + at::Tensor & sqrt_() const; + at::Tensor square() const; + at::Tensor & square_() const; + at::Tensor std(bool unbiased) const; + at::Tensor std(at::OptionalIntArrayRef dim, bool unbiased, bool keepdim=false) const; + at::Tensor std(at::OptionalIntArrayRef dim=::std::nullopt, const ::std::optional & correction=::std::nullopt, bool keepdim=false) const; + at::Tensor std(at::DimnameList dim, bool unbiased, bool keepdim=false) const; + at::Tensor std(at::DimnameList dim, const ::std::optional & correction=::std::nullopt, bool keepdim=false) const; + at::Tensor prod(::std::optional dtype=::std::nullopt) const; + at::Tensor prod(int64_t dim, bool keepdim=false, ::std::optional dtype=::std::nullopt) const; + at::Tensor prod(at::Dimname dim, bool keepdim=false, ::std::optional dtype=::std::nullopt) const; + at::Tensor t() const; + at::Tensor & t_() const; + at::Tensor tan() const; + at::Tensor & tan_() const; + at::Tensor tanh() const; + at::Tensor & tanh_() const; + at::Tensor tile(at::IntArrayRef dims) const; + at::Tensor tile_symint(c10::SymIntArrayRef dims) const; + at::Tensor transpose(int64_t dim0, int64_t dim1) const; + at::Tensor transpose(at::Dimname dim0, at::Dimname dim1) const; + at::Tensor & transpose_(int64_t dim0, int64_t dim1) const; + at::Tensor flip(at::IntArrayRef dims) const; + at::Tensor fliplr() const; + at::Tensor flipud() const; + at::Tensor roll(at::IntArrayRef shifts, at::IntArrayRef dims={}) const; + at::Tensor roll_symint(c10::SymIntArrayRef shifts, at::IntArrayRef dims={}) const; + at::Tensor rot90(int64_t k=1, at::IntArrayRef dims={0,1}) const; + at::Tensor _nested_tensor_size() const; + at::Tensor _nested_tensor_strides() const; + at::Tensor _nested_tensor_storage_offsets() const; + at::Tensor trunc() const; + at::Tensor & trunc_() const; + at::Tensor fix() const; + at::Tensor & fix_() const; + at::Tensor type_as(const at::Tensor & other) const; + at::Tensor unsqueeze(int64_t dim) const; + at::Tensor & unsqueeze_(int64_t dim) const; + at::Tensor var(bool unbiased) const; + at::Tensor var(at::OptionalIntArrayRef dim, bool unbiased, bool keepdim=false) const; + at::Tensor var(at::OptionalIntArrayRef dim=::std::nullopt, const ::std::optional & correction=::std::nullopt, bool keepdim=false) const; + at::Tensor var(at::DimnameList dim, bool unbiased, bool keepdim=false) const; + at::Tensor var(at::DimnameList dim, const ::std::optional & correction=::std::nullopt, bool keepdim=false) const; + at::Tensor view_as(const at::Tensor & other) const; + at::Tensor where(const at::Tensor & condition, const at::Tensor & other) const; + at::Tensor where(const at::Tensor & condition, const at::Scalar & other) const; + at::Tensor norm(const ::std::optional & p, at::ScalarType dtype) const; + at::Tensor norm(const at::Scalar & p=2) const; + at::Tensor norm(const ::std::optional & p, at::IntArrayRef dim, bool keepdim, at::ScalarType dtype) const; + at::Tensor norm(const ::std::optional & p, at::IntArrayRef dim, bool keepdim=false) const; + at::Tensor norm(const ::std::optional & p, at::DimnameList dim, bool keepdim, at::ScalarType dtype) const; + at::Tensor norm(const ::std::optional & p, at::DimnameList dim, bool keepdim=false) const; + ::std::tuple frexp() const; + at::Tensor clone(::std::optional memory_format=::std::nullopt) const; + at::Tensor positive() const; + const at::Tensor & resize_as_(const at::Tensor & the_template, ::std::optional memory_format=::std::nullopt) const; + const at::Tensor & resize_as_sparse_(const at::Tensor & the_template) const; + at::Tensor & zero_() const; + at::Tensor sub(const at::Tensor & other, const at::Scalar & alpha=1) const; + at::Tensor & sub_(const at::Tensor & other, const at::Scalar & alpha=1) const; + at::Tensor sub(const at::Scalar & other, const at::Scalar & alpha=1) const; + at::Tensor & sub_(const at::Scalar & other, const at::Scalar & alpha=1) const; + at::Tensor subtract(const at::Tensor & other, const at::Scalar & alpha=1) const; + at::Tensor & subtract_(const at::Tensor & other, const at::Scalar & alpha=1) const; + at::Tensor subtract(const at::Scalar & other, const at::Scalar & alpha=1) const; + at::Tensor & subtract_(const at::Scalar & other, const at::Scalar & alpha=1) const; + at::Tensor heaviside(const at::Tensor & values) const; + at::Tensor & heaviside_(const at::Tensor & values) const; + at::Tensor addmm(const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor & addmm_(const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor _addmm_activation(const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta=1, const at::Scalar & alpha=1, bool use_gelu=false) const; + const at::Tensor & sparse_resize_(at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim) const; + const at::Tensor & sparse_resize_and_clear_(at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim) const; + at::Tensor sparse_mask(const at::Tensor & mask) const; + at::Tensor _sparse_mask_projection(const at::Tensor & mask, bool accumulate_matches=false) const; + at::Tensor to_dense(::std::optional dtype=::std::nullopt, ::std::optional masked_grad=::std::nullopt) const; + at::Tensor _to_dense(::std::optional dtype=::std::nullopt, ::std::optional masked_grad=::std::nullopt) const; + int64_t sparse_dim() const; + int64_t _dimI() const; + int64_t dense_dim() const; + int64_t _dimV() const; + int64_t _nnz() const; + at::Tensor coalesce() const; + bool is_coalesced() const; + at::Tensor _indices() const; + at::Tensor _values() const; + at::Tensor & _coalesced_(bool coalesced) const; + at::Tensor indices() const; + at::Tensor values() const; + at::Tensor crow_indices() const; + at::Tensor col_indices() const; + at::Tensor ccol_indices() const; + at::Tensor row_indices() const; + ::std::vector unbind(int64_t dim=0) const; + ::std::vector unbind(at::Dimname dim) const; + at::Tensor to_sparse(int64_t sparse_dim) const; + at::Tensor _to_sparse(int64_t sparse_dim) const; + at::Tensor to_sparse(::std::optional layout=::std::nullopt, at::OptionalIntArrayRef blocksize=::std::nullopt, ::std::optional dense_dim=::std::nullopt) const; + at::Tensor _to_sparse(::std::optional layout=::std::nullopt, at::OptionalIntArrayRef blocksize=::std::nullopt, ::std::optional dense_dim=::std::nullopt) const; + at::Tensor to_sparse_csr(::std::optional dense_dim=::std::nullopt) const; + at::Tensor _to_sparse_csr(::std::optional dense_dim=::std::nullopt) const; + at::Tensor to_sparse_csc(::std::optional dense_dim=::std::nullopt) const; + at::Tensor _to_sparse_csc(::std::optional dense_dim=::std::nullopt) const; + at::Tensor to_sparse_bsr(at::IntArrayRef blocksize, ::std::optional dense_dim=::std::nullopt) const; + at::Tensor _to_sparse_bsr(at::IntArrayRef blocksize, ::std::optional dense_dim=::std::nullopt) const; + at::Tensor to_sparse_bsc(at::IntArrayRef blocksize, ::std::optional dense_dim=::std::nullopt) const; + at::Tensor _to_sparse_bsc(at::IntArrayRef blocksize, ::std::optional dense_dim=::std::nullopt) const; + at::Tensor to_mkldnn(::std::optional dtype=::std::nullopt) const; + at::Tensor dequantize() const; + double q_scale() const; + int64_t q_zero_point() const; + at::Tensor q_per_channel_scales() const; + at::Tensor q_per_channel_zero_points() const; + int64_t q_per_channel_axis() const; + at::Tensor int_repr() const; + at::QScheme qscheme() const; + at::Tensor _autocast_to_reduced_precision(bool cuda_enabled, bool cpu_enabled, at::ScalarType cuda_dtype, at::ScalarType cpu_dtype) const; + at::Tensor _autocast_to_full_precision(bool cuda_enabled, bool cpu_enabled) const; + at::Tensor to(at::TensorOptions options={}, bool non_blocking=false, bool copy=false, ::std::optional memory_format=::std::nullopt) const; + at::Tensor to(::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory, bool non_blocking, bool copy, ::std::optional memory_format) const; + at::Tensor to(at::Device device, at::ScalarType dtype, bool non_blocking=false, bool copy=false, ::std::optional memory_format=::std::nullopt) const; + at::Tensor to(at::ScalarType dtype, bool non_blocking=false, bool copy=false, ::std::optional memory_format=::std::nullopt) const; + at::Tensor to(const at::Tensor & other, bool non_blocking=false, bool copy=false, ::std::optional memory_format=::std::nullopt) const; + at::Scalar item() const; + at::Tensor & set_(at::Storage source) const; + at::Tensor & set_(at::Storage source, int64_t storage_offset, at::IntArrayRef size, at::IntArrayRef stride={}) const; + at::Tensor & set__symint(at::Storage source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride={}) const; + at::Tensor & set_(const at::Tensor & source, int64_t storage_offset, at::IntArrayRef size, at::IntArrayRef stride={}) const; + at::Tensor & set__symint(const at::Tensor & source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride={}) const; + at::Tensor & set_(const at::Tensor & source) const; + at::Tensor & set_() const; + bool is_set_to(const at::Tensor & tensor) const; + at::Tensor & masked_fill_(const at::Tensor & mask, const at::Scalar & value) const; + at::Tensor masked_fill(const at::Tensor & mask, const at::Scalar & value) const; + at::Tensor & masked_fill_(const at::Tensor & mask, const at::Tensor & value) const; + at::Tensor masked_fill(const at::Tensor & mask, const at::Tensor & value) const; + at::Tensor & masked_scatter_(const at::Tensor & mask, const at::Tensor & source) const; + at::Tensor masked_scatter(const at::Tensor & mask, const at::Tensor & source) const; + at::Tensor view(at::IntArrayRef size) const; + at::Tensor view_symint(c10::SymIntArrayRef size) const; + at::Tensor view(at::ScalarType dtype) const; + at::Tensor & put_(const at::Tensor & index, const at::Tensor & source, bool accumulate=false) const; + at::Tensor put(const at::Tensor & index, const at::Tensor & source, bool accumulate=false) const; + at::Tensor & index_add_(int64_t dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha=1) const; + at::Tensor index_add(int64_t dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha=1) const; + at::Tensor index_add(at::Dimname dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha=1) const; + at::Tensor & index_reduce_(int64_t dim, const at::Tensor & index, const at::Tensor & source, c10::string_view reduce, bool include_self=true) const; + at::Tensor index_reduce(int64_t dim, const at::Tensor & index, const at::Tensor & source, c10::string_view reduce, bool include_self=true) const; + at::Tensor & index_fill_(int64_t dim, const at::Tensor & index, const at::Scalar & value) const; + at::Tensor index_fill(int64_t dim, const at::Tensor & index, const at::Scalar & value) const; + at::Tensor & index_fill_(int64_t dim, const at::Tensor & index, const at::Tensor & value) const; + at::Tensor index_fill(int64_t dim, const at::Tensor & index, const at::Tensor & value) const; + at::Tensor & index_fill_(at::Dimname dim, const at::Tensor & index, const at::Scalar & value) const; + at::Tensor & index_fill_(at::Dimname dim, const at::Tensor & index, const at::Tensor & value) const; + at::Tensor index_fill(at::Dimname dim, const at::Tensor & index, const at::Scalar & value) const; + at::Tensor index_fill(at::Dimname dim, const at::Tensor & index, const at::Tensor & value) const; + at::Tensor scatter(int64_t dim, const at::Tensor & index, const at::Tensor & src) const; + at::Tensor & scatter_(int64_t dim, const at::Tensor & index, const at::Tensor & src) const; + at::Tensor scatter(int64_t dim, const at::Tensor & index, const at::Scalar & value) const; + at::Tensor & scatter_(int64_t dim, const at::Tensor & index, const at::Scalar & value) const; + at::Tensor scatter(int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce) const; + at::Tensor & scatter_(int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce) const; + at::Tensor scatter(int64_t dim, const at::Tensor & index, const at::Scalar & value, c10::string_view reduce) const; + at::Tensor & scatter_(int64_t dim, const at::Tensor & index, const at::Scalar & value, c10::string_view reduce) const; + at::Tensor scatter(at::Dimname dim, const at::Tensor & index, const at::Tensor & src) const; + at::Tensor scatter(at::Dimname dim, const at::Tensor & index, const at::Scalar & value) const; + at::Tensor scatter_add(int64_t dim, const at::Tensor & index, const at::Tensor & src) const; + at::Tensor & scatter_add_(int64_t dim, const at::Tensor & index, const at::Tensor & src) const; + at::Tensor scatter_add(at::Dimname dim, const at::Tensor & index, const at::Tensor & src) const; + at::Tensor scatter_reduce(int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, bool include_self=true) const; + at::Tensor & scatter_reduce_(int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, bool include_self=true) const; + at::Tensor & eq_(const at::Scalar & other) const; + at::Tensor & eq_(const at::Tensor & other) const; + at::Tensor bitwise_and(const at::Scalar & other) const; + at::Tensor bitwise_and(const at::Tensor & other) const; + at::Tensor & bitwise_and_(const at::Scalar & other) const; + at::Tensor & bitwise_and_(const at::Tensor & other) const; + at::Tensor __and__(const at::Scalar & other) const; + at::Tensor __and__(const at::Tensor & other) const; + at::Tensor & __iand__(const at::Scalar & other) const; + at::Tensor & __iand__(const at::Tensor & other) const; + at::Tensor bitwise_or(const at::Scalar & other) const; + at::Tensor bitwise_or(const at::Tensor & other) const; + at::Tensor & bitwise_or_(const at::Scalar & other) const; + at::Tensor & bitwise_or_(const at::Tensor & other) const; + at::Tensor __or__(const at::Scalar & other) const; + at::Tensor __or__(const at::Tensor & other) const; + at::Tensor & __ior__(const at::Scalar & other) const; + at::Tensor & __ior__(const at::Tensor & other) const; + at::Tensor bitwise_xor(const at::Scalar & other) const; + at::Tensor bitwise_xor(const at::Tensor & other) const; + at::Tensor & bitwise_xor_(const at::Scalar & other) const; + at::Tensor & bitwise_xor_(const at::Tensor & other) const; + at::Tensor __xor__(const at::Scalar & other) const; + at::Tensor __xor__(const at::Tensor & other) const; + at::Tensor & __ixor__(const at::Scalar & other) const; + at::Tensor & __ixor__(const at::Tensor & other) const; + at::Tensor __lshift__(const at::Scalar & other) const; + at::Tensor __lshift__(const at::Tensor & other) const; + at::Tensor & __ilshift__(const at::Scalar & other) const; + at::Tensor & __ilshift__(const at::Tensor & other) const; + at::Tensor bitwise_left_shift(const at::Tensor & other) const; + at::Tensor & bitwise_left_shift_(const at::Tensor & other) const; + at::Tensor bitwise_left_shift(const at::Scalar & other) const; + at::Tensor & bitwise_left_shift_(const at::Scalar & other) const; + at::Tensor __rshift__(const at::Scalar & other) const; + at::Tensor __rshift__(const at::Tensor & other) const; + at::Tensor & __irshift__(const at::Scalar & other) const; + at::Tensor & __irshift__(const at::Tensor & other) const; + at::Tensor bitwise_right_shift(const at::Tensor & other) const; + at::Tensor & bitwise_right_shift_(const at::Tensor & other) const; + at::Tensor bitwise_right_shift(const at::Scalar & other) const; + at::Tensor & bitwise_right_shift_(const at::Scalar & other) const; + at::Tensor & tril_(int64_t diagonal=0) const; + at::Tensor & tril__symint(c10::SymInt diagonal=0) const; + at::Tensor & triu_(int64_t diagonal=0) const; + at::Tensor & triu__symint(c10::SymInt diagonal=0) const; + at::Tensor & digamma_() const; + at::Tensor & lerp_(const at::Tensor & end, const at::Scalar & weight) const; + at::Tensor & lerp_(const at::Tensor & end, const at::Tensor & weight) const; + at::Tensor & addbmm_(const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor addbmm(const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta=1, const at::Scalar & alpha=1) const; + at::Tensor & random_(int64_t from, ::std::optional to, ::std::optional generator=::std::nullopt) const; + at::Tensor & random_(int64_t to, ::std::optional generator=::std::nullopt) const; + at::Tensor & random_(::std::optional generator=::std::nullopt) const; + at::Tensor & uniform_(double from=0, double to=1, ::std::optional generator=::std::nullopt) const; + at::Tensor & cauchy_(double median=0, double sigma=1, ::std::optional generator=::std::nullopt) const; + at::Tensor & log_normal_(double mean=1, double std=2, ::std::optional generator=::std::nullopt) const; + at::Tensor & exponential_(double lambd=1, ::std::optional generator=::std::nullopt) const; + at::Tensor & geometric_(double p, ::std::optional generator=::std::nullopt) const; + at::Tensor diag(int64_t diagonal=0) const; + at::Tensor cross(const at::Tensor & other, ::std::optional dim=::std::nullopt) const; + at::Tensor triu(int64_t diagonal=0) const; + at::Tensor triu_symint(c10::SymInt diagonal=0) const; + at::Tensor tril(int64_t diagonal=0) const; + at::Tensor tril_symint(c10::SymInt diagonal=0) const; + at::Tensor trace() const; + at::Tensor ne(const at::Scalar & other) const; + at::Tensor ne(const at::Tensor & other) const; + at::Tensor & ne_(const at::Scalar & other) const; + at::Tensor & ne_(const at::Tensor & other) const; + at::Tensor not_equal(const at::Scalar & other) const; + at::Tensor not_equal(const at::Tensor & other) const; + at::Tensor & not_equal_(const at::Scalar & other) const; + at::Tensor & not_equal_(const at::Tensor & other) const; + at::Tensor eq(const at::Scalar & other) const; + at::Tensor eq(const at::Tensor & other) const; + at::Tensor ge(const at::Scalar & other) const; + at::Tensor ge(const at::Tensor & other) const; + at::Tensor & ge_(const at::Scalar & other) const; + at::Tensor & ge_(const at::Tensor & other) const; + at::Tensor greater_equal(const at::Scalar & other) const; + at::Tensor greater_equal(const at::Tensor & other) const; + at::Tensor & greater_equal_(const at::Scalar & other) const; + at::Tensor & greater_equal_(const at::Tensor & other) const; + at::Tensor le(const at::Scalar & other) const; + at::Tensor le(const at::Tensor & other) const; + at::Tensor & le_(const at::Scalar & other) const; + at::Tensor & le_(const at::Tensor & other) const; + at::Tensor less_equal(const at::Scalar & other) const; + at::Tensor less_equal(const at::Tensor & other) const; + at::Tensor & less_equal_(const at::Scalar & other) const; + at::Tensor & less_equal_(const at::Tensor & other) const; + at::Tensor gt(const at::Scalar & other) const; + at::Tensor gt(const at::Tensor & other) const; + at::Tensor & gt_(const at::Scalar & other) const; + at::Tensor & gt_(const at::Tensor & other) const; + at::Tensor greater(const at::Scalar & other) const; + at::Tensor greater(const at::Tensor & other) const; + at::Tensor & greater_(const at::Scalar & other) const; + at::Tensor & greater_(const at::Tensor & other) const; + at::Tensor lt(const at::Scalar & other) const; + at::Tensor lt(const at::Tensor & other) const; + at::Tensor & lt_(const at::Scalar & other) const; + at::Tensor & lt_(const at::Tensor & other) const; + at::Tensor less(const at::Scalar & other) const; + at::Tensor less(const at::Tensor & other) const; + at::Tensor & less_(const at::Scalar & other) const; + at::Tensor & less_(const at::Tensor & other) const; + at::Tensor take(const at::Tensor & index) const; + at::Tensor take_along_dim(const at::Tensor & indices, ::std::optional dim=::std::nullopt) const; + at::Tensor index_select(int64_t dim, const at::Tensor & index) const; + at::Tensor index_select(at::Dimname dim, const at::Tensor & index) const; + at::Tensor masked_select(const at::Tensor & mask) const; + at::Tensor nonzero() const; + at::Tensor nonzero_static(int64_t size, int64_t fill_value=-1) const; + at::Tensor nonzero_static_symint(c10::SymInt size, int64_t fill_value=-1) const; + ::std::vector nonzero_numpy() const; + at::Tensor argwhere() const; + at::Tensor gather(int64_t dim, const at::Tensor & index, bool sparse_grad=false) const; + at::Tensor gather(at::Dimname dim, const at::Tensor & index, bool sparse_grad=false) const; + at::Tensor addcmul(const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value=1) const; + at::Tensor & addcmul_(const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value=1) const; + at::Tensor addcdiv(const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value=1) const; + at::Tensor & addcdiv_(const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value=1) const; + ::std::tuple triangular_solve(const at::Tensor & A, bool upper=true, bool transpose=false, bool unitriangular=false) const; + ::std::tuple svd(bool some=true, bool compute_uv=true) const; + at::Tensor swapaxes(int64_t axis0, int64_t axis1) const; + at::Tensor & swapaxes_(int64_t axis0, int64_t axis1) const; + at::Tensor swapdims(int64_t dim0, int64_t dim1) const; + at::Tensor & swapdims_(int64_t dim0, int64_t dim1) const; + at::Tensor cholesky(bool upper=false) const; + at::Tensor cholesky_solve(const at::Tensor & input2, bool upper=false) const; + at::Tensor cholesky_inverse(bool upper=false) const; + ::std::tuple qr(bool some=true) const; + ::std::tuple geqrf() const; + at::Tensor orgqr(const at::Tensor & input2) const; + at::Tensor ormqr(const at::Tensor & input2, const at::Tensor & input3, bool left=true, bool transpose=false) const; + at::Tensor lu_solve(const at::Tensor & LU_data, const at::Tensor & LU_pivots) const; + at::Tensor multinomial(int64_t num_samples, bool replacement=false, ::std::optional generator=::std::nullopt) const; + at::Tensor multinomial_symint(c10::SymInt num_samples, bool replacement=false, ::std::optional generator=::std::nullopt) const; + at::Tensor & lgamma_() const; + at::Tensor lgamma() const; + at::Tensor digamma() const; + at::Tensor polygamma(int64_t n) const; + at::Tensor & polygamma_(int64_t n) const; + at::Tensor erfinv() const; + at::Tensor & erfinv_() const; + at::Tensor i0() const; + at::Tensor & i0_() const; + at::Tensor sign() const; + at::Tensor & sign_() const; + at::Tensor signbit() const; + at::Tensor dist(const at::Tensor & other, const at::Scalar & p=2) const; + at::Tensor & atan2_(const at::Tensor & other) const; + at::Tensor atan2(const at::Tensor & other) const; + at::Tensor arctan2(const at::Tensor & other) const; + at::Tensor & arctan2_(const at::Tensor & other) const; + at::Tensor lerp(const at::Tensor & end, const at::Scalar & weight) const; + at::Tensor lerp(const at::Tensor & end, const at::Tensor & weight) const; + at::Tensor histc(int64_t bins=100, const at::Scalar & min=0, const at::Scalar & max=0) const; + ::std::tuple histogram(const at::Tensor & bins, const ::std::optional & weight={}, bool density=false) const; + ::std::tuple histogram(int64_t bins=100, ::std::optional> range=::std::nullopt, const ::std::optional & weight={}, bool density=false) const; + at::Tensor fmod(const at::Scalar & other) const; + at::Tensor & fmod_(const at::Scalar & other) const; + at::Tensor fmod(const at::Tensor & other) const; + at::Tensor & fmod_(const at::Tensor & other) const; + at::Tensor hypot(const at::Tensor & other) const; + at::Tensor & hypot_(const at::Tensor & other) const; + at::Tensor igamma(const at::Tensor & other) const; + at::Tensor & igamma_(const at::Tensor & other) const; + at::Tensor igammac(const at::Tensor & other) const; + at::Tensor & igammac_(const at::Tensor & other) const; + at::Tensor nextafter(const at::Tensor & other) const; + at::Tensor & nextafter_(const at::Tensor & other) const; + at::Tensor remainder(const at::Scalar & other) const; + at::Tensor & remainder_(const at::Scalar & other) const; + at::Tensor remainder(const at::Tensor & other) const; + at::Tensor & remainder_(const at::Tensor & other) const; + at::Tensor min() const; + at::Tensor fmin(const at::Tensor & other) const; + at::Tensor max() const; + at::Tensor fmax(const at::Tensor & other) const; + at::Tensor maximum(const at::Tensor & other) const; + at::Tensor max(const at::Tensor & other) const; + at::Tensor minimum(const at::Tensor & other) const; + at::Tensor min(const at::Tensor & other) const; + at::Tensor quantile(const at::Tensor & q, ::std::optional dim=::std::nullopt, bool keepdim=false, c10::string_view interpolation="linear") const; + at::Tensor quantile(double q, ::std::optional dim=::std::nullopt, bool keepdim=false, c10::string_view interpolation="linear") const; + at::Tensor nanquantile(const at::Tensor & q, ::std::optional dim=::std::nullopt, bool keepdim=false, c10::string_view interpolation="linear") const; + at::Tensor nanquantile(double q, ::std::optional dim=::std::nullopt, bool keepdim=false, c10::string_view interpolation="linear") const; + ::std::tuple sort(int64_t dim=-1, bool descending=false) const; + ::std::tuple sort(::std::optional stable, int64_t dim=-1, bool descending=false) const; + ::std::tuple sort(at::Dimname dim, bool descending=false) const; + ::std::tuple sort(::std::optional stable, at::Dimname dim, bool descending=false) const; + at::Tensor msort() const; + at::Tensor argsort(int64_t dim=-1, bool descending=false) const; + at::Tensor argsort(bool stable, int64_t dim=-1, bool descending=false) const; + at::Tensor argsort(at::Dimname dim, bool descending=false) const; + ::std::tuple topk(int64_t k, int64_t dim=-1, bool largest=true, bool sorted=true) const; + ::std::tuple topk_symint(c10::SymInt k, int64_t dim=-1, bool largest=true, bool sorted=true) const; + at::Tensor all() const; + at::Tensor any() const; + at::Tensor renorm(const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm) const; + at::Tensor & renorm_(const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm) const; + at::Tensor unfold(int64_t dimension, int64_t size, int64_t step) const; + bool equal(const at::Tensor & other) const; + at::Tensor pow(const at::Tensor & exponent) const; + at::Tensor pow(const at::Scalar & exponent) const; + at::Tensor & pow_(const at::Scalar & exponent) const; + at::Tensor & pow_(const at::Tensor & exponent) const; + at::Tensor float_power(const at::Tensor & exponent) const; + at::Tensor float_power(const at::Scalar & exponent) const; + at::Tensor & float_power_(const at::Scalar & exponent) const; + at::Tensor & float_power_(const at::Tensor & exponent) const; + at::Tensor & normal_(double mean=0, double std=1, ::std::optional generator=::std::nullopt) const; + at::Tensor alias() const; + at::Tensor isfinite() const; + at::Tensor isinf() const; + void record_stream(at::Stream s) const; + at::Tensor isposinf() const; + at::Tensor isneginf() const; + at::Tensor det() const; + ::std::tuple slogdet() const; + at::Tensor logdet() const; + at::Tensor inverse() const; + at::Tensor inner(const at::Tensor & other) const; + at::Tensor outer(const at::Tensor & vec2) const; + at::Tensor ger(const at::Tensor & vec2) const; + at::Tensor to_padded_tensor(double padding, at::OptionalIntArrayRef output_size=::std::nullopt) const; + at::Tensor to_padded_tensor_symint(double padding, at::OptionalSymIntArrayRef output_size=::std::nullopt) const; + + // Special C++ only overloads for std()-like functions (See gh-40287) + // These are needed because int -> bool conversion takes precedence over int -> IntArrayRef + // So, for example std(0) would select the std(unbiased=False) overload + + Tensor var(int dim) const { + return var(IntArrayRef{dim}); + } + + Tensor std(int dim) const { + return std(IntArrayRef{dim}); + } + + // We changed .dtype() to return a TypeMeta in #12766. Ideally, we want the + // at::kDouble and its friends to be TypeMeta's, but that hasn't happened yet. + // Before that change, we make this method to maintain BC for C++ usage like + // `x.to(y.dtype)`. + // TODO: remove following two after at::kDouble and its friends are TypeMeta's. + inline Tensor to(caffe2::TypeMeta type_meta, bool non_blocking=false, bool copy=false) const { + return this->to(/*scalar_type=*/typeMetaToScalarType(type_meta), non_blocking, copy); + } + inline Tensor to(Device device, caffe2::TypeMeta type_meta, bool non_blocking=false, bool copy=false) const { + return this->to(device, /*scalar_type=*/typeMetaToScalarType(type_meta), non_blocking, copy); + } + + template + decltype(auto) m(F func, Args&&... params) const { + return func(*this, std::forward(params)...); + } + + /// NOTE: This is similar to the legacy `.data()` function on `Variable`, and is intended + /// to be used from functions that need to access the `Variable`'s equivalent `Tensor` + /// (i.e. `Tensor` that shares the same storage and tensor metadata with the `Variable`). + /// + /// One notable difference with the legacy `.data()` function is that changes to the + /// returned `Tensor`'s tensor metadata (e.g. sizes / strides / storage / storage_offset) + /// will not update the original `Variable`, due to the fact that this function + /// shallow-copies the `Variable`'s underlying TensorImpl. + at::Tensor tensor_data() const { + return TensorBase::tensor_data(); + } + + /// NOTE: `var.variable_data()` in C++ has the same semantics as `tensor.data` + /// in Python, which create a new `Variable` that shares the same storage and + /// tensor metadata with the original `Variable`, but with a completely new + /// autograd history. + /// + /// NOTE: If we change the tensor metadata (e.g. sizes / strides / + /// storage / storage_offset) of a variable created from `var.variable_data()`, those + /// changes will not update the original variable `var`. In `.variable_data()`, we set + /// `allow_tensor_metadata_change_` to false to make such changes explicitly illegal, + /// in order to prevent users from changing metadata of `var.variable_data()` + /// and expecting the original variable `var` to also be updated. + at::Tensor variable_data() const { + return TensorBase::variable_data(); + } + + // Hooks + //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + template + using hook_return_void_t = std::enable_if_t>::value, unsigned>; + template + using hook_return_var_t = std::enable_if_t, Tensor>, unsigned>; + + /// Registers a backward hook. + /// + /// The hook will be called every time a gradient with respect to the Tensor is computed. + /// The hook should have one of the following signature: + /// ``` + /// hook(Tensor grad) -> Tensor + /// ``` + /// ``` + /// hook(Tensor grad) -> void + /// ``` + /// The hook should not modify its argument, but it can optionally return a new gradient + /// which will be used in place of `grad`. + /// + /// This function returns the index of the hook in the list which can be used to remove hook. + /// + /// Example: + /// @code + /// auto v = torch::tensor({0., 0., 0.}, torch::requires_grad()); + /// auto h = v.register_hook([](torch::Tensor grad){ return grad * 2; }); // double the gradient + /// v.backward(torch::tensor({1., 2., 3.})); + /// // This prints: + /// // ``` + /// // 2 + /// // 4 + /// // 6 + /// // [ CPUFloatType{3} ] + /// // ``` + /// std::cout << v.grad() << std::endl; + /// v.remove_hook(h); // removes the hook + /// @endcode + template + hook_return_void_t register_hook(T&& hook) const; + template + hook_return_var_t register_hook(T&& hook) const; + + // Variable methods + //~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + + Tensor data() const { + return TensorBase::data(); + } + + void _backward(TensorList inputs, const std::optional& gradient, std::optional keep_graph, bool create_graph) const; + + const Tensor& requires_grad_(bool _requires_grad=true) const { + TensorBase::requires_grad_(_requires_grad); + return *this; + } +}; + +namespace detail { +// Helper creator for Tensor class which doesn't requires the users to pass +// in an intrusive_ptr instead it just converts the argument passed to +// requested intrusive_ptr type. +template +Tensor make_tensor(Args&&... args) { + return Tensor(c10::make_intrusive(std::forward(args)...)); +} + +} // namespace detail + +} // namespace at + + +namespace at { + +// aten::_backward(Tensor self, Tensor[] inputs, Tensor? gradient=None, bool? retain_graph=None, bool create_graph=False) -> () +inline void Tensor::__dispatch__backward(at::TensorList inputs, const ::std::optional & gradient, ::std::optional retain_graph, bool create_graph) const { + return at::_ops::_backward::call(const_cast(*this), inputs, gradient, retain_graph, create_graph); +} + +// aten::set_data(Tensor(a!) self, Tensor new_data) -> () +inline void Tensor::__dispatch_set_data(const at::Tensor & new_data) const { + return at::_ops::set_data::call(const_cast(*this), new_data); +} + +// aten::data(Tensor self) -> Tensor +inline at::Tensor Tensor::__dispatch_data() const { + return at::_ops::data::call(const_cast(*this)); +} + +// aten::is_leaf(Tensor self) -> bool +inline bool Tensor::__dispatch_is_leaf() const { + return at::_ops::is_leaf::call(const_cast(*this)); +} + +// aten::output_nr(Tensor self) -> int +inline int64_t Tensor::__dispatch_output_nr() const { + return at::_ops::output_nr::call(const_cast(*this)); +} + +// aten::_version(Tensor self) -> int +inline int64_t Tensor::__dispatch__version() const { + return at::_ops::_version::call(const_cast(*this)); +} + +// aten::requires_grad_(Tensor(a!) self, bool requires_grad=True) -> Tensor(a!) +inline at::Tensor & Tensor::__dispatch_requires_grad_(bool requires_grad) const { + return at::_ops::requires_grad_::call(const_cast(*this), requires_grad); +} + +// aten::retain_grad(Tensor(a!) self) -> () +inline void Tensor::__dispatch_retain_grad() const { + return at::_ops::retain_grad::call(const_cast(*this)); +} + +// aten::retains_grad(Tensor self) -> bool +inline bool Tensor::__dispatch_retains_grad() const { + return at::_ops::retains_grad::call(const_cast(*this)); +} + +// aten::_fw_primal(Tensor(a) self, int level) -> Tensor(a) +inline at::Tensor Tensor::_fw_primal(int64_t level) const { + return at::_ops::_fw_primal::call(const_cast(*this), level); +} + +// aten::rename_(Tensor(a!) self, Dimname[]? names) -> Tensor(a!) +inline at::Tensor & Tensor::rename_(::std::optional names) const { + return at::_ops::rename_::call(const_cast(*this), names); +} + +// aten::rename(Tensor(a) self, Dimname[]? names) -> Tensor(a) +inline at::Tensor Tensor::rename(::std::optional names) const { + return at::_ops::rename::call(const_cast(*this), names); +} + +// aten::align_to(Tensor(a) self, Dimname[] names) -> Tensor(a) +inline at::Tensor Tensor::align_to(at::DimnameList names) const { + return at::_ops::align_to::call(const_cast(*this), names); +} + +// aten::align_to.ellipsis_idx(Tensor(a) self, Dimname[] order, int ellipsis_idx) -> Tensor(a) +inline at::Tensor Tensor::align_to(at::DimnameList order, int64_t ellipsis_idx) const { + return at::_ops::align_to_ellipsis_idx::call(const_cast(*this), order, ellipsis_idx); +} + +// aten::align_as(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::align_as(const at::Tensor & other) const { + return at::_ops::align_as::call(const_cast(*this), other); +} + +// aten::refine_names(Tensor(a) self, Dimname[] names) -> Tensor(a) +inline at::Tensor Tensor::refine_names(at::DimnameList names) const { + return at::_ops::refine_names::call(const_cast(*this), names); +} + +// aten::abs(Tensor self) -> Tensor +inline at::Tensor Tensor::abs() const { + return at::_ops::abs::call(const_cast(*this)); +} + +// aten::abs_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::abs_() const { + return at::_ops::abs_::call(const_cast(*this)); +} + +// aten::absolute(Tensor self) -> Tensor +inline at::Tensor Tensor::absolute() const { + return at::_ops::absolute::call(const_cast(*this)); +} + +// aten::absolute_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::absolute_() const { + return at::_ops::absolute_::call(const_cast(*this)); +} + +// aten::angle(Tensor self) -> Tensor +inline at::Tensor Tensor::angle() const { + return at::_ops::angle::call(const_cast(*this)); +} + +// aten::sgn(Tensor self) -> Tensor +inline at::Tensor Tensor::sgn() const { + return at::_ops::sgn::call(const_cast(*this)); +} + +// aten::sgn_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::sgn_() const { + return at::_ops::sgn_::call(const_cast(*this)); +} + +// aten::chalf(Tensor self, *, MemoryFormat? memory_format=None) -> Tensor +inline at::Tensor Tensor::chalf(::std::optional memory_format) const { + return at::_ops::chalf::call(const_cast(*this), memory_format); +} + +// aten::_conj(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::_conj() const { + return at::_ops::_conj::call(const_cast(*this)); +} + +// aten::conj(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::__dispatch_conj() const { + return at::_ops::conj::call(const_cast(*this)); +} + +// aten::_conj_physical(Tensor self) -> Tensor +inline at::Tensor Tensor::_conj_physical() const { + return at::_ops::_conj_physical::call(const_cast(*this)); +} + +// aten::conj_physical(Tensor self) -> Tensor +inline at::Tensor Tensor::conj_physical() const { + return at::_ops::conj_physical::call(const_cast(*this)); +} + +// aten::conj_physical_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::conj_physical_() const { + return at::_ops::conj_physical_::call(const_cast(*this)); +} + +// aten::resolve_conj(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::resolve_conj() const { + return at::_ops::resolve_conj::call(const_cast(*this)); +} + +// aten::resolve_neg(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::resolve_neg() const { + return at::_ops::resolve_neg::call(const_cast(*this)); +} + +// aten::_neg_view(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::_neg_view() const { + return at::_ops::_neg_view::call(const_cast(*this)); +} + +// aten::acos(Tensor self) -> Tensor +inline at::Tensor Tensor::acos() const { + return at::_ops::acos::call(const_cast(*this)); +} + +// aten::acos_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::acos_() const { + return at::_ops::acos_::call(const_cast(*this)); +} + +// aten::arccos(Tensor self) -> Tensor +inline at::Tensor Tensor::arccos() const { + return at::_ops::arccos::call(const_cast(*this)); +} + +// aten::arccos_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::arccos_() const { + return at::_ops::arccos_::call(const_cast(*this)); +} + +// aten::add.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::add(const at::Tensor & other, const at::Scalar & alpha) const { + return at::_ops::add_Tensor::call(const_cast(*this), other, alpha); +} + +// aten::add_.Tensor(Tensor(a!) self, Tensor other, *, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::add_(const at::Tensor & other, const at::Scalar & alpha) const { + return at::_ops::add__Tensor::call(const_cast(*this), other, alpha); +} + +// aten::add.Scalar(Tensor self, Scalar other, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::add(const at::Scalar & other, const at::Scalar & alpha) const { + return at::_ops::add_Scalar::call(const_cast(*this), other, alpha); +} + +// aten::add_.Scalar(Tensor(a!) self, Scalar other, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::add_(const at::Scalar & other, const at::Scalar & alpha) const { + return at::_ops::add__Scalar::call(const_cast(*this), other, alpha); +} + +// aten::addmv(Tensor self, Tensor mat, Tensor vec, *, Scalar beta=1, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::addmv(const at::Tensor & mat, const at::Tensor & vec, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::addmv::call(const_cast(*this), mat, vec, beta, alpha); +} + +// aten::addmv_(Tensor(a!) self, Tensor mat, Tensor vec, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::addmv_(const at::Tensor & mat, const at::Tensor & vec, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::addmv_::call(const_cast(*this), mat, vec, beta, alpha); +} + +// aten::addr(Tensor self, Tensor vec1, Tensor vec2, *, Scalar beta=1, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::addr(const at::Tensor & vec1, const at::Tensor & vec2, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::addr::call(const_cast(*this), vec1, vec2, beta, alpha); +} + +// aten::addr_(Tensor(a!) self, Tensor vec1, Tensor vec2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::addr_(const at::Tensor & vec1, const at::Tensor & vec2, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::addr_::call(const_cast(*this), vec1, vec2, beta, alpha); +} + +// aten::_is_all_true(Tensor self) -> Tensor +inline at::Tensor Tensor::_is_all_true() const { + return at::_ops::_is_all_true::call(const_cast(*this)); +} + +// aten::_is_any_true(Tensor self) -> Tensor +inline at::Tensor Tensor::_is_any_true() const { + return at::_ops::_is_any_true::call(const_cast(*this)); +} + +// aten::all.dim(Tensor self, int dim, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::all(int64_t dim, bool keepdim) const { + return at::_ops::all_dim::call(const_cast(*this), dim, keepdim); +} + +// aten::all.dims(Tensor self, int[]? dim=None, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::all(at::OptionalIntArrayRef dim, bool keepdim) const { + return at::_ops::all_dims::call(const_cast(*this), dim, keepdim); +} + +// aten::all.dimname(Tensor self, Dimname dim, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::all(at::Dimname dim, bool keepdim) const { + return at::_ops::all_dimname::call(const_cast(*this), dim, keepdim); +} + +// aten::allclose(Tensor self, Tensor other, float rtol=1e-05, float atol=1e-08, bool equal_nan=False) -> bool +inline bool Tensor::allclose(const at::Tensor & other, double rtol, double atol, bool equal_nan) const { + return at::_ops::allclose::call(const_cast(*this), other, rtol, atol, equal_nan); +} + +// aten::any.dim(Tensor self, int dim, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::any(int64_t dim, bool keepdim) const { + return at::_ops::any_dim::call(const_cast(*this), dim, keepdim); +} + +// aten::any.dims(Tensor self, int[]? dim=None, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::any(at::OptionalIntArrayRef dim, bool keepdim) const { + return at::_ops::any_dims::call(const_cast(*this), dim, keepdim); +} + +// aten::any.dimname(Tensor self, Dimname dim, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::any(at::Dimname dim, bool keepdim) const { + return at::_ops::any_dimname::call(const_cast(*this), dim, keepdim); +} + +// aten::argmax(Tensor self, int? dim=None, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::argmax(::std::optional dim, bool keepdim) const { + return at::_ops::argmax::call(const_cast(*this), dim, keepdim); +} + +// aten::argmin(Tensor self, int? dim=None, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::argmin(::std::optional dim, bool keepdim) const { + return at::_ops::argmin::call(const_cast(*this), dim, keepdim); +} + +// aten::acosh(Tensor self) -> Tensor +inline at::Tensor Tensor::acosh() const { + return at::_ops::acosh::call(const_cast(*this)); +} + +// aten::acosh_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::acosh_() const { + return at::_ops::acosh_::call(const_cast(*this)); +} + +// aten::arccosh(Tensor self) -> Tensor +inline at::Tensor Tensor::arccosh() const { + return at::_ops::arccosh::call(const_cast(*this)); +} + +// aten::arccosh_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::arccosh_() const { + return at::_ops::arccosh_::call(const_cast(*this)); +} + +// aten::asinh(Tensor self) -> Tensor +inline at::Tensor Tensor::asinh() const { + return at::_ops::asinh::call(const_cast(*this)); +} + +// aten::asinh_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::asinh_() const { + return at::_ops::asinh_::call(const_cast(*this)); +} + +// aten::arcsinh(Tensor self) -> Tensor +inline at::Tensor Tensor::arcsinh() const { + return at::_ops::arcsinh::call(const_cast(*this)); +} + +// aten::arcsinh_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::arcsinh_() const { + return at::_ops::arcsinh_::call(const_cast(*this)); +} + +// aten::atanh(Tensor self) -> Tensor +inline at::Tensor Tensor::atanh() const { + return at::_ops::atanh::call(const_cast(*this)); +} + +// aten::atanh_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::atanh_() const { + return at::_ops::atanh_::call(const_cast(*this)); +} + +// aten::arctanh(Tensor self) -> Tensor +inline at::Tensor Tensor::arctanh() const { + return at::_ops::arctanh::call(const_cast(*this)); +} + +// aten::arctanh_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::arctanh_() const { + return at::_ops::arctanh_::call(const_cast(*this)); +} + +// aten::as_strided(Tensor(a) self, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None) -> Tensor(a) +inline at::Tensor Tensor::as_strided(at::IntArrayRef size, at::IntArrayRef stride, ::std::optional storage_offset) const { + return at::_ops::as_strided::call(const_cast(*this), c10::fromIntArrayRefSlow(size), c10::fromIntArrayRefSlow(stride), storage_offset.has_value() ? ::std::make_optional(c10::SymInt(*storage_offset)) : ::std::nullopt); +} + +// aten::as_strided(Tensor(a) self, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None) -> Tensor(a) +inline at::Tensor Tensor::as_strided_symint(c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset) const { + return at::_ops::as_strided::call(const_cast(*this), size, stride, storage_offset); +} + +// aten::as_strided_(Tensor(a!) self, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None) -> Tensor(a!) +inline const at::Tensor & Tensor::as_strided_(at::IntArrayRef size, at::IntArrayRef stride, ::std::optional storage_offset) const { + return at::_ops::as_strided_::call(const_cast(*this), c10::fromIntArrayRefSlow(size), c10::fromIntArrayRefSlow(stride), storage_offset.has_value() ? ::std::make_optional(c10::SymInt(*storage_offset)) : ::std::nullopt); +} + +// aten::as_strided_(Tensor(a!) self, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None) -> Tensor(a!) +inline const at::Tensor & Tensor::as_strided__symint(c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset) const { + return at::_ops::as_strided_::call(const_cast(*this), size, stride, storage_offset); +} + +// aten::asin(Tensor self) -> Tensor +inline at::Tensor Tensor::asin() const { + return at::_ops::asin::call(const_cast(*this)); +} + +// aten::asin_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::asin_() const { + return at::_ops::asin_::call(const_cast(*this)); +} + +// aten::arcsin(Tensor self) -> Tensor +inline at::Tensor Tensor::arcsin() const { + return at::_ops::arcsin::call(const_cast(*this)); +} + +// aten::arcsin_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::arcsin_() const { + return at::_ops::arcsin_::call(const_cast(*this)); +} + +// aten::atan(Tensor self) -> Tensor +inline at::Tensor Tensor::atan() const { + return at::_ops::atan::call(const_cast(*this)); +} + +// aten::atan_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::atan_() const { + return at::_ops::atan_::call(const_cast(*this)); +} + +// aten::arctan(Tensor self) -> Tensor +inline at::Tensor Tensor::arctan() const { + return at::_ops::arctan::call(const_cast(*this)); +} + +// aten::arctan_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::arctan_() const { + return at::_ops::arctan_::call(const_cast(*this)); +} + +// aten::baddbmm(Tensor self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::baddbmm(const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::baddbmm::call(const_cast(*this), batch1, batch2, beta, alpha); +} + +// aten::baddbmm_(Tensor(a!) self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::baddbmm_(const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::baddbmm_::call(const_cast(*this), batch1, batch2, beta, alpha); +} + +// aten::bernoulli(Tensor self, *, Generator? generator=None) -> Tensor +inline at::Tensor Tensor::bernoulli(::std::optional generator) const { + return at::_ops::bernoulli::call(const_cast(*this), generator); +} + +// aten::bernoulli_.Tensor(Tensor(a!) self, Tensor p, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::bernoulli_(const at::Tensor & p, ::std::optional generator) const { + return at::_ops::bernoulli__Tensor::call(const_cast(*this), p, generator); +} + +// aten::bernoulli_.float(Tensor(a!) self, float p=0.5, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::bernoulli_(double p, ::std::optional generator) const { + return at::_ops::bernoulli__float::call(const_cast(*this), p, generator); +} + +// aten::bernoulli.p(Tensor self, float p, *, Generator? generator=None) -> Tensor +inline at::Tensor Tensor::bernoulli(double p, ::std::optional generator) const { + return at::_ops::bernoulli_p::call(const_cast(*this), p, generator); +} + +// aten::bincount(Tensor self, Tensor? weights=None, SymInt minlength=0) -> Tensor +inline at::Tensor Tensor::bincount(const ::std::optional & weights, int64_t minlength) const { + return at::_ops::bincount::call(const_cast(*this), weights, minlength); +} + +// aten::bincount(Tensor self, Tensor? weights=None, SymInt minlength=0) -> Tensor +inline at::Tensor Tensor::bincount_symint(const ::std::optional & weights, c10::SymInt minlength) const { + return at::_ops::bincount::call(const_cast(*this), weights, minlength); +} + +// aten::bitwise_not(Tensor self) -> Tensor +inline at::Tensor Tensor::bitwise_not() const { + return at::_ops::bitwise_not::call(const_cast(*this)); +} + +// aten::bitwise_not_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_not_() const { + return at::_ops::bitwise_not_::call(const_cast(*this)); +} + +// aten::copysign.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::copysign(const at::Tensor & other) const { + return at::_ops::copysign_Tensor::call(const_cast(*this), other); +} + +// aten::copysign_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::copysign_(const at::Tensor & other) const { + return at::_ops::copysign__Tensor::call(const_cast(*this), other); +} + +// aten::copysign.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::copysign(const at::Scalar & other) const { + return at::_ops::copysign_Scalar::call(const_cast(*this), other); +} + +// aten::copysign_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::copysign_(const at::Scalar & other) const { + return at::_ops::copysign__Scalar::call(const_cast(*this), other); +} + +// aten::_lazy_clone(Tensor self) -> Tensor +inline at::Tensor Tensor::_lazy_clone() const { + return at::_ops::_lazy_clone::call(const_cast(*this)); +} + +// aten::logical_not(Tensor self) -> Tensor +inline at::Tensor Tensor::logical_not() const { + return at::_ops::logical_not::call(const_cast(*this)); +} + +// aten::logical_not_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::logical_not_() const { + return at::_ops::logical_not_::call(const_cast(*this)); +} + +// aten::logical_xor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::logical_xor(const at::Tensor & other) const { + return at::_ops::logical_xor::call(const_cast(*this), other); +} + +// aten::logical_xor_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::logical_xor_(const at::Tensor & other) const { + return at::_ops::logical_xor_::call(const_cast(*this), other); +} + +// aten::logical_and(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::logical_and(const at::Tensor & other) const { + return at::_ops::logical_and::call(const_cast(*this), other); +} + +// aten::logical_and_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::logical_and_(const at::Tensor & other) const { + return at::_ops::logical_and_::call(const_cast(*this), other); +} + +// aten::logical_or(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::logical_or(const at::Tensor & other) const { + return at::_ops::logical_or::call(const_cast(*this), other); +} + +// aten::logical_or_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::logical_or_(const at::Tensor & other) const { + return at::_ops::logical_or_::call(const_cast(*this), other); +} + +// aten::bmm(Tensor self, Tensor mat2) -> Tensor +inline at::Tensor Tensor::bmm(const at::Tensor & mat2) const { + return at::_ops::bmm::call(const_cast(*this), mat2); +} + +// aten::broadcast_to(Tensor(a) self, SymInt[] size) -> Tensor(a) +inline at::Tensor Tensor::broadcast_to(at::IntArrayRef size) const { + return at::_ops::broadcast_to::call(const_cast(*this), c10::fromIntArrayRefSlow(size)); +} + +// aten::broadcast_to(Tensor(a) self, SymInt[] size) -> Tensor(a) +inline at::Tensor Tensor::broadcast_to_symint(c10::SymIntArrayRef size) const { + return at::_ops::broadcast_to::call(const_cast(*this), size); +} + +// aten::ceil(Tensor self) -> Tensor +inline at::Tensor Tensor::ceil() const { + return at::_ops::ceil::call(const_cast(*this)); +} + +// aten::ceil_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::ceil_() const { + return at::_ops::ceil_::call(const_cast(*this)); +} + +// aten::unsafe_chunk(Tensor self, int chunks, int dim=0) -> Tensor[] +inline ::std::vector Tensor::unsafe_chunk(int64_t chunks, int64_t dim) const { + return at::_ops::unsafe_chunk::call(const_cast(*this), chunks, dim); +} + +// aten::chunk(Tensor(a -> *) self, int chunks, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::chunk(int64_t chunks, int64_t dim) const { + return at::_ops::chunk::call(const_cast(*this), chunks, dim); +} + +// aten::tensor_split.sections(Tensor(a -> *) self, SymInt sections, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::tensor_split(int64_t sections, int64_t dim) const { + return at::_ops::tensor_split_sections::call(const_cast(*this), sections, dim); +} + +// aten::tensor_split.sections(Tensor(a -> *) self, SymInt sections, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::tensor_split_symint(c10::SymInt sections, int64_t dim) const { + return at::_ops::tensor_split_sections::call(const_cast(*this), sections, dim); +} + +// aten::tensor_split.indices(Tensor(a -> *) self, SymInt[] indices, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::tensor_split(at::IntArrayRef indices, int64_t dim) const { + return at::_ops::tensor_split_indices::call(const_cast(*this), c10::fromIntArrayRefSlow(indices), dim); +} + +// aten::tensor_split.indices(Tensor(a -> *) self, SymInt[] indices, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::tensor_split_symint(c10::SymIntArrayRef indices, int64_t dim) const { + return at::_ops::tensor_split_indices::call(const_cast(*this), indices, dim); +} + +// aten::tensor_split.tensor_indices_or_sections(Tensor(a -> *) self, Tensor tensor_indices_or_sections, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::tensor_split(const at::Tensor & tensor_indices_or_sections, int64_t dim) const { + return at::_ops::tensor_split_tensor_indices_or_sections::call(const_cast(*this), tensor_indices_or_sections, dim); +} + +// aten::clamp(Tensor self, Scalar? min=None, Scalar? max=None) -> Tensor +inline at::Tensor Tensor::clamp(const ::std::optional & min, const ::std::optional & max) const { + return at::_ops::clamp::call(const_cast(*this), min, max); +} + +// aten::clamp.Tensor(Tensor self, Tensor? min=None, Tensor? max=None) -> Tensor +inline at::Tensor Tensor::clamp(const ::std::optional & min, const ::std::optional & max) const { + return at::_ops::clamp_Tensor::call(const_cast(*this), min, max); +} + +// aten::clamp_(Tensor(a!) self, Scalar? min=None, Scalar? max=None) -> Tensor(a!) +inline at::Tensor & Tensor::clamp_(const ::std::optional & min, const ::std::optional & max) const { + return at::_ops::clamp_::call(const_cast(*this), min, max); +} + +// aten::clamp_.Tensor(Tensor(a!) self, Tensor? min=None, Tensor? max=None) -> Tensor(a!) +inline at::Tensor & Tensor::clamp_(const ::std::optional & min, const ::std::optional & max) const { + return at::_ops::clamp__Tensor::call(const_cast(*this), min, max); +} + +// aten::clamp_max(Tensor self, Scalar max) -> Tensor +inline at::Tensor Tensor::clamp_max(const at::Scalar & max) const { + return at::_ops::clamp_max::call(const_cast(*this), max); +} + +// aten::clamp_max.Tensor(Tensor self, Tensor max) -> Tensor +inline at::Tensor Tensor::clamp_max(const at::Tensor & max) const { + return at::_ops::clamp_max_Tensor::call(const_cast(*this), max); +} + +// aten::clamp_max_(Tensor(a!) self, Scalar max) -> Tensor(a!) +inline at::Tensor & Tensor::clamp_max_(const at::Scalar & max) const { + return at::_ops::clamp_max_::call(const_cast(*this), max); +} + +// aten::clamp_max_.Tensor(Tensor(a!) self, Tensor max) -> Tensor(a!) +inline at::Tensor & Tensor::clamp_max_(const at::Tensor & max) const { + return at::_ops::clamp_max__Tensor::call(const_cast(*this), max); +} + +// aten::clamp_min(Tensor self, Scalar min) -> Tensor +inline at::Tensor Tensor::clamp_min(const at::Scalar & min) const { + return at::_ops::clamp_min::call(const_cast(*this), min); +} + +// aten::clamp_min.Tensor(Tensor self, Tensor min) -> Tensor +inline at::Tensor Tensor::clamp_min(const at::Tensor & min) const { + return at::_ops::clamp_min_Tensor::call(const_cast(*this), min); +} + +// aten::clamp_min_(Tensor(a!) self, Scalar min) -> Tensor(a!) +inline at::Tensor & Tensor::clamp_min_(const at::Scalar & min) const { + return at::_ops::clamp_min_::call(const_cast(*this), min); +} + +// aten::clamp_min_.Tensor(Tensor(a!) self, Tensor min) -> Tensor(a!) +inline at::Tensor & Tensor::clamp_min_(const at::Tensor & min) const { + return at::_ops::clamp_min__Tensor::call(const_cast(*this), min); +} + +// aten::clip(Tensor self, Scalar? min=None, Scalar? max=None) -> Tensor +inline at::Tensor Tensor::clip(const ::std::optional & min, const ::std::optional & max) const { + return at::_ops::clip::call(const_cast(*this), min, max); +} + +// aten::clip.Tensor(Tensor self, Tensor? min=None, Tensor? max=None) -> Tensor +inline at::Tensor Tensor::clip(const ::std::optional & min, const ::std::optional & max) const { + return at::_ops::clip_Tensor::call(const_cast(*this), min, max); +} + +// aten::clip_(Tensor(a!) self, Scalar? min=None, Scalar? max=None) -> Tensor(a!) +inline at::Tensor & Tensor::clip_(const ::std::optional & min, const ::std::optional & max) const { + return at::_ops::clip_::call(const_cast(*this), min, max); +} + +// aten::clip_.Tensor(Tensor(a!) self, Tensor? min=None, Tensor? max=None) -> Tensor(a!) +inline at::Tensor & Tensor::clip_(const ::std::optional & min, const ::std::optional & max) const { + return at::_ops::clip__Tensor::call(const_cast(*this), min, max); +} + +// aten::contiguous(Tensor(a) self, *, MemoryFormat memory_format=contiguous_format) -> Tensor(a) +inline at::Tensor Tensor::__dispatch_contiguous(at::MemoryFormat memory_format) const { + return at::_ops::contiguous::call(const_cast(*this), memory_format); +} + +// aten::copy_(Tensor(a!) self, Tensor src, bool non_blocking=False) -> Tensor(a!) +inline at::Tensor & Tensor::copy_(const at::Tensor & src, bool non_blocking) const { + return at::_ops::copy_::call(const_cast(*this), src, non_blocking); +} + +// aten::cos(Tensor self) -> Tensor +inline at::Tensor Tensor::cos() const { + return at::_ops::cos::call(const_cast(*this)); +} + +// aten::cos_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::cos_() const { + return at::_ops::cos_::call(const_cast(*this)); +} + +// aten::cosh(Tensor self) -> Tensor +inline at::Tensor Tensor::cosh() const { + return at::_ops::cosh::call(const_cast(*this)); +} + +// aten::cosh_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::cosh_() const { + return at::_ops::cosh_::call(const_cast(*this)); +} + +// aten::count_nonzero.dim_IntList(Tensor self, int[] dim) -> Tensor +inline at::Tensor Tensor::count_nonzero(at::IntArrayRef dim) const { + return at::_ops::count_nonzero_dim_IntList::call(const_cast(*this), dim); +} + +// aten::count_nonzero(Tensor self, int? dim=None) -> Tensor +inline at::Tensor Tensor::count_nonzero(::std::optional dim) const { + return at::_ops::count_nonzero::call(const_cast(*this), dim); +} + +// aten::cov(Tensor self, *, int correction=1, Tensor? fweights=None, Tensor? aweights=None) -> Tensor +inline at::Tensor Tensor::cov(int64_t correction, const ::std::optional & fweights, const ::std::optional & aweights) const { + return at::_ops::cov::call(const_cast(*this), correction, fweights, aweights); +} + +// aten::corrcoef(Tensor self) -> Tensor +inline at::Tensor Tensor::corrcoef() const { + return at::_ops::corrcoef::call(const_cast(*this)); +} + +// aten::cummax(Tensor self, int dim) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::cummax(int64_t dim) const { + return at::_ops::cummax::call(const_cast(*this), dim); +} + +// aten::cummax.dimname(Tensor self, Dimname dim) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::cummax(at::Dimname dim) const { + return at::_ops::cummax_dimname::call(const_cast(*this), dim); +} + +// aten::cummin(Tensor self, int dim) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::cummin(int64_t dim) const { + return at::_ops::cummin::call(const_cast(*this), dim); +} + +// aten::cummin.dimname(Tensor self, Dimname dim) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::cummin(at::Dimname dim) const { + return at::_ops::cummin_dimname::call(const_cast(*this), dim); +} + +// aten::cumprod(Tensor self, int dim, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::cumprod(int64_t dim, ::std::optional dtype) const { + return at::_ops::cumprod::call(const_cast(*this), dim, dtype); +} + +// aten::cumprod_(Tensor(a!) self, int dim, *, ScalarType? dtype=None) -> Tensor(a!) +inline at::Tensor & Tensor::cumprod_(int64_t dim, ::std::optional dtype) const { + return at::_ops::cumprod_::call(const_cast(*this), dim, dtype); +} + +// aten::cumprod.dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::cumprod(at::Dimname dim, ::std::optional dtype) const { + return at::_ops::cumprod_dimname::call(const_cast(*this), dim, dtype); +} + +// aten::cumprod_.dimname(Tensor(a!) self, Dimname dim, *, ScalarType? dtype=None) -> Tensor(a!) +inline at::Tensor & Tensor::cumprod_(at::Dimname dim, ::std::optional dtype) const { + return at::_ops::cumprod__dimname::call(const_cast(*this), dim, dtype); +} + +// aten::cumsum(Tensor self, int dim, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::cumsum(int64_t dim, ::std::optional dtype) const { + return at::_ops::cumsum::call(const_cast(*this), dim, dtype); +} + +// aten::cumsum_(Tensor(a!) self, int dim, *, ScalarType? dtype=None) -> Tensor(a!) +inline at::Tensor & Tensor::cumsum_(int64_t dim, ::std::optional dtype) const { + return at::_ops::cumsum_::call(const_cast(*this), dim, dtype); +} + +// aten::cumsum.dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::cumsum(at::Dimname dim, ::std::optional dtype) const { + return at::_ops::cumsum_dimname::call(const_cast(*this), dim, dtype); +} + +// aten::cumsum_.dimname(Tensor(a!) self, Dimname dim, *, ScalarType? dtype=None) -> Tensor(a!) +inline at::Tensor & Tensor::cumsum_(at::Dimname dim, ::std::optional dtype) const { + return at::_ops::cumsum__dimname::call(const_cast(*this), dim, dtype); +} + +// aten::diag_embed(Tensor self, int offset=0, int dim1=-2, int dim2=-1) -> Tensor +inline at::Tensor Tensor::diag_embed(int64_t offset, int64_t dim1, int64_t dim2) const { + return at::_ops::diag_embed::call(const_cast(*this), offset, dim1, dim2); +} + +// aten::diagflat(Tensor self, int offset=0) -> Tensor +inline at::Tensor Tensor::diagflat(int64_t offset) const { + return at::_ops::diagflat::call(const_cast(*this), offset); +} + +// aten::diagonal(Tensor(a) self, int offset=0, int dim1=0, int dim2=1) -> Tensor(a) +inline at::Tensor Tensor::diagonal(int64_t offset, int64_t dim1, int64_t dim2) const { + return at::_ops::diagonal::call(const_cast(*this), offset, dim1, dim2); +} + +// aten::diagonal.Dimname(Tensor(a) self, *, Dimname outdim, Dimname dim1, Dimname dim2, int offset=0) -> Tensor(a) +inline at::Tensor Tensor::diagonal(at::Dimname outdim, at::Dimname dim1, at::Dimname dim2, int64_t offset) const { + return at::_ops::diagonal_Dimname::call(const_cast(*this), outdim, dim1, dim2, offset); +} + +// aten::fill_diagonal_(Tensor(a!) self, Scalar fill_value, bool wrap=False) -> Tensor(a!) +inline at::Tensor & Tensor::fill_diagonal_(const at::Scalar & fill_value, bool wrap) const { + return at::_ops::fill_diagonal_::call(const_cast(*this), fill_value, wrap); +} + +// aten::diff(Tensor self, int n=1, int dim=-1, Tensor? prepend=None, Tensor? append=None) -> Tensor +inline at::Tensor Tensor::diff(int64_t n, int64_t dim, const ::std::optional & prepend, const ::std::optional & append) const { + return at::_ops::diff::call(const_cast(*this), n, dim, prepend, append); +} + +// aten::div.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::div(const at::Tensor & other) const { + return at::_ops::div_Tensor::call(const_cast(*this), other); +} + +// aten::div_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::div_(const at::Tensor & other) const { + return at::_ops::div__Tensor::call(const_cast(*this), other); +} + +// aten::div.Tensor_mode(Tensor self, Tensor other, *, str? rounding_mode) -> Tensor +inline at::Tensor Tensor::div(const at::Tensor & other, ::std::optional rounding_mode) const { + return at::_ops::div_Tensor_mode::call(const_cast(*this), other, rounding_mode); +} + +// aten::div_.Tensor_mode(Tensor(a!) self, Tensor other, *, str? rounding_mode) -> Tensor(a!) +inline at::Tensor & Tensor::div_(const at::Tensor & other, ::std::optional rounding_mode) const { + return at::_ops::div__Tensor_mode::call(const_cast(*this), other, rounding_mode); +} + +// aten::div.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::div(const at::Scalar & other) const { + return at::_ops::div_Scalar::call(const_cast(*this), other); +} + +// aten::div_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::div_(const at::Scalar & other) const { + return at::_ops::div__Scalar::call(const_cast(*this), other); +} + +// aten::div.Scalar_mode(Tensor self, Scalar other, *, str? rounding_mode) -> Tensor +inline at::Tensor Tensor::div(const at::Scalar & other, ::std::optional rounding_mode) const { + return at::_ops::div_Scalar_mode::call(const_cast(*this), other, rounding_mode); +} + +// aten::div_.Scalar_mode(Tensor(a!) self, Scalar other, *, str? rounding_mode) -> Tensor(a!) +inline at::Tensor & Tensor::div_(const at::Scalar & other, ::std::optional rounding_mode) const { + return at::_ops::div__Scalar_mode::call(const_cast(*this), other, rounding_mode); +} + +// aten::divide.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::divide(const at::Tensor & other) const { + return at::_ops::divide_Tensor::call(const_cast(*this), other); +} + +// aten::divide_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::divide_(const at::Tensor & other) const { + return at::_ops::divide__Tensor::call(const_cast(*this), other); +} + +// aten::divide.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::divide(const at::Scalar & other) const { + return at::_ops::divide_Scalar::call(const_cast(*this), other); +} + +// aten::divide_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::divide_(const at::Scalar & other) const { + return at::_ops::divide__Scalar::call(const_cast(*this), other); +} + +// aten::divide.Tensor_mode(Tensor self, Tensor other, *, str? rounding_mode) -> Tensor +inline at::Tensor Tensor::divide(const at::Tensor & other, ::std::optional rounding_mode) const { + return at::_ops::divide_Tensor_mode::call(const_cast(*this), other, rounding_mode); +} + +// aten::divide_.Tensor_mode(Tensor(a!) self, Tensor other, *, str? rounding_mode) -> Tensor(a!) +inline at::Tensor & Tensor::divide_(const at::Tensor & other, ::std::optional rounding_mode) const { + return at::_ops::divide__Tensor_mode::call(const_cast(*this), other, rounding_mode); +} + +// aten::divide.Scalar_mode(Tensor self, Scalar other, *, str? rounding_mode) -> Tensor +inline at::Tensor Tensor::divide(const at::Scalar & other, ::std::optional rounding_mode) const { + return at::_ops::divide_Scalar_mode::call(const_cast(*this), other, rounding_mode); +} + +// aten::divide_.Scalar_mode(Tensor(a!) self, Scalar other, *, str? rounding_mode) -> Tensor(a!) +inline at::Tensor & Tensor::divide_(const at::Scalar & other, ::std::optional rounding_mode) const { + return at::_ops::divide__Scalar_mode::call(const_cast(*this), other, rounding_mode); +} + +// aten::true_divide.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::true_divide(const at::Tensor & other) const { + return at::_ops::true_divide_Tensor::call(const_cast(*this), other); +} + +// aten::true_divide_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::true_divide_(const at::Tensor & other) const { + return at::_ops::true_divide__Tensor::call(const_cast(*this), other); +} + +// aten::true_divide.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::true_divide(const at::Scalar & other) const { + return at::_ops::true_divide_Scalar::call(const_cast(*this), other); +} + +// aten::true_divide_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::true_divide_(const at::Scalar & other) const { + return at::_ops::true_divide__Scalar::call(const_cast(*this), other); +} + +// aten::dot(Tensor self, Tensor tensor) -> Tensor +inline at::Tensor Tensor::dot(const at::Tensor & tensor) const { + return at::_ops::dot::call(const_cast(*this), tensor); +} + +// aten::vdot(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::vdot(const at::Tensor & other) const { + return at::_ops::vdot::call(const_cast(*this), other); +} + +// aten::new_empty(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_empty(at::IntArrayRef size, at::TensorOptions options) const { + return at::_ops::new_empty::call(const_cast(*this), c10::fromIntArrayRefSlow(size), c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt()); +} + +// aten::new_empty(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_empty(at::IntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const { + return at::_ops::new_empty::call(const_cast(*this), c10::fromIntArrayRefSlow(size), dtype, layout, device, pin_memory); +} + +// aten::new_empty(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_empty_symint(c10::SymIntArrayRef size, at::TensorOptions options) const { + return at::_ops::new_empty::call(const_cast(*this), size, c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt()); +} + +// aten::new_empty(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_empty_symint(c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const { + return at::_ops::new_empty::call(const_cast(*this), size, dtype, layout, device, pin_memory); +} + +// aten::new_empty_strided(Tensor self, SymInt[] size, SymInt[] stride, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_empty_strided(at::IntArrayRef size, at::IntArrayRef stride, at::TensorOptions options) const { + return at::_ops::new_empty_strided::call(const_cast(*this), c10::fromIntArrayRefSlow(size), c10::fromIntArrayRefSlow(stride), c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt()); +} + +// aten::new_empty_strided(Tensor self, SymInt[] size, SymInt[] stride, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_empty_strided(at::IntArrayRef size, at::IntArrayRef stride, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const { + return at::_ops::new_empty_strided::call(const_cast(*this), c10::fromIntArrayRefSlow(size), c10::fromIntArrayRefSlow(stride), dtype, layout, device, pin_memory); +} + +// aten::new_empty_strided(Tensor self, SymInt[] size, SymInt[] stride, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_empty_strided_symint(c10::SymIntArrayRef size, c10::SymIntArrayRef stride, at::TensorOptions options) const { + return at::_ops::new_empty_strided::call(const_cast(*this), size, stride, c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt()); +} + +// aten::new_empty_strided(Tensor self, SymInt[] size, SymInt[] stride, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_empty_strided_symint(c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const { + return at::_ops::new_empty_strided::call(const_cast(*this), size, stride, dtype, layout, device, pin_memory); +} + +// aten::new_full(Tensor self, SymInt[] size, Scalar fill_value, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_full(at::IntArrayRef size, const at::Scalar & fill_value, at::TensorOptions options) const { + return at::_ops::new_full::call(const_cast(*this), c10::fromIntArrayRefSlow(size), fill_value, c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt()); +} + +// aten::new_full(Tensor self, SymInt[] size, Scalar fill_value, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_full(at::IntArrayRef size, const at::Scalar & fill_value, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const { + return at::_ops::new_full::call(const_cast(*this), c10::fromIntArrayRefSlow(size), fill_value, dtype, layout, device, pin_memory); +} + +// aten::new_full(Tensor self, SymInt[] size, Scalar fill_value, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_full_symint(c10::SymIntArrayRef size, const at::Scalar & fill_value, at::TensorOptions options) const { + return at::_ops::new_full::call(const_cast(*this), size, fill_value, c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt()); +} + +// aten::new_full(Tensor self, SymInt[] size, Scalar fill_value, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_full_symint(c10::SymIntArrayRef size, const at::Scalar & fill_value, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const { + return at::_ops::new_full::call(const_cast(*this), size, fill_value, dtype, layout, device, pin_memory); +} + +// aten::new_zeros(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_zeros(at::IntArrayRef size, at::TensorOptions options) const { + return at::_ops::new_zeros::call(const_cast(*this), c10::fromIntArrayRefSlow(size), c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt()); +} + +// aten::new_zeros(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_zeros(at::IntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const { + return at::_ops::new_zeros::call(const_cast(*this), c10::fromIntArrayRefSlow(size), dtype, layout, device, pin_memory); +} + +// aten::new_zeros(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_zeros_symint(c10::SymIntArrayRef size, at::TensorOptions options) const { + return at::_ops::new_zeros::call(const_cast(*this), size, c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt()); +} + +// aten::new_zeros(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_zeros_symint(c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const { + return at::_ops::new_zeros::call(const_cast(*this), size, dtype, layout, device, pin_memory); +} + +// aten::new_ones(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_ones(at::IntArrayRef size, at::TensorOptions options) const { + return at::_ops::new_ones::call(const_cast(*this), c10::fromIntArrayRefSlow(size), c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt()); +} + +// aten::new_ones(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_ones(at::IntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const { + return at::_ops::new_ones::call(const_cast(*this), c10::fromIntArrayRefSlow(size), dtype, layout, device, pin_memory); +} + +// aten::new_ones(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_ones_symint(c10::SymIntArrayRef size, at::TensorOptions options) const { + return at::_ops::new_ones::call(const_cast(*this), size, c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt()); +} + +// aten::new_ones(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor +inline at::Tensor Tensor::new_ones_symint(c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory) const { + return at::_ops::new_ones::call(const_cast(*this), size, dtype, layout, device, pin_memory); +} + +// aten::resize_(Tensor(a!) self, SymInt[] size, *, MemoryFormat? memory_format=None) -> Tensor(a!) +inline const at::Tensor & Tensor::resize_(at::IntArrayRef size, ::std::optional memory_format) const { + return at::_ops::resize_::call(const_cast(*this), c10::fromIntArrayRefSlow(size), memory_format); +} + +// aten::resize_(Tensor(a!) self, SymInt[] size, *, MemoryFormat? memory_format=None) -> Tensor(a!) +inline const at::Tensor & Tensor::resize__symint(c10::SymIntArrayRef size, ::std::optional memory_format) const { + return at::_ops::resize_::call(const_cast(*this), size, memory_format); +} + +// aten::erf(Tensor self) -> Tensor +inline at::Tensor Tensor::erf() const { + return at::_ops::erf::call(const_cast(*this)); +} + +// aten::erf_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::erf_() const { + return at::_ops::erf_::call(const_cast(*this)); +} + +// aten::erfc(Tensor self) -> Tensor +inline at::Tensor Tensor::erfc() const { + return at::_ops::erfc::call(const_cast(*this)); +} + +// aten::erfc_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::erfc_() const { + return at::_ops::erfc_::call(const_cast(*this)); +} + +// aten::exp(Tensor self) -> Tensor +inline at::Tensor Tensor::exp() const { + return at::_ops::exp::call(const_cast(*this)); +} + +// aten::exp_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::exp_() const { + return at::_ops::exp_::call(const_cast(*this)); +} + +// aten::exp2(Tensor self) -> Tensor +inline at::Tensor Tensor::exp2() const { + return at::_ops::exp2::call(const_cast(*this)); +} + +// aten::exp2_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::exp2_() const { + return at::_ops::exp2_::call(const_cast(*this)); +} + +// aten::expm1(Tensor self) -> Tensor +inline at::Tensor Tensor::expm1() const { + return at::_ops::expm1::call(const_cast(*this)); +} + +// aten::expm1_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::expm1_() const { + return at::_ops::expm1_::call(const_cast(*this)); +} + +// aten::expand(Tensor(a) self, SymInt[] size, *, bool implicit=False) -> Tensor(a) +inline at::Tensor Tensor::expand(at::IntArrayRef size, bool implicit) const { + return at::_ops::expand::call(const_cast(*this), c10::fromIntArrayRefSlow(size), implicit); +} + +// aten::expand(Tensor(a) self, SymInt[] size, *, bool implicit=False) -> Tensor(a) +inline at::Tensor Tensor::expand_symint(c10::SymIntArrayRef size, bool implicit) const { + return at::_ops::expand::call(const_cast(*this), size, implicit); +} + +// aten::expand_as(Tensor(a) self, Tensor other) -> Tensor(a) +inline at::Tensor Tensor::expand_as(const at::Tensor & other) const { + return at::_ops::expand_as::call(const_cast(*this), other); +} + +// aten::flatten.using_ints(Tensor(a) self, int start_dim=0, int end_dim=-1) -> Tensor(a) +inline at::Tensor Tensor::flatten(int64_t start_dim, int64_t end_dim) const { + return at::_ops::flatten_using_ints::call(const_cast(*this), start_dim, end_dim); +} + +// aten::flatten.named_out_dim(Tensor(a) self, int start_dim, int end_dim, Dimname out_dim) -> Tensor(a) +inline at::Tensor Tensor::flatten(int64_t start_dim, int64_t end_dim, at::Dimname out_dim) const { + return at::_ops::flatten_named_out_dim::call(const_cast(*this), start_dim, end_dim, out_dim); +} + +// aten::flatten.using_names(Tensor(a) self, Dimname start_dim, Dimname end_dim, Dimname out_dim) -> Tensor(a) +inline at::Tensor Tensor::flatten(at::Dimname start_dim, at::Dimname end_dim, at::Dimname out_dim) const { + return at::_ops::flatten_using_names::call(const_cast(*this), start_dim, end_dim, out_dim); +} + +// aten::flatten.DimnameList(Tensor(a) self, Dimname[] dims, Dimname out_dim) -> Tensor(a) +inline at::Tensor Tensor::flatten(at::DimnameList dims, at::Dimname out_dim) const { + return at::_ops::flatten_DimnameList::call(const_cast(*this), dims, out_dim); +} + +// aten::unflatten.int(Tensor(a) self, int dim, SymInt[] sizes) -> Tensor(a) +inline at::Tensor Tensor::unflatten(int64_t dim, at::IntArrayRef sizes) const { + return at::_ops::unflatten_int::call(const_cast(*this), dim, c10::fromIntArrayRefSlow(sizes)); +} + +// aten::unflatten.int(Tensor(a) self, int dim, SymInt[] sizes) -> Tensor(a) +inline at::Tensor Tensor::unflatten_symint(int64_t dim, c10::SymIntArrayRef sizes) const { + return at::_ops::unflatten_int::call(const_cast(*this), dim, sizes); +} + +// aten::unflatten.Dimname(Tensor(a) self, Dimname dim, SymInt[] sizes, Dimname[] names) -> Tensor(a) +inline at::Tensor Tensor::unflatten(at::Dimname dim, at::IntArrayRef sizes, at::DimnameList names) const { + return at::_ops::unflatten_Dimname::call(const_cast(*this), dim, c10::fromIntArrayRefSlow(sizes), names); +} + +// aten::unflatten.Dimname(Tensor(a) self, Dimname dim, SymInt[] sizes, Dimname[] names) -> Tensor(a) +inline at::Tensor Tensor::unflatten_symint(at::Dimname dim, c10::SymIntArrayRef sizes, at::DimnameList names) const { + return at::_ops::unflatten_Dimname::call(const_cast(*this), dim, sizes, names); +} + +// aten::fill_.Scalar(Tensor(a!) self, Scalar value) -> Tensor(a!) +inline at::Tensor & Tensor::fill_(const at::Scalar & value) const { + return at::_ops::fill__Scalar::call(const_cast(*this), value); +} + +// aten::fill_.Tensor(Tensor(a!) self, Tensor value) -> Tensor(a!) +inline at::Tensor & Tensor::fill_(const at::Tensor & value) const { + return at::_ops::fill__Tensor::call(const_cast(*this), value); +} + +// aten::floor(Tensor self) -> Tensor +inline at::Tensor Tensor::floor() const { + return at::_ops::floor::call(const_cast(*this)); +} + +// aten::floor_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::floor_() const { + return at::_ops::floor_::call(const_cast(*this)); +} + +// aten::floor_divide(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::floor_divide(const at::Tensor & other) const { + return at::_ops::floor_divide::call(const_cast(*this), other); +} + +// aten::floor_divide_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::floor_divide_(const at::Tensor & other) const { + return at::_ops::floor_divide__Tensor::call(const_cast(*this), other); +} + +// aten::floor_divide.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::floor_divide(const at::Scalar & other) const { + return at::_ops::floor_divide_Scalar::call(const_cast(*this), other); +} + +// aten::floor_divide_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::floor_divide_(const at::Scalar & other) const { + return at::_ops::floor_divide__Scalar::call(const_cast(*this), other); +} + +// aten::frac(Tensor self) -> Tensor +inline at::Tensor Tensor::frac() const { + return at::_ops::frac::call(const_cast(*this)); +} + +// aten::frac_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::frac_() const { + return at::_ops::frac_::call(const_cast(*this)); +} + +// aten::gcd(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::gcd(const at::Tensor & other) const { + return at::_ops::gcd::call(const_cast(*this), other); +} + +// aten::gcd_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::gcd_(const at::Tensor & other) const { + return at::_ops::gcd_::call(const_cast(*this), other); +} + +// aten::lcm(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::lcm(const at::Tensor & other) const { + return at::_ops::lcm::call(const_cast(*this), other); +} + +// aten::lcm_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::lcm_(const at::Tensor & other) const { + return at::_ops::lcm_::call(const_cast(*this), other); +} + +// aten::index.Tensor(Tensor self, Tensor?[] indices) -> Tensor +inline at::Tensor Tensor::index(const c10::List<::std::optional> & indices) const { + return at::_ops::index_Tensor::call(const_cast(*this), indices); +} + +// aten::index_copy_(Tensor(a!) self, int dim, Tensor index, Tensor source) -> Tensor(a!) +inline at::Tensor & Tensor::index_copy_(int64_t dim, const at::Tensor & index, const at::Tensor & source) const { + return at::_ops::index_copy_::call(const_cast(*this), dim, index, source); +} + +// aten::index_copy(Tensor self, int dim, Tensor index, Tensor source) -> Tensor +inline at::Tensor Tensor::index_copy(int64_t dim, const at::Tensor & index, const at::Tensor & source) const { + return at::_ops::index_copy::call(const_cast(*this), dim, index, source); +} + +// aten::index_copy_.dimname(Tensor(a!) self, Dimname dim, Tensor index, Tensor source) -> Tensor(a!) +inline at::Tensor & Tensor::index_copy_(at::Dimname dim, const at::Tensor & index, const at::Tensor & source) const { + return at::_ops::index_copy__dimname::call(const_cast(*this), dim, index, source); +} + +// aten::index_copy.dimname(Tensor self, Dimname dim, Tensor index, Tensor source) -> Tensor +inline at::Tensor Tensor::index_copy(at::Dimname dim, const at::Tensor & index, const at::Tensor & source) const { + return at::_ops::index_copy_dimname::call(const_cast(*this), dim, index, source); +} + +// aten::index_put_(Tensor(a!) self, Tensor?[] indices, Tensor values, bool accumulate=False) -> Tensor(a!) +inline at::Tensor & Tensor::index_put_(const c10::List<::std::optional> & indices, const at::Tensor & values, bool accumulate) const { + return at::_ops::index_put_::call(const_cast(*this), indices, values, accumulate); +} + +// aten::index_put(Tensor self, Tensor?[] indices, Tensor values, bool accumulate=False) -> Tensor +inline at::Tensor Tensor::index_put(const c10::List<::std::optional> & indices, const at::Tensor & values, bool accumulate) const { + return at::_ops::index_put::call(const_cast(*this), indices, values, accumulate); +} + +// aten::isclose(Tensor self, Tensor other, float rtol=1e-05, float atol=1e-08, bool equal_nan=False) -> Tensor +inline at::Tensor Tensor::isclose(const at::Tensor & other, double rtol, double atol, bool equal_nan) const { + return at::_ops::isclose::call(const_cast(*this), other, rtol, atol, equal_nan); +} + +// aten::isnan(Tensor self) -> Tensor +inline at::Tensor Tensor::isnan() const { + return at::_ops::isnan::call(const_cast(*this)); +} + +// aten::is_distributed(Tensor self) -> bool +inline bool Tensor::is_distributed() const { + return at::_ops::is_distributed::call(const_cast(*this)); +} + +// aten::is_floating_point(Tensor self) -> bool +inline bool Tensor::__dispatch_is_floating_point() const { + return at::_ops::is_floating_point::call(const_cast(*this)); +} + +// aten::is_complex(Tensor self) -> bool +inline bool Tensor::__dispatch_is_complex() const { + return at::_ops::is_complex::call(const_cast(*this)); +} + +// aten::is_conj(Tensor self) -> bool +inline bool Tensor::__dispatch_is_conj() const { + return at::_ops::is_conj::call(const_cast(*this)); +} + +// aten::_is_zerotensor(Tensor self) -> bool +inline bool Tensor::__dispatch__is_zerotensor() const { + return at::_ops::_is_zerotensor::call(const_cast(*this)); +} + +// aten::is_neg(Tensor self) -> bool +inline bool Tensor::__dispatch_is_neg() const { + return at::_ops::is_neg::call(const_cast(*this)); +} + +// aten::isreal(Tensor self) -> Tensor +inline at::Tensor Tensor::isreal() const { + return at::_ops::isreal::call(const_cast(*this)); +} + +// aten::is_nonzero(Tensor self) -> bool +inline bool Tensor::is_nonzero() const { + return at::_ops::is_nonzero::call(const_cast(*this)); +} + +// aten::is_same_size(Tensor self, Tensor other) -> bool +inline bool Tensor::is_same_size(const at::Tensor & other) const { + return at::_ops::is_same_size::call(const_cast(*this), other); +} + +// aten::is_signed(Tensor self) -> bool +inline bool Tensor::__dispatch_is_signed() const { + return at::_ops::is_signed::call(const_cast(*this)); +} + +// aten::is_inference(Tensor self) -> bool +inline bool Tensor::__dispatch_is_inference() const { + return at::_ops::is_inference::call(const_cast(*this)); +} + +// aten::kron(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::kron(const at::Tensor & other) const { + return at::_ops::kron::call(const_cast(*this), other); +} + +// aten::kthvalue(Tensor self, SymInt k, int dim=-1, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::kthvalue(int64_t k, int64_t dim, bool keepdim) const { + return at::_ops::kthvalue::call(const_cast(*this), k, dim, keepdim); +} + +// aten::kthvalue(Tensor self, SymInt k, int dim=-1, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::kthvalue_symint(c10::SymInt k, int64_t dim, bool keepdim) const { + return at::_ops::kthvalue::call(const_cast(*this), k, dim, keepdim); +} + +// aten::kthvalue.dimname(Tensor self, SymInt k, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::kthvalue(int64_t k, at::Dimname dim, bool keepdim) const { + return at::_ops::kthvalue_dimname::call(const_cast(*this), k, dim, keepdim); +} + +// aten::kthvalue.dimname(Tensor self, SymInt k, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::kthvalue_symint(c10::SymInt k, at::Dimname dim, bool keepdim) const { + return at::_ops::kthvalue_dimname::call(const_cast(*this), k, dim, keepdim); +} + +// aten::nan_to_num(Tensor self, float? nan=None, float? posinf=None, float? neginf=None) -> Tensor +inline at::Tensor Tensor::nan_to_num(::std::optional nan, ::std::optional posinf, ::std::optional neginf) const { + return at::_ops::nan_to_num::call(const_cast(*this), nan, posinf, neginf); +} + +// aten::nan_to_num_(Tensor(a!) self, float? nan=None, float? posinf=None, float? neginf=None) -> Tensor(a!) +inline at::Tensor & Tensor::nan_to_num_(::std::optional nan, ::std::optional posinf, ::std::optional neginf) const { + return at::_ops::nan_to_num_::call(const_cast(*this), nan, posinf, neginf); +} + +// aten::ldexp.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::ldexp(const at::Tensor & other) const { + return at::_ops::ldexp_Tensor::call(const_cast(*this), other); +} + +// aten::ldexp_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::ldexp_(const at::Tensor & other) const { + return at::_ops::ldexp_::call(const_cast(*this), other); +} + +// aten::log(Tensor self) -> Tensor +inline at::Tensor Tensor::log() const { + return at::_ops::log::call(const_cast(*this)); +} + +// aten::log_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::log_() const { + return at::_ops::log_::call(const_cast(*this)); +} + +// aten::log10(Tensor self) -> Tensor +inline at::Tensor Tensor::log10() const { + return at::_ops::log10::call(const_cast(*this)); +} + +// aten::log10_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::log10_() const { + return at::_ops::log10_::call(const_cast(*this)); +} + +// aten::log1p(Tensor self) -> Tensor +inline at::Tensor Tensor::log1p() const { + return at::_ops::log1p::call(const_cast(*this)); +} + +// aten::log1p_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::log1p_() const { + return at::_ops::log1p_::call(const_cast(*this)); +} + +// aten::log2(Tensor self) -> Tensor +inline at::Tensor Tensor::log2() const { + return at::_ops::log2::call(const_cast(*this)); +} + +// aten::log2_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::log2_() const { + return at::_ops::log2_::call(const_cast(*this)); +} + +// aten::logaddexp(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::logaddexp(const at::Tensor & other) const { + return at::_ops::logaddexp::call(const_cast(*this), other); +} + +// aten::logaddexp2(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::logaddexp2(const at::Tensor & other) const { + return at::_ops::logaddexp2::call(const_cast(*this), other); +} + +// aten::xlogy.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::xlogy(const at::Tensor & other) const { + return at::_ops::xlogy_Tensor::call(const_cast(*this), other); +} + +// aten::xlogy.Scalar_Other(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::xlogy(const at::Scalar & other) const { + return at::_ops::xlogy_Scalar_Other::call(const_cast(*this), other); +} + +// aten::xlogy_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::xlogy_(const at::Tensor & other) const { + return at::_ops::xlogy__Tensor::call(const_cast(*this), other); +} + +// aten::xlogy_.Scalar_Other(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::xlogy_(const at::Scalar & other) const { + return at::_ops::xlogy__Scalar_Other::call(const_cast(*this), other); +} + +// aten::log_softmax.int(Tensor self, int dim, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::log_softmax(int64_t dim, ::std::optional dtype) const { + return at::_ops::log_softmax_int::call(const_cast(*this), dim, dtype); +} + +// aten::log_softmax.Dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::log_softmax(at::Dimname dim, ::std::optional dtype) const { + return at::_ops::log_softmax_Dimname::call(const_cast(*this), dim, dtype); +} + +// aten::logcumsumexp(Tensor self, int dim) -> Tensor +inline at::Tensor Tensor::logcumsumexp(int64_t dim) const { + return at::_ops::logcumsumexp::call(const_cast(*this), dim); +} + +// aten::logcumsumexp.dimname(Tensor self, Dimname dim) -> Tensor +inline at::Tensor Tensor::logcumsumexp(at::Dimname dim) const { + return at::_ops::logcumsumexp_dimname::call(const_cast(*this), dim); +} + +// aten::logsumexp(Tensor self, int[1] dim, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::logsumexp(at::IntArrayRef dim, bool keepdim) const { + return at::_ops::logsumexp::call(const_cast(*this), dim, keepdim); +} + +// aten::logsumexp.names(Tensor self, Dimname[1] dim, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::logsumexp(at::DimnameList dim, bool keepdim) const { + return at::_ops::logsumexp_names::call(const_cast(*this), dim, keepdim); +} + +// aten::matmul(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::matmul(const at::Tensor & other) const { + return at::_ops::matmul::call(const_cast(*this), other); +} + +// aten::matrix_power(Tensor self, int n) -> Tensor +inline at::Tensor Tensor::matrix_power(int64_t n) const { + return at::_ops::matrix_power::call(const_cast(*this), n); +} + +// aten::matrix_exp(Tensor self) -> Tensor +inline at::Tensor Tensor::matrix_exp() const { + return at::_ops::matrix_exp::call(const_cast(*this)); +} + +// aten::aminmax(Tensor self, *, int? dim=None, bool keepdim=False) -> (Tensor min, Tensor max) +inline ::std::tuple Tensor::aminmax(::std::optional dim, bool keepdim) const { + return at::_ops::aminmax::call(const_cast(*this), dim, keepdim); +} + +// aten::max.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::max(int64_t dim, bool keepdim) const { + return at::_ops::max_dim::call(const_cast(*this), dim, keepdim); +} + +// aten::max.names_dim(Tensor self, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::max(at::Dimname dim, bool keepdim) const { + return at::_ops::max_names_dim::call(const_cast(*this), dim, keepdim); +} + +// aten::amax(Tensor self, int[1] dim=[], bool keepdim=False) -> Tensor +inline at::Tensor Tensor::amax(at::IntArrayRef dim, bool keepdim) const { + return at::_ops::amax::call(const_cast(*this), dim, keepdim); +} + +// aten::mean(Tensor self, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::mean(::std::optional dtype) const { + return at::_ops::mean::call(const_cast(*this), dtype); +} + +// aten::mean.dim(Tensor self, int[1]? dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::mean(at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype) const { + return at::_ops::mean_dim::call(const_cast(*this), dim, keepdim, dtype); +} + +// aten::mean.names_dim(Tensor self, Dimname[1] dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::mean(at::DimnameList dim, bool keepdim, ::std::optional dtype) const { + return at::_ops::mean_names_dim::call(const_cast(*this), dim, keepdim, dtype); +} + +// aten::nanmean(Tensor self, int[1]? dim=None, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::nanmean(at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype) const { + return at::_ops::nanmean::call(const_cast(*this), dim, keepdim, dtype); +} + +// aten::median(Tensor self) -> Tensor +inline at::Tensor Tensor::median() const { + return at::_ops::median::call(const_cast(*this)); +} + +// aten::median.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::median(int64_t dim, bool keepdim) const { + return at::_ops::median_dim::call(const_cast(*this), dim, keepdim); +} + +// aten::median.names_dim(Tensor self, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::median(at::Dimname dim, bool keepdim) const { + return at::_ops::median_names_dim::call(const_cast(*this), dim, keepdim); +} + +// aten::nanmedian(Tensor self) -> Tensor +inline at::Tensor Tensor::nanmedian() const { + return at::_ops::nanmedian::call(const_cast(*this)); +} + +// aten::nanmedian.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::nanmedian(int64_t dim, bool keepdim) const { + return at::_ops::nanmedian_dim::call(const_cast(*this), dim, keepdim); +} + +// aten::nanmedian.names_dim(Tensor self, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::nanmedian(at::Dimname dim, bool keepdim) const { + return at::_ops::nanmedian_names_dim::call(const_cast(*this), dim, keepdim); +} + +// aten::min.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::min(int64_t dim, bool keepdim) const { + return at::_ops::min_dim::call(const_cast(*this), dim, keepdim); +} + +// aten::min.names_dim(Tensor self, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::min(at::Dimname dim, bool keepdim) const { + return at::_ops::min_names_dim::call(const_cast(*this), dim, keepdim); +} + +// aten::amin(Tensor self, int[1] dim=[], bool keepdim=False) -> Tensor +inline at::Tensor Tensor::amin(at::IntArrayRef dim, bool keepdim) const { + return at::_ops::amin::call(const_cast(*this), dim, keepdim); +} + +// aten::mm(Tensor self, Tensor mat2) -> Tensor +inline at::Tensor Tensor::mm(const at::Tensor & mat2) const { + return at::_ops::mm::call(const_cast(*this), mat2); +} + +// aten::mode(Tensor self, int dim=-1, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::mode(int64_t dim, bool keepdim) const { + return at::_ops::mode::call(const_cast(*this), dim, keepdim); +} + +// aten::mode.dimname(Tensor self, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::mode(at::Dimname dim, bool keepdim) const { + return at::_ops::mode_dimname::call(const_cast(*this), dim, keepdim); +} + +// aten::mul.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::mul(const at::Tensor & other) const { + return at::_ops::mul_Tensor::call(const_cast(*this), other); +} + +// aten::mul_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::mul_(const at::Tensor & other) const { + return at::_ops::mul__Tensor::call(const_cast(*this), other); +} + +// aten::mul.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::mul(const at::Scalar & other) const { + return at::_ops::mul_Scalar::call(const_cast(*this), other); +} + +// aten::mul_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::mul_(const at::Scalar & other) const { + return at::_ops::mul__Scalar::call(const_cast(*this), other); +} + +// aten::multiply.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::multiply(const at::Tensor & other) const { + return at::_ops::multiply_Tensor::call(const_cast(*this), other); +} + +// aten::multiply_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::multiply_(const at::Tensor & other) const { + return at::_ops::multiply__Tensor::call(const_cast(*this), other); +} + +// aten::multiply.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::multiply(const at::Scalar & other) const { + return at::_ops::multiply_Scalar::call(const_cast(*this), other); +} + +// aten::multiply_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::multiply_(const at::Scalar & other) const { + return at::_ops::multiply__Scalar::call(const_cast(*this), other); +} + +// aten::mv(Tensor self, Tensor vec) -> Tensor +inline at::Tensor Tensor::mv(const at::Tensor & vec) const { + return at::_ops::mv::call(const_cast(*this), vec); +} + +// aten::mvlgamma(Tensor self, int p) -> Tensor +inline at::Tensor Tensor::mvlgamma(int64_t p) const { + return at::_ops::mvlgamma::call(const_cast(*this), p); +} + +// aten::mvlgamma_(Tensor(a!) self, int p) -> Tensor(a!) +inline at::Tensor & Tensor::mvlgamma_(int64_t p) const { + return at::_ops::mvlgamma_::call(const_cast(*this), p); +} + +// aten::narrow_copy(Tensor self, int dim, SymInt start, SymInt length) -> Tensor +inline at::Tensor Tensor::narrow_copy(int64_t dim, int64_t start, int64_t length) const { + return at::_ops::narrow_copy::call(const_cast(*this), dim, start, length); +} + +// aten::narrow_copy(Tensor self, int dim, SymInt start, SymInt length) -> Tensor +inline at::Tensor Tensor::narrow_copy_symint(int64_t dim, c10::SymInt start, c10::SymInt length) const { + return at::_ops::narrow_copy::call(const_cast(*this), dim, start, length); +} + +// aten::narrow(Tensor(a) self, int dim, SymInt start, SymInt length) -> Tensor(a) +inline at::Tensor Tensor::narrow(int64_t dim, int64_t start, int64_t length) const { + return at::_ops::narrow::call(const_cast(*this), dim, start, length); +} + +// aten::narrow(Tensor(a) self, int dim, SymInt start, SymInt length) -> Tensor(a) +inline at::Tensor Tensor::narrow_symint(int64_t dim, c10::SymInt start, c10::SymInt length) const { + return at::_ops::narrow::call(const_cast(*this), dim, start, length); +} + +// aten::narrow.Tensor(Tensor(a) self, int dim, Tensor start, SymInt length) -> Tensor(a) +inline at::Tensor Tensor::narrow(int64_t dim, const at::Tensor & start, int64_t length) const { + return at::_ops::narrow_Tensor::call(const_cast(*this), dim, start, length); +} + +// aten::narrow.Tensor(Tensor(a) self, int dim, Tensor start, SymInt length) -> Tensor(a) +inline at::Tensor Tensor::narrow_symint(int64_t dim, const at::Tensor & start, c10::SymInt length) const { + return at::_ops::narrow_Tensor::call(const_cast(*this), dim, start, length); +} + +// aten::permute(Tensor(a) self, int[] dims) -> Tensor(a) +inline at::Tensor Tensor::permute(at::IntArrayRef dims) const { + return at::_ops::permute::call(const_cast(*this), dims); +} + +// aten::movedim.intlist(Tensor(a) self, int[] source, int[] destination) -> Tensor(a) +inline at::Tensor Tensor::movedim(at::IntArrayRef source, at::IntArrayRef destination) const { + return at::_ops::movedim_intlist::call(const_cast(*this), source, destination); +} + +// aten::movedim.int(Tensor(a) self, int source, int destination) -> Tensor(a) +inline at::Tensor Tensor::movedim(int64_t source, int64_t destination) const { + return at::_ops::movedim_int::call(const_cast(*this), source, destination); +} + +// aten::moveaxis.intlist(Tensor(a) self, int[] source, int[] destination) -> Tensor(a) +inline at::Tensor Tensor::moveaxis(at::IntArrayRef source, at::IntArrayRef destination) const { + return at::_ops::moveaxis_intlist::call(const_cast(*this), source, destination); +} + +// aten::moveaxis.int(Tensor(a) self, int source, int destination) -> Tensor(a) +inline at::Tensor Tensor::moveaxis(int64_t source, int64_t destination) const { + return at::_ops::moveaxis_int::call(const_cast(*this), source, destination); +} + +// aten::numpy_T(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::numpy_T() const { + return at::_ops::numpy_T::call(const_cast(*this)); +} + +// aten::matrix_H(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::matrix_H() const { + return at::_ops::matrix_H::call(const_cast(*this)); +} + +// aten::mT(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::mT() const { + return at::_ops::mT::call(const_cast(*this)); +} + +// aten::mH(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::mH() const { + return at::_ops::mH::call(const_cast(*this)); +} + +// aten::adjoint(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::adjoint() const { + return at::_ops::adjoint::call(const_cast(*this)); +} + +// aten::is_pinned(Tensor self, Device? device=None) -> bool +inline bool Tensor::is_pinned(::std::optional device) const { + return at::_ops::is_pinned::call(const_cast(*this), device); +} + +// aten::pin_memory(Tensor(a) self, Device? device=None) -> Tensor(a) +inline at::Tensor Tensor::pin_memory(::std::optional device) const { + return at::_ops::pin_memory::call(const_cast(*this), device); +} + +// aten::pinverse(Tensor self, float rcond=1e-15) -> Tensor +inline at::Tensor Tensor::pinverse(double rcond) const { + return at::_ops::pinverse::call(const_cast(*this), rcond); +} + +// aten::rad2deg(Tensor self) -> Tensor +inline at::Tensor Tensor::rad2deg() const { + return at::_ops::rad2deg::call(const_cast(*this)); +} + +// aten::rad2deg_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::rad2deg_() const { + return at::_ops::rad2deg_::call(const_cast(*this)); +} + +// aten::deg2rad(Tensor self) -> Tensor +inline at::Tensor Tensor::deg2rad() const { + return at::_ops::deg2rad::call(const_cast(*this)); +} + +// aten::deg2rad_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::deg2rad_() const { + return at::_ops::deg2rad_::call(const_cast(*this)); +} + +// aten::ravel(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::ravel() const { + return at::_ops::ravel::call(const_cast(*this)); +} + +// aten::reciprocal(Tensor self) -> Tensor +inline at::Tensor Tensor::reciprocal() const { + return at::_ops::reciprocal::call(const_cast(*this)); +} + +// aten::reciprocal_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::reciprocal_() const { + return at::_ops::reciprocal_::call(const_cast(*this)); +} + +// aten::neg(Tensor self) -> Tensor +inline at::Tensor Tensor::neg() const { + return at::_ops::neg::call(const_cast(*this)); +} + +// aten::neg_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::neg_() const { + return at::_ops::neg_::call(const_cast(*this)); +} + +// aten::negative(Tensor self) -> Tensor +inline at::Tensor Tensor::negative() const { + return at::_ops::negative::call(const_cast(*this)); +} + +// aten::negative_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::negative_() const { + return at::_ops::negative_::call(const_cast(*this)); +} + +// aten::repeat(Tensor self, SymInt[] repeats) -> Tensor +inline at::Tensor Tensor::repeat(at::IntArrayRef repeats) const { + return at::_ops::repeat::call(const_cast(*this), c10::fromIntArrayRefSlow(repeats)); +} + +// aten::repeat(Tensor self, SymInt[] repeats) -> Tensor +inline at::Tensor Tensor::repeat_symint(c10::SymIntArrayRef repeats) const { + return at::_ops::repeat::call(const_cast(*this), repeats); +} + +// aten::repeat_interleave.self_Tensor(Tensor self, Tensor repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor +inline at::Tensor Tensor::repeat_interleave(const at::Tensor & repeats, ::std::optional dim, ::std::optional output_size) const { + return at::_ops::repeat_interleave_self_Tensor::call(const_cast(*this), repeats, dim, output_size.has_value() ? ::std::make_optional(c10::SymInt(*output_size)) : ::std::nullopt); +} + +// aten::repeat_interleave.self_Tensor(Tensor self, Tensor repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor +inline at::Tensor Tensor::repeat_interleave_symint(const at::Tensor & repeats, ::std::optional dim, ::std::optional output_size) const { + return at::_ops::repeat_interleave_self_Tensor::call(const_cast(*this), repeats, dim, output_size); +} + +// aten::repeat_interleave.self_int(Tensor self, SymInt repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor +inline at::Tensor Tensor::repeat_interleave(int64_t repeats, ::std::optional dim, ::std::optional output_size) const { + return at::_ops::repeat_interleave_self_int::call(const_cast(*this), repeats, dim, output_size.has_value() ? ::std::make_optional(c10::SymInt(*output_size)) : ::std::nullopt); +} + +// aten::repeat_interleave.self_int(Tensor self, SymInt repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor +inline at::Tensor Tensor::repeat_interleave_symint(c10::SymInt repeats, ::std::optional dim, ::std::optional output_size) const { + return at::_ops::repeat_interleave_self_int::call(const_cast(*this), repeats, dim, output_size); +} + +// aten::reshape(Tensor(a) self, SymInt[] shape) -> Tensor(a) +inline at::Tensor Tensor::reshape(at::IntArrayRef shape) const { + return at::_ops::reshape::call(const_cast(*this), c10::fromIntArrayRefSlow(shape)); +} + +// aten::reshape(Tensor(a) self, SymInt[] shape) -> Tensor(a) +inline at::Tensor Tensor::reshape_symint(c10::SymIntArrayRef shape) const { + return at::_ops::reshape::call(const_cast(*this), shape); +} + +// aten::_reshape_alias(Tensor(a) self, SymInt[] size, SymInt[] stride) -> Tensor(a) +inline at::Tensor Tensor::_reshape_alias(at::IntArrayRef size, at::IntArrayRef stride) const { + return at::_ops::_reshape_alias::call(const_cast(*this), c10::fromIntArrayRefSlow(size), c10::fromIntArrayRefSlow(stride)); +} + +// aten::_reshape_alias(Tensor(a) self, SymInt[] size, SymInt[] stride) -> Tensor(a) +inline at::Tensor Tensor::_reshape_alias_symint(c10::SymIntArrayRef size, c10::SymIntArrayRef stride) const { + return at::_ops::_reshape_alias::call(const_cast(*this), size, stride); +} + +// aten::reshape_as(Tensor(a) self, Tensor other) -> Tensor(a) +inline at::Tensor Tensor::reshape_as(const at::Tensor & other) const { + return at::_ops::reshape_as::call(const_cast(*this), other); +} + +// aten::round(Tensor self) -> Tensor +inline at::Tensor Tensor::round() const { + return at::_ops::round::call(const_cast(*this)); +} + +// aten::round_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::round_() const { + return at::_ops::round_::call(const_cast(*this)); +} + +// aten::round.decimals(Tensor self, *, int decimals) -> Tensor +inline at::Tensor Tensor::round(int64_t decimals) const { + return at::_ops::round_decimals::call(const_cast(*this), decimals); +} + +// aten::round_.decimals(Tensor(a!) self, *, int decimals) -> Tensor(a!) +inline at::Tensor & Tensor::round_(int64_t decimals) const { + return at::_ops::round__decimals::call(const_cast(*this), decimals); +} + +// aten::relu(Tensor self) -> Tensor +inline at::Tensor Tensor::relu() const { + return at::_ops::relu::call(const_cast(*this)); +} + +// aten::relu_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::relu_() const { + return at::_ops::relu_::call(const_cast(*this)); +} + +// aten::prelu(Tensor self, Tensor weight) -> Tensor +inline at::Tensor Tensor::prelu(const at::Tensor & weight) const { + return at::_ops::prelu::call(const_cast(*this), weight); +} + +// aten::hardshrink(Tensor self, Scalar lambd=0.5) -> Tensor +inline at::Tensor Tensor::hardshrink(const at::Scalar & lambd) const { + return at::_ops::hardshrink::call(const_cast(*this), lambd); +} + +// aten::hardshrink_backward(Tensor grad_out, Tensor self, Scalar lambd) -> Tensor +inline at::Tensor Tensor::hardshrink_backward(const at::Tensor & grad_out, const at::Scalar & lambd) const { + return at::_ops::hardshrink_backward::call(grad_out, const_cast(*this), lambd); +} + +// aten::rsqrt(Tensor self) -> Tensor +inline at::Tensor Tensor::rsqrt() const { + return at::_ops::rsqrt::call(const_cast(*this)); +} + +// aten::rsqrt_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::rsqrt_() const { + return at::_ops::rsqrt_::call(const_cast(*this)); +} + +// aten::select.Dimname(Tensor(a) self, Dimname dim, int index) -> Tensor(a) +inline at::Tensor Tensor::select(at::Dimname dim, int64_t index) const { + return at::_ops::select_Dimname::call(const_cast(*this), dim, index); +} + +// aten::select.int(Tensor(a) self, int dim, SymInt index) -> Tensor(a) +inline at::Tensor Tensor::select(int64_t dim, int64_t index) const { + return at::_ops::select_int::call(const_cast(*this), dim, index); +} + +// aten::select.int(Tensor(a) self, int dim, SymInt index) -> Tensor(a) +inline at::Tensor Tensor::select_symint(int64_t dim, c10::SymInt index) const { + return at::_ops::select_int::call(const_cast(*this), dim, index); +} + +// aten::sigmoid(Tensor self) -> Tensor +inline at::Tensor Tensor::sigmoid() const { + return at::_ops::sigmoid::call(const_cast(*this)); +} + +// aten::sigmoid_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::sigmoid_() const { + return at::_ops::sigmoid_::call(const_cast(*this)); +} + +// aten::logit(Tensor self, float? eps=None) -> Tensor +inline at::Tensor Tensor::logit(::std::optional eps) const { + return at::_ops::logit::call(const_cast(*this), eps); +} + +// aten::logit_(Tensor(a!) self, float? eps=None) -> Tensor(a!) +inline at::Tensor & Tensor::logit_(::std::optional eps) const { + return at::_ops::logit_::call(const_cast(*this), eps); +} + +// aten::sin(Tensor self) -> Tensor +inline at::Tensor Tensor::sin() const { + return at::_ops::sin::call(const_cast(*this)); +} + +// aten::sin_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::sin_() const { + return at::_ops::sin_::call(const_cast(*this)); +} + +// aten::sinc(Tensor self) -> Tensor +inline at::Tensor Tensor::sinc() const { + return at::_ops::sinc::call(const_cast(*this)); +} + +// aten::sinc_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::sinc_() const { + return at::_ops::sinc_::call(const_cast(*this)); +} + +// aten::sinh(Tensor self) -> Tensor +inline at::Tensor Tensor::sinh() const { + return at::_ops::sinh::call(const_cast(*this)); +} + +// aten::sinh_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::sinh_() const { + return at::_ops::sinh_::call(const_cast(*this)); +} + +// aten::detach(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::detach() const { + return at::_ops::detach::call(const_cast(*this)); +} + +// aten::detach_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::detach_() const { + return at::_ops::detach_::call(const_cast(*this)); +} + +// aten::size.Dimname(Tensor self, Dimname dim) -> int +inline int64_t Tensor::size(at::Dimname dim) const { + return at::_ops::size_Dimname::call(const_cast(*this), dim); +} + +// aten::slice.Tensor(Tensor(a) self, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a) +inline at::Tensor Tensor::slice(int64_t dim, ::std::optional start, ::std::optional end, int64_t step) const { + return at::_ops::slice_Tensor::call(const_cast(*this), dim, start.has_value() ? ::std::make_optional(c10::SymInt(*start)) : ::std::nullopt, end.has_value() ? ::std::make_optional(c10::SymInt(*end)) : ::std::nullopt, step); +} + +// aten::slice.Tensor(Tensor(a) self, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a) +inline at::Tensor Tensor::slice_symint(int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step) const { + return at::_ops::slice_Tensor::call(const_cast(*this), dim, start, end, step); +} + +// aten::slice_inverse(Tensor(a) self, Tensor src, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a) +inline at::Tensor Tensor::slice_inverse(const at::Tensor & src, int64_t dim, ::std::optional start, ::std::optional end, int64_t step) const { + return at::_ops::slice_inverse::call(const_cast(*this), src, dim, start.has_value() ? ::std::make_optional(c10::SymInt(*start)) : ::std::nullopt, end.has_value() ? ::std::make_optional(c10::SymInt(*end)) : ::std::nullopt, step); +} + +// aten::slice_inverse(Tensor(a) self, Tensor src, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a) +inline at::Tensor Tensor::slice_inverse_symint(const at::Tensor & src, int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step) const { + return at::_ops::slice_inverse::call(const_cast(*this), src, dim, start, end, step); +} + +// aten::slice_scatter(Tensor self, Tensor src, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor +inline at::Tensor Tensor::slice_scatter(const at::Tensor & src, int64_t dim, ::std::optional start, ::std::optional end, int64_t step) const { + return at::_ops::slice_scatter::call(const_cast(*this), src, dim, start.has_value() ? ::std::make_optional(c10::SymInt(*start)) : ::std::nullopt, end.has_value() ? ::std::make_optional(c10::SymInt(*end)) : ::std::nullopt, step); +} + +// aten::slice_scatter(Tensor self, Tensor src, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor +inline at::Tensor Tensor::slice_scatter_symint(const at::Tensor & src, int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step) const { + return at::_ops::slice_scatter::call(const_cast(*this), src, dim, start, end, step); +} + +// aten::select_scatter(Tensor self, Tensor src, int dim, SymInt index) -> Tensor +inline at::Tensor Tensor::select_scatter(const at::Tensor & src, int64_t dim, int64_t index) const { + return at::_ops::select_scatter::call(const_cast(*this), src, dim, index); +} + +// aten::select_scatter(Tensor self, Tensor src, int dim, SymInt index) -> Tensor +inline at::Tensor Tensor::select_scatter_symint(const at::Tensor & src, int64_t dim, c10::SymInt index) const { + return at::_ops::select_scatter::call(const_cast(*this), src, dim, index); +} + +// aten::diagonal_scatter(Tensor self, Tensor src, int offset=0, int dim1=0, int dim2=1) -> Tensor +inline at::Tensor Tensor::diagonal_scatter(const at::Tensor & src, int64_t offset, int64_t dim1, int64_t dim2) const { + return at::_ops::diagonal_scatter::call(const_cast(*this), src, offset, dim1, dim2); +} + +// aten::as_strided_scatter(Tensor self, Tensor src, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None) -> Tensor +inline at::Tensor Tensor::as_strided_scatter(const at::Tensor & src, at::IntArrayRef size, at::IntArrayRef stride, ::std::optional storage_offset) const { + return at::_ops::as_strided_scatter::call(const_cast(*this), src, c10::fromIntArrayRefSlow(size), c10::fromIntArrayRefSlow(stride), storage_offset.has_value() ? ::std::make_optional(c10::SymInt(*storage_offset)) : ::std::nullopt); +} + +// aten::as_strided_scatter(Tensor self, Tensor src, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None) -> Tensor +inline at::Tensor Tensor::as_strided_scatter_symint(const at::Tensor & src, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset) const { + return at::_ops::as_strided_scatter::call(const_cast(*this), src, size, stride, storage_offset); +} + +// aten::smm(Tensor self, Tensor mat2) -> Tensor +inline at::Tensor Tensor::smm(const at::Tensor & mat2) const { + return at::_ops::smm::call(const_cast(*this), mat2); +} + +// aten::softmax.int(Tensor self, int dim, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::softmax(int64_t dim, ::std::optional dtype) const { + return at::_ops::softmax_int::call(const_cast(*this), dim, dtype); +} + +// aten::softmax.Dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::softmax(at::Dimname dim, ::std::optional dtype) const { + return at::_ops::softmax_Dimname::call(const_cast(*this), dim, dtype); +} + +// aten::unsafe_split.Tensor(Tensor self, SymInt split_size, int dim=0) -> Tensor[] +inline ::std::vector Tensor::unsafe_split(int64_t split_size, int64_t dim) const { + return at::_ops::unsafe_split_Tensor::call(const_cast(*this), split_size, dim); +} + +// aten::unsafe_split.Tensor(Tensor self, SymInt split_size, int dim=0) -> Tensor[] +inline ::std::vector Tensor::unsafe_split_symint(c10::SymInt split_size, int64_t dim) const { + return at::_ops::unsafe_split_Tensor::call(const_cast(*this), split_size, dim); +} + +// aten::split.Tensor(Tensor(a -> *) self, SymInt split_size, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::split(int64_t split_size, int64_t dim) const { + return at::_ops::split_Tensor::call(const_cast(*this), split_size, dim); +} + +// aten::split.Tensor(Tensor(a -> *) self, SymInt split_size, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::split_symint(c10::SymInt split_size, int64_t dim) const { + return at::_ops::split_Tensor::call(const_cast(*this), split_size, dim); +} + +// aten::split.sizes(Tensor(a -> *) self, SymInt[] split_size, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::split(at::IntArrayRef split_size, int64_t dim) const { + return at::_ops::split_sizes::call(const_cast(*this), c10::fromIntArrayRefSlow(split_size), dim); +} + +// aten::split.sizes(Tensor(a -> *) self, SymInt[] split_size, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::split_symint(c10::SymIntArrayRef split_size, int64_t dim) const { + return at::_ops::split_sizes::call(const_cast(*this), split_size, dim); +} + +// aten::unsafe_split_with_sizes(Tensor self, SymInt[] split_sizes, int dim=0) -> Tensor[] +inline ::std::vector Tensor::unsafe_split_with_sizes(at::IntArrayRef split_sizes, int64_t dim) const { + return at::_ops::unsafe_split_with_sizes::call(const_cast(*this), c10::fromIntArrayRefSlow(split_sizes), dim); +} + +// aten::unsafe_split_with_sizes(Tensor self, SymInt[] split_sizes, int dim=0) -> Tensor[] +inline ::std::vector Tensor::unsafe_split_with_sizes_symint(c10::SymIntArrayRef split_sizes, int64_t dim) const { + return at::_ops::unsafe_split_with_sizes::call(const_cast(*this), split_sizes, dim); +} + +// aten::split_with_sizes(Tensor(a -> *) self, SymInt[] split_sizes, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::split_with_sizes(at::IntArrayRef split_sizes, int64_t dim) const { + return at::_ops::split_with_sizes::call(const_cast(*this), c10::fromIntArrayRefSlow(split_sizes), dim); +} + +// aten::split_with_sizes(Tensor(a -> *) self, SymInt[] split_sizes, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::split_with_sizes_symint(c10::SymIntArrayRef split_sizes, int64_t dim) const { + return at::_ops::split_with_sizes::call(const_cast(*this), split_sizes, dim); +} + +// aten::hsplit.int(Tensor(a -> *) self, int sections) -> Tensor(a)[] +inline ::std::vector Tensor::hsplit(int64_t sections) const { + return at::_ops::hsplit_int::call(const_cast(*this), sections); +} + +// aten::hsplit.array(Tensor(a -> *) self, int[] indices) -> Tensor(a)[] +inline ::std::vector Tensor::hsplit(at::IntArrayRef indices) const { + return at::_ops::hsplit_array::call(const_cast(*this), indices); +} + +// aten::vsplit.int(Tensor(a -> *) self, int sections) -> Tensor(a)[] +inline ::std::vector Tensor::vsplit(int64_t sections) const { + return at::_ops::vsplit_int::call(const_cast(*this), sections); +} + +// aten::vsplit.array(Tensor(a -> *) self, int[] indices) -> Tensor(a)[] +inline ::std::vector Tensor::vsplit(at::IntArrayRef indices) const { + return at::_ops::vsplit_array::call(const_cast(*this), indices); +} + +// aten::dsplit.int(Tensor(a -> *) self, int sections) -> Tensor(a)[] +inline ::std::vector Tensor::dsplit(int64_t sections) const { + return at::_ops::dsplit_int::call(const_cast(*this), sections); +} + +// aten::dsplit.array(Tensor(a -> *) self, int[] indices) -> Tensor(a)[] +inline ::std::vector Tensor::dsplit(at::IntArrayRef indices) const { + return at::_ops::dsplit_array::call(const_cast(*this), indices); +} + +// aten::squeeze(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::squeeze() const { + return at::_ops::squeeze::call(const_cast(*this)); +} + +// aten::squeeze.dim(Tensor(a) self, int dim) -> Tensor(a) +inline at::Tensor Tensor::squeeze(int64_t dim) const { + return at::_ops::squeeze_dim::call(const_cast(*this), dim); +} + +// aten::squeeze.dimname(Tensor(a) self, Dimname dim) -> Tensor(a) +inline at::Tensor Tensor::squeeze(at::Dimname dim) const { + return at::_ops::squeeze_dimname::call(const_cast(*this), dim); +} + +// aten::squeeze.dims(Tensor(a) self, int[] dim) -> Tensor(a) +inline at::Tensor Tensor::squeeze(at::IntArrayRef dim) const { + return at::_ops::squeeze_dims::call(const_cast(*this), dim); +} + +// aten::squeeze_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::squeeze_() const { + return at::_ops::squeeze_::call(const_cast(*this)); +} + +// aten::squeeze_.dim(Tensor(a!) self, int dim) -> Tensor(a!) +inline at::Tensor & Tensor::squeeze_(int64_t dim) const { + return at::_ops::squeeze__dim::call(const_cast(*this), dim); +} + +// aten::squeeze_.dims(Tensor(a!) self, int[] dim) -> Tensor(a!) +inline at::Tensor & Tensor::squeeze_(at::IntArrayRef dim) const { + return at::_ops::squeeze__dims::call(const_cast(*this), dim); +} + +// aten::squeeze_.dimname(Tensor(a!) self, Dimname dim) -> Tensor(a!) +inline at::Tensor & Tensor::squeeze_(at::Dimname dim) const { + return at::_ops::squeeze__dimname::call(const_cast(*this), dim); +} + +// aten::sspaddmm(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::sspaddmm(const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::sspaddmm::call(const_cast(*this), mat1, mat2, beta, alpha); +} + +// aten::stft(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool normalized=False, bool? onesided=None, bool? return_complex=None, bool? align_to_window=None) -> Tensor +inline at::Tensor Tensor::stft(int64_t n_fft, ::std::optional hop_length, ::std::optional win_length, const ::std::optional & window, bool normalized, ::std::optional onesided, ::std::optional return_complex, ::std::optional align_to_window) const { + return at::_ops::stft::call(const_cast(*this), n_fft, hop_length, win_length, window, normalized, onesided, return_complex, align_to_window); +} + +// aten::stft.center(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool center=True, str pad_mode="reflect", bool normalized=False, bool? onesided=None, bool? return_complex=None, bool? align_to_window=None) -> Tensor +inline at::Tensor Tensor::stft(int64_t n_fft, ::std::optional hop_length, ::std::optional win_length, const ::std::optional & window, bool center, c10::string_view pad_mode, bool normalized, ::std::optional onesided, ::std::optional return_complex, ::std::optional align_to_window) const { + return at::_ops::stft_center::call(const_cast(*this), n_fft, hop_length, win_length, window, center, pad_mode, normalized, onesided, return_complex, align_to_window); +} + +// aten::istft(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool center=True, bool normalized=False, bool? onesided=None, int? length=None, bool return_complex=False) -> Tensor +inline at::Tensor Tensor::istft(int64_t n_fft, ::std::optional hop_length, ::std::optional win_length, const ::std::optional & window, bool center, bool normalized, ::std::optional onesided, ::std::optional length, bool return_complex) const { + return at::_ops::istft::call(const_cast(*this), n_fft, hop_length, win_length, window, center, normalized, onesided, length, return_complex); +} + +// aten::stride.Dimname(Tensor self, Dimname dim) -> int +inline int64_t Tensor::stride(at::Dimname dim) const { + return at::_ops::stride_Dimname::call(const_cast(*this), dim); +} + +// aten::sum(Tensor self, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::sum(::std::optional dtype) const { + return at::_ops::sum::call(const_cast(*this), dtype); +} + +// aten::sum.dim_IntList(Tensor self, int[1]? dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::sum(at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype) const { + return at::_ops::sum_dim_IntList::call(const_cast(*this), dim, keepdim, dtype); +} + +// aten::sum.dim_DimnameList(Tensor self, Dimname[1] dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::sum(at::DimnameList dim, bool keepdim, ::std::optional dtype) const { + return at::_ops::sum_dim_DimnameList::call(const_cast(*this), dim, keepdim, dtype); +} + +// aten::nansum(Tensor self, int[1]? dim=None, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::nansum(at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype) const { + return at::_ops::nansum::call(const_cast(*this), dim, keepdim, dtype); +} + +// aten::hash_tensor(Tensor self, int[1] dim=[], *, bool keepdim=False, int mode=0) -> Tensor +inline at::Tensor Tensor::hash_tensor(at::IntArrayRef dim, bool keepdim, int64_t mode) const { + return at::_ops::hash_tensor::call(const_cast(*this), dim, keepdim, mode); +} + +// aten::sum_to_size(Tensor self, SymInt[] size) -> Tensor +inline at::Tensor Tensor::sum_to_size(at::IntArrayRef size) const { + return at::_ops::sum_to_size::call(const_cast(*this), c10::fromIntArrayRefSlow(size)); +} + +// aten::sum_to_size(Tensor self, SymInt[] size) -> Tensor +inline at::Tensor Tensor::sum_to_size_symint(c10::SymIntArrayRef size) const { + return at::_ops::sum_to_size::call(const_cast(*this), size); +} + +// aten::sqrt(Tensor self) -> Tensor +inline at::Tensor Tensor::sqrt() const { + return at::_ops::sqrt::call(const_cast(*this)); +} + +// aten::sqrt_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::sqrt_() const { + return at::_ops::sqrt_::call(const_cast(*this)); +} + +// aten::square(Tensor self) -> Tensor +inline at::Tensor Tensor::square() const { + return at::_ops::square::call(const_cast(*this)); +} + +// aten::square_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::square_() const { + return at::_ops::square_::call(const_cast(*this)); +} + +// aten::std(Tensor self, bool unbiased=True) -> Tensor +inline at::Tensor Tensor::std(bool unbiased) const { + return at::_ops::std::call(const_cast(*this), unbiased); +} + +// aten::std.dim(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::std(at::OptionalIntArrayRef dim, bool unbiased, bool keepdim) const { + return at::_ops::std_dim::call(const_cast(*this), dim, unbiased, keepdim); +} + +// aten::std.correction(Tensor self, int[1]? dim=None, *, Scalar? correction=None, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::std(at::OptionalIntArrayRef dim, const ::std::optional & correction, bool keepdim) const { + return at::_ops::std_correction::call(const_cast(*this), dim, correction, keepdim); +} + +// aten::std.names_dim(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::std(at::DimnameList dim, bool unbiased, bool keepdim) const { + return at::_ops::std_names_dim::call(const_cast(*this), dim, unbiased, keepdim); +} + +// aten::std.correction_names(Tensor self, Dimname[1] dim, *, Scalar? correction=None, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::std(at::DimnameList dim, const ::std::optional & correction, bool keepdim) const { + return at::_ops::std_correction_names::call(const_cast(*this), dim, correction, keepdim); +} + +// aten::prod(Tensor self, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::prod(::std::optional dtype) const { + return at::_ops::prod::call(const_cast(*this), dtype); +} + +// aten::prod.dim_int(Tensor self, int dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::prod(int64_t dim, bool keepdim, ::std::optional dtype) const { + return at::_ops::prod_dim_int::call(const_cast(*this), dim, keepdim, dtype); +} + +// aten::prod.dim_Dimname(Tensor self, Dimname dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::prod(at::Dimname dim, bool keepdim, ::std::optional dtype) const { + return at::_ops::prod_dim_Dimname::call(const_cast(*this), dim, keepdim, dtype); +} + +// aten::t(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::t() const { + return at::_ops::t::call(const_cast(*this)); +} + +// aten::t_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::t_() const { + return at::_ops::t_::call(const_cast(*this)); +} + +// aten::tan(Tensor self) -> Tensor +inline at::Tensor Tensor::tan() const { + return at::_ops::tan::call(const_cast(*this)); +} + +// aten::tan_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::tan_() const { + return at::_ops::tan_::call(const_cast(*this)); +} + +// aten::tanh(Tensor self) -> Tensor +inline at::Tensor Tensor::tanh() const { + return at::_ops::tanh::call(const_cast(*this)); +} + +// aten::tanh_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::tanh_() const { + return at::_ops::tanh_::call(const_cast(*this)); +} + +// aten::tile(Tensor self, SymInt[] dims) -> Tensor +inline at::Tensor Tensor::tile(at::IntArrayRef dims) const { + return at::_ops::tile::call(const_cast(*this), c10::fromIntArrayRefSlow(dims)); +} + +// aten::tile(Tensor self, SymInt[] dims) -> Tensor +inline at::Tensor Tensor::tile_symint(c10::SymIntArrayRef dims) const { + return at::_ops::tile::call(const_cast(*this), dims); +} + +// aten::transpose.int(Tensor(a) self, int dim0, int dim1) -> Tensor(a) +inline at::Tensor Tensor::transpose(int64_t dim0, int64_t dim1) const { + return at::_ops::transpose_int::call(const_cast(*this), dim0, dim1); +} + +// aten::transpose.Dimname(Tensor(a) self, Dimname dim0, Dimname dim1) -> Tensor(a) +inline at::Tensor Tensor::transpose(at::Dimname dim0, at::Dimname dim1) const { + return at::_ops::transpose_Dimname::call(const_cast(*this), dim0, dim1); +} + +// aten::transpose_(Tensor(a!) self, int dim0, int dim1) -> Tensor(a!) +inline at::Tensor & Tensor::transpose_(int64_t dim0, int64_t dim1) const { + return at::_ops::transpose_::call(const_cast(*this), dim0, dim1); +} + +// aten::flip(Tensor self, int[] dims) -> Tensor +inline at::Tensor Tensor::flip(at::IntArrayRef dims) const { + return at::_ops::flip::call(const_cast(*this), dims); +} + +// aten::fliplr(Tensor self) -> Tensor +inline at::Tensor Tensor::fliplr() const { + return at::_ops::fliplr::call(const_cast(*this)); +} + +// aten::flipud(Tensor self) -> Tensor +inline at::Tensor Tensor::flipud() const { + return at::_ops::flipud::call(const_cast(*this)); +} + +// aten::roll(Tensor self, SymInt[1] shifts, int[1] dims=[]) -> Tensor +inline at::Tensor Tensor::roll(at::IntArrayRef shifts, at::IntArrayRef dims) const { + return at::_ops::roll::call(const_cast(*this), c10::fromIntArrayRefSlow(shifts), dims); +} + +// aten::roll(Tensor self, SymInt[1] shifts, int[1] dims=[]) -> Tensor +inline at::Tensor Tensor::roll_symint(c10::SymIntArrayRef shifts, at::IntArrayRef dims) const { + return at::_ops::roll::call(const_cast(*this), shifts, dims); +} + +// aten::rot90(Tensor self, int k=1, int[] dims=[0,1]) -> Tensor +inline at::Tensor Tensor::rot90(int64_t k, at::IntArrayRef dims) const { + return at::_ops::rot90::call(const_cast(*this), k, dims); +} + +// aten::_nested_tensor_size(Tensor self) -> Tensor +inline at::Tensor Tensor::_nested_tensor_size() const { + return at::_ops::_nested_tensor_size::call(const_cast(*this)); +} + +// aten::_nested_tensor_strides(Tensor self) -> Tensor +inline at::Tensor Tensor::_nested_tensor_strides() const { + return at::_ops::_nested_tensor_strides::call(const_cast(*this)); +} + +// aten::_nested_tensor_storage_offsets(Tensor self) -> Tensor +inline at::Tensor Tensor::_nested_tensor_storage_offsets() const { + return at::_ops::_nested_tensor_storage_offsets::call(const_cast(*this)); +} + +// aten::trunc(Tensor self) -> Tensor +inline at::Tensor Tensor::trunc() const { + return at::_ops::trunc::call(const_cast(*this)); +} + +// aten::trunc_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::trunc_() const { + return at::_ops::trunc_::call(const_cast(*this)); +} + +// aten::fix(Tensor self) -> Tensor +inline at::Tensor Tensor::fix() const { + return at::_ops::fix::call(const_cast(*this)); +} + +// aten::fix_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::fix_() const { + return at::_ops::fix_::call(const_cast(*this)); +} + +// aten::type_as(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::type_as(const at::Tensor & other) const { + return at::_ops::type_as::call(const_cast(*this), other); +} + +// aten::unsqueeze(Tensor(a) self, int dim) -> Tensor(a) +inline at::Tensor Tensor::unsqueeze(int64_t dim) const { + return at::_ops::unsqueeze::call(const_cast(*this), dim); +} + +// aten::unsqueeze_(Tensor(a!) self, int dim) -> Tensor(a!) +inline at::Tensor & Tensor::unsqueeze_(int64_t dim) const { + return at::_ops::unsqueeze_::call(const_cast(*this), dim); +} + +// aten::var(Tensor self, bool unbiased=True) -> Tensor +inline at::Tensor Tensor::var(bool unbiased) const { + return at::_ops::var::call(const_cast(*this), unbiased); +} + +// aten::var.dim(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::var(at::OptionalIntArrayRef dim, bool unbiased, bool keepdim) const { + return at::_ops::var_dim::call(const_cast(*this), dim, unbiased, keepdim); +} + +// aten::var.correction(Tensor self, int[1]? dim=None, *, Scalar? correction=None, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::var(at::OptionalIntArrayRef dim, const ::std::optional & correction, bool keepdim) const { + return at::_ops::var_correction::call(const_cast(*this), dim, correction, keepdim); +} + +// aten::var.names_dim(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::var(at::DimnameList dim, bool unbiased, bool keepdim) const { + return at::_ops::var_names_dim::call(const_cast(*this), dim, unbiased, keepdim); +} + +// aten::var.correction_names(Tensor self, Dimname[1] dim, *, Scalar? correction=None, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::var(at::DimnameList dim, const ::std::optional & correction, bool keepdim) const { + return at::_ops::var_correction_names::call(const_cast(*this), dim, correction, keepdim); +} + +// aten::view_as(Tensor(a) self, Tensor other) -> Tensor(a) +inline at::Tensor Tensor::view_as(const at::Tensor & other) const { + return at::_ops::view_as::call(const_cast(*this), other); +} + +// aten::where.self(Tensor condition, Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::where(const at::Tensor & condition, const at::Tensor & other) const { + return at::_ops::where_self::call(condition, const_cast(*this), other); +} + +// aten::where.ScalarOther(Tensor condition, Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::where(const at::Tensor & condition, const at::Scalar & other) const { + return at::_ops::where_ScalarOther::call(condition, const_cast(*this), other); +} + +// aten::norm.ScalarOpt_dtype(Tensor self, Scalar? p, *, ScalarType dtype) -> Tensor +inline at::Tensor Tensor::norm(const ::std::optional & p, at::ScalarType dtype) const { + return at::_ops::norm_ScalarOpt_dtype::call(const_cast(*this), p, dtype); +} + +// aten::norm.Scalar(Tensor self, Scalar p=2) -> Tensor +inline at::Tensor Tensor::norm(const at::Scalar & p) const { + return at::_ops::norm_Scalar::call(const_cast(*this), p); +} + +// aten::norm.ScalarOpt_dim_dtype(Tensor self, Scalar? p, int[1] dim, bool keepdim, *, ScalarType dtype) -> Tensor +inline at::Tensor Tensor::norm(const ::std::optional & p, at::IntArrayRef dim, bool keepdim, at::ScalarType dtype) const { + return at::_ops::norm_ScalarOpt_dim_dtype::call(const_cast(*this), p, dim, keepdim, dtype); +} + +// aten::norm.ScalarOpt_dim(Tensor self, Scalar? p, int[1] dim, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::norm(const ::std::optional & p, at::IntArrayRef dim, bool keepdim) const { + return at::_ops::norm_ScalarOpt_dim::call(const_cast(*this), p, dim, keepdim); +} + +// aten::norm.names_ScalarOpt_dim_dtype(Tensor self, Scalar? p, Dimname[1] dim, bool keepdim, *, ScalarType dtype) -> Tensor +inline at::Tensor Tensor::norm(const ::std::optional & p, at::DimnameList dim, bool keepdim, at::ScalarType dtype) const { + return at::_ops::norm_names_ScalarOpt_dim_dtype::call(const_cast(*this), p, dim, keepdim, dtype); +} + +// aten::norm.names_ScalarOpt_dim(Tensor self, Scalar? p, Dimname[1] dim, bool keepdim=False) -> Tensor +inline at::Tensor Tensor::norm(const ::std::optional & p, at::DimnameList dim, bool keepdim) const { + return at::_ops::norm_names_ScalarOpt_dim::call(const_cast(*this), p, dim, keepdim); +} + +// aten::frexp.Tensor(Tensor self) -> (Tensor mantissa, Tensor exponent) +inline ::std::tuple Tensor::frexp() const { + return at::_ops::frexp_Tensor::call(const_cast(*this)); +} + +// aten::clone(Tensor self, *, MemoryFormat? memory_format=None) -> Tensor +inline at::Tensor Tensor::clone(::std::optional memory_format) const { + return at::_ops::clone::call(const_cast(*this), memory_format); +} + +// aten::positive(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::positive() const { + return at::_ops::positive::call(const_cast(*this)); +} + +// aten::resize_as_(Tensor(a!) self, Tensor the_template, *, MemoryFormat? memory_format=None) -> Tensor(a!) +inline const at::Tensor & Tensor::resize_as_(const at::Tensor & the_template, ::std::optional memory_format) const { + return at::_ops::resize_as_::call(const_cast(*this), the_template, memory_format); +} + +// aten::resize_as_sparse_(Tensor(a!) self, Tensor the_template) -> Tensor(a!) +inline const at::Tensor & Tensor::resize_as_sparse_(const at::Tensor & the_template) const { + return at::_ops::resize_as_sparse_::call(const_cast(*this), the_template); +} + +// aten::zero_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::zero_() const { + return at::_ops::zero_::call(const_cast(*this)); +} + +// aten::sub.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::sub(const at::Tensor & other, const at::Scalar & alpha) const { + return at::_ops::sub_Tensor::call(const_cast(*this), other, alpha); +} + +// aten::sub_.Tensor(Tensor(a!) self, Tensor other, *, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::sub_(const at::Tensor & other, const at::Scalar & alpha) const { + return at::_ops::sub__Tensor::call(const_cast(*this), other, alpha); +} + +// aten::sub.Scalar(Tensor self, Scalar other, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::sub(const at::Scalar & other, const at::Scalar & alpha) const { + return at::_ops::sub_Scalar::call(const_cast(*this), other, alpha); +} + +// aten::sub_.Scalar(Tensor(a!) self, Scalar other, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::sub_(const at::Scalar & other, const at::Scalar & alpha) const { + return at::_ops::sub__Scalar::call(const_cast(*this), other, alpha); +} + +// aten::subtract.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::subtract(const at::Tensor & other, const at::Scalar & alpha) const { + return at::_ops::subtract_Tensor::call(const_cast(*this), other, alpha); +} + +// aten::subtract_.Tensor(Tensor(a!) self, Tensor other, *, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::subtract_(const at::Tensor & other, const at::Scalar & alpha) const { + return at::_ops::subtract__Tensor::call(const_cast(*this), other, alpha); +} + +// aten::subtract.Scalar(Tensor self, Scalar other, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::subtract(const at::Scalar & other, const at::Scalar & alpha) const { + return at::_ops::subtract_Scalar::call(const_cast(*this), other, alpha); +} + +// aten::subtract_.Scalar(Tensor(a!) self, Scalar other, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::subtract_(const at::Scalar & other, const at::Scalar & alpha) const { + return at::_ops::subtract__Scalar::call(const_cast(*this), other, alpha); +} + +// aten::heaviside(Tensor self, Tensor values) -> Tensor +inline at::Tensor Tensor::heaviside(const at::Tensor & values) const { + return at::_ops::heaviside::call(const_cast(*this), values); +} + +// aten::heaviside_(Tensor(a!) self, Tensor values) -> Tensor(a!) +inline at::Tensor & Tensor::heaviside_(const at::Tensor & values) const { + return at::_ops::heaviside_::call(const_cast(*this), values); +} + +// aten::addmm(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::addmm(const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::addmm::call(const_cast(*this), mat1, mat2, beta, alpha); +} + +// aten::addmm_(Tensor(a!) self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::addmm_(const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::addmm_::call(const_cast(*this), mat1, mat2, beta, alpha); +} + +// aten::_addmm_activation(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1, bool use_gelu=False) -> Tensor +inline at::Tensor Tensor::_addmm_activation(const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha, bool use_gelu) const { + return at::_ops::_addmm_activation::call(const_cast(*this), mat1, mat2, beta, alpha, use_gelu); +} + +// aten::sparse_resize_(Tensor(a!) self, int[] size, int sparse_dim, int dense_dim) -> Tensor(a!) +inline const at::Tensor & Tensor::sparse_resize_(at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim) const { + return at::_ops::sparse_resize_::call(const_cast(*this), size, sparse_dim, dense_dim); +} + +// aten::sparse_resize_and_clear_(Tensor(a!) self, int[] size, int sparse_dim, int dense_dim) -> Tensor(a!) +inline const at::Tensor & Tensor::sparse_resize_and_clear_(at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim) const { + return at::_ops::sparse_resize_and_clear_::call(const_cast(*this), size, sparse_dim, dense_dim); +} + +// aten::sparse_mask(Tensor self, Tensor mask) -> Tensor +inline at::Tensor Tensor::sparse_mask(const at::Tensor & mask) const { + return at::_ops::sparse_mask::call(const_cast(*this), mask); +} + +// aten::_sparse_mask_projection(Tensor self, Tensor mask, bool accumulate_matches=False) -> Tensor +inline at::Tensor Tensor::_sparse_mask_projection(const at::Tensor & mask, bool accumulate_matches) const { + return at::_ops::_sparse_mask_projection::call(const_cast(*this), mask, accumulate_matches); +} + +// aten::to_dense(Tensor self, ScalarType? dtype=None, *, bool? masked_grad=None) -> Tensor +inline at::Tensor Tensor::to_dense(::std::optional dtype, ::std::optional masked_grad) const { + return at::_ops::to_dense::call(const_cast(*this), dtype, masked_grad); +} + +// aten::_to_dense(Tensor self, ScalarType? dtype=None, bool? masked_grad=None) -> Tensor +inline at::Tensor Tensor::_to_dense(::std::optional dtype, ::std::optional masked_grad) const { + return at::_ops::_to_dense::call(const_cast(*this), dtype, masked_grad); +} + +// aten::sparse_dim(Tensor self) -> int +inline int64_t Tensor::sparse_dim() const { + return at::_ops::sparse_dim::call(const_cast(*this)); +} + +// aten::_dimI(Tensor self) -> int +inline int64_t Tensor::_dimI() const { + return at::_ops::_dimI::call(const_cast(*this)); +} + +// aten::dense_dim(Tensor self) -> int +inline int64_t Tensor::dense_dim() const { + return at::_ops::dense_dim::call(const_cast(*this)); +} + +// aten::_dimV(Tensor self) -> int +inline int64_t Tensor::_dimV() const { + return at::_ops::_dimV::call(const_cast(*this)); +} + +// aten::_nnz(Tensor self) -> int +inline int64_t Tensor::_nnz() const { + return at::_ops::_nnz::call(const_cast(*this)); +} + +// aten::coalesce(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::coalesce() const { + return at::_ops::coalesce::call(const_cast(*this)); +} + +// aten::is_coalesced(Tensor self) -> bool +inline bool Tensor::is_coalesced() const { + return at::_ops::is_coalesced::call(const_cast(*this)); +} + +// aten::_indices(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::_indices() const { + return at::_ops::_indices::call(const_cast(*this)); +} + +// aten::_values(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::_values() const { + return at::_ops::_values::call(const_cast(*this)); +} + +// aten::_coalesced_(Tensor(a!) self, bool coalesced) -> Tensor(a!) +inline at::Tensor & Tensor::_coalesced_(bool coalesced) const { + return at::_ops::_coalesced_::call(const_cast(*this), coalesced); +} + +// aten::indices(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::indices() const { + return at::_ops::indices::call(const_cast(*this)); +} + +// aten::values(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::values() const { + return at::_ops::values::call(const_cast(*this)); +} + +// aten::crow_indices(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::crow_indices() const { + return at::_ops::crow_indices::call(const_cast(*this)); +} + +// aten::col_indices(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::col_indices() const { + return at::_ops::col_indices::call(const_cast(*this)); +} + +// aten::ccol_indices(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::ccol_indices() const { + return at::_ops::ccol_indices::call(const_cast(*this)); +} + +// aten::row_indices(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::row_indices() const { + return at::_ops::row_indices::call(const_cast(*this)); +} + +// aten::unbind.int(Tensor(a -> *) self, int dim=0) -> Tensor(a)[] +inline ::std::vector Tensor::unbind(int64_t dim) const { + return at::_ops::unbind_int::call(const_cast(*this), dim); +} + +// aten::unbind.Dimname(Tensor(a -> *) self, Dimname dim) -> Tensor(a)[] +inline ::std::vector Tensor::unbind(at::Dimname dim) const { + return at::_ops::unbind_Dimname::call(const_cast(*this), dim); +} + +// aten::to_sparse.sparse_dim(Tensor self, int sparse_dim) -> Tensor +inline at::Tensor Tensor::to_sparse(int64_t sparse_dim) const { + return at::_ops::to_sparse_sparse_dim::call(const_cast(*this), sparse_dim); +} + +// aten::_to_sparse.sparse_dim(Tensor self, int sparse_dim) -> Tensor +inline at::Tensor Tensor::_to_sparse(int64_t sparse_dim) const { + return at::_ops::_to_sparse_sparse_dim::call(const_cast(*this), sparse_dim); +} + +// aten::to_sparse(Tensor self, *, Layout? layout=None, int[2]? blocksize=None, int? dense_dim=None) -> Tensor +inline at::Tensor Tensor::to_sparse(::std::optional layout, at::OptionalIntArrayRef blocksize, ::std::optional dense_dim) const { + return at::_ops::to_sparse::call(const_cast(*this), layout, blocksize, dense_dim); +} + +// aten::_to_sparse(Tensor self, *, Layout? layout=None, int[2]? blocksize=None, int? dense_dim=None) -> Tensor +inline at::Tensor Tensor::_to_sparse(::std::optional layout, at::OptionalIntArrayRef blocksize, ::std::optional dense_dim) const { + return at::_ops::_to_sparse::call(const_cast(*this), layout, blocksize, dense_dim); +} + +// aten::to_sparse_csr(Tensor self, int? dense_dim=None) -> Tensor +inline at::Tensor Tensor::to_sparse_csr(::std::optional dense_dim) const { + return at::_ops::to_sparse_csr::call(const_cast(*this), dense_dim); +} + +// aten::_to_sparse_csr(Tensor self, int? dense_dim=None) -> Tensor +inline at::Tensor Tensor::_to_sparse_csr(::std::optional dense_dim) const { + return at::_ops::_to_sparse_csr::call(const_cast(*this), dense_dim); +} + +// aten::to_sparse_csc(Tensor self, int? dense_dim=None) -> Tensor +inline at::Tensor Tensor::to_sparse_csc(::std::optional dense_dim) const { + return at::_ops::to_sparse_csc::call(const_cast(*this), dense_dim); +} + +// aten::_to_sparse_csc(Tensor self, int? dense_dim=None) -> Tensor +inline at::Tensor Tensor::_to_sparse_csc(::std::optional dense_dim) const { + return at::_ops::_to_sparse_csc::call(const_cast(*this), dense_dim); +} + +// aten::to_sparse_bsr(Tensor self, int[2] blocksize, int? dense_dim=None) -> Tensor +inline at::Tensor Tensor::to_sparse_bsr(at::IntArrayRef blocksize, ::std::optional dense_dim) const { + return at::_ops::to_sparse_bsr::call(const_cast(*this), blocksize, dense_dim); +} + +// aten::_to_sparse_bsr(Tensor self, int[2] blocksize, int? dense_dim=None) -> Tensor +inline at::Tensor Tensor::_to_sparse_bsr(at::IntArrayRef blocksize, ::std::optional dense_dim) const { + return at::_ops::_to_sparse_bsr::call(const_cast(*this), blocksize, dense_dim); +} + +// aten::to_sparse_bsc(Tensor self, int[2] blocksize, int? dense_dim=None) -> Tensor +inline at::Tensor Tensor::to_sparse_bsc(at::IntArrayRef blocksize, ::std::optional dense_dim) const { + return at::_ops::to_sparse_bsc::call(const_cast(*this), blocksize, dense_dim); +} + +// aten::_to_sparse_bsc(Tensor self, int[2] blocksize, int? dense_dim=None) -> Tensor +inline at::Tensor Tensor::_to_sparse_bsc(at::IntArrayRef blocksize, ::std::optional dense_dim) const { + return at::_ops::_to_sparse_bsc::call(const_cast(*this), blocksize, dense_dim); +} + +// aten::to_mkldnn(Tensor self, ScalarType? dtype=None) -> Tensor +inline at::Tensor Tensor::to_mkldnn(::std::optional dtype) const { + return at::_ops::to_mkldnn::call(const_cast(*this), dtype); +} + +// aten::dequantize.self(Tensor self) -> Tensor +inline at::Tensor Tensor::dequantize() const { + return at::_ops::dequantize_self::call(const_cast(*this)); +} + +// aten::q_scale(Tensor self) -> float +inline double Tensor::q_scale() const { + return at::_ops::q_scale::call(const_cast(*this)); +} + +// aten::q_zero_point(Tensor self) -> int +inline int64_t Tensor::q_zero_point() const { + return at::_ops::q_zero_point::call(const_cast(*this)); +} + +// aten::q_per_channel_scales(Tensor self) -> Tensor +inline at::Tensor Tensor::q_per_channel_scales() const { + return at::_ops::q_per_channel_scales::call(const_cast(*this)); +} + +// aten::q_per_channel_zero_points(Tensor self) -> Tensor +inline at::Tensor Tensor::q_per_channel_zero_points() const { + return at::_ops::q_per_channel_zero_points::call(const_cast(*this)); +} + +// aten::q_per_channel_axis(Tensor self) -> int +inline int64_t Tensor::q_per_channel_axis() const { + return at::_ops::q_per_channel_axis::call(const_cast(*this)); +} + +// aten::int_repr(Tensor self) -> Tensor +inline at::Tensor Tensor::int_repr() const { + return at::_ops::int_repr::call(const_cast(*this)); +} + +// aten::qscheme(Tensor self) -> QScheme +inline at::QScheme Tensor::qscheme() const { + return at::_ops::qscheme::call(const_cast(*this)); +} + +// aten::_autocast_to_reduced_precision(Tensor(a) self, bool cuda_enabled, bool cpu_enabled, ScalarType cuda_dtype, ScalarType cpu_dtype) -> Tensor(a) +inline at::Tensor Tensor::_autocast_to_reduced_precision(bool cuda_enabled, bool cpu_enabled, at::ScalarType cuda_dtype, at::ScalarType cpu_dtype) const { + return at::_ops::_autocast_to_reduced_precision::call(const_cast(*this), cuda_enabled, cpu_enabled, cuda_dtype, cpu_dtype); +} + +// aten::_autocast_to_full_precision(Tensor(a) self, bool cuda_enabled, bool cpu_enabled) -> Tensor(a) +inline at::Tensor Tensor::_autocast_to_full_precision(bool cuda_enabled, bool cpu_enabled) const { + return at::_ops::_autocast_to_full_precision::call(const_cast(*this), cuda_enabled, cpu_enabled); +} + +// aten::to.dtype_layout(Tensor(a) self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a) +inline at::Tensor Tensor::to(at::TensorOptions options, bool non_blocking, bool copy, ::std::optional memory_format) const { + return at::_ops::to_dtype_layout::call(const_cast(*this), c10::optTypeMetaToScalarType(options.dtype_opt()), options.layout_opt(), options.device_opt(), options.pinned_memory_opt(), non_blocking, copy, c10::impl::check_tensor_options_and_extract_memory_format(options, memory_format)); +} + +// aten::to.dtype_layout(Tensor(a) self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a) +inline at::Tensor Tensor::to(::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory, bool non_blocking, bool copy, ::std::optional memory_format) const { + return at::_ops::to_dtype_layout::call(const_cast(*this), dtype, layout, device, pin_memory, non_blocking, copy, memory_format); +} + +// aten::to.device(Tensor(a) self, Device device, ScalarType dtype, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a) +inline at::Tensor Tensor::to(at::Device device, at::ScalarType dtype, bool non_blocking, bool copy, ::std::optional memory_format) const { + return at::_ops::to_device::call(const_cast(*this), device, dtype, non_blocking, copy, memory_format); +} + +// aten::to.dtype(Tensor(a) self, ScalarType dtype, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a) +inline at::Tensor Tensor::to(at::ScalarType dtype, bool non_blocking, bool copy, ::std::optional memory_format) const { + return at::_ops::to_dtype::call(const_cast(*this), dtype, non_blocking, copy, memory_format); +} + +// aten::to.other(Tensor(a) self, Tensor other, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a) +inline at::Tensor Tensor::to(const at::Tensor & other, bool non_blocking, bool copy, ::std::optional memory_format) const { + return at::_ops::to_other::call(const_cast(*this), other, non_blocking, copy, memory_format); +} + +// aten::item(Tensor self) -> Scalar +inline at::Scalar Tensor::item() const { + return at::_ops::item::call(const_cast(*this)); +} + +// aten::set_.source_Storage(Tensor(a!) self, Storage source) -> Tensor(a!) +inline at::Tensor & Tensor::set_(at::Storage source) const { + return at::_ops::set__source_Storage::call(const_cast(*this), source); +} + +// aten::set_.source_Storage_storage_offset(Tensor(a!) self, Storage source, SymInt storage_offset, SymInt[] size, SymInt[] stride=[]) -> Tensor(a!) +inline at::Tensor & Tensor::set_(at::Storage source, int64_t storage_offset, at::IntArrayRef size, at::IntArrayRef stride) const { + return at::_ops::set__source_Storage_storage_offset::call(const_cast(*this), source, storage_offset, c10::fromIntArrayRefSlow(size), c10::fromIntArrayRefSlow(stride)); +} + +// aten::set_.source_Storage_storage_offset(Tensor(a!) self, Storage source, SymInt storage_offset, SymInt[] size, SymInt[] stride=[]) -> Tensor(a!) +inline at::Tensor & Tensor::set__symint(at::Storage source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride) const { + return at::_ops::set__source_Storage_storage_offset::call(const_cast(*this), source, storage_offset, size, stride); +} + +// aten::set_.source_Tensor_storage_offset(Tensor(a!) self, Tensor source, SymInt storage_offset, SymInt[] size, SymInt[] stride=[]) -> Tensor(a!) +inline at::Tensor & Tensor::set_(const at::Tensor & source, int64_t storage_offset, at::IntArrayRef size, at::IntArrayRef stride) const { + return at::_ops::set__source_Tensor_storage_offset::call(const_cast(*this), source, storage_offset, c10::fromIntArrayRefSlow(size), c10::fromIntArrayRefSlow(stride)); +} + +// aten::set_.source_Tensor_storage_offset(Tensor(a!) self, Tensor source, SymInt storage_offset, SymInt[] size, SymInt[] stride=[]) -> Tensor(a!) +inline at::Tensor & Tensor::set__symint(const at::Tensor & source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride) const { + return at::_ops::set__source_Tensor_storage_offset::call(const_cast(*this), source, storage_offset, size, stride); +} + +// aten::set_.source_Tensor(Tensor(a!) self, Tensor source) -> Tensor(a!) +inline at::Tensor & Tensor::set_(const at::Tensor & source) const { + return at::_ops::set__source_Tensor::call(const_cast(*this), source); +} + +// aten::set_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::set_() const { + return at::_ops::set_::call(const_cast(*this)); +} + +// aten::is_set_to(Tensor self, Tensor tensor) -> bool +inline bool Tensor::is_set_to(const at::Tensor & tensor) const { + return at::_ops::is_set_to::call(const_cast(*this), tensor); +} + +// aten::masked_fill_.Scalar(Tensor(a!) self, Tensor mask, Scalar value) -> Tensor(a!) +inline at::Tensor & Tensor::masked_fill_(const at::Tensor & mask, const at::Scalar & value) const { + return at::_ops::masked_fill__Scalar::call(const_cast(*this), mask, value); +} + +// aten::masked_fill.Scalar(Tensor self, Tensor mask, Scalar value) -> Tensor +inline at::Tensor Tensor::masked_fill(const at::Tensor & mask, const at::Scalar & value) const { + return at::_ops::masked_fill_Scalar::call(const_cast(*this), mask, value); +} + +// aten::masked_fill_.Tensor(Tensor(a!) self, Tensor mask, Tensor value) -> Tensor(a!) +inline at::Tensor & Tensor::masked_fill_(const at::Tensor & mask, const at::Tensor & value) const { + return at::_ops::masked_fill__Tensor::call(const_cast(*this), mask, value); +} + +// aten::masked_fill.Tensor(Tensor self, Tensor mask, Tensor value) -> Tensor +inline at::Tensor Tensor::masked_fill(const at::Tensor & mask, const at::Tensor & value) const { + return at::_ops::masked_fill_Tensor::call(const_cast(*this), mask, value); +} + +// aten::masked_scatter_(Tensor(a!) self, Tensor mask, Tensor source) -> Tensor(a!) +inline at::Tensor & Tensor::masked_scatter_(const at::Tensor & mask, const at::Tensor & source) const { + return at::_ops::masked_scatter_::call(const_cast(*this), mask, source); +} + +// aten::masked_scatter(Tensor self, Tensor mask, Tensor source) -> Tensor +inline at::Tensor Tensor::masked_scatter(const at::Tensor & mask, const at::Tensor & source) const { + return at::_ops::masked_scatter::call(const_cast(*this), mask, source); +} + +// aten::view(Tensor(a) self, SymInt[] size) -> Tensor(a) +inline at::Tensor Tensor::view(at::IntArrayRef size) const { + return at::_ops::view::call(const_cast(*this), c10::fromIntArrayRefSlow(size)); +} + +// aten::view(Tensor(a) self, SymInt[] size) -> Tensor(a) +inline at::Tensor Tensor::view_symint(c10::SymIntArrayRef size) const { + return at::_ops::view::call(const_cast(*this), size); +} + +// aten::view.dtype(Tensor(a) self, ScalarType dtype) -> Tensor(a) +inline at::Tensor Tensor::view(at::ScalarType dtype) const { + return at::_ops::view_dtype::call(const_cast(*this), dtype); +} + +// aten::put_(Tensor(a!) self, Tensor index, Tensor source, bool accumulate=False) -> Tensor(a!) +inline at::Tensor & Tensor::put_(const at::Tensor & index, const at::Tensor & source, bool accumulate) const { + return at::_ops::put_::call(const_cast(*this), index, source, accumulate); +} + +// aten::put(Tensor self, Tensor index, Tensor source, bool accumulate=False) -> Tensor +inline at::Tensor Tensor::put(const at::Tensor & index, const at::Tensor & source, bool accumulate) const { + return at::_ops::put::call(const_cast(*this), index, source, accumulate); +} + +// aten::index_add_(Tensor(a!) self, int dim, Tensor index, Tensor source, *, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::index_add_(int64_t dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha) const { + return at::_ops::index_add_::call(const_cast(*this), dim, index, source, alpha); +} + +// aten::index_add(Tensor self, int dim, Tensor index, Tensor source, *, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::index_add(int64_t dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha) const { + return at::_ops::index_add::call(const_cast(*this), dim, index, source, alpha); +} + +// aten::index_add.dimname(Tensor self, Dimname dim, Tensor index, Tensor source, *, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::index_add(at::Dimname dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha) const { + return at::_ops::index_add_dimname::call(const_cast(*this), dim, index, source, alpha); +} + +// aten::index_reduce_(Tensor(a!) self, int dim, Tensor index, Tensor source, str reduce, *, bool include_self=True) -> Tensor(a!) +inline at::Tensor & Tensor::index_reduce_(int64_t dim, const at::Tensor & index, const at::Tensor & source, c10::string_view reduce, bool include_self) const { + return at::_ops::index_reduce_::call(const_cast(*this), dim, index, source, reduce, include_self); +} + +// aten::index_reduce(Tensor self, int dim, Tensor index, Tensor source, str reduce, *, bool include_self=True) -> Tensor +inline at::Tensor Tensor::index_reduce(int64_t dim, const at::Tensor & index, const at::Tensor & source, c10::string_view reduce, bool include_self) const { + return at::_ops::index_reduce::call(const_cast(*this), dim, index, source, reduce, include_self); +} + +// aten::index_fill_.int_Scalar(Tensor(a!) self, int dim, Tensor index, Scalar value) -> Tensor(a!) +inline at::Tensor & Tensor::index_fill_(int64_t dim, const at::Tensor & index, const at::Scalar & value) const { + return at::_ops::index_fill__int_Scalar::call(const_cast(*this), dim, index, value); +} + +// aten::index_fill.int_Scalar(Tensor self, int dim, Tensor index, Scalar value) -> Tensor +inline at::Tensor Tensor::index_fill(int64_t dim, const at::Tensor & index, const at::Scalar & value) const { + return at::_ops::index_fill_int_Scalar::call(const_cast(*this), dim, index, value); +} + +// aten::index_fill_.int_Tensor(Tensor(a!) self, int dim, Tensor index, Tensor value) -> Tensor(a!) +inline at::Tensor & Tensor::index_fill_(int64_t dim, const at::Tensor & index, const at::Tensor & value) const { + return at::_ops::index_fill__int_Tensor::call(const_cast(*this), dim, index, value); +} + +// aten::index_fill.int_Tensor(Tensor self, int dim, Tensor index, Tensor value) -> Tensor +inline at::Tensor Tensor::index_fill(int64_t dim, const at::Tensor & index, const at::Tensor & value) const { + return at::_ops::index_fill_int_Tensor::call(const_cast(*this), dim, index, value); +} + +// aten::index_fill_.Dimname_Scalar(Tensor(a!) self, Dimname dim, Tensor index, Scalar value) -> Tensor(a!) +inline at::Tensor & Tensor::index_fill_(at::Dimname dim, const at::Tensor & index, const at::Scalar & value) const { + return at::_ops::index_fill__Dimname_Scalar::call(const_cast(*this), dim, index, value); +} + +// aten::index_fill_.Dimname_Tensor(Tensor(a!) self, Dimname dim, Tensor index, Tensor value) -> Tensor(a!) +inline at::Tensor & Tensor::index_fill_(at::Dimname dim, const at::Tensor & index, const at::Tensor & value) const { + return at::_ops::index_fill__Dimname_Tensor::call(const_cast(*this), dim, index, value); +} + +// aten::index_fill.Dimname_Scalar(Tensor self, Dimname dim, Tensor index, Scalar value) -> Tensor +inline at::Tensor Tensor::index_fill(at::Dimname dim, const at::Tensor & index, const at::Scalar & value) const { + return at::_ops::index_fill_Dimname_Scalar::call(const_cast(*this), dim, index, value); +} + +// aten::index_fill.Dimname_Tensor(Tensor self, Dimname dim, Tensor index, Tensor value) -> Tensor +inline at::Tensor Tensor::index_fill(at::Dimname dim, const at::Tensor & index, const at::Tensor & value) const { + return at::_ops::index_fill_Dimname_Tensor::call(const_cast(*this), dim, index, value); +} + +// aten::scatter.src(Tensor self, int dim, Tensor index, Tensor src) -> Tensor +inline at::Tensor Tensor::scatter(int64_t dim, const at::Tensor & index, const at::Tensor & src) const { + return at::_ops::scatter_src::call(const_cast(*this), dim, index, src); +} + +// aten::scatter_.src(Tensor(a!) self, int dim, Tensor index, Tensor src) -> Tensor(a!) +inline at::Tensor & Tensor::scatter_(int64_t dim, const at::Tensor & index, const at::Tensor & src) const { + return at::_ops::scatter__src::call(const_cast(*this), dim, index, src); +} + +// aten::scatter.value(Tensor self, int dim, Tensor index, Scalar value) -> Tensor +inline at::Tensor Tensor::scatter(int64_t dim, const at::Tensor & index, const at::Scalar & value) const { + return at::_ops::scatter_value::call(const_cast(*this), dim, index, value); +} + +// aten::scatter_.value(Tensor(a!) self, int dim, Tensor index, Scalar value) -> Tensor(a!) +inline at::Tensor & Tensor::scatter_(int64_t dim, const at::Tensor & index, const at::Scalar & value) const { + return at::_ops::scatter__value::call(const_cast(*this), dim, index, value); +} + +// aten::scatter.reduce(Tensor self, int dim, Tensor index, Tensor src, *, str reduce) -> Tensor +inline at::Tensor Tensor::scatter(int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce) const { + return at::_ops::scatter_reduce::call(const_cast(*this), dim, index, src, reduce); +} + +// aten::scatter_.reduce(Tensor(a!) self, int dim, Tensor index, Tensor src, *, str reduce) -> Tensor(a!) +inline at::Tensor & Tensor::scatter_(int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce) const { + return at::_ops::scatter__reduce::call(const_cast(*this), dim, index, src, reduce); +} + +// aten::scatter.value_reduce(Tensor self, int dim, Tensor index, Scalar value, *, str reduce) -> Tensor +inline at::Tensor Tensor::scatter(int64_t dim, const at::Tensor & index, const at::Scalar & value, c10::string_view reduce) const { + return at::_ops::scatter_value_reduce::call(const_cast(*this), dim, index, value, reduce); +} + +// aten::scatter_.value_reduce(Tensor(a!) self, int dim, Tensor index, Scalar value, *, str reduce) -> Tensor(a!) +inline at::Tensor & Tensor::scatter_(int64_t dim, const at::Tensor & index, const at::Scalar & value, c10::string_view reduce) const { + return at::_ops::scatter__value_reduce::call(const_cast(*this), dim, index, value, reduce); +} + +// aten::scatter.dimname_src(Tensor self, Dimname dim, Tensor index, Tensor src) -> Tensor +inline at::Tensor Tensor::scatter(at::Dimname dim, const at::Tensor & index, const at::Tensor & src) const { + return at::_ops::scatter_dimname_src::call(const_cast(*this), dim, index, src); +} + +// aten::scatter.dimname_value(Tensor self, Dimname dim, Tensor index, Scalar value) -> Tensor +inline at::Tensor Tensor::scatter(at::Dimname dim, const at::Tensor & index, const at::Scalar & value) const { + return at::_ops::scatter_dimname_value::call(const_cast(*this), dim, index, value); +} + +// aten::scatter_add(Tensor self, int dim, Tensor index, Tensor src) -> Tensor +inline at::Tensor Tensor::scatter_add(int64_t dim, const at::Tensor & index, const at::Tensor & src) const { + return at::_ops::scatter_add::call(const_cast(*this), dim, index, src); +} + +// aten::scatter_add_(Tensor(a!) self, int dim, Tensor index, Tensor src) -> Tensor(a!) +inline at::Tensor & Tensor::scatter_add_(int64_t dim, const at::Tensor & index, const at::Tensor & src) const { + return at::_ops::scatter_add_::call(const_cast(*this), dim, index, src); +} + +// aten::scatter_add.dimname(Tensor self, Dimname dim, Tensor index, Tensor src) -> Tensor +inline at::Tensor Tensor::scatter_add(at::Dimname dim, const at::Tensor & index, const at::Tensor & src) const { + return at::_ops::scatter_add_dimname::call(const_cast(*this), dim, index, src); +} + +// aten::scatter_reduce.two(Tensor self, int dim, Tensor index, Tensor src, str reduce, *, bool include_self=True) -> Tensor +inline at::Tensor Tensor::scatter_reduce(int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, bool include_self) const { + return at::_ops::scatter_reduce_two::call(const_cast(*this), dim, index, src, reduce, include_self); +} + +// aten::scatter_reduce_.two(Tensor(a!) self, int dim, Tensor index, Tensor src, str reduce, *, bool include_self=True) -> Tensor(a!) +inline at::Tensor & Tensor::scatter_reduce_(int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, bool include_self) const { + return at::_ops::scatter_reduce__two::call(const_cast(*this), dim, index, src, reduce, include_self); +} + +// aten::eq_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::eq_(const at::Scalar & other) const { + return at::_ops::eq__Scalar::call(const_cast(*this), other); +} + +// aten::eq_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::eq_(const at::Tensor & other) const { + return at::_ops::eq__Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_and.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::bitwise_and(const at::Scalar & other) const { + return at::_ops::bitwise_and_Scalar::call(const_cast(*this), other); +} + +// aten::bitwise_and.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::bitwise_and(const at::Tensor & other) const { + return at::_ops::bitwise_and_Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_and_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_and_(const at::Scalar & other) const { + return at::_ops::bitwise_and__Scalar::call(const_cast(*this), other); +} + +// aten::bitwise_and_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_and_(const at::Tensor & other) const { + return at::_ops::bitwise_and__Tensor::call(const_cast(*this), other); +} + +// aten::__and__.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::__and__(const at::Scalar & other) const { + return at::_ops::__and___Scalar::call(const_cast(*this), other); +} + +// aten::__and__.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::__and__(const at::Tensor & other) const { + return at::_ops::__and___Tensor::call(const_cast(*this), other); +} + +// aten::__iand__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::__iand__(const at::Scalar & other) const { + return at::_ops::__iand___Scalar::call(const_cast(*this), other); +} + +// aten::__iand__.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::__iand__(const at::Tensor & other) const { + return at::_ops::__iand___Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_or.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::bitwise_or(const at::Scalar & other) const { + return at::_ops::bitwise_or_Scalar::call(const_cast(*this), other); +} + +// aten::bitwise_or.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::bitwise_or(const at::Tensor & other) const { + return at::_ops::bitwise_or_Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_or_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_or_(const at::Scalar & other) const { + return at::_ops::bitwise_or__Scalar::call(const_cast(*this), other); +} + +// aten::bitwise_or_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_or_(const at::Tensor & other) const { + return at::_ops::bitwise_or__Tensor::call(const_cast(*this), other); +} + +// aten::__or__.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::__or__(const at::Scalar & other) const { + return at::_ops::__or___Scalar::call(const_cast(*this), other); +} + +// aten::__or__.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::__or__(const at::Tensor & other) const { + return at::_ops::__or___Tensor::call(const_cast(*this), other); +} + +// aten::__ior__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::__ior__(const at::Scalar & other) const { + return at::_ops::__ior___Scalar::call(const_cast(*this), other); +} + +// aten::__ior__.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::__ior__(const at::Tensor & other) const { + return at::_ops::__ior___Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_xor.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::bitwise_xor(const at::Scalar & other) const { + return at::_ops::bitwise_xor_Scalar::call(const_cast(*this), other); +} + +// aten::bitwise_xor.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::bitwise_xor(const at::Tensor & other) const { + return at::_ops::bitwise_xor_Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_xor_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_xor_(const at::Scalar & other) const { + return at::_ops::bitwise_xor__Scalar::call(const_cast(*this), other); +} + +// aten::bitwise_xor_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_xor_(const at::Tensor & other) const { + return at::_ops::bitwise_xor__Tensor::call(const_cast(*this), other); +} + +// aten::__xor__.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::__xor__(const at::Scalar & other) const { + return at::_ops::__xor___Scalar::call(const_cast(*this), other); +} + +// aten::__xor__.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::__xor__(const at::Tensor & other) const { + return at::_ops::__xor___Tensor::call(const_cast(*this), other); +} + +// aten::__ixor__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::__ixor__(const at::Scalar & other) const { + return at::_ops::__ixor___Scalar::call(const_cast(*this), other); +} + +// aten::__ixor__.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::__ixor__(const at::Tensor & other) const { + return at::_ops::__ixor___Tensor::call(const_cast(*this), other); +} + +// aten::__lshift__.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::__lshift__(const at::Scalar & other) const { + return at::_ops::__lshift___Scalar::call(const_cast(*this), other); +} + +// aten::__lshift__.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::__lshift__(const at::Tensor & other) const { + return at::_ops::__lshift___Tensor::call(const_cast(*this), other); +} + +// aten::__ilshift__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::__ilshift__(const at::Scalar & other) const { + return at::_ops::__ilshift___Scalar::call(const_cast(*this), other); +} + +// aten::__ilshift__.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::__ilshift__(const at::Tensor & other) const { + return at::_ops::__ilshift___Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_left_shift.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::bitwise_left_shift(const at::Tensor & other) const { + return at::_ops::bitwise_left_shift_Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_left_shift_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_left_shift_(const at::Tensor & other) const { + return at::_ops::bitwise_left_shift__Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_left_shift.Tensor_Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::bitwise_left_shift(const at::Scalar & other) const { + return at::_ops::bitwise_left_shift_Tensor_Scalar::call(const_cast(*this), other); +} + +// aten::bitwise_left_shift_.Tensor_Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_left_shift_(const at::Scalar & other) const { + return at::_ops::bitwise_left_shift__Tensor_Scalar::call(const_cast(*this), other); +} + +// aten::__rshift__.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::__rshift__(const at::Scalar & other) const { + return at::_ops::__rshift___Scalar::call(const_cast(*this), other); +} + +// aten::__rshift__.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::__rshift__(const at::Tensor & other) const { + return at::_ops::__rshift___Tensor::call(const_cast(*this), other); +} + +// aten::__irshift__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::__irshift__(const at::Scalar & other) const { + return at::_ops::__irshift___Scalar::call(const_cast(*this), other); +} + +// aten::__irshift__.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::__irshift__(const at::Tensor & other) const { + return at::_ops::__irshift___Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_right_shift.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::bitwise_right_shift(const at::Tensor & other) const { + return at::_ops::bitwise_right_shift_Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_right_shift_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_right_shift_(const at::Tensor & other) const { + return at::_ops::bitwise_right_shift__Tensor::call(const_cast(*this), other); +} + +// aten::bitwise_right_shift.Tensor_Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::bitwise_right_shift(const at::Scalar & other) const { + return at::_ops::bitwise_right_shift_Tensor_Scalar::call(const_cast(*this), other); +} + +// aten::bitwise_right_shift_.Tensor_Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::bitwise_right_shift_(const at::Scalar & other) const { + return at::_ops::bitwise_right_shift__Tensor_Scalar::call(const_cast(*this), other); +} + +// aten::tril_(Tensor(a!) self, SymInt diagonal=0) -> Tensor(a!) +inline at::Tensor & Tensor::tril_(int64_t diagonal) const { + return at::_ops::tril_::call(const_cast(*this), diagonal); +} + +// aten::tril_(Tensor(a!) self, SymInt diagonal=0) -> Tensor(a!) +inline at::Tensor & Tensor::tril__symint(c10::SymInt diagonal) const { + return at::_ops::tril_::call(const_cast(*this), diagonal); +} + +// aten::triu_(Tensor(a!) self, SymInt diagonal=0) -> Tensor(a!) +inline at::Tensor & Tensor::triu_(int64_t diagonal) const { + return at::_ops::triu_::call(const_cast(*this), diagonal); +} + +// aten::triu_(Tensor(a!) self, SymInt diagonal=0) -> Tensor(a!) +inline at::Tensor & Tensor::triu__symint(c10::SymInt diagonal) const { + return at::_ops::triu_::call(const_cast(*this), diagonal); +} + +// aten::digamma_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::digamma_() const { + return at::_ops::digamma_::call(const_cast(*this)); +} + +// aten::lerp_.Scalar(Tensor(a!) self, Tensor end, Scalar weight) -> Tensor(a!) +inline at::Tensor & Tensor::lerp_(const at::Tensor & end, const at::Scalar & weight) const { + return at::_ops::lerp__Scalar::call(const_cast(*this), end, weight); +} + +// aten::lerp_.Tensor(Tensor(a!) self, Tensor end, Tensor weight) -> Tensor(a!) +inline at::Tensor & Tensor::lerp_(const at::Tensor & end, const at::Tensor & weight) const { + return at::_ops::lerp__Tensor::call(const_cast(*this), end, weight); +} + +// aten::addbmm_(Tensor(a!) self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!) +inline at::Tensor & Tensor::addbmm_(const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::addbmm_::call(const_cast(*this), batch1, batch2, beta, alpha); +} + +// aten::addbmm(Tensor self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor +inline at::Tensor Tensor::addbmm(const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha) const { + return at::_ops::addbmm::call(const_cast(*this), batch1, batch2, beta, alpha); +} + +// aten::random_.from(Tensor(a!) self, int from, int? to, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::random_(int64_t from, ::std::optional to, ::std::optional generator) const { + return at::_ops::random__from::call(const_cast(*this), from, to, generator); +} + +// aten::random_.to(Tensor(a!) self, int to, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::random_(int64_t to, ::std::optional generator) const { + return at::_ops::random__to::call(const_cast(*this), to, generator); +} + +// aten::random_(Tensor(a!) self, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::random_(::std::optional generator) const { + return at::_ops::random_::call(const_cast(*this), generator); +} + +// aten::uniform_(Tensor(a!) self, float from=0, float to=1, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::uniform_(double from, double to, ::std::optional generator) const { + return at::_ops::uniform_::call(const_cast(*this), from, to, generator); +} + +// aten::cauchy_(Tensor(a!) self, float median=0, float sigma=1, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::cauchy_(double median, double sigma, ::std::optional generator) const { + return at::_ops::cauchy_::call(const_cast(*this), median, sigma, generator); +} + +// aten::log_normal_(Tensor(a!) self, float mean=1, float std=2, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::log_normal_(double mean, double std, ::std::optional generator) const { + return at::_ops::log_normal_::call(const_cast(*this), mean, std, generator); +} + +// aten::exponential_(Tensor(a!) self, float lambd=1, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::exponential_(double lambd, ::std::optional generator) const { + return at::_ops::exponential_::call(const_cast(*this), lambd, generator); +} + +// aten::geometric_(Tensor(a!) self, float p, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::geometric_(double p, ::std::optional generator) const { + return at::_ops::geometric_::call(const_cast(*this), p, generator); +} + +// aten::diag(Tensor self, int diagonal=0) -> Tensor +inline at::Tensor Tensor::diag(int64_t diagonal) const { + return at::_ops::diag::call(const_cast(*this), diagonal); +} + +// aten::cross(Tensor self, Tensor other, int? dim=None) -> Tensor +inline at::Tensor Tensor::cross(const at::Tensor & other, ::std::optional dim) const { + return at::_ops::cross::call(const_cast(*this), other, dim); +} + +// aten::triu(Tensor self, SymInt diagonal=0) -> Tensor +inline at::Tensor Tensor::triu(int64_t diagonal) const { + return at::_ops::triu::call(const_cast(*this), diagonal); +} + +// aten::triu(Tensor self, SymInt diagonal=0) -> Tensor +inline at::Tensor Tensor::triu_symint(c10::SymInt diagonal) const { + return at::_ops::triu::call(const_cast(*this), diagonal); +} + +// aten::tril(Tensor self, SymInt diagonal=0) -> Tensor +inline at::Tensor Tensor::tril(int64_t diagonal) const { + return at::_ops::tril::call(const_cast(*this), diagonal); +} + +// aten::tril(Tensor self, SymInt diagonal=0) -> Tensor +inline at::Tensor Tensor::tril_symint(c10::SymInt diagonal) const { + return at::_ops::tril::call(const_cast(*this), diagonal); +} + +// aten::trace(Tensor self) -> Tensor +inline at::Tensor Tensor::trace() const { + return at::_ops::trace::call(const_cast(*this)); +} + +// aten::ne.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::ne(const at::Scalar & other) const { + return at::_ops::ne_Scalar::call(const_cast(*this), other); +} + +// aten::ne.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::ne(const at::Tensor & other) const { + return at::_ops::ne_Tensor::call(const_cast(*this), other); +} + +// aten::ne_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::ne_(const at::Scalar & other) const { + return at::_ops::ne__Scalar::call(const_cast(*this), other); +} + +// aten::ne_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::ne_(const at::Tensor & other) const { + return at::_ops::ne__Tensor::call(const_cast(*this), other); +} + +// aten::not_equal.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::not_equal(const at::Scalar & other) const { + return at::_ops::not_equal_Scalar::call(const_cast(*this), other); +} + +// aten::not_equal.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::not_equal(const at::Tensor & other) const { + return at::_ops::not_equal_Tensor::call(const_cast(*this), other); +} + +// aten::not_equal_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::not_equal_(const at::Scalar & other) const { + return at::_ops::not_equal__Scalar::call(const_cast(*this), other); +} + +// aten::not_equal_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::not_equal_(const at::Tensor & other) const { + return at::_ops::not_equal__Tensor::call(const_cast(*this), other); +} + +// aten::eq.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::eq(const at::Scalar & other) const { + return at::_ops::eq_Scalar::call(const_cast(*this), other); +} + +// aten::eq.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::eq(const at::Tensor & other) const { + return at::_ops::eq_Tensor::call(const_cast(*this), other); +} + +// aten::ge.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::ge(const at::Scalar & other) const { + return at::_ops::ge_Scalar::call(const_cast(*this), other); +} + +// aten::ge.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::ge(const at::Tensor & other) const { + return at::_ops::ge_Tensor::call(const_cast(*this), other); +} + +// aten::ge_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::ge_(const at::Scalar & other) const { + return at::_ops::ge__Scalar::call(const_cast(*this), other); +} + +// aten::ge_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::ge_(const at::Tensor & other) const { + return at::_ops::ge__Tensor::call(const_cast(*this), other); +} + +// aten::greater_equal.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::greater_equal(const at::Scalar & other) const { + return at::_ops::greater_equal_Scalar::call(const_cast(*this), other); +} + +// aten::greater_equal.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::greater_equal(const at::Tensor & other) const { + return at::_ops::greater_equal_Tensor::call(const_cast(*this), other); +} + +// aten::greater_equal_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::greater_equal_(const at::Scalar & other) const { + return at::_ops::greater_equal__Scalar::call(const_cast(*this), other); +} + +// aten::greater_equal_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::greater_equal_(const at::Tensor & other) const { + return at::_ops::greater_equal__Tensor::call(const_cast(*this), other); +} + +// aten::le.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::le(const at::Scalar & other) const { + return at::_ops::le_Scalar::call(const_cast(*this), other); +} + +// aten::le.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::le(const at::Tensor & other) const { + return at::_ops::le_Tensor::call(const_cast(*this), other); +} + +// aten::le_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::le_(const at::Scalar & other) const { + return at::_ops::le__Scalar::call(const_cast(*this), other); +} + +// aten::le_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::le_(const at::Tensor & other) const { + return at::_ops::le__Tensor::call(const_cast(*this), other); +} + +// aten::less_equal.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::less_equal(const at::Scalar & other) const { + return at::_ops::less_equal_Scalar::call(const_cast(*this), other); +} + +// aten::less_equal.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::less_equal(const at::Tensor & other) const { + return at::_ops::less_equal_Tensor::call(const_cast(*this), other); +} + +// aten::less_equal_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::less_equal_(const at::Scalar & other) const { + return at::_ops::less_equal__Scalar::call(const_cast(*this), other); +} + +// aten::less_equal_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::less_equal_(const at::Tensor & other) const { + return at::_ops::less_equal__Tensor::call(const_cast(*this), other); +} + +// aten::gt.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::gt(const at::Scalar & other) const { + return at::_ops::gt_Scalar::call(const_cast(*this), other); +} + +// aten::gt.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::gt(const at::Tensor & other) const { + return at::_ops::gt_Tensor::call(const_cast(*this), other); +} + +// aten::gt_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::gt_(const at::Scalar & other) const { + return at::_ops::gt__Scalar::call(const_cast(*this), other); +} + +// aten::gt_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::gt_(const at::Tensor & other) const { + return at::_ops::gt__Tensor::call(const_cast(*this), other); +} + +// aten::greater.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::greater(const at::Scalar & other) const { + return at::_ops::greater_Scalar::call(const_cast(*this), other); +} + +// aten::greater.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::greater(const at::Tensor & other) const { + return at::_ops::greater_Tensor::call(const_cast(*this), other); +} + +// aten::greater_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::greater_(const at::Scalar & other) const { + return at::_ops::greater__Scalar::call(const_cast(*this), other); +} + +// aten::greater_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::greater_(const at::Tensor & other) const { + return at::_ops::greater__Tensor::call(const_cast(*this), other); +} + +// aten::lt.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::lt(const at::Scalar & other) const { + return at::_ops::lt_Scalar::call(const_cast(*this), other); +} + +// aten::lt.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::lt(const at::Tensor & other) const { + return at::_ops::lt_Tensor::call(const_cast(*this), other); +} + +// aten::lt_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::lt_(const at::Scalar & other) const { + return at::_ops::lt__Scalar::call(const_cast(*this), other); +} + +// aten::lt_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::lt_(const at::Tensor & other) const { + return at::_ops::lt__Tensor::call(const_cast(*this), other); +} + +// aten::less.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::less(const at::Scalar & other) const { + return at::_ops::less_Scalar::call(const_cast(*this), other); +} + +// aten::less.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::less(const at::Tensor & other) const { + return at::_ops::less_Tensor::call(const_cast(*this), other); +} + +// aten::less_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::less_(const at::Scalar & other) const { + return at::_ops::less__Scalar::call(const_cast(*this), other); +} + +// aten::less_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::less_(const at::Tensor & other) const { + return at::_ops::less__Tensor::call(const_cast(*this), other); +} + +// aten::take(Tensor self, Tensor index) -> Tensor +inline at::Tensor Tensor::take(const at::Tensor & index) const { + return at::_ops::take::call(const_cast(*this), index); +} + +// aten::take_along_dim(Tensor self, Tensor indices, int? dim=None) -> Tensor +inline at::Tensor Tensor::take_along_dim(const at::Tensor & indices, ::std::optional dim) const { + return at::_ops::take_along_dim::call(const_cast(*this), indices, dim); +} + +// aten::index_select(Tensor self, int dim, Tensor index) -> Tensor +inline at::Tensor Tensor::index_select(int64_t dim, const at::Tensor & index) const { + return at::_ops::index_select::call(const_cast(*this), dim, index); +} + +// aten::index_select.dimname(Tensor self, Dimname dim, Tensor index) -> Tensor +inline at::Tensor Tensor::index_select(at::Dimname dim, const at::Tensor & index) const { + return at::_ops::index_select_dimname::call(const_cast(*this), dim, index); +} + +// aten::masked_select(Tensor self, Tensor mask) -> Tensor +inline at::Tensor Tensor::masked_select(const at::Tensor & mask) const { + return at::_ops::masked_select::call(const_cast(*this), mask); +} + +// aten::nonzero(Tensor self) -> Tensor +inline at::Tensor Tensor::nonzero() const { + return at::_ops::nonzero::call(const_cast(*this)); +} + +// aten::nonzero_static(Tensor self, *, SymInt size, int fill_value=-1) -> Tensor +inline at::Tensor Tensor::nonzero_static(int64_t size, int64_t fill_value) const { + return at::_ops::nonzero_static::call(const_cast(*this), size, fill_value); +} + +// aten::nonzero_static(Tensor self, *, SymInt size, int fill_value=-1) -> Tensor +inline at::Tensor Tensor::nonzero_static_symint(c10::SymInt size, int64_t fill_value) const { + return at::_ops::nonzero_static::call(const_cast(*this), size, fill_value); +} + +// aten::nonzero_numpy(Tensor self) -> Tensor[] +inline ::std::vector Tensor::nonzero_numpy() const { + return at::_ops::nonzero_numpy::call(const_cast(*this)); +} + +// aten::argwhere(Tensor self) -> Tensor +inline at::Tensor Tensor::argwhere() const { + return at::_ops::argwhere::call(const_cast(*this)); +} + +// aten::gather(Tensor self, int dim, Tensor index, *, bool sparse_grad=False) -> Tensor +inline at::Tensor Tensor::gather(int64_t dim, const at::Tensor & index, bool sparse_grad) const { + return at::_ops::gather::call(const_cast(*this), dim, index, sparse_grad); +} + +// aten::gather.dimname(Tensor self, Dimname dim, Tensor index, *, bool sparse_grad=False) -> Tensor +inline at::Tensor Tensor::gather(at::Dimname dim, const at::Tensor & index, bool sparse_grad) const { + return at::_ops::gather_dimname::call(const_cast(*this), dim, index, sparse_grad); +} + +// aten::addcmul(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor +inline at::Tensor Tensor::addcmul(const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value) const { + return at::_ops::addcmul::call(const_cast(*this), tensor1, tensor2, value); +} + +// aten::addcmul_(Tensor(a!) self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor(a!) +inline at::Tensor & Tensor::addcmul_(const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value) const { + return at::_ops::addcmul_::call(const_cast(*this), tensor1, tensor2, value); +} + +// aten::addcdiv(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor +inline at::Tensor Tensor::addcdiv(const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value) const { + return at::_ops::addcdiv::call(const_cast(*this), tensor1, tensor2, value); +} + +// aten::addcdiv_(Tensor(a!) self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor(a!) +inline at::Tensor & Tensor::addcdiv_(const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value) const { + return at::_ops::addcdiv_::call(const_cast(*this), tensor1, tensor2, value); +} + +// aten::triangular_solve(Tensor self, Tensor A, bool upper=True, bool transpose=False, bool unitriangular=False) -> (Tensor solution, Tensor cloned_coefficient) +inline ::std::tuple Tensor::triangular_solve(const at::Tensor & A, bool upper, bool transpose, bool unitriangular) const { + return at::_ops::triangular_solve::call(const_cast(*this), A, upper, transpose, unitriangular); +} + +// aten::svd(Tensor self, bool some=True, bool compute_uv=True) -> (Tensor U, Tensor S, Tensor V) +inline ::std::tuple Tensor::svd(bool some, bool compute_uv) const { + return at::_ops::svd::call(const_cast(*this), some, compute_uv); +} + +// aten::swapaxes(Tensor(a) self, int axis0, int axis1) -> Tensor(a) +inline at::Tensor Tensor::swapaxes(int64_t axis0, int64_t axis1) const { + return at::_ops::swapaxes::call(const_cast(*this), axis0, axis1); +} + +// aten::swapaxes_(Tensor(a!) self, int axis0, int axis1) -> Tensor(a!) +inline at::Tensor & Tensor::swapaxes_(int64_t axis0, int64_t axis1) const { + return at::_ops::swapaxes_::call(const_cast(*this), axis0, axis1); +} + +// aten::swapdims(Tensor(a) self, int dim0, int dim1) -> Tensor(a) +inline at::Tensor Tensor::swapdims(int64_t dim0, int64_t dim1) const { + return at::_ops::swapdims::call(const_cast(*this), dim0, dim1); +} + +// aten::swapdims_(Tensor(a!) self, int dim0, int dim1) -> Tensor(a!) +inline at::Tensor & Tensor::swapdims_(int64_t dim0, int64_t dim1) const { + return at::_ops::swapdims_::call(const_cast(*this), dim0, dim1); +} + +// aten::cholesky(Tensor self, bool upper=False) -> Tensor +inline at::Tensor Tensor::cholesky(bool upper) const { + return at::_ops::cholesky::call(const_cast(*this), upper); +} + +// aten::cholesky_solve(Tensor self, Tensor input2, bool upper=False) -> Tensor +inline at::Tensor Tensor::cholesky_solve(const at::Tensor & input2, bool upper) const { + return at::_ops::cholesky_solve::call(const_cast(*this), input2, upper); +} + +// aten::cholesky_inverse(Tensor self, bool upper=False) -> Tensor +inline at::Tensor Tensor::cholesky_inverse(bool upper) const { + return at::_ops::cholesky_inverse::call(const_cast(*this), upper); +} + +// aten::qr(Tensor self, bool some=True) -> (Tensor Q, Tensor R) +inline ::std::tuple Tensor::qr(bool some) const { + return at::_ops::qr::call(const_cast(*this), some); +} + +// aten::geqrf(Tensor self) -> (Tensor a, Tensor tau) +inline ::std::tuple Tensor::geqrf() const { + return at::_ops::geqrf::call(const_cast(*this)); +} + +// aten::orgqr(Tensor self, Tensor input2) -> Tensor +inline at::Tensor Tensor::orgqr(const at::Tensor & input2) const { + return at::_ops::orgqr::call(const_cast(*this), input2); +} + +// aten::ormqr(Tensor self, Tensor input2, Tensor input3, bool left=True, bool transpose=False) -> Tensor +inline at::Tensor Tensor::ormqr(const at::Tensor & input2, const at::Tensor & input3, bool left, bool transpose) const { + return at::_ops::ormqr::call(const_cast(*this), input2, input3, left, transpose); +} + +// aten::lu_solve(Tensor self, Tensor LU_data, Tensor LU_pivots) -> Tensor +inline at::Tensor Tensor::lu_solve(const at::Tensor & LU_data, const at::Tensor & LU_pivots) const { + return at::_ops::lu_solve::call(const_cast(*this), LU_data, LU_pivots); +} + +// aten::multinomial(Tensor self, SymInt num_samples, bool replacement=False, *, Generator? generator=None) -> Tensor +inline at::Tensor Tensor::multinomial(int64_t num_samples, bool replacement, ::std::optional generator) const { + return at::_ops::multinomial::call(const_cast(*this), num_samples, replacement, generator); +} + +// aten::multinomial(Tensor self, SymInt num_samples, bool replacement=False, *, Generator? generator=None) -> Tensor +inline at::Tensor Tensor::multinomial_symint(c10::SymInt num_samples, bool replacement, ::std::optional generator) const { + return at::_ops::multinomial::call(const_cast(*this), num_samples, replacement, generator); +} + +// aten::lgamma_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::lgamma_() const { + return at::_ops::lgamma_::call(const_cast(*this)); +} + +// aten::lgamma(Tensor self) -> Tensor +inline at::Tensor Tensor::lgamma() const { + return at::_ops::lgamma::call(const_cast(*this)); +} + +// aten::digamma(Tensor self) -> Tensor +inline at::Tensor Tensor::digamma() const { + return at::_ops::digamma::call(const_cast(*this)); +} + +// aten::polygamma(int n, Tensor self) -> Tensor +inline at::Tensor Tensor::polygamma(int64_t n) const { + return at::_ops::polygamma::call(n, const_cast(*this)); +} + +// aten::polygamma_(Tensor(a!) self, int n) -> Tensor(a!) +inline at::Tensor & Tensor::polygamma_(int64_t n) const { + return at::_ops::polygamma_::call(const_cast(*this), n); +} + +// aten::erfinv(Tensor self) -> Tensor +inline at::Tensor Tensor::erfinv() const { + return at::_ops::erfinv::call(const_cast(*this)); +} + +// aten::erfinv_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::erfinv_() const { + return at::_ops::erfinv_::call(const_cast(*this)); +} + +// aten::i0(Tensor self) -> Tensor +inline at::Tensor Tensor::i0() const { + return at::_ops::i0::call(const_cast(*this)); +} + +// aten::i0_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::i0_() const { + return at::_ops::i0_::call(const_cast(*this)); +} + +// aten::sign(Tensor self) -> Tensor +inline at::Tensor Tensor::sign() const { + return at::_ops::sign::call(const_cast(*this)); +} + +// aten::sign_(Tensor(a!) self) -> Tensor(a!) +inline at::Tensor & Tensor::sign_() const { + return at::_ops::sign_::call(const_cast(*this)); +} + +// aten::signbit(Tensor self) -> Tensor +inline at::Tensor Tensor::signbit() const { + return at::_ops::signbit::call(const_cast(*this)); +} + +// aten::dist(Tensor self, Tensor other, Scalar p=2) -> Tensor +inline at::Tensor Tensor::dist(const at::Tensor & other, const at::Scalar & p) const { + return at::_ops::dist::call(const_cast(*this), other, p); +} + +// aten::atan2_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::atan2_(const at::Tensor & other) const { + return at::_ops::atan2_::call(const_cast(*this), other); +} + +// aten::atan2(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::atan2(const at::Tensor & other) const { + return at::_ops::atan2::call(const_cast(*this), other); +} + +// aten::arctan2(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::arctan2(const at::Tensor & other) const { + return at::_ops::arctan2::call(const_cast(*this), other); +} + +// aten::arctan2_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::arctan2_(const at::Tensor & other) const { + return at::_ops::arctan2_::call(const_cast(*this), other); +} + +// aten::lerp.Scalar(Tensor self, Tensor end, Scalar weight) -> Tensor +inline at::Tensor Tensor::lerp(const at::Tensor & end, const at::Scalar & weight) const { + return at::_ops::lerp_Scalar::call(const_cast(*this), end, weight); +} + +// aten::lerp.Tensor(Tensor self, Tensor end, Tensor weight) -> Tensor +inline at::Tensor Tensor::lerp(const at::Tensor & end, const at::Tensor & weight) const { + return at::_ops::lerp_Tensor::call(const_cast(*this), end, weight); +} + +// aten::histc(Tensor self, int bins=100, Scalar min=0, Scalar max=0) -> Tensor +inline at::Tensor Tensor::histc(int64_t bins, const at::Scalar & min, const at::Scalar & max) const { + return at::_ops::histc::call(const_cast(*this), bins, min, max); +} + +// aten::histogram.bins_tensor(Tensor self, Tensor bins, *, Tensor? weight=None, bool density=False) -> (Tensor hist, Tensor bin_edges) +inline ::std::tuple Tensor::histogram(const at::Tensor & bins, const ::std::optional & weight, bool density) const { + return at::_ops::histogram_bins_tensor::call(const_cast(*this), bins, weight, density); +} + +// aten::histogram.bin_ct(Tensor self, int bins=100, *, float[]? range=None, Tensor? weight=None, bool density=False) -> (Tensor hist, Tensor bin_edges) +inline ::std::tuple Tensor::histogram(int64_t bins, ::std::optional> range, const ::std::optional & weight, bool density) const { + return at::_ops::histogram_bin_ct::call(const_cast(*this), bins, range, weight, density); +} + +// aten::fmod.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::fmod(const at::Scalar & other) const { + return at::_ops::fmod_Scalar::call(const_cast(*this), other); +} + +// aten::fmod_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::fmod_(const at::Scalar & other) const { + return at::_ops::fmod__Scalar::call(const_cast(*this), other); +} + +// aten::fmod.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::fmod(const at::Tensor & other) const { + return at::_ops::fmod_Tensor::call(const_cast(*this), other); +} + +// aten::fmod_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::fmod_(const at::Tensor & other) const { + return at::_ops::fmod__Tensor::call(const_cast(*this), other); +} + +// aten::hypot(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::hypot(const at::Tensor & other) const { + return at::_ops::hypot::call(const_cast(*this), other); +} + +// aten::hypot_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::hypot_(const at::Tensor & other) const { + return at::_ops::hypot_::call(const_cast(*this), other); +} + +// aten::igamma(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::igamma(const at::Tensor & other) const { + return at::_ops::igamma::call(const_cast(*this), other); +} + +// aten::igamma_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::igamma_(const at::Tensor & other) const { + return at::_ops::igamma_::call(const_cast(*this), other); +} + +// aten::igammac(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::igammac(const at::Tensor & other) const { + return at::_ops::igammac::call(const_cast(*this), other); +} + +// aten::igammac_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::igammac_(const at::Tensor & other) const { + return at::_ops::igammac_::call(const_cast(*this), other); +} + +// aten::nextafter(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::nextafter(const at::Tensor & other) const { + return at::_ops::nextafter::call(const_cast(*this), other); +} + +// aten::nextafter_(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::nextafter_(const at::Tensor & other) const { + return at::_ops::nextafter_::call(const_cast(*this), other); +} + +// aten::remainder.Scalar(Tensor self, Scalar other) -> Tensor +inline at::Tensor Tensor::remainder(const at::Scalar & other) const { + return at::_ops::remainder_Scalar::call(const_cast(*this), other); +} + +// aten::remainder_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!) +inline at::Tensor & Tensor::remainder_(const at::Scalar & other) const { + return at::_ops::remainder__Scalar::call(const_cast(*this), other); +} + +// aten::remainder.Tensor(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::remainder(const at::Tensor & other) const { + return at::_ops::remainder_Tensor::call(const_cast(*this), other); +} + +// aten::remainder_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!) +inline at::Tensor & Tensor::remainder_(const at::Tensor & other) const { + return at::_ops::remainder__Tensor::call(const_cast(*this), other); +} + +// aten::min(Tensor self) -> Tensor +inline at::Tensor Tensor::min() const { + return at::_ops::min::call(const_cast(*this)); +} + +// aten::fmin(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::fmin(const at::Tensor & other) const { + return at::_ops::fmin::call(const_cast(*this), other); +} + +// aten::max(Tensor self) -> Tensor +inline at::Tensor Tensor::max() const { + return at::_ops::max::call(const_cast(*this)); +} + +// aten::fmax(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::fmax(const at::Tensor & other) const { + return at::_ops::fmax::call(const_cast(*this), other); +} + +// aten::maximum(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::maximum(const at::Tensor & other) const { + return at::_ops::maximum::call(const_cast(*this), other); +} + +// aten::max.other(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::max(const at::Tensor & other) const { + return at::_ops::max_other::call(const_cast(*this), other); +} + +// aten::minimum(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::minimum(const at::Tensor & other) const { + return at::_ops::minimum::call(const_cast(*this), other); +} + +// aten::min.other(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::min(const at::Tensor & other) const { + return at::_ops::min_other::call(const_cast(*this), other); +} + +// aten::quantile(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor +inline at::Tensor Tensor::quantile(const at::Tensor & q, ::std::optional dim, bool keepdim, c10::string_view interpolation) const { + return at::_ops::quantile::call(const_cast(*this), q, dim, keepdim, interpolation); +} + +// aten::quantile.scalar(Tensor self, float q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor +inline at::Tensor Tensor::quantile(double q, ::std::optional dim, bool keepdim, c10::string_view interpolation) const { + return at::_ops::quantile_scalar::call(const_cast(*this), q, dim, keepdim, interpolation); +} + +// aten::nanquantile(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor +inline at::Tensor Tensor::nanquantile(const at::Tensor & q, ::std::optional dim, bool keepdim, c10::string_view interpolation) const { + return at::_ops::nanquantile::call(const_cast(*this), q, dim, keepdim, interpolation); +} + +// aten::nanquantile.scalar(Tensor self, float q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor +inline at::Tensor Tensor::nanquantile(double q, ::std::optional dim, bool keepdim, c10::string_view interpolation) const { + return at::_ops::nanquantile_scalar::call(const_cast(*this), q, dim, keepdim, interpolation); +} + +// aten::sort(Tensor self, int dim=-1, bool descending=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::sort(int64_t dim, bool descending) const { + return at::_ops::sort::call(const_cast(*this), dim, descending); +} + +// aten::sort.stable(Tensor self, *, bool? stable, int dim=-1, bool descending=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::sort(::std::optional stable, int64_t dim, bool descending) const { + return at::_ops::sort_stable::call(const_cast(*this), stable, dim, descending); +} + +// aten::sort.dimname(Tensor self, Dimname dim, bool descending=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::sort(at::Dimname dim, bool descending) const { + return at::_ops::sort_dimname::call(const_cast(*this), dim, descending); +} + +// aten::sort.dimname_stable(Tensor self, *, bool? stable, Dimname dim, bool descending=False) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::sort(::std::optional stable, at::Dimname dim, bool descending) const { + return at::_ops::sort_dimname_stable::call(const_cast(*this), stable, dim, descending); +} + +// aten::msort(Tensor self) -> Tensor +inline at::Tensor Tensor::msort() const { + return at::_ops::msort::call(const_cast(*this)); +} + +// aten::argsort(Tensor self, int dim=-1, bool descending=False) -> Tensor +inline at::Tensor Tensor::argsort(int64_t dim, bool descending) const { + return at::_ops::argsort::call(const_cast(*this), dim, descending); +} + +// aten::argsort.stable(Tensor self, *, bool stable, int dim=-1, bool descending=False) -> Tensor +inline at::Tensor Tensor::argsort(bool stable, int64_t dim, bool descending) const { + return at::_ops::argsort_stable::call(const_cast(*this), stable, dim, descending); +} + +// aten::argsort.dimname(Tensor self, Dimname dim, bool descending=False) -> Tensor +inline at::Tensor Tensor::argsort(at::Dimname dim, bool descending) const { + return at::_ops::argsort_dimname::call(const_cast(*this), dim, descending); +} + +// aten::topk(Tensor self, SymInt k, int dim=-1, bool largest=True, bool sorted=True) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::topk(int64_t k, int64_t dim, bool largest, bool sorted) const { + return at::_ops::topk::call(const_cast(*this), k, dim, largest, sorted); +} + +// aten::topk(Tensor self, SymInt k, int dim=-1, bool largest=True, bool sorted=True) -> (Tensor values, Tensor indices) +inline ::std::tuple Tensor::topk_symint(c10::SymInt k, int64_t dim, bool largest, bool sorted) const { + return at::_ops::topk::call(const_cast(*this), k, dim, largest, sorted); +} + +// aten::all(Tensor self) -> Tensor +inline at::Tensor Tensor::all() const { + return at::_ops::all::call(const_cast(*this)); +} + +// aten::any(Tensor self) -> Tensor +inline at::Tensor Tensor::any() const { + return at::_ops::any::call(const_cast(*this)); +} + +// aten::renorm(Tensor self, Scalar p, int dim, Scalar maxnorm) -> Tensor +inline at::Tensor Tensor::renorm(const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm) const { + return at::_ops::renorm::call(const_cast(*this), p, dim, maxnorm); +} + +// aten::renorm_(Tensor(a!) self, Scalar p, int dim, Scalar maxnorm) -> Tensor(a!) +inline at::Tensor & Tensor::renorm_(const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm) const { + return at::_ops::renorm_::call(const_cast(*this), p, dim, maxnorm); +} + +// aten::unfold(Tensor(a) self, int dimension, int size, int step) -> Tensor(a) +inline at::Tensor Tensor::unfold(int64_t dimension, int64_t size, int64_t step) const { + return at::_ops::unfold::call(const_cast(*this), dimension, size, step); +} + +// aten::equal(Tensor self, Tensor other) -> bool +inline bool Tensor::equal(const at::Tensor & other) const { + return at::_ops::equal::call(const_cast(*this), other); +} + +// aten::pow.Tensor_Tensor(Tensor self, Tensor exponent) -> Tensor +inline at::Tensor Tensor::pow(const at::Tensor & exponent) const { + return at::_ops::pow_Tensor_Tensor::call(const_cast(*this), exponent); +} + +// aten::pow.Tensor_Scalar(Tensor self, Scalar exponent) -> Tensor +inline at::Tensor Tensor::pow(const at::Scalar & exponent) const { + return at::_ops::pow_Tensor_Scalar::call(const_cast(*this), exponent); +} + +// aten::pow_.Scalar(Tensor(a!) self, Scalar exponent) -> Tensor(a!) +inline at::Tensor & Tensor::pow_(const at::Scalar & exponent) const { + return at::_ops::pow__Scalar::call(const_cast(*this), exponent); +} + +// aten::pow_.Tensor(Tensor(a!) self, Tensor exponent) -> Tensor(a!) +inline at::Tensor & Tensor::pow_(const at::Tensor & exponent) const { + return at::_ops::pow__Tensor::call(const_cast(*this), exponent); +} + +// aten::float_power.Tensor_Tensor(Tensor self, Tensor exponent) -> Tensor +inline at::Tensor Tensor::float_power(const at::Tensor & exponent) const { + return at::_ops::float_power_Tensor_Tensor::call(const_cast(*this), exponent); +} + +// aten::float_power.Tensor_Scalar(Tensor self, Scalar exponent) -> Tensor +inline at::Tensor Tensor::float_power(const at::Scalar & exponent) const { + return at::_ops::float_power_Tensor_Scalar::call(const_cast(*this), exponent); +} + +// aten::float_power_.Scalar(Tensor(a!) self, Scalar exponent) -> Tensor(a!) +inline at::Tensor & Tensor::float_power_(const at::Scalar & exponent) const { + return at::_ops::float_power__Scalar::call(const_cast(*this), exponent); +} + +// aten::float_power_.Tensor(Tensor(a!) self, Tensor exponent) -> Tensor(a!) +inline at::Tensor & Tensor::float_power_(const at::Tensor & exponent) const { + return at::_ops::float_power__Tensor::call(const_cast(*this), exponent); +} + +// aten::normal_(Tensor(a!) self, float mean=0, float std=1, *, Generator? generator=None) -> Tensor(a!) +inline at::Tensor & Tensor::normal_(double mean, double std, ::std::optional generator) const { + return at::_ops::normal_::call(const_cast(*this), mean, std, generator); +} + +// aten::alias(Tensor(a) self) -> Tensor(a) +inline at::Tensor Tensor::alias() const { + return at::_ops::alias::call(const_cast(*this)); +} + +// aten::isfinite(Tensor self) -> Tensor +inline at::Tensor Tensor::isfinite() const { + return at::_ops::isfinite::call(const_cast(*this)); +} + +// aten::isinf(Tensor self) -> Tensor +inline at::Tensor Tensor::isinf() const { + return at::_ops::isinf::call(const_cast(*this)); +} + +// aten::record_stream(Tensor(a!) self, Stream s) -> () +inline void Tensor::record_stream(at::Stream s) const { + return at::_ops::record_stream::call(const_cast(*this), s); +} + +// aten::isposinf(Tensor self) -> Tensor +inline at::Tensor Tensor::isposinf() const { + return at::_ops::isposinf::call(const_cast(*this)); +} + +// aten::isneginf(Tensor self) -> Tensor +inline at::Tensor Tensor::isneginf() const { + return at::_ops::isneginf::call(const_cast(*this)); +} + +// aten::det(Tensor self) -> Tensor +inline at::Tensor Tensor::det() const { + return at::_ops::det::call(const_cast(*this)); +} + +// aten::slogdet(Tensor self) -> (Tensor sign, Tensor logabsdet) +inline ::std::tuple Tensor::slogdet() const { + return at::_ops::slogdet::call(const_cast(*this)); +} + +// aten::logdet(Tensor self) -> Tensor +inline at::Tensor Tensor::logdet() const { + return at::_ops::logdet::call(const_cast(*this)); +} + +// aten::inverse(Tensor self) -> Tensor +inline at::Tensor Tensor::inverse() const { + return at::_ops::inverse::call(const_cast(*this)); +} + +// aten::inner(Tensor self, Tensor other) -> Tensor +inline at::Tensor Tensor::inner(const at::Tensor & other) const { + return at::_ops::inner::call(const_cast(*this), other); +} + +// aten::outer(Tensor self, Tensor vec2) -> Tensor +inline at::Tensor Tensor::outer(const at::Tensor & vec2) const { + return at::_ops::outer::call(const_cast(*this), vec2); +} + +// aten::ger(Tensor self, Tensor vec2) -> Tensor +inline at::Tensor Tensor::ger(const at::Tensor & vec2) const { + return at::_ops::ger::call(const_cast(*this), vec2); +} + +// aten::to_padded_tensor(Tensor self, float padding, SymInt[]? output_size=None) -> Tensor +inline at::Tensor Tensor::to_padded_tensor(double padding, at::OptionalIntArrayRef output_size) const { + return at::_ops::to_padded_tensor::call(const_cast(*this), padding, output_size.has_value() ? ::std::make_optional(c10::fromIntArrayRefSlow(*output_size)) : ::std::nullopt); +} + +// aten::to_padded_tensor(Tensor self, float padding, SymInt[]? output_size=None) -> Tensor +inline at::Tensor Tensor::to_padded_tensor_symint(double padding, at::OptionalSymIntArrayRef output_size) const { + return at::_ops::to_padded_tensor::call(const_cast(*this), padding, output_size); +} +} // namespace at + + +namespace c10 { +template <> +struct MaybeOwnedTraits { + using owned_type = at::Tensor; + using borrow_type = at::Tensor; + + static borrow_type createBorrow(const owned_type& from) { + // NOTE: this can be implemented without the special + // unsafe_borrow_t Tensor constructor as + // + // return borrow_type(c10::intrusive_ptr::reclaim(from.unsafeGetTensorImpl())); + // + // but that hurts inlining due to the nullptr check in the + // Tensor(c10::intrusive_ptr<...>) constructor. We already know + // that from.impl_ isn't null because from is a valid Tensor, so + // we needn't do the check again. (using __builtin_assume can + // avoid this, but wouldn't be portable to MSVC.) + return borrow_type(borrow_type::unsafe_borrow_t{}, from); + } + + static void assignBorrow(borrow_type& lhs, const borrow_type& rhs) { + lhs.unsafeReleaseTensorImpl(); + // See above note: this can be implemented with public API + // similarly to createBorrow(), but that would hurt inlining. + lhs = borrow_type(borrow_type::unsafe_borrow_t{}, rhs); + } + + static void destroyBorrow(borrow_type& toDestroy) { + toDestroy.unsafeReleaseTensorImpl(); // "leak" it, but it was already +0. + } + + static const owned_type& referenceFromBorrow(const borrow_type& borrow) { + return borrow; + } + + static const owned_type* pointerFromBorrow(const borrow_type& borrow) { + return &borrow; + } + + static bool debugBorrowIsValid(const borrow_type& /*borrow*/) { + return true; + } +}; + +template <> +struct ExclusivelyOwnedTraits { + using repr_type = at::Tensor; + using pointer_type = at::Tensor*; + using const_pointer_type = const at::Tensor*; + + static repr_type nullRepr() { + return at::Tensor(); + } + + template + static repr_type createInPlace(Args&&... args) { + return at::Tensor(std::forward(args)...); + } + + static repr_type moveToRepr(at::Tensor&& x) { + return std::move(x); + } + + static void destroyOwned(at::Tensor& x) { + return ExclusivelyOwnedTraits::destroyOwned(x); + } + + static at::Tensor take(at::Tensor& x) { + return std::move(x); + } + + static pointer_type getImpl(repr_type& x) { + return &x; + } + + static const_pointer_type getImpl(const repr_type& x) { + return &x; + } +}; +} // namespace c10 + +namespace at { + +inline c10::MaybeOwned borrow_from_optional_tensor( + const std::optional& opt) { + return opt.has_value() + ? c10::MaybeOwned::borrowed(*opt) + : c10::MaybeOwned::owned(std::in_place); +} + +inline c10::MaybeOwned Tensor::expect_contiguous(MemoryFormat memory_format) const & { + if (is_contiguous(memory_format)) { + return c10::MaybeOwned::borrowed(*this); + } else { + return c10::MaybeOwned::owned(__dispatch_contiguous(memory_format)); + } +} +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/blob.h b/backends/pytorch/include/ATen/core/blob.h new file mode 100644 index 000000000..f2e5e419a --- /dev/null +++ b/backends/pytorch/include/ATen/core/blob.h @@ -0,0 +1,209 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include + +#include +#include +#include + +namespace caffe2 { + +class Tensor; + +/** + * @brief Blob is a general container that hosts a typed pointer. + * + * A Blob hosts a pointer as well as its type, and takes charge of deleting it + * properly when the blob is deallocated or re-allocated with a new type. A blob + * could contain anything, although the most common case is to contain a Tensor. + */ +class TORCH_API Blob final : public c10::intrusive_ptr_target { + public: + /** + * Initializes an empty Blob. + */ + Blob() noexcept = default; + ~Blob() override { + Reset(); + } + + Blob(Blob&& other) noexcept : Blob() { + swap(other); + } + + Blob& operator=(Blob&& other) noexcept { + Blob(std::move(other)).swap(*this); + return *this; + } + + /** + * Checks if the content stored in the blob is of type T. + */ + template + bool IsType() const noexcept { + return meta_.Match(); + } + + /** + * Returns the meta info of the blob. + */ + const TypeMeta meta() const noexcept { + return meta_; + } + + /** + * Returns a printable typename of the blob. + */ + std::string_view TypeName() const noexcept { + return meta_.name(); + } + + /** + * @brief Gets the const reference of the stored object. The code checks if + * the stored object is of the desired type. + */ + // TODO(jerryzh): add a Get(c10::DeviceType) function? + template + const T& Get() const { + TORCH_INTERNAL_ASSERT( + IsType(), + "wrong type for the Blob instance. Blob contains ", + meta_.name(), + " while caller expects ", + TypeMeta::TypeName()); + // TODO: after we add Get(c10::DeviceType) + // and changed all the callsites, we can add + // a static assert here to enforce T != Tensor + return *static_cast(pointer_); + } + + const void* GetRaw() const noexcept { + return pointer_; + } + void* GetRaw() noexcept { + return pointer_; + } + + /** + * @brief Gets a mutable pointer to the stored object. + * + * If the current object is not of the right type, a new object is created + * and the old object is freed. Note that type T should have a default + * constructor. Otherwise, create the object yourself first, and use + * Reset(). + */ + template + T* GetMutable() { + static_assert( + std::is_default_constructible_v, + "GetMutable can't be called with non-default-constructible types. " + "Try using specialized methods"); + if (IsType()) { + return static_cast(pointer_); + } else { + // TODO Re-enable logging + // VLOG(1) << "Create new mutable object " << TypeMeta::TypeName(); + return Reset(new T()); + } + } + + template + T* GetMutableOrNull() { + if (IsType()) { + return static_cast(pointer_); + } else { + return nullptr; + } + } + + /** + * Sets the underlying object to the allocated one. The Blob then takes over + * the ownership of the passed in pointer. If there is already an object in + * the Blob, the old object is freed. + * + * This is used when the underlying class T does not have a default ctor, or + * complex initializations needs to be done outside the blob. + */ + template + T* Reset(T* allocated) { + free_(); + meta_ = TypeMeta::Make(); + pointer_ = static_cast(allocated); + has_ownership_ = true; + return allocated; + } + + /** + * Sets the underlying object to the allocated one, but does not take over + * the ownership of the passed in pointer. If there is already an object in + * the Blob, the old object is freed. + * + * Unlike Reset, this does not take over the ownership of the pointer and the + * caller is responsible for making sure that the lifetime of the allocated + * blob outlasts the lifetime of any access to this blob, until another Reset + * call is made or the blob is destructed. + */ + template + std::remove_const_t* ShareExternal( + std::remove_const_t* allocated) { + return static_cast(ShareExternal( + static_cast(allocated), + TypeMeta::Make>())); + } + + void* ShareExternal(void* allocated, const TypeMeta meta) { + free_(); + meta_ = meta; + pointer_ = allocated; + has_ownership_ = false; + return allocated; + } + + /** + * Resets the Blob to an empty one. + */ + void Reset() { + free_(); + pointer_ = nullptr; + meta_ = TypeMeta(); + has_ownership_ = false; + } + + /** + * @brief Swaps the underlying storage of two blobs. + */ + void swap(Blob& rhs) noexcept { + using std::swap; + swap(meta_, rhs.meta_); + swap(pointer_, rhs.pointer_); + swap(has_ownership_, rhs.has_ownership_); + } + + private: + void free_() { + if (has_ownership_ && pointer_ != nullptr) { + (*meta_.deleteFn())(pointer_); + } + } + + TypeMeta meta_; + void* pointer_{nullptr}; + bool has_ownership_{false}; + + C10_DISABLE_COPY_AND_ASSIGN(Blob); +}; + +inline void swap(Blob& lhs, Blob& rhs) noexcept { + lhs.swap(rhs); +} + +inline std::ostream& operator<<(std::ostream& out, const Blob& v) { + return out << "Blob[" << v.TypeName() << ']'; +} + +} // namespace caffe2 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/custom_class.h b/backends/pytorch/include/ATen/core/custom_class.h new file mode 100644 index 000000000..2811256fb --- /dev/null +++ b/backends/pytorch/include/ATen/core/custom_class.h @@ -0,0 +1,33 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include +#include +#include + +namespace c10 { + +struct ClassType; +using ClassTypePtr = std::shared_ptr; + +TORCH_API c10::ClassTypePtr getCustomClassTypeImpl(const std::type_index &tindex); + +template +const c10::ClassTypePtr& getCustomClassType() { + // Classes are never unregistered from getCustomClassTypeMap and the + // hash lookup can be a hot path, so just cache. + // For the same reason, it's fine If this ends up getting duplicated across + // DSO boundaries for whatever reason. + static c10::ClassTypePtr cache = getCustomClassTypeImpl( + std::type_index(typeid(T))); + return cache; +} + +} + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/dynamic_type.h b/backends/pytorch/include/ATen/core/dynamic_type.h new file mode 100644 index 000000000..9d44d9ae8 --- /dev/null +++ b/backends/pytorch/include/ATen/core/dynamic_type.h @@ -0,0 +1,252 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +#include +#include + +namespace c10 { + +using DynamicTypeBits = std::uint32_t; +#define DYNAMIC_TYPE_BIT(x) (1u << x) + +constexpr DynamicTypeBits kDynamicCovariantTypeBit = DYNAMIC_TYPE_BIT(31); +constexpr DynamicTypeBits kDynamicAnyTypeBit = DYNAMIC_TYPE_BIT(30); + +constexpr DynamicTypeBits kDynamicNoneTypeBit = DYNAMIC_TYPE_BIT(1); +constexpr DynamicTypeBits kDynamicIntTypeBit = DYNAMIC_TYPE_BIT(3); +constexpr DynamicTypeBits kDynamicFloatTypeBit = DYNAMIC_TYPE_BIT(4); +constexpr DynamicTypeBits kDynamicComplexTypeBit = DYNAMIC_TYPE_BIT(5); +constexpr DynamicTypeBits kDynamicListTypeBit = DYNAMIC_TYPE_BIT(7); +constexpr DynamicTypeBits kDynamicTupleTypeBit = DYNAMIC_TYPE_BIT(8); +constexpr DynamicTypeBits kDynamicClassTypeBit = DYNAMIC_TYPE_BIT(10); + +#define FORALL_DYNAMIC_TYPES(_) \ + _(Tensor, DYNAMIC_TYPE_BIT(0), 1) \ + _(None, kDynamicNoneTypeBit, 1) \ + _(Bool, DYNAMIC_TYPE_BIT(2), 1) \ + _(Int, kDynamicIntTypeBit, 1) \ + _(Float, kDynamicFloatTypeBit, 1) \ + _(Complex, kDynamicComplexTypeBit, 1) \ + _(Number, \ + (kDynamicIntTypeBit | kDynamicFloatTypeBit | kDynamicComplexTypeBit), \ + 1) \ + _(String, DYNAMIC_TYPE_BIT(6), 1) \ + _(List, kDynamicListTypeBit, 0) \ + _(Tuple, (kDynamicTupleTypeBit | kDynamicCovariantTypeBit), 0) \ + _(Dict, DYNAMIC_TYPE_BIT(9), 0) \ + _(Class, kDynamicClassTypeBit, 0) \ + _(Optional, \ + (DYNAMIC_TYPE_BIT(11) | kDynamicNoneTypeBit | kDynamicCovariantTypeBit), \ + 0) \ + _(AnyList, (kDynamicListTypeBit | kDynamicAnyTypeBit), 1) \ + _(AnyTuple, \ + (kDynamicTupleTypeBit | kDynamicCovariantTypeBit | kDynamicAnyTypeBit), \ + 1) \ + _(DeviceObj, DYNAMIC_TYPE_BIT(12), 1) \ + _(StreamObj, DYNAMIC_TYPE_BIT(13), 1) \ + _(Capsule, DYNAMIC_TYPE_BIT(14), 1) \ + _(Generator, DYNAMIC_TYPE_BIT(15), 1) \ + _(Storage, DYNAMIC_TYPE_BIT(16), 1) \ + _(Var, DYNAMIC_TYPE_BIT(17), 0) \ + _(AnyClass, (kDynamicClassTypeBit | kDynamicAnyTypeBit), 1) \ + _(QScheme, DYNAMIC_TYPE_BIT(18), 1) \ + _(Quantizer, DYNAMIC_TYPE_BIT(19), 1) \ + _(AnyEnum, DYNAMIC_TYPE_BIT(20), 1) \ + _(RRef, DYNAMIC_TYPE_BIT(21), 0) \ + _(Future, DYNAMIC_TYPE_BIT(22), 0) \ + _(Await, DYNAMIC_TYPE_BIT(23), 0) \ + _(Any, 0xffffffff, 1) + +#define FORALL_DYNAMIC_TYPES_FAKE(_) \ + _(ScalarType, kDynamicIntTypeBit, 1) \ + _(Layout, kDynamicIntTypeBit, 1) \ + _(SymInt, kDynamicIntTypeBit, 1) \ + _(SymBool, kDynamicIntTypeBit, 1) \ + _(MemoryFormat, kDynamicIntTypeBit, 1) + +#define FORWARD_DECL_TYPE(NAME, _, __) struct NAME ## Type; + FORALL_DYNAMIC_TYPES(FORWARD_DECL_TYPE) + FORALL_DYNAMIC_TYPES_FAKE(FORWARD_DECL_TYPE) +#undef FORWARD_DECL_TYPE + +class DynamicType; +using DynamicTypePtr = std::shared_ptr; + +/** + * DynamicType is designed as a low dependency type system for TorchScript. The + * existing JIT types are used for both compilation and runtime, which makes + * sense for server contexts because we often compile and run the model in + * the same process, however this doesn't hold for mobile devices where we + * always compiles a model ahead of time, therefore there will be dependencies + * which are not needed, but built with mobile runtime causing binary size + * bloat, by design. Every basic type like Int, Bool or String will bring their + * vtable, typeinfo, constructor, destructor and even more data from their + * specializations for STL types to the binary causing a long tail bloat. + * + * The core problem is about the complexity to implement and maintain a single + * type system for both analysis and execution purposes. Although they should + * have the exactly same semantics, in practice implement a unified abstraction + * adds conceptual and representational overhead for both sides of the world. + * + * To address the issues, DynamicType implements a minimal subset of JIT types + * and uses a generic algorithm to test all subtyping relations. To achieve + * this, we assign each dynamic type a single integer tag to represent its + * semantics. More specifically, a dynamic type is defined as a set of "control + * bits" and "data bits", where control bits describe the special behavior when + * testing a type and data bits map to identity of each nominal type. We use bit + * operations to perform all the tests. + * + * For example, a "covariant bit" is a control bit used to describe if a type + * is covariant, right now the most used one is tuple type, and in addition to + * the control bit, tuple type's data bit is the 8th bit from the LSB. Control + * bits start from MSB and data bits start from LSB. + * + * If two types are equal, then they are subtype of each other, also if the bits + * from one type tag is subset of the other tag, it automatically becomes a + * subtype of the other. This simplifies the subtyping logic a lot, and over the + * long term it is possible to adopt this scheme on the server side as well. + * Special cases can be added but they generally should not take too much code + * size. + * + * DynamicType may or may not inherit from c10::Type because it's not the core + * requirement of DynamicType to interface with existing JIT types, but we might + * want to inherit from c10::Type to reduce the migration cost. + */ +class DynamicType : public SharedType { + using ClassTypePtr = std::shared_ptr; + + /** + * A implementation detail to support NamedTuple. + */ + struct LabeledDynamicType { + std::optional label; + DynamicTypePtr ty; + explicit LabeledDynamicType(DynamicTypePtr t) : ty(std::move(t)) {} + + bool equals(const LabeledDynamicType& other) const; + bool isSubtypeOf(const LabeledDynamicType& other) const; + }; + + public: + // TODO Change Ptr to DynamicTypePtr when all migrations are done. + using Ptr = TypePtr; + using ElementType = DynamicType; + ~DynamicType() override; + + struct Arguments { + Arguments() = default; + Arguments(c10::ArrayRef /*args*/); + Arguments(const std::vector& /*names*/, c10::ArrayRef /*args*/); + std::vector elems; + }; + + enum class Tag : DynamicTypeBits { +#define DYNAMIC_TYPE_ITEM(NAME, VAL, _) NAME = VAL, + FORALL_DYNAMIC_TYPES(DYNAMIC_TYPE_ITEM) + FORALL_DYNAMIC_TYPES_FAKE(DYNAMIC_TYPE_ITEM) +#undef DYNAMIC_TYPE_ITEM + }; + + bool equals(const Type& rhs) const override; + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override; + std::string str() const override; + static const TypeKind Kind = TypeKind::DynamicType; + static TORCH_API DynamicTypePtr create(Type& ty); + + explicit DynamicType(Tag /*tag*/, Arguments /*arguments*/); + explicit DynamicType(Tag /*tag*/, std::string_view /*name*/, Arguments /*arguments*/); + + DynamicType(DynamicType&& other) = delete; + DynamicType(const DynamicType&) = delete; + DynamicType& operator=(const DynamicType&) = delete; + DynamicType& operator=(DynamicType&&) = delete; + + TypePtr containedType(size_t /*i*/) const override; + size_t containedTypeSize() const override; + Tag tag() const { + return tag_; + } + const std::optional& name() const { + return name_; + } + const Arguments& arguments() const { + return arguments_; + } + TORCH_API TypeKind dynamicKind() const; + + // Should be used only on the server side to restore static type information. +#ifndef C10_MOBILE + TORCH_API +#endif + TypePtr fallback() const; + + private: + bool symmetric() const override { + return false; + } + friend struct Type; + // NOTE: Here we are using SingletonOrSharedTypePtr to mean + // "original-type-because-it-was-actually-a-DynamicType or shared". + static SingletonOrSharedTypePtr create(const Type& ty); + DynamicType(const Type& other); + bool equals(const DynamicType& other) const; + + template + bool compareArguments(const DynamicType& other, const F& f) const { + if (arguments_.elems.size() != other.arguments_.elems.size()) { + return false; + } + for (size_t i = 0; i < arguments_.elems.size(); i++) { + if (!f(arguments_.elems[i], other.arguments_.elems[i])) { + return false; + } + } + return true; + } + + Tag tag_; + std::optional name_; + union { + Arguments arguments_; + ClassTypePtr class_; + }; +}; + +template +struct DynamicTypeTrait { + C10_NOINLINE static auto tagValue() { + TORCH_CHECK(false); + return DynamicType::Tag::Any; + } +}; + +namespace detail { +C10_NOINLINE DynamicTypePtr makeBaseType(DynamicType::Tag tag); +} + +#define DYNAMIC_TYPE_TAG_VALUE(NAME, _, IS_BASE_TYPE) \ + template <> \ + struct TORCH_API DynamicTypeTrait { \ + C10_ERASE static auto tagValue() { \ + return DynamicType::Tag::NAME; \ + } \ + static constexpr bool isBaseType = IS_BASE_TYPE; \ + template \ + static std::enable_if_t getBaseType() { \ + static auto type = detail::makeBaseType(tagValue()); \ + return type; \ + } \ + }; // namespace c10 +FORALL_DYNAMIC_TYPES(DYNAMIC_TYPE_TAG_VALUE) +FORALL_DYNAMIC_TYPES_FAKE(DYNAMIC_TYPE_TAG_VALUE) +#undef DYNAMIC_TYPE_TAG_VALUE + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/functional.h b/backends/pytorch/include/ATen/core/functional.h new file mode 100644 index 000000000..c5bf67656 --- /dev/null +++ b/backends/pytorch/include/ATen/core/functional.h @@ -0,0 +1,59 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace c10 { + +// The passed in function must take T by value (T), or by +// const reference (const T&); taking T by non-const reference +// will result in an error like: +// +// error: no type named 'type' in 'class std::invoke_result' +// +// No explicit template parameters are required. + +// Overload for explicit function and ArrayRef +template +inline auto fmap(const T& inputs, const F& fn) -> std::vector { + std::vector r; + r.reserve(inputs.size()); + for(const auto & input : inputs) + r.push_back(fn(input)); + return r; +} + +// C++ forbids taking an address of a constructor, so here's a workaround... +// Overload for constructor (R) application +template +inline std::vector fmap(const T& inputs) { + std::vector r; + r.reserve(inputs.size()); + for(auto & input : inputs) + r.push_back(R(input)); + return r; +} + +template +inline std::vector filter(at::ArrayRef inputs, const F& fn) { + std::vector r; + r.reserve(inputs.size()); + for(auto & input : inputs) { + if (fn(input)) { + r.push_back(input); + } + } + return r; +} + +template +inline std::vector filter(const std::vector& inputs, const F& fn) { + return filter(static_cast>(inputs), fn); +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/ivalue.h b/backends/pytorch/include/ATen/core/ivalue.h new file mode 100644 index 000000000..1c0d3221f --- /dev/null +++ b/backends/pytorch/include/ATen/core/ivalue.h @@ -0,0 +1,1642 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default") + +namespace torch { +class TORCH_API CustomClassHolder : public c10::intrusive_ptr_target {}; +namespace jit { +using ::torch::CustomClassHolder; +struct Function; +struct CompilationUnit; +struct Module; +} // namespace jit +} // namespace torch +namespace c10 { +template +class Dict; +template +class List; +template +class IListRef; +struct IValue; +struct ClassType; +struct Type; +class RRefInterface; + +struct ClassType; +using ClassTypePtr = std::shared_ptr; + +TORCH_API bool _fastEqualsForContainer(const IValue& lhs, const IValue& rhs); + +TORCH_API torch::jit::Function* checkObjectSortSchema( + const c10::ClassTypePtr& t, + std::stringstream& why_not); + +// A comparator that checks ordering of two IValues of same type. +typedef std::function IValueComparator; + +TORCH_API IValueComparator getLessThanComparator(const IValue& v); +TORCH_API IValueComparator getGreaterThanComparator(const IValue& v); + +namespace ivalue { +struct Tuple; +struct Future; +struct Await; +struct ConstantString; +struct GenericDict; +struct Object; +struct PyObjectHolder; +struct EnumHolder; +// We need a ComplexHolder because currently the payloads in the Union +// only take 64 bits. Since ComplexDouble takes up 128 bits, and is too big +// to fit in the IValue directly, we indirect complex numbers through an +// intrusive pointer to ComplexHolder (which contains a c10::complex). +struct ComplexHolder : c10::intrusive_ptr_target { + public: + template + ComplexHolder(c10::complex c) { + val = convert>(c); + } + ComplexHolder() = default; + c10::complex val; +}; + +// Similar to ComplexHolder, for StreamData3 +struct StreamData3Holder : c10::intrusive_ptr_target { + public: + StreamData3Holder(struct c10::StreamData3 d) : val(d) {} + StreamData3Holder() = delete; + struct c10::StreamData3 val; +}; + +} // namespace ivalue + +// This is an owning wrapper for a std::optional> +// that can be implicitly converted to a (non-owning) std::optional>. +// Its purpose is to be used in generated code to keep the vector alive +// either until the end of a statement (as a temporary), or as a saved arg +// in autograd. +template +struct OptionalArray { + std::optional> list; + + OptionalArray() = default; + OptionalArray(std::vector val) : list(std::move(val)) {} + + // Used when saving an argument for the backwards pass. + OptionalArray& operator=(std::optional> ref) { + if (ref) { + list = std::vector(ref->begin(), ref->end()); + } else { + list = std::nullopt; + } + return *this; + } + + // Used when saving an argument for the backwards pass. + OptionalArray& operator=(c10::OptionalArrayRef ref) { + if (ref) { + list = std::vector(ref->begin(), ref->end()); + } else { + list = std::nullopt; + } + return *this; + } + + operator std::optional>() { + if (!list) { + return std::nullopt; + } + return *list; + } + + operator c10::OptionalArrayRef() { + if (!list) { + return std::nullopt; + } + return *list; + } +}; + +// Capsule is an internal implementation detail of custom C++ classes. We +// define it as an owning wrapper for +// c10::intrusive_ptr This wrapper is here to serve as +// an abstraction of the type erased custom class object pointer. It also allow +// pybind11 to treat this as a standalone class to register as a separate type +// caster, instead of a custom pointer holder which the pointer holder type +// caster try to "unwrap" it automatically. +struct Capsule { + c10::intrusive_ptr obj_ptr; + explicit Capsule(c10::intrusive_ptr ptr) + : obj_ptr(std::move(ptr)) {} +}; + +// IValue is the generic tagged union used by the interpreter to hold +// all value types. +// It is a 16-byte object with an 8-byte payload and an 8-byte tag. +// The tag is currently 4 bytes to determine the type, and 1 byte +// to mark whether that type is a subtype of c10::intrusive_ptr_target and needs +// retain/release calls. + +#define TORCH_FORALL_TAGS(_) \ + _(None) \ + _(Tensor) \ + _(Storage) \ + _(Double) \ + _(ComplexDouble) \ + _(Int) \ + _(UInt) \ + _(SymInt) \ + _(SymFloat) \ + _(SymBool) \ + _(Bool) \ + _(Tuple) \ + _(String) \ + _(Blob) \ + _(GenericList) \ + _(GenericDict) \ + _(Future) \ + _(Await) \ + _(Device) \ + _(Stream) \ + _(Object) \ + _(PyObject) \ + _(Uninitialized) \ + _(Capsule) \ + _(RRef) \ + _(Quantizer) \ + _(Generator) \ + _(Enum) + +// [doxygen private] +// These methods are not actually private but we don't want to document them, so +// they are marked `@private`, which hides them on the doxygen documentation for +// this page. + +/// IValue (Interpreter Value) is a tagged union over the types +/// supported by the TorchScript interpreter. IValues contain their +/// values as an `IValue::Payload`, which holds primitive types +/// (`int64_t`, `bool`, `double`, `Device`) and `Tensor` as values, +/// and all other types as a `c10::intrusive_ptr`. In order to +/// optimize performance of the destructor and related operations by +/// making the `Tensor` and `c10::intrusive_ptr` paths generate the +/// same code, we represent a null `c10::intrusive_ptr` as +/// `UndefinedTensorImpl::singleton()`, *not* `nullptr`. +/// +/// IValues are used as inputs to and outputs from the TorchScript interpreter. +/// To retrieve the value contained within an IValue, use the `.toX()` methods, +/// where `X` is the type you are trying to get. Note that neither the `.toX()` +/// methods nor the templated `.to` functions do any kind of casting, they +/// only unwrap the contained value. For example: +/// +/// \rst +/// .. code-block:: cpp +/// +/// // Make the IValue +/// torch::IValue my_ivalue(26); +/// std::cout << my_ivalue << "\n"; +/// +/// // Unwrap the IValue +/// int64_t my_int = my_ivalue.toInt(); +/// std::cout << my_int << "\n"; +/// +/// // This will throw an error! +/// // `my_ivalue` is tagged as an int and cannot be used as another type +/// torch::Tensor my_tensor = my_ivalue.toTensor(); +/// \endrst +struct TORCH_API IValue final { + IValue(const IValue& rhs) : IValue(rhs.payload, rhs.tag) { + if (isIntrusivePtr() && + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton()) { + c10::raw::intrusive_ptr::incref(payload.u.as_intrusive_ptr); + } + } + + IValue(IValue&& rhs) noexcept : tag(rhs.tag) { + moveFrom(std::move(rhs)); + } + + /// @private [doxygen private] + ~IValue() { + destroy(); + } + + C10_ALWAYS_INLINE IValue& operator=(IValue&& rhs) & noexcept { + if (&rhs == this) { + return *this; + } + + destroy(); + moveFrom(std::move(rhs)); + return *this; + } + + IValue& operator=(IValue const& rhs) & { + *this = IValue(rhs); + return *this; + } + + void dump() const; + + /** + * Equality comparison. The semantics are the same as Python's `==`: + * 1. Numerical types are compared by value. + * 2. Tensors compute element-wise equality, returning a BoolTensor (see: + * `torch.eq()`) + * 3. Strings are compared by value. + * 4. Sequence types (list, tuple) are compared lexicographically by + * comparing their elements. Different sequence types never compare equal. + * 5. Mappings (dict) must have equal (key, value) pairs. + * 6. If not listed above, the default behavior for is to test identity + * equality (e.g. pointer equality). + * + * Why does this return an IValue instead of a bool? Because in PyTorch, + * `tensor1 == tensor2` returns a `BoolTensor`, not a bool. + * + * NOTE: we (like Python) assume that identity equality implies value equality + * for efficiency. + * TODO: need to support customizing equality + */ + IValue equals(const IValue& rhs) const; + /** + * This implements the same semantics as `bool(lhs == rhs)` in Python. which + * is the same as `equals()` except for Tensor types. + */ + TORCH_API friend bool operator==(const IValue& lhs, const IValue& rhs); + TORCH_API friend bool operator!=(const IValue& lhs, const IValue& rhs); + + /** + * Identity comparison. Checks if `this` is the same object as `rhs`. The + * semantics are the same as Python's `is` operator. + * + * NOTE: Like in Python, this operation is poorly defined for primitive types + * like numbers and strings. Prefer to use `==` unless you really want to + * check identity equality. + */ + bool is(const IValue& rhs) const; + + /** + * Hashing for IValues. Returns an IValue-boxed int. + * + * Some notes: + * - Like eager, Tensors are hashed by looking at the pointer. This is not + * strictly correct because two value-equal tensors with different tensor + * pointers will hash differently, but we choose to reproduce the eager + * semantics. + * - Hashing is not defined on all built-in IValue types (e.g. list and + * dict), following Python. Calling `hash()` on these types will throw. + */ + IValue hash() const { + return (int64_t)IValue::hash(*this); + } + // This is defined because `c10::hash` dispatches to a function of this + // signature. See the member function `hash()`. + static size_t hash(const IValue& iv); + + /** + * @private [doxygen private] + * [container equality] + * This is an equality implementation that assumes objects with the same + * identity equal themselves, for efficiency reasons. We primarily have this + * for consistency, because Python does the same thing. This actually + * provokes user-visible changes in behavior due to quirks in torch: + * [tensor1] == [tensor1] -> True (because container equality will first + * compare identity) [tensor1] == [tensor1_copy] -> RuntimeError: + * Boolean value of Tensor with more than one value is ambiguous + */ + TORCH_API friend bool _fastEqualsForContainer( + const IValue& lhs, + const IValue& rhs); + + private: + static bool isAliasOf(const at::Tensor& a, const at::Tensor& b) { + if (a.is_sparse()) { + return isAliasOf(a._values(), b) || isAliasOf(a._indices(), b); + } + if (b.is_sparse()) { + return isAliasOf(a, b._values()) || isAliasOf(a, b._indices()); + } + if (a.is_sparse_csr()) { + return isAliasOf(a.values(), b) || isAliasOf(a.crow_indices(), b) || + isAliasOf(a.col_indices(), b); + } + if (b.is_sparse_csr()) { + return isAliasOf(a, b.values()) || isAliasOf(a, b.crow_indices()) || + isAliasOf(a, b.col_indices()); + } + + // Opaque tensors such as the ones constructed by the MKL-DNN backend + // don't have storage so we just compare their TensorImpls. + // TODO: Find way to expose alias info for opaque tensors. + if (!a.has_storage() || !b.has_storage()) { + return a.unsafeGetTensorImpl() == b.unsafeGetTensorImpl(); + } + + return a.is_alias_of(b); + } + + template + bool isListOf() const; + + public: + /// @private [doxygen private] + bool isAliasOf(const IValue& rhs) const { + if (this->tag != rhs.tag) { + // Trivially don't alias if the type is different + return false; + } + + // Tensors should be compared based on internal storage + if (this->isTensor()) { + return isAliasOf(this->toTensor(), rhs.toTensor()); + } + + if (!isIntrusivePtr()) { + // Primitive types don't alias anything + return false; + } + + AT_ASSERT(rhs.isIntrusivePtr()); + + // Other types can be compared by their ptr value + return this->payload.u.as_intrusive_ptr == rhs.payload.u.as_intrusive_ptr; + } + + /// @private [doxygen private] + size_t use_count() const noexcept { + if (isTensor()) { + return payload.as_tensor.use_count(); + } + + if (!isIntrusivePtrLegacyBehavior()) { + return 1; + } + + if (payload.u.as_intrusive_ptr == c10::UndefinedTensorImpl::singleton()) { + return 0; + } + return c10::raw::intrusive_ptr::use_count(payload.u.as_intrusive_ptr); + } + + /// @private [doxygen private] + void swap(IValue& rhs) noexcept { + if (isTensor() && rhs.isTensor()) { + std::swap(payload.as_tensor, rhs.payload.as_tensor); + } else if (isTensor()) { + at::Tensor t = std::move(payload.as_tensor); + // As far as I can tell, omitting the usual explicit destructor call + // is not UB in and of itself, and it's a slight perf win. The + // destructor is a no-op, because the moved-from Tensor is + // effectively an intrusive_ptr in the null state, so we don't need + // the behavior for correctness reasons either. Leaving this + // explanatory comment, including commented-out destructor call, to + // make this abundantly clear. + // + // payload.as_tensor.~Tensor(); + payload.u = rhs.payload.u; + new (&rhs.payload.as_tensor) at::Tensor(std::move(t)); + } else if (rhs.isTensor()) { + rhs.swap(*this); + return; + } else { + std::swap(payload.u, rhs.payload.u); + } + std::swap(tag, rhs.tag); + } + + // Accessors for subtypes are arranged together below + // While some of these accessors could be generated through templates, + // we prefer to write them manually for clarity + + IValue(at::TensorBase t) : tag(Tag::Tensor) { + new (&payload.as_tensor) at::Tensor(std::move(t)); + } + bool isTensor() const { + return Tag::Tensor == tag; + } + + private: + // Outlined error path so that toTensor() can be inlined. + [[noreturn]] void reportToTensorTypeError() const; + + public: + at::Tensor toTensor() &&; + at::Tensor& toTensor() &; + const at::Tensor& toTensor() const&; + at::TensorImpl* unsafeToTensorImpl() const { + TORCH_INTERNAL_ASSERT(isTensor()); + return payload.as_tensor.unsafeGetTensorImpl(); + } + + IValue(at::Storage s) : tag(Tag::Storage) { + payload.u.as_intrusive_ptr = + null_to_undefined_tensor(s.unsafeReleaseStorageImpl()); + } + bool isStorage() const { + return Tag::Storage == tag; + } + c10::Storage toStorage() &&; + c10::Storage toStorage() const&; + + const IValue& toIValue() const { + return *this; + } + IValue& toIValue() { + return *this; + } + + /// @private [doxygen private] + IValue(intrusive_ptr blob) : tag(Tag::Blob) { + // TODO (after Tensor merge) If we pass in a Blob holding a Tensor, extract + // and store it as a Tensor instead. + payload.u.as_intrusive_ptr = null_to_undefined_tensor(blob.release()); + } + + /// @private [doxygen private] + bool isBlob() const { + return Tag::Blob == tag; + } + + /// @private [doxygen private] + c10::intrusive_ptr toBlob() &&; + + /// @private [doxygen private] + c10::intrusive_ptr toBlob() const&; + + // Capsule. No new callsites of these APIs should + // be introduced. + static inline IValue make_capsule( + intrusive_ptr blob); + bool isCapsule() const { + return Tag::Capsule == tag; + } + c10::intrusive_ptr toCapsule() &&; + c10::intrusive_ptr toCapsule() const&; + + // Custom C++ classes + template < + typename T, + std::enable_if_t, int> = 0> + IValue(intrusive_ptr custom_class); + bool isCustomClass() const; + template + c10::intrusive_ptr toCustomClass() &&; + template + c10::intrusive_ptr toCustomClass() const&; + + // Tuple + IValue(c10::intrusive_ptr v); + + template < + typename... Args, + std::enable_if_t< + !std::disjunction_v< + std::is_lvalue_reference..., + std::negation>...>, + std::nullptr_t> = nullptr> + IValue(const std::tuple& t); + template < + typename... Args, + std::enable_if_t< + !std::disjunction_v< + std::is_lvalue_reference..., + std::negation>...>, + std::nullptr_t> = nullptr> + IValue(std::tuple&& t); + bool isTuple() const { + return Tag::Tuple == tag; + } + c10::intrusive_ptr toTuple() &&; + c10::intrusive_ptr toTuple() const&; + [[nodiscard]] ivalue::Tuple& toTupleRef() const; + + // Double + IValue(double d) : tag(Tag::Double) { + payload.u.as_double = d; + } + bool isDouble() const { + return Tag::Double == tag; + } + double toDouble() const { + if (isDouble()) { + return payload.u.as_double; + } else if (isSymFloat()) { + return toSymFloat().guard_float(__FILE__, __LINE__); + } else { + TORCH_INTERNAL_ASSERT(0, "expected double"); + } + } + + // ComplexDouble + template + IValue(c10::complex c); + bool isComplexDouble() const { + return Tag::ComplexDouble == tag; + } + c10::complex toComplexDouble() const; + + // Future + IValue(c10::intrusive_ptr v); + bool isFuture() const { + return Tag::Future == tag; + } + c10::intrusive_ptr toFuture() &&; + c10::intrusive_ptr toFuture() const&; + + IValue(c10::intrusive_ptr v); + bool isAwait() const { + return Tag::Await == tag; + } + c10::intrusive_ptr toAwait() &&; + c10::intrusive_ptr toAwait() const&; + + // RRef + IValue(c10::intrusive_ptr v); + bool isRRef() const { + return Tag::RRef == tag; + } + c10::intrusive_ptr toRRef() &&; + c10::intrusive_ptr toRRef() const&; + + // Quantizer + IValue(c10::intrusive_ptr v); + bool isQuantizer() const { + return Tag::Quantizer == tag; + } + c10::intrusive_ptr toQuantizer() &&; + c10::intrusive_ptr toQuantizer() const&; + + // Int + IValue(int64_t i) : tag(Tag::Int) { + payload.u.as_int = i; + } + + IValue(const c10::SymInt& i) { + if (auto mi = i.maybe_as_int()) { + tag = Tag::Int; + payload.u.as_int = *mi; + } else { + tag = Tag::SymInt; + payload.u.as_intrusive_ptr = i.toSymNode().release(); + } + } + + bool isSymInt() const { + return Tag::SymInt == tag; + } + + c10::SymInt toSymInt() &&; + c10::SymInt toSymInt() const&; + + IValue(const c10::SymFloat& i) { + if (i.is_symbolic()) { + tag = Tag::SymFloat; + payload.u.as_intrusive_ptr = i.toSymNodeImpl().release(); + } else { + tag = Tag::Double; + payload.u.as_double = i.as_float_unchecked(); + } + } + + bool isSymFloat() const { + return Tag::SymFloat == tag; + } + + c10::SymFloat toSymFloat() &&; + c10::SymFloat toSymFloat() const&; + + IValue(const c10::SymBool& i) { + if (auto mi = i.maybe_as_bool()) { + tag = Tag::Bool; +#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ + payload.u.as_int = *mi; +#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ + /* due to byteorder if value assigned as_int, as_bool actually is not set correctly */ + payload.u.as_bool = *mi; +#else +#error Unexpected or undefined __BYTE_ORDER__ +#endif + } else { + tag = Tag::SymBool; + payload.u.as_intrusive_ptr = i.toSymNodeImpl().release(); + } + } + + bool isSymBool() const { + return Tag::SymBool == tag; + } + + c10::SymBool toSymBool() &&; + c10::SymBool toSymBool() const&; + + // allow you to pass literals (3, 4) without ambiguity + IValue(int32_t i) : IValue(static_cast(i)) {} + + bool isInt() const { + return Tag::Int == tag; + } + + int64_t toInt() const { + if (isInt()) { + return payload.u.as_int; + } else if (isSymInt()) { + return toSymInt().guard_int(__FILE__, __LINE__); + } else { + TORCH_INTERNAL_ASSERT(0, "expected int"); + } + } + + // Unsigned + IValue(uint64_t u) : tag( u <= std::numeric_limits::max() ? Tag::Int : Tag::UInt) { + payload.u.as_uint = u; + } + + + // See Note [Meaning of HAS_u] + // IValue type model closely follows that of c10::Scalar + // Where all integers are upcast to 64-bit representation, and `as_int` is used as default + // representation unless value could not be represented as signed int + bool isUnsigned() const { + return Tag::UInt == tag || (Tag::Int == tag && payload.u.as_int >= 0); + } + + uint64_t toUInt() const { + if (isUnsigned()) { + return payload.u.as_uint; + } else { + TORCH_INTERNAL_ASSERT(0, "expected unsigned int"); + } + } + + + // Bool + IValue(bool b) : tag(Tag::Bool) { +#if defined(__clang__) && defined(__x86_64__) + // Initializing entire payload stops valgrind's from reporting + // "jump or move depends on uninitialised value" in IValue copy constructor + // See https://github.com/pytorch/pytorch/issues/37117 + payload.u.as_int = b; +#else + payload.u.as_bool = b; +#endif + } + bool isBool() const { + return Tag::Bool == tag; + } + bool toBool() const { + if (isBool()) { + return payload.u.as_bool; + } else if (isSymBool()) { + return toSymBool().guard_bool(__FILE__, __LINE__); + } else { + TORCH_INTERNAL_ASSERT(0, "expected bool"); + } + } + + // IntList + bool isIntList() const; + bool isSymIntList() const; + c10::List toIntList() &&; + c10::List toIntList() const&; + std::vector toIntVector() const; + c10::List toSymIntList() &&; + c10::List toSymIntList() const&; + std::vector toSymIntVector() const; + at::DimVector toDimVector() const; + + // ConstantString + IValue(c10::intrusive_ptr v); + IValue(std::string v); + IValue(const char* v) : IValue(std::string(v)) {} + IValue(std::string_view v) : IValue(std::string(v)){} + bool isString() const { + return Tag::String == tag; + } + c10::intrusive_ptr toString() &&; + c10::intrusive_ptr toString() const&; + const std::string& toStringRef() const; + std::optional> toOptionalStringRef() + const; + std::string_view toStringView() const; + + // DoubleList + bool isDoubleList() const; + c10::List toDoubleList() &&; + c10::List toDoubleList() const&; + std::vector toDoubleVector() const; + + // ComplexDoubleList + bool isComplexDoubleList() const; + c10::List> toComplexDoubleList() &&; + c10::List> toComplexDoubleList() const&; + std::vector> toComplexDoubleVector() const; + + // BoolList + bool isBoolList() const; + c10::List toBoolList() &&; + c10::List toBoolList() const&; + + // TensorList + bool isTensorList() const; + c10::List toTensorList() &&; + c10::List toTensorList() const&; + std::vector toTensorVector() const; + + // OptionalTensorList + bool isOptionalTensorList() const; + c10::List> toOptionalTensorList() &&; + c10::List> toOptionalTensorList() const&; + std::vector> toOptionalTensorVector() const; + + // GenericList + IValue(c10::List v); + bool isList() const { + return Tag::GenericList == tag; + } + c10::List toList() &&; + c10::List toList() const&; + c10::ArrayRef toListRef() const; + + // Some template constructors of IValue calls another constructor recursively. + // This SFINAEs the called constructor exists. + template + using enable_if_ivalue_constructible = + std::enable_if_t, std::nullptr_t>; + + // The rule for lists is more complicated; the generic constructor is only + // acceptable if your element isn't SymInt. If you do have a SymInt element, + // then you must also, at construction time, check if you can decay the list + // into an int list (this is MANDATORY, as at a use site we may expect + // toIntList to work even if at the call site you had a SymIntArrayRef + // argument). In practice, only SymIntArrayRef is used this way, so we + // didn't bother making it work for the other constructors, we just make sure + // they're not selectable. + template + using enable_if_list_is_ivalue_constructible = std::enable_if_t< + std::is_constructible_v && !std::is_same_v, + std::nullptr_t>; + + template = nullptr> + IValue(c10::List&& v); + template = nullptr> + IValue(const c10::List& v); + template = nullptr> + IValue(at::ArrayRef v); + template = nullptr> + IValue(const std::vector& v); + template = nullptr> + IValue(std::vector&& v); + template + IValue(std::array v); + + // Manual constructors for lists of symints, which decay to int list if + // possible. To avoid ambiguous overload situations, we template them + // to prevent implicit conversions + template + using enable_if_symint = + std::enable_if_t, std::nullptr_t>; + + template = nullptr> + IValue(at::ArrayRef v); + template = nullptr> + IValue(at::OptionalArrayRef v); + template = nullptr> + IValue(const std::vector& v); + template = nullptr> + IValue(std::vector&& v); + + template + using enable_if_ilist_is_ivalue_constructible = std::enable_if_t< + std::is_constructible_v && + std::is_constructible_v::boxed_type> && + !std::is_same_v, + std::nullptr_t>; + + template = nullptr> + IValue(c10::IListRef v); + + // GenericDict + IValue(c10::Dict v); + bool isGenericDict() const { + return Tag::GenericDict == tag; + } + c10::Dict toGenericDict() &&; + c10::Dict toGenericDict() const&; + + template + IValue(c10::Dict v); + + template + /// \cond + /// DOXYGEN_CANNOT_HANDLE_CONSTRUCTORS_WITH_MACROS_SO_EXCLUDE_THIS_LINE_FROM_DOXYGEN + C10_DEPRECATED_MESSAGE( + "IValues based on std::unordered_map are slow and deprecated. Please use c10::Dict instead.") + /// \endcond + IValue(std::unordered_map v); + + template = nullptr> + IValue(std::optional v); + template = nullptr> + IValue(c10::OptionalArrayRef v); + IValue(std::nullopt_t /*unused*/); + + // ClassType + IValue(c10::intrusive_ptr v); + bool isObject() const { + return tag == Tag::Object; + } + c10::intrusive_ptr toObject() &&; + c10::intrusive_ptr toObject() const&; + ivalue::Object& toObjectRef() const; + + torch::jit::Module toModule() const; + bool isModule() const; + + // PyObject + IValue(c10::intrusive_ptr v); + bool isPyObject() const { + return tag == Tag::PyObject; + } + c10::intrusive_ptr toPyObjectHolder() &&; + c10::intrusive_ptr toPyObjectHolder() const&; + PyObject* toPyObject() const; + + // Enum + explicit IValue(c10::intrusive_ptr v); + bool isEnum() const { + return tag == Tag::Enum; + } + c10::intrusive_ptr toEnumHolder() &&; + c10::intrusive_ptr toEnumHolder() const&; + + // None + IValue() = default; + bool isNone() const { + return Tag::None == tag; + } + std::string toNone() const { + AT_ASSERT(isNone()); + return "None"; + } + + static IValue uninitialized() { + auto i = IValue(); + i.tag = Tag::Uninitialized; + return i; + } + + // Scalar, which gets encoded as either an Int, a Double or a ComplexDouble + IValue(const at::Scalar& s) : IValue() { + // NB: do the symbolic versions first, as isFloatingPoint is true + // for both SymFloat and double + if (s.isSymInt()) { + tag = Tag::SymInt; + payload.u.as_intrusive_ptr = s.toSymInt().toSymNode().release(); + } else if (s.isSymFloat()) { + tag = Tag::SymFloat; + payload.u.as_intrusive_ptr = s.toSymFloat().toSymNodeImpl().release(); + } else if (s.isSymBool()) { + tag = Tag::SymBool; + payload.u.as_intrusive_ptr = s.toSymBool().toSymNodeImpl().release(); + } else if (s.isFloatingPoint()) { + tag = Tag::Double; + payload.u.as_double = s.toDouble(); + } else if (s.isComplex()) { + *this = s.toComplexDouble(); + } else if (s.isBoolean()) { + tag = Tag::Bool; + payload.u.as_bool = s.toBool(); + } else { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + s.isIntegral(false), "Unknown type in Scalar"); + if (s.isUnsigned()) { + const auto val = s.toUInt64(); + payload.u.as_uint = val; + tag = val <= std::numeric_limits::max() ? Tag::Int : Tag::UInt; + } else { + payload.u.as_int = s.toLong(); + tag = Tag::Int; + } + } + } + + bool isScalar() const { + return isDouble() || isInt() || isComplexDouble() || isBool() || + isSymInt() || isSymFloat() || isSymBool(); + } + + at::Scalar toScalar() const { + if (isDouble()) + return toDouble(); + else if (isInt()) + return toInt(); + else if (isComplexDouble()) + return toComplexDouble(); + else if (isBool()) + return toBool(); + else if (isSymInt()) + return toSymInt(); + else if (isSymFloat()) + return toSymFloat(); + else if (isSymBool()) + return toSymBool(); + else if (isUnsigned()) + return toUInt(); + TORCH_CHECK(false, "IValue is not a Scalar"); + } + + // Device + IValue(c10::Device d) : tag(Tag::Device) { + payload.u.as_device.type = d.type(); + payload.u.as_device.index = d.index(); + } + bool isDevice() const { + return Tag::Device == tag; + } + c10::Device toDevice() const { + AT_ASSERT(isDevice()); + return c10::Device(payload.u.as_device.type, payload.u.as_device.index); + } + + // Stream + IValue(c10::Stream s) : tag(Tag::Stream) { + auto v = c10::make_intrusive(s.pack3()); + payload.u.as_intrusive_ptr = v.release(); + } + c10::Stream toStream() &&; + c10::Stream toStream() const&; + bool isStream() const { + return Tag::Stream == tag; + } + + // ScalarType + IValue(ScalarType t) + : IValue(static_cast>(t)) {} + at::ScalarType toScalarType() const { + return static_cast(toInt()); + } + + // Layout + IValue(Layout l) : IValue(static_cast>(l)) {} + at::Layout toLayout() const { + return static_cast(toInt()); + } + + // MemoryFormat + IValue(MemoryFormat m) + : IValue(static_cast>(m)) {} + at::MemoryFormat toMemoryFormat() const { + return static_cast(toInt()); + } + + // QScheme + IValue(at::QScheme qscheme) : tag(Tag::Int) { + payload.u.as_int = static_cast(qscheme); + } + + at::QScheme toQScheme() const { + return static_cast(toInt()); + } + + // Dimname + IValue(at::Dimname dimname) : IValue(dimname.symbol().toQualString()) {} + + at::Dimname toDimname() const { + return at::Dimname::fromSymbol(Symbol::fromQualString(toStringRef())); + } + + // Generator + IValue(at::Generator g) : tag(Tag::Generator) { + payload.u.as_intrusive_ptr = + null_to_undefined_tensor(g.unsafeReleaseGeneratorImpl()); + } + bool isGenerator() const { + return Tag::Generator == tag; + } + at::Generator toGenerator() &&; + at::Generator toGenerator() const&; + + // for debugging + std::string tagKind() const { + switch (tag) { +#define DEFINE_CASE(x) \ + case Tag::x: \ + return #x; + TORCH_FORALL_TAGS(DEFINE_CASE) +#undef DEFINE_CASE + } + return "InvalidTag(" + std::to_string(static_cast(tag)) + ")"; + } + + // generic v.to() implementations + // that can be used in special functions like pop/push + // that use template meta-programming. + // prefer the directly named methods when you can, + // since they are simpler to understand + + // Note: if you get linker errors saying one of these is missing, + // change it to ... && = delete; and you will see better error messages for + // why However, we cannot commit this because some compiler versions barf on + // it. + template + T to() &&; + template + typename c10::detail::ivalue_to_const_ref_overload_return::type to() + const&; + + // ToOptional: convert a IValue to the Optional obj that accepts both T and + // None + template + std::optional toOptional(); + template + std::optional toOptional() const; + + /// @private [doxygen private] + /// this is a shallow comparison of two IValues to test the object identity + bool isSameIdentity(const IValue& rhs) const; + + // Computes the "official" string representation of an IValue. This produces a + // TorchScript expression that can be used to recreate an IValue with the same + // value (e.g. when we are printing constants in the serializer). + // + // Callers can use `customFormatter` to override how `repr()` prints out an + // IValue. This is useful if you have some other environment where you can + // look up values, and you want to print a reference to that environment (like + // the serializer's constant table). + // + // repr() is not necessarily defined on all objects! + std::ostream& repr( + std::ostream& stream, + std::function customFormatter) + const; + + // Computes an "informal" string representation of an IValue. This should be + // used for debugging, or servicing `print()`-like functions. + // This is different from `repr()` in that there is no expectation that we can + // exactly reconstruct an IValue from the output; feel free to use a + // concise/pretty form + TORCH_API friend std::ostream& operator<<(std::ostream& out, const IValue& v); + + bool isPtrType() const { + if (isTensor()) { + return payload.as_tensor.defined(); + } + return isIntrusivePtrLegacyBehavior(); + } + + /// @private [doxygen private] + const void* internalToPointer() const { + TORCH_INTERNAL_ASSERT( + isPtrType(), "Can only call internalToPointer() for pointer types"); + if (isTensor()) { + return payload.as_tensor.unsafeGetTensorImpl(); + } else { + return payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton() + ? payload.u.as_intrusive_ptr + : nullptr; + } + } + + template + TypePtr type() const; + + // Detect aliased tensors. + struct HashAliasedIValue { + size_t hashTensor(const at::Tensor& ten) const { + if (ten.is_sparse()) { + // COO sparse tensors have a "values" tensor and an "indices" tensor + // so this will detect overlap of sparse tensors that share a values + // tensor, but not sparse tensors that share an indices tensor. + return hashTensor(ten._values()); + } else if (ten.is_sparse_csr()) { + // COO sparse tensors have a "values" tensor and an "indices" tensor + // so this will detect overlap of sparse tensors that share a values + // tensor, but not sparse tensors that share an indices tensor. + return hashTensor(ten.values()); + } else if (!ten.has_storage()) { + // Opaque tensors such as the ones constructed by the MKL-DNN backend + // don't have storage so we just use their TensorImpls. + // TODO: Find way to expose alias info for opaque tensors. + return reinterpret_cast(ten.unsafeGetTensorImpl()); + } else { + return reinterpret_cast(ten.storage().unsafeGetStorageImpl()); + } + } + size_t operator()(const IValue& val) const { + if (val.isTensor()) { + return hashTensor(val.toTensor()); + } + // If it is not a Tensor, then two mutable IValues alias each other only + // if they are the same pointer. + return val.payload.u.as_int; + } + }; + + struct CompAliasedIValues { + bool operator()(const IValue& lhs, const IValue& rhs) const { + return lhs.isAliasOf(rhs); + } + }; + + using HashAliasedIValues = + std::unordered_set; + using HashAliasedIValueMap = + std::unordered_map; + + struct HashIdentityIValue { + size_t operator()(const IValue& val) const { + return val.payload.u.as_int; + } + }; + + struct CompIdentityIValues { + bool operator()(const IValue& lhs, const IValue& rhs) const { + return lhs.is(rhs); + } + }; + + using HashIdentityIValues = + std::unordered_set; + using HashIdentityIValueMap = + std::unordered_map; + + // Checks if this and rhs has a subvalues in common. + // [t1,t2] and [t2, t3] returns true. + bool overlaps(const IValue& rhs) const; + + // Inserts all subvalues of this in subValues. + void getSubValues(HashAliasedIValues& subValues) const; + + // Apply visitor to every subvalue. + // TODO: There are several places that recurse over IValue. This is fragile. + // This visitor should be used to recurse over ivalues. + void visit(const std::function& visitor) const; + IValue deepcopy(std::optional device = std::nullopt) const; + IValue deepcopy( + HashIdentityIValueMap& memo, + std::optional device = std::nullopt) const; + + private: + static c10::intrusive_ptr_target* null_to_undefined_tensor( + c10::intrusive_ptr_target* p) { + return p ? p + : static_cast( + c10::UndefinedTensorImpl::singleton()); + } + + static bool ptrEqual(const IValue& lhs, const IValue& rhs); + // NOTE: IValue tags are intentionally private. In the future we may encode + // this value different (e.g. using NaN boxing), and this would make it more + // costly to determine the tag for all types vs just determining if something + // is a particular type. Instead we want clients to use the `isX` methods when + // possible. If for performance reasons you really, absolutely, must have a jump + // table, then we can revisit this. + enum class Tag : uint32_t { +#define DEFINE_TAG(x) x, + TORCH_FORALL_TAGS(DEFINE_TAG) +#undef DEFINE_TAG + }; + +#define COUNT_TAG(x) 1 + + static constexpr auto kNumTags = TORCH_FORALL_TAGS(COUNT_TAG) 0; +#undef COUNT_TAG + + template < + class T, + class NullType = c10::detail::intrusive_target_default_null_type> + c10::intrusive_ptr moveToIntrusivePtr(); + template < + typename T, + class NullType = c10::detail::intrusive_target_default_null_type> + c10::intrusive_ptr toIntrusivePtr() const; + + void destroy() { + // We carefully construct this call to both 1) avoid UB by using + // the "wrong" one of as_tensor and as_intrusive_ptr and 2) enable + // the compiler to generate the same code for each case. It is + // surprisingly difficult to get this right. + if (isTensor() || isIntrusivePtr()) { + c10::intrusive_ptr_target* p = isTensor() + ? payload.as_tensor.unsafeGetTensorImpl() + : payload.u.as_intrusive_ptr; + c10::intrusive_ptr:: + reclaim(p); + // No need to make this destructor call! + // payload.as_tensor.~Tensor(); + } + } + + // NOLINTNEXTLINE(cppcoreguidelines-rvalue-reference-param-not-moved) + C10_ALWAYS_INLINE void moveFrom(IValue&& rhs) noexcept { + if (rhs.isTensor()) { + new (&payload.as_tensor) at::Tensor(std::move(rhs.payload.as_tensor)); + // As far as I can tell, omitting the usual explicit destructor call + // is not UB in and of itself, and it's a slight perf win. The + // destructor is a no-op, because the moved-from Tensor is + // effectively an intrusive_ptr in the null state, so we don't need + // the behavior for correctness reasons either. Leaving this + // explanatory comment, including commented-out destructor call, to + // make this abundantly clear. + // + // rhs.payload.as_tensor.~Tensor(); + } else { + payload.u = rhs.payload.u; + } + tag = rhs.tag; + rhs.clearToNone(); + } + + void clearToNone() noexcept { + payload.u.as_int = 0; + tag = Tag::None; + } + + private: + // This is the source of truth for isIntrusivePtr; edit results here + // as needed and isIntrusivePtr will pick them up. + // NOLINTBEGIN(bugprone-branch-clone) + static constexpr bool isIntrusivePtrConstexpr(Tag tag) { + switch (tag) { + case Tag::None: + return false; + case Tag::Tensor: + return false; + case Tag::Storage: + return true; + case Tag::Generator: + return true; + case Tag::Double: + return false; + case Tag::ComplexDouble: + return true; + case Tag::Int: + return false; + case Tag::UInt: + return false; + case Tag::SymInt: + return true; + case Tag::SymFloat: + return true; + case Tag::SymBool: + return true; + case Tag::Bool: + return false; + case Tag::Tuple: + return true; + case Tag::String: + return true; + case Tag::Blob: + return true; + case Tag::GenericList: + return true; + case Tag::GenericDict: + return true; + case Tag::Future: + return true; + case Tag::Await: + return true; + case Tag::Device: + return false; + case Tag::Stream: + return true; + case Tag::Object: + return true; + case Tag::PyObject: + return true; + case Tag::Uninitialized: + return false; + case Tag::Capsule: + return true; + case Tag::RRef: + return true; + case Tag::Quantizer: + return true; + case Tag::Enum: + return true; + } + return false; + } + // NOLINTEND(bugprone-branch-clone) + + public: + // Don't edit this just to add results for new tags; edit + // isIntrusivePtrConstexpr above. + bool isIntrusivePtr() const { + // Implementation NOTE: the switch in isIntrusivePtrConstexpr + // above is the previous production implementation of this + // function. We observed that, at least on x86_64, the generated + // instruction sequence was a similar bit vector test to what we + // have manually implemented below, except that there was an extra + // "bounds check" branch confirming, essentially, that `tag < + // kNumTags` and providing a consistent result in that case. We + // don't care about the result if tag is out of bounds, so we'd + // like to eliminate that comparison and branch; manually + // implementing this function as a bit test is the simplest way I + // could find to accomplish that elimination. + static constexpr uint32_t kTruthTableBitVector = +#define TRUTH_TABLE_ENTRY(tag) \ + (uint32_t(isIntrusivePtrConstexpr(Tag::tag)) << uint32_t(Tag::tag)) | + TORCH_FORALL_TAGS(TRUTH_TABLE_ENTRY) +#undef TRUTH_TABLE_ENTRY + 0; + + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + static_cast(tag) < kNumTags, + "unexpected tag ", + static_cast(tag)); + return kTruthTableBitVector & (1 << (uint32_t(tag) % 32)); + } + + // Storage and Generator were treated specially when + // is_intrusive_ptr was stored as explicit state. This getter + // preserves the old behavior for use with WeakIValue for now. + bool isIntrusivePtrLegacyBehavior() const { + if (tag == Tag::Storage || tag == Tag::Generator) { + return payload.u.as_intrusive_ptr != + c10::UndefinedTensorImpl::singleton(); + } else { + return isIntrusivePtr(); + } + } + + union Payload { + // [TriviallyCopyablePayload] + // We use a nested union here so that we can make the copy easy + // and efficient in the non-tensor (i.e., trivially copyable) + // case. Specifically, we do not have to do a switch-on-tag to + // figure out which union member to assign; we can just use + // TriviallyCopyablePayload::operator=. + union TriviallyCopyablePayload { + TriviallyCopyablePayload() : as_int(0) {} + int64_t as_int; + // See Note [Meaning of HAS_u] + uint64_t as_uint; + double as_double; + bool as_bool; + // Invariant: never nullptr; null state is represented as + // c10::UndefinedTensorImpl::singleton() for consistency of + // representation with Tensor. + c10::intrusive_ptr_target* as_intrusive_ptr; + struct { + c10::DeviceType type; + DeviceIndex index; + } as_device; + } u; + static_assert(std::is_trivially_copyable_v); + at::Tensor as_tensor; + Payload() : u() {} + Payload(const Payload&) = delete; + Payload(Payload&&) = delete; + Payload& operator=(const Payload&) = delete; + Payload& operator=(Payload&&) = delete; + // NOLINTNEXTLINE(modernize-use-equals-default) + ~Payload() {} + }; + + IValue(const Payload& p, Tag t) : tag(t) { + if (isTensor()) { + new (&payload.as_tensor) at::Tensor(p.as_tensor); + } else { + payload.u = p.u; + } + } + + template + struct TagType {}; + + friend MaybeOwnedTraits; + + Payload payload; + Tag tag{IValue::Tag::None}; + friend struct WeakIValue; +}; + +struct TORCH_API WeakIValue final { + WeakIValue() = default; + + WeakIValue(const WeakIValue& rhs) + : payload(rhs.payload), + tag(rhs.tag), + is_intrusive_ptr(rhs.is_intrusive_ptr) { + if (is_intrusive_ptr && + payload.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton()) { + c10::raw::weak_intrusive_ptr::incref(payload.as_intrusive_ptr); + } + } + WeakIValue(const IValue& rhs) + : tag(rhs.tag), is_intrusive_ptr(rhs.isIntrusivePtrLegacyBehavior()) { + if (rhs.isTensor()) { + payload.as_intrusive_ptr = rhs.unsafeToTensorImpl(); + is_intrusive_ptr = true; + } else { + payload = rhs.payload.u; + } + if (is_intrusive_ptr) { + if (payload.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton()) { + c10::raw::weak_intrusive_ptr::incref(payload.as_intrusive_ptr); + } + } + } + WeakIValue(WeakIValue&& rhs) noexcept : WeakIValue() { + swap(rhs); + } + ~WeakIValue() { + if (is_intrusive_ptr && + payload.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton()) { + c10::raw::weak_intrusive_ptr::decref(payload.as_intrusive_ptr); + } + } + WeakIValue& operator=(WeakIValue&& rhs) & noexcept { + WeakIValue(std::move(rhs)).swap(*this); // this also sets rhs to None + return *this; + } + WeakIValue& operator=(WeakIValue const& rhs) & { + WeakIValue(rhs).swap(*this); + return *this; + } + void swap(WeakIValue& rhs) noexcept { + std::swap(payload, rhs.payload); + std::swap(is_intrusive_ptr, rhs.is_intrusive_ptr); + std::swap(tag, rhs.tag); + } + + bool isSameIdentity(const WeakIValue& rhs) const { + return payload.as_int == rhs.payload.as_int && tag == rhs.tag && + is_intrusive_ptr == rhs.is_intrusive_ptr; + } + + IValue lock() const { + if (!is_intrusive_ptr) { + IValue::Payload newPayload; + newPayload.u = payload; + return IValue(newPayload, tag); + } + if (IValue::Tag::Tensor == tag) { + auto temp = + c10::weak_intrusive_ptr:: + reclaim(static_cast(payload.as_intrusive_ptr)); + c10::intrusive_ptr ip( + temp.lock()); + temp.release(); + if (!ip) { + return IValue(); + } else { + return IValue(at::Tensor(std::move(ip))); + } + } else { + auto temp = c10::weak_intrusive_ptr::reclaim( + payload.as_intrusive_ptr == c10::UndefinedTensorImpl::singleton() + ? nullptr + : payload.as_intrusive_ptr); + IValue::Payload pl; + pl.u.as_intrusive_ptr = temp.lock().release(); + temp.release(); + if (!pl.u.as_intrusive_ptr) { + return IValue(); + } else { + return IValue(pl, tag); + } + } + } + + size_t use_count() const noexcept { + if (!is_intrusive_ptr) { + return 1; + } + auto temp = c10::weak_intrusive_ptr< + c10::intrusive_ptr_target, + c10::UndefinedTensorImpl>::reclaim(payload.as_intrusive_ptr); + size_t result = temp.use_count(); + temp.release(); + return result; + } + + size_t weak_use_count() const noexcept { + if (!is_intrusive_ptr) { + return 1; + } + auto temp = c10::weak_intrusive_ptr< + c10::intrusive_ptr_target, + c10::UndefinedTensorImpl>::reclaim(payload.as_intrusive_ptr); + size_t result = temp.weak_use_count(); + temp.release(); + return result; + } + size_t hash() const { + return payload.as_int; + } + + private: + using Payload = IValue::Payload::TriviallyCopyablePayload; + Payload payload; + IValue::Tag tag{IValue::Tag::None}; + bool is_intrusive_ptr{false}; +}; + +// An owning pointer to a type. When the type is class type, it requires a pair +// of shared_ptrs to the class type and its owning CU, so that the class type is +// guaranteed to stay alive as long as we hold this object. +struct TORCH_API StrongTypePtr { + StrongTypePtr(std::shared_ptr cu, TypePtr type); + + std::shared_ptr cu_; + TypePtr type_; +}; + +// [Constant Object Weak CompilationUnit Reference] +// A non owning pointer to a type. When a class get inserted as a constant +// into a graph, if we used a strong pointer we would have a circular reference +// from Object -> CompilationUnit and CompilationUnit -> Graph (which owns the +// Constant Object) +struct TORCH_API WeakTypePtr { + WeakTypePtr(std::weak_ptr cu, TypePtr type); + + std::weak_ptr cu_; + TypePtr type_; +}; + +// internal build errors with std::variant :/ +struct WeakOrStrongCompilationUnit { + explicit WeakOrStrongCompilationUnit( + std::shared_ptr shared_cu) + : strong_ptr_(std::move(shared_cu)), weak_ptr_(std::nullopt) {} + + explicit WeakOrStrongCompilationUnit( + std::weak_ptr weak_cu) + : strong_ptr_(std::nullopt), weak_ptr_(std::move(weak_cu)) {} + + std::shared_ptr getStrongRefOrThrow() const { + TORCH_INTERNAL_ASSERT(strong_ptr_.has_value()); + return *strong_ptr_; + } + + std::weak_ptr getWeakRefOrThrow() const { + TORCH_INTERNAL_ASSERT(weak_ptr_.has_value()); + return *weak_ptr_; + } + + bool holdingStrongRef() const { + return strong_ptr_.has_value(); + } + + bool holdingEmptyStrongRef() const { + return strong_ptr_ == nullptr; + } + + std::optional> strong_ptr_; + std::optional> weak_ptr_; +}; + +// An Object will hold a non-owning Compilation Unit reference if it is a +// Constant in the graph and a Owning reference otherwise +struct TORCH_API WeakOrStrongTypePtr { + explicit WeakOrStrongTypePtr(WeakTypePtr weak) + : cu_(WeakOrStrongCompilationUnit(std::move(weak.cu_))), + type_(std::move(weak.type_)) {} + explicit WeakOrStrongTypePtr(StrongTypePtr strong) + : cu_(WeakOrStrongCompilationUnit(std::move(strong.cu_))), + type_(std::move(strong.type_)) {} + explicit WeakOrStrongTypePtr(WeakOrStrongCompilationUnit cu, TypePtr type) + : cu_(std::move(cu)), type_(std::move(type)) {} + WeakTypePtr asWeakTypePtr() const; + + WeakOrStrongCompilationUnit cu_; + TypePtr type_; + + bool holds_strong_ref() const { + return cu_.holdingStrongRef(); + } + + bool holds_empty_strong_ref() const { + return cu_.holdingEmptyStrongRef(); + } +}; + +} // namespace c10 + +C10_DIAGNOSTIC_POP() + +#include // IWYU pragma: keep + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/ivalue_inl.h b/backends/pytorch/include/ATen/core/ivalue_inl.h new file mode 100644 index 000000000..e68c8ba81 --- /dev/null +++ b/backends/pytorch/include/ATen/core/ivalue_inl.h @@ -0,0 +1,2578 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default") + +namespace torch { +namespace jit { +struct Function; +struct CompilationUnit; +} // namespace jit +TORCH_API bool isCustomClass(const c10::IValue& v); +} // namespace torch +namespace c10 { +struct IValue; +struct ClassType; +struct TupleType; +struct EnumType; +struct InferredType; + +// For custom class __init__ registration, we need to pass in a function +// that looks like this: [](IValue x, args...) + +// However, make_boxed_from_unboxed_functor.h automatically sets the input types +// of the function by introspecting the types of the functor (which is IValue in +// this case). However, we need the type it binds to be Foo. + +// Instead, we pass in a lambda [](ivalue_holder x, args...) from +// which getTypePtr can recover the original class pointer. + +template +struct tagged_capsule { + IValue ivalue; +}; + +template +c10::intrusive_ptr IValue::moveToIntrusivePtr() { + auto t = c10::intrusive_ptr::reclaim( + payload.u.as_intrusive_ptr == c10::UndefinedTensorImpl::singleton() + ? NullType::singleton() + : static_cast(payload.u.as_intrusive_ptr)); + clearToNone(); + return t; +} +template +c10::intrusive_ptr IValue::toIntrusivePtr() const { + if (payload.u.as_intrusive_ptr == c10::UndefinedTensorImpl::singleton()) { + return c10::intrusive_ptr(); + } + c10::raw::intrusive_ptr::incref(payload.u.as_intrusive_ptr); + return c10::intrusive_ptr::reclaim( + static_cast(payload.u.as_intrusive_ptr)); +} + +template +intrusive_ptr static_intrusive_pointer_cast(intrusive_ptr r) { + return intrusive_ptr::reclaim(static_cast(r.release())); +} + +template +intrusive_ptr dynamic_intrusive_pointer_cast(intrusive_ptr r) { + return intrusive_ptr::reclaim(dynamic_cast(r.release())); +} + +inline c10::intrusive_ptr IValue::toFuture() && { + AT_ASSERT(isFuture(), "Expected Future but got ", tagKind()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toFuture() const& { + AT_ASSERT(isFuture(), "Expected Future but got ", tagKind()); + return toIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toAwait() && { + AT_ASSERT(isAwait(), "Expected Await but got ", tagKind()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toAwait() const& { + AT_ASSERT(isAwait(), "Expected Await but got ", tagKind()); + return toIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toRRef() && { + AT_ASSERT(isRRef(), "Expected RRef but got ", tagKind()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toRRef() const& { + AT_ASSERT(isRRef(), "Expected RRef but got ", tagKind()); + return toIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toQuantizer() && { + AT_ASSERT(isQuantizer(), "Expected Quantizer but got ", tagKind()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toQuantizer() const& { + AT_ASSERT(isQuantizer(), "Expected Quantizer but got ", tagKind()); + return toIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toString() && { + AT_ASSERT(isString(), "Expected String but got ", tagKind()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toString() const& { + AT_ASSERT(isString(), "Expected String but got ", tagKind()); + return toIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toObject() && { + AT_ASSERT(isObject(), "Expected Object but got ", tagKind()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toObject() const& { + AT_ASSERT(isObject(), "Expected Object but got ", tagKind()); + return toIntrusivePtr(); +} +inline c10::intrusive_ptr IValue:: + toPyObjectHolder() && { + TORCH_INTERNAL_ASSERT(isPyObject(), "Expected PyObject but got ", tagKind()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toPyObjectHolder() + const& { + TORCH_INTERNAL_ASSERT(isPyObject(), "Expected PyObject but got ", tagKind()); + return toIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toEnumHolder() && { + TORCH_INTERNAL_ASSERT(isEnum(), "Expected Enum but got ", tagKind()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toEnumHolder() const& { + TORCH_INTERNAL_ASSERT(isEnum(), "Expected Enum but got ", tagKind()); + return toIntrusivePtr(); +} +inline c10::complex IValue::toComplexDouble() const { + TORCH_INTERNAL_ASSERT(isComplexDouble(), "Expected ComplexDouble but got ", tagKind()); + auto ptr = toIntrusivePtr(); + return (*ptr).val; +} +inline at::Tensor IValue::toTensor() && { + if (C10_UNLIKELY(!isTensor())) { + reportToTensorTypeError(); + } + auto result = std::move(payload.as_tensor); + // As far as I can tell, omitting the usual explicit destructor call + // is not UB in and of itself, and it's a slight perf win. The + // destructor is a no-op, because the moved-from Tensor is + // effectively an intrusive_ptr in the null state, so we don't need + // the behavior for correctness reasons either. Leaving this + // explanatory comment, including commented-out destructor call, to + // make this abundantly clear. + // + // payload.as_tensor.~Tensor(); + clearToNone(); + return result; +} +inline at::Tensor& IValue::toTensor() & { + if (C10_UNLIKELY(!isTensor())) { + reportToTensorTypeError(); + } + return payload.as_tensor; +} +inline const at::Tensor& IValue::toTensor() const& { + if (C10_UNLIKELY(!isTensor())) { + reportToTensorTypeError(); + } + return payload.as_tensor; +} +inline c10::Storage IValue::toStorage() && { + AT_ASSERT(isStorage(), "Expected Storage but got ", tagKind()); + return c10::Storage( + moveToIntrusivePtr()); +} +inline c10::Storage IValue::toStorage() const& { + AT_ASSERT(isStorage(), "Expected Storage but got ", tagKind()); + return c10::Storage(toIntrusivePtr()); +} +inline c10::Stream IValue::toStream() && { + AT_ASSERT(isStream(), "Expected Stream but got ", tagKind()); + auto ptr = toIntrusivePtr(); + return c10::Stream::unpack3((*ptr).val.stream_id, + (*ptr).val.device_index, + (*ptr).val.device_type); +} +inline c10::Stream IValue::toStream() const& { + AT_ASSERT(isStream(), "Expected Stream but got ", tagKind()); + auto ptr = toIntrusivePtr(); + return c10::Stream::unpack3((*ptr).val.stream_id, + (*ptr).val.device_index, + (*ptr).val.device_type); +} +inline c10::intrusive_ptr IValue::toBlob() && { + AT_ASSERT(isBlob(), "Expected Blob but got ", tagKind()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toBlob() const& { + AT_ASSERT(isBlob(), "Expected Blob but got ", tagKind()); + return toIntrusivePtr(); + ; +} +inline c10::intrusive_ptr IValue::toCapsule() && { + TORCH_INTERNAL_ASSERT(isCapsule()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toCapsule() const& { + TORCH_INTERNAL_ASSERT(isCapsule()); + return toIntrusivePtr(); +} +inline at::Generator IValue::toGenerator() && { + AT_ASSERT(isGenerator(), "Expected Generator but got ", tagKind()); + return at::Generator(moveToIntrusivePtr()); +} +inline at::Generator IValue::toGenerator() const& { + AT_ASSERT(isGenerator(), "Expected Generator but got ", tagKind()); + return at::Generator(toIntrusivePtr()); +} +inline c10::SymInt IValue::toSymInt() && { + AT_ASSERT(isSymInt() || isInt(), "Expected SymInt or int but got ", tagKind()); + if (isSymInt()) { + return c10::SymInt(moveToIntrusivePtr()); + } else { + return c10::SymInt(payload.u.as_int); + } +} +inline c10::SymInt IValue::toSymInt() const& { + AT_ASSERT(isSymInt() || isInt(), "Expected SymInt or int but got ", tagKind()); + if (isSymInt()) { + return c10::SymInt(toIntrusivePtr()); + } else { + return c10::SymInt(payload.u.as_int); + } +} +inline c10::SymFloat IValue::toSymFloat() && { + AT_ASSERT(isSymFloat() || isDouble(), "Expected SymFloat or double but got ", tagKind()); + if (isSymFloat()) { + return c10::SymFloat(moveToIntrusivePtr()); + } else { + return c10::SymFloat(payload.u.as_double); + } +} +inline c10::SymFloat IValue::toSymFloat() const& { + AT_ASSERT(isSymFloat() || isDouble(), "Expected SymFloat or double but got ", tagKind()); + if (isSymFloat()) { + return c10::SymFloat(toIntrusivePtr()); + } else { + return c10::SymFloat(payload.u.as_double); + } +} +inline c10::SymBool IValue::toSymBool() && { + AT_ASSERT(isSymBool() || isBool(), "Expected SymBool or boolean but got ", tagKind()); + if (isSymBool()) { + return c10::SymBool(moveToIntrusivePtr()); + } else { + return c10::SymBool(payload.u.as_bool); + } +} + +inline c10::SymBool IValue::toSymBool() const& { + AT_ASSERT(isSymBool() || isBool(), "Expected SymBool or boolean but got ", tagKind()); + if (isSymBool()) { + return c10::SymBool(toIntrusivePtr()); + } else { + return c10::SymBool(payload.u.as_bool); + } +} + +namespace ivalue { + +void TORCH_API +checkCustomClassType(const ClassType* expected_type, const Type* actual_type); + +template +using Shared = c10::intrusive_ptr; + +// string +struct TORCH_API ConstantString final : c10::intrusive_ptr_target { + private: + // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members) + const std::string str_; + + public: + ConstantString(std::string str) : str_(std::move(str)) {} + ConstantString(std::string_view str) : str_(std::string(str)) {} + static c10::intrusive_ptr create(std::string str_); + static c10::intrusive_ptr create(std::string_view str_); + static c10::intrusive_ptr create(const char* str_); + + const std::string& string() const { + return str_; + } + std::string_view string_view() const { + return str_; + } + + operator const std::string&() const { + return string(); + } + TORCH_API friend std::ostream& operator<<( + std::ostream& out, + const ConstantString& v); +}; + +struct Future; + +struct TORCH_API TupleElements { + private: + size_t inlineSize_; + // We represent TupleElements this way to save doing a heap + // allocation in the common (at least for unpickling) case where we + // have only 3 elements. We have our own union instead of + // c10::SmallVector because c10::SmallVector always + // stores the begin/end/capacity pointers, which would be a waste of + // space in our use case. + union { + std::vector elementsVector_; + // Don't want to declare a std::array because the convenient + // iteration and size members are a footgun in this case -- the + // actual size of the array may be smaller than 3! + // NOLINTNEXTLINE(*c-arrays*) + IValue elementsInline_[3]; + }; + + void destroyInline() { + for (const auto ii : c10::irange(inlineSize_)) { + elementsInline_[ii].~IValue(); + } + } + public: + + using iterator = IValue*; + using const_iterator = const IValue*; + + TupleElements() : inlineSize_(0) { + new (&elementsVector_) std::vector(); + } + + explicit TupleElements(std::vector elements) + : inlineSize_(0), elementsVector_(std::move(elements)) {} + + explicit TupleElements(c10::ArrayRef elements) + : inlineSize_(elements.size() <= 3 ? elements.size() : 0) { + switch (inlineSize_) { + case 3: + new (&elementsInline_[2]) IValue(elements[2]); + [[fallthrough]]; + case 2: + new (&elementsInline_[1]) IValue(elements[1]); + [[fallthrough]]; + case 1: + new (&elementsInline_[0]) IValue(elements[0]); + break; + case 0: + new (&elementsVector_) std::vector(elements.begin(), elements.end()); + break; + } + } + + explicit TupleElements(IValue&& e1) + : inlineSize_(1) { + new (&elementsInline_[0]) IValue(std::move(e1)); + } + + explicit TupleElements(IValue&& e1, IValue&& e2) + : inlineSize_(2) { + new (&elementsInline_[0]) IValue(std::move(e1)); + new (&elementsInline_[1]) IValue(std::move(e2)); + } + + explicit TupleElements(IValue&& e1, IValue&& e2, IValue&& e3) + : inlineSize_(3) { + new (&elementsInline_[0]) IValue(std::move(e1)); + new (&elementsInline_[1]) IValue(std::move(e2)); + new (&elementsInline_[2]) IValue(std::move(e3)); + } + + ~TupleElements() { + if (inlineSize_) { + destroyInline(); + } else { + elementsVector_.~vector(); + } + } + + // It would be nice to make this noncopyable to prevent people from + // writing code like `auto output = + // forward(...).toTupleRef().elements()` (which does refcount bumps on + // each element, unlike the more efficient but verbose + // ``` + // auto outputIntrusivePtr = forward(...).toTuple(); + // const auto& output = outputIntrusivePtr->elements(); + // ``` + // ), but there is simply an overwhelming amount of code that does + // it the inefficient way. + // See also operator std::vector below. + TupleElements(const TupleElements& rhs) + : inlineSize_(rhs.inlineSize_) { + if (rhs.inlineSize_) { + for (const auto ii : c10::irange(inlineSize_)) { + new (&elementsInline_[ii]) IValue(rhs.elementsInline_[ii]); + } + } else { + new (&elementsVector_) std::vector(rhs.elementsVector_); + } + } + + TupleElements& operator=(const TupleElements& rhs) { + if (inlineSize_) { + if (rhs.inlineSize_) { + for (const auto ii : c10::irange(std::min(inlineSize_, rhs.inlineSize_))) { + elementsInline_[ii] = rhs.elementsInline_[ii]; + } + if (rhs.inlineSize_ > inlineSize_) { + for (const auto ii : c10::irange(inlineSize_, rhs.inlineSize_)) { + new (&elementsInline_[ii]) IValue(rhs.elementsInline_[ii]); + } + } else { + for (const auto ii : c10::irange(rhs.inlineSize_, inlineSize_)) { + elementsInline_[ii].~IValue(); + } + } + } else { + destroyInline(); + new (&elementsVector_) std::vector(rhs.elementsVector_); + } + } else { + if (rhs.inlineSize_) { + elementsVector_.~vector(); + for (const auto ii : c10::irange(rhs.inlineSize_)) { + new (&elementsInline_[ii]) IValue(rhs.elementsInline_[ii]); + } + } else { + elementsVector_ = rhs.elementsVector_; + } + } + inlineSize_ = rhs.inlineSize_; + return *this; + } + + TupleElements(TupleElements&& rhs) noexcept + : inlineSize_(rhs.inlineSize_) { + if (inlineSize_) { + for (const auto ii : c10::irange(inlineSize_)) { + new (&elementsInline_[ii]) IValue(std::move(rhs.elementsInline_[ii])); + } + } else { + new (&elementsVector_) std::vector(std::move(rhs.elementsVector_)); + } + } + + TupleElements& operator=(TupleElements&& rhs) noexcept { + if (inlineSize_) { + if (rhs.inlineSize_) { + for (const auto ii : c10::irange(std::min(inlineSize_, rhs.inlineSize_))) { + elementsInline_[ii] = std::move(rhs.elementsInline_[ii]); + } + if (rhs.inlineSize_ > inlineSize_) { + for (const auto ii : c10::irange(inlineSize_, rhs.inlineSize_)) { + new (&elementsInline_[ii]) IValue(std::move(rhs.elementsInline_[ii])); + } + } else { + for (const auto ii : c10::irange(rhs.inlineSize_, inlineSize_)) { + elementsInline_[ii].~IValue(); + } + } + } else { + destroyInline(); + new (&elementsVector_) std::vector(std::move(rhs.elementsVector_)); + } + } else { + if (rhs.inlineSize_) { + elementsVector_.~vector(); + for (const auto ii : c10::irange(rhs.inlineSize_)) { + new (&elementsInline_[ii]) IValue(std::move(rhs.elementsInline_[ii])); + } + } else { + elementsVector_ = std::move(rhs.elementsVector_); + } + } + inlineSize_ = rhs.inlineSize_; + return *this; + } + + [[nodiscard]] c10::ArrayRef asArrayRef() const { + if (inlineSize_) { + return c10::ArrayRef(elementsInline_, inlineSize_); + } else { + return elementsVector_; + } + } + + // Mimic implicit conversion from std::vector to ArrayRef. + operator c10::ArrayRef() const { + return asArrayRef(); + } + + static size_t hash(const TupleElements& v) { + return c10::hash>()(v.asArrayRef()); + } + + void setContents(std::vector&& contents) { + if (inlineSize_) { + destroyInline(); + new (&elementsVector_) std::vector(std::move(contents)); + inlineSize_ = 0; + } else { + elementsVector_ = std::move(contents); + } + } + + [[nodiscard]] bool empty() const { + return inlineSize_ ? false : elementsVector_.empty(); + } + + [[nodiscard]] size_t size() const { + return inlineSize_ ? inlineSize_ : elementsVector_.size(); + } + + [[nodiscard]] IValue& operator[](size_t idx) { + if (inlineSize_) { + return elementsInline_[idx]; + } else { + return elementsVector_[idx]; + } + } + + [[nodiscard]] const IValue& operator[](size_t idx) const { + if (inlineSize_) { + return elementsInline_[idx]; + } else { + return elementsVector_[idx]; + } + } + + [[nodiscard]] IValue& at(size_t idx) { + if (inlineSize_) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(inlineSize_ <= 3); + TORCH_CHECK(idx < inlineSize_, "TupleElements: invalid index Index = ", idx, "; Length = ", inlineSize_); + return elementsInline_[idx]; + } else { + return elementsVector_.at(idx); + } + } + + [[nodiscard]] const IValue& at(size_t idx) const { + if (inlineSize_) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(inlineSize_ <= 3); + TORCH_CHECK(idx < inlineSize_, "TupleElements: invalid index Index = ", idx, "; Length = ", inlineSize_); + return elementsInline_[idx]; + } else { + TORCH_CHECK(idx < elementsVector_.size(), "TupleElements: invalid index Index = ", idx, "; Length = ", elementsVector_.size()); + return elementsVector_.at(idx); + } + } + + [[nodiscard]] iterator begin() { + if (inlineSize_) { + return elementsInline_; + } else { + return elementsVector_.data(); + } + } + + [[nodiscard]] iterator end() { + if (inlineSize_) { + return elementsInline_ + inlineSize_; + } else { + return elementsVector_.data() + elementsVector_.size(); + } + } + + [[nodiscard]] const_iterator begin() const { + if (inlineSize_) { + return elementsInline_; + } else { + return elementsVector_.data(); + } + } + + [[nodiscard]] const_iterator end() const { + if (inlineSize_) { + return elementsInline_ + inlineSize_; + } else { + return elementsVector_.data() + elementsVector_.size(); + } + } + + [[nodiscard]] const_iterator cbegin() const { + return begin(); + } + + [[nodiscard]] const_iterator cend() const { + return end(); + } + + [[nodiscard]] std::vector vec() const& { + return asArrayRef().vec(); + } + + [[nodiscard]] IValue& back() { + return *(end() - 1); + } + + [[nodiscard]] const IValue& back() const { + return *(end() - 1); + } + + [[nodiscard]] std::vector vec() && { + std::vector result; + result.reserve(size()); + for (auto&& iv : *this) { + result.push_back(std::move(iv)); + } + return result; + } + + // More compatibility shims for the overwhelming amount of code that + // likes to copy tuple elements into a vector; see comment above the + // copy constructor. + operator std::vector() const & { + return vec(); + } + + operator std::vector() && { + return vec(); + } +}; + +template +struct TupleTypeFactory {}; + +template <> +struct TORCH_API TupleTypeFactory { + static TupleTypePtr create(std::vector types) { + return TupleType::create(std::move(types)); + } + static TupleTypePtr fallback(const Type& type); +}; + +template <> +struct TORCH_API TupleTypeFactory { + static DynamicTypePtr create(const std::vector& elemTypes); + static DynamicTypePtr fallback(const Type& /*unused*/); +}; + +struct TORCH_API Tuple : c10::intrusive_ptr_target { + private: + TupleElements elements_; + mutable c10::TypePtr type_; // lazily computed for unnamed tuples + + public: + // named tuples have additional type information, so we + // directly create them tagged + static c10::intrusive_ptr createNamed( + std::vector elements_, + c10::TypePtr type_) { + return c10::make_intrusive(std::move(elements_), std::move(type_)); + } + + static c10::intrusive_ptr createNamed( + TupleElements elements_, + std::shared_ptr type_) { + return c10::make_intrusive(std::move(elements_), std::move(type_)); + } + + static c10::intrusive_ptr createNamed( + std::initializer_list elements_, + std::shared_ptr type_) { + return createNamed(TupleElements(c10::ArrayRef(elements_)), std::move(type_)); + } + + // MSVC apparently can't disambiguate the other two overloads of + // create when passed an initializer_list without this. + static c10::intrusive_ptr create(std::initializer_list elements_) { + return create(c10::ArrayRef(elements_)); + } + + static c10::intrusive_ptr create(std::vector elements_) { + return c10::make_intrusive(std::move(elements_)); + } + + static c10::intrusive_ptr create(TupleElements elements_) { + return c10::make_intrusive(std::move(elements_)); + } + + static c10::intrusive_ptr create(c10::ArrayRef elements_) { + return create(TupleElements(elements_)); + } + + static c10::intrusive_ptr create(IValue e1) { + return c10::make_intrusive(std::move(e1)); + } + + static c10::intrusive_ptr create(IValue e1, IValue e2) { + return c10::make_intrusive(std::move(e1), std::move(e2)); + } + + static c10::intrusive_ptr create(IValue e1, IValue e2, IValue e3) { + return c10::make_intrusive(std::move(e1), std::move(e2), std::move(e3)); + } + + private: + // Workaround inability to use `>` operator in template argument list. + template + static constexpr bool hasMoreThanThreeArgs() { + return sizeof...(Args) > 3; + } + + public: + template + static c10::intrusive_ptr create(Args&&... elements_) { + switch (sizeof...(Args)) { + case 1: + case 2: + case 3: + return create(IValue(std::forward(elements_))...); + default: + return create( + std::vector{IValue(std::forward(elements_))...}); + } + } + + // Again, it would be nice to make this noncopyable, but there's a + // lot of extant code that copies Tuples. + // Tuple(const Tuple& rhs) = delete; + + const TupleElements& elements() const& { + return elements_; + } + + TupleElements elements() && { + return std::move(elements_); + } + + void setElements(std::vector&& elements) { + elements_.setContents(std::move(elements)); + } + + void setElements(TupleElements&& elements) { + elements_ = std::move(elements); + } + + void unsafeSetElement(size_t idx, const IValue& element) { + elements_[idx] = element; + } + + void unsafeSetElement(size_t idx, IValue&& element) { + elements_[idx] = std::move(element); + } + + size_t size() const { + return elements_.size(); + } + + template + std::shared_ptr type() const { + if (!type_) { + type_ = TupleTypeFactory::create(fmap(elements(), [&](const IValue& v) { + return v.type(); + })); + } + if (auto t = type_->cast()) { + return t; + } + return TupleTypeFactory::fallback(*type_); + } + + static size_t hash(const Tuple& t) { + return c10::get_hash(t.elements()); + } + + TORCH_API friend bool operator==( + const ivalue::Tuple& lhs, + const ivalue::Tuple& rhs); + + private: + // NOTE: If we try to avoid the overloads without + // `std::shared_ptr type` by defaulting it to nullptr, we + // end up having to call (part of) the shared_ptr destructor for + // `type` even though we should know statically it won't do + // anything. + explicit Tuple(std::vector elements) + : elements_(std::move(elements)){} + + explicit Tuple(std::vector elements, c10::TypePtr type) + : elements_(std::move(elements)), type_(std::move(type)) {} + + explicit Tuple(TupleElements&& elements) + : elements_(std::move(elements)) {} + + explicit Tuple(TupleElements&& elements, std::shared_ptr type) + : elements_(std::move(elements)), type_(std::move(type)) {} + + explicit Tuple(IValue&& e1) + : elements_(std::move(e1)) {} + + explicit Tuple(IValue&& e1, std::shared_ptr type) + : elements_(std::move(e1)), type_(std::move(type)) {} + + explicit Tuple(IValue&& e1, IValue&& e2) + : elements_(std::move(e1), std::move(e2)) {} + + explicit Tuple(IValue&& e1, IValue&& e2, std::shared_ptr type) + : elements_(std::move(e1), std::move(e2)), type_(std::move(type)) {} + + explicit Tuple(IValue&& e1, IValue&& e2, IValue&& e3) + : elements_(std::move(e1), std::move(e2), std::move(e3)) {} + + explicit Tuple(IValue&& e1, IValue&& e2, IValue&& e3, std::shared_ptr type) + : elements_(std::move(e1), std::move(e2), std::move(e3)), type_(std::move(type)) {} + + friend class c10::intrusive_ptr; +}; + +struct Object; +struct PyObjectHolder; +struct EnumHolder; +} // namespace ivalue + +// Future +struct C10_EXPORT ivalue::Future final : c10::intrusive_ptr_target { + private: + // Keep this private in order to force users to go through make_intrusive and + // thus prevent creating a Future that's not held by an intrusive_ptr. + explicit Future(TypePtr type, std::vector devices={}) + : type_(std::move(type)), + impl_(getTypeOfDevices(devices)), + devices_(sortAndDeduplicateDevices(impl_, std::move(devices))) {} + + friend c10::intrusive_ptr; + + struct FutureCallback { + std::function callback; + bool uses_future; // whether the Future& passed in is actually used + + template + FutureCallback(T callback, bool uses_future) + : callback(std::move(callback)), uses_future(uses_future) {} + }; + + public: + Future(const Future&) = delete; + Future(Future&&) = delete; + Future& operator=(const Future&) = delete; + Future& operator=(Future&&) = delete; + + // Destructor + // Explicitly destroy events under device guard, otherwise it can lead to + // extra context being created on device 0. Reason: python garbage collector + // calls this destructor, but python GC does not have a device context, so a + // "default" one (usually on device 0) could be created when we go down the + // line of event destroy. + ~Future() override { + while (!events_.empty()) { + c10::OptionalDeviceGuard deviceGuard(events_.back().device()); + events_.pop_back(); + } + } + + struct TORCH_API FutureError final : public std::exception { + explicit FutureError(std::string&& error_msg_) + : error_msg(std::move(error_msg_)) {} + + FutureError() = default; + + const char* what() const noexcept override { + return error_msg.c_str(); + } + + std::string error_msg; + }; + + /** + * Wait on the future until it completes. + */ + void wait() { + std::unique_lock lock(mutex_); + finished_cv_.wait(lock, [&]() -> bool { return completed_; }); + synchronizeWithCurrentStreams(); + } + + /** + * Wait on the future until it completes and throw an + * exception if an error exists. + */ + void waitAndThrow() { + wait(); + + if (eptr_) { + std::rethrow_exception(eptr_); + } + } + + /** + * Explicitly mark the future as completed with the output value. Optionally, + * the storages for all tensors in IValue can be passed as well. The DataPtrs + * of these storages are used to synchronize CUDA streams. If storages isn't + * given we will attempt to extract it from the value, if we need to (this + * happens if a non-empty set of devices was given to the constructor). Thus + * one only needs to provide storages when 1) they cannot be extracted through + * IValue::getSubValues() or through pickling in case of Python object; or + * when 2) customized storage extraction is more efficient. + */ + using WeakStorage = c10::weak_intrusive_ptr; + void markCompleted( + IValue value, + std::optional> storages = std::nullopt) { + // Start by performing all steps that can throw, before setting any field. + // Do this before even acquiring the mutex, because extractStorages might + // acquire the GIL, which could lead to a lock inversion with our mutex. + // See https://github.com/pytorch/pytorch/issues/58239. + std::vector actualStorages; + std::vector usedDevices; + try { + // FIXME We should always extract DataPtrs, in order to catch the case of + // users using CUDA values but forgetting to set devices, which currently + // leads to a silent synchronization/correctness issue. However, as this + // might worsen perf in CPU-only cases, we should only do so after careful + // benchmarks. + if (impl_.type() != c10::kCPU) { + actualStorages = + storages.has_value() ? std::move(*storages) : extractStorages(value); + usedDevices = getDevicesOfStorages(impl_, actualStorages); + ensureIsSubsetOfDevices(usedDevices, devices_); + } + } catch (const std::exception&) { + setError(std::current_exception()); + return; + } + + std::unique_lock lock(mutex_); + TORCH_CHECK( + !completed(), + "Attempting to mark a completed Future as complete again. Note that " + "a Future can only be marked completed once."); + + // Only set value_ and completed_ flag once all checks and preparation steps + // have returned successfully to allow for proper error propagation. + value_ = std::move(value); + completed_ = true; + + currentDevice_ = impl_.getDevice(); + storages_ = std::move(actualStorages); + for (const c10::Device& device : usedDevices) { + c10::Event event(impl_.type()); + event.record(impl_.getStream(device)); + events_.push_back(std::move(event)); + } + + std::vector cbs; + cbs.swap(callbacks_); + lock.unlock(); + + finished_cv_.notify_all(); + for (const auto& callback : cbs) { + invokeCallback(callback.callback, callback.uses_future); + } + } + + void markCompleted() { + markCompleted(IValue{}); + } + + void setError(std::exception_ptr eptr) { + std::unique_lock lock(mutex_); + setErrorInternal(std::move(eptr), lock); + } + + void setErrorIfNeeded(std::exception_ptr eptr) { + std::unique_lock lock(mutex_); + if (completed_) { + // This should be rare and shouldn't cause log spew. Its important to + // log errors and that's why we have this log here. + std::string msg = c10::str( + "Skipping setting following error on the Future since " + "it is already marked completed (this is not necessarily " + "an error):\n", + tryRetrieveErrorMessageInternal(std::move(eptr))); + if (eptr_) { + msg += c10::str( + ", \nOriginal exception:\n", + tryRetrieveErrorMessageInternal(eptr_)); + } + LOG(INFO) << msg; + return; + } else { + setErrorInternal(std::move(eptr), lock); + } + } + + // Get the result of the current future. + IValue value() { + std::unique_lock lock(mutex_); + AT_ASSERT(completed()); + if (eptr_) { + std::rethrow_exception(eptr_); + } + return value_; + } + + // This accessor should only be used if we know that the future is + // completed() with no error. + const IValue& constValue() const { + std::unique_lock lock(mutex_); + AT_ASSERT(completed()); + TORCH_INTERNAL_ASSERT( + !eptr_, + "value() accessor should only be used when future is not completed with ", + "an error, but future had the following error: ", + tryRetrieveErrorMessageInternal(eptr_) + ); + return value_; + } + + // This accessor should only be used if we know that the future is + // completed() with no error. + const std::vector& storages() const { + std::unique_lock lock(mutex_); + AT_ASSERT(completed()); + AT_ASSERT(!eptr_); + return storages_; + } + + /** + * Add a callback to the future. + * The callbacks will be executed once the future completes. + * If the future has already completed, + * this function will execute the callback immediately. + */ + template + void addCallback(T callback, bool uses_future = true) { + static_assert( + std::is_invocable_r_v, + "The callback must have signature void(Future&)"); + + std::unique_lock lock(mutex_); + if (completed()) { + lock.unlock(); + invokeCallback(callback, uses_future); + return; + } + callbacks_.emplace_back(std::move(callback), uses_future); + } + + /** + * Add a callback to the future, and return another Future to hold the return + * value of the callback. This is necessary when the callback provider needs + * to know for sure when the callback has finished. + */ + template + c10::intrusive_ptr then(T callback, TypePtr type) { + using IValueWithStorages = std::tuple>; + static_assert( + std::disjunction_v< + std::is_invocable_r, + std::is_invocable_r>, + "The callback must have signature IValue(Future&) or " + "std::tuple>(Future&)"); + + auto childFut = createInstance(::std::move(type)); + addCallback([childFut, + cb = std::move(callback)](Future& parentFut) { + try { + if constexpr (::std::is_convertible_v, IValueWithStorages>) { + auto [ivalue, storages] = cb(parentFut); + childFut->markCompleted(::std::move(ivalue), ::std::move(storages)); + } else { + childFut->markCompleted(cb(parentFut)); + } + } catch (std::exception&) { + childFut->setError(std::current_exception()); + } + }); + return childFut; + } + + template + c10::intrusive_ptr thenAsync(T callback, TypePtr type) { + static_assert( + std::is_invocable_r_v, T, Future&>, + "The callback must have signature c10::intrusive_ptr(Future&)"); + + auto childFut = createInstance(std::move(type)); + addCallback( + [childFut, cb = std::move(callback)](Future& parentFut) mutable { + c10::intrusive_ptr intermediateFut; + try { + intermediateFut = cb(parentFut); + } catch (std::exception&) { + childFut->setError(std::current_exception()); + return; + } + intermediateFut->addCallback( + [childFut = std::move(childFut)](Future& intermediateFut) { + if (intermediateFut.hasError()) { + childFut->setError(intermediateFut.exception_ptr()); + } else { + childFut->markCompleted( + intermediateFut.value(), intermediateFut.storages()); + } + }); + }); + return childFut; + } + + // Tries to retrieve the error message from std::exception_ptr. + std::string tryRetrieveErrorMessage() const { + TORCH_CHECK(hasError(), "No error present on the future."); + std::unique_lock lock(mutex_); + return tryRetrieveErrorMessageInternal(eptr_); + } + + // Check if the current future has completed + bool completed() const { + return completed_; + } + + bool hasValue() const { + std::unique_lock lock(mutex_); + return completed_ && !eptr_; + } + + bool hasError() const { + std::unique_lock lock(mutex_); + return eptr_ ? true : false; + } + + std::exception_ptr exception_ptr() const { + std::unique_lock lock(mutex_); + return eptr_; + } + + TORCH_API friend std::ostream& operator<<( + std::ostream& out, + const Future& v); + + const TypePtr& elementType() const { + return type_; + } + + const std::vector& devices() const { + return devices_; + } + + // This method should be used when one intends to manually create a child + // future, for example when implementing a customized version of then(). + c10::intrusive_ptr createInstance(at::TypePtr type) { + return c10::make_intrusive(std::move(type), devices_); + } + + private: + + // This method should always be used when invoking a callback (regardless of + // how/when that happens) as it will ensure that the proper "environment" is + // set up before running the callback, as in, it will set up the CUDA streams, + // synchronize them with the value, and so on (if needed). + template + void invokeCallback(T& callback, bool uses_future) { + static_assert( + std::is_invocable_r_v, + "The callback must have signature void(Future&)"); + + // The synchronization performed below shouldn't be needed when the future + // is not used by the callback. + if (uses_future) { + c10::OptionalDeviceGuard deviceGuard(currentDevice_); + + std::vector streams; + streams.reserve(devices_.size()); + for (const c10::Device& device : devices_) { + streams.push_back(impl_.getStreamFromGlobalPool(device)); + } + c10::MultiStreamGuard streamGuard(streams); + synchronizeWithCurrentStreams(); + callback(*this); + } else { + callback(*this); + } + } + + // This method should be called before this future's value is used, as it + // ensures that the CUDA streams that are "current" at the callsite properly + // synchronize with the value. + void synchronizeWithCurrentStreams() { + for (c10::Event& event : events_) { + event.block(impl_.getStream(event.device())); + } + + for (const WeakStorage& weak_storage : storages_) { + c10::intrusive_ptr storage = weak_storage.lock(); + if (!storage) { + continue; + } + if (!storage->device().is_cpu()) { + impl_.recordDataPtrOnStream( + storage->data_ptr(), impl_.getStream(storage->device())); + } + } + } + + void setErrorInternal( + std::exception_ptr eptr, + std::unique_lock& lock) { + TORCH_CHECK( + !eptr_, + "Error already set on this Future: ", + tryRetrieveErrorMessageInternal(eptr_), + ", trying to set error: ", + tryRetrieveErrorMessageInternal(eptr)); + TORCH_INTERNAL_ASSERT(!completed(), "Future is already marked completed"); + completed_ = true; + eptr_ = std::move(eptr); + + std::vector cbs; + cbs.swap(callbacks_); + lock.unlock(); + + finished_cv_.notify_all(); + for (const auto& callback : cbs) { + invokeCallback(callback.callback, callback.uses_future); + } + } + + // Tries to retrieve the error message from std::exception_ptr. + std::string tryRetrieveErrorMessageInternal(std::exception_ptr eptr) const { + try { + std::rethrow_exception(std::move(eptr)); + } catch (const std::exception& e) { + return e.what(); + } catch (...) { + return "Unknown Exception Type"; + } + } + + // Defined in ivalue.cpp. + static std::vector extractStorages( + const at::IValue& value); + + static std::vector getDevicesOfStorages( + const c10::impl::VirtualGuardImpl& impl, + const std::vector& storages) { + c10::DeviceIndex deviceCount = impl.deviceCount(); + std::vector isDeviceUsed(deviceCount, false); + for (const WeakStorage& weak_storage : storages) { + c10::intrusive_ptr storage = weak_storage.lock(); + if (!storage) { + continue; + } + c10::Device device = storage->device(); + if (!device.is_cpu()) { + TORCH_CHECK_VALUE( + device.type() == impl.type(), + "Expected all data ptrs to be on a device of type ", + impl.type(), + ", got one on device ", + device); + isDeviceUsed[device.index()] = true; + } + } + std::vector devices; + for (c10::DeviceIndex idx = 0; idx < deviceCount; idx++) { + if (isDeviceUsed[idx]) { + devices.emplace_back(impl.type(), idx); + } + } + return devices; + } + + static std::string formatSetOfDevices( + const std::vector& devices) { + if (devices.empty()) { + return "(none)"; + } + std::ostringstream oss; + oss << devices[0]; + for (const auto idx : c10::irange(1, devices.size())) { + if (idx == devices.size() - 1) { + oss << " and "; + } else { + oss << ", "; + } + oss << devices[idx]; + } + return oss.str(); + } + + static c10::DeviceType getTypeOfDevices( + const std::vector& devices) { + if (devices.empty()) { + return c10::kCPU; + } + c10::DeviceType deviceType = devices[0].type(); + for (const auto idx : c10::irange(1, devices.size())) { + TORCH_CHECK_VALUE( + devices[idx].type() == deviceType, + "Expected all devices to be of the same type, but got a mismatch between ", + devices[0], + " and ", + devices[idx]); + } + return deviceType; + } + + // We need devices to be sorted in order to use ensureIsSubsetOfDevices. + static std::vector sortAndDeduplicateDevices( + const c10::impl::VirtualGuardImpl& /*impl*/, + std::vector devices) { + std::sort( + devices.begin(), devices.end(), + [](const c10::Device& a, const c10::Device& b) { return a.index() < b.index(); }); + // Deduplicate by compacting. + size_t targetIdx = 0; + for (const auto sourceIdx : c10::irange(devices.size())) { + TORCH_CHECK_VALUE( + devices[sourceIdx].has_index(), + "Expected devices to have indices, got ", devices[sourceIdx]); + if (targetIdx > 0 && devices[targetIdx - 1].index() == devices[sourceIdx].index()) { + // It's a duplicate, skip it. + continue; + } + if (sourceIdx != targetIdx) { + devices[targetIdx] = devices[sourceIdx]; + } + targetIdx++; + } + // If there were duplicates there's now a gap at the end: trim it. Resizing + // requires the item type to be default-constructible (which c10::Device is + // not) because in principle it could be required to create new items. Since + // we know we'll shrink the vector, we provide a custom dummy value instead. + devices.resize(targetIdx, c10::Device(c10::kCPU)); + return devices; + } + + static void ensureIsSubsetOfDevices( + const std::vector& subset, + const std::vector& superset) { + // We assume the devices in both vectors have the same consistent type, and + // their indices are unique and sorted. + std::vector excessDevices; + std::set_difference( + subset.begin(), + subset.end(), + superset.begin(), + superset.end(), + std::back_inserter(excessDevices), + [](const c10::Device& a, const c10::Device& b) { return a.index() < b.index(); }); + TORCH_CHECK_VALUE( + excessDevices.empty(), + "The result contained tensors residing on device(s) ", + formatSetOfDevices(excessDevices), + " which are not among the expected device(s) ", + formatSetOfDevices(superset)); + } + + mutable std::mutex mutex_; + std::atomic_bool completed_ = {false}; // is this future complete + std::condition_variable finished_cv_; + + IValue value_; // when finished the value + TypePtr type_; + std::vector callbacks_; + std::exception_ptr eptr_; + + // An upcast pointer to a virtual class which allows us to manipulate events, + // streams, ... in a generic way, without an explicit dependency on CUDA. + // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members) + const c10::impl::VirtualGuardImpl impl_; + + // The device that was current when markCompleted was called, which we'll + // restore when invoking callbacks. It's optional because we'll only store it + // if the future completes successfully. + std::optional currentDevice_; + + // The events that correspond to the completion of the async I/O kernels. They + // are recorded on the appropriate streams when the future is marked completed + // and can then be queried/waited/blocked on. There is one event for each + // distinct device on which the value's tensors reside. + std::vector events_; + + // A cached version of the storages extracted from the value when the future + // is first marked completed. + std::vector storages_; + + // The bounding set of devices that this future, and any of its children, is + // allowed to use. This is a superset of the set of devices used by the events + // above. We need this to know what streams (for which devices) to set as + // current when invoking a callback, thus allowing the callback to use devices + // that the parent future didn't use. This field is set to the value provided + // in the constructor and will be "inherited" by all child futures. + // NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members) + const std::vector devices_; +}; + +struct C10_EXPORT ivalue::Await final : c10::intrusive_ptr_target { + private: + explicit Await(TypePtr elType, std::function fn) + : elType_(std::move(elType)), type_(AwaitType::create(elType_)), fn_(std::move(fn)) {} + + explicit Await(TypePtr elType) : elType_(std::move(elType)), type_(AwaitType::create(elType_)) { } + + friend c10::intrusive_ptr; + + public: + Await(const Await&) = delete; + Await(Await&&) = delete; + Await& operator=(const Await&) = delete; + Await& operator=(Await&&) = delete; + ~Await() override = default; + + IValue wait() { + if (!completed_) { + TORCH_CHECK(fn_, "Incompleted Await: fn can't be None"); + value_ = fn_(); + completed_ = true; + args_ = {}; + } + return value_; + } + + IValue value() { + TORCH_CHECK(completed_, "Await must be completed"); + return value_; + } + + void setFn(std::function fn) { + fn_ = std::move(fn); + } + + bool completed() { + return completed_; + } + + void markCompleted(IValue value) { + value_ = std::move(value); + completed_ = true; + } + + TORCH_API friend std::ostream& operator<<( + std::ostream& out, + const Await& v); + + const TypePtr& elementType() const { + return elType_; + } + + const TypePtr& type() const { + return type_; + } + + void setArgs(std::vector args) { + args_ = std::move(args); + } + + std::vector& args() { + return args_; + } + + private: + TypePtr elType_; + TypePtr type_; + std::vector args_; + std::function fn_; + IValue value_; + bool completed_{}; +}; + +// Input is a list of Futures with the same target type. +// Output is a Future to the List of completed Futures. +TORCH_API intrusive_ptr collectAll( + const c10::List>& srcs); +// Input is a List of Futures with the same target type. +// Output is a Future that will be updated with a seen value. +TORCH_API intrusive_ptr collectAny( + const c10::List>& srcs); + +// User-defined object. +struct C10_EXPORT ivalue::Object final : c10::intrusive_ptr_target { + public: + // In general, class types hold a shared_ptr to its owning CompilationUnit, + // so that its type and methods do not get deallocated while the class exists. + // However, the CompilationUnit holds ownership of the type's graphs, so + // inserting a constant object into a Graph would create a reference cycle if + // that constant object held a shared_ptr to its CU. For these objects we + // instantiate them with non-owning references to its CU + Object(WeakOrStrongTypePtr type, size_t numSlots) : type_(std::move(type)) { + slots_.resize(numSlots); + } + + Object(StrongTypePtr type, size_t numSlots) + : type_(WeakOrStrongTypePtr(std::move(type))) { + slots_.resize(numSlots); + } + + static c10::intrusive_ptr create( + WeakOrStrongTypePtr type, + size_t numSlots) { + return c10::make_intrusive(std::move(type), numSlots); + } + + static c10::intrusive_ptr create( + StrongTypePtr type, + size_t numSlots) { + return c10::make_intrusive(std::move(type), numSlots); + } + + static c10::intrusive_ptr create(ClassTypePtr classType, size_t numSlots); + + /** + * Slot API. + * + * Attributes are stored as a simple vector so that lookups are fast at + * runtime. A "slot" is just an index into that vector, which can be computed + * statically if you have access to the class type. Use this API if you are + * writing compiler stuff. + */ + void setSlot(size_t slot, IValue v) { + if (slot >= slots_.size()) { + // for module types, it is possible that the members of the class have + // expanded after the object was created. In this case, we expand + // the slots to the right size + resizeObject(slot); + } + slots_[slot] = std::move(v); + } + + const IValue& getSlot(size_t slot) const { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(slot < slots_.size()); + // NOTE: This lookup is fairly hot, so we use unchecked access to the + // vector. Errors should still be detectable with ASan. + return slots_[slot]; + } + + void unsafeRemoveSlot(size_t slot) { + TORCH_CHECK(slot < slots_.size()); + slots_.erase(slots_.begin() + static_cast(slot)); + } + + /** + * Attribute API. + * + * Wrappers around the slot stuff so that users can access attributes + * directly. Use this API if you are a user. + * + * Note: Unlike in Python, TorchScript must make a distinction between + * attributes (which are IValues) and methods (which are Methods). If you + * want a method, use `obj.type()->getMethod()` + */ + IValue getAttr(const std::string& name) const; + void setAttr(const std::string& name, IValue v); + // Remove attribute by name, caller is responsible for + // the safety of this operation + // We didn't remove the attribute in the type because the type + // might be shared by multiple objects. + // Therefore after removing attribute, the object is in an inconsistent + // state where it has more attribute types in its Type than + // the attribute slots it has, user needs to make sure the object + // has consistent by removing the attribute in type as well + void unsafeRemoveAttr(const std::string& name); + + std::string name() const; + + const std::vector& slots() const { + return slots_; + } + std::shared_ptr type() const; + + std::shared_ptr compilation_unit() { + if (type_.holds_strong_ref()) { + return type_.cu_.getStrongRefOrThrow(); + } else { + auto weak_ptr = type_.cu_.getWeakRefOrThrow(); + return std::shared_ptr(weak_ptr); + } + } + + c10::intrusive_ptr copy_to_weak_compilation_ref() const; + + void unsafe_make_weak_compilation_ref() { + type_ = WeakOrStrongTypePtr(type_.asWeakTypePtr()); + } + + c10::intrusive_ptr copy() const; + + c10::intrusive_ptr deepcopy( + std::optional device = std::nullopt) const; + + c10::intrusive_ptr deepcopy( + IValue::HashIdentityIValueMap& memo, + std::optional device = std::nullopt) const; + + bool is_weak_compilation_ref() const { + return !type_.holds_strong_ref(); + } + + bool is_empty_strong_compilation_ref() const { + return type_.holds_empty_strong_ref(); + } + + private: + void resizeObject(size_t slot); + WeakOrStrongTypePtr type_; + std::vector slots_; +}; + +// virtual ivalue PyObjectHolder that hold a py::object, we make this virtual +// because the py::object and refcounting logic should happen in libtorch_python +// see concrete implementation in python_ivalue.h +struct ivalue::PyObjectHolder : c10::intrusive_ptr_target { + public: + virtual PyObject* getPyObject() = 0; + virtual c10::InferredType tryToInferType() = 0; + virtual IValue toIValue(const TypePtr& type, std::optional N = std::nullopt) = 0; + virtual std::string toStr() = 0; + virtual std::vector extractTensors() = 0; + + ~PyObjectHolder() override = default; +}; + +struct ivalue::EnumHolder : c10::intrusive_ptr_target { + public: + EnumHolder(std::shared_ptr type, std::string name, IValue value) + : type_(std::move(type)), + name_(std::move(name)), + value_(std::move(value)) {} + + bool is(const ivalue::EnumHolder& rhs) { + return *this == rhs; + } + + friend bool operator==( + const ivalue::EnumHolder& lhs, + const ivalue::EnumHolder& rhs); + + TORCH_API friend std::ostream& operator<<( + std::ostream& out, + const ivalue::EnumHolder& v); + + TORCH_API const std::string& qualifiedClassName() const; + + const std::string& unqualifiedClassName() const; + + const std::string& name() const { + return name_; + } + + const IValue& value() const { + return value_; + } + + std::shared_ptr type() const { + return type_; + } + + private: + std::shared_ptr type_; + std::string name_; + IValue value_; +}; + +#undef TORCH_FORALL_TAGS + +namespace detail { + +struct _guarded_unsigned_long_unique_dummy final { + _guarded_unsigned_long_unique_dummy(int64_t /*unused*/){} +}; +using _guarded_unsigned_long = std::conditional_t< + std::is_same_v || + std::is_same_v, + _guarded_unsigned_long_unique_dummy, + unsigned long>; + +} // namespace detail + +inline ivalue::Object& IValue::toObjectRef() const { + AT_ASSERT(isObject(), "Expected Object but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), "Attempted to create null reference"); + return *static_cast(payload.u.as_intrusive_ptr); +} + +// note: when adding a DEFINE_TO case here you should also add a +// toX method to IValue. These named methods are much more discoverable +// than the to templated function. + +#define DEFINE_TO(T, method_name) \ + template <> \ + inline T IValue::to()&& { \ + return static_cast(std::move(*this).method_name()); \ + } \ + template <> \ + inline c10::detail::ivalue_to_const_ref_overload_return::type IValue::to() const& { \ + typedef c10::detail::ivalue_to_const_ref_overload_return::type return_type; \ + return static_cast(this->method_name()); \ + } + +DEFINE_TO(at::Tensor, toTensor) +DEFINE_TO(at::Storage, toStorage) +DEFINE_TO(c10::Stream, toStream) +DEFINE_TO(float, toDouble) +DEFINE_TO(double, toDouble) +DEFINE_TO(c10::complex, toComplexDouble) +DEFINE_TO(unsigned char, toInt) +DEFINE_TO(signed char, toInt) +DEFINE_TO(unsigned short, toInt) +DEFINE_TO(short, toInt) +DEFINE_TO(int, toInt) +DEFINE_TO(uint32_t, toInt) +DEFINE_TO(uint64_t, toInt) +DEFINE_TO(detail::_guarded_unsigned_long, toInt) +DEFINE_TO(int64_t, toInt) +DEFINE_TO(bool, toBool) +DEFINE_TO(c10::intrusive_ptr, toBlob) +DEFINE_TO(c10::intrusive_ptr, toString) +DEFINE_TO(c10::intrusive_ptr, toObject) +DEFINE_TO(at::Scalar, toScalar) +DEFINE_TO(c10::List, toIntList) +DEFINE_TO(c10::List, toSymIntList) +DEFINE_TO(c10::List, toDoubleList) +DEFINE_TO(c10::List>, toComplexDoubleList) +DEFINE_TO(c10::List, toBoolList) +DEFINE_TO(c10::List, toTensorList) +DEFINE_TO(c10::impl::GenericList, toList) +DEFINE_TO(c10::impl::GenericDict, toGenericDict) +DEFINE_TO(c10::intrusive_ptr, toTuple) +DEFINE_TO(std::string, toStringRef) +DEFINE_TO(std::string_view, toStringView) +DEFINE_TO(c10::intrusive_ptr, toFuture) +DEFINE_TO(c10::intrusive_ptr, toAwait) +DEFINE_TO(c10::intrusive_ptr, toRRef) +DEFINE_TO(c10::intrusive_ptr, toQuantizer) +DEFINE_TO(IValue, toIValue) +DEFINE_TO(c10::Device, toDevice) +DEFINE_TO(at::ScalarType, toScalarType) +DEFINE_TO(at::Layout, toLayout) +DEFINE_TO(at::MemoryFormat, toMemoryFormat) +DEFINE_TO(at::QScheme, toQScheme) +DEFINE_TO(at::Dimname, toDimname) +DEFINE_TO(at::Generator, toGenerator) +DEFINE_TO(c10::SymInt, toSymInt) +DEFINE_TO(c10::SymFloat, toSymFloat) +DEFINE_TO(c10::SymBool, toSymBool) + +template +struct _fake_type {}; + +// generic_to converts an IValue from a generic list or generic dict +// to a concrete list/dict type likelike List, Dict<...> or std::optional. +// Note that in the case of lists, this only works for IValue-based lists, +// i.e. not for int64_t, double, ... +// generic_to is an implementation detail of IValue::to and not +// supposed to be called directly. +// The _fake_type parameter allows us to overload +// based on the return type. +template +// TODO this is deprecated but we don't throw a warning because a lot of ops in +// native_functions.yaml still return std::vector. +// C10_DEPRECATED_MESSAGE("IValues based on std::vector are potentially slow +// and deprecated. Please use torch::List instead.") +std::vector generic_to(IValue ivalue, _fake_type> /*unused*/) { + // We need to do a deep copy of the vector because there might be other + // references to this same IValue that also use the list. We can't just + // move the elements out. + auto list = std::move(ivalue).template to>(); + std::vector result; + result.reserve(list.size()); + for (Elem v : list) { + result.push_back(std::move(v)); + } + return result; +} + +template +c10::intrusive_ptr IValue::toCustomClass() && { + static_assert( + std::is_base_of_v == true, + "toCustomClass requires that template parameter T must inherit " + "from torch::CustomClassHolder"); + auto obj = toObject(); + TORCH_CHECK( + obj->slots().size() == 1, + "Tried to cast IValue to custom class but it did " + "not contain a custom class!"); + const auto* expected_type = c10::getCustomClassType>().get(); + ivalue::checkCustomClassType(expected_type, type().get()); + auto userObj = + c10::static_intrusive_pointer_cast(obj->getSlot(0).toCapsule()); + return userObj; +} + +template +c10::intrusive_ptr IValue::toCustomClass() const& { + static_assert( + std::is_base_of_v == true, + "toCustomClass requires that template parameter T must inherit " + "from torch::CustomClassHolder"); + auto obj = toObject(); + TORCH_CHECK( + obj->slots().size() == 1, + "Tried to cast IValue to custom class but it did " + "not contain a custom class!"); + const auto* expected_type = c10::getCustomClassType>().get(); + ivalue::checkCustomClassType(expected_type, type().get()); + auto userObj = + c10::static_intrusive_pointer_cast(obj->getSlot(0).toCapsule()); + return userObj; +} + +template +T generic_to(IValue ivalue, _fake_type /*unused*/) { + using ElemType = typename std::remove_pointer::type::element_type; + return std::move(ivalue).template toCustomClass(); +} + +template +tagged_capsule generic_to(IValue ivalue, _fake_type> /*unused*/) { + return tagged_capsule{std::move(ivalue)}; +} + +template +c10::List generic_to(IValue ivalue, _fake_type> /*unused*/) { + return impl::toTypedList(std::move(ivalue).toList()); +} + +template +static T createVectorLikeFromList(const c10::detail::ListImpl* impl) { + T result; + result.reserve(impl->list.size()); + for (const auto & i : impl->list) { + result.push_back(i.to()); + } + return result; +} + +template +static std::vector createVectorFromList(const c10::detail::ListImpl* impl) { + return createVectorLikeFromList>(impl); +} + +template +std::vector createVectorFromList(const c10::List& impl) { + std::vector result; + result.reserve(impl.size()); + for (size_t i = 0, N = impl.size(); i < N; ++i) { + result.push_back(impl[i]); + } + return result; +} + +template +OptionalArray generic_to(IValue ivalue, _fake_type> /*unused*/) { + if (ivalue.isNone()) { + return {}; + } + return createVectorFromList( + std::move(ivalue).template to>() + ); +} + +namespace detail { +template +std::array generic_to_array( + IValue ivalue, + _fake_type> /*unused*/, + std::index_sequence /*unused*/) { + // We need to do a deep copy of the array because there might be other + // references to this same IValue that also use the list. We can't just + // move the elements out. + auto list = std::move(ivalue).template to>(); + TORCH_CHECK( + list.size() == sizeof...(I), + "Tried to convert a List with ", + list.size(), + " elements to a fixed-size array of size ", + sizeof...(I)); + return {list[I]...}; +} +} // namespace detail + +template +std::array generic_to( + IValue ivalue, + _fake_type> ft) { + return detail::generic_to_array(ivalue, ft, std::make_index_sequence()); +} + +template +c10::Dict generic_to( + IValue ivalue, + _fake_type> /*unused*/) { + return impl::toTypedDict(std::move(ivalue).toGenericDict()); +} + +template +C10_DEPRECATED_MESSAGE( + "IValues based on std::unordered_map are slow and deprecated. Please use c10::Dict instead.") +std::unordered_map generic_to( + IValue ivalue, + _fake_type> /*unused*/) { + std::unordered_map specialized_dict; + + for (const auto& item : std::move(ivalue).toGenericDict()) { + specialized_dict[item.key().template to()] = item.value().template to(); + } + + return specialized_dict; +} + +template +std::optional generic_to(IValue ivalue, _fake_type> /*unused*/) { + if (ivalue.isNone()) { + return std::nullopt; + } + return std::move(ivalue).template to(); +} + +namespace detail { +template +Tuple generic_to_tuple_impl( + const ivalue::TupleElements& t, + std::index_sequence /*unused*/) { + return std::make_tuple( + t[INDEX].to::type>()...); +} +} // namespace detail + +template < + typename... Args, + typename Indices = std::make_index_sequence, + std::enable_if_t< + !std::disjunction_v< + std::is_lvalue_reference..., + std::negation>...>, + std::nullptr_t> = nullptr> +std::tuple generic_to(const IValue& ivalue, _fake_type> /*unused*/) { + const auto& vals = ivalue.toTupleRef().elements(); + TORCH_CHECK(vals.size() == sizeof...(Args)); + return detail::generic_to_tuple_impl>(vals, Indices{}); +} + +template +inline T IValue::to() && { + return generic_to(std::move(*this), _fake_type{}); +} + +template <> +inline std::optional IValue::to() && { + // In the default implementation, the IValue is destroyed with std::move. + // But if the unboxed type is std::optional we cannot destroy + // the IValue. + return generic_to(*this, _fake_type>{}); +} + +template +inline typename c10::detail::ivalue_to_const_ref_overload_return::type IValue::to() const& { + return generic_to(*this, _fake_type{}); +} + +inline c10::List IValue::toIntList() && { + AT_ASSERT(isIntList(), "Expected IntList but got ", tagKind()); + return c10::List(moveToIntrusivePtr()); +} +inline c10::List IValue::toIntList() const& { + AT_ASSERT(isIntList(), "Expected IntList but got ", tagKind()); + return c10::List(toIntrusivePtr()); +} +inline std::vector IValue::toIntVector() const { + AT_ASSERT(isIntList(), "Expected IntList but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toIntVector on null intrusive_ptr IValue"); + return createVectorFromList( + static_cast(payload.u.as_intrusive_ptr)); +} +inline c10::List IValue::toSymIntList() && { + AT_ASSERT( + isSymIntList() || isIntList(), + "Expected SymIntList or IntList but got ", + tagKind()); + return c10::List(moveToIntrusivePtr()); +} +inline c10::List IValue::toSymIntList() const& { + AT_ASSERT( + isSymIntList() || isIntList(), + "Expected SymIntList or IntList but got ", + tagKind()); + return c10::List(toIntrusivePtr()); +} +inline std::vector IValue::toSymIntVector() const { + AT_ASSERT(isSymIntList() || isIntList(), "Expected SymIntList or IntList but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toSymIntVector on null intrusive_ptr IValue"); + return createVectorFromList( + static_cast(payload.u.as_intrusive_ptr)); +} +inline at::DimVector IValue::toDimVector() const { + AT_ASSERT(isIntList(), "Expected IntList but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toDimVector on null intrusive_ptr IValue"); + return createVectorLikeFromList( + static_cast(payload.u.as_intrusive_ptr)); +} +inline c10::List IValue::toDoubleList() && { + AT_ASSERT(isDoubleList(), "Expected DoubleList but got ", tagKind()); + return c10::List(moveToIntrusivePtr()); +} +inline c10::List IValue::toDoubleList() const& { + AT_ASSERT(isDoubleList(), "Expected DoubleList but got ", tagKind()); + return c10::List(toIntrusivePtr()); +} +inline std::vector IValue::toDoubleVector() const { + AT_ASSERT(isDoubleList(), "Expected DoubleList but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toDoubleVector on null intrusive_ptr IValue"); + return createVectorFromList( + static_cast(payload.u.as_intrusive_ptr)); +} +inline c10::List> IValue::toComplexDoubleList() && { + AT_ASSERT(isComplexDoubleList(), "Expected ComplexDoubleList but got ", tagKind()); + return c10::List>(moveToIntrusivePtr()); +} +inline c10::List> IValue::toComplexDoubleList() const& { + AT_ASSERT(isComplexDoubleList(), "Expected ComplexDoubleList but got ", tagKind()); + return c10::List>(toIntrusivePtr()); +} +inline std::vector> IValue::toComplexDoubleVector() const { + AT_ASSERT(isComplexDoubleList(), "Expected ComplexDoubleList but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toComplexDoubleVector on null intrusive_ptr IValue"); + return createVectorFromList>( + static_cast(payload.u.as_intrusive_ptr)); +} +inline c10::List IValue::toBoolList() && { + AT_ASSERT(isBoolList(), "Expected BoolList but got ", tagKind()); + return c10::List(moveToIntrusivePtr()); +} +inline c10::List IValue::toBoolList() const& { + AT_ASSERT(isBoolList(), "Expected BoolList but got ", tagKind()); + return c10::List(toIntrusivePtr()); +} +inline c10::List IValue::toTensorList() && { + AT_ASSERT(isTensorList(), "Expected TensorList but got ", tagKind()); + return c10::List(moveToIntrusivePtr()); +} +inline c10::List IValue::toTensorList() const& { + AT_ASSERT(isTensorList(), "Expected TensorList but got ", tagKind()); + return c10::List(toIntrusivePtr()); +} +inline std::vector IValue::toTensorVector() const { + AT_ASSERT(isTensorList(), "Expected TensorList but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toTensorVector on null intrusive_ptr IValue"); + return createVectorFromList( + static_cast(payload.u.as_intrusive_ptr)); +} +inline c10::List> IValue::toOptionalTensorList() && { + AT_ASSERT(isOptionalTensorList(), "Expected OptionalTensorList but got ", tagKind()); + return c10::List>(moveToIntrusivePtr()); +} +inline c10::List> IValue::toOptionalTensorList() const& { + AT_ASSERT(isOptionalTensorList(), "Expected OptionalTensorList but got ", tagKind()); + return c10::List>(toIntrusivePtr()); +} +inline std::vector> IValue::toOptionalTensorVector() const { + AT_ASSERT(isOptionalTensorList(), "Expected OptionalTensorList but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toOptionalTensorVector on null intrusive_ptr IValue"); + return createVectorFromList>( + static_cast(payload.u.as_intrusive_ptr)); +} +inline c10::List IValue::toList() && { + AT_ASSERT(isList(), "Expected GenericList but got ", tagKind()); + return c10::List(moveToIntrusivePtr()); +} +inline c10::List IValue::toList() const& { + AT_ASSERT(isList(), "Expected GenericList but got ", tagKind()); + return c10::List(toIntrusivePtr()); +} +inline c10::ArrayRef IValue::toListRef() const { + AT_ASSERT(isList(), "Expected GenericList but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toListRef on null intrusive_ptr IValue"); + return static_cast(payload.u.as_intrusive_ptr) + ->list; +} +inline c10::Dict IValue::toGenericDict() && { + AT_ASSERT(isGenericDict(), "Expected GenericDict but got ", tagKind()); + return c10::Dict(moveToIntrusivePtr()); +} +inline c10::Dict IValue::toGenericDict() const& { + AT_ASSERT(isGenericDict(), "Expected GenericDict but got ", tagKind()); + return c10::Dict(toIntrusivePtr()); +} +inline c10::intrusive_ptr IValue::toTuple() && { + AT_ASSERT(isTuple(), "Expected Tuple but got ", tagKind()); + return moveToIntrusivePtr(); +} +inline c10::intrusive_ptr IValue::toTuple() const& { + AT_ASSERT(isTuple(), "Expected Tuple but got ", tagKind()); + return toIntrusivePtr(); +} +inline ivalue::Tuple& IValue::toTupleRef() const { + AT_ASSERT(isTuple(), "Expected Tuple but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toTupleRef on null intrusive_ptr IValue"); + return *static_cast( + payload.u.as_intrusive_ptr); +} + +inline IValue::IValue(c10::intrusive_ptr v) + : tag(Tag::Tuple) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.release()); +} +template < + typename... Args, + std::enable_if_t< + !std::disjunction_v< + std::is_lvalue_reference..., + std::negation>...>, + std::nullptr_t>> +inline IValue::IValue(const std::tuple& t) + : IValue(std::apply(c10::ivalue::Tuple::create, t)) { +} + +template < + typename... Args, + std::enable_if_t< + !std::disjunction_v< + std::is_lvalue_reference..., + std::negation>...>, + std::nullptr_t>> +inline IValue::IValue(std::tuple&& t) + : IValue(std::apply(c10::ivalue::Tuple::create, std::move(t))) { +} + +inline IValue::IValue(c10::intrusive_ptr v) + : tag(Tag::String) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.release()); +} +inline IValue::IValue(std::string v) + : IValue(ivalue::ConstantString::create(std::move(v))) {} + +inline IValue::IValue(c10::impl::GenericList v) + : tag(Tag::GenericList) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.impl_.release()); +} + +template > +inline IValue::IValue(c10::List&& v) : IValue(impl::toList(std::move(v))) {} +template > +inline IValue::IValue(const c10::List& v) : IValue(impl::toList(v)) {} +template > +inline IValue::IValue(at::ArrayRef v) : IValue(c10::List()) { + auto list = to>(); + list.reserve(v.size()); + for (const auto& e : v) { + list.push_back(e); + } +} +template > +inline IValue::IValue(at::ArrayRef v) : IValue() { + auto vi = c10::asIntArrayRefSlowOpt(v); + if (vi.has_value()) { + // This list is entirely integers; ensure it is typed as + // an IntList so toIntList works + *this = IValue(*vi); + } else { + // This list has SymInts; type it as a SymInt + *this = IValue(impl::toList(c10::List())); + auto list = to>(); + list.reserve(v.size()); + for (const auto& e : v) { + list.push_back(e); + } + } +} +template > +inline IValue::IValue(at::OptionalArrayRef mb_v) : IValue() { + if (!mb_v.has_value()) return; + *this = IValue(*mb_v); +} +template > +inline IValue::IValue(const std::vector& v) : IValue() { + *this = IValue(at::ArrayRef(v)); +} +template > +inline IValue::IValue(std::vector&& v) : IValue() { + auto vi = c10::asIntArrayRefSlowOpt(v); + if (vi.has_value()) { + // This list is entirely integers; ensure it is typed as + // an IntList so toIntList works + *this = IValue(*vi); + } else { + // This list has SymInts; type it as a SymInt + *this = IValue(impl::toList(c10::List())); + auto list = to>(); + list.reserve(v.size()); + for (auto&& e : std::move(v)) { + list.push_back(std::move(e)); + } + } +} +template > +inline IValue::IValue(const std::vector& v) : IValue(c10::List()) { + auto list = to>(); + list.reserve(v.size()); + for (const auto& e : v) { + list.push_back(e); + } +} + +template > +inline IValue::IValue(std::vector&& v) : IValue(c10::List()) { + auto list = to>(); + list.reserve(v.size()); + if constexpr (std::is_same_v) { + for (auto e : v) { + list.push_back(e); + } + } else { + for (auto&& e : std::move(v)) { + list.push_back(std::move(e)); + } + } +} + +template > +inline IValue::IValue(c10::OptionalArrayRef v) : IValue() { + if (v.has_value()) { + *this = IValue(std::move(*v)); + } +} + +template +inline IValue::IValue(std::array v) : IValue(c10::List()) { + auto list = to>(); + list.reserve(v.size()); + for (auto& e : v) { + list.push_back(std::move(e)); + } +} + +template > +inline IValue::IValue(c10::IListRef v) : IValue() { + constexpr bool boxed_type_constructs_ivalue = + std::is_constructible_v::boxed_type>; + // First, we try to use the boxed value. + // If we fail (either it's not in the boxed state, or its boxed type + // can not construct an IValue), we fallback to copying the list. + if (boxed_type_constructs_ivalue && v.isBoxed()) { + *this = IValue(impl::toList(v.toBoxed())); + } else { + c10::List list; + list.reserve(v.size()); + for (const auto& t : v) { + list.push_back(t); + } + *this = IValue(impl::toList(std::move(list))); + } +} + +inline IValue::IValue(c10::impl::GenericDict v) + : tag(Tag::GenericDict) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.impl_.release()); +} +template +inline IValue::IValue(c10::Dict v) + : IValue(impl::toGenericDict(std::move(v))) {} + +template +inline IValue::IValue(std::unordered_map v) + : IValue(Dict()) { + auto dict = to>(); + dict.reserve(v.size()); + for (auto& e : v) { + dict.insert(std::move(e.first), std::move(e.second)); + } +} + +template > +inline IValue::IValue(std::optional v) : IValue() { + if (v.has_value()) { + *this = IValue(std::move(*v)); + } +} + +inline IValue::IValue(std::nullopt_t /*unused*/) : IValue() {} + +inline IValue::IValue(c10::intrusive_ptr v) + : tag(Tag::Object) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.release()); +} + +inline IValue::IValue(c10::intrusive_ptr v) + : tag(Tag::PyObject) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.release()); +} + +inline IValue::IValue(c10::intrusive_ptr v) + : tag(Tag::Enum) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.release()); +} + +inline IValue IValue::make_capsule( + intrusive_ptr blob) { + IValue iv; + iv.tag = Tag::Capsule; + iv.payload.u.as_intrusive_ptr = null_to_undefined_tensor(blob.release()); + return iv; +} + +template < + typename T, + std::enable_if_t, int>> +IValue::IValue(c10::intrusive_ptr custom_class) : tag(Tag::Object) { + auto classType = []() { + try { + return c10::getCustomClassType>(); + } catch (const c10::Error&) { + throw c10::Error( + "Trying to instantiate a class that isn't a registered custom class: " + + std::string(c10::util::get_fully_qualified_type_name())); + } + }(); + auto ivalue_obj = c10::ivalue::Object::create(std::move(classType), /* numSlots */1); + ivalue_obj->setSlot(0, IValue::make_capsule(std::move(custom_class))); + payload.u.as_intrusive_ptr = null_to_undefined_tensor(ivalue_obj.release()); + +} + +inline IValue::IValue(c10::intrusive_ptr v) + : tag(Tag::Future) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.release()); +} + +inline IValue::IValue(c10::intrusive_ptr v) + : tag(Tag::Await) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.release()); +} + +inline IValue::IValue(c10::intrusive_ptr v) + : tag(Tag::RRef) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.release()); +} + +inline IValue::IValue(c10::intrusive_ptr v) + : tag(Tag::Quantizer) { + payload.u.as_intrusive_ptr = null_to_undefined_tensor(v.release()); +} + +template +inline IValue::IValue(c10::complex c) + : tag(Tag::ComplexDouble) { + auto v = c10::make_intrusive(c); + payload.u.as_intrusive_ptr = v.release(); +} + +inline const std::string& IValue::toStringRef() const { + AT_ASSERT(isString(), "Expected String but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toStringRef on null intrusive_ptr IValue"); + return static_cast( + payload.u.as_intrusive_ptr) + ->string(); +} +inline std::optional> IValue:: + toOptionalStringRef() const { + if (isNone()) { + return std::nullopt; + } + AT_ASSERT(isString(), "Expected std::optional but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toOptionalStringRef on null intrusive_ptr IValue"); + return std::reference_wrapper( + static_cast(payload.u.as_intrusive_ptr) + ->string()); +} + +inline std::string_view IValue::toStringView() const { + AT_ASSERT(isString(), "Expected String but got ", tagKind()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + payload.u.as_intrusive_ptr != c10::UndefinedTensorImpl::singleton(), + "called toStringView on null intrusive_ptr IValue"); + return static_cast( + payload.u.as_intrusive_ptr) + ->string_view(); +} + +inline PyObject* IValue::toPyObject() const { + return toPyObjectHolder()->getPyObject(); +} + +template +inline std::optional IValue::toOptional() { + if (this->isNone()) { + return std::nullopt; + } + return this->to(); +} + +template +inline std::optional IValue::toOptional() const { + if (this->isNone()) { + return std::nullopt; + } + return this->to(); +} + +inline bool IValue::isCustomClass() const { + return torch::isCustomClass(*this); +} + +inline bool IValue::isSameIdentity(const IValue& rhs) const { + // We choose to not use memcmp for payload check due to potential random + // padding characters on union type + + // Semantics: + // 1. Immutable primitive values of the same type (Int, Double, None, Bool, + // Str) return value equality + // 2. If it is a tensor type, we need to take undefined tensor into account + // 3. Undefined_tensor is None and vice versa should be true + // 4. If it is a reference type (i.e. isIntrusivePtr()), then is True when + // the pointed-to object is the same. + // 5. False for all other comparisons. + if (this->isNone() && rhs.isNone()) { + return true; + } else if (this->isBool() && rhs.isBool()) { + // for bool type, do equality check + return this->toBool() == rhs.toBool(); + } else if (this->isTensor() && rhs.isTensor()) { + return this->payload.as_tensor.is_same(rhs.payload.as_tensor); + } else if (this->isTensor() && rhs.isNone()) { + // special case: undefined tensor and None are the same identity + return !this->payload.as_tensor.defined(); + } else if (this->isNone() && rhs.isTensor()) { + // special case: undefined tensor and None are the same identity + return !rhs.payload.as_tensor.defined(); + } else if (this->isInt() && rhs.isInt()) { + return this->toInt() == rhs.toInt(); + } else if (this->isDouble() && rhs.isDouble()) { + return this->toDouble() == rhs.toDouble(); + } else if (this->isString() && rhs.isString()) { + return this->toStringRef() == rhs.toStringRef(); + } else { + // for objects holding in IValue, do shallow compare on pointer address to + // testify the identity + return this->isIntrusivePtr() && rhs.isIntrusivePtr() && + this->payload.u.as_intrusive_ptr == rhs.payload.u.as_intrusive_ptr; + } +} + +namespace ivalue { +namespace detail { + +template +IValue from_(T&& x, std::true_type /*unused*/) { + return IValue(std::forward(x)); +} +template +IValue from_(c10::intrusive_ptr x, std::false_type /*unused*/) { + return IValue(std::move(x)); +} +template +IValue from_(T&& /*x*/, std::false_type /*unused*/) { + static_assert( + guts::false_t::value, + "You are calling from with a type that it doesn't support, and isn't a potential custom class (ie: is an intrusive_ptr)"); + return IValue(); +} +} // namespace detail + +template +IValue from(T&& x) { + return detail::from_( + std::forward(x), typename std::is_constructible::type{}); +} + +} // namespace ivalue + + +template <> +struct MaybeOwnedTraits { + using owned_type = IValue; + using borrow_type = IValue; + + static borrow_type createBorrow(const owned_type& from) { + if (!from.isPtrType()) { + return from; + } + if (from.isTensor()) { + return IValue(MaybeOwnedTraits::createBorrow(from.toTensor())); + } else { + return IValue(from.payload, from.tag); + } + } + + static void assignBorrow(borrow_type& lhs, const borrow_type& rhs) { + lhs.clearToNone(); + if (!rhs.isPtrType()) { + lhs = rhs; + } else if (rhs.isTensor()) { + lhs = IValue(MaybeOwnedTraits::createBorrow(rhs.toTensor())); + } else { + lhs = IValue(rhs.payload, rhs.tag); + } + } + + static void destroyBorrow(borrow_type& toDestroy) { + toDestroy.clearToNone(); + } + + static const owned_type& referenceFromBorrow(const borrow_type& borrow) { + return borrow; + } + + static const owned_type* pointerFromBorrow(const borrow_type& borrow) { + return &borrow; + } + + static bool debugBorrowIsValid(const borrow_type& /*unused*/) { + return true; + } +}; + +template <> +struct IValue::TagType { + static TORCH_API c10::TypePtr get(const IValue& /*v*/); +}; + +template <> +struct IValue::TagType { + static TORCH_API c10::TypePtr get(const IValue& /*v*/); +}; + +template +TypePtr IValue::type() const { + return IValue::TagType::get(*this); +} + +} // namespace c10 + +C10_DIAGNOSTIC_POP() + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/ivalue_to.h b/backends/pytorch/include/ATen/core/ivalue_to.h new file mode 100644 index 000000000..52a3f23bb --- /dev/null +++ b/backends/pytorch/include/ATen/core/ivalue_to.h @@ -0,0 +1,41 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include + +namespace at { +class Tensor; +} // namespace at + +namespace c10 { +struct IValue; +namespace detail { +// Determine the return type of `IValue::to() const &`. It's a const +// reference when possible and a copy otherwise. It is in this +// separate header so that List can use it as well. +template +struct ivalue_to_const_ref_overload_return { + using type = T; +}; + +template<> +struct ivalue_to_const_ref_overload_return { + using type = const at::Tensor&; +}; + +template<> +struct ivalue_to_const_ref_overload_return { + using type = const std::string&; +}; + +template<> +struct ivalue_to_const_ref_overload_return { + using type = const IValue&; +}; + +} // namespace detail +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/jit_type.h b/backends/pytorch/include/ATen/core/jit_type.h new file mode 100644 index 000000000..bec7f2e58 --- /dev/null +++ b/backends/pytorch/include/ATen/core/jit_type.h @@ -0,0 +1,2435 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include + + +namespace torch::jit { +struct Function; +} // namespace torch::jit + + +namespace c10 { + +template +class Dict; +struct IValue; +struct FunctionSchema; +struct NamedType; +using OptNameList = std::optional>; + +void standardizeVectorForUnion(std::vector& reference, std::vector* to_fill); +void standardizeVectorForUnion(std::vector* to_flatten); + +inline bool is_contiguous_strides( + const IntArrayRef sizes, + const IntArrayRef strides) { + size_t n_dim = sizes.size(); + if (n_dim == 0) { + return true; + } + + if (strides[n_dim - 1] != 1) { + return false; + } + + for (int i = static_cast(n_dim) - 2; i >= 0; i--) { + if (strides[i] != strides[i + 1] * sizes[i + 1]) { + return false; + } + } + return true; +} + +struct AnyType; +using AnyTypePtr = SingletonTypePtr; +// Any is the top of the type hierarchy, all other types are subtypes +// T <: Any, forall T +struct TORCH_API AnyType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "Any"; + } + static const TypeKind Kind = TypeKind::AnyType; + // global singleton + static AnyTypePtr get(); + + private: + AnyType() : Type(TypeKind::AnyType) {} +}; + +inline std::string toString(const Type& type) { + return type.str(); +} + +// Shim for compatibility with code that uses TypePtr. +inline std::string toString(const TypePtr& typePtr) { + return toString(*typePtr); +} + +inline bool operator!=(const Type& lhs, const Type& rhs) { + return !(lhs == rhs); +} + +// common base for all types that have a single sub element +// e.g. Future[T], Optional[T], List[T] +template +struct SingleElementType : public SharedType { + static const TypeKind Kind = K; + + const TypePtr& getElementType() const { + return elem; + } + + bool hasFreeVariables() const override { + return getElementType()->hasFreeVariables(); + } + + at::ArrayRef containedTypes() const override { + return elem; + } + + bool equals(const Type& rhs) const override { + if (auto rhs_ = rhs.cast()) { + return *getElementType() == *rhs_->getElementType(); + } + return false; + } + + protected: + SingleElementType(TypePtr elem) : SharedType(Kind), elem(std::move(elem)) { + TORCH_CHECK(this->elem, c10::str( + "Can not create ", typeKindToString(Kind), " with None type")); + } + + private: + TypePtr elem; +}; + +struct UnionType; +using UnionTypePtr = std::shared_ptr; +struct TORCH_API UnionType : public SharedType { + friend struct Type; + + static const TypeKind Kind = TypeKind::UnionType; + + bool isSubtypeOfExt(const Type& rhs_, std::ostream* why_not) const override; + + std::string str() const override; + + static UnionTypePtr create(std::vector reference); + + bool equals(const Type& rhs) const override; + + bool isUnionType() const override { + return true; + } + + at::ArrayRef containedTypes() const override { + return types_; + } + + // For testing purposes only + at::ArrayRef getTypes() const { + return types_; + } + + TypePtr createWithContained(std::vector contained_types) const override { + return create(std::move(contained_types)); + } + + bool canHoldType(const Type& type) const; + + bool hasFreeVariables() const override { + return has_free_variables_; + } + + std::optional toOptional() const; + + std::optional subtractTypeSet(std::vector& to_subtract) const; + + protected: + explicit UnionType(std::vector types, TypeKind kind=TypeKind::UnionType); + std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override; + std::string unionStr( + const TypePrinter& printer = nullptr, + bool is_annotation_str = false) const; + // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes) + bool has_free_variables_; + // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes) + std::vector types_; + // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes) + bool can_hold_none_; + +}; + +struct OptionalType; +using OptionalTypePtr = std::shared_ptr; +// This type represents an optional type. There is one `Optional` for +// each element type. `Optional[T]` can accept both `T` and +// `None`(`std::nullopt` in C++) +// Subtype hierarchy for Optional: +// - Optional[T] <: Optional[R] iff T <: R +// - T <: Optional[R] if T <: R +// - None <: Optional[T] for all T +// - Optional[T] == Union[T, None] for all T +struct TORCH_API OptionalType : public UnionType { + static OptionalTypePtr create(const TypePtr& contained); + + static const TypeKind Kind = TypeKind::OptionalType; + + friend struct Type; + + bool equals(const Type& rhs) const override; + + const TypePtr& getElementType() const { + return contained_; + } + + at::ArrayRef containedTypes() const override { + return contained_; + } + + std::string str() const override { + std::stringstream ss; + ss << getElementType()->str() << '?'; + return ss.str(); + } + + TypePtr createWithContained( + std::vector contained_types) const override { + AT_ASSERT(contained_types.size() == 1); + return create(contained_types[0]); + } + + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override; + + bool isUnionType() const override { + return true; + } + + // common cast Optional[Tensor] for undefined tensor type + static TypePtr ofTensor(); + // + // global singleton + static TypePtr get(TypePtr inner); + + private: + explicit OptionalType(const TypePtr& contained); + + TypePtr contained_; + + std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override { + std::stringstream ss; + ss << "Optional[" << getElementType()->annotation_str(printer) << ']'; + return ss.str(); + } +}; + +template +inline std::optional merge_primitive( + const std::optional& a, + const std::optional& b) { + if (a.has_value() && b.has_value() && a.value() == b.value()) { + return a; + } + return std::optional{}; +} + +// If we see `a + b + c` and know that a, b, and c are the same size and have +// two dimensions (WxH), then we can generate a fused kernel for them. That +// fused kernel would likely have indexing math to handling both the W and H +// dimensions. However, if we knew the WxH dimensions were contiguous, we can +// pretend like we only have a single dimension, simplifying the indexing logic. +// This can be performed even if the dimensions are transposed, +// as long as a, b, and c are transposed in the same way. +// We'd like to have the compiler be able to do this dimensionality reduction, +// but simply knowing sizes is not enough. +// We can extend profiling to also record stride information. +// Rather than recording specific strides, +// we can simply order the strides from smallest to largest with +// `stride_indices` A contiguity marker on the smallest stride (c0) indicates +// the stride is precisely 1, otherwise a contiguity marker means that $stride_n +// = size_{n-1}*stride_{n-1}$ +struct TORCH_API Stride { + Stride() = default; + Stride( + const std::optional& stride_index, + std::optional contiguous, + const std::optional& stride) + : stride_index_(stride_index), contiguous_(contiguous), stride_(stride) {} + + bool operator==(const Stride& b) const { + return stride_index_ == b.stride_index_ && contiguous_ == b.contiguous_ && + stride_ == b.stride_; + } + + bool isComplete() const { + return stride_index_ && contiguous_ && stride_; + } + + std::optional stride_index_; + std::optional contiguous_; + std::optional stride_; +}; + +template <> +inline std::optional merge_primitive( + const std::optional& a, + const std::optional& b) { + std::optional left = a; + std::optional right = b; + if (!left.has_value()) { + left = {Stride()}; + } + if (!right.has_value()) { + right = {Stride()}; + } + + auto merged_index = + merge_primitive(left->stride_index_, right->stride_index_); + auto merged_cont = merge_primitive(left->contiguous_, right->contiguous_); + auto merged_stride = merge_primitive(left->stride_, right->stride_); + auto r = Stride(merged_index, merged_cont, merged_stride); + // normalize + if (!r.stride_index_.has_value() && !r.contiguous_.has_value() && + !r.stride_.has_value()) { + return std::optional{}; + } + + return r; +} + +struct TORCH_API ShapeSymbol { + // needed for use in `std::map` + ShapeSymbol() : value_(-1) {} + // is this symbol a fixed/static dimension + bool is_static() const { + return value_ >= 0; + } + bool operator==(const ShapeSymbol& b) const { + return value_ == b.value_; + } + bool operator<(const ShapeSymbol& b) const { + return value_ < b.value_; + } + + static ShapeSymbol fromStaticSize(int64_t val) { + return ShapeSymbol(val); + } + int64_t static_size() const { + TORCH_CHECK(is_static()); + return value_; + } + + int64_t value() const { + return value_; + } + + static ShapeSymbol newSymbol() { + return fromStaticSize(-static_cast(++num_symbols)); + } + friend TORCH_API std::ostream& operator<<( + std::ostream& os, + const ShapeSymbol& s); + + private: + ShapeSymbol(int64_t val) : value_(val) {} + int64_t value_; + static std::atomic num_symbols; +}; + +inline ShapeSymbol merge_primitive( + const ShapeSymbol& a, + const ShapeSymbol& b) { + if (a.is_static() && b.is_static() && a == b) { + return a; + } + return ShapeSymbol::newSymbol(); +} + +// Shape of a Tensor represented with ShapeSymbol's. Unranked, ranked unknown +// dims, partially known and fully known shapes are all supported. +struct TORCH_API SymbolicShape { + // Unranked shape constructor. + SymbolicShape() : dims_(std::nullopt) {} + + // Known rank but unknown dimensions. + SymbolicShape(std::optional rank) : dims_(std::nullopt) { + if(!rank) { + return; + } + + std::vector shape_symbols; + shape_symbols.reserve(*rank); + for(size_t i = 0; i < *rank; ++i) { + shape_symbols.push_back(ShapeSymbol::newSymbol()); + } + dims_ = shape_symbols; + } + + // Mix of known and unknown ranks + SymbolicShape(const std::vector>& dims) { + std::vector shape_symbols; + shape_symbols.reserve(dims.size()); + for(std::optional dim: dims) { + if(!dim) { + shape_symbols.push_back(ShapeSymbol::newSymbol()); + } else { + shape_symbols.push_back(ShapeSymbol::fromStaticSize(*dim)); + } + } + dims_ = shape_symbols; + } + + void dump() const; + + SymbolicShape(std::vector dims) : dims_(std::move(dims)) {} + + SymbolicShape(c10::IntArrayRef dims) { + std::vector shape_symbols; + shape_symbols.reserve(dims.size()); + for(int64_t dim : dims) { + shape_symbols.push_back(ShapeSymbol::fromStaticSize(dim)); + } + dims_ = shape_symbols; + } + + ShapeSymbol operator[](size_t i) const { + TORCH_CHECK(dims_, "Rank isn't fixed"); + return (*dims_).at(i); + } + + ShapeSymbol at(size_t i) const { + TORCH_CHECK(dims_, "Rank isn't fixed"); + return (*dims_).at(i); + } + + // Returns rank or nullopt in case of unranked shape. + std::optional rank() const { + if(!dims_) { + return std::nullopt; + } + return dims_->size(); + } + + std::optional> sizes() const { + return dims_; + } + + std::optional> symbolicDims() const { + if (!dims_) { + return std::nullopt; + } + auto symbolic_dims = std::vector(); + for (const ShapeSymbol& s : *dims_) { + symbolic_dims.push_back(!s.is_static()); + } + return symbolic_dims; + } + + // Checks whether the shape is fully defined/complete, ie. rank and sizes + // of every dimension are known. + bool isComplete() const { + if(!dims_) { + return false; + } + for(auto d : *dims_) { + if(!d.is_static()) { + return false; + } + } + return true; + } + + // Create new SymbolicShape that is result of merging self and another + // SymbolicShape. Only dimensions that are static and equal will be + // preserved. + // If either of two shapes are of unknown rank or they have unmatching rank, + // result will be unranked. + SymbolicShape merge(const SymbolicShape& other) const; + + friend bool operator==(const SymbolicShape& lhs, const SymbolicShape& rhs) { + return lhs.dims_ == rhs.dims_; + } + + friend bool operator!=(const SymbolicShape& lhs, const SymbolicShape& rhs) { + return !(lhs == rhs); + } + + private: + std::optional> dims_; +}; + +namespace detail { +inline bool isComplete(const Stride& s) { + return s.isComplete(); +} + +template +inline bool isComplete(const T& /*t*/) { + return true; +} +} + +template +struct VaryingShape { + using ListOfOptionalElements = std::vector>; + VaryingShape(const std::vector& vec) + : VaryingShape(ListOfOptionalElements(vec.begin(), vec.end())) {} + + VaryingShape(c10::ArrayRef vec) + : VaryingShape(ListOfOptionalElements(vec.begin(), vec.end())) {} + + VaryingShape(std::optional size = std::nullopt) : dims_(std::nullopt) { + if (size) { + dims_ = ListOfOptionalElements(*size); + } + } + + VaryingShape(ListOfOptionalElements dims) : dims_(std::move(dims)) {} + + VaryingShape(size_t size) : VaryingShape(std::optional(size)) {} + + bool operator==(const VaryingShape& other) const { + return dims_ == other.dims_; + } + + const std::optional &operator[](size_t i) const { + TORCH_CHECK(dims_, "Rank isn't fixed"); + return (*dims_).at(i); + } + + std::optional size() const { + if (!dims_) { + return std::nullopt; + } + const auto& dims = dims_.value(); + return dims.size(); + } + + const std::optional& sizes() const { + return dims_; + } + + TORCH_API VaryingShape merge(const VaryingShape& other) const; + + std::optional> concrete_sizes() const { + if (!dims_) { + return std::nullopt; + } + std::vector sizes; + sizes.reserve(dims_.value().size()); + for (auto d : *dims_) { + if (!d) { + return std::nullopt; + } + sizes.push_back(d.value()); + } + return sizes; + } + + bool isComplete() const { + if (!dims_) { + return false; + } + for (auto d : *dims_) { + if (!d || !detail::isComplete(*d)) { + return false; + } + } + return true; + } + + private: + std::optional dims_; +}; + +struct TensorType; +// TODO: investigate making this SingletonOrSharedTypePtr +using TensorTypePtr = std::shared_ptr; +// This type represents a single Tensor with a specific size +struct TORCH_API TensorType : public SharedType { + static TensorTypePtr create(const at::Tensor& t); + + // used by TensorType::create(size_t dim) which in turn used by + // shape_analysis.cpp + static TensorTypePtr create( + std::optional scalar_type, + std::optional device, + const VaryingShape& sizes, + const VaryingShape& strides, + std::optional requires_grad, + std::optional undefined = false, + bool tensor_contiguity = false); + + static TensorTypePtr create( + std::optional scalar_type, + std::optional device, + SymbolicShape sizes, + VaryingShape stride_, + std::optional requires_grad, + std::optional undefined = false); + + static TensorTypePtr create( + std::optional scalar_type, + std::optional device, + std::optional dim, + std::optional requires_grad); + + // overloaded create variadic template argument as it could not distinguish + // initializer list + static TensorTypePtr createContiguous( + at::ScalarType scalar_type, + at::Device device, + at::IntArrayRef sizes); + + static TypePtr fromNumberType(const Type& typ); + static TypePtr fromBoolType(); + + std::optional dim() const { + return sizes().size(); + } + + VaryingShape sizes() const; + + VaryingShape strides() const; + + const VaryingShape& stride_properties() const { + return strides_; + } + + const std::optional& device() const { + return device_; + } + const std::optional& scalarType() const { + return scalar_type_; + } + const std::optional& requiresGrad() const { + return requires_grad_; + } + bool requires_grad() const override { + return requires_grad_ ? *requires_grad_ : true; + } + + bool equals(const Type& rhs) const override; + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override; + + std::string str() const override; + + std::string repr_str() const override { + if (isInferredType()) { + return str() + " (inferred)"; + } else { + return str(); + } + } + + std::optional numel() const { + size_t prod = 1; + const auto& shape = sizes(); + + for (size_t i = 0; i < shape.size(); i++) { + auto const &s = shape[i]; + if (!s.has_value()) { + return std::optional{}; + } + prod *= s.value(); + } + return prod; + } + + TensorTypePtr withRequiresGrad(std::optional s) { + auto copy = clone(); + copy->requires_grad_ = s; + return copy; + } + + TensorTypePtr withScalarType(std::optional st) { + auto copy = clone(); + copy->scalar_type_ = st; + return copy; + } + + TensorTypePtr withDim(std::optional d) { + auto copy = clone(); + // withDim is only used by the legacy executor + // that only cares about the rank, so create dummy symbols)) : + copy->sizes_ = SymbolicShape(d); + copy->strides_ = VaryingShape(d); + return copy; + } + + TensorTypePtr withStrides(VaryingShape sstrides) const { + auto cloned = clone(); + cloned->strides_ = std::move(sstrides); + return cloned; + } + + TensorTypePtr withSizesStrides( + at::IntArrayRef sizes, + at::IntArrayRef strides) const { + auto cloned = clone(); + auto ssizes = SymbolicShape(sizes); + cloned->sizes_ = ssizes; + cloned->strides_ = computeStrideProps(sizes, strides); + return cloned; + } + + TensorTypePtr withSymbolicShapes(SymbolicShape ssizes) const { + auto cloned = clone(); + cloned->sizes_ = std::move(ssizes); + return cloned; + } + + TensorTypePtr withSizes(at::IntArrayRef sizes) const { + return withSizesStrides( + sizes, contiguousStridesOf(sizes)); + } + + TensorTypePtr withDevice(const std::optional device) const { + auto copy = clone(); + copy->device_ = device; + return copy; + } + + TensorTypePtr dimensionedOnly() const { + auto copy = clone(); + copy->sizes_ = SymbolicShape(sizes().size()); + copy->strides_ = VaryingShape(sizes().size()); + return copy; + } + + TensorTypePtr contiguous() const { + auto cloned = clone(); + auto concrete_sizes = sizes().concrete_sizes(); + TORCH_INTERNAL_ASSERT(concrete_sizes.has_value()); + auto strides = computeStrideProps( + *concrete_sizes, + contiguousStridesOf(*concrete_sizes)); + cloned->strides_ = strides; + return cloned; + } + + const SymbolicShape& symbolic_sizes() const; + + TensorTypePtr merge(const TensorType& other, bool merge_sizes = true) const; + + bool matchTensor(const at::Tensor& t); + + // is all information about the type specified except for autograd? + // This replaces the notion of a 'CompleteTensorType' that used to exist + // in the type-hierarchy. Excluding require_grad and undefined allows + // this to match the old behavior. + bool isComplete() const { + return scalar_type_ && device_ && sizes_.isComplete() && strides_.isComplete(); + } + + bool isInferredType() const { + return is_inferred_; + } + + static TensorTypePtr getInferred() { + static auto valueInferred = TensorType::create( + /*scalar_type=*/{}, + /*device=*/{}, + /*sizes=*/SymbolicShape(), + /*stride=*/VaryingShape{}, + /*requires_grad=*/{}, + /*undefined=*/false); + valueInferred->is_inferred_ = true; + return valueInferred; + } + + // this property is used by GuardElimination + // please see `checkInputs` for more details + bool isSummarized() const { + return !(isComplete() && requiresGrad().has_value() && + undefined().has_value()); + } + + TensorTypePtr withUndefined() { + auto r = clone(); + r->undefined_ = true; + return r; + } + + TensorTypePtr withPossiblyUndefined() { + auto r = clone(); + r->undefined_ = std::nullopt; + return r; + } + + std::optional undefined() const { return undefined_; } + + static const TensorTypePtr& get(); + + static const TypeKind Kind = TypeKind::TensorType; + + static std::vector contiguousStridesOf( + at::IntArrayRef in_sizes, + at::MemoryFormat memory_format = MemoryFormat::Contiguous) { + auto contiguous_fn = [](const at::IntArrayRef& sizes, + const std::vector& dim_order) { + std::vector strides(sizes.size()); + if (sizes.empty()) // zero-dim case + return strides; + + strides[dim_order[0]] = 1; + for (size_t i = 1; i < dim_order.size(); i++) { + auto cur_dim = dim_order[i]; + auto pre_dim = dim_order[i - 1]; + strides[cur_dim] = strides[pre_dim] * sizes[pre_dim]; + } + return strides; + }; + + std::vector dim_order(in_sizes.size()); + if (memory_format == MemoryFormat::ChannelsLast) { + dim_order = {1, 3, 2, 0}; + } else if (memory_format == MemoryFormat::ChannelsLast3d) { + dim_order = {1, 4, 3, 2, 0}; + } else { + auto ndims = in_sizes.size(); + for (size_t i = 0; i < ndims; i++) { + dim_order[i] = static_cast(ndims - i - 1); // Reverse + } + } + return contiguous_fn(in_sizes, dim_order); + } + + private: + TensorType( + std::optional scalar_type, + std::optional device, + SymbolicShape sizes, + VaryingShape strides, + std::optional requires_grad, + std::optional undefined = false); + + TensorTypePtr clone() const { + return TensorTypePtr(new TensorType( + scalar_type_, device_, sizes_, strides_, requires_grad_, undefined_)); + } + + static VaryingShape computeStrideProps( + at::IntArrayRef sizes, + at::IntArrayRef strides, + bool tensor_contiguity = false); + + std::optional scalar_type_; + std::optional device_; + SymbolicShape sizes_; + VaryingShape strides_; + std::optional requires_grad_; + // we exploit the fact certain tensors must be zero in the autograd to + // optimize gradient computation. Such zero tensors are currently implemented + // with `UndefinedTensorImpl.` They can be handled only by special operators + // (e.g. `AutogradAdd`) and their `Tensor::defined()` property returns false. + // Normally, `undefined_` is set to false, unless a type was created + // with `withUndefined` + // This will also mean that `undefined` tensors will fail + // `subtypeOf(TensorType::get())` check + // undefined_ may become `std::nullopt` if the tensor was observed to be both + // defined and undefined. However, no tensor type starts out with + // `undefined_` set to `std::nullopt` + std::optional undefined_; + // Represents whether or not this type was inferred. + bool is_inferred_ = false; +}; + +struct ListType; +using ListTypePtr = std::shared_ptr; +struct TORCH_API ListType + : public SingleElementType { + // It's not exactly a singleton, but there should be exactly one instance of + // List[T] for every T + friend struct Type; + template + static ListTypePtr create(T&&... all) { + return ListTypePtr( + new ListType(std::forward(all)...)); // NOLINT(modernize-make-shared) + } + + std::string str() const override { + std::stringstream ss; + ss << getElementType()->str() << "[]"; + return ss.str(); + } + TypePtr createWithContained( + std::vector contained_types) const override { + return create(std::move(contained_types.at(0))); + } + + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override; + + // global singleton + // Given an inner type T and an identifier, + // this function will return the global singleton type pointer + // the type List. + // The extra "identifier" argument is needed because we have multiple container types + // that all reuse this function (List, array, etc.) + static TypePtr get(const std::string& identifier, TypePtr inner); + + // common cast List[Tensor] + static ListTypePtr ofTensors(); + static ListTypePtr ofOptionalTensors(); + static ListTypePtr ofInts(); + static ListTypePtr ofSymInts(); + static ListTypePtr ofFloats(); + static ListTypePtr ofComplexDoubles(); + static ListTypePtr ofBools(); + static ListTypePtr ofStrings(); + static ListTypePtr ofNumbers(); + + private: + ListType(TypePtr elem) : SingleElementType(std::move(elem)) {} + + std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override { + std::stringstream ss; + ss << "List[" << getElementType()->annotation_str(printer) << ']'; + return ss.str(); + } +}; + +struct DictType; +using DictTypePtr = std::shared_ptr; +struct TORCH_API DictType : public SharedType { + friend struct Type; + static const TypeKind Kind = TypeKind::DictType; + + static DictTypePtr create(TypePtr key, TypePtr value) { + auto kind = key->kind(); + if (auto dyn = key->castRaw()) { + kind = dyn->dynamicKind(); + } + C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum") + switch (kind) { + case TypeKind::AnyType: + case TypeKind::IntType: + case TypeKind::BoolType: + case TypeKind::FloatType: + case TypeKind::ComplexType: + case TypeKind::StringType: + case TypeKind::TensorType: + case TypeKind::DeviceObjType: + return DictTypePtr(new DictType(std::move(key), std::move(value))); + default: + TORCH_CHECK(false, + "Cannot create dict for key type '", + key->str(), + "', only int, float, complex, Tensor, device and string keys are supported"); + } + C10_DIAGNOSTIC_POP() + } + + // aligned with the format in FunctionSchema + std::string str() const override { + std::stringstream ss; + ss << "Dict(" << getKeyType()->str() << ", " << getValueType()->str() + << ')'; + return ss.str(); + } + + TypePtr createWithContained( + std::vector contained_types) const override { + TORCH_CHECK(contained_types.size() == 2, "Expected 2 contained types"); + return create(std::move(contained_types.at(0)), std::move(contained_types.at(1))); + } + + const TypePtr& getKeyType() const { + return types.at(0); + } + + const TypePtr& getValueType() const { + return types.at(1); + } + + bool hasFreeVariables() const override { + return has_free_variables; + } + + at::ArrayRef containedTypes() const override { + return types; + } + + bool equals(const Type& rhs) const override { + if (auto* dict_rhs = rhs.castRaw()) { + return *getKeyType() == *(dict_rhs->getKeyType()) && + *getValueType() == *(dict_rhs->getValueType()); + } + return false; + } + + // global singleton + // Given an inner type T and an identifier, + // this function will return the global singleton type pointer + // the type List. + // The extra "identifier" argument is needed because we have multiple container types + // that all reuse this function (Dict and unordered_map) + static TypePtr get(const std::string& identifier, TypePtr key, TypePtr val); + + private: + DictType(TypePtr key, TypePtr value) + : SharedType(TypeKind::DictType), + has_free_variables( + key->hasFreeVariables() || value->hasFreeVariables()) { + types.reserve(2); + types.push_back(std::move(key)); + types.push_back(std::move(value)); + } + + std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override; + + std::vector types; + bool has_free_variables; +}; + +struct FutureType; +using FutureTypePtr = std::shared_ptr; + +struct TORCH_API FutureType + : public SingleElementType { + friend struct Type; + template + static FutureTypePtr create(TypePtr elem) { + return FutureTypePtr( + new FutureType(std::move(elem))); // NOLINT(modernize-make-shared) + } + + std::string str() const override { + std::stringstream ss; + ss << "Future(" << getElementType()->str() << ')'; + return ss.str(); + } + TypePtr createWithContained( + std::vector contained_types) const override { + return create(std::move(contained_types.at(0))); + } + + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override { + if (Type::isSubtypeOfExt(rhs, why_not)) { + return true; + } + if (auto rhs_ = rhs.castRaw()) { + return getElementType()->isSubtypeOfExt(*rhs_->getElementType(), why_not); + } + return false; + } + + private: + FutureType(TypePtr elem) : SingleElementType(std::move(elem)) {} + + std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override { + std::stringstream ss; + ss << "Future[" << getElementType()->annotation_str(printer) << ']'; + return ss.str(); + } +}; + +struct AwaitType; +using AwaitTypePtr = std::shared_ptr; + +struct TORCH_API AwaitType + : public SingleElementType { + friend struct Type; + template + static AwaitTypePtr create(TypePtr elem) { + return AwaitTypePtr( + new AwaitType(std::move(elem))); // NOLINT(modernize-make-shared) + } + + std::string str() const override { + std::stringstream ss; + ss << "Await(" << getElementType()->str() << ')'; + return ss.str(); + } + TypePtr createWithContained( + std::vector contained_types) const override { + return create(std::move(contained_types.at(0))); + } + + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override { + if (Type::isSubtypeOfExt(rhs, why_not)) { + return true; + } + if (auto rhs_ = rhs.castRaw()) { + return getElementType()->isSubtypeOfExt(*rhs_->getElementType(), why_not); + } + return false; + } + + private: + AwaitType(TypePtr elem) : SingleElementType(std::move(elem)) {} + + std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override { + std::stringstream ss; + ss << "Await[" << getElementType()->annotation_str(printer) << ']'; + return ss.str(); + } +}; + +struct RRefType; +using RRefTypePtr = std::shared_ptr; + +struct TORCH_API RRefType + : public SingleElementType { + friend struct Type; + template + static RRefTypePtr create(TypePtr elem) { + return RRefTypePtr( + new RRefType(std::move(elem))); // NOLINT(modernize-make-shared) + } + + std::string str() const override { + std::stringstream ss; + ss << "RRef(" << getElementType()->str() << ')'; + return ss.str(); + } + TypePtr createWithContained( + std::vector contained_types) const override { + return create(std::move(contained_types.at(0))); + } + + private: + RRefType(TypePtr elem) : SingleElementType(std::move(elem)) {} + + std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override { + std::stringstream ss; + ss << "RRef[" << getElementType()->annotation_str(printer) << ']'; + return ss.str(); + } +}; + +// Any should never appear in a named type like a class, namedtuple or +// interface. If it does, then dynamic type information will be lost in the +// Pickler, leading to hard-to-track-down bugs that will only occur +// after saving or loading a model. This is because we rely on the +// static types in named types to reconstruct type tags of loaded +// values. Lifting this restriction requires solving the serialization +// problem first. +TORCH_API void checkNoAny( + const Type& base, + const char* what, + const std::string& attrname, + const TypePtr& attrtype); + +struct TupleType; +using TupleTypePtr = std::shared_ptr; +using NameList = std::vector; +// This type represents a Tuple +struct TORCH_API TupleType : public NamedType { + + static TupleTypePtr createNamed(const std::optional& name, + const std::vector& field_names, + const std::vector& field_types, + std::vector& field_defaults); + + static TupleTypePtr createNamed(const std::optional& name, + const std::vector& field_names, + const std::vector& field_types); + + static TupleTypePtr createNamed(const std::optional& name, + const std::vector& field_names, + const std::vector& field_types); + + static TupleTypePtr create( + std::vector types) { + return TupleTypePtr(new TupleType( + std::move(types), + std::nullopt, + nullptr)); // NOLINT(modernize-make-shared) + } + static TupleTypePtr create() { + return create({}); + } + + at::ArrayRef elements() const { + return elements_; + } + + bool equals(const Type& rhs) const override; + bool isSubtypeOfExt(const Type& rhs_, std::ostream* why_not) const override; + + std::string str() const override; + bool hasFreeVariables() const override { + return has_free_variables_; + } + at::ArrayRef containedTypes() const override { + return elements_; + } + TypePtr createWithContained( + std::vector contained_types) const override { + return std::shared_ptr( + new TupleType(std::move(contained_types), name(), schema())); + } + const std::shared_ptr& schema() const { + return schema_; + } + std::optional> names() const; + + static const TypeKind Kind = TypeKind::TupleType; + + private: + template + static TupleTypePtr createWithSpec( + const std::optional& name, + const std::vector& field_names, + const std::vector& field_types, + std::vector& field_defaults); + + TupleType( + std::vector elements_, + std::optional name, + std::shared_ptr schema); + + bool compare( + const Type& rhs, + const std::function& fn) const { + if (rhs.kind() != kind()) { + return false; + } + + const auto& l_elements = elements(); + const auto& r_elements = rhs.castRaw()->elements(); + if (l_elements.size() != r_elements.size()) + return false; + for (size_t i = 0; i < l_elements.size(); ++i) { + if (!fn(*l_elements[i], *r_elements[i])) + return false; + } + return true; + } + + std::string annotation_str_impl(const TypePrinter& printer = nullptr) const override; + + std::vector elements_; + bool has_free_variables_; + std::shared_ptr schema_; +}; + +// the common supertype of all Enums, only used in operator registration. +// EnumType <: AnyEnumType for all Enums +struct AnyEnumType; +using AnyEnumTypePtr = SingletonTypePtr; +struct TORCH_API AnyEnumType final : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "AnyEnumType"; + } + static const TypeKind Kind = TypeKind::AnyEnumType; + // global singleton + static AnyEnumTypePtr get(); +private: + AnyEnumType() + : Type(TypeKind::AnyEnumType) {} +}; + +struct NumberType; +using NumberTypePtr = SingletonTypePtr; +// This type represents a Python number +// Subtype hierarchy for Number Types (NumberType as the base type): +// IntType <: NumberType +// FloatType <: NumberType +// ComplexType <:NumberType +// +// WARNING: if you add a new subtype of NumberType that is not +// represented by a global singleton, you need to change NumberTypePtr +// to a SingletonOrSharedTypePtr and deal with NumberType needing to +// both inherit and not inherit from SharedType! +struct TORCH_API NumberType : public Type { + bool equals(const Type& rhs) const override; + + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override; + + std::string str() const override { + return "Scalar"; // match what PythonArgParser says for clarity + } + static const TypeKind Kind = TypeKind::NumberType; + // global singleton + static NumberTypePtr get(); + + protected: + NumberType(TypeKind kind = TypeKind::NumberType) : Type(kind) {} + + std::string annotation_str_impl( + [[maybe_unused]] const TypePrinter& printer = nullptr) const override { + return "number"; // technically not a valid python type, but + // we need to use it when parsing back in annotations + // for implicit conversions + } +}; + +struct FloatType; +using FloatTypePtr = SingletonTypePtr; +// This type represents a Python float number +struct TORCH_API FloatType : public NumberType { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "float"; + } + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override { + // NOLINTNEXTLINE(bugprone-parent-virtual-call) + return rhs.kind() == TypeKind::NumberType || Type::isSubtypeOfExt(rhs, why_not); + } + static const TypeKind Kind = TypeKind::FloatType; + // global singleton + static FloatTypePtr get(); + + private: + FloatType() : NumberType(TypeKind::FloatType) {} + std::string annotation_str_impl( + [[maybe_unused]] const TypePrinter& printer = nullptr) const override { + return "float"; + } +}; + +struct ComplexType; +using ComplexTypePtr = SingletonTypePtr; +// This type represents a Python float number +struct TORCH_API ComplexType : public NumberType { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "complex"; + } + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override { + // NOLINTNEXTLINE(bugprone-parent-virtual-call) + return rhs.kind() == TypeKind::NumberType || Type::isSubtypeOfExt(rhs, why_not); + } + static const TypeKind Kind = TypeKind::ComplexType; + // global singleton + static ComplexTypePtr get(); + + private: + ComplexType() : NumberType(TypeKind::ComplexType) {} + std::string annotation_str_impl( + [[maybe_unused]] const TypePrinter& printer = nullptr) const override { + return "complex"; + } +}; + +// We need to introduce `SymIntType` to represent the `SymInt` type +// used in function schemas e.g. `aten::narrow_copy(... SymInt length) +// `SymInt` will be used to enable tracing arithmetic operations on +// dimension values. Please see [SymInt.h] for more information +struct SymIntType; +using SymIntTypePtr = SingletonTypePtr; +struct TORCH_API SymIntType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "SymInt"; + } + std::string annotation_str_impl(const TypePrinter& printer [[maybe_unused]] = nullptr) const override { + return "int"; + } + static const TypeKind Kind = TypeKind::SymIntType; + // global singleton + static SymIntTypePtr get(); + + private: + SymIntType() : Type(TypeKind::SymIntType) {} +}; + +struct SymFloatType; +using SymFloatTypePtr = SingletonTypePtr; +struct TORCH_API SymFloatType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "SymFloat"; + } + std::string annotation_str_impl(const TypePrinter& printer [[maybe_unused]] = nullptr) const override { + return "float"; + } + static const TypeKind Kind = TypeKind::SymFloatType; + // global singleton + static SymFloatTypePtr get(); + + private: + SymFloatType() : Type(TypeKind::SymFloatType) {} +}; + +struct SymBoolType; +using SymBoolTypePtr = SingletonTypePtr; +struct TORCH_API SymBoolType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "SymBool"; + } + std::string annotation_str_impl(const TypePrinter& printer [[maybe_unused]] = nullptr) const override { + return "bool"; + } + static const TypeKind Kind = TypeKind::SymBoolType; + // global singleton + static SymBoolTypePtr get(); + + private: + SymBoolType() : Type(TypeKind::SymBoolType) {} +}; + +struct IntType; +using IntTypePtr = SingletonTypePtr; +// This type represents a Python int number +struct TORCH_API IntType : public NumberType { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "int"; + } + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override { + // NOLINTNEXTLINE(bugprone-parent-virtual-call) + return rhs.kind() == TypeKind::NumberType || Type::isSubtypeOfExt(rhs, why_not); + } + static const TypeKind Kind = TypeKind::IntType; + // global singleton + static IntTypePtr get(); + + private: + IntType() : NumberType(TypeKind::IntType) {} + std::string annotation_str_impl( + [[maybe_unused]] const TypePrinter& printer = nullptr) const override { + return "int"; + } +}; + +struct BoolType; +using BoolTypePtr = SingletonTypePtr; +// This node represents a Python bool value +struct TORCH_API BoolType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "bool"; + } + static const TypeKind Kind = TypeKind::BoolType; + // global singleton + static BoolTypePtr get(); + + private: + BoolType() : Type(TypeKind::BoolType) {} +}; + +struct StringType; +using StringTypePtr = SingletonTypePtr; +// This type represents a Python string +struct TORCH_API StringType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + // we only use "str" (not "string") in both FunctionSchema and script + return annotation_str(); + } + std::string annotation_str_impl( + [[maybe_unused]] const TypePrinter& printer = nullptr) const override { + return "str"; + } + static const TypeKind Kind = TypeKind::StringType; + // global singleton + static StringTypePtr get(); + + private: + StringType() : Type(TypeKind::StringType) {} +}; + +struct StorageType; +using StorageTypePtr = SingletonTypePtr; +struct TORCH_API StorageType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return annotation_str(); + } + std::string annotation_str_impl( + [[maybe_unused]] const TypePrinter& printer = nullptr) const override { + return "Storage"; + } + static const TypeKind Kind = TypeKind::StorageType; + // global singleton + static StorageTypePtr get(); + + private: + StorageType() : Type(TypeKind::StorageType) {} +}; + +struct FunctionType; +using FunctionTypePtr = std::shared_ptr; +struct TORCH_API FunctionType : public NamedType { + static FunctionTypePtr create(torch::jit::Function* function) { + return FunctionTypePtr( + new FunctionType(function)); // NOLINT(modernize-make-shared) + } + bool equals(const Type& rhs) const override { + if (auto func_type = rhs.cast()) { + return func_type->function_ == function_; + } + + return false; + } + std::string str() const override { + return "Function"; + } + torch::jit::Function* function() const { + return function_; + } + static const TypeKind Kind = TypeKind::FunctionType; + + private: + FunctionType(torch::jit::Function* function); + std::string annotation_str_impl( + [[maybe_unused]] const TypePrinter& printer = nullptr) const override { + // NOLINTNEXTLINE(bugprone-unchecked-optional-access) + return name()->qualifiedName(); + } + torch::jit::Function* function_; +}; + +struct NoneType; +using NoneTypePtr = SingletonTypePtr; +// This type represents a Python None +struct TORCH_API NoneType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "NoneType"; + } + bool isSubtypeOfExt(const Type& rhs, std::ostream *why_not) const override; + + static const TypeKind Kind = TypeKind::NoneType; + // global singleton + static NoneTypePtr get(); + + private: + NoneType() : Type(TypeKind::NoneType) {} +}; + +struct GeneratorType; +using GeneratorTypePtr = SingletonTypePtr; +// This type represents a Generator +struct TORCH_API GeneratorType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "Generator"; + } + static const TypeKind Kind = TypeKind::GeneratorType; + // global singleton + static GeneratorTypePtr get(); + + private: + GeneratorType() : Type(TypeKind::GeneratorType) {} +}; + +struct QuantizerType; +using QuantizerTypePtr = SingletonTypePtr; +// This type represents a Quantizer +struct TORCH_API QuantizerType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "Quantizer"; + } + static const TypeKind Kind = TypeKind::QuantizerType; + // global singleton + static QuantizerTypePtr get(); + + private: + QuantizerType() : Type(TypeKind::QuantizerType) {} +}; + +struct QSchemeType; +using QSchemeTypePtr = SingletonTypePtr; +// This type represents a QScheme +struct TORCH_API QSchemeType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "QScheme"; + } + static const TypeKind Kind = TypeKind::QSchemeType; + // global singleton + static QSchemeTypePtr get(); + + private: + QSchemeType() : Type(TypeKind::QSchemeType) {} +}; + +struct DeviceObjType; +using DeviceObjTypePtr = SingletonTypePtr; +// This type represents a Device +struct TORCH_API DeviceObjType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "Device"; + } + static const TypeKind Kind = TypeKind::DeviceObjType; + // global singleton + static DeviceObjTypePtr get(); + + private: + DeviceObjType() : Type(TypeKind::DeviceObjType) {} +}; + +struct StreamObjType; +using StreamObjTypePtr = SingletonTypePtr; +// This type represents a Generator +struct TORCH_API StreamObjType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "Stream"; + } + static const TypeKind Kind = TypeKind::StreamObjType; + // global singleton + static StreamObjTypePtr get(); + +private: + StreamObjType() : Type(TypeKind::StreamObjType) {} +}; + +struct VarType; +using VarTypePtr = std::shared_ptr; +// This type represents a type variable, used in FunctionSchema +struct VarType : public SharedType { + static VarTypePtr create(std::string name_) { + return VarTypePtr(new VarType(std::move(name_))); + } + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return name(); + } + const std::string& name() const { + return name_; + } + bool hasFreeVariables() const override { + return true; + } + static const TypeKind Kind = TypeKind::VarType; + + private: + VarType(std::string name_) + : SharedType(TypeKind::VarType), name_(std::move(name_)) {} + std::string name_; +}; + +struct CapsuleType; +using CapsuleTypePtr = SingletonTypePtr; +// This type represents a Python Capsule. +// It does not appear in the IR and is only used during runtime +struct TORCH_API CapsuleType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "Capsule"; + } + static const TypeKind Kind = TypeKind::CapsuleType; + // global singleton + static CapsuleTypePtr get(); +private: + CapsuleType() + : Type(TypeKind::CapsuleType) {} +}; + +struct PyObjectType; +using PyObjectTypePtr = SingletonTypePtr; +// This type represents a PyObject Type +struct TORCH_API PyObjectType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "PyObject"; + } + static const TypeKind Kind = TypeKind::PyObjectType; + // global singleton + static PyObjectTypePtr get(); +private: + PyObjectType() + : Type(TypeKind::PyObjectType) {} +}; + +enum class TypeVerbosity { + None, + Type, + TypeAndStride, + Full, + Symbolic, + Default = Full, +}; + +TORCH_API TypeVerbosity type_verbosity(); + +TORCH_API std::ostream& operator<<(std::ostream& out, const Type& t); +template +TORCH_API std::ostream& operator<<( + std::ostream& out, + const VaryingShape& t); +TORCH_API std::ostream& operator<<(std::ostream& os, const SymbolicShape& s); +TORCH_API std::ostream& operator<<(std::ostream& os, const ShapeSymbol& s); +TORCH_API std::ostream& operator<<(std::ostream& os, const Stride& s); +// what is the type, ignoring extra size/shape information? +// e.g. Tensor(2x3) -> Dynamic, and Tuple(Tensor(2x3),...) -> Tuple(Dynamic,...) + +// `unshapedType` is used to remove Tensor subtypes. We treat all Tensor +// subtypes as simply "Tensor"; we also create a new version of any +// container types in which internal Tensors have undergone the same +// operation. This is used for type comparisons between two Tensor types +// (`unshapedType` means that we don't falsely return `false` for e.g. +// Tensors of different dimensions). It's also used in the alias +// analysis pass. +// Be careful with calls because this can be very slow. If calling this +// on a graph, use `EraseShapeInformation` in shape_analysis.h +inline TypePtr unshapedType(const TypePtr& type) { + if (type->isSubtypeOf(*TensorType::get())) { + return TensorType::get(); + } + at::ArrayRef contained = type->containedTypes(); + if (contained.empty()) { + return type; + } + return type->withContained(fmap(type->containedTypes(), unshapedType)); +} + +inline TypePtr TensorType::fromNumberType(const Type& typ) { + if (typ.isSubtypeOf(*IntType::get())) { + return TensorType::createContiguous(at::kLong, at::kCPU, {}); + } else if (typ.isSubtypeOf(*FloatType::get())) { + return TensorType::createContiguous(at::kDouble, at::kCPU, {}); + } else if (typ.isSubtypeOf(*BoolType::get())) { + return TensorType::createContiguous(at::kBool, at::kCPU, {}); + } else if (typ.kind() == NumberType::Kind) { + return TensorType::create(std::nullopt, at::kCPU, {}, std::nullopt); + } + TORCH_CHECK(false, "Unknown number type: ", typ.str()); +} +inline TypePtr TensorType::fromBoolType() { + return TensorType::createContiguous(at::kBool, at::kCPU, {}); +} + +inline std::optional tryScalarTypeFromJitType(const Type& type) { + if (type == *FloatType::get()) { + return at::typeMetaToScalarType(c10::get_default_dtype()); + } else if (type == *IntType::get()) { + return at::ScalarType::Long; + } else if (type == *BoolType::get()) { + return at::ScalarType::Bool; + } + return std::nullopt; +} + +inline at::ScalarType scalarTypeFromJitType(const Type& type) { + auto result = tryScalarTypeFromJitType(type); + TORCH_CHECK( + result, + "Add new condition, expected Float, Complex, Int, or Bool but got", + type.str()); + return *result; +} + +// Attempt to find the correct supertype of the two types `t1` and `t2`. +// If no supertype is found, then nullopt will be returned if +// `default_to_union` is false, and `Union[t1, t2]` will be returned +// if it is true. If `t1 == t2`, or `t1` is a type refinement of `t2`, +// then `t2` will be returned (and vice versa). +// +// Two different tensortypes will return dynamic. +// +// Currently we chose not to support returning a NumberType for +// two types from the set of {FloatType, IntType, ComplexType}, because +// there is a lack of operator support for NumberType. +// +// If `type_hint` is an `InterfaceType`, then we can use that as a +// potential supertype for `ClassType`s in the list. Otherwise, we have +// no way to find and use some common interface type +TORCH_API std::optional unifyTypes( + const TypePtr& t1, + const TypePtr& t2, + bool default_to_union = false, + const TypePtr& type_hint = nullptr); + +TORCH_API std::optional unifyTypeList( + at::ArrayRef elements, + std::ostream& why_not, + bool default_to_union = false, + const TypePtr& type_hint = nullptr); + +namespace detail { +template +struct getTypePtr_ final { + static decltype(auto) call() { + return ([]() { + try { + return getCustomClassType(); + } catch(const c10::Error&) { + TORCH_CHECK( + false, + "Type ", + c10::util::get_fully_qualified_type_name(), + " could not be converted to any of the known types." + ); + } + }()); + } +}; + +template +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return getTypePtr_::call(); + } +}; + +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return AnyType::get(); + } +}; + +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return TensorType::get(); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return StorageType::get(); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return StreamObjType::get(); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return FloatType::get(); + } +}; +template <> +struct getTypePtr_> final { + static decltype(auto) call() { + return ComplexType::get(); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return IntType::get(); + } +}; + +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return IntType::get(); + } +}; + +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return SymIntType::get(); + } +}; +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return IntType::get(); + } +}; + +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return SymFloatType::get(); + } +}; +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return FloatType::get(); + } +}; + +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return SymBoolType::get(); + } +}; +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return BoolType::get(); + } +}; + +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return DeviceObjType::get(); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return BoolType::get(); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return NumberType::get(); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return QSchemeType::get(); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return TypeFactory::create( + TypeFactory::get()); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return StringType::get(); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return StringType::get(); + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return StringType::get(); + } +}; +template +struct getMaybeFakeTypePtr_, fake> final { + static const auto& call() { + static auto inner_type = getMaybeFakeTypePtr_::call(); + // The "per vector" static singleton needs to live in a .cpp file, + // otherwise we'll end up with one singleton instance per shared library. + static auto type = ListType::get("vector", inner_type); + return type; + } +}; +template +struct getMaybeFakeTypePtr_, fake> final { + static const auto& call() { + static auto inner_type = getMaybeFakeTypePtr_::call(); + // The "per ArrayRef" static singleton needs to live in a .cpp file, + // otherwise we'll end up with one singleton instance per shared library. + static auto type = ListType::get("ArrayRef", inner_type); + return type; + } +}; +template +struct getMaybeFakeTypePtr_ final { + static const auto& call() { + static auto type = ListType::create(getMaybeFakeTypePtr_::call()); + return type; + } +}; +template +struct getMaybeFakeTypePtr_, fake> final { + static const auto& call() { + static auto inner_type = getMaybeFakeTypePtr_::call(); + // The "per List" static singleton needs to live in a .cpp file, + // otherwise we'll end up with one singleton instance per shared library. + static auto type = ListType::get("List", inner_type); + return type; + } +}; +template +struct getMaybeFakeTypePtr_, fake> final { + static const auto& call() { + static auto inner_type = getMaybeFakeTypePtr_::call(); + static auto type = ListType::get("List", inner_type); + return type; + } +}; +template +struct getMaybeFakeTypePtr_, fake> final { + static const auto& call() { + static auto inner_type = getMaybeFakeTypePtr_::call(); + // The "per array" static singleton needs to live in a .cpp file, + // otherwise we'll end up with one singleton instance per shared library. + // (Concatenating the length onto the end of the string because we want a unique + // type_ptr created for every std::array type). + static auto type = ListType::get(std::string("array") + std::to_string(N), inner_type); + return type; + } +}; +template +struct getMaybeFakeTypePtr_, fake> final { + static const auto& call() { + static auto inner_key_type = getMaybeFakeTypePtr_::call(); + static auto inner_val_type = getMaybeFakeTypePtr_::call(); + // The "per unordered_map" static singleton needs to live in a .cpp file, + // otherwise we'll end up with one singleton instance per shared library. + static auto type = DictType::get("unordered_map", inner_key_type, inner_val_type); + return type; + } +}; +template +struct getMaybeFakeTypePtr_, fake> final { + static const auto& call() { + static auto inner_key_type = getMaybeFakeTypePtr_::call(); + static auto inner_val_type = getMaybeFakeTypePtr_::call(); + // The "per Dict" static singleton needs to live in a .cpp file, + // otherwise we'll end up with one singleton instance per shared library. + static auto type = DictType::get("Dict", inner_key_type, inner_val_type); + return type; + } +}; + +template +struct getMaybeFakeTypePtr_, fake> final { + static const auto& call() { + static auto inner_type = getMaybeFakeTypePtr_::call(); + // The "per std::optional" static singleton needs to live in a .cpp file, + // otherwise we'll end up with one singleton instance per shared library. + static auto type = OptionalType::get(inner_type); + return type; + } +}; + + +template<> +struct getTypePtr_ final { + static const auto& call() { + static auto inner_type = getMaybeFakeTypePtr_::call(); + // The "per std::optional" static singleton needs to live in a .cpp file, + // otherwise we'll end up with one singleton instance per shared library. + static auto type = OptionalType::get(inner_type); + return type; + } +}; + +template +struct getMaybeFakeTypePtr_ final { + static const auto& call() { + // The "per std::optional" static singleton needs to live in a .cpp file, + // otherwise we'll end up with one singleton instance per shared library. + static auto inner_type = getMaybeFakeTypePtr_::call(); + static auto type = OptionalType::get(inner_type); + return type; + } +}; + +template +struct getMaybeFakeTypePtr_, fake> final { + static const auto& call() { + static auto type = ([]() { + std::vector contained_types = { + (getMaybeFakeTypePtr_::call())... + }; + return TupleType::create(std::move(contained_types)); + })(); + return type; + } +}; +template <> +struct getTypePtr_ final { + static decltype(auto) call() { + return NoneType::get(); + } +}; +} // namespace detail +template +inline decltype(auto) getTypePtr() { + // TODO: static_assert that a templated function exists, and throw a friendly + // error message if not + return detail::getMaybeFakeTypePtr_::call(); +} + +template +inline TypePtr getTypePtrCopy() { + // TODO: static_assert that a templated function exists, and throw a friendly + // error message if not + return getTypePtr(); +} + +template +inline decltype(auto) getFakeTypePtr() { + return detail::getMaybeFakeTypePtr_::call(); +} + +template +inline TypePtr getFakeTypePtrCopy() { + return getFakeTypePtr(); +} + +using TypeEnv = std::unordered_map; +struct MatchTypeReturn { + MatchTypeReturn(std::string reason) : reason_(std::move(reason)) {} + static MatchTypeReturn Success() { + return MatchTypeReturn(); + } + bool success() const { + return !reason_.has_value(); + } + const std::string& reason() const { + // NOLINTNEXTLINE(bugprone-unchecked-optional-access) + return reason_.value(); + } + + private: + MatchTypeReturn() + : reason_(std::nullopt) {} + std::optional reason_; // is there is no match, this contains the reason +}; + +// attempt to match the type variables in formal to actual, adding them to type_env. +// If no match is possible this returns a MatchTypeReturn with r.success() == false +// and a r.reason() that describes why it could not match. +// note: It is possible to successfully match a formal, but for type variables +// in the formal to still not be defined. In particular, None matches Optional[T] +// but does not define the value of T. +TORCH_API MatchTypeReturn +matchTypeVariables(const TypePtr& formal, const TypePtr& actual, TypeEnv& type_env); + +// replace type variables appearing in `type` with the values in +// `type_env`. Returns nullptr if a variable used in `type` +// does not appear in `type_env` +TORCH_API TypePtr tryEvalTypeVariables(const TypePtr& type, TypeEnv& type_env); + +TORCH_API bool elementTypeCanBeInferredFromMembers(const TypePtr& elem_type); + +struct InterfaceType; +using InterfaceTypePtr = std::shared_ptr; + +// Interfaces are a list of abstract methods that a class might meet. +// If a class provides those methods, it implicitly meets the interface. + +// Subtype relations for Interface with ClassType: +// lhs (ClassType or InterfaceType) is a subtype of rhs if: +// 1. lhs methods are a superset of rhs methods +// 2. if rhs is module interface, the lhs must be module interface or module itself +struct TORCH_API InterfaceType : public NamedType { + static InterfaceTypePtr create( + QualifiedName qualifiedName, bool is_module=false); + + bool equals(const Type& rhs) const override { + if (auto user_rhs = rhs.castRaw()) { + return isSubTypeImpl(*this, *user_rhs, nullptr) && + isSubTypeImpl(*user_rhs, *this, nullptr); + } + return false; + } + + std::string str() const override { + // NOLINTNEXTLINE(bugprone-unchecked-optional-access) + return std::string("InterfaceType<") + name()->name() + ">"; + } + + bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const override; + + // try to find a method of this interface, + // returns nullptr if not found. + const FunctionSchema* getMethod(const std::string& name) const; + void addMethod(FunctionSchema schema); + const std::vector& methods() const { + return *methods_; + } + + bool is_module() const override{ + return is_module_; + } + static const TypeKind Kind = TypeKind::InterfaceType; + ~InterfaceType() override = default; + private: + InterfaceType(QualifiedName name, bool is_module); + static bool isSubTypeImpl( + const InterfaceType& lhs, + const InterfaceType& rhs, + std::ostream* why_not); + + std::string annotation_str_impl( + [[maybe_unused]] const TypePrinter& printer = nullptr) const override { + // NOLINTNEXTLINE(bugprone-unchecked-optional-access) + return name()->qualifiedName(); + } + + // shared_ptr so that this header does not have to depend on + // FunctionSchema.h + std::shared_ptr> methods_; + // flag to distinguish if it's an interface type from a module or not + bool is_module_; +}; + +template +struct EnumerationType : public Type { +static const TypeKind Kind = K; + +bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); +} + +protected: +EnumerationType() : Type(Kind) {} +}; + +// WARNING: These enumeration types below DO NOT actually get parsed out +// from the logical schema strings, instead they are mapped as ints. To +// observe these types, use real_type() instead of type() on Argument + +struct ScalarTypeType; +using ScalarTypeTypePtr = SingletonTypePtr; +struct TORCH_API ScalarTypeType : public EnumerationType { +std::string str() const override { +return "ScalarType"; +} +static const TypeKind Kind = TypeKind::ScalarTypeType; +// global singleton +static ScalarTypeTypePtr get(); + +private: +ScalarTypeType() {} +}; + +struct MemoryFormatType; +using MemoryFormatTypePtr = SingletonTypePtr; +struct TORCH_API MemoryFormatType : public EnumerationType { +std::string str() const override { +return "MemoryFormat"; +} +static const TypeKind Kind = TypeKind::MemoryFormatType; +// global singleton +static MemoryFormatTypePtr get(); + +private: +MemoryFormatType() {} +}; + +struct LayoutType; +using LayoutTypePtr = SingletonTypePtr; +struct TORCH_API LayoutType : public EnumerationType { +std::string str() const override { +return "Layout"; +} +static const TypeKind Kind = TypeKind::LayoutType; +// global singleton +static LayoutTypePtr get(); + +private: +LayoutType() {} +}; + +namespace detail { +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return ScalarTypeType::get(); + } +}; +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return LayoutType::get(); + } +}; +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return MemoryFormatType::get(); + } +}; +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return IntType::get(); + } +}; +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return IntType::get(); + } +}; +template <> +struct getMaybeFakeTypePtr_ final { + static decltype(auto) call() { + return IntType::get(); + } +}; +} // namespace detail + +// the common supertype of all lists, +// List[T] <: AnyList for all T +struct AnyListType; +using AnyListTypePtr = SingletonTypePtr; +struct TORCH_API AnyListType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "list"; + } + static const TypeKind Kind = TypeKind::AnyListType; + // global singleton + static AnyListTypePtr get(); +private: + AnyListType() + : Type(TypeKind::AnyListType) {} +}; + +// the common supertype of all tuples, +// Tuple[T...] <: AnyTuple for all T +struct AnyTupleType; +using AnyTupleTypePtr = SingletonTypePtr; +struct TORCH_API AnyTupleType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + + std::string str() const override { + return "tuple"; + } + static const TypeKind Kind = TypeKind::AnyTupleType; + + // global singleton + static AnyTupleTypePtr get(); +private: + AnyTupleType() + : Type(TypeKind::AnyTupleType) {} +}; + +// the common supertype of all classes, +// ClassType <: AnyClassType for all classes +struct AnyClassType; +using AnyClassTypePtr = SingletonTypePtr; +struct TORCH_API AnyClassType : public Type { + bool equals(const Type& rhs) const override { + return rhs.kind() == kind(); + } + std::string str() const override { + return "AnyClassType"; + } + static const TypeKind Kind = TypeKind::AnyClassType; + // global singleton + static AnyClassTypePtr get(); +private: + AnyClassType() + : Type(TypeKind::AnyClassType) {} +}; + +template<> +inline detail::CastReturnType::type Type::cast() { + if (kind() == TypeKind::TupleType || kind() == TypeKind::FunctionType || + kind() == TypeKind::ClassType || kind() == TypeKind::InterfaceType) { + return std::static_pointer_cast(static_cast(this)->shared_from_this()); + } + return nullptr; +} + +template<> +inline detail::CastConstReturnType::type Type::cast() const { + if (kind() == TypeKind::TupleType || kind() == TypeKind::FunctionType || + kind() == TypeKind::ClassType || kind() == TypeKind::InterfaceType) { + return std::static_pointer_cast(static_cast(this)->shared_from_this()); + } + return nullptr; +} + +template<> +inline const NamedType* Type::castRaw() const { + if (kind() == TypeKind::TupleType || kind() == TypeKind::FunctionType || + kind() == TypeKind::ClassType || kind() == TypeKind::InterfaceType) { + return static_cast(this); + } + return nullptr; +} + +// Used as a return type when inferring the IValue type of a Python object. +struct InferredType { + /* implicit */ InferredType(TypePtr type) : type_(std::move(type)) {} + /* implicit */ InferredType(std::string reason) + : type_(nullptr), reason_(std::move(reason)) {} + TypePtr type() const { + TORCH_INTERNAL_ASSERT( + type_, + "Tried to get the type from an InferredType but the type is null. ", + "Reason: ", + reason_); + return type_; + } + bool success() const { + return type_ != nullptr; + } + const std::string& reason() const { + TORCH_INTERNAL_ASSERT(!type_); + return reason_; + } + +private: + TypePtr type_; + std::string reason_; +}; + +TORCH_API bool containsAnyType(const TypePtr& type); + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/jit_type_base.h b/backends/pytorch/include/ATen/core/jit_type_base.h new file mode 100644 index 000000000..c7ffbb337 --- /dev/null +++ b/backends/pytorch/include/ATen/core/jit_type_base.h @@ -0,0 +1,602 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +#define C10_FORALL_TYPES(_) \ + _(AnyType) \ + _(EnumType) \ + _(AnyEnumType) \ + _(TensorType) \ + _(StorageType) \ + _(TupleType) \ + _(ListType) \ + _(DictType) \ + _(NumberType) \ + _(FloatType) \ + _(ComplexType) \ + _(FutureType) \ + _(AwaitType) \ + _(RRefType) \ + _(IntType) \ + _(NoneType) \ + _(StringType) \ + _(GeneratorType) \ + _(QuantizerType) \ + _(BoolType) \ + _(OptionalType) \ + _(VarType) \ + _(DeviceObjType) \ + _(StreamObjType) \ + _(FunctionType) \ + _(ClassType) \ + _(PyObjectType) \ + _(CapsuleType) \ + _(InterfaceType) \ + _(QSchemeType) \ + _(ScalarTypeType) \ + _(LayoutType) \ + _(MemoryFormatType) \ + _(AnyListType) \ + _(AnyTupleType) \ + _(AnyClassType) \ + _(SymIntType) \ + _(SymFloatType) \ + _(SymBoolType) \ + _(UnionType) \ + _(DynamicType) + +enum class TypeKind { +#define DEFINE_TYPE(T) T, + C10_FORALL_TYPES(DEFINE_TYPE) +#undef DEFINE_TYPE +}; + +TORCH_API const char* typeKindToString(TypeKind kind); + +struct Type; +struct SharedType; + +// Use this to customize how a Type is printed using `annotation_str()`. If +// std::nullopt is returned, `annotation_str()` falls through to its default +// implementation. +using TypePrinter = std::function(const Type&)>; + +namespace detail { +template +struct IsSingletonType : public std::integral_constant {}; +} // namespace detail +#define TORCH_DECLARE_SINGLETON(Type) \ + struct Type; \ + namespace detail { \ + template <> struct IsSingletonType : public std::integral_constant {}; \ + } + +TORCH_DECLARE_SINGLETON(AnyType) +TORCH_DECLARE_SINGLETON(AnyEnumType) +TORCH_DECLARE_SINGLETON(NumberType) +TORCH_DECLARE_SINGLETON(FloatType) +TORCH_DECLARE_SINGLETON(ComplexType) +TORCH_DECLARE_SINGLETON(IntType) +TORCH_DECLARE_SINGLETON(BoolType) +TORCH_DECLARE_SINGLETON(StringType) +TORCH_DECLARE_SINGLETON(StorageType) +TORCH_DECLARE_SINGLETON(NoneType) +TORCH_DECLARE_SINGLETON(GeneratorType) +TORCH_DECLARE_SINGLETON(QuantizerType) +TORCH_DECLARE_SINGLETON(QSchemeType) +TORCH_DECLARE_SINGLETON(DeviceObjType) +TORCH_DECLARE_SINGLETON(StreamObjType) +TORCH_DECLARE_SINGLETON(CapsuleType) +TORCH_DECLARE_SINGLETON(PyObjectType) +TORCH_DECLARE_SINGLETON(ScalarTypeType) +TORCH_DECLARE_SINGLETON(LayoutType) +TORCH_DECLARE_SINGLETON(MemoryFormatType) +TORCH_DECLARE_SINGLETON(AnyListType) +TORCH_DECLARE_SINGLETON(AnyTupleType) +TORCH_DECLARE_SINGLETON(AnyClassType) + +namespace detail { +template +struct CastReturnType { + using type = std::shared_ptr; +}; + +template +struct CastReturnType::value>> { + using type = SingletonTypePtr; +}; + +template +struct CastConstReturnType { + using type = std::shared_ptr; +}; + +template +struct CastConstReturnType::value>> { + using type = SingletonTypePtr; +}; + +template +struct as_shared_type { + using type = SharedType*; +}; + +template +struct as_shared_type { + using type = const SharedType *; +}; +} // namespace detail + +struct TORCH_API Type { + friend TORCH_API bool operator==(const Type& lhs, const Type& rhs); + private: + TypeKind kind_; + + protected: + Type(TypeKind kind) : kind_(kind) {} + + Type(const Type&) = default; + Type& operator=(const Type&) = default; + Type(Type&&) noexcept = default; + Type& operator=(Type&&) noexcept = default; + + virtual std::string annotation_str_impl(const TypePrinter& /*printer*/) const { + return str(); + } + // a == b + virtual bool equals(const Type& rhs) const = 0; + // a == b <=> b == a + virtual bool symmetric() const { + return true; + } + + public: + template + class SingletonOrSharedTypePtr { + public: + using element_type = typename std::shared_ptr::element_type; + + SingletonOrSharedTypePtr() = default; + + /* implicit */ SingletonOrSharedTypePtr(std::shared_ptr x) + : repr_(std::move(x)) {} + + template , bool> = true> + /* implicit */ SingletonOrSharedTypePtr(std::shared_ptr x) + : repr_(std::move(x)) {} + + /* implicit */ SingletonOrSharedTypePtr(std::nullptr_t) + : repr_(nullptr) {} + + /* implicit */ SingletonOrSharedTypePtr(SingletonTypePtr p) + : repr_(makeSingletonSharedPtr(p.get())) {} + + template , bool> = true> + /* implicit */ SingletonOrSharedTypePtr(SingletonTypePtr p) + : repr_(makeSingletonSharedPtr(static_cast(p.get()))) {} + + + // We need to support construction from T* for pybind. The problem + // is that it's not clear if we are supposed to be taking shared + // ownership or not. + // + // Case 1: if T is known statically to derive from SharedType, we should use + // shared_from_this() and take shared_ownership. + // + // Case 2: if T is exactly Type, we need to do a dynamic_cast to + // check if it's a SharedType and do the right thing. + // + // Case 3: Otherwise, T is not a SharedType. Use a singleton + // pointer. + + template , bool> = true> + /* implicit */ SingletonOrSharedTypePtr(T* p) : SingletonOrSharedTypePtr(static_cast::type>(p)->shared_from_this()) {} + + template , bool> = true> + /* implicit */ SingletonOrSharedTypePtr(T* p) { + if (auto* shared_p = dynamic_cast::type>(p)) { + repr_ = shared_p->shared_from_this(); + } else { + repr_ = makeSingletonSharedPtr(p); + } + } + + template && !std::is_base_of_v, bool> = true> + /* implicit */ SingletonOrSharedTypePtr(T* p) + : repr_(makeSingletonSharedPtr(p)) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(dynamic_cast::type>(p) == nullptr); + } + + SingletonOrSharedTypePtr(const SingletonOrSharedTypePtr&) = default; + SingletonOrSharedTypePtr(SingletonOrSharedTypePtr&&) noexcept = default; + SingletonOrSharedTypePtr& operator=(const SingletonOrSharedTypePtr&) = default; + SingletonOrSharedTypePtr& operator=(SingletonOrSharedTypePtr&&) noexcept = default; + ~SingletonOrSharedTypePtr() = default; + + T* get() const { + return repr_.get(); + } + + operator bool() const { + return repr_ != nullptr; + } + + bool operator==(std::nullptr_t) const { + return repr_ == nullptr; + } + + bool operator!=(std::nullptr_t) const { + return repr_ != nullptr; + } + + template , void>, bool> = true> + U& operator*() const { + return *get(); + } + + T* operator->() const { + return get(); + } + + private: + // Use shared_ptr's aliasing constructor to create a non-owning pointer + // to a singleton. The lifetime is tied to the null shared_ptr, so there's + // no reference counting overhead for the singleton itself. + static std::shared_ptr makeSingletonSharedPtr(T* ptr) { + return std::shared_ptr(std::shared_ptr(), ptr); + } + + std::shared_ptr repr_; + }; + + using TypePtr = SingletonOrSharedTypePtr; + using Ptr = TypePtr; + using ElementType = Type; + + // subtyping relation. By default, we return true for the case + // when the type is exactly equal or if this <: T where rhs = Optional[T] + + // if this returns false and the why_not stream is non-null, it contains + // additional details that describe why this is not a subtype of 'rhs'. + // This additional information should only contain details that are not + // obvious from the annotation_str() that describes the type. For instance it + // is clear that `int <: str` is false but not clear why `Foo <: InterfaceBar` + // might be false. + virtual bool isSubtypeOfExt(const Type& rhs, std::ostream* why_not) const; + virtual bool is_module() const; + bool isSubtypeOf(const Type& rhs) const { + return isSubtypeOfExt(rhs, nullptr); + } + // Compatibility shims to accommodate existing code that passes shared_ptrs + // around. Ideally, we would just delete this, but it should be harmless. + template + std::enable_if_t, bool> + isSubtypeOf(const std::shared_ptr& rhs) const { + return isSubtypeOf(*rhs); + } + + template + std::enable_if_t, bool> + isSubtypeOf(const SingletonOrSharedTypePtr& rhs) const { + return isSubtypeOf(*rhs); + } + + template + std::enable_if_t, bool> + isSubtypeOf(SingletonTypePtr rhs) const { + return isSubtypeOf(*rhs); + } + + template + std::enable_if_t, bool> + isSubtypeOfExt(const SingletonOrSharedTypePtr& rhs, std::ostream* why_not) const { + return isSubtypeOfExt(*rhs, why_not); + } + + template + std::enable_if_t, bool> + isSubtypeOfExt(const std::shared_ptr& rhs, std::ostream* why_not) const { + return isSubtypeOfExt(*rhs, why_not); + } + + template + std::enable_if_t, bool> + isSubtypeOfExt(SingletonTypePtr rhs, std::ostream* why_not) const { + return isSubtypeOfExt(*rhs, why_not); + } + + // How this type will appear in FunctionSchema declarations + virtual std::string str() const = 0; + + // How this type will appear as if it were a type annotation in Python + // which is sometimes different than how it appears in declarations (e.g. + // int[] vs List[int]) + // + // Takes a custom printer that users can pass in to customize the output of + // this method. + std::string annotation_str(const TypePrinter& printer) const { + if (printer) { + // the printer can return std::nullopt to fall through to the default impl + if (auto renamed = printer(*this)) { + return *renamed; + } + } + return annotation_str_impl(printer); + } + std::string annotation_str() const { + // Overload instead of define a default value for `printer` to help + // debuggers out. + return annotation_str(nullptr); + } + + // Returns a human readable string that includes additional information like + // "type is inferred rather than explicitly defined" to help construct more + // user-friendly messages. + virtual std::string repr_str() const { + return annotation_str(); + } + + TypeKind kind() const { + return kind_; + } + + virtual bool isUnionType() const { + return false; + } + + virtual bool requires_grad() const { + for (const auto& ct : containedTypes()) { + if (ct->requires_grad()) { + return true; + } + } + return false; + } + + // Dynamically cast this object to the subclass indicated by the + // template variable, returning nullptr if the cast is invalid. + template ::value, bool> = true> + typename detail::CastReturnType::type cast() { + if (T::Kind == kind()) { + return std::static_pointer_cast(static_cast(this)->shared_from_this()); + } + return nullptr; + } + template ::value, bool> = true> + typename detail::CastReturnType::type cast() { + if (T::Kind == kind()) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(this == T::get().get()); + return typename detail::CastReturnType::type(static_cast(this)); + } + return nullptr; + } + template ::value, bool> = true> + typename detail::CastConstReturnType::type cast() const { + if (T::Kind == kind()) { + return std::static_pointer_cast(static_cast(this)->shared_from_this()); + } + return nullptr; + } + template ::value, bool> = true> + typename detail::CastConstReturnType::type cast() const { + if (T::Kind == kind()) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(this == T::get().get()); + return typename detail::CastConstReturnType::type(static_cast(this)); + } + return nullptr; + } + template + T* castRaw() { + if (T::Kind == kind()) { + return static_cast(this); + } + return nullptr; + } + template + const T* castRaw() const { + if (T::Kind == kind()) { + return static_cast(this); + } + return nullptr; + } + template + auto expect() { + auto r = cast(); + AT_ASSERT(r); + return r; + } + template + auto expect() const { + auto r = cast(); + AT_ASSERT(r); + return r; + } + template + T& expectRef() { + auto* r = castRaw(); + AT_ASSERT(r); + return *r; + } + template + const T& expectRef() const { + auto* r = castRaw(); + AT_ASSERT(r); + return *r; + } + virtual ~Type() = default; + virtual bool hasFreeVariables() const { + return false; + } + // list of types this type contains, e.g. for a List then element type of a + // list for a tuple, the types of the tuple elements + virtual at::ArrayRef containedTypes() const { + return {}; + } + virtual TypePtr containedType(size_t i) const { + return containedTypes().at(i); + } + virtual size_t containedTypeSize() const { + return containedTypes().size(); + } + // create a new version of this type, replacing its contained types with + // contained_types + TypePtr withContained(std::vector contained_types); + // per-type constructor, you only need to override this if the + // containedTypes() is not empty + virtual TypePtr createWithContained( + // NOLINTNEXTLINE(performance-unnecessary-value-param) + std::vector /*contained_types*/) const { + TORCH_CHECK(false, + "type with contained types did not overload createWithContained: ", + str()); + } + +}; + +template +using SingletonOrSharedTypePtr = Type::SingletonOrSharedTypePtr; + + +template +bool operator==(const SingletonOrSharedTypePtr& x, const SingletonOrSharedTypePtr& y) { + return (void*)x.get() == (void*)y.get(); +} + +template +bool operator==(const SingletonOrSharedTypePtr& x, const std::shared_ptr& y) { + return (void*)x.get() == (void*)y.get(); +} + +template +bool operator==(const std::shared_ptr& x, const SingletonOrSharedTypePtr& y) { + return (void*)x.get() == (void*)y.get(); +} + +template +bool operator==(const SingletonOrSharedTypePtr& x, const SingletonTypePtr& y) { + return (void*)x.get() == (void*)y.get(); +} + +template +bool operator==(const SingletonTypePtr& x, const SingletonOrSharedTypePtr& y) { + return (void*)x.get() == (void*)y.get(); +} + +template +bool operator!=(const SingletonOrSharedTypePtr& x, const SingletonOrSharedTypePtr& y) { + return !(x == y); +} + +template +bool operator!=(const SingletonOrSharedTypePtr& x, const std::shared_ptr& y) { + return !(x == y); +} + +template +bool operator!=(const std::shared_ptr& x, const SingletonOrSharedTypePtr& y) { + return !(x == y); +} + +template +bool operator!=(const SingletonOrSharedTypePtr& x, const SingletonTypePtr& y) { + return !(x == y); +} + +template +bool operator!=(const SingletonTypePtr& x, const SingletonOrSharedTypePtr& y) { + return !(x == y); +} + +using TypePtr = SingletonOrSharedTypePtr; +using ConstTypePtr = SingletonOrSharedTypePtr; + +// Explicitly enable MaybeOwned>, rather than allowing +// MaybeOwned to be used for any type right away. +template +struct MaybeOwnedTraits> + : public MaybeOwnedTraitsGenericImpl> {}; + +// Base class for Types that are guaranteed to be owned by std::shared_ptr. +struct TORCH_API SharedType : public Type, public std::enable_shared_from_this { + using Type::Type; +}; + +inline TypePtr Type::withContained(std::vector contained_types) { + auto current_contained = containedTypes(); + // Types with no contained_types don't need this call. Check before calling! + // + // (We can't support this efficiently because types without + // contained types may be singletons, in which case + // shared_from_this will crash; we would have to provide a virtual + // typeptr_from_this or isSingleton.) + TORCH_INTERNAL_ASSERT(!current_contained.empty() && current_contained.size() == contained_types.size()); + if (current_contained.equals(contained_types)) { + return std::static_pointer_cast(static_cast(this)->shared_from_this()); + } + return createWithContained(std::move(contained_types)); +} + + +inline bool operator==(const Type& lhs, const Type& rhs) { + if (C10_UNLIKELY(!rhs.symmetric())) { + return rhs.equals(lhs); + } + return lhs.equals(rhs); +} + +struct NamedType; +using NamedTypePtr = std::shared_ptr; +using ConstNamedTypePtr = std::shared_ptr; + +struct TORCH_API NamedType : public SharedType { + NamedType(TypeKind tk, std::optional name) + : SharedType(tk), name_(std::move(name)) { + TORCH_INTERNAL_ASSERT( + tk == TypeKind::TupleType || tk == TypeKind::FunctionType || + tk == TypeKind::ClassType || tk == TypeKind::InterfaceType || + tk == TypeKind::EnumType, + "If you add a new kind of NamedType, ", + "please update the cast specialization and this assert"); + } + + // Fully qualified name of type + // Looks like: "foo.bar.Baz". + const std::optional& name() const { + return name_; + } + + private: + std::optional name_; +}; + +} // namespace c10 + +namespace std { +template +struct hash> { + size_t operator()(const c10::SingletonOrSharedTypePtr& x) const { + return std::hash()(x.get()); + } +}; +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/operator_name.h b/backends/pytorch/include/ATen/core/operator_name.h new file mode 100644 index 000000000..3dbd04bde --- /dev/null +++ b/backends/pytorch/include/ATen/core/operator_name.h @@ -0,0 +1,103 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include +#include +#include +#include +#include +#include + +namespace c10 { + +// TODO: consider storing namespace separately too +struct OperatorName final { + std::string name; + std::string overload_name; + OperatorName(std::string name, std::string overload_name) + : name(std::move(name)), overload_name(std::move(overload_name)) {} + + // TODO: These two functions below are slow! Fix internal data structures so + // I don't have to manually reconstruct the namespaces! + + // Return the namespace of this OperatorName, if it exists. The + // returned string_view is only live as long as the OperatorName + // exists and name is not mutated + std::optional getNamespace() const { + auto pos = name.find("::"); + if (pos == std::string::npos) { + return std::nullopt; + } else { + return std::string_view(name.data(), pos); + } + } + + // Returns true if we successfully set the namespace + bool setNamespaceIfNotSet(const char* ns) { + if (!getNamespace().has_value()) { + const auto ns_len = strlen(ns); + const auto old_name_size = name.size(); + name.resize(ns_len + 2 + old_name_size); + // Shift current value of name to the end of the new space. + name.replace( + name.size() - old_name_size, old_name_size, name, 0, old_name_size); + name.replace(0, ns_len, ns, ns_len); + name[ns_len] = ':'; + name[ns_len + 1] = ':'; + return true; + } else { + return false; + } + } +}; + +// Non-owning view of an OperatorName. Unlike OperatorName, most of +// its functions are constexpr, so it can be used for compile time +// computations +struct OperatorNameView final { + std::string_view name; + std::string_view overload_name; + constexpr OperatorNameView( + std::string_view name, + std::string_view overload_name) + : name(name), overload_name(overload_name) {} + // Parses strings like "foo.overload" and also "foo" + constexpr static OperatorNameView parse(std::string_view full_name) { + auto i = full_name.find('.'); + if (i == std::string_view::npos) { + return OperatorNameView(full_name, std::string_view()); + } else { + return OperatorNameView(full_name.substr(0, i), full_name.substr(i + 1)); + } + } +}; + +inline bool operator==(const OperatorName& lhs, const OperatorName& rhs) { + return lhs.name == rhs.name && lhs.overload_name == rhs.overload_name; +} + +inline bool operator!=(const OperatorName& lhs, const OperatorName& rhs) { + return !operator==(lhs, rhs); +} + +TORCH_API std::string toString(const OperatorName& opName); +TORCH_API std::ostream& operator<<(std::ostream& /*os*/, const OperatorName& /*opName*/); + +} // namespace c10 + +namespace std { +template <> +struct hash<::c10::OperatorName> { + size_t operator()(const ::c10::OperatorName& x) const { + return std::hash()(x.name) ^ + (~std::hash()(x.overload_name)); + } +}; +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/qualified_name.h b/backends/pytorch/include/ATen/core/qualified_name.h new file mode 100644 index 000000000..60e05fd90 --- /dev/null +++ b/backends/pytorch/include/ATen/core/qualified_name.h @@ -0,0 +1,166 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include + +namespace c10 { + +// Represents a name of the form "foo.bar.baz" +struct QualifiedName { + QualifiedName() = default; + + // `name` can be a dotted string, like "foo.bar.baz", or just a bare name. + /* implicit */ QualifiedName(const std::string& name) { + TORCH_CHECK(!name.empty()); + // split the string into its atoms. + size_t startSearchFrom = 0; + size_t pos = name.find(delimiter_, startSearchFrom); + + while (pos != std::string::npos) { + auto atom = name.substr(startSearchFrom, pos - startSearchFrom); + TORCH_INTERNAL_ASSERT( + !atom.empty(), "Invalid name for qualified name: '", name, "'"); + atoms_.push_back(std::move(atom)); + startSearchFrom = pos + 1; + pos = name.find(delimiter_, startSearchFrom); + } + + auto finalAtom = name.substr(startSearchFrom); + TORCH_INTERNAL_ASSERT( + !finalAtom.empty(), "Invalid name for qualified name: '", name, "'"); + atoms_.emplace_back(std::move(finalAtom)); + + cacheAccessors(); + } + + explicit QualifiedName(std::vector atoms) : atoms_(std::move(atoms)) { + for (const auto& atom : atoms_) { + TORCH_CHECK(!atom.empty(), "Atom cannot be empty"); + TORCH_CHECK( + atom.find(delimiter_) == std::string::npos, + "Delimiter not allowed in atom"); + } + + cacheAccessors(); + } + // Unnecessary copy. Ideally we'd use something like std::string_view. + /* implicit */ QualifiedName(const char* name) + : QualifiedName(std::string(name)) {} + + // `name` must be a bare name (no dots!) + explicit QualifiedName(const QualifiedName& prefix, std::string name) { + TORCH_INTERNAL_ASSERT(!name.empty()); + TORCH_INTERNAL_ASSERT(name.find(delimiter_) == std::string::npos); + atoms_.insert(atoms_.begin(), prefix.atoms_.begin(), prefix.atoms_.end()); + atoms_.push_back(std::move(name)); + + cacheAccessors(); + } + + // Is `this` a prefix of `other`? + // For example, "foo.bar" is a prefix of "foo.bar.baz" + bool isPrefixOf(const QualifiedName& other) const { + const auto& thisAtoms = atoms_; + const auto& otherAtoms = other.atoms_; + + if (thisAtoms.size() > otherAtoms.size()) { + // Can't be a prefix if it's bigger + return false; + } + for (const auto i : c10::irange(thisAtoms.size())) { + if (thisAtoms[i] != otherAtoms[i]) { + return false; + } + } + return true; + } + + // The fully qualified name, like "foo.bar.baz" + const std::string& qualifiedName() const { + return qualifiedName_; + } + + // The leading qualifier, like "foo.bar" + const std::string& prefix() const { + return prefix_; + } + + // The base name, like "baz" + const std::string& name() const { + return name_; + } + + const std::vector& atoms() const { + return atoms_; + } + + bool operator==(const QualifiedName& other) const { + return this->qualifiedName_ == other.qualifiedName_; + } + + bool operator!=(const QualifiedName& other) const { + return !(*this == other); + } + + private: + static constexpr char delimiter_ = '.'; + + // Helper for cacheAccessors() below. + template + std::string join(char delimiter, const T& v) { + std::string out; + size_t reserve = 0; + for (const auto& e : v) { + reserve += e.size() + 1; + } + out.reserve(reserve); + for (const auto i : c10::irange(v.size())) { + if (i != 0) { + out.push_back(delimiter); + } + out.append(v[i]); + } + return out; + } + + void cacheAccessors() { + qualifiedName_ = join(delimiter_, atoms_); + if (atoms_.size() > 1) { + ArrayRef view(atoms_); + const auto prefixView = view.slice(0, view.size() - 1); + prefix_ = join(delimiter_, prefixView); + } + + if (!atoms_.empty()) { + name_ = atoms_.back(); + } + } + + // The actual list of names, like "{foo, bar, baz}" + std::vector atoms_; + + /* + * Cached accessors, derived from `atoms_`. + */ + std::string qualifiedName_; + std::string prefix_; + std::string name_; +}; +} // namespace c10 + +namespace std { +template <> +struct hash { + size_t operator()(const c10::QualifiedName& n) const noexcept { + return std::hash()(n.qualifiedName()); + } +}; +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/rref_interface.h b/backends/pytorch/include/ATen/core/rref_interface.h new file mode 100644 index 000000000..b40fe9ed5 --- /dev/null +++ b/backends/pytorch/include/ATen/core/rref_interface.h @@ -0,0 +1,46 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace c10 { + +struct Type; +using worker_id_t = int16_t; + +// This abstract class contains only user-facing APIs, and will be shared +// between jit and distributed to implement TorchScript support. +class C10_EXPORT RRefInterface : public c10::intrusive_ptr_target { + public: + RRefInterface() = default; + // RRef is made NOT copyable NOT movable to prevent messing up reference + // counting. + RRefInterface(const RRefInterface& other) = delete; + RRefInterface(RRefInterface&& other) = delete; + RRefInterface& operator=(const RRefInterface& other) = delete; + RRefInterface& operator=(RRefInterface&& other) = delete; + + ~RRefInterface() override = default; + + // returns the worker id of the owner + virtual worker_id_t owner() const = 0; + + // returns the worker name of the owner + virtual std::string ownerName() const = 0; + + // Returns true if this is the ``OwnerRRef`` + virtual bool isOwner() const = 0; + + // Returns true if this is an ``OwnerRRef`` or if this ``UserRRef`` has been + // confirmed by its owner. + virtual bool confirmedByOwner() const = 0; + + virtual const TypePtr type() const = 0; +}; + +} + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/symbol.h b/backends/pytorch/include/ATen/core/symbol.h new file mode 100644 index 000000000..759d2ae76 --- /dev/null +++ b/backends/pytorch/include/ATen/core/symbol.h @@ -0,0 +1,152 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once +#include +#include +#include // For std::hash +#include + + +namespace c10 { + +// 'prim' symbols are synthetic operators that occur only in the IR +// and don't have corresponding implementations in ATen. + +// 'onnx' symbols correspond to ONNX operators. Their semantics +// are defined in https://github.com/onnx/onnx/blob/master/docs/Operators.md +// The particular version we are targeting is specified by '_onnx_opset_version' +// in torch.onnx.symbolic_helper +// +// In general, most ONNX operators won't get an entry here, because they +// are handled from the Python end. However, you may occasionally need +// to intern an ONNX symbol here so that you can conveniently write an +// optimization on ONNX operations. + +// 'attr' symbols are attribute keys. They are shared between both ONNX and ATen +// operators (you disambiguate their meaning by looking at the operator itself). +// In general, you only need to define attribute keys that are used by +// onnx or prim; ATen attributes are automatically generated in FORALL_ATTR_BASE_SYMBOLS. + +// Note [Symbol allocation] +// ~~~~~~~~~~~~~~~~~~~~~~~~ +// +// 1. Symbol namespace is split up into namespaces. +// +// 2. The intended access pattern for built-in symbols is onnx::MatMul +// in the c10 namespace (this is a Symbol). +// + +// Built-in constant definition strategy: +// - Enum is the most convenient way to generate a contiguous sequence +// of numbers for an identifier. +// - However, an enum gives you a fresh type. We want onnx::MatMul to +// be type Symbol, not some random enum type! +// - Therefore, after using enums to generate the sequence of integers, +// we then declare constexpr Symbols to get everything the actual Symbol +// type we want. Symbols must be constexpr to be valid to be "case"ed on. + +using unique_t = uint32_t; + +const std::string& domain_prefix(); + +// A Symbol is like an interned string, but with a little extra +// structure; it is namespaced via SymbolNamespace and the resulting +// intern pointers support efficient namespace testing. +struct TORCH_API Symbol { + explicit constexpr Symbol() : value(0) {} + explicit constexpr Symbol(unique_t uniq) + : value(uniq) {} + + // Get a Symbol for a qualified string like "attr::bar" + static Symbol fromQualString(const std::string & s); + + // Get a Symbol from a domain and an unqualified string like "org.pytorch.attr" and "bar" + static Symbol fromDomainAndUnqualString(const std::string & d, const std::string & s); + + // Constructors for our various namespaced strings. This will construct + // the appropriate namespaced string, e.g., "attr::foo" for the + // argument "foo", and then attempt to intern it. DO NOT USE THIS + // with a string literal; attr::foo should be available in that case + // (and if it's not, you should add it to the built-ins list above.) + static Symbol attr(const std::string & s); + static Symbol aten(const std::string & s); + static Symbol cuda(const std::string & s); + static Symbol onnx(const std::string & s); + static Symbol prim(const std::string & s); + static Symbol user(const std::string & s); + static Symbol caffe2(const std::string & s); + static Symbol dimname(const std::string & s); + // TODO: eliminate me + static Symbol scope(const std::string & s); + + bool is_attr() const; + bool is_aten() const; + bool is_cuda() const; + bool is_prim() const; + bool is_prims() const; + bool is_nvprims() const; + bool is_onnx() const; + bool is_user() const; + bool is_caffe2() const; + bool is_dimname() const; + + // So we can switch on this + constexpr operator unique_t() const { + return value; + } + + Symbol ns() const; + + // Give a string corresponding to the unqualified version of this name, e.g., + // "mm". Use this in a context where the intended namespace of the string is + // obvious; this is a *lossy* conversion. + const char * toUnqualString() const; + + // Give a string corresponding to the qualified version of this name, + // e.g., "aten::mm". This string format is made available to Python bindings + // (so we know how to parse it.) + const char * toQualString() const; + + // This describes a symbol in a case where humans read it. At the moment it's + // the same as toQualString. This has to be a const char* returned because + // a lot of printf style macros use it. + const char * toDisplayString() const; + + // Give a string corresponding to the domain name for the symbol, + // e.g., "org.pytorch.aten". + std::string domainString() const; + +private: + + explicit Symbol(Symbol ns, const std::string & s); + unique_t value; +}; + +static inline bool operator==(Symbol lhs, Symbol rhs) { + return static_cast(lhs) == static_cast(rhs); +} + +inline Symbol Symbol::attr(const std::string & s) { return Symbol::fromQualString("attr::" + s); } +inline Symbol Symbol::aten(const std::string & s) { return Symbol::fromQualString("aten::" + s); } +inline Symbol Symbol::cuda(const std::string & s) { return Symbol::fromQualString("cuda::" + s); } +inline Symbol Symbol::onnx(const std::string & s) { return Symbol::fromQualString("onnx::" + s); } +inline Symbol Symbol::prim(const std::string & s) { return Symbol::fromQualString("prim::" + s); } +inline Symbol Symbol::scope(const std::string & s) { return Symbol::fromQualString("scope::" + s); } +inline Symbol Symbol::user(const std::string & s) { return Symbol::fromQualString("user::" + s); } +inline Symbol Symbol::caffe2(const std::string & s) { return Symbol::fromQualString("_caffe2::" + s); } +inline Symbol Symbol::dimname(const std::string & s) { return Symbol::fromQualString("dimname::" + s); } + +} // namespace c10 + +// make symbol behave like an integer in hash tables +namespace std { +template <> +struct hash { + size_t operator()(c10::Symbol s) const { + return std::hash()(static_cast(s)); + } +}; +} + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/type_factory.h b/backends/pytorch/include/ATen/core/type_factory.h new file mode 100644 index 000000000..a0ffab285 --- /dev/null +++ b/backends/pytorch/include/ATen/core/type_factory.h @@ -0,0 +1,113 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include +#include +#include + +namespace c10 { + +template +struct TORCH_API TypeFactoryBase {}; + +template <> +struct TORCH_API TypeFactoryBase { + template + static c10::DynamicTypePtr create(TypePtr ty, Args&&... args) { + return std::make_shared( + c10::DynamicTypeTrait::tagValue(), + c10::DynamicType::Arguments(c10::ArrayRef( + {std::move(ty), std::forward(args)...}))); + } + template + static c10::DynamicTypePtr create(const std::vector& types) { + return std::make_shared( + c10::DynamicTypeTrait::tagValue(), + c10::DynamicType::Arguments(types)); + } + static c10::DynamicTypePtr createNamedTuple( + const std::string& name, + const std::vector& fields, + const std::vector& types) { + return std::make_shared( + c10::DynamicType::Tag::Tuple, + name, + c10::DynamicType::Arguments(fields, types)); + } + template + C10_ERASE static c10::DynamicTypePtr createNamed(const std::string& name) { + return std::make_shared( + c10::DynamicTypeTrait::tagValue(), + name, + c10::DynamicType::Arguments{}); + } + template + C10_ERASE static decltype(auto) get() { + return DynamicTypeTrait::getBaseType(); + } + static const std::unordered_map& basePythonTypes(); +}; + +using DynamicTypeFactory = TypeFactoryBase; + +// Helper functions for constructing DynamicTypes inline. +template < + typename T, + std::enable_if_t::isBaseType, int> = 0> +C10_ERASE DynamicTypePtr dynT() { + return DynamicTypeFactory::get(); +} + +template < + typename T, + typename... Args, + std::enable_if_t::isBaseType, int> = 0> +C10_ERASE DynamicTypePtr dynT(Args&&... args) { + return DynamicTypeFactory::create(std::forward(args)...); +} + +template <> +struct TORCH_API TypeFactoryBase { + template + static c10::TypePtr create(TypePtr ty, Args&&... args) { + return T::create(std::move(ty), std::forward(args)...); + } + template + static c10::TypePtr create(std::vector types) { + return T::create(std::move(types)); + } + static c10::TypePtr createNamedTuple( + const std::string& name, + const std::vector& fields, + const std::vector& types); + template + C10_ERASE static c10::TypePtr createNamed(const std::string& name) { + return T::create(name); + } + static const std::unordered_map& basePythonTypes(); + template + C10_ERASE static c10::TypePtr get() { + return T::get(); + } +}; + +using DefaultTypeFactory = TypeFactoryBase; + +using PlatformType = +#ifdef C10_MOBILE + c10::DynamicType +#else + c10::Type +#endif + ; + +using TypeFactory = TypeFactoryBase; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/core/type_ptr.h b/backends/pytorch/include/ATen/core/type_ptr.h new file mode 100644 index 000000000..5574060d2 --- /dev/null +++ b/backends/pytorch/include/ATen/core/type_ptr.h @@ -0,0 +1,59 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include +#include + +namespace c10 { + +// Compatibility wrapper around a raw pointer so that existing code +// written to deal with a shared_ptr can keep working. +template +class SingletonTypePtr { + public: + /* implicit */ SingletonTypePtr(T* p) : repr_(p) {} + + // We need this to satisfy Pybind11, but it shouldn't be hit. + explicit SingletonTypePtr(std::shared_ptr /*unused*/) { TORCH_CHECK(false); } + + using element_type = typename std::shared_ptr::element_type; + + template , void>, bool> = true> + T& operator*() const { + return *repr_; + } + + T* get() const { + return repr_; + } + + T* operator->() const { + return repr_; + } + + operator bool() const { + return repr_ != nullptr; + } + + private: + T* repr_{nullptr}; +}; + +template +bool operator==(SingletonTypePtr lhs, SingletonTypePtr rhs) { + return (void*)lhs.get() == (void*)rhs.get(); +} + +template +bool operator!=(SingletonTypePtr lhs, SingletonTypePtr rhs) { + return !(lhs == rhs); +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_addmm_activation_ops.h b/backends/pytorch/include/ATen/ops/_addmm_activation_ops.h new file mode 100644 index 000000000..095e60934 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_addmm_activation_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _addmm_activation_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_addmm_activation"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_addmm_activation.out(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1, bool use_gelu=False, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha, bool use_gelu, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha, bool use_gelu, at::Tensor & out); +}; + +struct TORCH_API _addmm_activation { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_addmm_activation"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_addmm_activation(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1, bool use_gelu=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha, bool use_gelu); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha, bool use_gelu); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_autocast_to_full_precision_ops.h b/backends/pytorch/include/ATen/ops/_autocast_to_full_precision_ops.h new file mode 100644 index 000000000..f505b7df7 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_autocast_to_full_precision_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _autocast_to_full_precision { + using schema = at::Tensor (const at::Tensor &, bool, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_autocast_to_full_precision"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_autocast_to_full_precision(Tensor(a) self, bool cuda_enabled, bool cpu_enabled) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, bool cuda_enabled, bool cpu_enabled); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool cuda_enabled, bool cpu_enabled); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_autocast_to_reduced_precision_ops.h b/backends/pytorch/include/ATen/ops/_autocast_to_reduced_precision_ops.h new file mode 100644 index 000000000..4f6f669e2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_autocast_to_reduced_precision_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _autocast_to_reduced_precision { + using schema = at::Tensor (const at::Tensor &, bool, bool, at::ScalarType, at::ScalarType); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_autocast_to_reduced_precision"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_autocast_to_reduced_precision(Tensor(a) self, bool cuda_enabled, bool cpu_enabled, ScalarType cuda_dtype, ScalarType cpu_dtype) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, bool cuda_enabled, bool cpu_enabled, at::ScalarType cuda_dtype, at::ScalarType cpu_dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool cuda_enabled, bool cpu_enabled, at::ScalarType cuda_dtype, at::ScalarType cpu_dtype); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_backward_ops.h b/backends/pytorch/include/ATen/ops/_backward_ops.h new file mode 100644 index 000000000..f41b6017f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_backward_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _backward { + using schema = void (const at::Tensor &, at::TensorList, const ::std::optional &, ::std::optional, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_backward"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_backward(Tensor self, Tensor[] inputs, Tensor? gradient=None, bool? retain_graph=None, bool create_graph=False) -> ()"; + static void call(const at::Tensor & self, at::TensorList inputs, const ::std::optional & gradient, ::std::optional retain_graph, bool create_graph); + static void redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::TensorList inputs, const ::std::optional & gradient, ::std::optional retain_graph, bool create_graph); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_coalesced_ops.h b/backends/pytorch/include/ATen/ops/_coalesced_ops.h new file mode 100644 index 000000000..42fab17ac --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_coalesced_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _coalesced_ { + using schema = at::Tensor & (at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_coalesced_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_coalesced_(Tensor(a!) self, bool coalesced) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, bool coalesced); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, bool coalesced); +}; + +struct TORCH_API _coalesced_out { + using schema = at::Tensor & (const at::Tensor &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_coalesced"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_coalesced.out(Tensor self, bool coalesced, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, bool coalesced, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool coalesced, at::Tensor & out); +}; + +struct TORCH_API _coalesced { + using schema = at::Tensor (const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_coalesced"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_coalesced(Tensor self, bool coalesced) -> Tensor"; + static at::Tensor call(const at::Tensor & self, bool coalesced); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool coalesced); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_conj_ops.h b/backends/pytorch/include/ATen/ops/_conj_ops.h new file mode 100644 index 000000000..2741d5ba8 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_conj_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _conj { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_conj"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_conj(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_conj_physical_ops.h b/backends/pytorch/include/ATen/ops/_conj_physical_ops.h new file mode 100644 index 000000000..3ddeb28a2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_conj_physical_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _conj_physical { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_conj_physical"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_conj_physical(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API _conj_physical_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_conj_physical"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_conj_physical.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_dimI_ops.h b/backends/pytorch/include/ATen/ops/_dimI_ops.h new file mode 100644 index 000000000..e1864bed0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_dimI_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _dimI { + using schema = int64_t (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_dimI"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_dimI(Tensor self) -> int"; + static int64_t call(const at::Tensor & self); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_dimV_ops.h b/backends/pytorch/include/ATen/ops/_dimV_ops.h new file mode 100644 index 000000000..faaff9c8b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_dimV_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _dimV { + using schema = int64_t (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_dimV"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_dimV(Tensor self) -> int"; + static int64_t call(const at::Tensor & self); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_fw_primal_ops.h b/backends/pytorch/include/ATen/ops/_fw_primal_ops.h new file mode 100644 index 000000000..8a294abaf --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_fw_primal_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _fw_primal { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_fw_primal"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_fw_primal(Tensor(a) self, int level) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t level); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t level); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_indices_ops.h b/backends/pytorch/include/ATen/ops/_indices_ops.h new file mode 100644 index 000000000..1d6e2f15a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_indices_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _indices { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_indices"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_indices(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_is_all_true_ops.h b/backends/pytorch/include/ATen/ops/_is_all_true_ops.h new file mode 100644 index 000000000..165fbaf05 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_is_all_true_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _is_all_true { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_is_all_true"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_is_all_true(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_is_any_true_ops.h b/backends/pytorch/include/ATen/ops/_is_any_true_ops.h new file mode 100644 index 000000000..e4ab25ecf --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_is_any_true_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _is_any_true { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_is_any_true"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_is_any_true(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_is_zerotensor_ops.h b/backends/pytorch/include/ATen/ops/_is_zerotensor_ops.h new file mode 100644 index 000000000..2c62f2c89 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_is_zerotensor_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _is_zerotensor { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_is_zerotensor"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_is_zerotensor(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_lazy_clone_ops.h b/backends/pytorch/include/ATen/ops/_lazy_clone_ops.h new file mode 100644 index 000000000..6f53940c4 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_lazy_clone_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _lazy_clone { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_lazy_clone"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_lazy_clone(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_neg_view_ops.h b/backends/pytorch/include/ATen/ops/_neg_view_ops.h new file mode 100644 index 000000000..d85f4b622 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_neg_view_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _neg_view { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_neg_view"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_neg_view(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_nested_tensor_size_ops.h b/backends/pytorch/include/ATen/ops/_nested_tensor_size_ops.h new file mode 100644 index 000000000..f3151d080 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_nested_tensor_size_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _nested_tensor_size { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_nested_tensor_size"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_nested_tensor_size(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API _nested_tensor_size_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_nested_tensor_size"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_nested_tensor_size.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_nested_tensor_storage_offsets_ops.h b/backends/pytorch/include/ATen/ops/_nested_tensor_storage_offsets_ops.h new file mode 100644 index 000000000..67e166da2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_nested_tensor_storage_offsets_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _nested_tensor_storage_offsets { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_nested_tensor_storage_offsets"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_nested_tensor_storage_offsets(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API _nested_tensor_storage_offsets_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_nested_tensor_storage_offsets"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_nested_tensor_storage_offsets.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_nested_tensor_strides_ops.h b/backends/pytorch/include/ATen/ops/_nested_tensor_strides_ops.h new file mode 100644 index 000000000..78d5d11e0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_nested_tensor_strides_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _nested_tensor_strides { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_nested_tensor_strides"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_nested_tensor_strides(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API _nested_tensor_strides_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_nested_tensor_strides"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_nested_tensor_strides.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_nnz_ops.h b/backends/pytorch/include/ATen/ops/_nnz_ops.h new file mode 100644 index 000000000..15b6eb465 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_nnz_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _nnz { + using schema = int64_t (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_nnz"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_nnz(Tensor self) -> int"; + static int64_t call(const at::Tensor & self); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_reshape_alias_ops.h b/backends/pytorch/include/ATen/ops/_reshape_alias_ops.h new file mode 100644 index 000000000..4a5916d37 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_reshape_alias_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _reshape_alias { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_reshape_alias"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_reshape_alias(Tensor(a) self, SymInt[] size, SymInt[] stride) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size, c10::SymIntArrayRef stride); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, c10::SymIntArrayRef stride); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_sparse_mask_projection_ops.h b/backends/pytorch/include/ATen/ops/_sparse_mask_projection_ops.h new file mode 100644 index 000000000..adfb5f7d6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_sparse_mask_projection_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _sparse_mask_projection { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_sparse_mask_projection"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_sparse_mask_projection(Tensor self, Tensor mask, bool accumulate_matches=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mask, bool accumulate_matches); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask, bool accumulate_matches); +}; + +struct TORCH_API _sparse_mask_projection_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_sparse_mask_projection"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_sparse_mask_projection.out(Tensor self, Tensor mask, bool accumulate_matches=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mask, bool accumulate_matches, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask, bool accumulate_matches, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_to_dense_ops.h b/backends/pytorch/include/ATen/ops/_to_dense_ops.h new file mode 100644 index 000000000..1fd4a4b8e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_to_dense_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _to_dense { + using schema = at::Tensor (const at::Tensor &, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_dense"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_to_dense(Tensor self, ScalarType? dtype=None, bool? masked_grad=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dtype, ::std::optional masked_grad); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype, ::std::optional masked_grad); +}; + +struct TORCH_API _to_dense_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_dense"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_to_dense.out(Tensor self, ScalarType? dtype=None, bool? masked_grad=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional dtype, ::std::optional masked_grad, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype, ::std::optional masked_grad, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_to_sparse_bsc_ops.h b/backends/pytorch/include/ATen/ops/_to_sparse_bsc_ops.h new file mode 100644 index 000000000..51c32da71 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_to_sparse_bsc_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _to_sparse_bsc { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse_bsc"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_to_sparse_bsc(Tensor self, int[2] blocksize, int? dense_dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim); +}; + +struct TORCH_API _to_sparse_bsc_out { + using schema = at::Tensor & (const at::Tensor &, at::IntArrayRef, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse_bsc"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_to_sparse_bsc.out(Tensor self, int[2] blocksize, int? dense_dim=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_to_sparse_bsr_ops.h b/backends/pytorch/include/ATen/ops/_to_sparse_bsr_ops.h new file mode 100644 index 000000000..7e38fce86 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_to_sparse_bsr_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _to_sparse_bsr { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse_bsr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_to_sparse_bsr(Tensor self, int[2] blocksize, int? dense_dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim); +}; + +struct TORCH_API _to_sparse_bsr_out { + using schema = at::Tensor & (const at::Tensor &, at::IntArrayRef, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse_bsr"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_to_sparse_bsr.out(Tensor self, int[2] blocksize, int? dense_dim=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_to_sparse_csc_ops.h b/backends/pytorch/include/ATen/ops/_to_sparse_csc_ops.h new file mode 100644 index 000000000..05c7b7cb2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_to_sparse_csc_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _to_sparse_csc { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse_csc"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_to_sparse_csc(Tensor self, int? dense_dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dense_dim); +}; + +struct TORCH_API _to_sparse_csc_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse_csc"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_to_sparse_csc.out(Tensor self, int? dense_dim=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional dense_dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dense_dim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_to_sparse_csr_ops.h b/backends/pytorch/include/ATen/ops/_to_sparse_csr_ops.h new file mode 100644 index 000000000..e6b79e106 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_to_sparse_csr_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _to_sparse_csr { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse_csr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_to_sparse_csr(Tensor self, int? dense_dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dense_dim); +}; + +struct TORCH_API _to_sparse_csr_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse_csr"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_to_sparse_csr.out(Tensor self, int? dense_dim=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional dense_dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dense_dim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_to_sparse_ops.h b/backends/pytorch/include/ATen/ops/_to_sparse_ops.h new file mode 100644 index 000000000..8d4acd384 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_to_sparse_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _to_sparse_sparse_dim { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse"; + static constexpr const char* overload_name = "sparse_dim"; + static constexpr const char* schema_str = "_to_sparse.sparse_dim(Tensor self, int sparse_dim) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t sparse_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t sparse_dim); +}; + +struct TORCH_API _to_sparse { + using schema = at::Tensor (const at::Tensor &, ::std::optional, at::OptionalIntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_to_sparse(Tensor self, *, Layout? layout=None, int[2]? blocksize=None, int? dense_dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional layout, at::OptionalIntArrayRef blocksize, ::std::optional dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional layout, at::OptionalIntArrayRef blocksize, ::std::optional dense_dim); +}; + +struct TORCH_API _to_sparse_sparse_dim_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse"; + static constexpr const char* overload_name = "sparse_dim_out"; + static constexpr const char* schema_str = "_to_sparse.sparse_dim_out(Tensor self, int sparse_dim, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t sparse_dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t sparse_dim, at::Tensor & out); +}; + +struct TORCH_API _to_sparse_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::OptionalIntArrayRef, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_to_sparse"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "_to_sparse.out(Tensor self, *, Layout? layout=None, int[2]? blocksize=None, int? dense_dim=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional layout, at::OptionalIntArrayRef blocksize, ::std::optional dense_dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional layout, at::OptionalIntArrayRef blocksize, ::std::optional dense_dim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_values_ops.h b/backends/pytorch/include/ATen/ops/_values_ops.h new file mode 100644 index 000000000..3f90a341a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_values_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _values { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_values"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_values(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/_version_ops.h b/backends/pytorch/include/ATen/ops/_version_ops.h new file mode 100644 index 000000000..8176a1769 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/_version_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API _version { + using schema = int64_t (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::_version"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "_version(Tensor self) -> int"; + static int64_t call(const at::Tensor & self); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/abs_ops.h b/backends/pytorch/include/ATen/ops/abs_ops.h new file mode 100644 index 000000000..033f19aad --- /dev/null +++ b/backends/pytorch/include/ATen/ops/abs_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API abs { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::abs"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "abs(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API abs_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::abs_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "abs_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API abs_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::abs"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "abs.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/absolute_ops.h b/backends/pytorch/include/ATen/ops/absolute_ops.h new file mode 100644 index 000000000..39eda76b0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/absolute_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API absolute { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::absolute"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "absolute(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API absolute_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::absolute_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "absolute_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API absolute_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::absolute"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "absolute.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/acos_ops.h b/backends/pytorch/include/ATen/ops/acos_ops.h new file mode 100644 index 000000000..00463b8fc --- /dev/null +++ b/backends/pytorch/include/ATen/ops/acos_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API acos { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::acos"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "acos(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API acos_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::acos_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "acos_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API acos_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::acos"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "acos.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/acosh_ops.h b/backends/pytorch/include/ATen/ops/acosh_ops.h new file mode 100644 index 000000000..a4759644b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/acosh_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API acosh { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::acosh"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "acosh(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API acosh_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::acosh_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "acosh_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API acosh_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::acosh"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "acosh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/add_ops.h b/backends/pytorch/include/ATen/ops/add_ops.h new file mode 100644 index 000000000..8a244bf1c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/add_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API add_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::add"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "add.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); +}; + +struct TORCH_API add__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::add_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "add_.Tensor(Tensor(a!) self, Tensor other, *, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); +}; + +struct TORCH_API add_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::add"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "add.out(Tensor self, Tensor other, *, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha, at::Tensor & out); +}; + +struct TORCH_API add_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::add"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "add.Scalar(Tensor self, Scalar other, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); +}; + +struct TORCH_API add__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::add_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "add_.Scalar(Tensor(a!) self, Scalar other, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); +}; + +struct TORCH_API add_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::add"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "add.Scalar_out(Tensor self, Scalar other, Scalar alpha=1, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/addbmm_ops.h b/backends/pytorch/include/ATen/ops/addbmm_ops.h new file mode 100644 index 000000000..ebaeb99e8 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/addbmm_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API addbmm_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addbmm_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addbmm_(Tensor(a!) self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API addbmm_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addbmm"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "addbmm.out(Tensor self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); +}; + +struct TORCH_API addbmm { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addbmm"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addbmm(Tensor self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/addcdiv_ops.h b/backends/pytorch/include/ATen/ops/addcdiv_ops.h new file mode 100644 index 000000000..1838761a1 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/addcdiv_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API addcdiv_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addcdiv"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "addcdiv.out(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value, at::Tensor & out); +}; + +struct TORCH_API addcdiv { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addcdiv"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addcdiv(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value); +}; + +struct TORCH_API addcdiv_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addcdiv_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addcdiv_(Tensor(a!) self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/addcmul_ops.h b/backends/pytorch/include/ATen/ops/addcmul_ops.h new file mode 100644 index 000000000..08d72dacf --- /dev/null +++ b/backends/pytorch/include/ATen/ops/addcmul_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API addcmul_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addcmul"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "addcmul.out(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value, at::Tensor & out); +}; + +struct TORCH_API addcmul { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addcmul"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addcmul(Tensor self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value); +}; + +struct TORCH_API addcmul_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addcmul_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addcmul_(Tensor(a!) self, Tensor tensor1, Tensor tensor2, *, Scalar value=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & tensor1, const at::Tensor & tensor2, const at::Scalar & value); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/addmm_ops.h b/backends/pytorch/include/ATen/ops/addmm_ops.h new file mode 100644 index 000000000..ed7d4ff3f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/addmm_ops.h @@ -0,0 +1,78 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API addmm_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addmm"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "addmm.out(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); +}; + +struct TORCH_API addmm { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addmm"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addmm(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API addmm_dtype { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, at::ScalarType, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addmm"; + static constexpr const char* overload_name = "dtype"; + static constexpr const char* schema_str = "addmm.dtype(Tensor self, Tensor mat1, Tensor mat2, ScalarType out_dtype, *, Scalar beta=1, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, at::ScalarType out_dtype, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, at::ScalarType out_dtype, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API addmm_dtype_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, at::ScalarType, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addmm"; + static constexpr const char* overload_name = "dtype_out"; + static constexpr const char* schema_str = "addmm.dtype_out(Tensor self, Tensor mat1, Tensor mat2, ScalarType out_dtype, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, at::ScalarType out_dtype, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, at::ScalarType out_dtype, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); +}; + +struct TORCH_API addmm_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addmm_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addmm_(Tensor(a!) self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/addmv_ops.h b/backends/pytorch/include/ATen/ops/addmv_ops.h new file mode 100644 index 000000000..c56ab77c8 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/addmv_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API addmv { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addmv"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addmv(Tensor self, Tensor mat, Tensor vec, *, Scalar beta=1, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mat, const at::Tensor & vec, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat, const at::Tensor & vec, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API addmv_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addmv_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addmv_(Tensor(a!) self, Tensor mat, Tensor vec, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & mat, const at::Tensor & vec, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & mat, const at::Tensor & vec, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API addmv_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addmv"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "addmv.out(Tensor self, Tensor mat, Tensor vec, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mat, const at::Tensor & vec, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat, const at::Tensor & vec, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/addr_ops.h b/backends/pytorch/include/ATen/ops/addr_ops.h new file mode 100644 index 000000000..a143bcaf7 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/addr_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API addr { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addr(Tensor self, Tensor vec1, Tensor vec2, *, Scalar beta=1, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & vec1, const at::Tensor & vec2, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & vec1, const at::Tensor & vec2, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API addr_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addr_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "addr_(Tensor(a!) self, Tensor vec1, Tensor vec2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & vec1, const at::Tensor & vec2, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & vec1, const at::Tensor & vec2, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API addr_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::addr"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "addr.out(Tensor self, Tensor vec1, Tensor vec2, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & vec1, const at::Tensor & vec2, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & vec1, const at::Tensor & vec2, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/adjoint_ops.h b/backends/pytorch/include/ATen/ops/adjoint_ops.h new file mode 100644 index 000000000..bc2ac0f87 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/adjoint_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API adjoint { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::adjoint"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "adjoint(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/alias_ops.h b/backends/pytorch/include/ATen/ops/alias_ops.h new file mode 100644 index 000000000..18ce1337d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/alias_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API alias { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::alias"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "alias(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/align_as_ops.h b/backends/pytorch/include/ATen/ops/align_as_ops.h new file mode 100644 index 000000000..d3b1f0f25 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/align_as_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API align_as { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::align_as"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "align_as(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/align_to_ops.h b/backends/pytorch/include/ATen/ops/align_to_ops.h new file mode 100644 index 000000000..a3c0c7820 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/align_to_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API align_to { + using schema = at::Tensor (const at::Tensor &, at::DimnameList); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::align_to"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "align_to(Tensor(a) self, Dimname[] names) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::DimnameList names); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList names); +}; + +struct TORCH_API align_to_ellipsis_idx { + using schema = at::Tensor (const at::Tensor &, at::DimnameList, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::align_to"; + static constexpr const char* overload_name = "ellipsis_idx"; + static constexpr const char* schema_str = "align_to.ellipsis_idx(Tensor(a) self, Dimname[] order, int ellipsis_idx) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::DimnameList order, int64_t ellipsis_idx); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList order, int64_t ellipsis_idx); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/all_ops.h b/backends/pytorch/include/ATen/ops/all_ops.h new file mode 100644 index 000000000..087f5eec4 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/all_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API all_dim { + using schema = at::Tensor (const at::Tensor &, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::all"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "all.dim(Tensor self, int dim, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim); +}; + +struct TORCH_API all_dims { + using schema = at::Tensor (const at::Tensor &, at::OptionalIntArrayRef, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::all"; + static constexpr const char* overload_name = "dims"; + static constexpr const char* schema_str = "all.dims(Tensor self, int[]? dim=None, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim); +}; + +struct TORCH_API all_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::all"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "all.out(Tensor self, int dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API all_dims_out { + using schema = at::Tensor & (const at::Tensor &, at::OptionalIntArrayRef, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::all"; + static constexpr const char* overload_name = "dims_out"; + static constexpr const char* schema_str = "all.dims_out(Tensor self, int[]? dim=None, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API all_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::all"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "all.dimname(Tensor self, Dimname dim, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim); +}; + +struct TORCH_API all_dimname_out { + using schema = at::Tensor & (const at::Tensor &, at::Dimname, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::all"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "all.dimname_out(Tensor self, Dimname dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API all { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::all"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "all(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API all_all_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::all"; + static constexpr const char* overload_name = "all_out"; + static constexpr const char* schema_str = "all.all_out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/allclose_ops.h b/backends/pytorch/include/ATen/ops/allclose_ops.h new file mode 100644 index 000000000..5ba5108bc --- /dev/null +++ b/backends/pytorch/include/ATen/ops/allclose_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API allclose { + using schema = bool (const at::Tensor &, const at::Tensor &, double, double, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::allclose"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "allclose(Tensor self, Tensor other, float rtol=1e-05, float atol=1e-08, bool equal_nan=False) -> bool"; + static bool call(const at::Tensor & self, const at::Tensor & other, double rtol, double atol, bool equal_nan); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, double rtol, double atol, bool equal_nan); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/amax_ops.h b/backends/pytorch/include/ATen/ops/amax_ops.h new file mode 100644 index 000000000..172e88f15 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/amax_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API amax { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::amax"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "amax(Tensor self, int[1] dim=[], bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim, bool keepdim); +}; + +struct TORCH_API amax_out { + using schema = at::Tensor & (const at::Tensor &, at::IntArrayRef, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::amax"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "amax.out(Tensor self, int[1] dim=[], bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::IntArrayRef dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim, bool keepdim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/amin_ops.h b/backends/pytorch/include/ATen/ops/amin_ops.h new file mode 100644 index 000000000..f384c9f05 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/amin_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API amin { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::amin"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "amin(Tensor self, int[1] dim=[], bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim, bool keepdim); +}; + +struct TORCH_API amin_out { + using schema = at::Tensor & (const at::Tensor &, at::IntArrayRef, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::amin"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "amin.out(Tensor self, int[1] dim=[], bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::IntArrayRef dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim, bool keepdim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/aminmax_ops.h b/backends/pytorch/include/ATen/ops/aminmax_ops.h new file mode 100644 index 000000000..8c77d4f15 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/aminmax_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API aminmax { + using schema = ::std::tuple (const at::Tensor &, ::std::optional, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::aminmax"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "aminmax(Tensor self, *, int? dim=None, bool keepdim=False) -> (Tensor min, Tensor max)"; + static ::std::tuple call(const at::Tensor & self, ::std::optional dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dim, bool keepdim); +}; + +struct TORCH_API aminmax_out { + using schema = ::std::tuple (const at::Tensor &, ::std::optional, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::aminmax"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "aminmax.out(Tensor self, *, int? dim=None, bool keepdim=False, Tensor(a!) min, Tensor(b!) max) -> (Tensor(a!) min, Tensor(b!) max)"; + static ::std::tuple call(const at::Tensor & self, ::std::optional dim, bool keepdim, at::Tensor & min, at::Tensor & max); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dim, bool keepdim, at::Tensor & min, at::Tensor & max); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/and_ops.h b/backends/pytorch/include/ATen/ops/and_ops.h new file mode 100644 index 000000000..32eb545e4 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/and_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API __and___Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__and__"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "__and__.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API __and___Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__and__"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "__and__.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API __iand___Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__iand__"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "__iand__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API __iand___Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__iand__"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "__iand__.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/angle_ops.h b/backends/pytorch/include/ATen/ops/angle_ops.h new file mode 100644 index 000000000..4209716da --- /dev/null +++ b/backends/pytorch/include/ATen/ops/angle_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API angle { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::angle"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "angle(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API angle_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::angle"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "angle.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/any_ops.h b/backends/pytorch/include/ATen/ops/any_ops.h new file mode 100644 index 000000000..d703e2cfb --- /dev/null +++ b/backends/pytorch/include/ATen/ops/any_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API any_dim { + using schema = at::Tensor (const at::Tensor &, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::any"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "any.dim(Tensor self, int dim, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim); +}; + +struct TORCH_API any_dims { + using schema = at::Tensor (const at::Tensor &, at::OptionalIntArrayRef, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::any"; + static constexpr const char* overload_name = "dims"; + static constexpr const char* schema_str = "any.dims(Tensor self, int[]? dim=None, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim); +}; + +struct TORCH_API any_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::any"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "any.out(Tensor self, int dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API any_dims_out { + using schema = at::Tensor & (const at::Tensor &, at::OptionalIntArrayRef, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::any"; + static constexpr const char* overload_name = "dims_out"; + static constexpr const char* schema_str = "any.dims_out(Tensor self, int[]? dim=None, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API any_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::any"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "any.dimname(Tensor self, Dimname dim, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim); +}; + +struct TORCH_API any_dimname_out { + using schema = at::Tensor & (const at::Tensor &, at::Dimname, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::any"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "any.dimname_out(Tensor self, Dimname dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API any { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::any"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "any(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API any_all_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::any"; + static constexpr const char* overload_name = "all_out"; + static constexpr const char* schema_str = "any.all_out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/arccos_ops.h b/backends/pytorch/include/ATen/ops/arccos_ops.h new file mode 100644 index 000000000..07103c0c4 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/arccos_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API arccos { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arccos"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arccos(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API arccos_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arccos_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arccos_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API arccos_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arccos"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "arccos.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/arccosh_ops.h b/backends/pytorch/include/ATen/ops/arccosh_ops.h new file mode 100644 index 000000000..564c5161a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/arccosh_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API arccosh { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arccosh"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arccosh(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API arccosh_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arccosh_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arccosh_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API arccosh_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arccosh"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "arccosh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/arcsin_ops.h b/backends/pytorch/include/ATen/ops/arcsin_ops.h new file mode 100644 index 000000000..9c59ea88c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/arcsin_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API arcsin { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arcsin"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arcsin(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API arcsin_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arcsin_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arcsin_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API arcsin_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arcsin"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "arcsin.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/arcsinh_ops.h b/backends/pytorch/include/ATen/ops/arcsinh_ops.h new file mode 100644 index 000000000..2f5fdefd3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/arcsinh_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API arcsinh { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arcsinh"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arcsinh(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API arcsinh_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arcsinh_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arcsinh_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API arcsinh_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arcsinh"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "arcsinh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/arctan2_ops.h b/backends/pytorch/include/ATen/ops/arctan2_ops.h new file mode 100644 index 000000000..bba676197 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/arctan2_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API arctan2 { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arctan2"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arctan2(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API arctan2_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arctan2"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "arctan2.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API arctan2_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arctan2_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arctan2_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/arctan_ops.h b/backends/pytorch/include/ATen/ops/arctan_ops.h new file mode 100644 index 000000000..c5b92b61a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/arctan_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API arctan { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arctan"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arctan(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API arctan_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arctan_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arctan_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API arctan_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arctan"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "arctan.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/arctanh_ops.h b/backends/pytorch/include/ATen/ops/arctanh_ops.h new file mode 100644 index 000000000..d98ecc829 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/arctanh_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API arctanh { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arctanh"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arctanh(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API arctanh_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arctanh_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "arctanh_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API arctanh_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::arctanh"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "arctanh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/argmax_ops.h b/backends/pytorch/include/ATen/ops/argmax_ops.h new file mode 100644 index 000000000..b1c340596 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/argmax_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API argmax { + using schema = at::Tensor (const at::Tensor &, ::std::optional, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::argmax"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "argmax(Tensor self, int? dim=None, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dim, bool keepdim); +}; + +struct TORCH_API argmax_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::argmax"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "argmax.out(Tensor self, int? dim=None, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dim, bool keepdim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/argmin_ops.h b/backends/pytorch/include/ATen/ops/argmin_ops.h new file mode 100644 index 000000000..7800d435d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/argmin_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API argmin { + using schema = at::Tensor (const at::Tensor &, ::std::optional, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::argmin"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "argmin(Tensor self, int? dim=None, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dim, bool keepdim); +}; + +struct TORCH_API argmin_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::argmin"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "argmin.out(Tensor self, int? dim=None, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dim, bool keepdim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/argsort_ops.h b/backends/pytorch/include/ATen/ops/argsort_ops.h new file mode 100644 index 000000000..e73698995 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/argsort_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API argsort { + using schema = at::Tensor (const at::Tensor &, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::argsort"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "argsort(Tensor self, int dim=-1, bool descending=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, bool descending); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool descending); +}; + +struct TORCH_API argsort_stable { + using schema = at::Tensor (const at::Tensor &, bool, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::argsort"; + static constexpr const char* overload_name = "stable"; + static constexpr const char* schema_str = "argsort.stable(Tensor self, *, bool stable, int dim=-1, bool descending=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, bool stable, int64_t dim, bool descending); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool stable, int64_t dim, bool descending); +}; + +struct TORCH_API argsort_stable_out { + using schema = at::Tensor & (const at::Tensor &, bool, int64_t, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::argsort"; + static constexpr const char* overload_name = "stable_out"; + static constexpr const char* schema_str = "argsort.stable_out(Tensor self, *, bool stable, int dim=-1, bool descending=False, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, bool stable, int64_t dim, bool descending, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool stable, int64_t dim, bool descending, at::Tensor & out); +}; + +struct TORCH_API argsort_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::argsort"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "argsort.dimname(Tensor self, Dimname dim, bool descending=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, bool descending); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool descending); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/argwhere_ops.h b/backends/pytorch/include/ATen/ops/argwhere_ops.h new file mode 100644 index 000000000..2b5c0b3f8 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/argwhere_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API argwhere { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::argwhere"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "argwhere(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/as_strided_ops.h b/backends/pytorch/include/ATen/ops/as_strided_ops.h new file mode 100644 index 000000000..8bd63b91c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/as_strided_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API as_strided { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef, c10::SymIntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::as_strided"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "as_strided(Tensor(a) self, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset); +}; + +struct TORCH_API as_strided_ { + using schema = const at::Tensor & (const at::Tensor &, c10::SymIntArrayRef, c10::SymIntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::as_strided_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "as_strided_(Tensor(a!) self, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/as_strided_scatter_ops.h b/backends/pytorch/include/ATen/ops/as_strided_scatter_ops.h new file mode 100644 index 000000000..9642f80b4 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/as_strided_scatter_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API as_strided_scatter { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, c10::SymIntArrayRef, c10::SymIntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::as_strided_scatter"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "as_strided_scatter(Tensor self, Tensor src, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & src, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset); +}; + +struct TORCH_API as_strided_scatter_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, c10::SymIntArrayRef, c10::SymIntArrayRef, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::as_strided_scatter"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "as_strided_scatter.out(Tensor self, Tensor src, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & src, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional storage_offset, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/asin_ops.h b/backends/pytorch/include/ATen/ops/asin_ops.h new file mode 100644 index 000000000..a4f3e5366 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/asin_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API asin { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::asin"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "asin(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API asin_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::asin_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "asin_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API asin_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::asin"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "asin.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/asinh_ops.h b/backends/pytorch/include/ATen/ops/asinh_ops.h new file mode 100644 index 000000000..90ee4de61 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/asinh_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API asinh { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::asinh"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "asinh(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API asinh_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::asinh_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "asinh_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API asinh_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::asinh"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "asinh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/atan2_ops.h b/backends/pytorch/include/ATen/ops/atan2_ops.h new file mode 100644 index 000000000..0c304e6de --- /dev/null +++ b/backends/pytorch/include/ATen/ops/atan2_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API atan2_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::atan2"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "atan2.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API atan2_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::atan2_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "atan2_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API atan2 { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::atan2"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "atan2(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/atan_ops.h b/backends/pytorch/include/ATen/ops/atan_ops.h new file mode 100644 index 000000000..c993cad67 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/atan_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API atan { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::atan"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "atan(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API atan_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::atan_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "atan_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API atan_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::atan"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "atan.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/atanh_ops.h b/backends/pytorch/include/ATen/ops/atanh_ops.h new file mode 100644 index 000000000..e04862670 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/atanh_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API atanh { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::atanh"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "atanh(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API atanh_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::atanh_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "atanh_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API atanh_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::atanh"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "atanh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/baddbmm_ops.h b/backends/pytorch/include/ATen/ops/baddbmm_ops.h new file mode 100644 index 000000000..18d0abe09 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/baddbmm_ops.h @@ -0,0 +1,78 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API baddbmm { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::baddbmm"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "baddbmm(Tensor self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API baddbmm_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::baddbmm_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "baddbmm_(Tensor(a!) self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API baddbmm_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::baddbmm"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "baddbmm.out(Tensor self, Tensor batch1, Tensor batch2, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); +}; + +struct TORCH_API baddbmm_dtype { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, at::ScalarType, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::baddbmm"; + static constexpr const char* overload_name = "dtype"; + static constexpr const char* schema_str = "baddbmm.dtype(Tensor self, Tensor batch1, Tensor batch2, ScalarType out_dtype, *, Scalar beta=1, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, at::ScalarType out_dtype, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, at::ScalarType out_dtype, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API baddbmm_dtype_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, at::ScalarType, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::baddbmm"; + static constexpr const char* overload_name = "dtype_out"; + static constexpr const char* schema_str = "baddbmm.dtype_out(Tensor self, Tensor batch1, Tensor batch2, ScalarType out_dtype, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, at::ScalarType out_dtype, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & batch1, const at::Tensor & batch2, at::ScalarType out_dtype, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/bernoulli_ops.h b/backends/pytorch/include/ATen/ops/bernoulli_ops.h new file mode 100644 index 000000000..65c7027d2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/bernoulli_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API bernoulli { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bernoulli"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "bernoulli(Tensor self, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional generator); +}; + +struct TORCH_API bernoulli_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bernoulli"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "bernoulli.out(Tensor self, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API bernoulli__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bernoulli_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bernoulli_.Tensor(Tensor(a!) self, Tensor p, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & p, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & p, ::std::optional generator); +}; + +struct TORCH_API bernoulli__float { + using schema = at::Tensor & (at::Tensor &, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bernoulli_"; + static constexpr const char* overload_name = "float"; + static constexpr const char* schema_str = "bernoulli_.float(Tensor(a!) self, float p=0.5, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, double p, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, double p, ::std::optional generator); +}; + +struct TORCH_API bernoulli_p { + using schema = at::Tensor (const at::Tensor &, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bernoulli"; + static constexpr const char* overload_name = "p"; + static constexpr const char* schema_str = "bernoulli.p(Tensor self, float p, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, double p, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double p, ::std::optional generator); +}; + +struct TORCH_API bernoulli_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bernoulli"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "bernoulli.Tensor_out(Tensor self, Tensor p, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & p, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & p, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API bernoulli_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bernoulli"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bernoulli.Tensor(Tensor self, Tensor p, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & p, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & p, ::std::optional generator); +}; + +struct TORCH_API bernoulli_float_out { + using schema = at::Tensor & (const at::Tensor &, double, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bernoulli"; + static constexpr const char* overload_name = "float_out"; + static constexpr const char* schema_str = "bernoulli.float_out(Tensor self, float p=0.5, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, double p, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double p, ::std::optional generator, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/bincount_ops.h b/backends/pytorch/include/ATen/ops/bincount_ops.h new file mode 100644 index 000000000..d826da3ac --- /dev/null +++ b/backends/pytorch/include/ATen/ops/bincount_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API bincount { + using schema = at::Tensor (const at::Tensor &, const ::std::optional &, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bincount"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "bincount(Tensor self, Tensor? weights=None, SymInt minlength=0) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const ::std::optional & weights, c10::SymInt minlength); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & weights, c10::SymInt minlength); +}; + +struct TORCH_API bincount_out { + using schema = at::Tensor & (const at::Tensor &, const ::std::optional &, c10::SymInt, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bincount"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "bincount.out(Tensor self, Tensor? weights=None, SymInt minlength=0, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const ::std::optional & weights, c10::SymInt minlength, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & weights, c10::SymInt minlength, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/bitwise_and_ops.h b/backends/pytorch/include/ATen/ops/bitwise_and_ops.h new file mode 100644 index 000000000..8d49f3e5f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/bitwise_and_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API bitwise_and_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_and"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "bitwise_and.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API bitwise_and_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_and"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "bitwise_and.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API bitwise_and_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_and"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "bitwise_and.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API bitwise_and_Scalar_Tensor { + using schema = at::Tensor (const at::Scalar &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_and"; + static constexpr const char* overload_name = "Scalar_Tensor"; + static constexpr const char* schema_str = "bitwise_and.Scalar_Tensor(Scalar self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Scalar & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_and_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_and"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bitwise_and.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_and__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_and_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "bitwise_and_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API bitwise_and__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_and_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bitwise_and_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_and_Scalar_Tensor_out { + using schema = at::Tensor & (const at::Scalar &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_and"; + static constexpr const char* overload_name = "Scalar_Tensor_out"; + static constexpr const char* schema_str = "bitwise_and.Scalar_Tensor_out(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Scalar & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/bitwise_left_shift_ops.h b/backends/pytorch/include/ATen/ops/bitwise_left_shift_ops.h new file mode 100644 index 000000000..aaf33aa2f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/bitwise_left_shift_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API bitwise_left_shift_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_left_shift"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bitwise_left_shift.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_left_shift__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_left_shift_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bitwise_left_shift_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_left_shift_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_left_shift"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "bitwise_left_shift.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API bitwise_left_shift_Tensor_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_left_shift"; + static constexpr const char* overload_name = "Tensor_Scalar"; + static constexpr const char* schema_str = "bitwise_left_shift.Tensor_Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API bitwise_left_shift__Tensor_Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_left_shift_"; + static constexpr const char* overload_name = "Tensor_Scalar"; + static constexpr const char* schema_str = "bitwise_left_shift_.Tensor_Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API bitwise_left_shift_Tensor_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_left_shift"; + static constexpr const char* overload_name = "Tensor_Scalar_out"; + static constexpr const char* schema_str = "bitwise_left_shift.Tensor_Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API bitwise_left_shift_Scalar_Tensor { + using schema = at::Tensor (const at::Scalar &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_left_shift"; + static constexpr const char* overload_name = "Scalar_Tensor"; + static constexpr const char* schema_str = "bitwise_left_shift.Scalar_Tensor(Scalar self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Scalar & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_left_shift_Scalar_Tensor_out { + using schema = at::Tensor & (const at::Scalar &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_left_shift"; + static constexpr const char* overload_name = "Scalar_Tensor_out"; + static constexpr const char* schema_str = "bitwise_left_shift.Scalar_Tensor_out(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Scalar & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/bitwise_not_ops.h b/backends/pytorch/include/ATen/ops/bitwise_not_ops.h new file mode 100644 index 000000000..3e752585b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/bitwise_not_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API bitwise_not { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_not"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "bitwise_not(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API bitwise_not_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_not_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "bitwise_not_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API bitwise_not_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_not"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "bitwise_not.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/bitwise_or_ops.h b/backends/pytorch/include/ATen/ops/bitwise_or_ops.h new file mode 100644 index 000000000..aca3c86f5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/bitwise_or_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API bitwise_or_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_or"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "bitwise_or.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API bitwise_or_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_or"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "bitwise_or.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API bitwise_or_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_or"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "bitwise_or.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API bitwise_or_Scalar_Tensor { + using schema = at::Tensor (const at::Scalar &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_or"; + static constexpr const char* overload_name = "Scalar_Tensor"; + static constexpr const char* schema_str = "bitwise_or.Scalar_Tensor(Scalar self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Scalar & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_or_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_or"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bitwise_or.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_or__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_or_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "bitwise_or_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API bitwise_or__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_or_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bitwise_or_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_or_Scalar_Tensor_out { + using schema = at::Tensor & (const at::Scalar &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_or"; + static constexpr const char* overload_name = "Scalar_Tensor_out"; + static constexpr const char* schema_str = "bitwise_or.Scalar_Tensor_out(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Scalar & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/bitwise_right_shift_ops.h b/backends/pytorch/include/ATen/ops/bitwise_right_shift_ops.h new file mode 100644 index 000000000..74b5793d5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/bitwise_right_shift_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API bitwise_right_shift_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_right_shift"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bitwise_right_shift.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_right_shift__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_right_shift_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bitwise_right_shift_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_right_shift_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_right_shift"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "bitwise_right_shift.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API bitwise_right_shift_Tensor_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_right_shift"; + static constexpr const char* overload_name = "Tensor_Scalar"; + static constexpr const char* schema_str = "bitwise_right_shift.Tensor_Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API bitwise_right_shift__Tensor_Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_right_shift_"; + static constexpr const char* overload_name = "Tensor_Scalar"; + static constexpr const char* schema_str = "bitwise_right_shift_.Tensor_Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API bitwise_right_shift_Tensor_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_right_shift"; + static constexpr const char* overload_name = "Tensor_Scalar_out"; + static constexpr const char* schema_str = "bitwise_right_shift.Tensor_Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API bitwise_right_shift_Scalar_Tensor { + using schema = at::Tensor (const at::Scalar &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_right_shift"; + static constexpr const char* overload_name = "Scalar_Tensor"; + static constexpr const char* schema_str = "bitwise_right_shift.Scalar_Tensor(Scalar self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Scalar & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_right_shift_Scalar_Tensor_out { + using schema = at::Tensor & (const at::Scalar &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_right_shift"; + static constexpr const char* overload_name = "Scalar_Tensor_out"; + static constexpr const char* schema_str = "bitwise_right_shift.Scalar_Tensor_out(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Scalar & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/bitwise_xor_ops.h b/backends/pytorch/include/ATen/ops/bitwise_xor_ops.h new file mode 100644 index 000000000..f9a1a75f8 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/bitwise_xor_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API bitwise_xor_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_xor"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "bitwise_xor.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API bitwise_xor_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_xor"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "bitwise_xor.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API bitwise_xor_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_xor"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "bitwise_xor.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API bitwise_xor_Scalar_Tensor { + using schema = at::Tensor (const at::Scalar &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_xor"; + static constexpr const char* overload_name = "Scalar_Tensor"; + static constexpr const char* schema_str = "bitwise_xor.Scalar_Tensor(Scalar self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Scalar & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_xor_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_xor"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bitwise_xor.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_xor__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_xor_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "bitwise_xor_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API bitwise_xor__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_xor_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "bitwise_xor_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API bitwise_xor_Scalar_Tensor_out { + using schema = at::Tensor & (const at::Scalar &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bitwise_xor"; + static constexpr const char* overload_name = "Scalar_Tensor_out"; + static constexpr const char* schema_str = "bitwise_xor.Scalar_Tensor_out(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Scalar & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/bmm_ops.h b/backends/pytorch/include/ATen/ops/bmm_ops.h new file mode 100644 index 000000000..c4990e1c7 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/bmm_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API bmm { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bmm"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "bmm(Tensor self, Tensor mat2) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mat2); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat2); +}; + +struct TORCH_API bmm_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bmm"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "bmm.out(Tensor self, Tensor mat2, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mat2, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat2, at::Tensor & out); +}; + +struct TORCH_API bmm_dtype { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, at::ScalarType); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bmm"; + static constexpr const char* overload_name = "dtype"; + static constexpr const char* schema_str = "bmm.dtype(Tensor self, Tensor mat2, ScalarType out_dtype) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mat2, at::ScalarType out_dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat2, at::ScalarType out_dtype); +}; + +struct TORCH_API bmm_dtype_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::ScalarType, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::bmm"; + static constexpr const char* overload_name = "dtype_out"; + static constexpr const char* schema_str = "bmm.dtype_out(Tensor self, Tensor mat2, ScalarType out_dtype, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mat2, at::ScalarType out_dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat2, at::ScalarType out_dtype, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/broadcast_to_ops.h b/backends/pytorch/include/ATen/ops/broadcast_to_ops.h new file mode 100644 index 000000000..3cc210048 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/broadcast_to_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API broadcast_to { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::broadcast_to"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "broadcast_to(Tensor(a) self, SymInt[] size) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cauchy_ops.h b/backends/pytorch/include/ATen/ops/cauchy_ops.h new file mode 100644 index 000000000..4a9b11773 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cauchy_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cauchy_ { + using schema = at::Tensor & (at::Tensor &, double, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cauchy_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cauchy_(Tensor(a!) self, float median=0, float sigma=1, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, double median, double sigma, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, double median, double sigma, ::std::optional generator); +}; + +struct TORCH_API cauchy_out { + using schema = at::Tensor & (const at::Tensor &, double, double, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cauchy"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cauchy.out(Tensor self, float median=0, float sigma=1, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, double median, double sigma, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double median, double sigma, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API cauchy { + using schema = at::Tensor (const at::Tensor &, double, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cauchy"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cauchy(Tensor self, float median=0, float sigma=1, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, double median, double sigma, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double median, double sigma, ::std::optional generator); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/ccol_indices_ops.h b/backends/pytorch/include/ATen/ops/ccol_indices_ops.h new file mode 100644 index 000000000..fee4463fc --- /dev/null +++ b/backends/pytorch/include/ATen/ops/ccol_indices_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API ccol_indices { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ccol_indices"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "ccol_indices(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/ceil_ops.h b/backends/pytorch/include/ATen/ops/ceil_ops.h new file mode 100644 index 000000000..d3ada1869 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/ceil_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API ceil { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ceil"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "ceil(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API ceil_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ceil_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "ceil_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API ceil_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ceil"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "ceil.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/chalf_ops.h b/backends/pytorch/include/ATen/ops/chalf_ops.h new file mode 100644 index 000000000..5ffb04ca0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/chalf_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API chalf { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::chalf"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "chalf(Tensor self, *, MemoryFormat? memory_format=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional memory_format); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional memory_format); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cholesky_inverse_ops.h b/backends/pytorch/include/ATen/ops/cholesky_inverse_ops.h new file mode 100644 index 000000000..bea385f39 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cholesky_inverse_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cholesky_inverse { + using schema = at::Tensor (const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cholesky_inverse"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cholesky_inverse(Tensor self, bool upper=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, bool upper); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool upper); +}; + +struct TORCH_API cholesky_inverse_out { + using schema = at::Tensor & (const at::Tensor &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cholesky_inverse"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cholesky_inverse.out(Tensor self, bool upper=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, bool upper, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool upper, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cholesky_ops.h b/backends/pytorch/include/ATen/ops/cholesky_ops.h new file mode 100644 index 000000000..c49fe7997 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cholesky_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cholesky_out { + using schema = at::Tensor & (const at::Tensor &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cholesky"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cholesky.out(Tensor self, bool upper=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, bool upper, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool upper, at::Tensor & out); +}; + +struct TORCH_API cholesky { + using schema = at::Tensor (const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cholesky"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cholesky(Tensor self, bool upper=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, bool upper); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool upper); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cholesky_solve_ops.h b/backends/pytorch/include/ATen/ops/cholesky_solve_ops.h new file mode 100644 index 000000000..019e0f4ac --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cholesky_solve_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cholesky_solve_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cholesky_solve"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cholesky_solve.out(Tensor self, Tensor input2, bool upper=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & input2, bool upper, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & input2, bool upper, at::Tensor & out); +}; + +struct TORCH_API cholesky_solve { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cholesky_solve"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cholesky_solve(Tensor self, Tensor input2, bool upper=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & input2, bool upper); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & input2, bool upper); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/chunk_ops.h b/backends/pytorch/include/ATen/ops/chunk_ops.h new file mode 100644 index 000000000..272a44791 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/chunk_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API chunk { + using schema = ::std::vector (const at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::chunk"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "chunk(Tensor(a -> *) self, int chunks, int dim=0) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, int64_t chunks, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t chunks, int64_t dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/clamp_max_ops.h b/backends/pytorch/include/ATen/ops/clamp_max_ops.h new file mode 100644 index 000000000..4aa59d015 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/clamp_max_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API clamp_max { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_max"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "clamp_max(Tensor self, Scalar max) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & max); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & max); +}; + +struct TORCH_API clamp_max_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_max"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "clamp_max.Tensor(Tensor self, Tensor max) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & max); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & max); +}; + +struct TORCH_API clamp_max_ { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_max_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "clamp_max_(Tensor(a!) self, Scalar max) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & max); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & max); +}; + +struct TORCH_API clamp_max__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_max_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "clamp_max_.Tensor(Tensor(a!) self, Tensor max) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & max); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & max); +}; + +struct TORCH_API clamp_max_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_max"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "clamp_max.out(Tensor self, Scalar max, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & max, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & max, at::Tensor & out); +}; + +struct TORCH_API clamp_max_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_max"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "clamp_max.Tensor_out(Tensor self, Tensor max, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & max, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & max, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/clamp_min_ops.h b/backends/pytorch/include/ATen/ops/clamp_min_ops.h new file mode 100644 index 000000000..6cb8c00b3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/clamp_min_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API clamp_min { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_min"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "clamp_min(Tensor self, Scalar min) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & min); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & min); +}; + +struct TORCH_API clamp_min_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_min"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "clamp_min.Tensor(Tensor self, Tensor min) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & min); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & min); +}; + +struct TORCH_API clamp_min_ { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_min_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "clamp_min_(Tensor(a!) self, Scalar min) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & min); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & min); +}; + +struct TORCH_API clamp_min__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_min_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "clamp_min_.Tensor(Tensor(a!) self, Tensor min) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & min); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & min); +}; + +struct TORCH_API clamp_min_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_min"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "clamp_min.out(Tensor self, Scalar min, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & min, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & min, at::Tensor & out); +}; + +struct TORCH_API clamp_min_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_min"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "clamp_min.Tensor_out(Tensor self, Tensor min, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & min, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & min, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/clamp_ops.h b/backends/pytorch/include/ATen/ops/clamp_ops.h new file mode 100644 index 000000000..900b8fc47 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/clamp_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API clamp { + using schema = at::Tensor (const at::Tensor &, const ::std::optional &, const ::std::optional &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "clamp(Tensor self, Scalar? min=None, Scalar? max=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const ::std::optional & min, const ::std::optional & max); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & min, const ::std::optional & max); +}; + +struct TORCH_API clamp_Tensor { + using schema = at::Tensor (const at::Tensor &, const ::std::optional &, const ::std::optional &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "clamp.Tensor(Tensor self, Tensor? min=None, Tensor? max=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const ::std::optional & min, const ::std::optional & max); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & min, const ::std::optional & max); +}; + +struct TORCH_API clamp_ { + using schema = at::Tensor & (at::Tensor &, const ::std::optional &, const ::std::optional &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "clamp_(Tensor(a!) self, Scalar? min=None, Scalar? max=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const ::std::optional & min, const ::std::optional & max); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const ::std::optional & min, const ::std::optional & max); +}; + +struct TORCH_API clamp__Tensor { + using schema = at::Tensor & (at::Tensor &, const ::std::optional &, const ::std::optional &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "clamp_.Tensor(Tensor(a!) self, Tensor? min=None, Tensor? max=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const ::std::optional & min, const ::std::optional & max); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const ::std::optional & min, const ::std::optional & max); +}; + +struct TORCH_API clamp_out { + using schema = at::Tensor & (const at::Tensor &, const ::std::optional &, const ::std::optional &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "clamp.out(Tensor self, Scalar? min=None, Scalar? max=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const ::std::optional & min, const ::std::optional & max, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & min, const ::std::optional & max, at::Tensor & out); +}; + +struct TORCH_API clamp_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const ::std::optional &, const ::std::optional &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clamp"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "clamp.Tensor_out(Tensor self, Tensor? min=None, Tensor? max=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const ::std::optional & min, const ::std::optional & max, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & min, const ::std::optional & max, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/clip_ops.h b/backends/pytorch/include/ATen/ops/clip_ops.h new file mode 100644 index 000000000..e26321154 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/clip_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API clip { + using schema = at::Tensor (const at::Tensor &, const ::std::optional &, const ::std::optional &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clip"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "clip(Tensor self, Scalar? min=None, Scalar? max=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const ::std::optional & min, const ::std::optional & max); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & min, const ::std::optional & max); +}; + +struct TORCH_API clip_Tensor { + using schema = at::Tensor (const at::Tensor &, const ::std::optional &, const ::std::optional &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clip"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "clip.Tensor(Tensor self, Tensor? min=None, Tensor? max=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const ::std::optional & min, const ::std::optional & max); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & min, const ::std::optional & max); +}; + +struct TORCH_API clip_ { + using schema = at::Tensor & (at::Tensor &, const ::std::optional &, const ::std::optional &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clip_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "clip_(Tensor(a!) self, Scalar? min=None, Scalar? max=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const ::std::optional & min, const ::std::optional & max); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const ::std::optional & min, const ::std::optional & max); +}; + +struct TORCH_API clip__Tensor { + using schema = at::Tensor & (at::Tensor &, const ::std::optional &, const ::std::optional &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clip_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "clip_.Tensor(Tensor(a!) self, Tensor? min=None, Tensor? max=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const ::std::optional & min, const ::std::optional & max); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const ::std::optional & min, const ::std::optional & max); +}; + +struct TORCH_API clip_out { + using schema = at::Tensor & (const at::Tensor &, const ::std::optional &, const ::std::optional &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clip"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "clip.out(Tensor self, Scalar? min=None, Scalar? max=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const ::std::optional & min, const ::std::optional & max, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & min, const ::std::optional & max, at::Tensor & out); +}; + +struct TORCH_API clip_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const ::std::optional &, const ::std::optional &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clip"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "clip.Tensor_out(Tensor self, Tensor? min=None, Tensor? max=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const ::std::optional & min, const ::std::optional & max, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & min, const ::std::optional & max, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/clone_ops.h b/backends/pytorch/include/ATen/ops/clone_ops.h new file mode 100644 index 000000000..267df2359 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/clone_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API clone { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clone"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "clone(Tensor self, *, MemoryFormat? memory_format=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional memory_format); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional memory_format); +}; + +struct TORCH_API clone_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::clone"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "clone.out(Tensor self, *, MemoryFormat? memory_format=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional memory_format, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional memory_format, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/coalesce_ops.h b/backends/pytorch/include/ATen/ops/coalesce_ops.h new file mode 100644 index 000000000..034555ca3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/coalesce_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API coalesce { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::coalesce"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "coalesce(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/col_indices_ops.h b/backends/pytorch/include/ATen/ops/col_indices_ops.h new file mode 100644 index 000000000..bcb787b12 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/col_indices_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API col_indices { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::col_indices"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "col_indices(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/conj_ops.h b/backends/pytorch/include/ATen/ops/conj_ops.h new file mode 100644 index 000000000..ef366ceaa --- /dev/null +++ b/backends/pytorch/include/ATen/ops/conj_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API conj { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::conj"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "conj(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/conj_physical_ops.h b/backends/pytorch/include/ATen/ops/conj_physical_ops.h new file mode 100644 index 000000000..811ec8436 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/conj_physical_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API conj_physical { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::conj_physical"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "conj_physical(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API conj_physical_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::conj_physical"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "conj_physical.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API conj_physical_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::conj_physical_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "conj_physical_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/contiguous_ops.h b/backends/pytorch/include/ATen/ops/contiguous_ops.h new file mode 100644 index 000000000..ca21276fb --- /dev/null +++ b/backends/pytorch/include/ATen/ops/contiguous_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API contiguous { + using schema = at::Tensor (const at::Tensor &, at::MemoryFormat); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::contiguous"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "contiguous(Tensor(a) self, *, MemoryFormat memory_format=contiguous_format) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::MemoryFormat memory_format); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::MemoryFormat memory_format); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/copy_ops.h b/backends/pytorch/include/ATen/ops/copy_ops.h new file mode 100644 index 000000000..6388b3af7 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/copy_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API copy { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::copy"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "copy(Tensor self, Tensor src, bool non_blocking=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & src, bool non_blocking); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, bool non_blocking); +}; + +struct TORCH_API copy_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::copy_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "copy_(Tensor(a!) self, Tensor src, bool non_blocking=False) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & src, bool non_blocking); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & src, bool non_blocking); +}; + +struct TORCH_API copy_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::copy"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "copy.out(Tensor self, Tensor src, bool non_blocking=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & src, bool non_blocking, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, bool non_blocking, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/copysign_ops.h b/backends/pytorch/include/ATen/ops/copysign_ops.h new file mode 100644 index 000000000..42bb59042 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/copysign_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API copysign_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::copysign"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "copysign.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API copysign_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::copysign"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "copysign.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API copysign__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::copysign_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "copysign_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API copysign_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::copysign"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "copysign.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API copysign__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::copysign_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "copysign_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API copysign_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::copysign"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "copysign.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/corrcoef_ops.h b/backends/pytorch/include/ATen/ops/corrcoef_ops.h new file mode 100644 index 000000000..7bf201946 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/corrcoef_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API corrcoef { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::corrcoef"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "corrcoef(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cos_ops.h b/backends/pytorch/include/ATen/ops/cos_ops.h new file mode 100644 index 000000000..6a8dacd76 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cos_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cos { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cos"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cos(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API cos_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cos_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cos_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API cos_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cos"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cos.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cosh_ops.h b/backends/pytorch/include/ATen/ops/cosh_ops.h new file mode 100644 index 000000000..9e564e13e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cosh_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cosh { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cosh"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cosh(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API cosh_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cosh_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cosh_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API cosh_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cosh"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cosh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/count_nonzero_ops.h b/backends/pytorch/include/ATen/ops/count_nonzero_ops.h new file mode 100644 index 000000000..22a77be80 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/count_nonzero_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API count_nonzero_dim_IntList { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::count_nonzero"; + static constexpr const char* overload_name = "dim_IntList"; + static constexpr const char* schema_str = "count_nonzero.dim_IntList(Tensor self, int[] dim) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim); +}; + +struct TORCH_API count_nonzero { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::count_nonzero"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "count_nonzero(Tensor self, int? dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dim); +}; + +struct TORCH_API count_nonzero_dim_IntList_out { + using schema = at::Tensor & (const at::Tensor &, at::IntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::count_nonzero"; + static constexpr const char* overload_name = "dim_IntList_out"; + static constexpr const char* schema_str = "count_nonzero.dim_IntList_out(Tensor self, int[] dim, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::IntArrayRef dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim, at::Tensor & out); +}; + +struct TORCH_API count_nonzero_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::count_nonzero"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "count_nonzero.out(Tensor self, int? dim=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cov_ops.h b/backends/pytorch/include/ATen/ops/cov_ops.h new file mode 100644 index 000000000..ead826c68 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cov_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cov { + using schema = at::Tensor (const at::Tensor &, int64_t, const ::std::optional &, const ::std::optional &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cov"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cov(Tensor self, *, int correction=1, Tensor? fweights=None, Tensor? aweights=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t correction, const ::std::optional & fweights, const ::std::optional & aweights); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t correction, const ::std::optional & fweights, const ::std::optional & aweights); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cross_ops.h b/backends/pytorch/include/ATen/ops/cross_ops.h new file mode 100644 index 000000000..fd531409d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cross_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cross_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cross"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cross.out(Tensor self, Tensor other, int? dim=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, ::std::optional dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, ::std::optional dim, at::Tensor & out); +}; + +struct TORCH_API cross { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cross"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cross(Tensor self, Tensor other, int? dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other, ::std::optional dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, ::std::optional dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/crow_indices_ops.h b/backends/pytorch/include/ATen/ops/crow_indices_ops.h new file mode 100644 index 000000000..0da380e04 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/crow_indices_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API crow_indices { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::crow_indices"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "crow_indices(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cummax_ops.h b/backends/pytorch/include/ATen/ops/cummax_ops.h new file mode 100644 index 000000000..921d2bc79 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cummax_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cummax { + using schema = ::std::tuple (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cummax"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cummax(Tensor self, int dim) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim); +}; + +struct TORCH_API cummax_out { + using schema = ::std::tuple (const at::Tensor &, int64_t, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cummax"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cummax.out(Tensor self, int dim, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API cummax_dimname { + using schema = ::std::tuple (const at::Tensor &, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cummax"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "cummax.dimname(Tensor self, Dimname dim) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim); +}; + +struct TORCH_API cummax_dimname_out { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cummax"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "cummax.dimname_out(Tensor self, Dimname dim, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, at::Tensor & values, at::Tensor & indices); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cummin_ops.h b/backends/pytorch/include/ATen/ops/cummin_ops.h new file mode 100644 index 000000000..904d5cddd --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cummin_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cummin { + using schema = ::std::tuple (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cummin"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cummin(Tensor self, int dim) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim); +}; + +struct TORCH_API cummin_out { + using schema = ::std::tuple (const at::Tensor &, int64_t, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cummin"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cummin.out(Tensor self, int dim, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API cummin_dimname { + using schema = ::std::tuple (const at::Tensor &, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cummin"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "cummin.dimname(Tensor self, Dimname dim) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim); +}; + +struct TORCH_API cummin_dimname_out { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cummin"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "cummin.dimname_out(Tensor self, Dimname dim, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, at::Tensor & values, at::Tensor & indices); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cumprod_ops.h b/backends/pytorch/include/ATen/ops/cumprod_ops.h new file mode 100644 index 000000000..4b0b32e9b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cumprod_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cumprod { + using schema = at::Tensor (const at::Tensor &, int64_t, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumprod"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cumprod(Tensor self, int dim, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, ::std::optional dtype); +}; + +struct TORCH_API cumprod_ { + using schema = at::Tensor & (at::Tensor &, int64_t, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumprod_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cumprod_(Tensor(a!) self, int dim, *, ScalarType? dtype=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, ::std::optional dtype); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, ::std::optional dtype); +}; + +struct TORCH_API cumprod_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumprod"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cumprod.out(Tensor self, int dim, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, ::std::optional dtype, at::Tensor & out); +}; + +struct TORCH_API cumprod_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumprod"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "cumprod.dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, ::std::optional dtype); +}; + +struct TORCH_API cumprod__dimname { + using schema = at::Tensor & (at::Tensor &, at::Dimname, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumprod_"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "cumprod_.dimname(Tensor(a!) self, Dimname dim, *, ScalarType? dtype=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, at::Dimname dim, ::std::optional dtype); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, at::Dimname dim, ::std::optional dtype); +}; + +struct TORCH_API cumprod_dimname_out { + using schema = at::Tensor & (const at::Tensor &, at::Dimname, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumprod"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "cumprod.dimname_out(Tensor self, Dimname dim, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Dimname dim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, ::std::optional dtype, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/cumsum_ops.h b/backends/pytorch/include/ATen/ops/cumsum_ops.h new file mode 100644 index 000000000..b57c53b11 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/cumsum_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API cumsum { + using schema = at::Tensor (const at::Tensor &, int64_t, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumsum"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cumsum(Tensor self, int dim, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, ::std::optional dtype); +}; + +struct TORCH_API cumsum_ { + using schema = at::Tensor & (at::Tensor &, int64_t, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumsum_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "cumsum_(Tensor(a!) self, int dim, *, ScalarType? dtype=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, ::std::optional dtype); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, ::std::optional dtype); +}; + +struct TORCH_API cumsum_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumsum"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "cumsum.out(Tensor self, int dim, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, ::std::optional dtype, at::Tensor & out); +}; + +struct TORCH_API cumsum_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumsum"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "cumsum.dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, ::std::optional dtype); +}; + +struct TORCH_API cumsum__dimname { + using schema = at::Tensor & (at::Tensor &, at::Dimname, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumsum_"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "cumsum_.dimname(Tensor(a!) self, Dimname dim, *, ScalarType? dtype=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, at::Dimname dim, ::std::optional dtype); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, at::Dimname dim, ::std::optional dtype); +}; + +struct TORCH_API cumsum_dimname_out { + using schema = at::Tensor & (const at::Tensor &, at::Dimname, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::cumsum"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "cumsum.dimname_out(Tensor self, Dimname dim, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Dimname dim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, ::std::optional dtype, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/data_ops.h b/backends/pytorch/include/ATen/ops/data_ops.h new file mode 100644 index 000000000..2e0fce2f1 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/data_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API data { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::data"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "data(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/deg2rad_ops.h b/backends/pytorch/include/ATen/ops/deg2rad_ops.h new file mode 100644 index 000000000..2c62b9979 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/deg2rad_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API deg2rad { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::deg2rad"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "deg2rad(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API deg2rad_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::deg2rad_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "deg2rad_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API deg2rad_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::deg2rad"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "deg2rad.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/dense_dim_ops.h b/backends/pytorch/include/ATen/ops/dense_dim_ops.h new file mode 100644 index 000000000..b38ecfc12 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/dense_dim_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API dense_dim { + using schema = int64_t (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dense_dim"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "dense_dim(Tensor self) -> int"; + static int64_t call(const at::Tensor & self); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/dequantize_ops.h b/backends/pytorch/include/ATen/ops/dequantize_ops.h new file mode 100644 index 000000000..1fda5ebe6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/dequantize_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API dequantize_self { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dequantize"; + static constexpr const char* overload_name = "self"; + static constexpr const char* schema_str = "dequantize.self(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API dequantize_tensors { + using schema = ::std::vector (at::TensorList); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dequantize"; + static constexpr const char* overload_name = "tensors"; + static constexpr const char* schema_str = "dequantize.tensors(Tensor[] tensors) -> Tensor[]"; + static ::std::vector call(at::TensorList tensors); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, at::TensorList tensors); +}; + +struct TORCH_API dequantize_self_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dequantize"; + static constexpr const char* overload_name = "self_out"; + static constexpr const char* schema_str = "dequantize.self_out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API dequantize_tensors_out { + using schema = void (at::TensorList, at::TensorList); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dequantize"; + static constexpr const char* overload_name = "tensors_out"; + static constexpr const char* schema_str = "dequantize.tensors_out(Tensor[] tensors, *, Tensor(a!)[] out) -> ()"; + static void call(at::TensorList tensors, at::TensorList out); + static void redispatch(c10::DispatchKeySet dispatchKeySet, at::TensorList tensors, at::TensorList out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/det_ops.h b/backends/pytorch/include/ATen/ops/det_ops.h new file mode 100644 index 000000000..d77132450 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/det_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API det { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::det"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "det(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/detach_ops.h b/backends/pytorch/include/ATen/ops/detach_ops.h new file mode 100644 index 000000000..f7be889ca --- /dev/null +++ b/backends/pytorch/include/ATen/ops/detach_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API detach { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::detach"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "detach(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API detach_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::detach_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "detach_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/diag_embed_ops.h b/backends/pytorch/include/ATen/ops/diag_embed_ops.h new file mode 100644 index 000000000..10ce1be5c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/diag_embed_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API diag_embed { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diag_embed"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "diag_embed(Tensor self, int offset=0, int dim1=-2, int dim2=-1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t offset, int64_t dim1, int64_t dim2); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t offset, int64_t dim1, int64_t dim2); +}; + +struct TORCH_API diag_embed_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, int64_t, int64_t, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diag_embed"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "diag_embed.out(Tensor self, int offset=0, int dim1=-2, int dim2=-1, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t offset, int64_t dim1, int64_t dim2, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t offset, int64_t dim1, int64_t dim2, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/diag_ops.h b/backends/pytorch/include/ATen/ops/diag_ops.h new file mode 100644 index 000000000..1c65b1268 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/diag_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API diag_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diag"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "diag.out(Tensor self, int diagonal=0, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t diagonal, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t diagonal, at::Tensor & out); +}; + +struct TORCH_API diag { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diag"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "diag(Tensor self, int diagonal=0) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t diagonal); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t diagonal); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/diagflat_ops.h b/backends/pytorch/include/ATen/ops/diagflat_ops.h new file mode 100644 index 000000000..22c6111b5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/diagflat_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API diagflat { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diagflat"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "diagflat(Tensor self, int offset=0) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t offset); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t offset); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/diagonal_ops.h b/backends/pytorch/include/ATen/ops/diagonal_ops.h new file mode 100644 index 000000000..533a4230c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/diagonal_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API diagonal { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diagonal"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "diagonal(Tensor(a) self, int offset=0, int dim1=0, int dim2=1) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t offset, int64_t dim1, int64_t dim2); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t offset, int64_t dim1, int64_t dim2); +}; + +struct TORCH_API diagonal_Dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, at::Dimname, at::Dimname, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diagonal"; + static constexpr const char* overload_name = "Dimname"; + static constexpr const char* schema_str = "diagonal.Dimname(Tensor(a) self, *, Dimname outdim, Dimname dim1, Dimname dim2, int offset=0) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::Dimname outdim, at::Dimname dim1, at::Dimname dim2, int64_t offset); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname outdim, at::Dimname dim1, at::Dimname dim2, int64_t offset); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/diagonal_scatter_ops.h b/backends/pytorch/include/ATen/ops/diagonal_scatter_ops.h new file mode 100644 index 000000000..53dbeb613 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/diagonal_scatter_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API diagonal_scatter { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, int64_t, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diagonal_scatter"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "diagonal_scatter(Tensor self, Tensor src, int offset=0, int dim1=0, int dim2=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & src, int64_t offset, int64_t dim1, int64_t dim2); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, int64_t offset, int64_t dim1, int64_t dim2); +}; + +struct TORCH_API diagonal_scatter_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, int64_t, int64_t, int64_t, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diagonal_scatter"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "diagonal_scatter.out(Tensor self, Tensor src, int offset=0, int dim1=0, int dim2=1, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & src, int64_t offset, int64_t dim1, int64_t dim2, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, int64_t offset, int64_t dim1, int64_t dim2, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/diff_ops.h b/backends/pytorch/include/ATen/ops/diff_ops.h new file mode 100644 index 000000000..8842ffe3e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/diff_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API diff { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t, const ::std::optional &, const ::std::optional &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diff"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "diff(Tensor self, int n=1, int dim=-1, Tensor? prepend=None, Tensor? append=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t n, int64_t dim, const ::std::optional & prepend, const ::std::optional & append); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t n, int64_t dim, const ::std::optional & prepend, const ::std::optional & append); +}; + +struct TORCH_API diff_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, int64_t, const ::std::optional &, const ::std::optional &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::diff"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "diff.out(Tensor self, int n=1, int dim=-1, Tensor? prepend=None, Tensor? append=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t n, int64_t dim, const ::std::optional & prepend, const ::std::optional & append, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t n, int64_t dim, const ::std::optional & prepend, const ::std::optional & append, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/digamma_ops.h b/backends/pytorch/include/ATen/ops/digamma_ops.h new file mode 100644 index 000000000..69a2a0bd0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/digamma_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API digamma_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::digamma_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "digamma_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API digamma_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::digamma"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "digamma.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API digamma { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::digamma"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "digamma(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/dist_ops.h b/backends/pytorch/include/ATen/ops/dist_ops.h new file mode 100644 index 000000000..816b1d8d5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/dist_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API dist { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dist"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "dist(Tensor self, Tensor other, Scalar p=2) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other, const at::Scalar & p); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, const at::Scalar & p); +}; + +struct TORCH_API dist_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dist"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "dist.out(Tensor self, Tensor other, Scalar p=2, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, const at::Scalar & p, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, const at::Scalar & p, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/div_ops.h b/backends/pytorch/include/ATen/ops/div_ops.h new file mode 100644 index 000000000..88e18449b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/div_ops.h @@ -0,0 +1,155 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API div_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "div.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API div__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "div_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API div_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "div.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API div_Tensor_mode { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div"; + static constexpr const char* overload_name = "Tensor_mode"; + static constexpr const char* schema_str = "div.Tensor_mode(Tensor self, Tensor other, *, str? rounding_mode) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode); +}; + +struct TORCH_API div__Tensor_mode { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div_"; + static constexpr const char* overload_name = "Tensor_mode"; + static constexpr const char* schema_str = "div_.Tensor_mode(Tensor(a!) self, Tensor other, *, str? rounding_mode) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode); +}; + +struct TORCH_API div_out_mode { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div"; + static constexpr const char* overload_name = "out_mode"; + static constexpr const char* schema_str = "div.out_mode(Tensor self, Tensor other, *, str? rounding_mode, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode, at::Tensor & out); +}; + +struct TORCH_API div_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "div.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API div__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "div_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API div_Scalar_mode { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div"; + static constexpr const char* overload_name = "Scalar_mode"; + static constexpr const char* schema_str = "div.Scalar_mode(Tensor self, Scalar other, *, str? rounding_mode) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other, ::std::optional rounding_mode); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, ::std::optional rounding_mode); +}; + +struct TORCH_API div__Scalar_mode { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div_"; + static constexpr const char* overload_name = "Scalar_mode"; + static constexpr const char* schema_str = "div_.Scalar_mode(Tensor(a!) self, Scalar other, *, str? rounding_mode) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other, ::std::optional rounding_mode); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other, ::std::optional rounding_mode); +}; + +struct TORCH_API div_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "div.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API div_Scalar_mode_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::div"; + static constexpr const char* overload_name = "Scalar_mode_out"; + static constexpr const char* schema_str = "div.Scalar_mode_out(Tensor self, Scalar other, *, str? rounding_mode, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, ::std::optional rounding_mode, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, ::std::optional rounding_mode, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/divide_ops.h b/backends/pytorch/include/ATen/ops/divide_ops.h new file mode 100644 index 000000000..b92af7e3e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/divide_ops.h @@ -0,0 +1,133 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API divide_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::divide"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "divide.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API divide__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::divide_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "divide_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API divide_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::divide"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "divide.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API divide_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::divide"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "divide.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API divide__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::divide_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "divide_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API divide_Tensor_mode { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::divide"; + static constexpr const char* overload_name = "Tensor_mode"; + static constexpr const char* schema_str = "divide.Tensor_mode(Tensor self, Tensor other, *, str? rounding_mode) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode); +}; + +struct TORCH_API divide__Tensor_mode { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::divide_"; + static constexpr const char* overload_name = "Tensor_mode"; + static constexpr const char* schema_str = "divide_.Tensor_mode(Tensor(a!) self, Tensor other, *, str? rounding_mode) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode); +}; + +struct TORCH_API divide_out_mode { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::divide"; + static constexpr const char* overload_name = "out_mode"; + static constexpr const char* schema_str = "divide.out_mode(Tensor self, Tensor other, *, str? rounding_mode, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, ::std::optional rounding_mode, at::Tensor & out); +}; + +struct TORCH_API divide_Scalar_mode { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::divide"; + static constexpr const char* overload_name = "Scalar_mode"; + static constexpr const char* schema_str = "divide.Scalar_mode(Tensor self, Scalar other, *, str? rounding_mode) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other, ::std::optional rounding_mode); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, ::std::optional rounding_mode); +}; + +struct TORCH_API divide__Scalar_mode { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::divide_"; + static constexpr const char* overload_name = "Scalar_mode"; + static constexpr const char* schema_str = "divide_.Scalar_mode(Tensor(a!) self, Scalar other, *, str? rounding_mode) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other, ::std::optional rounding_mode); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other, ::std::optional rounding_mode); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/dot_ops.h b/backends/pytorch/include/ATen/ops/dot_ops.h new file mode 100644 index 000000000..4fd0999cd --- /dev/null +++ b/backends/pytorch/include/ATen/ops/dot_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API dot { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dot"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "dot(Tensor self, Tensor tensor) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & tensor); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & tensor); +}; + +struct TORCH_API dot_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dot"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "dot.out(Tensor self, Tensor tensor, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & tensor, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & tensor, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/dsplit_ops.h b/backends/pytorch/include/ATen/ops/dsplit_ops.h new file mode 100644 index 000000000..b6573be81 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/dsplit_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API dsplit_int { + using schema = ::std::vector (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dsplit"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "dsplit.int(Tensor(a -> *) self, int sections) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, int64_t sections); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t sections); +}; + +struct TORCH_API dsplit_array { + using schema = ::std::vector (const at::Tensor &, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::dsplit"; + static constexpr const char* overload_name = "array"; + static constexpr const char* schema_str = "dsplit.array(Tensor(a -> *) self, int[] indices) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, at::IntArrayRef indices); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef indices); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/eq_ops.h b/backends/pytorch/include/ATen/ops/eq_ops.h new file mode 100644 index 000000000..0088635bc --- /dev/null +++ b/backends/pytorch/include/ATen/ops/eq_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API eq__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::eq_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "eq_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API eq__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::eq_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "eq_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API eq_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::eq"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "eq.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API eq_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::eq"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "eq.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API eq_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::eq"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "eq.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API eq_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::eq"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "eq.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/equal_ops.h b/backends/pytorch/include/ATen/ops/equal_ops.h new file mode 100644 index 000000000..1dea09cc9 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/equal_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API equal { + using schema = bool (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::equal"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "equal(Tensor self, Tensor other) -> bool"; + static bool call(const at::Tensor & self, const at::Tensor & other); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/erf_ops.h b/backends/pytorch/include/ATen/ops/erf_ops.h new file mode 100644 index 000000000..9f5d736ad --- /dev/null +++ b/backends/pytorch/include/ATen/ops/erf_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API erf { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::erf"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "erf(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API erf_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::erf_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "erf_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API erf_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::erf"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "erf.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/erfc_ops.h b/backends/pytorch/include/ATen/ops/erfc_ops.h new file mode 100644 index 000000000..f0638b6f0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/erfc_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API erfc { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::erfc"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "erfc(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API erfc_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::erfc_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "erfc_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API erfc_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::erfc"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "erfc.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/erfinv_ops.h b/backends/pytorch/include/ATen/ops/erfinv_ops.h new file mode 100644 index 000000000..bfef84efa --- /dev/null +++ b/backends/pytorch/include/ATen/ops/erfinv_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API erfinv { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::erfinv"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "erfinv(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API erfinv_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::erfinv_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "erfinv_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API erfinv_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::erfinv"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "erfinv.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/exp2_ops.h b/backends/pytorch/include/ATen/ops/exp2_ops.h new file mode 100644 index 000000000..a343cd621 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/exp2_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API exp2 { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::exp2"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "exp2(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API exp2_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::exp2_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "exp2_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API exp2_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::exp2"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "exp2.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/exp_ops.h b/backends/pytorch/include/ATen/ops/exp_ops.h new file mode 100644 index 000000000..a7bab235d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/exp_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API exp { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::exp"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "exp(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API exp_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::exp_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "exp_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API exp_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::exp"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "exp.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/expand_as_ops.h b/backends/pytorch/include/ATen/ops/expand_as_ops.h new file mode 100644 index 000000000..e52465792 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/expand_as_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API expand_as { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::expand_as"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "expand_as(Tensor(a) self, Tensor other) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/expand_ops.h b/backends/pytorch/include/ATen/ops/expand_ops.h new file mode 100644 index 000000000..6f97b62d3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/expand_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API expand { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::expand"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "expand(Tensor(a) self, SymInt[] size, *, bool implicit=False) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size, bool implicit); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, bool implicit); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/expm1_ops.h b/backends/pytorch/include/ATen/ops/expm1_ops.h new file mode 100644 index 000000000..60b20ccc6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/expm1_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API expm1 { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::expm1"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "expm1(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API expm1_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::expm1_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "expm1_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API expm1_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::expm1"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "expm1.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/exponential_ops.h b/backends/pytorch/include/ATen/ops/exponential_ops.h new file mode 100644 index 000000000..2e15072f3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/exponential_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API exponential_ { + using schema = at::Tensor & (at::Tensor &, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::exponential_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "exponential_(Tensor(a!) self, float lambd=1, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, double lambd, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, double lambd, ::std::optional generator); +}; + +struct TORCH_API exponential_out { + using schema = at::Tensor & (const at::Tensor &, double, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::exponential"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "exponential.out(Tensor self, float lambd=1, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, double lambd, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double lambd, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API exponential { + using schema = at::Tensor (const at::Tensor &, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::exponential"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "exponential(Tensor self, float lambd=1, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, double lambd, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double lambd, ::std::optional generator); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/fill_diagonal_ops.h b/backends/pytorch/include/ATen/ops/fill_diagonal_ops.h new file mode 100644 index 000000000..757a579e5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/fill_diagonal_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API fill_diagonal_ { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fill_diagonal_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "fill_diagonal_(Tensor(a!) self, Scalar fill_value, bool wrap=False) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & fill_value, bool wrap); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & fill_value, bool wrap); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/fill_ops.h b/backends/pytorch/include/ATen/ops/fill_ops.h new file mode 100644 index 000000000..2a178007a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/fill_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API fill_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fill"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "fill.Scalar(Tensor self, Scalar value) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & value); +}; + +struct TORCH_API fill_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fill"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "fill.Tensor(Tensor self, Tensor value) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & value); +}; + +struct TORCH_API fill__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fill_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "fill_.Scalar(Tensor(a!) self, Scalar value) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & value); +}; + +struct TORCH_API fill__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fill_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "fill_.Tensor(Tensor(a!) self, Tensor value) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & value); +}; + +struct TORCH_API fill_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fill"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "fill.Scalar_out(Tensor self, Scalar value, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & value, at::Tensor & out); +}; + +struct TORCH_API fill_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fill"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "fill.Tensor_out(Tensor self, Tensor value, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & value, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/fix_ops.h b/backends/pytorch/include/ATen/ops/fix_ops.h new file mode 100644 index 000000000..0c990a40d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/fix_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API fix { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fix"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "fix(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API fix_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fix_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "fix_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API fix_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fix"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "fix.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/flatten_ops.h b/backends/pytorch/include/ATen/ops/flatten_ops.h new file mode 100644 index 000000000..736074937 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/flatten_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API flatten_using_ints { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::flatten"; + static constexpr const char* overload_name = "using_ints"; + static constexpr const char* schema_str = "flatten.using_ints(Tensor(a) self, int start_dim=0, int end_dim=-1) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t start_dim, int64_t end_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t start_dim, int64_t end_dim); +}; + +struct TORCH_API flatten_named_out_dim { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::flatten"; + static constexpr const char* overload_name = "named_out_dim"; + static constexpr const char* schema_str = "flatten.named_out_dim(Tensor(a) self, int start_dim, int end_dim, Dimname out_dim) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t start_dim, int64_t end_dim, at::Dimname out_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t start_dim, int64_t end_dim, at::Dimname out_dim); +}; + +struct TORCH_API flatten_using_names { + using schema = at::Tensor (const at::Tensor &, at::Dimname, at::Dimname, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::flatten"; + static constexpr const char* overload_name = "using_names"; + static constexpr const char* schema_str = "flatten.using_names(Tensor(a) self, Dimname start_dim, Dimname end_dim, Dimname out_dim) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::Dimname start_dim, at::Dimname end_dim, at::Dimname out_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname start_dim, at::Dimname end_dim, at::Dimname out_dim); +}; + +struct TORCH_API flatten_DimnameList { + using schema = at::Tensor (const at::Tensor &, at::DimnameList, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::flatten"; + static constexpr const char* overload_name = "DimnameList"; + static constexpr const char* schema_str = "flatten.DimnameList(Tensor(a) self, Dimname[] dims, Dimname out_dim) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::DimnameList dims, at::Dimname out_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dims, at::Dimname out_dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/flip_ops.h b/backends/pytorch/include/ATen/ops/flip_ops.h new file mode 100644 index 000000000..c40d48103 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/flip_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API flip { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::flip"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "flip(Tensor self, int[] dims) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef dims); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dims); +}; + +struct TORCH_API flip_out { + using schema = at::Tensor & (const at::Tensor &, at::IntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::flip"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "flip.out(Tensor self, int[] dims, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::IntArrayRef dims, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dims, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/fliplr_ops.h b/backends/pytorch/include/ATen/ops/fliplr_ops.h new file mode 100644 index 000000000..9996687ec --- /dev/null +++ b/backends/pytorch/include/ATen/ops/fliplr_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API fliplr { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fliplr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "fliplr(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/flipud_ops.h b/backends/pytorch/include/ATen/ops/flipud_ops.h new file mode 100644 index 000000000..029b522d1 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/flipud_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API flipud { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::flipud"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "flipud(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/float_power_ops.h b/backends/pytorch/include/ATen/ops/float_power_ops.h new file mode 100644 index 000000000..219150037 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/float_power_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API float_power_Tensor_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::float_power"; + static constexpr const char* overload_name = "Tensor_Tensor_out"; + static constexpr const char* schema_str = "float_power.Tensor_Tensor_out(Tensor self, Tensor exponent, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & exponent, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & exponent, at::Tensor & out); +}; + +struct TORCH_API float_power_Tensor_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::float_power"; + static constexpr const char* overload_name = "Tensor_Tensor"; + static constexpr const char* schema_str = "float_power.Tensor_Tensor(Tensor self, Tensor exponent) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & exponent); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & exponent); +}; + +struct TORCH_API float_power_Scalar_out { + using schema = at::Tensor & (const at::Scalar &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::float_power"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "float_power.Scalar_out(Scalar self, Tensor exponent, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Scalar & self, const at::Tensor & exponent, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & exponent, at::Tensor & out); +}; + +struct TORCH_API float_power_Scalar { + using schema = at::Tensor (const at::Scalar &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::float_power"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "float_power.Scalar(Scalar self, Tensor exponent) -> Tensor"; + static at::Tensor call(const at::Scalar & self, const at::Tensor & exponent); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & exponent); +}; + +struct TORCH_API float_power_Tensor_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::float_power"; + static constexpr const char* overload_name = "Tensor_Scalar_out"; + static constexpr const char* schema_str = "float_power.Tensor_Scalar_out(Tensor self, Scalar exponent, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & exponent, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & exponent, at::Tensor & out); +}; + +struct TORCH_API float_power_Tensor_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::float_power"; + static constexpr const char* overload_name = "Tensor_Scalar"; + static constexpr const char* schema_str = "float_power.Tensor_Scalar(Tensor self, Scalar exponent) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & exponent); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & exponent); +}; + +struct TORCH_API float_power__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::float_power_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "float_power_.Scalar(Tensor(a!) self, Scalar exponent) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & exponent); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & exponent); +}; + +struct TORCH_API float_power__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::float_power_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "float_power_.Tensor(Tensor(a!) self, Tensor exponent) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & exponent); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & exponent); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/floor_divide_ops.h b/backends/pytorch/include/ATen/ops/floor_divide_ops.h new file mode 100644 index 000000000..c0220f4f5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/floor_divide_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API floor_divide { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::floor_divide"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "floor_divide(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API floor_divide__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::floor_divide_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "floor_divide_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API floor_divide_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::floor_divide"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "floor_divide.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API floor_divide_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::floor_divide"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "floor_divide.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API floor_divide__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::floor_divide_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "floor_divide_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API floor_divide_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::floor_divide"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "floor_divide.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/floor_ops.h b/backends/pytorch/include/ATen/ops/floor_ops.h new file mode 100644 index 000000000..79f59ba05 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/floor_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API floor { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::floor"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "floor(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API floor_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::floor_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "floor_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API floor_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::floor"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "floor.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/fmax_ops.h b/backends/pytorch/include/ATen/ops/fmax_ops.h new file mode 100644 index 000000000..bcdb52884 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/fmax_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API fmax { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fmax"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "fmax(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API fmax_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fmax"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "fmax.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/fmin_ops.h b/backends/pytorch/include/ATen/ops/fmin_ops.h new file mode 100644 index 000000000..ef06696f5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/fmin_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API fmin { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fmin"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "fmin(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API fmin_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fmin"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "fmin.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/fmod_ops.h b/backends/pytorch/include/ATen/ops/fmod_ops.h new file mode 100644 index 000000000..5bd71436e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/fmod_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API fmod_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fmod"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "fmod.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API fmod_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fmod"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "fmod.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API fmod__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fmod_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "fmod_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API fmod_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fmod"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "fmod.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API fmod_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fmod"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "fmod.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API fmod__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::fmod_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "fmod_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/frac_ops.h b/backends/pytorch/include/ATen/ops/frac_ops.h new file mode 100644 index 000000000..73e47f448 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/frac_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API frac { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::frac"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "frac(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API frac_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::frac_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "frac_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API frac_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::frac"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "frac.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/frexp_ops.h b/backends/pytorch/include/ATen/ops/frexp_ops.h new file mode 100644 index 000000000..d93cda66b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/frexp_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API frexp_Tensor { + using schema = ::std::tuple (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::frexp"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "frexp.Tensor(Tensor self) -> (Tensor mantissa, Tensor exponent)"; + static ::std::tuple call(const at::Tensor & self); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API frexp_Tensor_out { + using schema = ::std::tuple (const at::Tensor &, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::frexp"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "frexp.Tensor_out(Tensor self, *, Tensor(a!) mantissa, Tensor(b!) exponent) -> (Tensor(a!) mantissa, Tensor(b!) exponent)"; + static ::std::tuple call(const at::Tensor & self, at::Tensor & mantissa, at::Tensor & exponent); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & mantissa, at::Tensor & exponent); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/gather_ops.h b/backends/pytorch/include/ATen/ops/gather_ops.h new file mode 100644 index 000000000..02ce2c00f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/gather_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API gather_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gather"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "gather.out(Tensor self, int dim, Tensor index, *, bool sparse_grad=False, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, bool sparse_grad, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, bool sparse_grad, at::Tensor & out); +}; + +struct TORCH_API gather { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gather"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "gather(Tensor self, int dim, Tensor index, *, bool sparse_grad=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, bool sparse_grad); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, bool sparse_grad); +}; + +struct TORCH_API gather_dimname_out { + using schema = at::Tensor & (const at::Tensor &, at::Dimname, const at::Tensor &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gather"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "gather.dimname_out(Tensor self, Dimname dim, Tensor index, *, bool sparse_grad=False, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index, bool sparse_grad, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index, bool sparse_grad, at::Tensor & out); +}; + +struct TORCH_API gather_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gather"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "gather.dimname(Tensor self, Dimname dim, Tensor index, *, bool sparse_grad=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index, bool sparse_grad); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index, bool sparse_grad); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/gcd_ops.h b/backends/pytorch/include/ATen/ops/gcd_ops.h new file mode 100644 index 000000000..e77c8557e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/gcd_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API gcd_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gcd"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "gcd.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API gcd { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gcd"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "gcd(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API gcd_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gcd_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "gcd_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/ge_ops.h b/backends/pytorch/include/ATen/ops/ge_ops.h new file mode 100644 index 000000000..d72c47f75 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/ge_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API ge_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ge"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "ge.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API ge_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ge"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "ge.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API ge_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ge"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "ge.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API ge_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ge"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "ge.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API ge__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ge_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "ge_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API ge__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ge_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "ge_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/geometric_ops.h b/backends/pytorch/include/ATen/ops/geometric_ops.h new file mode 100644 index 000000000..0540b61d1 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/geometric_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API geometric_ { + using schema = at::Tensor & (at::Tensor &, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::geometric_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "geometric_(Tensor(a!) self, float p, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, double p, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, double p, ::std::optional generator); +}; + +struct TORCH_API geometric_out { + using schema = at::Tensor & (const at::Tensor &, double, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::geometric"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "geometric.out(Tensor self, float p, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, double p, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double p, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API geometric { + using schema = at::Tensor (const at::Tensor &, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::geometric"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "geometric(Tensor self, float p, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, double p, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double p, ::std::optional generator); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/geqrf_ops.h b/backends/pytorch/include/ATen/ops/geqrf_ops.h new file mode 100644 index 000000000..5d83fd58a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/geqrf_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API geqrf_a { + using schema = ::std::tuple (const at::Tensor &, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::geqrf"; + static constexpr const char* overload_name = "a"; + static constexpr const char* schema_str = "geqrf.a(Tensor self, *, Tensor(a!) a, Tensor(b!) tau) -> (Tensor(a!) a, Tensor(b!) tau)"; + static ::std::tuple call(const at::Tensor & self, at::Tensor & a, at::Tensor & tau); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & a, at::Tensor & tau); +}; + +struct TORCH_API geqrf { + using schema = ::std::tuple (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::geqrf"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "geqrf(Tensor self) -> (Tensor a, Tensor tau)"; + static ::std::tuple call(const at::Tensor & self); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/ger_ops.h b/backends/pytorch/include/ATen/ops/ger_ops.h new file mode 100644 index 000000000..2b02a7b9b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/ger_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API ger { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ger"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "ger(Tensor self, Tensor vec2) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & vec2); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & vec2); +}; + +struct TORCH_API ger_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ger"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "ger.out(Tensor self, Tensor vec2, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & vec2, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & vec2, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/greater_equal_ops.h b/backends/pytorch/include/ATen/ops/greater_equal_ops.h new file mode 100644 index 000000000..739164459 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/greater_equal_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API greater_equal_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater_equal"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "greater_equal.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API greater_equal_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater_equal"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "greater_equal.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API greater_equal_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater_equal"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "greater_equal.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API greater_equal_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater_equal"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "greater_equal.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API greater_equal__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater_equal_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "greater_equal_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API greater_equal__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater_equal_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "greater_equal_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/greater_ops.h b/backends/pytorch/include/ATen/ops/greater_ops.h new file mode 100644 index 000000000..b1c4de7df --- /dev/null +++ b/backends/pytorch/include/ATen/ops/greater_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API greater_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "greater.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API greater_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "greater.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API greater_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "greater.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API greater_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "greater.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API greater__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "greater_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API greater__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::greater_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "greater_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/gt_ops.h b/backends/pytorch/include/ATen/ops/gt_ops.h new file mode 100644 index 000000000..c4499fba0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/gt_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API gt_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gt"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "gt.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API gt_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gt"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "gt.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API gt_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gt"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "gt.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API gt_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gt"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "gt.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API gt__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gt_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "gt_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API gt__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::gt_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "gt_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/hardshrink_backward_ops.h b/backends/pytorch/include/ATen/ops/hardshrink_backward_ops.h new file mode 100644 index 000000000..4570853b5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/hardshrink_backward_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API hardshrink_backward_grad_input { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hardshrink_backward"; + static constexpr const char* overload_name = "grad_input"; + static constexpr const char* schema_str = "hardshrink_backward.grad_input(Tensor grad_out, Tensor self, Scalar lambd, *, Tensor(a!) grad_input) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & grad_out, const at::Tensor & self, const at::Scalar & lambd, at::Tensor & grad_input); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & grad_out, const at::Tensor & self, const at::Scalar & lambd, at::Tensor & grad_input); +}; + +struct TORCH_API hardshrink_backward { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hardshrink_backward"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "hardshrink_backward(Tensor grad_out, Tensor self, Scalar lambd) -> Tensor"; + static at::Tensor call(const at::Tensor & grad_out, const at::Tensor & self, const at::Scalar & lambd); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & grad_out, const at::Tensor & self, const at::Scalar & lambd); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/hardshrink_ops.h b/backends/pytorch/include/ATen/ops/hardshrink_ops.h new file mode 100644 index 000000000..89d7e97dc --- /dev/null +++ b/backends/pytorch/include/ATen/ops/hardshrink_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API hardshrink_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hardshrink"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "hardshrink.out(Tensor self, Scalar lambd=0.5, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & lambd, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & lambd, at::Tensor & out); +}; + +struct TORCH_API hardshrink { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hardshrink"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "hardshrink(Tensor self, Scalar lambd=0.5) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & lambd); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & lambd); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/hash_tensor_ops.h b/backends/pytorch/include/ATen/ops/hash_tensor_ops.h new file mode 100644 index 000000000..ad770ae3a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/hash_tensor_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API hash_tensor { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, bool, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hash_tensor"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "hash_tensor(Tensor self, int[1] dim=[], *, bool keepdim=False, int mode=0) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef dim, bool keepdim, int64_t mode); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim, bool keepdim, int64_t mode); +}; + +struct TORCH_API hash_tensor_out { + using schema = at::Tensor & (const at::Tensor &, at::IntArrayRef, bool, int64_t, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hash_tensor"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "hash_tensor.out(Tensor self, int[1] dim=[], *, bool keepdim=False, int mode=0, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::IntArrayRef dim, bool keepdim, int64_t mode, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim, bool keepdim, int64_t mode, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/heaviside_ops.h b/backends/pytorch/include/ATen/ops/heaviside_ops.h new file mode 100644 index 000000000..2d1ad993f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/heaviside_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API heaviside_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::heaviside"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "heaviside.out(Tensor self, Tensor values, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & values, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & values, at::Tensor & out); +}; + +struct TORCH_API heaviside { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::heaviside"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "heaviside(Tensor self, Tensor values) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & values); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & values); +}; + +struct TORCH_API heaviside_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::heaviside_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "heaviside_(Tensor(a!) self, Tensor values) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & values); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & values); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/histc_ops.h b/backends/pytorch/include/ATen/ops/histc_ops.h new file mode 100644 index 000000000..ee5a6d80d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/histc_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API histc_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::histc"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "histc.out(Tensor self, int bins=100, Scalar min=0, Scalar max=0, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t bins, const at::Scalar & min, const at::Scalar & max, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t bins, const at::Scalar & min, const at::Scalar & max, at::Tensor & out); +}; + +struct TORCH_API histc { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::histc"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "histc(Tensor self, int bins=100, Scalar min=0, Scalar max=0) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t bins, const at::Scalar & min, const at::Scalar & max); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t bins, const at::Scalar & min, const at::Scalar & max); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/histogram_ops.h b/backends/pytorch/include/ATen/ops/histogram_ops.h new file mode 100644 index 000000000..38e84158e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/histogram_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API histogram_bins_tensor_out { + using schema = ::std::tuple (const at::Tensor &, const at::Tensor &, const ::std::optional &, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::histogram"; + static constexpr const char* overload_name = "bins_tensor_out"; + static constexpr const char* schema_str = "histogram.bins_tensor_out(Tensor self, Tensor bins, *, Tensor? weight=None, bool density=False, Tensor(a!) hist, Tensor(b!) bin_edges) -> (Tensor(a!) hist, Tensor(b!) bin_edges)"; + static ::std::tuple call(const at::Tensor & self, const at::Tensor & bins, const ::std::optional & weight, bool density, at::Tensor & hist, at::Tensor & bin_edges); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & bins, const ::std::optional & weight, bool density, at::Tensor & hist, at::Tensor & bin_edges); +}; + +struct TORCH_API histogram_bins_tensor { + using schema = ::std::tuple (const at::Tensor &, const at::Tensor &, const ::std::optional &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::histogram"; + static constexpr const char* overload_name = "bins_tensor"; + static constexpr const char* schema_str = "histogram.bins_tensor(Tensor self, Tensor bins, *, Tensor? weight=None, bool density=False) -> (Tensor hist, Tensor bin_edges)"; + static ::std::tuple call(const at::Tensor & self, const at::Tensor & bins, const ::std::optional & weight, bool density); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & bins, const ::std::optional & weight, bool density); +}; + +struct TORCH_API histogram_bin_ct_out { + using schema = ::std::tuple (const at::Tensor &, int64_t, ::std::optional>, const ::std::optional &, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::histogram"; + static constexpr const char* overload_name = "bin_ct_out"; + static constexpr const char* schema_str = "histogram.bin_ct_out(Tensor self, int bins=100, *, float[]? range=None, Tensor? weight=None, bool density=False, Tensor(a!) hist, Tensor(b!) bin_edges) -> (Tensor(a!) hist, Tensor(b!) bin_edges)"; + static ::std::tuple call(const at::Tensor & self, int64_t bins, ::std::optional> range, const ::std::optional & weight, bool density, at::Tensor & hist, at::Tensor & bin_edges); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t bins, ::std::optional> range, const ::std::optional & weight, bool density, at::Tensor & hist, at::Tensor & bin_edges); +}; + +struct TORCH_API histogram_bin_ct { + using schema = ::std::tuple (const at::Tensor &, int64_t, ::std::optional>, const ::std::optional &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::histogram"; + static constexpr const char* overload_name = "bin_ct"; + static constexpr const char* schema_str = "histogram.bin_ct(Tensor self, int bins=100, *, float[]? range=None, Tensor? weight=None, bool density=False) -> (Tensor hist, Tensor bin_edges)"; + static ::std::tuple call(const at::Tensor & self, int64_t bins, ::std::optional> range, const ::std::optional & weight, bool density); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t bins, ::std::optional> range, const ::std::optional & weight, bool density); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/hsplit_ops.h b/backends/pytorch/include/ATen/ops/hsplit_ops.h new file mode 100644 index 000000000..2a209e5d6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/hsplit_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API hsplit_int { + using schema = ::std::vector (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hsplit"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "hsplit.int(Tensor(a -> *) self, int sections) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, int64_t sections); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t sections); +}; + +struct TORCH_API hsplit_array { + using schema = ::std::vector (const at::Tensor &, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hsplit"; + static constexpr const char* overload_name = "array"; + static constexpr const char* schema_str = "hsplit.array(Tensor(a -> *) self, int[] indices) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, at::IntArrayRef indices); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef indices); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/hypot_ops.h b/backends/pytorch/include/ATen/ops/hypot_ops.h new file mode 100644 index 000000000..406a0a28c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/hypot_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API hypot_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hypot"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "hypot.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API hypot { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hypot"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "hypot(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API hypot_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::hypot_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "hypot_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/i0_ops.h b/backends/pytorch/include/ATen/ops/i0_ops.h new file mode 100644 index 000000000..fb580ba0e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/i0_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API i0 { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::i0"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "i0(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API i0_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::i0_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "i0_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API i0_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::i0"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "i0.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/igamma_ops.h b/backends/pytorch/include/ATen/ops/igamma_ops.h new file mode 100644 index 000000000..353cc1aa2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/igamma_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API igamma_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::igamma"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "igamma.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API igamma { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::igamma"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "igamma(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API igamma_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::igamma_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "igamma_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/igammac_ops.h b/backends/pytorch/include/ATen/ops/igammac_ops.h new file mode 100644 index 000000000..cf0086675 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/igammac_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API igammac_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::igammac"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "igammac.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API igammac { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::igammac"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "igammac(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API igammac_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::igammac_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "igammac_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/index_add_ops.h b/backends/pytorch/include/ATen/ops/index_add_ops.h new file mode 100644 index 000000000..10b2b0fe3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/index_add_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API index_add_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_add"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "index_add.out(Tensor self, int dim, Tensor index, Tensor source, *, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha, at::Tensor & out); +}; + +struct TORCH_API index_add_ { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_add_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "index_add_(Tensor(a!) self, int dim, Tensor index, Tensor source, *, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha); +}; + +struct TORCH_API index_add { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_add"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "index_add(Tensor self, int dim, Tensor index, Tensor source, *, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha); +}; + +struct TORCH_API index_add_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_add"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "index_add.dimname(Tensor self, Dimname dim, Tensor index, Tensor source, *, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & source, const at::Scalar & alpha); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/index_copy_ops.h b/backends/pytorch/include/ATen/ops/index_copy_ops.h new file mode 100644 index 000000000..3863b2dea --- /dev/null +++ b/backends/pytorch/include/ATen/ops/index_copy_ops.h @@ -0,0 +1,78 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API index_copy_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_copy"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "index_copy.out(Tensor self, int dim, Tensor index, Tensor source, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, at::Tensor & out); +}; + +struct TORCH_API index_copy_ { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_copy_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "index_copy_(Tensor(a!) self, int dim, Tensor index, Tensor source) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source); +}; + +struct TORCH_API index_copy { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_copy"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "index_copy(Tensor self, int dim, Tensor index, Tensor source) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source); +}; + +struct TORCH_API index_copy__dimname { + using schema = at::Tensor & (at::Tensor &, at::Dimname, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_copy_"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "index_copy_.dimname(Tensor(a!) self, Dimname dim, Tensor index, Tensor source) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & source); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & source); +}; + +struct TORCH_API index_copy_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_copy"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "index_copy.dimname(Tensor self, Dimname dim, Tensor index, Tensor source) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & source); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & source); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/index_fill_ops.h b/backends/pytorch/include/ATen/ops/index_fill_ops.h new file mode 100644 index 000000000..14f99eb5e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/index_fill_ops.h @@ -0,0 +1,133 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API index_fill__int_Scalar { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_fill_"; + static constexpr const char* overload_name = "int_Scalar"; + static constexpr const char* schema_str = "index_fill_.int_Scalar(Tensor(a!) self, int dim, Tensor index, Scalar value) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value); +}; + +struct TORCH_API index_fill_int_Scalar { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_fill"; + static constexpr const char* overload_name = "int_Scalar"; + static constexpr const char* schema_str = "index_fill.int_Scalar(Tensor self, int dim, Tensor index, Scalar value) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value); +}; + +struct TORCH_API index_fill__int_Tensor { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_fill_"; + static constexpr const char* overload_name = "int_Tensor"; + static constexpr const char* schema_str = "index_fill_.int_Tensor(Tensor(a!) self, int dim, Tensor index, Tensor value) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & value); +}; + +struct TORCH_API index_fill_int_Tensor { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_fill"; + static constexpr const char* overload_name = "int_Tensor"; + static constexpr const char* schema_str = "index_fill.int_Tensor(Tensor self, int dim, Tensor index, Tensor value) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & value); +}; + +struct TORCH_API index_fill__Dimname_Scalar { + using schema = at::Tensor & (at::Tensor &, at::Dimname, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_fill_"; + static constexpr const char* overload_name = "Dimname_Scalar"; + static constexpr const char* schema_str = "index_fill_.Dimname_Scalar(Tensor(a!) self, Dimname dim, Tensor index, Scalar value) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Scalar & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Scalar & value); +}; + +struct TORCH_API index_fill__Dimname_Tensor { + using schema = at::Tensor & (at::Tensor &, at::Dimname, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_fill_"; + static constexpr const char* overload_name = "Dimname_Tensor"; + static constexpr const char* schema_str = "index_fill_.Dimname_Tensor(Tensor(a!) self, Dimname dim, Tensor index, Tensor value) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & value); +}; + +struct TORCH_API index_fill_Dimname_Scalar { + using schema = at::Tensor (const at::Tensor &, at::Dimname, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_fill"; + static constexpr const char* overload_name = "Dimname_Scalar"; + static constexpr const char* schema_str = "index_fill.Dimname_Scalar(Tensor self, Dimname dim, Tensor index, Scalar value) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Scalar & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Scalar & value); +}; + +struct TORCH_API index_fill_Dimname_Tensor { + using schema = at::Tensor (const at::Tensor &, at::Dimname, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_fill"; + static constexpr const char* overload_name = "Dimname_Tensor"; + static constexpr const char* schema_str = "index_fill.Dimname_Tensor(Tensor self, Dimname dim, Tensor index, Tensor value) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & value); +}; + +struct TORCH_API index_fill_int_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_fill"; + static constexpr const char* overload_name = "int_Scalar_out"; + static constexpr const char* schema_str = "index_fill.int_Scalar_out(Tensor self, int dim, Tensor index, Scalar value, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value, at::Tensor & out); +}; + +struct TORCH_API index_fill_int_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_fill"; + static constexpr const char* overload_name = "int_Tensor_out"; + static constexpr const char* schema_str = "index_fill.int_Tensor_out(Tensor self, int dim, Tensor index, Tensor value, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & value, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/index_ops.h b/backends/pytorch/include/ATen/ops/index_ops.h new file mode 100644 index 000000000..716e25f91 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/index_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API index_Tensor { + using schema = at::Tensor (const at::Tensor &, const c10::List<::std::optional> &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "index.Tensor(Tensor self, Tensor?[] indices) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const c10::List<::std::optional> & indices); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const c10::List<::std::optional> & indices); +}; + +struct TORCH_API index_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const c10::List<::std::optional> &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "index.Tensor_out(Tensor self, Tensor?[] indices, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const c10::List<::std::optional> & indices, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const c10::List<::std::optional> & indices, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/index_put_ops.h b/backends/pytorch/include/ATen/ops/index_put_ops.h new file mode 100644 index 000000000..d07305d00 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/index_put_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API index_put_ { + using schema = at::Tensor & (at::Tensor &, const c10::List<::std::optional> &, const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_put_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "index_put_(Tensor(a!) self, Tensor?[] indices, Tensor values, bool accumulate=False) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const c10::List<::std::optional> & indices, const at::Tensor & values, bool accumulate); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const c10::List<::std::optional> & indices, const at::Tensor & values, bool accumulate); +}; + +struct TORCH_API index_put { + using schema = at::Tensor (const at::Tensor &, const c10::List<::std::optional> &, const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_put"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "index_put(Tensor self, Tensor?[] indices, Tensor values, bool accumulate=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const c10::List<::std::optional> & indices, const at::Tensor & values, bool accumulate); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const c10::List<::std::optional> & indices, const at::Tensor & values, bool accumulate); +}; + +struct TORCH_API index_put_out { + using schema = at::Tensor & (const at::Tensor &, const c10::List<::std::optional> &, const at::Tensor &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_put"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "index_put.out(Tensor self, Tensor?[] indices, Tensor values, bool accumulate=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const c10::List<::std::optional> & indices, const at::Tensor & values, bool accumulate, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const c10::List<::std::optional> & indices, const at::Tensor & values, bool accumulate, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/index_reduce_ops.h b/backends/pytorch/include/ATen/ops/index_reduce_ops.h new file mode 100644 index 000000000..4dc3b3db0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/index_reduce_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API index_reduce_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, c10::string_view, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_reduce"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "index_reduce.out(Tensor self, int dim, Tensor index, Tensor source, str reduce, *, bool include_self=True, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, c10::string_view reduce, bool include_self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, c10::string_view reduce, bool include_self, at::Tensor & out); +}; + +struct TORCH_API index_reduce_ { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, c10::string_view, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_reduce_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "index_reduce_(Tensor(a!) self, int dim, Tensor index, Tensor source, str reduce, *, bool include_self=True) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, c10::string_view reduce, bool include_self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, c10::string_view reduce, bool include_self); +}; + +struct TORCH_API index_reduce { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, c10::string_view, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_reduce"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "index_reduce(Tensor self, int dim, Tensor index, Tensor source, str reduce, *, bool include_self=True) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, c10::string_view reduce, bool include_self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & source, c10::string_view reduce, bool include_self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/index_select_ops.h b/backends/pytorch/include/ATen/ops/index_select_ops.h new file mode 100644 index 000000000..883863c98 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/index_select_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API index_select_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_select"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "index_select.out(Tensor self, int dim, Tensor index, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, at::Tensor & out); +}; + +struct TORCH_API index_select { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_select"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "index_select(Tensor self, int dim, Tensor index) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index); +}; + +struct TORCH_API index_select_dimname_out { + using schema = at::Tensor & (const at::Tensor &, at::Dimname, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_select"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "index_select.dimname_out(Tensor self, Dimname dim, Tensor index, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index, at::Tensor & out); +}; + +struct TORCH_API index_select_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::index_select"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "index_select.dimname(Tensor self, Dimname dim, Tensor index) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/indices_ops.h b/backends/pytorch/include/ATen/ops/indices_ops.h new file mode 100644 index 000000000..501687a0a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/indices_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API indices { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::indices"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "indices(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/inner_ops.h b/backends/pytorch/include/ATen/ops/inner_ops.h new file mode 100644 index 000000000..3916ecdf2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/inner_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API inner { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::inner"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "inner(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API inner_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::inner"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "inner.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/int_repr_ops.h b/backends/pytorch/include/ATen/ops/int_repr_ops.h new file mode 100644 index 000000000..8a326b66d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/int_repr_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API int_repr { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::int_repr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "int_repr(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API int_repr_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::int_repr"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "int_repr.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/inverse_ops.h b/backends/pytorch/include/ATen/ops/inverse_ops.h new file mode 100644 index 000000000..9d55bc97c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/inverse_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API inverse { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::inverse"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "inverse(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API inverse_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::inverse"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "inverse.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_coalesced_ops.h b/backends/pytorch/include/ATen/ops/is_coalesced_ops.h new file mode 100644 index 000000000..c956cecf0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_coalesced_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_coalesced { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_coalesced"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_coalesced(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_complex_ops.h b/backends/pytorch/include/ATen/ops/is_complex_ops.h new file mode 100644 index 000000000..39618a4c9 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_complex_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_complex { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_complex"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_complex(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_conj_ops.h b/backends/pytorch/include/ATen/ops/is_conj_ops.h new file mode 100644 index 000000000..80b868cbc --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_conj_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_conj { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_conj"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_conj(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_distributed_ops.h b/backends/pytorch/include/ATen/ops/is_distributed_ops.h new file mode 100644 index 000000000..aa77604aa --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_distributed_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_distributed { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_distributed"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_distributed(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_floating_point_ops.h b/backends/pytorch/include/ATen/ops/is_floating_point_ops.h new file mode 100644 index 000000000..86be94692 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_floating_point_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_floating_point { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_floating_point"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_floating_point(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_inference_ops.h b/backends/pytorch/include/ATen/ops/is_inference_ops.h new file mode 100644 index 000000000..8759c7f6b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_inference_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_inference { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_inference"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_inference(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_leaf_ops.h b/backends/pytorch/include/ATen/ops/is_leaf_ops.h new file mode 100644 index 000000000..1bcb22e66 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_leaf_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_leaf { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_leaf"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_leaf(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_neg_ops.h b/backends/pytorch/include/ATen/ops/is_neg_ops.h new file mode 100644 index 000000000..a0b840877 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_neg_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_neg { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_neg"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_neg(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_nonzero_ops.h b/backends/pytorch/include/ATen/ops/is_nonzero_ops.h new file mode 100644 index 000000000..e455790b6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_nonzero_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_nonzero { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_nonzero"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_nonzero(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_pinned_ops.h b/backends/pytorch/include/ATen/ops/is_pinned_ops.h new file mode 100644 index 000000000..352153849 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_pinned_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_pinned { + using schema = bool (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_pinned"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_pinned(Tensor self, Device? device=None) -> bool"; + static bool call(const at::Tensor & self, ::std::optional device); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional device); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_same_size_ops.h b/backends/pytorch/include/ATen/ops/is_same_size_ops.h new file mode 100644 index 000000000..8eaf6f7f0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_same_size_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_same_size { + using schema = bool (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_same_size"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_same_size(Tensor self, Tensor other) -> bool"; + static bool call(const at::Tensor & self, const at::Tensor & other); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_set_to_ops.h b/backends/pytorch/include/ATen/ops/is_set_to_ops.h new file mode 100644 index 000000000..723374e73 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_set_to_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_set_to { + using schema = bool (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_set_to"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_set_to(Tensor self, Tensor tensor) -> bool"; + static bool call(const at::Tensor & self, const at::Tensor & tensor); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & tensor); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/is_signed_ops.h b/backends/pytorch/include/ATen/ops/is_signed_ops.h new file mode 100644 index 000000000..d25b945e3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/is_signed_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API is_signed { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::is_signed"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "is_signed(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/isclose_ops.h b/backends/pytorch/include/ATen/ops/isclose_ops.h new file mode 100644 index 000000000..bfa994e71 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/isclose_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API isclose { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, double, double, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isclose"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "isclose(Tensor self, Tensor other, float rtol=1e-05, float atol=1e-08, bool equal_nan=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other, double rtol, double atol, bool equal_nan); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, double rtol, double atol, bool equal_nan); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/isfinite_ops.h b/backends/pytorch/include/ATen/ops/isfinite_ops.h new file mode 100644 index 000000000..d251cecd8 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/isfinite_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API isfinite { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isfinite"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "isfinite(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/isinf_ops.h b/backends/pytorch/include/ATen/ops/isinf_ops.h new file mode 100644 index 000000000..976dc2c12 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/isinf_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API isinf { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isinf"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "isinf(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API isinf_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isinf"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "isinf.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/isnan_ops.h b/backends/pytorch/include/ATen/ops/isnan_ops.h new file mode 100644 index 000000000..02f17f66c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/isnan_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API isnan { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isnan"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "isnan(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API isnan_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isnan"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "isnan.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/isneginf_ops.h b/backends/pytorch/include/ATen/ops/isneginf_ops.h new file mode 100644 index 000000000..ee502e46a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/isneginf_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API isneginf { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isneginf"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "isneginf(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API isneginf_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isneginf"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "isneginf.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/isposinf_ops.h b/backends/pytorch/include/ATen/ops/isposinf_ops.h new file mode 100644 index 000000000..b214fa3a6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/isposinf_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API isposinf { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isposinf"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "isposinf(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API isposinf_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isposinf"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "isposinf.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/isreal_ops.h b/backends/pytorch/include/ATen/ops/isreal_ops.h new file mode 100644 index 000000000..d5d6a9847 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/isreal_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API isreal { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::isreal"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "isreal(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/istft_ops.h b/backends/pytorch/include/ATen/ops/istft_ops.h new file mode 100644 index 000000000..509cf2a07 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/istft_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API istft { + using schema = at::Tensor (const at::Tensor &, int64_t, ::std::optional, ::std::optional, const ::std::optional &, bool, bool, ::std::optional, ::std::optional, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::istft"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "istft(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool center=True, bool normalized=False, bool? onesided=None, int? length=None, bool return_complex=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t n_fft, ::std::optional hop_length, ::std::optional win_length, const ::std::optional & window, bool center, bool normalized, ::std::optional onesided, ::std::optional length, bool return_complex); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t n_fft, ::std::optional hop_length, ::std::optional win_length, const ::std::optional & window, bool center, bool normalized, ::std::optional onesided, ::std::optional length, bool return_complex); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/item_ops.h b/backends/pytorch/include/ATen/ops/item_ops.h new file mode 100644 index 000000000..d2e9d43d7 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/item_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API item { + using schema = at::Scalar (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::item"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "item(Tensor self) -> Scalar"; + static at::Scalar call(const at::Tensor & self); + static at::Scalar redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/kron_ops.h b/backends/pytorch/include/ATen/ops/kron_ops.h new file mode 100644 index 000000000..2dd231ac1 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/kron_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API kron { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::kron"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "kron(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API kron_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::kron"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "kron.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/kthvalue_ops.h b/backends/pytorch/include/ATen/ops/kthvalue_ops.h new file mode 100644 index 000000000..895466421 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/kthvalue_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API kthvalue { + using schema = ::std::tuple (const at::Tensor &, c10::SymInt, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::kthvalue"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "kthvalue(Tensor self, SymInt k, int dim=-1, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, c10::SymInt k, int64_t dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt k, int64_t dim, bool keepdim); +}; + +struct TORCH_API kthvalue_values { + using schema = ::std::tuple (const at::Tensor &, c10::SymInt, int64_t, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::kthvalue"; + static constexpr const char* overload_name = "values"; + static constexpr const char* schema_str = "kthvalue.values(Tensor self, SymInt k, int dim=-1, bool keepdim=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, c10::SymInt k, int64_t dim, bool keepdim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt k, int64_t dim, bool keepdim, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API kthvalue_dimname { + using schema = ::std::tuple (const at::Tensor &, c10::SymInt, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::kthvalue"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "kthvalue.dimname(Tensor self, SymInt k, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, c10::SymInt k, at::Dimname dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt k, at::Dimname dim, bool keepdim); +}; + +struct TORCH_API kthvalue_dimname_out { + using schema = ::std::tuple (const at::Tensor &, c10::SymInt, at::Dimname, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::kthvalue"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "kthvalue.dimname_out(Tensor self, SymInt k, Dimname dim, bool keepdim=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, c10::SymInt k, at::Dimname dim, bool keepdim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt k, at::Dimname dim, bool keepdim, at::Tensor & values, at::Tensor & indices); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/lcm_ops.h b/backends/pytorch/include/ATen/ops/lcm_ops.h new file mode 100644 index 000000000..dc236343f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/lcm_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API lcm_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lcm"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "lcm.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API lcm { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lcm"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "lcm(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API lcm_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lcm_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "lcm_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/ldexp_ops.h b/backends/pytorch/include/ATen/ops/ldexp_ops.h new file mode 100644 index 000000000..ded3b02e5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/ldexp_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API ldexp_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ldexp"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "ldexp.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API ldexp_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ldexp_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "ldexp_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API ldexp_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ldexp"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "ldexp.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/le_ops.h b/backends/pytorch/include/ATen/ops/le_ops.h new file mode 100644 index 000000000..3901bdcee --- /dev/null +++ b/backends/pytorch/include/ATen/ops/le_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API le_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::le"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "le.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API le_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::le"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "le.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API le_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::le"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "le.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API le_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::le"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "le.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API le__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::le_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "le_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API le__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::le_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "le_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/lerp_ops.h b/backends/pytorch/include/ATen/ops/lerp_ops.h new file mode 100644 index 000000000..0393c817a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/lerp_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API lerp__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lerp_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "lerp_.Scalar(Tensor(a!) self, Tensor end, Scalar weight) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & end, const at::Scalar & weight); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & end, const at::Scalar & weight); +}; + +struct TORCH_API lerp__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lerp_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "lerp_.Tensor(Tensor(a!) self, Tensor end, Tensor weight) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & end, const at::Tensor & weight); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & end, const at::Tensor & weight); +}; + +struct TORCH_API lerp_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lerp"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "lerp.Scalar_out(Tensor self, Tensor end, Scalar weight, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & end, const at::Scalar & weight, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & end, const at::Scalar & weight, at::Tensor & out); +}; + +struct TORCH_API lerp_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lerp"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "lerp.Tensor_out(Tensor self, Tensor end, Tensor weight, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & end, const at::Tensor & weight, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & end, const at::Tensor & weight, at::Tensor & out); +}; + +struct TORCH_API lerp_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lerp"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "lerp.Scalar(Tensor self, Tensor end, Scalar weight) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & end, const at::Scalar & weight); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & end, const at::Scalar & weight); +}; + +struct TORCH_API lerp_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lerp"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "lerp.Tensor(Tensor self, Tensor end, Tensor weight) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & end, const at::Tensor & weight); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & end, const at::Tensor & weight); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/less_equal_ops.h b/backends/pytorch/include/ATen/ops/less_equal_ops.h new file mode 100644 index 000000000..d469b82d6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/less_equal_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API less_equal_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less_equal"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "less_equal.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API less_equal_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less_equal"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "less_equal.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API less_equal_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less_equal"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "less_equal.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API less_equal_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less_equal"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "less_equal.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API less_equal__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less_equal_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "less_equal_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API less_equal__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less_equal_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "less_equal_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/less_ops.h b/backends/pytorch/include/ATen/ops/less_ops.h new file mode 100644 index 000000000..b9860532b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/less_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API less_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "less.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API less_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "less.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API less_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "less.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API less_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "less.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API less__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "less_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API less__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::less_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "less_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/lgamma_ops.h b/backends/pytorch/include/ATen/ops/lgamma_ops.h new file mode 100644 index 000000000..5b7967e4e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/lgamma_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API lgamma_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lgamma"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "lgamma.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API lgamma_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lgamma_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "lgamma_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API lgamma { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lgamma"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "lgamma(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/log10_ops.h b/backends/pytorch/include/ATen/ops/log10_ops.h new file mode 100644 index 000000000..066e7df74 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/log10_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API log10 { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log10"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "log10(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API log10_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log10_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "log10_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API log10_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log10"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "log10.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/log1p_ops.h b/backends/pytorch/include/ATen/ops/log1p_ops.h new file mode 100644 index 000000000..c2de140ae --- /dev/null +++ b/backends/pytorch/include/ATen/ops/log1p_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API log1p { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log1p"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "log1p(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API log1p_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log1p_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "log1p_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API log1p_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log1p"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "log1p.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/log2_ops.h b/backends/pytorch/include/ATen/ops/log2_ops.h new file mode 100644 index 000000000..8d47653a1 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/log2_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API log2 { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log2"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "log2(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API log2_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log2_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "log2_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API log2_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log2"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "log2.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/log_normal_ops.h b/backends/pytorch/include/ATen/ops/log_normal_ops.h new file mode 100644 index 000000000..43926175d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/log_normal_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API log_normal_ { + using schema = at::Tensor & (at::Tensor &, double, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log_normal_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "log_normal_(Tensor(a!) self, float mean=1, float std=2, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, double mean, double std, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, double mean, double std, ::std::optional generator); +}; + +struct TORCH_API log_normal_out { + using schema = at::Tensor & (const at::Tensor &, double, double, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log_normal"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "log_normal.out(Tensor self, float mean=1, float std=2, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, double mean, double std, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double mean, double std, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API log_normal { + using schema = at::Tensor (const at::Tensor &, double, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log_normal"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "log_normal(Tensor self, float mean=1, float std=2, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, double mean, double std, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double mean, double std, ::std::optional generator); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/log_ops.h b/backends/pytorch/include/ATen/ops/log_ops.h new file mode 100644 index 000000000..8c58863c3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/log_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API log { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "log(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API log_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "log_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API log_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "log.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/log_softmax_ops.h b/backends/pytorch/include/ATen/ops/log_softmax_ops.h new file mode 100644 index 000000000..5b03c80e9 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/log_softmax_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API log_softmax_int { + using schema = at::Tensor (const at::Tensor &, int64_t, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log_softmax"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "log_softmax.int(Tensor self, int dim, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, ::std::optional dtype); +}; + +struct TORCH_API log_softmax_int_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log_softmax"; + static constexpr const char* overload_name = "int_out"; + static constexpr const char* schema_str = "log_softmax.int_out(Tensor self, int dim, ScalarType? dtype=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, ::std::optional dtype, at::Tensor & out); +}; + +struct TORCH_API log_softmax_Dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::log_softmax"; + static constexpr const char* overload_name = "Dimname"; + static constexpr const char* schema_str = "log_softmax.Dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, ::std::optional dtype); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/logaddexp2_ops.h b/backends/pytorch/include/ATen/ops/logaddexp2_ops.h new file mode 100644 index 000000000..170fd9850 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/logaddexp2_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API logaddexp2_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logaddexp2"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "logaddexp2.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API logaddexp2 { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logaddexp2"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logaddexp2(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/logaddexp_ops.h b/backends/pytorch/include/ATen/ops/logaddexp_ops.h new file mode 100644 index 000000000..178835bee --- /dev/null +++ b/backends/pytorch/include/ATen/ops/logaddexp_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API logaddexp_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logaddexp"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "logaddexp.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API logaddexp { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logaddexp"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logaddexp(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/logcumsumexp_ops.h b/backends/pytorch/include/ATen/ops/logcumsumexp_ops.h new file mode 100644 index 000000000..de14206d0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/logcumsumexp_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API logcumsumexp { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logcumsumexp"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logcumsumexp(Tensor self, int dim) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim); +}; + +struct TORCH_API logcumsumexp_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logcumsumexp"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "logcumsumexp.out(Tensor self, int dim, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, at::Tensor & out); +}; + +struct TORCH_API logcumsumexp_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logcumsumexp"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "logcumsumexp.dimname(Tensor self, Dimname dim) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim); +}; + +struct TORCH_API logcumsumexp_dimname_out { + using schema = at::Tensor & (const at::Tensor &, at::Dimname, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logcumsumexp"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "logcumsumexp.dimname_out(Tensor self, Dimname dim, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Dimname dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/logdet_ops.h b/backends/pytorch/include/ATen/ops/logdet_ops.h new file mode 100644 index 000000000..b1414f715 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/logdet_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API logdet { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logdet"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logdet(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/logical_and_ops.h b/backends/pytorch/include/ATen/ops/logical_and_ops.h new file mode 100644 index 000000000..9d20d2b14 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/logical_and_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API logical_and { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_and"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logical_and(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API logical_and_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_and_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logical_and_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API logical_and_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_and"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "logical_and.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/logical_not_ops.h b/backends/pytorch/include/ATen/ops/logical_not_ops.h new file mode 100644 index 000000000..4517cb5e9 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/logical_not_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API logical_not { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_not"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logical_not(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API logical_not_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_not_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logical_not_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API logical_not_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_not"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "logical_not.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/logical_or_ops.h b/backends/pytorch/include/ATen/ops/logical_or_ops.h new file mode 100644 index 000000000..63e658d63 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/logical_or_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API logical_or { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_or"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logical_or(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API logical_or_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_or_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logical_or_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API logical_or_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_or"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "logical_or.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/logical_xor_ops.h b/backends/pytorch/include/ATen/ops/logical_xor_ops.h new file mode 100644 index 000000000..e2ff73f56 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/logical_xor_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API logical_xor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_xor"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logical_xor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API logical_xor_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_xor_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logical_xor_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API logical_xor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logical_xor"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "logical_xor.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/logit_ops.h b/backends/pytorch/include/ATen/ops/logit_ops.h new file mode 100644 index 000000000..fa426ce05 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/logit_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API logit { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logit"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logit(Tensor self, float? eps=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional eps); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional eps); +}; + +struct TORCH_API logit_ { + using schema = at::Tensor & (at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logit_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logit_(Tensor(a!) self, float? eps=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, ::std::optional eps); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, ::std::optional eps); +}; + +struct TORCH_API logit_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logit"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "logit.out(Tensor self, float? eps=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional eps, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional eps, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/logsumexp_ops.h b/backends/pytorch/include/ATen/ops/logsumexp_ops.h new file mode 100644 index 000000000..a94ce6845 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/logsumexp_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API logsumexp { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logsumexp"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "logsumexp(Tensor self, int[1] dim, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim, bool keepdim); +}; + +struct TORCH_API logsumexp_out { + using schema = at::Tensor & (const at::Tensor &, at::IntArrayRef, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logsumexp"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "logsumexp.out(Tensor self, int[1] dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::IntArrayRef dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API logsumexp_names { + using schema = at::Tensor (const at::Tensor &, at::DimnameList, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logsumexp"; + static constexpr const char* overload_name = "names"; + static constexpr const char* schema_str = "logsumexp.names(Tensor self, Dimname[1] dim, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::DimnameList dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, bool keepdim); +}; + +struct TORCH_API logsumexp_names_out { + using schema = at::Tensor & (const at::Tensor &, at::DimnameList, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::logsumexp"; + static constexpr const char* overload_name = "names_out"; + static constexpr const char* schema_str = "logsumexp.names_out(Tensor self, Dimname[1] dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::DimnameList dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, bool keepdim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/lshift_ops.h b/backends/pytorch/include/ATen/ops/lshift_ops.h new file mode 100644 index 000000000..bddac5879 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/lshift_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API __lshift___Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__lshift__"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "__lshift__.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API __lshift___Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__lshift__"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "__lshift__.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API __ilshift___Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__ilshift__"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "__ilshift__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API __ilshift___Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__ilshift__"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "__ilshift__.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API __lshift___Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__lshift__"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "__lshift__.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API __lshift___Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__lshift__"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "__lshift__.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/lt_ops.h b/backends/pytorch/include/ATen/ops/lt_ops.h new file mode 100644 index 000000000..56be923a8 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/lt_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API lt_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lt"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "lt.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API lt_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lt"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "lt.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API lt_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lt"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "lt.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API lt_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lt"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "lt.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API lt__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lt_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "lt_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API lt__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lt_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "lt_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/lu_solve_ops.h b/backends/pytorch/include/ATen/ops/lu_solve_ops.h new file mode 100644 index 000000000..f36976b95 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/lu_solve_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API lu_solve_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lu_solve"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "lu_solve.out(Tensor self, Tensor LU_data, Tensor LU_pivots, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & LU_data, const at::Tensor & LU_pivots, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & LU_data, const at::Tensor & LU_pivots, at::Tensor & out); +}; + +struct TORCH_API lu_solve { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::lu_solve"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "lu_solve(Tensor self, Tensor LU_data, Tensor LU_pivots) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & LU_data, const at::Tensor & LU_pivots); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & LU_data, const at::Tensor & LU_pivots); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/mH_ops.h b/backends/pytorch/include/ATen/ops/mH_ops.h new file mode 100644 index 000000000..fbddffafb --- /dev/null +++ b/backends/pytorch/include/ATen/ops/mH_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API mH { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mH"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "mH(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/mT_ops.h b/backends/pytorch/include/ATen/ops/mT_ops.h new file mode 100644 index 000000000..3c900320a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/mT_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API mT { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mT"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "mT(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/masked_fill_ops.h b/backends/pytorch/include/ATen/ops/masked_fill_ops.h new file mode 100644 index 000000000..24ce78320 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/masked_fill_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API masked_fill__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_fill_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "masked_fill_.Scalar(Tensor(a!) self, Tensor mask, Scalar value) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & mask, const at::Scalar & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & mask, const at::Scalar & value); +}; + +struct TORCH_API masked_fill_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_fill"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "masked_fill.Scalar(Tensor self, Tensor mask, Scalar value) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mask, const at::Scalar & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask, const at::Scalar & value); +}; + +struct TORCH_API masked_fill__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_fill_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "masked_fill_.Tensor(Tensor(a!) self, Tensor mask, Tensor value) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & mask, const at::Tensor & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & mask, const at::Tensor & value); +}; + +struct TORCH_API masked_fill_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_fill"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "masked_fill.Tensor(Tensor self, Tensor mask, Tensor value) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mask, const at::Tensor & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask, const at::Tensor & value); +}; + +struct TORCH_API masked_fill_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_fill"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "masked_fill.Scalar_out(Tensor self, Tensor mask, Scalar value, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mask, const at::Scalar & value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask, const at::Scalar & value, at::Tensor & out); +}; + +struct TORCH_API masked_fill_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_fill"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "masked_fill.Tensor_out(Tensor self, Tensor mask, Tensor value, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mask, const at::Tensor & value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask, const at::Tensor & value, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/masked_scatter_ops.h b/backends/pytorch/include/ATen/ops/masked_scatter_ops.h new file mode 100644 index 000000000..52d5e8cfb --- /dev/null +++ b/backends/pytorch/include/ATen/ops/masked_scatter_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API masked_scatter_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_scatter_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "masked_scatter_(Tensor(a!) self, Tensor mask, Tensor source) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & mask, const at::Tensor & source); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & mask, const at::Tensor & source); +}; + +struct TORCH_API masked_scatter { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_scatter"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "masked_scatter(Tensor self, Tensor mask, Tensor source) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mask, const at::Tensor & source); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask, const at::Tensor & source); +}; + +struct TORCH_API masked_scatter_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_scatter"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "masked_scatter.out(Tensor self, Tensor mask, Tensor source, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mask, const at::Tensor & source, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask, const at::Tensor & source, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/masked_select_ops.h b/backends/pytorch/include/ATen/ops/masked_select_ops.h new file mode 100644 index 000000000..85706ff98 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/masked_select_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API masked_select_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_select"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "masked_select.out(Tensor self, Tensor mask, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mask, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask, at::Tensor & out); +}; + +struct TORCH_API masked_select { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::masked_select"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "masked_select(Tensor self, Tensor mask) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mask); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/matmul_ops.h b/backends/pytorch/include/ATen/ops/matmul_ops.h new file mode 100644 index 000000000..7c3508971 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/matmul_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API matmul { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::matmul"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "matmul(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API matmul_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::matmul"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "matmul.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/matrix_H_ops.h b/backends/pytorch/include/ATen/ops/matrix_H_ops.h new file mode 100644 index 000000000..b80ff0921 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/matrix_H_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API matrix_H { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::matrix_H"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "matrix_H(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/matrix_exp_ops.h b/backends/pytorch/include/ATen/ops/matrix_exp_ops.h new file mode 100644 index 000000000..f0e8c5f45 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/matrix_exp_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API matrix_exp { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::matrix_exp"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "matrix_exp(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/matrix_power_ops.h b/backends/pytorch/include/ATen/ops/matrix_power_ops.h new file mode 100644 index 000000000..3aa1dac22 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/matrix_power_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API matrix_power { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::matrix_power"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "matrix_power(Tensor self, int n) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t n); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t n); +}; + +struct TORCH_API matrix_power_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::matrix_power"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "matrix_power.out(Tensor self, int n, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t n, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t n, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/max_ops.h b/backends/pytorch/include/ATen/ops/max_ops.h new file mode 100644 index 000000000..6962de106 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/max_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API max_dim { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::max"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "max.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim); +}; + +struct TORCH_API max_dim_max { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::max"; + static constexpr const char* overload_name = "dim_max"; + static constexpr const char* schema_str = "max.dim_max(Tensor self, int dim, bool keepdim=False, *, Tensor(a!) max, Tensor(b!) max_values) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & max, at::Tensor & max_values); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & max, at::Tensor & max_values); +}; + +struct TORCH_API max_names_dim { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::max"; + static constexpr const char* overload_name = "names_dim"; + static constexpr const char* schema_str = "max.names_dim(Tensor self, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim); +}; + +struct TORCH_API max_names_dim_max { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::max"; + static constexpr const char* overload_name = "names_dim_max"; + static constexpr const char* schema_str = "max.names_dim_max(Tensor self, Dimname dim, bool keepdim=False, *, Tensor(a!) max, Tensor(b!) max_values) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & max, at::Tensor & max_values); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & max, at::Tensor & max_values); +}; + +struct TORCH_API max { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::max"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "max(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API max_other { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::max"; + static constexpr const char* overload_name = "other"; + static constexpr const char* schema_str = "max.other(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API max_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::max"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "max.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API max_unary_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::max"; + static constexpr const char* overload_name = "unary_out"; + static constexpr const char* schema_str = "max.unary_out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/maximum_ops.h b/backends/pytorch/include/ATen/ops/maximum_ops.h new file mode 100644 index 000000000..bf636b10d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/maximum_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API maximum { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::maximum"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "maximum(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API maximum_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::maximum"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "maximum.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/mean_ops.h b/backends/pytorch/include/ATen/ops/mean_ops.h new file mode 100644 index 000000000..ce974cc26 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/mean_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API mean { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mean"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "mean(Tensor self, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype); +}; + +struct TORCH_API mean_dtype_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mean"; + static constexpr const char* overload_name = "dtype_out"; + static constexpr const char* schema_str = "mean.dtype_out(Tensor self, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype, at::Tensor & out); +}; + +struct TORCH_API mean_dim { + using schema = at::Tensor (const at::Tensor &, at::OptionalIntArrayRef, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mean"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "mean.dim(Tensor self, int[1]? dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype); +}; + +struct TORCH_API mean_out { + using schema = at::Tensor & (const at::Tensor &, at::OptionalIntArrayRef, bool, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mean"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "mean.out(Tensor self, int[1]? dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype, at::Tensor & out); +}; + +struct TORCH_API mean_names_dim { + using schema = at::Tensor (const at::Tensor &, at::DimnameList, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mean"; + static constexpr const char* overload_name = "names_dim"; + static constexpr const char* schema_str = "mean.names_dim(Tensor self, Dimname[1] dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::DimnameList dim, bool keepdim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, bool keepdim, ::std::optional dtype); +}; + +struct TORCH_API mean_names_out { + using schema = at::Tensor & (const at::Tensor &, at::DimnameList, bool, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mean"; + static constexpr const char* overload_name = "names_out"; + static constexpr const char* schema_str = "mean.names_out(Tensor self, Dimname[1] dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::DimnameList dim, bool keepdim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, bool keepdim, ::std::optional dtype, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/median_ops.h b/backends/pytorch/include/ATen/ops/median_ops.h new file mode 100644 index 000000000..47b96ce12 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/median_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API median { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::median"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "median(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API median_dim { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::median"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "median.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim); +}; + +struct TORCH_API median_dim_values { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::median"; + static constexpr const char* overload_name = "dim_values"; + static constexpr const char* schema_str = "median.dim_values(Tensor self, int dim, bool keepdim=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API median_names_dim { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::median"; + static constexpr const char* overload_name = "names_dim"; + static constexpr const char* schema_str = "median.names_dim(Tensor self, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim); +}; + +struct TORCH_API median_names_dim_values { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::median"; + static constexpr const char* overload_name = "names_dim_values"; + static constexpr const char* schema_str = "median.names_dim_values(Tensor self, Dimname dim, bool keepdim=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API median_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::median"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "median.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/min_ops.h b/backends/pytorch/include/ATen/ops/min_ops.h new file mode 100644 index 000000000..2ef662138 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/min_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API min_dim { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::min"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "min.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim); +}; + +struct TORCH_API min_dim_min { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::min"; + static constexpr const char* overload_name = "dim_min"; + static constexpr const char* schema_str = "min.dim_min(Tensor self, int dim, bool keepdim=False, *, Tensor(a!) min, Tensor(b!) min_indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & min, at::Tensor & min_indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & min, at::Tensor & min_indices); +}; + +struct TORCH_API min_names_dim { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::min"; + static constexpr const char* overload_name = "names_dim"; + static constexpr const char* schema_str = "min.names_dim(Tensor self, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim); +}; + +struct TORCH_API min_names_dim_min { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::min"; + static constexpr const char* overload_name = "names_dim_min"; + static constexpr const char* schema_str = "min.names_dim_min(Tensor self, Dimname dim, bool keepdim=False, *, Tensor(a!) min, Tensor(b!) min_indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & min, at::Tensor & min_indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & min, at::Tensor & min_indices); +}; + +struct TORCH_API min { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::min"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "min(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API min_unary_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::min"; + static constexpr const char* overload_name = "unary_out"; + static constexpr const char* schema_str = "min.unary_out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API min_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::min"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "min.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API min_other { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::min"; + static constexpr const char* overload_name = "other"; + static constexpr const char* schema_str = "min.other(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/minimum_ops.h b/backends/pytorch/include/ATen/ops/minimum_ops.h new file mode 100644 index 000000000..a25d3c595 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/minimum_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API minimum { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::minimum"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "minimum(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API minimum_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::minimum"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "minimum.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/mm_ops.h b/backends/pytorch/include/ATen/ops/mm_ops.h new file mode 100644 index 000000000..15dbd2b99 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/mm_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API mm { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mm"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "mm(Tensor self, Tensor mat2) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mat2); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat2); +}; + +struct TORCH_API mm_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mm"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "mm.out(Tensor self, Tensor mat2, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mat2, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat2, at::Tensor & out); +}; + +struct TORCH_API mm_dtype { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, at::ScalarType); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mm"; + static constexpr const char* overload_name = "dtype"; + static constexpr const char* schema_str = "mm.dtype(Tensor self, Tensor mat2, ScalarType out_dtype) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mat2, at::ScalarType out_dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat2, at::ScalarType out_dtype); +}; + +struct TORCH_API mm_dtype_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::ScalarType, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mm"; + static constexpr const char* overload_name = "dtype_out"; + static constexpr const char* schema_str = "mm.dtype_out(Tensor self, Tensor mat2, ScalarType out_dtype, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mat2, at::ScalarType out_dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat2, at::ScalarType out_dtype, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/mode_ops.h b/backends/pytorch/include/ATen/ops/mode_ops.h new file mode 100644 index 000000000..1a47cb43a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/mode_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API mode { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mode"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "mode(Tensor self, int dim=-1, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim); +}; + +struct TORCH_API mode_values { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mode"; + static constexpr const char* overload_name = "values"; + static constexpr const char* schema_str = "mode.values(Tensor self, int dim=-1, bool keepdim=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API mode_dimname { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mode"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "mode.dimname(Tensor self, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim); +}; + +struct TORCH_API mode_dimname_out { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mode"; + static constexpr const char* overload_name = "dimname_out"; + static constexpr const char* schema_str = "mode.dimname_out(Tensor self, Dimname dim, bool keepdim=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & values, at::Tensor & indices); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/moveaxis_ops.h b/backends/pytorch/include/ATen/ops/moveaxis_ops.h new file mode 100644 index 000000000..0d620442c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/moveaxis_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API moveaxis_intlist { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::moveaxis"; + static constexpr const char* overload_name = "intlist"; + static constexpr const char* schema_str = "moveaxis.intlist(Tensor(a) self, int[] source, int[] destination) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef source, at::IntArrayRef destination); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef source, at::IntArrayRef destination); +}; + +struct TORCH_API moveaxis_int { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::moveaxis"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "moveaxis.int(Tensor(a) self, int source, int destination) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t source, int64_t destination); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t source, int64_t destination); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/movedim_ops.h b/backends/pytorch/include/ATen/ops/movedim_ops.h new file mode 100644 index 000000000..3c8e01ba8 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/movedim_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API movedim_intlist { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::movedim"; + static constexpr const char* overload_name = "intlist"; + static constexpr const char* schema_str = "movedim.intlist(Tensor(a) self, int[] source, int[] destination) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef source, at::IntArrayRef destination); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef source, at::IntArrayRef destination); +}; + +struct TORCH_API movedim_int { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::movedim"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "movedim.int(Tensor(a) self, int source, int destination) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t source, int64_t destination); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t source, int64_t destination); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/msort_ops.h b/backends/pytorch/include/ATen/ops/msort_ops.h new file mode 100644 index 000000000..b5102ac25 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/msort_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API msort_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::msort"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "msort.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API msort { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::msort"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "msort(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/mul_ops.h b/backends/pytorch/include/ATen/ops/mul_ops.h new file mode 100644 index 000000000..881ca1495 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/mul_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API mul_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mul"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "mul.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API mul__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mul_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "mul_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API mul_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mul"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "mul.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API mul_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mul"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "mul.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API mul__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mul_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "mul_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API mul_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mul"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "mul.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/multinomial_ops.h b/backends/pytorch/include/ATen/ops/multinomial_ops.h new file mode 100644 index 000000000..c178a8bde --- /dev/null +++ b/backends/pytorch/include/ATen/ops/multinomial_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API multinomial_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymInt, bool, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::multinomial"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "multinomial.out(Tensor self, SymInt num_samples, bool replacement=False, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymInt num_samples, bool replacement, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt num_samples, bool replacement, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API multinomial { + using schema = at::Tensor (const at::Tensor &, c10::SymInt, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::multinomial"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "multinomial(Tensor self, SymInt num_samples, bool replacement=False, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymInt num_samples, bool replacement, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt num_samples, bool replacement, ::std::optional generator); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/multiply_ops.h b/backends/pytorch/include/ATen/ops/multiply_ops.h new file mode 100644 index 000000000..76ecb8d9b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/multiply_ops.h @@ -0,0 +1,78 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API multiply_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::multiply"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "multiply.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API multiply__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::multiply_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "multiply_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API multiply_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::multiply"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "multiply.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API multiply_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::multiply"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "multiply.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API multiply__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::multiply_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "multiply_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/mv_ops.h b/backends/pytorch/include/ATen/ops/mv_ops.h new file mode 100644 index 000000000..f2b07981f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/mv_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API mv { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mv"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "mv(Tensor self, Tensor vec) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & vec); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & vec); +}; + +struct TORCH_API mv_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mv"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "mv.out(Tensor self, Tensor vec, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & vec, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & vec, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/mvlgamma_ops.h b/backends/pytorch/include/ATen/ops/mvlgamma_ops.h new file mode 100644 index 000000000..be4e2c4f0 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/mvlgamma_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API mvlgamma_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mvlgamma"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "mvlgamma.out(Tensor self, int p, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t p, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t p, at::Tensor & out); +}; + +struct TORCH_API mvlgamma { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mvlgamma"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "mvlgamma(Tensor self, int p) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t p); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t p); +}; + +struct TORCH_API mvlgamma_ { + using schema = at::Tensor & (at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::mvlgamma_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "mvlgamma_(Tensor(a!) self, int p) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t p); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t p); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/nan_to_num_ops.h b/backends/pytorch/include/ATen/ops/nan_to_num_ops.h new file mode 100644 index 000000000..5bfc74f72 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/nan_to_num_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API nan_to_num { + using schema = at::Tensor (const at::Tensor &, ::std::optional, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nan_to_num"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nan_to_num(Tensor self, float? nan=None, float? posinf=None, float? neginf=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional nan, ::std::optional posinf, ::std::optional neginf); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional nan, ::std::optional posinf, ::std::optional neginf); +}; + +struct TORCH_API nan_to_num_ { + using schema = at::Tensor & (at::Tensor &, ::std::optional, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nan_to_num_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nan_to_num_(Tensor(a!) self, float? nan=None, float? posinf=None, float? neginf=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, ::std::optional nan, ::std::optional posinf, ::std::optional neginf); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, ::std::optional nan, ::std::optional posinf, ::std::optional neginf); +}; + +struct TORCH_API nan_to_num_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, ::std::optional, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nan_to_num"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "nan_to_num.out(Tensor self, float? nan=None, float? posinf=None, float? neginf=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional nan, ::std::optional posinf, ::std::optional neginf, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional nan, ::std::optional posinf, ::std::optional neginf, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/nanmean_ops.h b/backends/pytorch/include/ATen/ops/nanmean_ops.h new file mode 100644 index 000000000..7dc8113ad --- /dev/null +++ b/backends/pytorch/include/ATen/ops/nanmean_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API nanmean { + using schema = at::Tensor (const at::Tensor &, at::OptionalIntArrayRef, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanmean"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nanmean(Tensor self, int[1]? dim=None, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype); +}; + +struct TORCH_API nanmean_out { + using schema = at::Tensor & (const at::Tensor &, at::OptionalIntArrayRef, bool, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanmean"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "nanmean.out(Tensor self, int[1]? dim=None, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/nanmedian_ops.h b/backends/pytorch/include/ATen/ops/nanmedian_ops.h new file mode 100644 index 000000000..bb03c3a86 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/nanmedian_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API nanmedian { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanmedian"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nanmedian(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API nanmedian_dim { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanmedian"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "nanmedian.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim); +}; + +struct TORCH_API nanmedian_dim_values { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanmedian"; + static constexpr const char* overload_name = "dim_values"; + static constexpr const char* schema_str = "nanmedian.dim_values(Tensor self, int dim, bool keepdim=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API nanmedian_names_dim { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanmedian"; + static constexpr const char* overload_name = "names_dim"; + static constexpr const char* schema_str = "nanmedian.names_dim(Tensor self, Dimname dim, bool keepdim=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool keepdim); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim); +}; + +struct TORCH_API nanmedian_names_dim_values { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanmedian"; + static constexpr const char* overload_name = "names_dim_values"; + static constexpr const char* schema_str = "nanmedian.names_dim_values(Tensor self, Dimname dim, bool keepdim=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API nanmedian_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanmedian"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "nanmedian.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/nanquantile_ops.h b/backends/pytorch/include/ATen/ops/nanquantile_ops.h new file mode 100644 index 000000000..3910aa82b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/nanquantile_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API nanquantile { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, ::std::optional, bool, c10::string_view); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanquantile"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nanquantile(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & q, ::std::optional dim, bool keepdim, c10::string_view interpolation); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & q, ::std::optional dim, bool keepdim, c10::string_view interpolation); +}; + +struct TORCH_API nanquantile_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, ::std::optional, bool, c10::string_view, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanquantile"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "nanquantile.out(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, str interpolation='linear', Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & q, ::std::optional dim, bool keepdim, c10::string_view interpolation, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & q, ::std::optional dim, bool keepdim, c10::string_view interpolation, at::Tensor & out); +}; + +struct TORCH_API nanquantile_scalar { + using schema = at::Tensor (const at::Tensor &, double, ::std::optional, bool, c10::string_view); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanquantile"; + static constexpr const char* overload_name = "scalar"; + static constexpr const char* schema_str = "nanquantile.scalar(Tensor self, float q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor"; + static at::Tensor call(const at::Tensor & self, double q, ::std::optional dim, bool keepdim, c10::string_view interpolation); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double q, ::std::optional dim, bool keepdim, c10::string_view interpolation); +}; + +struct TORCH_API nanquantile_scalar_out { + using schema = at::Tensor & (const at::Tensor &, double, ::std::optional, bool, c10::string_view, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nanquantile"; + static constexpr const char* overload_name = "scalar_out"; + static constexpr const char* schema_str = "nanquantile.scalar_out(Tensor self, float q, int? dim=None, bool keepdim=False, *, str interpolation='linear', Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, double q, ::std::optional dim, bool keepdim, c10::string_view interpolation, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double q, ::std::optional dim, bool keepdim, c10::string_view interpolation, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/nansum_ops.h b/backends/pytorch/include/ATen/ops/nansum_ops.h new file mode 100644 index 000000000..a4ec056cd --- /dev/null +++ b/backends/pytorch/include/ATen/ops/nansum_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API nansum { + using schema = at::Tensor (const at::Tensor &, at::OptionalIntArrayRef, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nansum"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nansum(Tensor self, int[1]? dim=None, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype); +}; + +struct TORCH_API nansum_out { + using schema = at::Tensor & (const at::Tensor &, at::OptionalIntArrayRef, bool, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nansum"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "nansum.out(Tensor self, int[1]? dim=None, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/narrow_copy_ops.h b/backends/pytorch/include/ATen/ops/narrow_copy_ops.h new file mode 100644 index 000000000..2b570ae75 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/narrow_copy_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API narrow_copy { + using schema = at::Tensor (const at::Tensor &, int64_t, c10::SymInt, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::narrow_copy"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "narrow_copy(Tensor self, int dim, SymInt start, SymInt length) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, c10::SymInt start, c10::SymInt length); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, c10::SymInt start, c10::SymInt length); +}; + +struct TORCH_API narrow_copy_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, c10::SymInt, c10::SymInt, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::narrow_copy"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "narrow_copy.out(Tensor self, int dim, SymInt start, SymInt length, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, c10::SymInt start, c10::SymInt length, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, c10::SymInt start, c10::SymInt length, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/narrow_ops.h b/backends/pytorch/include/ATen/ops/narrow_ops.h new file mode 100644 index 000000000..29fc0a495 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/narrow_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API narrow { + using schema = at::Tensor (const at::Tensor &, int64_t, c10::SymInt, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::narrow"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "narrow(Tensor(a) self, int dim, SymInt start, SymInt length) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t dim, c10::SymInt start, c10::SymInt length); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, c10::SymInt start, c10::SymInt length); +}; + +struct TORCH_API narrow_Tensor { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::narrow"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "narrow.Tensor(Tensor(a) self, int dim, Tensor start, SymInt length) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & start, c10::SymInt length); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & start, c10::SymInt length); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/ne_ops.h b/backends/pytorch/include/ATen/ops/ne_ops.h new file mode 100644 index 000000000..441572137 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/ne_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API ne_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ne"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "ne.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API ne_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ne"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "ne.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API ne_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ne"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "ne.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API ne_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ne"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "ne.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API ne__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ne_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "ne_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API ne__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ne_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "ne_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/neg_ops.h b/backends/pytorch/include/ATen/ops/neg_ops.h new file mode 100644 index 000000000..7915844db --- /dev/null +++ b/backends/pytorch/include/ATen/ops/neg_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API neg { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::neg"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "neg(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API neg_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::neg_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "neg_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API neg_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::neg"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "neg.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/negative_ops.h b/backends/pytorch/include/ATen/ops/negative_ops.h new file mode 100644 index 000000000..cbe1fc736 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/negative_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API negative { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::negative"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "negative(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API negative_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::negative_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "negative_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API negative_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::negative"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "negative.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/new_empty_ops.h b/backends/pytorch/include/ATen/ops/new_empty_ops.h new file mode 100644 index 000000000..32a3d169d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/new_empty_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API new_empty { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef, ::std::optional, ::std::optional, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::new_empty"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "new_empty(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); +}; + +struct TORCH_API new_empty_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymIntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::new_empty"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "new_empty.out(Tensor self, SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymIntArrayRef size, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/new_empty_strided_ops.h b/backends/pytorch/include/ATen/ops/new_empty_strided_ops.h new file mode 100644 index 000000000..8ce923376 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/new_empty_strided_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API new_empty_strided { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef, c10::SymIntArrayRef, ::std::optional, ::std::optional, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::new_empty_strided"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "new_empty_strided(Tensor self, SymInt[] size, SymInt[] stride, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); +}; + +struct TORCH_API new_empty_strided_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymIntArrayRef, c10::SymIntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::new_empty_strided"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "new_empty_strided.out(Tensor self, SymInt[] size, SymInt[] stride, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/new_full_ops.h b/backends/pytorch/include/ATen/ops/new_full_ops.h new file mode 100644 index 000000000..741a6352c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/new_full_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API new_full { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef, const at::Scalar &, ::std::optional, ::std::optional, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::new_full"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "new_full(Tensor self, SymInt[] size, Scalar fill_value, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size, const at::Scalar & fill_value, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, const at::Scalar & fill_value, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); +}; + +struct TORCH_API new_full_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymIntArrayRef, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::new_full"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "new_full.out(Tensor self, SymInt[] size, Scalar fill_value, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymIntArrayRef size, const at::Scalar & fill_value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, const at::Scalar & fill_value, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/new_ones_ops.h b/backends/pytorch/include/ATen/ops/new_ones_ops.h new file mode 100644 index 000000000..75d558f74 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/new_ones_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API new_ones { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef, ::std::optional, ::std::optional, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::new_ones"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "new_ones(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); +}; + +struct TORCH_API new_ones_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymIntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::new_ones"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "new_ones.out(Tensor self, SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymIntArrayRef size, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/new_zeros_ops.h b/backends/pytorch/include/ATen/ops/new_zeros_ops.h new file mode 100644 index 000000000..4dce875da --- /dev/null +++ b/backends/pytorch/include/ATen/ops/new_zeros_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API new_zeros { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef, ::std::optional, ::std::optional, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::new_zeros"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "new_zeros(Tensor self, SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); +}; + +struct TORCH_API new_zeros_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymIntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::new_zeros"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "new_zeros.out(Tensor self, SymInt[] size, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymIntArrayRef size, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/nextafter_ops.h b/backends/pytorch/include/ATen/ops/nextafter_ops.h new file mode 100644 index 000000000..3e8942097 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/nextafter_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API nextafter_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nextafter"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "nextafter.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API nextafter { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nextafter"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nextafter(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API nextafter_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nextafter_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nextafter_(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/nonzero_numpy_ops.h b/backends/pytorch/include/ATen/ops/nonzero_numpy_ops.h new file mode 100644 index 000000000..daa4908b7 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/nonzero_numpy_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API nonzero_numpy { + using schema = ::std::vector (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nonzero_numpy"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nonzero_numpy(Tensor self) -> Tensor[]"; + static ::std::vector call(const at::Tensor & self); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/nonzero_ops.h b/backends/pytorch/include/ATen/ops/nonzero_ops.h new file mode 100644 index 000000000..9537c6dc2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/nonzero_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API nonzero_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nonzero"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "nonzero.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API nonzero { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nonzero"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nonzero(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/nonzero_static_ops.h b/backends/pytorch/include/ATen/ops/nonzero_static_ops.h new file mode 100644 index 000000000..5482710f9 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/nonzero_static_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API nonzero_static_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymInt, int64_t, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nonzero_static"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "nonzero_static.out(Tensor self, *, SymInt size, int fill_value=-1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymInt size, int64_t fill_value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt size, int64_t fill_value, at::Tensor & out); +}; + +struct TORCH_API nonzero_static { + using schema = at::Tensor (const at::Tensor &, c10::SymInt, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::nonzero_static"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "nonzero_static(Tensor self, *, SymInt size, int fill_value=-1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymInt size, int64_t fill_value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt size, int64_t fill_value); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/norm_ops.h b/backends/pytorch/include/ATen/ops/norm_ops.h new file mode 100644 index 000000000..a14b7ff37 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/norm_ops.h @@ -0,0 +1,155 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API norm_ScalarOpt_dtype { + using schema = at::Tensor (const at::Tensor &, const ::std::optional &, at::ScalarType); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "ScalarOpt_dtype"; + static constexpr const char* schema_str = "norm.ScalarOpt_dtype(Tensor self, Scalar? p, *, ScalarType dtype) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const ::std::optional & p, at::ScalarType dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & p, at::ScalarType dtype); +}; + +struct TORCH_API norm_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "norm.Scalar(Tensor self, Scalar p=2) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & p); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & p); +}; + +struct TORCH_API norm_ScalarOpt_dim_dtype { + using schema = at::Tensor (const at::Tensor &, const ::std::optional &, at::IntArrayRef, bool, at::ScalarType); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "ScalarOpt_dim_dtype"; + static constexpr const char* schema_str = "norm.ScalarOpt_dim_dtype(Tensor self, Scalar? p, int[1] dim, bool keepdim, *, ScalarType dtype) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const ::std::optional & p, at::IntArrayRef dim, bool keepdim, at::ScalarType dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & p, at::IntArrayRef dim, bool keepdim, at::ScalarType dtype); +}; + +struct TORCH_API norm_ScalarOpt_dim { + using schema = at::Tensor (const at::Tensor &, const ::std::optional &, at::IntArrayRef, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "ScalarOpt_dim"; + static constexpr const char* schema_str = "norm.ScalarOpt_dim(Tensor self, Scalar? p, int[1] dim, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const ::std::optional & p, at::IntArrayRef dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & p, at::IntArrayRef dim, bool keepdim); +}; + +struct TORCH_API norm_dtype_out { + using schema = at::Tensor & (const at::Tensor &, const ::std::optional &, at::IntArrayRef, bool, at::ScalarType, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "dtype_out"; + static constexpr const char* schema_str = "norm.dtype_out(Tensor self, Scalar? p, int[1] dim, bool keepdim, *, ScalarType dtype, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const ::std::optional & p, at::IntArrayRef dim, bool keepdim, at::ScalarType dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & p, at::IntArrayRef dim, bool keepdim, at::ScalarType dtype, at::Tensor & out); +}; + +struct TORCH_API norm_out { + using schema = at::Tensor & (const at::Tensor &, const ::std::optional &, at::IntArrayRef, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "norm.out(Tensor self, Scalar? p, int[1] dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const ::std::optional & p, at::IntArrayRef dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & p, at::IntArrayRef dim, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API norm_names_ScalarOpt_dim_dtype { + using schema = at::Tensor (const at::Tensor &, const ::std::optional &, at::DimnameList, bool, at::ScalarType); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "names_ScalarOpt_dim_dtype"; + static constexpr const char* schema_str = "norm.names_ScalarOpt_dim_dtype(Tensor self, Scalar? p, Dimname[1] dim, bool keepdim, *, ScalarType dtype) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const ::std::optional & p, at::DimnameList dim, bool keepdim, at::ScalarType dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & p, at::DimnameList dim, bool keepdim, at::ScalarType dtype); +}; + +struct TORCH_API norm_names_ScalarOpt_dim { + using schema = at::Tensor (const at::Tensor &, const ::std::optional &, at::DimnameList, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "names_ScalarOpt_dim"; + static constexpr const char* schema_str = "norm.names_ScalarOpt_dim(Tensor self, Scalar? p, Dimname[1] dim, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const ::std::optional & p, at::DimnameList dim, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & p, at::DimnameList dim, bool keepdim); +}; + +struct TORCH_API norm_names_dtype_out { + using schema = at::Tensor & (const at::Tensor &, const ::std::optional &, at::DimnameList, bool, at::ScalarType, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "names_dtype_out"; + static constexpr const char* schema_str = "norm.names_dtype_out(Tensor self, Scalar? p, Dimname[1] dim, bool keepdim, *, ScalarType dtype, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const ::std::optional & p, at::DimnameList dim, bool keepdim, at::ScalarType dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & p, at::DimnameList dim, bool keepdim, at::ScalarType dtype, at::Tensor & out); +}; + +struct TORCH_API norm_names_out { + using schema = at::Tensor & (const at::Tensor &, const ::std::optional &, at::DimnameList, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "names_out"; + static constexpr const char* schema_str = "norm.names_out(Tensor self, Scalar? p, Dimname[1] dim, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const ::std::optional & p, at::DimnameList dim, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & p, at::DimnameList dim, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API norm_ScalarOpt_dtype_out { + using schema = at::Tensor & (const at::Tensor &, const ::std::optional &, at::ScalarType, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "ScalarOpt_dtype_out"; + static constexpr const char* schema_str = "norm.ScalarOpt_dtype_out(Tensor self, Scalar? p, *, ScalarType dtype, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const ::std::optional & p, at::ScalarType dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const ::std::optional & p, at::ScalarType dtype, at::Tensor & out); +}; + +struct TORCH_API norm_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::norm"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "norm.Scalar_out(Tensor self, Scalar p=2, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & p, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & p, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/normal_ops.h b/backends/pytorch/include/ATen/ops/normal_ops.h new file mode 100644 index 000000000..4c61ff79b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/normal_ops.h @@ -0,0 +1,144 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API normal_ { + using schema = at::Tensor & (at::Tensor &, double, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "normal_(Tensor(a!) self, float mean=0, float std=1, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, double mean, double std, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, double mean, double std, ::std::optional generator); +}; + +struct TORCH_API normal_functional { + using schema = at::Tensor (const at::Tensor &, double, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal_functional"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "normal_functional(Tensor self, float mean=0, float std=1, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, double mean, double std, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double mean, double std, ::std::optional generator); +}; + +struct TORCH_API normal_Tensor_float_out { + using schema = at::Tensor & (const at::Tensor &, double, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal"; + static constexpr const char* overload_name = "Tensor_float_out"; + static constexpr const char* schema_str = "normal.Tensor_float_out(Tensor mean, float std=1, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & mean, double std, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & mean, double std, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API normal_Tensor_float { + using schema = at::Tensor (const at::Tensor &, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal"; + static constexpr const char* overload_name = "Tensor_float"; + static constexpr const char* schema_str = "normal.Tensor_float(Tensor mean, float std=1, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & mean, double std, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & mean, double std, ::std::optional generator); +}; + +struct TORCH_API normal_float_Tensor_out { + using schema = at::Tensor & (double, const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal"; + static constexpr const char* overload_name = "float_Tensor_out"; + static constexpr const char* schema_str = "normal.float_Tensor_out(float mean, Tensor std, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(double mean, const at::Tensor & std, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, double mean, const at::Tensor & std, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API normal_float_Tensor { + using schema = at::Tensor (double, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal"; + static constexpr const char* overload_name = "float_Tensor"; + static constexpr const char* schema_str = "normal.float_Tensor(float mean, Tensor std, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(double mean, const at::Tensor & std, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, double mean, const at::Tensor & std, ::std::optional generator); +}; + +struct TORCH_API normal_Tensor_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal"; + static constexpr const char* overload_name = "Tensor_Tensor_out"; + static constexpr const char* schema_str = "normal.Tensor_Tensor_out(Tensor mean, Tensor std, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & mean, const at::Tensor & std, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & mean, const at::Tensor & std, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API normal_Tensor_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal"; + static constexpr const char* overload_name = "Tensor_Tensor"; + static constexpr const char* schema_str = "normal.Tensor_Tensor(Tensor mean, Tensor std, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & mean, const at::Tensor & std, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & mean, const at::Tensor & std, ::std::optional generator); +}; + +struct TORCH_API normal_float_float { + using schema = at::Tensor (double, double, c10::SymIntArrayRef, ::std::optional, ::std::optional, ::std::optional, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal"; + static constexpr const char* overload_name = "float_float"; + static constexpr const char* schema_str = "normal.float_float(float mean, float std, SymInt[] size, *, Generator? generator=None, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor"; + static at::Tensor call(double mean, double std, c10::SymIntArrayRef size, ::std::optional generator, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, double mean, double std, c10::SymIntArrayRef size, ::std::optional generator, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory); +}; + +struct TORCH_API normal_float_float_out { + using schema = at::Tensor & (double, double, c10::SymIntArrayRef, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal"; + static constexpr const char* overload_name = "float_float_out"; + static constexpr const char* schema_str = "normal.float_float_out(float mean, float std, SymInt[] size, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(double mean, double std, c10::SymIntArrayRef size, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, double mean, double std, c10::SymIntArrayRef size, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API normal_out { + using schema = at::Tensor & (const at::Tensor &, double, double, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::normal"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "normal.out(Tensor self, float mean=0, float std=1, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, double mean, double std, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double mean, double std, ::std::optional generator, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/not_equal_ops.h b/backends/pytorch/include/ATen/ops/not_equal_ops.h new file mode 100644 index 000000000..472b56b90 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/not_equal_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API not_equal_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::not_equal"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "not_equal.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API not_equal_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::not_equal"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "not_equal.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API not_equal_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::not_equal"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "not_equal.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API not_equal_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::not_equal"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "not_equal.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API not_equal__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::not_equal_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "not_equal_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API not_equal__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::not_equal_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "not_equal_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/numpy_T_ops.h b/backends/pytorch/include/ATen/ops/numpy_T_ops.h new file mode 100644 index 000000000..a0dedb30d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/numpy_T_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API numpy_T { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::numpy_T"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "numpy_T(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/or_ops.h b/backends/pytorch/include/ATen/ops/or_ops.h new file mode 100644 index 000000000..5f74e0dd6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/or_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API __or___Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__or__"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "__or__.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API __or___Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__or__"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "__or__.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API __ior___Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__ior__"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "__ior__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API __ior___Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__ior__"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "__ior__.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/orgqr_ops.h b/backends/pytorch/include/ATen/ops/orgqr_ops.h new file mode 100644 index 000000000..145c518d8 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/orgqr_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API orgqr { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::orgqr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "orgqr(Tensor self, Tensor input2) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & input2); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & input2); +}; + +struct TORCH_API orgqr_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::orgqr"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "orgqr.out(Tensor self, Tensor input2, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & input2, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & input2, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/ormqr_ops.h b/backends/pytorch/include/ATen/ops/ormqr_ops.h new file mode 100644 index 000000000..4899fc120 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/ormqr_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API ormqr_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, bool, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ormqr"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "ormqr.out(Tensor self, Tensor input2, Tensor input3, bool left=True, bool transpose=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & input2, const at::Tensor & input3, bool left, bool transpose, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & input2, const at::Tensor & input3, bool left, bool transpose, at::Tensor & out); +}; + +struct TORCH_API ormqr { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, bool, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ormqr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "ormqr(Tensor self, Tensor input2, Tensor input3, bool left=True, bool transpose=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & input2, const at::Tensor & input3, bool left, bool transpose); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & input2, const at::Tensor & input3, bool left, bool transpose); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/outer_ops.h b/backends/pytorch/include/ATen/ops/outer_ops.h new file mode 100644 index 000000000..4efbdb16d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/outer_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API outer { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::outer"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "outer(Tensor self, Tensor vec2) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & vec2); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & vec2); +}; + +struct TORCH_API outer_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::outer"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "outer.out(Tensor self, Tensor vec2, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & vec2, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & vec2, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/output_nr_ops.h b/backends/pytorch/include/ATen/ops/output_nr_ops.h new file mode 100644 index 000000000..929fe45bd --- /dev/null +++ b/backends/pytorch/include/ATen/ops/output_nr_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API output_nr { + using schema = int64_t (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::output_nr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "output_nr(Tensor self) -> int"; + static int64_t call(const at::Tensor & self); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/permute_ops.h b/backends/pytorch/include/ATen/ops/permute_ops.h new file mode 100644 index 000000000..1cd3ab218 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/permute_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API permute { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::permute"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "permute(Tensor(a) self, int[] dims) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef dims); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dims); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/pin_memory_ops.h b/backends/pytorch/include/ATen/ops/pin_memory_ops.h new file mode 100644 index 000000000..0d88e78a9 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/pin_memory_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API pin_memory { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::pin_memory"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "pin_memory(Tensor(a) self, Device? device=None) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, ::std::optional device); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional device); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/pinverse_ops.h b/backends/pytorch/include/ATen/ops/pinverse_ops.h new file mode 100644 index 000000000..7d97fec0f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/pinverse_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API pinverse { + using schema = at::Tensor (const at::Tensor &, double); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::pinverse"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "pinverse(Tensor self, float rcond=1e-15) -> Tensor"; + static at::Tensor call(const at::Tensor & self, double rcond); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double rcond); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/polygamma_ops.h b/backends/pytorch/include/ATen/ops/polygamma_ops.h new file mode 100644 index 000000000..156c7d40d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/polygamma_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API polygamma_out { + using schema = at::Tensor & (int64_t, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::polygamma"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "polygamma.out(int n, Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(int64_t n, const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, int64_t n, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API polygamma { + using schema = at::Tensor (int64_t, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::polygamma"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "polygamma(int n, Tensor self) -> Tensor"; + static at::Tensor call(int64_t n, const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, int64_t n, const at::Tensor & self); +}; + +struct TORCH_API polygamma_ { + using schema = at::Tensor & (at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::polygamma_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "polygamma_(Tensor(a!) self, int n) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t n); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t n); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/positive_ops.h b/backends/pytorch/include/ATen/ops/positive_ops.h new file mode 100644 index 000000000..7299086c6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/positive_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API positive { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::positive"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "positive(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/pow_ops.h b/backends/pytorch/include/ATen/ops/pow_ops.h new file mode 100644 index 000000000..44bf09bc1 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/pow_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API pow_Tensor_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::pow"; + static constexpr const char* overload_name = "Tensor_Tensor_out"; + static constexpr const char* schema_str = "pow.Tensor_Tensor_out(Tensor self, Tensor exponent, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & exponent, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & exponent, at::Tensor & out); +}; + +struct TORCH_API pow_Tensor_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::pow"; + static constexpr const char* overload_name = "Tensor_Tensor"; + static constexpr const char* schema_str = "pow.Tensor_Tensor(Tensor self, Tensor exponent) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & exponent); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & exponent); +}; + +struct TORCH_API pow_Scalar_out { + using schema = at::Tensor & (const at::Scalar &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::pow"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "pow.Scalar_out(Scalar self, Tensor exponent, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Scalar & self, const at::Tensor & exponent, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & exponent, at::Tensor & out); +}; + +struct TORCH_API pow_Scalar { + using schema = at::Tensor (const at::Scalar &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::pow"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "pow.Scalar(Scalar self, Tensor exponent) -> Tensor"; + static at::Tensor call(const at::Scalar & self, const at::Tensor & exponent); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & exponent); +}; + +struct TORCH_API pow_Tensor_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::pow"; + static constexpr const char* overload_name = "Tensor_Scalar_out"; + static constexpr const char* schema_str = "pow.Tensor_Scalar_out(Tensor self, Scalar exponent, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & exponent, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & exponent, at::Tensor & out); +}; + +struct TORCH_API pow_Tensor_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::pow"; + static constexpr const char* overload_name = "Tensor_Scalar"; + static constexpr const char* schema_str = "pow.Tensor_Scalar(Tensor self, Scalar exponent) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & exponent); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & exponent); +}; + +struct TORCH_API pow__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::pow_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "pow_.Scalar(Tensor(a!) self, Scalar exponent) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & exponent); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & exponent); +}; + +struct TORCH_API pow__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::pow_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "pow_.Tensor(Tensor(a!) self, Tensor exponent) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & exponent); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & exponent); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/prelu_ops.h b/backends/pytorch/include/ATen/ops/prelu_ops.h new file mode 100644 index 000000000..e6e3ad1bd --- /dev/null +++ b/backends/pytorch/include/ATen/ops/prelu_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API prelu { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::prelu"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "prelu(Tensor self, Tensor weight) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & weight); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & weight); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/prod_ops.h b/backends/pytorch/include/ATen/ops/prod_ops.h new file mode 100644 index 000000000..030a8f0c7 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/prod_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API prod { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::prod"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "prod(Tensor self, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype); +}; + +struct TORCH_API prod_dim_int { + using schema = at::Tensor (const at::Tensor &, int64_t, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::prod"; + static constexpr const char* overload_name = "dim_int"; + static constexpr const char* schema_str = "prod.dim_int(Tensor self, int dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, bool keepdim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim, ::std::optional dtype); +}; + +struct TORCH_API prod_int_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, bool, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::prod"; + static constexpr const char* overload_name = "int_out"; + static constexpr const char* schema_str = "prod.int_out(Tensor self, int dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, bool keepdim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool keepdim, ::std::optional dtype, at::Tensor & out); +}; + +struct TORCH_API prod_dim_Dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::prod"; + static constexpr const char* overload_name = "dim_Dimname"; + static constexpr const char* schema_str = "prod.dim_Dimname(Tensor self, Dimname dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, bool keepdim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim, ::std::optional dtype); +}; + +struct TORCH_API prod_Dimname_out { + using schema = at::Tensor & (const at::Tensor &, at::Dimname, bool, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::prod"; + static constexpr const char* overload_name = "Dimname_out"; + static constexpr const char* schema_str = "prod.Dimname_out(Tensor self, Dimname dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Dimname dim, bool keepdim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool keepdim, ::std::optional dtype, at::Tensor & out); +}; + +struct TORCH_API prod_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::prod"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "prod.out(Tensor self, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/put_ops.h b/backends/pytorch/include/ATen/ops/put_ops.h new file mode 100644 index 000000000..b8823b5ae --- /dev/null +++ b/backends/pytorch/include/ATen/ops/put_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API put_ { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::put_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "put_(Tensor(a!) self, Tensor index, Tensor source, bool accumulate=False) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & index, const at::Tensor & source, bool accumulate); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & index, const at::Tensor & source, bool accumulate); +}; + +struct TORCH_API put { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::put"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "put(Tensor self, Tensor index, Tensor source, bool accumulate=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & index, const at::Tensor & source, bool accumulate); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & index, const at::Tensor & source, bool accumulate); +}; + +struct TORCH_API put_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::put"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "put.out(Tensor self, Tensor index, Tensor source, bool accumulate=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & index, const at::Tensor & source, bool accumulate, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & index, const at::Tensor & source, bool accumulate, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/q_per_channel_axis_ops.h b/backends/pytorch/include/ATen/ops/q_per_channel_axis_ops.h new file mode 100644 index 000000000..a39cb13bc --- /dev/null +++ b/backends/pytorch/include/ATen/ops/q_per_channel_axis_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API q_per_channel_axis { + using schema = int64_t (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::q_per_channel_axis"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "q_per_channel_axis(Tensor self) -> int"; + static int64_t call(const at::Tensor & self); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/q_per_channel_scales_ops.h b/backends/pytorch/include/ATen/ops/q_per_channel_scales_ops.h new file mode 100644 index 000000000..acf8ee9fc --- /dev/null +++ b/backends/pytorch/include/ATen/ops/q_per_channel_scales_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API q_per_channel_scales { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::q_per_channel_scales"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "q_per_channel_scales(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API q_per_channel_scales_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::q_per_channel_scales"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "q_per_channel_scales.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/q_per_channel_zero_points_ops.h b/backends/pytorch/include/ATen/ops/q_per_channel_zero_points_ops.h new file mode 100644 index 000000000..0fd82edca --- /dev/null +++ b/backends/pytorch/include/ATen/ops/q_per_channel_zero_points_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API q_per_channel_zero_points { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::q_per_channel_zero_points"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "q_per_channel_zero_points(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API q_per_channel_zero_points_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::q_per_channel_zero_points"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "q_per_channel_zero_points.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/q_scale_ops.h b/backends/pytorch/include/ATen/ops/q_scale_ops.h new file mode 100644 index 000000000..25b82ea77 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/q_scale_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API q_scale { + using schema = double (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::q_scale"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "q_scale(Tensor self) -> float"; + static double call(const at::Tensor & self); + static double redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/q_zero_point_ops.h b/backends/pytorch/include/ATen/ops/q_zero_point_ops.h new file mode 100644 index 000000000..78db1add5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/q_zero_point_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API q_zero_point { + using schema = int64_t (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::q_zero_point"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "q_zero_point(Tensor self) -> int"; + static int64_t call(const at::Tensor & self); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/qr_ops.h b/backends/pytorch/include/ATen/ops/qr_ops.h new file mode 100644 index 000000000..1c833024c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/qr_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API qr_Q { + using schema = ::std::tuple (const at::Tensor &, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::qr"; + static constexpr const char* overload_name = "Q"; + static constexpr const char* schema_str = "qr.Q(Tensor self, bool some=True, *, Tensor(a!) Q, Tensor(b!) R) -> (Tensor(a!) Q, Tensor(b!) R)"; + static ::std::tuple call(const at::Tensor & self, bool some, at::Tensor & Q, at::Tensor & R); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool some, at::Tensor & Q, at::Tensor & R); +}; + +struct TORCH_API qr { + using schema = ::std::tuple (const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::qr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "qr(Tensor self, bool some=True) -> (Tensor Q, Tensor R)"; + static ::std::tuple call(const at::Tensor & self, bool some); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool some); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/qscheme_ops.h b/backends/pytorch/include/ATen/ops/qscheme_ops.h new file mode 100644 index 000000000..9073d8009 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/qscheme_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API qscheme { + using schema = at::QScheme (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::qscheme"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "qscheme(Tensor self) -> QScheme"; + static at::QScheme call(const at::Tensor & self); + static at::QScheme redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/quantile_ops.h b/backends/pytorch/include/ATen/ops/quantile_ops.h new file mode 100644 index 000000000..337805bbf --- /dev/null +++ b/backends/pytorch/include/ATen/ops/quantile_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API quantile { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, ::std::optional, bool, c10::string_view); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::quantile"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "quantile(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & q, ::std::optional dim, bool keepdim, c10::string_view interpolation); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & q, ::std::optional dim, bool keepdim, c10::string_view interpolation); +}; + +struct TORCH_API quantile_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, ::std::optional, bool, c10::string_view, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::quantile"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "quantile.out(Tensor self, Tensor q, int? dim=None, bool keepdim=False, *, str interpolation='linear', Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & q, ::std::optional dim, bool keepdim, c10::string_view interpolation, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & q, ::std::optional dim, bool keepdim, c10::string_view interpolation, at::Tensor & out); +}; + +struct TORCH_API quantile_scalar { + using schema = at::Tensor (const at::Tensor &, double, ::std::optional, bool, c10::string_view); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::quantile"; + static constexpr const char* overload_name = "scalar"; + static constexpr const char* schema_str = "quantile.scalar(Tensor self, float q, int? dim=None, bool keepdim=False, *, str interpolation='linear') -> Tensor"; + static at::Tensor call(const at::Tensor & self, double q, ::std::optional dim, bool keepdim, c10::string_view interpolation); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double q, ::std::optional dim, bool keepdim, c10::string_view interpolation); +}; + +struct TORCH_API quantile_scalar_out { + using schema = at::Tensor & (const at::Tensor &, double, ::std::optional, bool, c10::string_view, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::quantile"; + static constexpr const char* overload_name = "scalar_out"; + static constexpr const char* schema_str = "quantile.scalar_out(Tensor self, float q, int? dim=None, bool keepdim=False, *, str interpolation='linear', Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, double q, ::std::optional dim, bool keepdim, c10::string_view interpolation, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double q, ::std::optional dim, bool keepdim, c10::string_view interpolation, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/rad2deg_ops.h b/backends/pytorch/include/ATen/ops/rad2deg_ops.h new file mode 100644 index 000000000..4d263ceda --- /dev/null +++ b/backends/pytorch/include/ATen/ops/rad2deg_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API rad2deg { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::rad2deg"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "rad2deg(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API rad2deg_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::rad2deg_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "rad2deg_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API rad2deg_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::rad2deg"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "rad2deg.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/random_ops.h b/backends/pytorch/include/ATen/ops/random_ops.h new file mode 100644 index 000000000..2910e9d04 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/random_ops.h @@ -0,0 +1,122 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API random__from { + using schema = at::Tensor & (at::Tensor &, int64_t, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::random_"; + static constexpr const char* overload_name = "from"; + static constexpr const char* schema_str = "random_.from(Tensor(a!) self, int from, int? to, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t from, ::std::optional to, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t from, ::std::optional to, ::std::optional generator); +}; + +struct TORCH_API random__to { + using schema = at::Tensor & (at::Tensor &, int64_t, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::random_"; + static constexpr const char* overload_name = "to"; + static constexpr const char* schema_str = "random_.to(Tensor(a!) self, int to, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t to, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t to, ::std::optional generator); +}; + +struct TORCH_API random_ { + using schema = at::Tensor & (at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::random_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "random_(Tensor(a!) self, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, ::std::optional generator); +}; + +struct TORCH_API random_from_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, ::std::optional, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::random"; + static constexpr const char* overload_name = "from_out"; + static constexpr const char* schema_str = "random.from_out(Tensor self, int from, int? to, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t from, ::std::optional to, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t from, ::std::optional to, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API random_from { + using schema = at::Tensor (const at::Tensor &, int64_t, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::random"; + static constexpr const char* overload_name = "from"; + static constexpr const char* schema_str = "random.from(Tensor self, int from, int? to, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t from, ::std::optional to, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t from, ::std::optional to, ::std::optional generator); +}; + +struct TORCH_API random_to_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::random"; + static constexpr const char* overload_name = "to_out"; + static constexpr const char* schema_str = "random.to_out(Tensor self, int to, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t to, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t to, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API random_to { + using schema = at::Tensor (const at::Tensor &, int64_t, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::random"; + static constexpr const char* overload_name = "to"; + static constexpr const char* schema_str = "random.to(Tensor self, int to, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t to, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t to, ::std::optional generator); +}; + +struct TORCH_API random_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::random"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "random.out(Tensor self, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API random { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::random"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "random(Tensor self, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional generator); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/ravel_ops.h b/backends/pytorch/include/ATen/ops/ravel_ops.h new file mode 100644 index 000000000..439ae73db --- /dev/null +++ b/backends/pytorch/include/ATen/ops/ravel_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API ravel { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::ravel"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "ravel(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/reciprocal_ops.h b/backends/pytorch/include/ATen/ops/reciprocal_ops.h new file mode 100644 index 000000000..ec9f1fe86 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/reciprocal_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API reciprocal { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::reciprocal"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "reciprocal(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API reciprocal_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::reciprocal_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "reciprocal_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API reciprocal_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::reciprocal"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "reciprocal.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/record_stream_ops.h b/backends/pytorch/include/ATen/ops/record_stream_ops.h new file mode 100644 index 000000000..50bd0acc1 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/record_stream_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API record_stream { + using schema = void (at::Tensor &, at::Stream); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::record_stream"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "record_stream(Tensor(a!) self, Stream s) -> ()"; + static void call(at::Tensor & self, at::Stream s); + static void redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, at::Stream s); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/refine_names_ops.h b/backends/pytorch/include/ATen/ops/refine_names_ops.h new file mode 100644 index 000000000..aa579bb8e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/refine_names_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API refine_names { + using schema = at::Tensor (const at::Tensor &, at::DimnameList); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::refine_names"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "refine_names(Tensor(a) self, Dimname[] names) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::DimnameList names); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList names); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/relu_ops.h b/backends/pytorch/include/ATen/ops/relu_ops.h new file mode 100644 index 000000000..4bc5dff24 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/relu_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API relu { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::relu"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "relu(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API relu_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::relu_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "relu_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API relu_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::relu"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "relu.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/remainder_ops.h b/backends/pytorch/include/ATen/ops/remainder_ops.h new file mode 100644 index 000000000..f39156d55 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/remainder_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API remainder_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::remainder"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "remainder.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API remainder_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::remainder"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "remainder.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API remainder__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::remainder_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "remainder_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API remainder_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::remainder"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "remainder.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API remainder_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::remainder"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "remainder.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API remainder__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::remainder_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "remainder_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API remainder_Scalar_Tensor { + using schema = at::Tensor (const at::Scalar &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::remainder"; + static constexpr const char* overload_name = "Scalar_Tensor"; + static constexpr const char* schema_str = "remainder.Scalar_Tensor(Scalar self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Scalar & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other); +}; + +struct TORCH_API remainder_Scalar_Tensor_out { + using schema = at::Tensor & (const at::Scalar &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::remainder"; + static constexpr const char* overload_name = "Scalar_Tensor_out"; + static constexpr const char* schema_str = "remainder.Scalar_Tensor_out(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Scalar & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/rename_ops.h b/backends/pytorch/include/ATen/ops/rename_ops.h new file mode 100644 index 000000000..40e63700f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/rename_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API rename_ { + using schema = at::Tensor & (at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::rename_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "rename_(Tensor(a!) self, Dimname[]? names) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, ::std::optional names); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, ::std::optional names); +}; + +struct TORCH_API rename { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::rename"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "rename(Tensor(a) self, Dimname[]? names) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, ::std::optional names); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional names); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/renorm_ops.h b/backends/pytorch/include/ATen/ops/renorm_ops.h new file mode 100644 index 000000000..6d40801fa --- /dev/null +++ b/backends/pytorch/include/ATen/ops/renorm_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API renorm_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, int64_t, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::renorm"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "renorm.out(Tensor self, Scalar p, int dim, Scalar maxnorm, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm, at::Tensor & out); +}; + +struct TORCH_API renorm { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &, int64_t, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::renorm"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "renorm(Tensor self, Scalar p, int dim, Scalar maxnorm) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm); +}; + +struct TORCH_API renorm_ { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &, int64_t, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::renorm_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "renorm_(Tensor(a!) self, Scalar p, int dim, Scalar maxnorm) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & p, int64_t dim, const at::Scalar & maxnorm); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/repeat_interleave_ops.h b/backends/pytorch/include/ATen/ops/repeat_interleave_ops.h new file mode 100644 index 000000000..a994ac0ab --- /dev/null +++ b/backends/pytorch/include/ATen/ops/repeat_interleave_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API repeat_interleave_Tensor { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::repeat_interleave"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "repeat_interleave.Tensor(Tensor repeats, *, SymInt? output_size=None) -> Tensor"; + static at::Tensor call(const at::Tensor & repeats, ::std::optional output_size); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & repeats, ::std::optional output_size); +}; + +struct TORCH_API repeat_interleave_self_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::repeat_interleave"; + static constexpr const char* overload_name = "self_Tensor"; + static constexpr const char* schema_str = "repeat_interleave.self_Tensor(Tensor self, Tensor repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & repeats, ::std::optional dim, ::std::optional output_size); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & repeats, ::std::optional dim, ::std::optional output_size); +}; + +struct TORCH_API repeat_interleave_self_int { + using schema = at::Tensor (const at::Tensor &, c10::SymInt, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::repeat_interleave"; + static constexpr const char* overload_name = "self_int"; + static constexpr const char* schema_str = "repeat_interleave.self_int(Tensor self, SymInt repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymInt repeats, ::std::optional dim, ::std::optional output_size); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt repeats, ::std::optional dim, ::std::optional output_size); +}; + +struct TORCH_API repeat_interleave_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::repeat_interleave"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "repeat_interleave.Tensor_out(Tensor repeats, *, SymInt? output_size=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & repeats, ::std::optional output_size, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & repeats, ::std::optional output_size, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/repeat_ops.h b/backends/pytorch/include/ATen/ops/repeat_ops.h new file mode 100644 index 000000000..1fe323549 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/repeat_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API repeat { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::repeat"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "repeat(Tensor self, SymInt[] repeats) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef repeats); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef repeats); +}; + +struct TORCH_API repeat_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymIntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::repeat"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "repeat.out(Tensor self, SymInt[] repeats, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymIntArrayRef repeats, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef repeats, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/requires_grad_ops.h b/backends/pytorch/include/ATen/ops/requires_grad_ops.h new file mode 100644 index 000000000..62b6ae799 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/requires_grad_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API requires_grad_ { + using schema = at::Tensor & (at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::requires_grad_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "requires_grad_(Tensor(a!) self, bool requires_grad=True) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, bool requires_grad); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, bool requires_grad); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/reshape_as_ops.h b/backends/pytorch/include/ATen/ops/reshape_as_ops.h new file mode 100644 index 000000000..cb11f5de3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/reshape_as_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API reshape_as { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::reshape_as"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "reshape_as(Tensor(a) self, Tensor other) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/reshape_ops.h b/backends/pytorch/include/ATen/ops/reshape_ops.h new file mode 100644 index 000000000..cef9dcc5a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/reshape_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API reshape { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::reshape"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "reshape(Tensor(a) self, SymInt[] shape) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef shape); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef shape); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/resize_as_ops.h b/backends/pytorch/include/ATen/ops/resize_as_ops.h new file mode 100644 index 000000000..95f53507b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/resize_as_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API resize_as_ { + using schema = const at::Tensor & (const at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resize_as_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "resize_as_(Tensor(a!) self, Tensor the_template, *, MemoryFormat? memory_format=None) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, const at::Tensor & the_template, ::std::optional memory_format); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & the_template, ::std::optional memory_format); +}; + +struct TORCH_API resize_as_out { + using schema = const at::Tensor & (const at::Tensor &, const at::Tensor &, ::std::optional, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resize_as"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "resize_as.out(Tensor self, Tensor the_template, *, MemoryFormat? memory_format=None, Tensor(a!) out) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, const at::Tensor & the_template, ::std::optional memory_format, const at::Tensor & out); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & the_template, ::std::optional memory_format, const at::Tensor & out); +}; + +struct TORCH_API resize_as { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resize_as"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "resize_as(Tensor self, Tensor the_template, *, MemoryFormat? memory_format=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & the_template, ::std::optional memory_format); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & the_template, ::std::optional memory_format); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/resize_as_sparse_ops.h b/backends/pytorch/include/ATen/ops/resize_as_sparse_ops.h new file mode 100644 index 000000000..c41bac473 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/resize_as_sparse_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API resize_as_sparse_ { + using schema = const at::Tensor & (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resize_as_sparse_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "resize_as_sparse_(Tensor(a!) self, Tensor the_template) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, const at::Tensor & the_template); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & the_template); +}; + +struct TORCH_API resize_as_sparse_out { + using schema = const at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resize_as_sparse"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "resize_as_sparse.out(Tensor self, Tensor the_template, *, Tensor(a!) out) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, const at::Tensor & the_template, const at::Tensor & out); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & the_template, const at::Tensor & out); +}; + +struct TORCH_API resize_as_sparse { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resize_as_sparse"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "resize_as_sparse(Tensor self, Tensor the_template) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & the_template); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & the_template); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/resize_ops.h b/backends/pytorch/include/ATen/ops/resize_ops.h new file mode 100644 index 000000000..c1f51b8d6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/resize_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API resize_ { + using schema = const at::Tensor & (const at::Tensor &, c10::SymIntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resize_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "resize_(Tensor(a!) self, SymInt[] size, *, MemoryFormat? memory_format=None) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional memory_format); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional memory_format); +}; + +struct TORCH_API resize_out { + using schema = const at::Tensor & (const at::Tensor &, c10::SymIntArrayRef, ::std::optional, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resize"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "resize.out(Tensor self, SymInt[] size, *, MemoryFormat? memory_format=None, Tensor(a!) out) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional memory_format, const at::Tensor & out); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional memory_format, const at::Tensor & out); +}; + +struct TORCH_API resize { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resize"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "resize(Tensor self, SymInt[] size, *, MemoryFormat? memory_format=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional memory_format); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size, ::std::optional memory_format); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/resolve_conj_ops.h b/backends/pytorch/include/ATen/ops/resolve_conj_ops.h new file mode 100644 index 000000000..e06fb0e94 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/resolve_conj_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API resolve_conj { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resolve_conj"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "resolve_conj(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/resolve_neg_ops.h b/backends/pytorch/include/ATen/ops/resolve_neg_ops.h new file mode 100644 index 000000000..931615b23 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/resolve_neg_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API resolve_neg { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::resolve_neg"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "resolve_neg(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/retain_grad_ops.h b/backends/pytorch/include/ATen/ops/retain_grad_ops.h new file mode 100644 index 000000000..8f1195069 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/retain_grad_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API retain_grad { + using schema = void (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::retain_grad"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "retain_grad(Tensor(a!) self) -> ()"; + static void call(at::Tensor & self); + static void redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/retains_grad_ops.h b/backends/pytorch/include/ATen/ops/retains_grad_ops.h new file mode 100644 index 000000000..fa2fb00e3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/retains_grad_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API retains_grad { + using schema = bool (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::retains_grad"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "retains_grad(Tensor self) -> bool"; + static bool call(const at::Tensor & self); + static bool redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/roll_ops.h b/backends/pytorch/include/ATen/ops/roll_ops.h new file mode 100644 index 000000000..8077c42cd --- /dev/null +++ b/backends/pytorch/include/ATen/ops/roll_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API roll { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::roll"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "roll(Tensor self, SymInt[1] shifts, int[1] dims=[]) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef shifts, at::IntArrayRef dims); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef shifts, at::IntArrayRef dims); +}; + +struct TORCH_API roll_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymIntArrayRef, at::IntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::roll"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "roll.out(Tensor self, SymInt[1] shifts, int[1] dims=[], *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymIntArrayRef shifts, at::IntArrayRef dims, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef shifts, at::IntArrayRef dims, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/rot90_ops.h b/backends/pytorch/include/ATen/ops/rot90_ops.h new file mode 100644 index 000000000..057a686c3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/rot90_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API rot90 { + using schema = at::Tensor (const at::Tensor &, int64_t, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::rot90"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "rot90(Tensor self, int k=1, int[] dims=[0,1]) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t k, at::IntArrayRef dims); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t k, at::IntArrayRef dims); +}; + +struct TORCH_API rot90_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, at::IntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::rot90"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "rot90.out(Tensor self, int k=1, int[] dims=[0,1], *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t k, at::IntArrayRef dims, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t k, at::IntArrayRef dims, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/round_ops.h b/backends/pytorch/include/ATen/ops/round_ops.h new file mode 100644 index 000000000..c6950953b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/round_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API round { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::round"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "round(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API round_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::round_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "round_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API round_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::round"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "round.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API round_decimals { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::round"; + static constexpr const char* overload_name = "decimals"; + static constexpr const char* schema_str = "round.decimals(Tensor self, *, int decimals) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t decimals); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t decimals); +}; + +struct TORCH_API round__decimals { + using schema = at::Tensor & (at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::round_"; + static constexpr const char* overload_name = "decimals"; + static constexpr const char* schema_str = "round_.decimals(Tensor(a!) self, *, int decimals) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t decimals); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t decimals); +}; + +struct TORCH_API round_decimals_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::round"; + static constexpr const char* overload_name = "decimals_out"; + static constexpr const char* schema_str = "round.decimals_out(Tensor self, *, int decimals, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t decimals, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t decimals, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/row_indices_ops.h b/backends/pytorch/include/ATen/ops/row_indices_ops.h new file mode 100644 index 000000000..171e4bde6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/row_indices_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API row_indices { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::row_indices"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "row_indices(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/rshift_ops.h b/backends/pytorch/include/ATen/ops/rshift_ops.h new file mode 100644 index 000000000..ff0338022 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/rshift_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API __rshift___Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__rshift__"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "__rshift__.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API __rshift___Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__rshift__"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "__rshift__.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API __irshift___Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__irshift__"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "__irshift__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API __irshift___Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__irshift__"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "__irshift__.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API __rshift___Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__rshift__"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "__rshift__.Scalar_out(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +struct TORCH_API __rshift___Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__rshift__"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "__rshift__.Tensor_out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/rsqrt_ops.h b/backends/pytorch/include/ATen/ops/rsqrt_ops.h new file mode 100644 index 000000000..2c57f1b5d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/rsqrt_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API rsqrt { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::rsqrt"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "rsqrt(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API rsqrt_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::rsqrt_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "rsqrt_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API rsqrt_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::rsqrt"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "rsqrt.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/scatter_add_ops.h b/backends/pytorch/include/ATen/ops/scatter_add_ops.h new file mode 100644 index 000000000..bd1910c00 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/scatter_add_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API scatter_add { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_add"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "scatter_add(Tensor self, int dim, Tensor index, Tensor src) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src); +}; + +struct TORCH_API scatter_add_ { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_add_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "scatter_add_(Tensor(a!) self, int dim, Tensor index, Tensor src) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src); +}; + +struct TORCH_API scatter_add_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_add"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "scatter_add.out(Tensor self, int dim, Tensor index, Tensor src, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, at::Tensor & out); +}; + +struct TORCH_API scatter_add_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_add"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "scatter_add.dimname(Tensor self, Dimname dim, Tensor index, Tensor src) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & src); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & src); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/scatter_ops.h b/backends/pytorch/include/ATen/ops/scatter_ops.h new file mode 100644 index 000000000..49f977d86 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/scatter_ops.h @@ -0,0 +1,177 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API scatter_src { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter"; + static constexpr const char* overload_name = "src"; + static constexpr const char* schema_str = "scatter.src(Tensor self, int dim, Tensor index, Tensor src) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src); +}; + +struct TORCH_API scatter__src { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_"; + static constexpr const char* overload_name = "src"; + static constexpr const char* schema_str = "scatter_.src(Tensor(a!) self, int dim, Tensor index, Tensor src) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src); +}; + +struct TORCH_API scatter_src_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter"; + static constexpr const char* overload_name = "src_out"; + static constexpr const char* schema_str = "scatter.src_out(Tensor self, int dim, Tensor index, Tensor src, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, at::Tensor & out); +}; + +struct TORCH_API scatter_value { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter"; + static constexpr const char* overload_name = "value"; + static constexpr const char* schema_str = "scatter.value(Tensor self, int dim, Tensor index, Scalar value) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value); +}; + +struct TORCH_API scatter__value { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_"; + static constexpr const char* overload_name = "value"; + static constexpr const char* schema_str = "scatter_.value(Tensor(a!) self, int dim, Tensor index, Scalar value) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value); +}; + +struct TORCH_API scatter_value_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter"; + static constexpr const char* overload_name = "value_out"; + static constexpr const char* schema_str = "scatter.value_out(Tensor self, int dim, Tensor index, Scalar value, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value, at::Tensor & out); +}; + +struct TORCH_API scatter_reduce { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, c10::string_view); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter"; + static constexpr const char* overload_name = "reduce"; + static constexpr const char* schema_str = "scatter.reduce(Tensor self, int dim, Tensor index, Tensor src, *, str reduce) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce); +}; + +struct TORCH_API scatter__reduce { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, c10::string_view); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_"; + static constexpr const char* overload_name = "reduce"; + static constexpr const char* schema_str = "scatter_.reduce(Tensor(a!) self, int dim, Tensor index, Tensor src, *, str reduce) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce); +}; + +struct TORCH_API scatter_reduce_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, c10::string_view, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter"; + static constexpr const char* overload_name = "reduce_out"; + static constexpr const char* schema_str = "scatter.reduce_out(Tensor self, int dim, Tensor index, Tensor src, *, str reduce, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, at::Tensor & out); +}; + +struct TORCH_API scatter_value_reduce { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Scalar &, c10::string_view); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter"; + static constexpr const char* overload_name = "value_reduce"; + static constexpr const char* schema_str = "scatter.value_reduce(Tensor self, int dim, Tensor index, Scalar value, *, str reduce) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value, c10::string_view reduce); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value, c10::string_view reduce); +}; + +struct TORCH_API scatter__value_reduce { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Scalar &, c10::string_view); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_"; + static constexpr const char* overload_name = "value_reduce"; + static constexpr const char* schema_str = "scatter_.value_reduce(Tensor(a!) self, int dim, Tensor index, Scalar value, *, str reduce) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value, c10::string_view reduce); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value, c10::string_view reduce); +}; + +struct TORCH_API scatter_value_reduce_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Scalar &, c10::string_view, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter"; + static constexpr const char* overload_name = "value_reduce_out"; + static constexpr const char* schema_str = "scatter.value_reduce_out(Tensor self, int dim, Tensor index, Scalar value, *, str reduce, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value, c10::string_view reduce, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Scalar & value, c10::string_view reduce, at::Tensor & out); +}; + +struct TORCH_API scatter_dimname_src { + using schema = at::Tensor (const at::Tensor &, at::Dimname, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter"; + static constexpr const char* overload_name = "dimname_src"; + static constexpr const char* schema_str = "scatter.dimname_src(Tensor self, Dimname dim, Tensor index, Tensor src) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & src); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Tensor & src); +}; + +struct TORCH_API scatter_dimname_value { + using schema = at::Tensor (const at::Tensor &, at::Dimname, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter"; + static constexpr const char* overload_name = "dimname_value"; + static constexpr const char* schema_str = "scatter.dimname_value(Tensor self, Dimname dim, Tensor index, Scalar value) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Scalar & value); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, const at::Tensor & index, const at::Scalar & value); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/scatter_reduce_ops.h b/backends/pytorch/include/ATen/ops/scatter_reduce_ops.h new file mode 100644 index 000000000..7ece27168 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/scatter_reduce_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API scatter_reduce_two { + using schema = at::Tensor (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, c10::string_view, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_reduce"; + static constexpr const char* overload_name = "two"; + static constexpr const char* schema_str = "scatter_reduce.two(Tensor self, int dim, Tensor index, Tensor src, str reduce, *, bool include_self=True) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, bool include_self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, bool include_self); +}; + +struct TORCH_API scatter_reduce__two { + using schema = at::Tensor & (at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, c10::string_view, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_reduce_"; + static constexpr const char* overload_name = "two"; + static constexpr const char* schema_str = "scatter_reduce_.two(Tensor(a!) self, int dim, Tensor index, Tensor src, str reduce, *, bool include_self=True) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, bool include_self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, bool include_self); +}; + +struct TORCH_API scatter_reduce_two_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, const at::Tensor &, const at::Tensor &, c10::string_view, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::scatter_reduce"; + static constexpr const char* overload_name = "two_out"; + static constexpr const char* schema_str = "scatter_reduce.two_out(Tensor self, int dim, Tensor index, Tensor src, str reduce, *, bool include_self=True, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, bool include_self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, const at::Tensor & index, const at::Tensor & src, c10::string_view reduce, bool include_self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/select_ops.h b/backends/pytorch/include/ATen/ops/select_ops.h new file mode 100644 index 000000000..6e1a10e35 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/select_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API select_Dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::select"; + static constexpr const char* overload_name = "Dimname"; + static constexpr const char* schema_str = "select.Dimname(Tensor(a) self, Dimname dim, int index) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, int64_t index); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, int64_t index); +}; + +struct TORCH_API select_int { + using schema = at::Tensor (const at::Tensor &, int64_t, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::select"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "select.int(Tensor(a) self, int dim, SymInt index) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t dim, c10::SymInt index); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, c10::SymInt index); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/select_scatter_ops.h b/backends/pytorch/include/ATen/ops/select_scatter_ops.h new file mode 100644 index 000000000..af127aaca --- /dev/null +++ b/backends/pytorch/include/ATen/ops/select_scatter_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API select_scatter { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, int64_t, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::select_scatter"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "select_scatter(Tensor self, Tensor src, int dim, SymInt index) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & src, int64_t dim, c10::SymInt index); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, int64_t dim, c10::SymInt index); +}; + +struct TORCH_API select_scatter_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, int64_t, c10::SymInt, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::select_scatter"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "select_scatter.out(Tensor self, Tensor src, int dim, SymInt index, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & src, int64_t dim, c10::SymInt index, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, int64_t dim, c10::SymInt index, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/set_data_ops.h b/backends/pytorch/include/ATen/ops/set_data_ops.h new file mode 100644 index 000000000..d827388d4 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/set_data_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API set_data { + using schema = void (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set_data"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "set_data(Tensor(a!) self, Tensor new_data) -> ()"; + static void call(at::Tensor & self, const at::Tensor & new_data); + static void redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & new_data); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/set_ops.h b/backends/pytorch/include/ATen/ops/set_ops.h new file mode 100644 index 000000000..49701c440 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/set_ops.h @@ -0,0 +1,166 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API set__source_Storage { + using schema = at::Tensor & (at::Tensor &, at::Storage); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set_"; + static constexpr const char* overload_name = "source_Storage"; + static constexpr const char* schema_str = "set_.source_Storage(Tensor(a!) self, Storage source) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, at::Storage source); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, at::Storage source); +}; + +struct TORCH_API set__source_Storage_storage_offset { + using schema = at::Tensor & (at::Tensor &, at::Storage, c10::SymInt, c10::SymIntArrayRef, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set_"; + static constexpr const char* overload_name = "source_Storage_storage_offset"; + static constexpr const char* schema_str = "set_.source_Storage_storage_offset(Tensor(a!) self, Storage source, SymInt storage_offset, SymInt[] size, SymInt[] stride=[]) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, at::Storage source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, at::Storage source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride); +}; + +struct TORCH_API set__source_Tensor_storage_offset { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, c10::SymInt, c10::SymIntArrayRef, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set_"; + static constexpr const char* overload_name = "source_Tensor_storage_offset"; + static constexpr const char* schema_str = "set_.source_Tensor_storage_offset(Tensor(a!) self, Tensor source, SymInt storage_offset, SymInt[] size, SymInt[] stride=[]) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride); +}; + +struct TORCH_API set__source_Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set_"; + static constexpr const char* overload_name = "source_Tensor"; + static constexpr const char* schema_str = "set_.source_Tensor(Tensor(a!) self, Tensor source) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & source); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & source); +}; + +struct TORCH_API set_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "set_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API set_source_Storage_out { + using schema = at::Tensor & (const at::Tensor &, at::Storage, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set"; + static constexpr const char* overload_name = "source_Storage_out"; + static constexpr const char* schema_str = "set.source_Storage_out(Tensor self, Storage source, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Storage source, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Storage source, at::Tensor & out); +}; + +struct TORCH_API set_source_Storage { + using schema = at::Tensor (const at::Tensor &, at::Storage); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set"; + static constexpr const char* overload_name = "source_Storage"; + static constexpr const char* schema_str = "set.source_Storage(Tensor self, Storage source) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Storage source); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Storage source); +}; + +struct TORCH_API set_source_Storage_storage_offset_out { + using schema = at::Tensor & (const at::Tensor &, at::Storage, c10::SymInt, c10::SymIntArrayRef, c10::SymIntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set"; + static constexpr const char* overload_name = "source_Storage_storage_offset_out"; + static constexpr const char* schema_str = "set.source_Storage_storage_offset_out(Tensor self, Storage source, SymInt storage_offset, SymInt[] size, SymInt[] stride=[], *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Storage source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Storage source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride, at::Tensor & out); +}; + +struct TORCH_API set_source_Storage_storage_offset { + using schema = at::Tensor (const at::Tensor &, at::Storage, c10::SymInt, c10::SymIntArrayRef, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set"; + static constexpr const char* overload_name = "source_Storage_storage_offset"; + static constexpr const char* schema_str = "set.source_Storage_storage_offset(Tensor self, Storage source, SymInt storage_offset, SymInt[] size, SymInt[] stride=[]) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Storage source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Storage source, c10::SymInt storage_offset, c10::SymIntArrayRef size, c10::SymIntArrayRef stride); +}; + +struct TORCH_API set_source_Tensor_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set"; + static constexpr const char* overload_name = "source_Tensor_out"; + static constexpr const char* schema_str = "set.source_Tensor_out(Tensor self, Tensor source, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & source, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & source, at::Tensor & out); +}; + +struct TORCH_API set_source_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set"; + static constexpr const char* overload_name = "source_Tensor"; + static constexpr const char* schema_str = "set.source_Tensor(Tensor self, Tensor source) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & source); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & source); +}; + +struct TORCH_API set_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "set.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API set { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::set"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "set(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sgn_ops.h b/backends/pytorch/include/ATen/ops/sgn_ops.h new file mode 100644 index 000000000..d140bb76f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sgn_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sgn { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sgn"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sgn(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API sgn_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sgn_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sgn_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API sgn_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sgn"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sgn.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sigmoid_ops.h b/backends/pytorch/include/ATen/ops/sigmoid_ops.h new file mode 100644 index 000000000..4cffec145 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sigmoid_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sigmoid { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sigmoid"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sigmoid(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API sigmoid_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sigmoid_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sigmoid_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API sigmoid_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sigmoid"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sigmoid.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sign_ops.h b/backends/pytorch/include/ATen/ops/sign_ops.h new file mode 100644 index 000000000..c0b3dd98e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sign_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sign { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sign"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sign(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API sign_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sign_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sign_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API sign_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sign"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sign.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/signbit_ops.h b/backends/pytorch/include/ATen/ops/signbit_ops.h new file mode 100644 index 000000000..08f0a18bf --- /dev/null +++ b/backends/pytorch/include/ATen/ops/signbit_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API signbit { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::signbit"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "signbit(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API signbit_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::signbit"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "signbit.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sin_ops.h b/backends/pytorch/include/ATen/ops/sin_ops.h new file mode 100644 index 000000000..62d1c634b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sin_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sin { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sin"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sin(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API sin_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sin_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sin_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API sin_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sin"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sin.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sinc_ops.h b/backends/pytorch/include/ATen/ops/sinc_ops.h new file mode 100644 index 000000000..cff637c82 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sinc_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sinc { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sinc"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sinc(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API sinc_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sinc_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sinc_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API sinc_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sinc"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sinc.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sinh_ops.h b/backends/pytorch/include/ATen/ops/sinh_ops.h new file mode 100644 index 000000000..9f9969473 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sinh_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sinh { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sinh"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sinh(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API sinh_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sinh_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sinh_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API sinh_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sinh"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sinh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/size_ops.h b/backends/pytorch/include/ATen/ops/size_ops.h new file mode 100644 index 000000000..79827da20 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/size_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API size_int { + using schema = int64_t (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::size"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "size.int(Tensor self, int dim) -> int"; + static int64_t call(const at::Tensor & self, int64_t dim); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim); +}; + +struct TORCH_API size_Dimname { + using schema = int64_t (const at::Tensor &, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::size"; + static constexpr const char* overload_name = "Dimname"; + static constexpr const char* schema_str = "size.Dimname(Tensor self, Dimname dim) -> int"; + static int64_t call(const at::Tensor & self, at::Dimname dim); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/slice_inverse_ops.h b/backends/pytorch/include/ATen/ops/slice_inverse_ops.h new file mode 100644 index 000000000..653767db3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/slice_inverse_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API slice_inverse { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, int64_t, ::std::optional, ::std::optional, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::slice_inverse"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "slice_inverse(Tensor(a) self, Tensor src, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & src, int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/slice_ops.h b/backends/pytorch/include/ATen/ops/slice_ops.h new file mode 100644 index 000000000..aa85c91ca --- /dev/null +++ b/backends/pytorch/include/ATen/ops/slice_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API slice_Tensor { + using schema = at::Tensor (const at::Tensor &, int64_t, ::std::optional, ::std::optional, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::slice"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "slice.Tensor(Tensor(a) self, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/slice_scatter_ops.h b/backends/pytorch/include/ATen/ops/slice_scatter_ops.h new file mode 100644 index 000000000..3e02da96a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/slice_scatter_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API slice_scatter { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, int64_t, ::std::optional, ::std::optional, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::slice_scatter"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "slice_scatter(Tensor self, Tensor src, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & src, int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step); +}; + +struct TORCH_API slice_scatter_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, int64_t, ::std::optional, ::std::optional, c10::SymInt, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::slice_scatter"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "slice_scatter.out(Tensor self, Tensor src, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & src, int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & src, int64_t dim, ::std::optional start, ::std::optional end, c10::SymInt step, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/slogdet_ops.h b/backends/pytorch/include/ATen/ops/slogdet_ops.h new file mode 100644 index 000000000..6ed8f8308 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/slogdet_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API slogdet { + using schema = ::std::tuple (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::slogdet"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "slogdet(Tensor self) -> (Tensor sign, Tensor logabsdet)"; + static ::std::tuple call(const at::Tensor & self); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API slogdet_out { + using schema = ::std::tuple (const at::Tensor &, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::slogdet"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "slogdet.out(Tensor self, *, Tensor(a!) sign, Tensor(b!) logabsdet) -> (Tensor(a!) sign, Tensor(b!) logabsdet)"; + static ::std::tuple call(const at::Tensor & self, at::Tensor & sign, at::Tensor & logabsdet); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & sign, at::Tensor & logabsdet); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/smm_ops.h b/backends/pytorch/include/ATen/ops/smm_ops.h new file mode 100644 index 000000000..dd6d37784 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/smm_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API smm { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::smm"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "smm(Tensor self, Tensor mat2) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mat2); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat2); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/softmax_ops.h b/backends/pytorch/include/ATen/ops/softmax_ops.h new file mode 100644 index 000000000..aaa386b06 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/softmax_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API softmax_int { + using schema = at::Tensor (const at::Tensor &, int64_t, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::softmax"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "softmax.int(Tensor self, int dim, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t dim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, ::std::optional dtype); +}; + +struct TORCH_API softmax_int_out { + using schema = at::Tensor & (const at::Tensor &, int64_t, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::softmax"; + static constexpr const char* overload_name = "int_out"; + static constexpr const char* schema_str = "softmax.int_out(Tensor self, int dim, ScalarType? dtype=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, int64_t dim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, ::std::optional dtype, at::Tensor & out); +}; + +struct TORCH_API softmax_Dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::softmax"; + static constexpr const char* overload_name = "Dimname"; + static constexpr const char* schema_str = "softmax.Dimname(Tensor self, Dimname dim, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, ::std::optional dtype); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sort_ops.h b/backends/pytorch/include/ATen/ops/sort_ops.h new file mode 100644 index 000000000..7db065e20 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sort_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sort_values { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sort"; + static constexpr const char* overload_name = "values"; + static constexpr const char* schema_str = "sort.values(Tensor self, int dim=-1, bool descending=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool descending, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool descending, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API sort_values_stable { + using schema = ::std::tuple (const at::Tensor &, ::std::optional, int64_t, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sort"; + static constexpr const char* overload_name = "values_stable"; + static constexpr const char* schema_str = "sort.values_stable(Tensor self, *, bool? stable, int dim=-1, bool descending=False, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, ::std::optional stable, int64_t dim, bool descending, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional stable, int64_t dim, bool descending, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API sort { + using schema = ::std::tuple (const at::Tensor &, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sort"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sort(Tensor self, int dim=-1, bool descending=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, int64_t dim, bool descending); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, bool descending); +}; + +struct TORCH_API sort_stable { + using schema = ::std::tuple (const at::Tensor &, ::std::optional, int64_t, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sort"; + static constexpr const char* overload_name = "stable"; + static constexpr const char* schema_str = "sort.stable(Tensor self, *, bool? stable, int dim=-1, bool descending=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, ::std::optional stable, int64_t dim, bool descending); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional stable, int64_t dim, bool descending); +}; + +struct TORCH_API sort_dimname_values { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sort"; + static constexpr const char* overload_name = "dimname_values"; + static constexpr const char* schema_str = "sort.dimname_values(Tensor self, Dimname dim, bool descending=False, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool descending, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool descending, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API sort_dimname_values_stable { + using schema = ::std::tuple (const at::Tensor &, ::std::optional, at::Dimname, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sort"; + static constexpr const char* overload_name = "dimname_values_stable"; + static constexpr const char* schema_str = "sort.dimname_values_stable(Tensor self, *, bool? stable, Dimname dim, bool descending=False, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, ::std::optional stable, at::Dimname dim, bool descending, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional stable, at::Dimname dim, bool descending, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API sort_dimname { + using schema = ::std::tuple (const at::Tensor &, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sort"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "sort.dimname(Tensor self, Dimname dim, bool descending=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, at::Dimname dim, bool descending); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, bool descending); +}; + +struct TORCH_API sort_dimname_stable { + using schema = ::std::tuple (const at::Tensor &, ::std::optional, at::Dimname, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sort"; + static constexpr const char* overload_name = "dimname_stable"; + static constexpr const char* schema_str = "sort.dimname_stable(Tensor self, *, bool? stable, Dimname dim, bool descending=False) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, ::std::optional stable, at::Dimname dim, bool descending); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional stable, at::Dimname dim, bool descending); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sparse_dim_ops.h b/backends/pytorch/include/ATen/ops/sparse_dim_ops.h new file mode 100644 index 000000000..f914dc872 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sparse_dim_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sparse_dim { + using schema = int64_t (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sparse_dim"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sparse_dim(Tensor self) -> int"; + static int64_t call(const at::Tensor & self); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sparse_mask_ops.h b/backends/pytorch/include/ATen/ops/sparse_mask_ops.h new file mode 100644 index 000000000..d6c490e38 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sparse_mask_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sparse_mask { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sparse_mask"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sparse_mask(Tensor self, Tensor mask) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mask); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask); +}; + +struct TORCH_API sparse_mask_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sparse_mask"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sparse_mask.out(Tensor self, Tensor mask, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mask, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mask, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sparse_resize_and_clear_ops.h b/backends/pytorch/include/ATen/ops/sparse_resize_and_clear_ops.h new file mode 100644 index 000000000..7d1e1c225 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sparse_resize_and_clear_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sparse_resize_and_clear_ { + using schema = const at::Tensor & (const at::Tensor &, at::IntArrayRef, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sparse_resize_and_clear_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sparse_resize_and_clear_(Tensor(a!) self, int[] size, int sparse_dim, int dense_dim) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim); +}; + +struct TORCH_API sparse_resize_and_clear_out { + using schema = const at::Tensor & (const at::Tensor &, at::IntArrayRef, int64_t, int64_t, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sparse_resize_and_clear"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sparse_resize_and_clear.out(Tensor self, int[] size, int sparse_dim, int dense_dim, *, Tensor(a!) out) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim, const at::Tensor & out); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim, const at::Tensor & out); +}; + +struct TORCH_API sparse_resize_and_clear { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sparse_resize_and_clear"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sparse_resize_and_clear(Tensor self, int[] size, int sparse_dim, int dense_dim) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sparse_resize_ops.h b/backends/pytorch/include/ATen/ops/sparse_resize_ops.h new file mode 100644 index 000000000..787765e91 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sparse_resize_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sparse_resize_ { + using schema = const at::Tensor & (const at::Tensor &, at::IntArrayRef, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sparse_resize_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sparse_resize_(Tensor(a!) self, int[] size, int sparse_dim, int dense_dim) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim); +}; + +struct TORCH_API sparse_resize_out { + using schema = const at::Tensor & (const at::Tensor &, at::IntArrayRef, int64_t, int64_t, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sparse_resize"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sparse_resize.out(Tensor self, int[] size, int sparse_dim, int dense_dim, *, Tensor(a!) out) -> Tensor(a!)"; + static const at::Tensor & call(const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim, const at::Tensor & out); + static const at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim, const at::Tensor & out); +}; + +struct TORCH_API sparse_resize { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sparse_resize"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sparse_resize(Tensor self, int[] size, int sparse_dim, int dense_dim) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef size, int64_t sparse_dim, int64_t dense_dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/split_ops.h b/backends/pytorch/include/ATen/ops/split_ops.h new file mode 100644 index 000000000..6d993c889 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/split_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API split_Tensor { + using schema = ::std::vector (const at::Tensor &, c10::SymInt, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::split"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "split.Tensor(Tensor(a -> *) self, SymInt split_size, int dim=0) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, c10::SymInt split_size, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt split_size, int64_t dim); +}; + +struct TORCH_API split_sizes { + using schema = ::std::vector (const at::Tensor &, c10::SymIntArrayRef, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::split"; + static constexpr const char* overload_name = "sizes"; + static constexpr const char* schema_str = "split.sizes(Tensor(a -> *) self, SymInt[] split_size, int dim=0) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, c10::SymIntArrayRef split_size, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef split_size, int64_t dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/split_with_sizes_ops.h b/backends/pytorch/include/ATen/ops/split_with_sizes_ops.h new file mode 100644 index 000000000..49713bc55 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/split_with_sizes_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API split_with_sizes { + using schema = ::std::vector (const at::Tensor &, c10::SymIntArrayRef, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::split_with_sizes"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "split_with_sizes(Tensor(a -> *) self, SymInt[] split_sizes, int dim=0) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, c10::SymIntArrayRef split_sizes, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef split_sizes, int64_t dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sqrt_ops.h b/backends/pytorch/include/ATen/ops/sqrt_ops.h new file mode 100644 index 000000000..3d68e8878 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sqrt_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sqrt { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sqrt"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sqrt(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API sqrt_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sqrt_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sqrt_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API sqrt_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sqrt"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sqrt.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/square_ops.h b/backends/pytorch/include/ATen/ops/square_ops.h new file mode 100644 index 000000000..692cb08ad --- /dev/null +++ b/backends/pytorch/include/ATen/ops/square_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API square { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::square"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "square(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API square_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::square_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "square_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API square_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::square"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "square.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/squeeze_ops.h b/backends/pytorch/include/ATen/ops/squeeze_ops.h new file mode 100644 index 000000000..f1ac1fa61 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/squeeze_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API squeeze { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::squeeze"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "squeeze(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API squeeze_dim { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::squeeze"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "squeeze.dim(Tensor(a) self, int dim) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim); +}; + +struct TORCH_API squeeze_dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::squeeze"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "squeeze.dimname(Tensor(a) self, Dimname dim) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim); +}; + +struct TORCH_API squeeze_dims { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::squeeze"; + static constexpr const char* overload_name = "dims"; + static constexpr const char* schema_str = "squeeze.dims(Tensor(a) self, int[] dim) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef dim); +}; + +struct TORCH_API squeeze_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::squeeze_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "squeeze_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API squeeze__dim { + using schema = at::Tensor & (at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::squeeze_"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "squeeze_.dim(Tensor(a!) self, int dim) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim); +}; + +struct TORCH_API squeeze__dims { + using schema = at::Tensor & (at::Tensor &, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::squeeze_"; + static constexpr const char* overload_name = "dims"; + static constexpr const char* schema_str = "squeeze_.dims(Tensor(a!) self, int[] dim) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, at::IntArrayRef dim); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, at::IntArrayRef dim); +}; + +struct TORCH_API squeeze__dimname { + using schema = at::Tensor & (at::Tensor &, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::squeeze_"; + static constexpr const char* overload_name = "dimname"; + static constexpr const char* schema_str = "squeeze_.dimname(Tensor(a!) self, Dimname dim) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, at::Dimname dim); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, at::Dimname dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sspaddmm_ops.h b/backends/pytorch/include/ATen/ops/sspaddmm_ops.h new file mode 100644 index 000000000..6b080fad6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sspaddmm_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sspaddmm { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sspaddmm"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sspaddmm(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha); +}; + +struct TORCH_API sspaddmm_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sspaddmm"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sspaddmm.out(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & mat1, const at::Tensor & mat2, const at::Scalar & beta, const at::Scalar & alpha, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/std_ops.h b/backends/pytorch/include/ATen/ops/std_ops.h new file mode 100644 index 000000000..5b9f6cbcc --- /dev/null +++ b/backends/pytorch/include/ATen/ops/std_ops.h @@ -0,0 +1,122 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API std { + using schema = at::Tensor (const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::std"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "std(Tensor self, bool unbiased=True) -> Tensor"; + static at::Tensor call(const at::Tensor & self, bool unbiased); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool unbiased); +}; + +struct TORCH_API std_dim { + using schema = at::Tensor (const at::Tensor &, at::OptionalIntArrayRef, bool, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::std"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "std.dim(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool unbiased, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool unbiased, bool keepdim); +}; + +struct TORCH_API std_correction { + using schema = at::Tensor (const at::Tensor &, at::OptionalIntArrayRef, const ::std::optional &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::std"; + static constexpr const char* overload_name = "correction"; + static constexpr const char* schema_str = "std.correction(Tensor self, int[1]? dim=None, *, Scalar? correction=None, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::OptionalIntArrayRef dim, const ::std::optional & correction, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, const ::std::optional & correction, bool keepdim); +}; + +struct TORCH_API std_out { + using schema = at::Tensor & (const at::Tensor &, at::OptionalIntArrayRef, bool, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::std"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "std.out(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool unbiased, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool unbiased, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API std_correction_out { + using schema = at::Tensor & (const at::Tensor &, at::OptionalIntArrayRef, const ::std::optional &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::std"; + static constexpr const char* overload_name = "correction_out"; + static constexpr const char* schema_str = "std.correction_out(Tensor self, int[1]? dim=None, *, Scalar? correction=None, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::OptionalIntArrayRef dim, const ::std::optional & correction, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, const ::std::optional & correction, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API std_names_dim { + using schema = at::Tensor (const at::Tensor &, at::DimnameList, bool, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::std"; + static constexpr const char* overload_name = "names_dim"; + static constexpr const char* schema_str = "std.names_dim(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::DimnameList dim, bool unbiased, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, bool unbiased, bool keepdim); +}; + +struct TORCH_API std_names_out { + using schema = at::Tensor & (const at::Tensor &, at::DimnameList, bool, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::std"; + static constexpr const char* overload_name = "names_out"; + static constexpr const char* schema_str = "std.names_out(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::DimnameList dim, bool unbiased, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, bool unbiased, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API std_correction_names { + using schema = at::Tensor (const at::Tensor &, at::DimnameList, const ::std::optional &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::std"; + static constexpr const char* overload_name = "correction_names"; + static constexpr const char* schema_str = "std.correction_names(Tensor self, Dimname[1] dim, *, Scalar? correction=None, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::DimnameList dim, const ::std::optional & correction, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, const ::std::optional & correction, bool keepdim); +}; + +struct TORCH_API std_correction_names_out { + using schema = at::Tensor & (const at::Tensor &, at::DimnameList, const ::std::optional &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::std"; + static constexpr const char* overload_name = "correction_names_out"; + static constexpr const char* schema_str = "std.correction_names_out(Tensor self, Dimname[1] dim, *, Scalar? correction=None, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::DimnameList dim, const ::std::optional & correction, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, const ::std::optional & correction, bool keepdim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/stft_ops.h b/backends/pytorch/include/ATen/ops/stft_ops.h new file mode 100644 index 000000000..4b840fbc6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/stft_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API stft { + using schema = at::Tensor (const at::Tensor &, int64_t, ::std::optional, ::std::optional, const ::std::optional &, bool, ::std::optional, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::stft"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "stft(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool normalized=False, bool? onesided=None, bool? return_complex=None, bool? align_to_window=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t n_fft, ::std::optional hop_length, ::std::optional win_length, const ::std::optional & window, bool normalized, ::std::optional onesided, ::std::optional return_complex, ::std::optional align_to_window); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t n_fft, ::std::optional hop_length, ::std::optional win_length, const ::std::optional & window, bool normalized, ::std::optional onesided, ::std::optional return_complex, ::std::optional align_to_window); +}; + +struct TORCH_API stft_center { + using schema = at::Tensor (const at::Tensor &, int64_t, ::std::optional, ::std::optional, const ::std::optional &, bool, c10::string_view, bool, ::std::optional, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::stft"; + static constexpr const char* overload_name = "center"; + static constexpr const char* schema_str = "stft.center(Tensor self, int n_fft, int? hop_length=None, int? win_length=None, Tensor? window=None, bool center=True, str pad_mode=\"reflect\", bool normalized=False, bool? onesided=None, bool? return_complex=None, bool? align_to_window=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t n_fft, ::std::optional hop_length, ::std::optional win_length, const ::std::optional & window, bool center, c10::string_view pad_mode, bool normalized, ::std::optional onesided, ::std::optional return_complex, ::std::optional align_to_window); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t n_fft, ::std::optional hop_length, ::std::optional win_length, const ::std::optional & window, bool center, c10::string_view pad_mode, bool normalized, ::std::optional onesided, ::std::optional return_complex, ::std::optional align_to_window); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/stride_ops.h b/backends/pytorch/include/ATen/ops/stride_ops.h new file mode 100644 index 000000000..09f6647b2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/stride_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API stride_int { + using schema = int64_t (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::stride"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "stride.int(Tensor self, int dim) -> int"; + static int64_t call(const at::Tensor & self, int64_t dim); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim); +}; + +struct TORCH_API stride_Dimname { + using schema = int64_t (const at::Tensor &, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::stride"; + static constexpr const char* overload_name = "Dimname"; + static constexpr const char* schema_str = "stride.Dimname(Tensor self, Dimname dim) -> int"; + static int64_t call(const at::Tensor & self, at::Dimname dim); + static int64_t redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sub_ops.h b/backends/pytorch/include/ATen/ops/sub_ops.h new file mode 100644 index 000000000..eee4188a7 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sub_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sub_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sub"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sub.out(Tensor self, Tensor other, *, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha, at::Tensor & out); +}; + +struct TORCH_API sub_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sub"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "sub.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); +}; + +struct TORCH_API sub__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sub_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "sub_.Tensor(Tensor(a!) self, Tensor other, *, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); +}; + +struct TORCH_API sub_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sub"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "sub.Scalar(Tensor self, Scalar other, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); +}; + +struct TORCH_API sub__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sub_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "sub_.Scalar(Tensor(a!) self, Scalar other, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); +}; + +struct TORCH_API sub_Scalar_out { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sub"; + static constexpr const char* overload_name = "Scalar_out"; + static constexpr const char* schema_str = "sub.Scalar_out(Tensor self, Scalar other, Scalar alpha=1, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/subtract_ops.h b/backends/pytorch/include/ATen/ops/subtract_ops.h new file mode 100644 index 000000000..ef3728b2d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/subtract_ops.h @@ -0,0 +1,78 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API subtract_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::subtract"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "subtract.out(Tensor self, Tensor other, *, Scalar alpha=1, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha, at::Tensor & out); +}; + +struct TORCH_API subtract_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::subtract"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "subtract.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); +}; + +struct TORCH_API subtract__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::subtract_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "subtract_.Tensor(Tensor(a!) self, Tensor other, *, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other, const at::Scalar & alpha); +}; + +struct TORCH_API subtract_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::subtract"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "subtract.Scalar(Tensor self, Scalar other, Scalar alpha=1) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); +}; + +struct TORCH_API subtract__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::subtract_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "subtract_.Scalar(Tensor(a!) self, Scalar other, Scalar alpha=1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other, const at::Scalar & alpha); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sum_ops.h b/backends/pytorch/include/ATen/ops/sum_ops.h new file mode 100644 index 000000000..eca85719c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sum_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sum { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sum"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sum(Tensor self, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype); +}; + +struct TORCH_API sum_dim_IntList { + using schema = at::Tensor (const at::Tensor &, at::OptionalIntArrayRef, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sum"; + static constexpr const char* overload_name = "dim_IntList"; + static constexpr const char* schema_str = "sum.dim_IntList(Tensor self, int[1]? dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype); +}; + +struct TORCH_API sum_dim_DimnameList { + using schema = at::Tensor (const at::Tensor &, at::DimnameList, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sum"; + static constexpr const char* overload_name = "dim_DimnameList"; + static constexpr const char* schema_str = "sum.dim_DimnameList(Tensor self, Dimname[1] dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::DimnameList dim, bool keepdim, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, bool keepdim, ::std::optional dtype); +}; + +struct TORCH_API sum_IntList_out { + using schema = at::Tensor & (const at::Tensor &, at::OptionalIntArrayRef, bool, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sum"; + static constexpr const char* overload_name = "IntList_out"; + static constexpr const char* schema_str = "sum.IntList_out(Tensor self, int[1]? dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool keepdim, ::std::optional dtype, at::Tensor & out); +}; + +struct TORCH_API sum_DimnameList_out { + using schema = at::Tensor & (const at::Tensor &, at::DimnameList, bool, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sum"; + static constexpr const char* overload_name = "DimnameList_out"; + static constexpr const char* schema_str = "sum.DimnameList_out(Tensor self, Dimname[1] dim, bool keepdim=False, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::DimnameList dim, bool keepdim, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, bool keepdim, ::std::optional dtype, at::Tensor & out); +}; + +struct TORCH_API sum_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sum"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "sum.out(Tensor self, *, ScalarType? dtype=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/sum_to_size_ops.h b/backends/pytorch/include/ATen/ops/sum_to_size_ops.h new file mode 100644 index 000000000..9b5b0436a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/sum_to_size_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API sum_to_size { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::sum_to_size"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "sum_to_size(Tensor self, SymInt[] size) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/svd_ops.h b/backends/pytorch/include/ATen/ops/svd_ops.h new file mode 100644 index 000000000..fb362b3a3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/svd_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API svd_U { + using schema = ::std::tuple (const at::Tensor &, bool, bool, at::Tensor &, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::svd"; + static constexpr const char* overload_name = "U"; + static constexpr const char* schema_str = "svd.U(Tensor self, bool some=True, bool compute_uv=True, *, Tensor(a!) U, Tensor(b!) S, Tensor(c!) V) -> (Tensor(a!) U, Tensor(b!) S, Tensor(c!) V)"; + static ::std::tuple call(const at::Tensor & self, bool some, bool compute_uv, at::Tensor & U, at::Tensor & S, at::Tensor & V); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool some, bool compute_uv, at::Tensor & U, at::Tensor & S, at::Tensor & V); +}; + +struct TORCH_API svd { + using schema = ::std::tuple (const at::Tensor &, bool, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::svd"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "svd(Tensor self, bool some=True, bool compute_uv=True) -> (Tensor U, Tensor S, Tensor V)"; + static ::std::tuple call(const at::Tensor & self, bool some, bool compute_uv); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool some, bool compute_uv); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/swapaxes_ops.h b/backends/pytorch/include/ATen/ops/swapaxes_ops.h new file mode 100644 index 000000000..a13d2c95c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/swapaxes_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API swapaxes { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::swapaxes"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "swapaxes(Tensor(a) self, int axis0, int axis1) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t axis0, int64_t axis1); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t axis0, int64_t axis1); +}; + +struct TORCH_API swapaxes_ { + using schema = at::Tensor & (at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::swapaxes_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "swapaxes_(Tensor(a!) self, int axis0, int axis1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t axis0, int64_t axis1); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t axis0, int64_t axis1); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/swapdims_ops.h b/backends/pytorch/include/ATen/ops/swapdims_ops.h new file mode 100644 index 000000000..b4f76235d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/swapdims_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API swapdims { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::swapdims"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "swapdims(Tensor(a) self, int dim0, int dim1) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t dim0, int64_t dim1); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim0, int64_t dim1); +}; + +struct TORCH_API swapdims_ { + using schema = at::Tensor & (at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::swapdims_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "swapdims_(Tensor(a!) self, int dim0, int dim1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim0, int64_t dim1); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim0, int64_t dim1); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/t_ops.h b/backends/pytorch/include/ATen/ops/t_ops.h new file mode 100644 index 000000000..7db02e0b2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/t_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API t { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::t"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "t(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API t_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::t_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "t_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/take_along_dim_ops.h b/backends/pytorch/include/ATen/ops/take_along_dim_ops.h new file mode 100644 index 000000000..9a3bd7c99 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/take_along_dim_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API take_along_dim_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::take_along_dim"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "take_along_dim.out(Tensor self, Tensor indices, int? dim=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & indices, ::std::optional dim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & indices, ::std::optional dim, at::Tensor & out); +}; + +struct TORCH_API take_along_dim { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::take_along_dim"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "take_along_dim(Tensor self, Tensor indices, int? dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & indices, ::std::optional dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & indices, ::std::optional dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/take_ops.h b/backends/pytorch/include/ATen/ops/take_ops.h new file mode 100644 index 000000000..5cb77898a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/take_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API take_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::take"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "take.out(Tensor self, Tensor index, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & index, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & index, at::Tensor & out); +}; + +struct TORCH_API take { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::take"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "take(Tensor self, Tensor index) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & index); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & index); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/tan_ops.h b/backends/pytorch/include/ATen/ops/tan_ops.h new file mode 100644 index 000000000..69d098be3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/tan_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API tan { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tan"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "tan(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API tan_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tan_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "tan_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API tan_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tan"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "tan.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/tanh_ops.h b/backends/pytorch/include/ATen/ops/tanh_ops.h new file mode 100644 index 000000000..d8e5e92ce --- /dev/null +++ b/backends/pytorch/include/ATen/ops/tanh_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API tanh { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tanh"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "tanh(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API tanh_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tanh_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "tanh_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API tanh_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tanh"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "tanh.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/tensor_split_ops.h b/backends/pytorch/include/ATen/ops/tensor_split_ops.h new file mode 100644 index 000000000..15f9079c6 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/tensor_split_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API tensor_split_sections { + using schema = ::std::vector (const at::Tensor &, c10::SymInt, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tensor_split"; + static constexpr const char* overload_name = "sections"; + static constexpr const char* schema_str = "tensor_split.sections(Tensor(a -> *) self, SymInt sections, int dim=0) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, c10::SymInt sections, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt sections, int64_t dim); +}; + +struct TORCH_API tensor_split_indices { + using schema = ::std::vector (const at::Tensor &, c10::SymIntArrayRef, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tensor_split"; + static constexpr const char* overload_name = "indices"; + static constexpr const char* schema_str = "tensor_split.indices(Tensor(a -> *) self, SymInt[] indices, int dim=0) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, c10::SymIntArrayRef indices, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef indices, int64_t dim); +}; + +struct TORCH_API tensor_split_tensor_indices_or_sections { + using schema = ::std::vector (const at::Tensor &, const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tensor_split"; + static constexpr const char* overload_name = "tensor_indices_or_sections"; + static constexpr const char* schema_str = "tensor_split.tensor_indices_or_sections(Tensor(a -> *) self, Tensor tensor_indices_or_sections, int dim=0) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, const at::Tensor & tensor_indices_or_sections, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & tensor_indices_or_sections, int64_t dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/tile_ops.h b/backends/pytorch/include/ATen/ops/tile_ops.h new file mode 100644 index 000000000..2ef8f168d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/tile_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API tile { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tile"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "tile(Tensor self, SymInt[] dims) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef dims); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef dims); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/to_dense_ops.h b/backends/pytorch/include/ATen/ops/to_dense_ops.h new file mode 100644 index 000000000..7cea797e5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/to_dense_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API to_dense { + using schema = at::Tensor (const at::Tensor &, ::std::optional, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_dense"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "to_dense(Tensor self, ScalarType? dtype=None, *, bool? masked_grad=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dtype, ::std::optional masked_grad); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype, ::std::optional masked_grad); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/to_mkldnn_ops.h b/backends/pytorch/include/ATen/ops/to_mkldnn_ops.h new file mode 100644 index 000000000..4fae366ca --- /dev/null +++ b/backends/pytorch/include/ATen/ops/to_mkldnn_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API to_mkldnn { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_mkldnn"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "to_mkldnn(Tensor self, ScalarType? dtype=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype); +}; + +struct TORCH_API to_mkldnn_out { + using schema = at::Tensor & (const at::Tensor &, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_mkldnn"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "to_mkldnn.out(Tensor self, ScalarType? dtype=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, ::std::optional dtype, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/to_ops.h b/backends/pytorch/include/ATen/ops/to_ops.h new file mode 100644 index 000000000..e0eea9994 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/to_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API to_dtype_layout { + using schema = at::Tensor (const at::Tensor &, ::std::optional, ::std::optional, ::std::optional, ::std::optional, bool, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to"; + static constexpr const char* overload_name = "dtype_layout"; + static constexpr const char* schema_str = "to.dtype_layout(Tensor(a) self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory, bool non_blocking, bool copy, ::std::optional memory_format); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dtype, ::std::optional layout, ::std::optional device, ::std::optional pin_memory, bool non_blocking, bool copy, ::std::optional memory_format); +}; + +struct TORCH_API to_device { + using schema = at::Tensor (const at::Tensor &, at::Device, at::ScalarType, bool, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to"; + static constexpr const char* overload_name = "device"; + static constexpr const char* schema_str = "to.device(Tensor(a) self, Device device, ScalarType dtype, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::Device device, at::ScalarType dtype, bool non_blocking, bool copy, ::std::optional memory_format); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Device device, at::ScalarType dtype, bool non_blocking, bool copy, ::std::optional memory_format); +}; + +struct TORCH_API to_dtype { + using schema = at::Tensor (const at::Tensor &, at::ScalarType, bool, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to"; + static constexpr const char* overload_name = "dtype"; + static constexpr const char* schema_str = "to.dtype(Tensor(a) self, ScalarType dtype, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::ScalarType dtype, bool non_blocking, bool copy, ::std::optional memory_format); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::ScalarType dtype, bool non_blocking, bool copy, ::std::optional memory_format); +}; + +struct TORCH_API to_other { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, bool, bool, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to"; + static constexpr const char* overload_name = "other"; + static constexpr const char* schema_str = "to.other(Tensor(a) self, Tensor other, bool non_blocking=False, bool copy=False, MemoryFormat? memory_format=None) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other, bool non_blocking, bool copy, ::std::optional memory_format); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, bool non_blocking, bool copy, ::std::optional memory_format); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/to_padded_tensor_ops.h b/backends/pytorch/include/ATen/ops/to_padded_tensor_ops.h new file mode 100644 index 000000000..33f898037 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/to_padded_tensor_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API to_padded_tensor { + using schema = at::Tensor (const at::Tensor &, double, at::OptionalSymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_padded_tensor"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "to_padded_tensor(Tensor self, float padding, SymInt[]? output_size=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, double padding, at::OptionalSymIntArrayRef output_size); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double padding, at::OptionalSymIntArrayRef output_size); +}; + +struct TORCH_API to_padded_tensor_out { + using schema = at::Tensor & (const at::Tensor &, double, at::OptionalSymIntArrayRef, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_padded_tensor"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "to_padded_tensor.out(Tensor self, float padding, SymInt[]? output_size=None, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, double padding, at::OptionalSymIntArrayRef output_size, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double padding, at::OptionalSymIntArrayRef output_size, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/to_sparse_bsc_ops.h b/backends/pytorch/include/ATen/ops/to_sparse_bsc_ops.h new file mode 100644 index 000000000..a06435c67 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/to_sparse_bsc_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API to_sparse_bsc { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_sparse_bsc"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "to_sparse_bsc(Tensor self, int[2] blocksize, int? dense_dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/to_sparse_bsr_ops.h b/backends/pytorch/include/ATen/ops/to_sparse_bsr_ops.h new file mode 100644 index 000000000..1e6effd90 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/to_sparse_bsr_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API to_sparse_bsr { + using schema = at::Tensor (const at::Tensor &, at::IntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_sparse_bsr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "to_sparse_bsr(Tensor self, int[2] blocksize, int? dense_dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef blocksize, ::std::optional dense_dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/to_sparse_csc_ops.h b/backends/pytorch/include/ATen/ops/to_sparse_csc_ops.h new file mode 100644 index 000000000..a211c797e --- /dev/null +++ b/backends/pytorch/include/ATen/ops/to_sparse_csc_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API to_sparse_csc { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_sparse_csc"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "to_sparse_csc(Tensor self, int? dense_dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dense_dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/to_sparse_csr_ops.h b/backends/pytorch/include/ATen/ops/to_sparse_csr_ops.h new file mode 100644 index 000000000..9e4ad0ef5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/to_sparse_csr_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API to_sparse_csr { + using schema = at::Tensor (const at::Tensor &, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_sparse_csr"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "to_sparse_csr(Tensor self, int? dense_dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional dense_dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/to_sparse_ops.h b/backends/pytorch/include/ATen/ops/to_sparse_ops.h new file mode 100644 index 000000000..b01320e25 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/to_sparse_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API to_sparse_sparse_dim { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_sparse"; + static constexpr const char* overload_name = "sparse_dim"; + static constexpr const char* schema_str = "to_sparse.sparse_dim(Tensor self, int sparse_dim) -> Tensor"; + static at::Tensor call(const at::Tensor & self, int64_t sparse_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t sparse_dim); +}; + +struct TORCH_API to_sparse { + using schema = at::Tensor (const at::Tensor &, ::std::optional, at::OptionalIntArrayRef, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::to_sparse"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "to_sparse(Tensor self, *, Layout? layout=None, int[2]? blocksize=None, int? dense_dim=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, ::std::optional layout, at::OptionalIntArrayRef blocksize, ::std::optional dense_dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, ::std::optional layout, at::OptionalIntArrayRef blocksize, ::std::optional dense_dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/topk_ops.h b/backends/pytorch/include/ATen/ops/topk_ops.h new file mode 100644 index 000000000..7cdc7006b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/topk_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API topk_values { + using schema = ::std::tuple (const at::Tensor &, c10::SymInt, int64_t, bool, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::topk"; + static constexpr const char* overload_name = "values"; + static constexpr const char* schema_str = "topk.values(Tensor self, SymInt k, int dim=-1, bool largest=True, bool sorted=True, *, Tensor(a!) values, Tensor(b!) indices) -> (Tensor(a!) values, Tensor(b!) indices)"; + static ::std::tuple call(const at::Tensor & self, c10::SymInt k, int64_t dim, bool largest, bool sorted, at::Tensor & values, at::Tensor & indices); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt k, int64_t dim, bool largest, bool sorted, at::Tensor & values, at::Tensor & indices); +}; + +struct TORCH_API topk { + using schema = ::std::tuple (const at::Tensor &, c10::SymInt, int64_t, bool, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::topk"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "topk(Tensor self, SymInt k, int dim=-1, bool largest=True, bool sorted=True) -> (Tensor values, Tensor indices)"; + static ::std::tuple call(const at::Tensor & self, c10::SymInt k, int64_t dim, bool largest, bool sorted); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt k, int64_t dim, bool largest, bool sorted); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/trace_ops.h b/backends/pytorch/include/ATen/ops/trace_ops.h new file mode 100644 index 000000000..08b55f2a4 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/trace_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API trace { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::trace"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "trace(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API trace_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::trace"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "trace.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/transpose_ops.h b/backends/pytorch/include/ATen/ops/transpose_ops.h new file mode 100644 index 000000000..a9f458df3 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/transpose_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API transpose_int { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::transpose"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "transpose.int(Tensor(a) self, int dim0, int dim1) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t dim0, int64_t dim1); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim0, int64_t dim1); +}; + +struct TORCH_API transpose_Dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::transpose"; + static constexpr const char* overload_name = "Dimname"; + static constexpr const char* schema_str = "transpose.Dimname(Tensor(a) self, Dimname dim0, Dimname dim1) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim0, at::Dimname dim1); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim0, at::Dimname dim1); +}; + +struct TORCH_API transpose_ { + using schema = at::Tensor & (at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::transpose_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "transpose_(Tensor(a!) self, int dim0, int dim1) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim0, int64_t dim1); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim0, int64_t dim1); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/triangular_solve_ops.h b/backends/pytorch/include/ATen/ops/triangular_solve_ops.h new file mode 100644 index 000000000..7c0e88d52 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/triangular_solve_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API triangular_solve_X { + using schema = ::std::tuple (const at::Tensor &, const at::Tensor &, bool, bool, bool, at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::triangular_solve"; + static constexpr const char* overload_name = "X"; + static constexpr const char* schema_str = "triangular_solve.X(Tensor self, Tensor A, bool upper=True, bool transpose=False, bool unitriangular=False, *, Tensor(a!) X, Tensor(b!) M) -> (Tensor(a!) solution, Tensor(b!) cloned_coefficient)"; + static ::std::tuple call(const at::Tensor & self, const at::Tensor & A, bool upper, bool transpose, bool unitriangular, at::Tensor & X, at::Tensor & M); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & A, bool upper, bool transpose, bool unitriangular, at::Tensor & X, at::Tensor & M); +}; + +struct TORCH_API triangular_solve { + using schema = ::std::tuple (const at::Tensor &, const at::Tensor &, bool, bool, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::triangular_solve"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "triangular_solve(Tensor self, Tensor A, bool upper=True, bool transpose=False, bool unitriangular=False) -> (Tensor solution, Tensor cloned_coefficient)"; + static ::std::tuple call(const at::Tensor & self, const at::Tensor & A, bool upper, bool transpose, bool unitriangular); + static ::std::tuple redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & A, bool upper, bool transpose, bool unitriangular); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/tril_ops.h b/backends/pytorch/include/ATen/ops/tril_ops.h new file mode 100644 index 000000000..3fbf6401b --- /dev/null +++ b/backends/pytorch/include/ATen/ops/tril_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API tril_ { + using schema = at::Tensor & (at::Tensor &, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tril_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "tril_(Tensor(a!) self, SymInt diagonal=0) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, c10::SymInt diagonal); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, c10::SymInt diagonal); +}; + +struct TORCH_API tril_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymInt, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tril"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "tril.out(Tensor self, SymInt diagonal=0, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymInt diagonal, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt diagonal, at::Tensor & out); +}; + +struct TORCH_API tril { + using schema = at::Tensor (const at::Tensor &, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::tril"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "tril(Tensor self, SymInt diagonal=0) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymInt diagonal); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt diagonal); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/triu_ops.h b/backends/pytorch/include/ATen/ops/triu_ops.h new file mode 100644 index 000000000..cc585ffa4 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/triu_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API triu_ { + using schema = at::Tensor & (at::Tensor &, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::triu_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "triu_(Tensor(a!) self, SymInt diagonal=0) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, c10::SymInt diagonal); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, c10::SymInt diagonal); +}; + +struct TORCH_API triu_out { + using schema = at::Tensor & (const at::Tensor &, c10::SymInt, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::triu"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "triu.out(Tensor self, SymInt diagonal=0, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, c10::SymInt diagonal, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt diagonal, at::Tensor & out); +}; + +struct TORCH_API triu { + using schema = at::Tensor (const at::Tensor &, c10::SymInt); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::triu"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "triu(Tensor self, SymInt diagonal=0) -> Tensor"; + static at::Tensor call(const at::Tensor & self, c10::SymInt diagonal); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt diagonal); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/true_divide_ops.h b/backends/pytorch/include/ATen/ops/true_divide_ops.h new file mode 100644 index 000000000..4c2a6f786 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/true_divide_ops.h @@ -0,0 +1,78 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API true_divide_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::true_divide"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "true_divide.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API true_divide__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::true_divide_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "true_divide_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API true_divide_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::true_divide"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "true_divide.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API true_divide_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::true_divide"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "true_divide.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API true_divide__Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::true_divide_"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "true_divide_.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/trunc_ops.h b/backends/pytorch/include/ATen/ops/trunc_ops.h new file mode 100644 index 000000000..0794d4424 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/trunc_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API trunc { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::trunc"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "trunc(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +struct TORCH_API trunc_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::trunc_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "trunc_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API trunc_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::trunc"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "trunc.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/type_as_ops.h b/backends/pytorch/include/ATen/ops/type_as_ops.h new file mode 100644 index 000000000..b75388dc5 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/type_as_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API type_as { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::type_as"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "type_as(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/unbind_ops.h b/backends/pytorch/include/ATen/ops/unbind_ops.h new file mode 100644 index 000000000..d04c89d57 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/unbind_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API unbind_int { + using schema = ::std::vector (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unbind"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "unbind.int(Tensor(a -> *) self, int dim=0) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim); +}; + +struct TORCH_API unbind_Dimname { + using schema = ::std::vector (const at::Tensor &, at::Dimname); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unbind"; + static constexpr const char* overload_name = "Dimname"; + static constexpr const char* schema_str = "unbind.Dimname(Tensor(a -> *) self, Dimname dim) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, at::Dimname dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/unflatten_ops.h b/backends/pytorch/include/ATen/ops/unflatten_ops.h new file mode 100644 index 000000000..c8492037c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/unflatten_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API unflatten_int { + using schema = at::Tensor (const at::Tensor &, int64_t, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unflatten"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "unflatten.int(Tensor(a) self, int dim, SymInt[] sizes) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t dim, c10::SymIntArrayRef sizes); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim, c10::SymIntArrayRef sizes); +}; + +struct TORCH_API unflatten_Dimname { + using schema = at::Tensor (const at::Tensor &, at::Dimname, c10::SymIntArrayRef, at::DimnameList); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unflatten"; + static constexpr const char* overload_name = "Dimname"; + static constexpr const char* schema_str = "unflatten.Dimname(Tensor(a) self, Dimname dim, SymInt[] sizes, Dimname[] names) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::Dimname dim, c10::SymIntArrayRef sizes, at::DimnameList names); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Dimname dim, c10::SymIntArrayRef sizes, at::DimnameList names); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/unfold_ops.h b/backends/pytorch/include/ATen/ops/unfold_ops.h new file mode 100644 index 000000000..bc67d28ed --- /dev/null +++ b/backends/pytorch/include/ATen/ops/unfold_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API unfold { + using schema = at::Tensor (const at::Tensor &, int64_t, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unfold"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "unfold(Tensor(a) self, int dimension, int size, int step) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t dimension, int64_t size, int64_t step); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dimension, int64_t size, int64_t step); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/uniform_ops.h b/backends/pytorch/include/ATen/ops/uniform_ops.h new file mode 100644 index 000000000..d33ffa207 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/uniform_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API uniform_ { + using schema = at::Tensor & (at::Tensor &, double, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::uniform_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "uniform_(Tensor(a!) self, float from=0, float to=1, *, Generator? generator=None) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, double from, double to, ::std::optional generator); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, double from, double to, ::std::optional generator); +}; + +struct TORCH_API uniform_out { + using schema = at::Tensor & (const at::Tensor &, double, double, ::std::optional, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::uniform"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "uniform.out(Tensor self, float from=0, float to=1, *, Generator? generator=None, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, double from, double to, ::std::optional generator, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double from, double to, ::std::optional generator, at::Tensor & out); +}; + +struct TORCH_API uniform { + using schema = at::Tensor (const at::Tensor &, double, double, ::std::optional); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::uniform"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "uniform(Tensor self, float from=0, float to=1, *, Generator? generator=None) -> Tensor"; + static at::Tensor call(const at::Tensor & self, double from, double to, ::std::optional generator); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, double from, double to, ::std::optional generator); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/unsafe_chunk_ops.h b/backends/pytorch/include/ATen/ops/unsafe_chunk_ops.h new file mode 100644 index 000000000..d32a86ddf --- /dev/null +++ b/backends/pytorch/include/ATen/ops/unsafe_chunk_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API unsafe_chunk { + using schema = ::std::vector (const at::Tensor &, int64_t, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unsafe_chunk"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "unsafe_chunk(Tensor self, int chunks, int dim=0) -> Tensor[]"; + static ::std::vector call(const at::Tensor & self, int64_t chunks, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t chunks, int64_t dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/unsafe_split_ops.h b/backends/pytorch/include/ATen/ops/unsafe_split_ops.h new file mode 100644 index 000000000..785055511 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/unsafe_split_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API unsafe_split_Tensor { + using schema = ::std::vector (const at::Tensor &, c10::SymInt, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unsafe_split"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "unsafe_split.Tensor(Tensor self, SymInt split_size, int dim=0) -> Tensor[]"; + static ::std::vector call(const at::Tensor & self, c10::SymInt split_size, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt split_size, int64_t dim); +}; + +struct TORCH_API unsafe_split_Tensor_out { + using schema = void (const at::Tensor &, c10::SymInt, int64_t, at::TensorList); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unsafe_split"; + static constexpr const char* overload_name = "Tensor_out"; + static constexpr const char* schema_str = "unsafe_split.Tensor_out(Tensor self, SymInt split_size, int dim=0, *, Tensor(a!)[] out) -> ()"; + static void call(const at::Tensor & self, c10::SymInt split_size, int64_t dim, at::TensorList out); + static void redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymInt split_size, int64_t dim, at::TensorList out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/unsafe_split_with_sizes_ops.h b/backends/pytorch/include/ATen/ops/unsafe_split_with_sizes_ops.h new file mode 100644 index 000000000..31b1ef26c --- /dev/null +++ b/backends/pytorch/include/ATen/ops/unsafe_split_with_sizes_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API unsafe_split_with_sizes { + using schema = ::std::vector (const at::Tensor &, c10::SymIntArrayRef, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unsafe_split_with_sizes"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "unsafe_split_with_sizes(Tensor self, SymInt[] split_sizes, int dim=0) -> Tensor[]"; + static ::std::vector call(const at::Tensor & self, c10::SymIntArrayRef split_sizes, int64_t dim); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef split_sizes, int64_t dim); +}; + +struct TORCH_API unsafe_split_with_sizes_out { + using schema = void (const at::Tensor &, c10::SymIntArrayRef, int64_t, at::TensorList); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unsafe_split_with_sizes"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "unsafe_split_with_sizes.out(Tensor self, SymInt[] split_sizes, int dim=0, *, Tensor(a!)[] out) -> ()"; + static void call(const at::Tensor & self, c10::SymIntArrayRef split_sizes, int64_t dim, at::TensorList out); + static void redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef split_sizes, int64_t dim, at::TensorList out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/unsqueeze_ops.h b/backends/pytorch/include/ATen/ops/unsqueeze_ops.h new file mode 100644 index 000000000..dbffca2bb --- /dev/null +++ b/backends/pytorch/include/ATen/ops/unsqueeze_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API unsqueeze { + using schema = at::Tensor (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unsqueeze"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "unsqueeze(Tensor(a) self, int dim) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, int64_t dim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t dim); +}; + +struct TORCH_API unsqueeze_ { + using schema = at::Tensor & (at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::unsqueeze_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "unsqueeze_(Tensor(a!) self, int dim) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, int64_t dim); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, int64_t dim); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/values_ops.h b/backends/pytorch/include/ATen/ops/values_ops.h new file mode 100644 index 000000000..bd6f7f3c2 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/values_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API values { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::values"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "values(Tensor(a) self) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/var_ops.h b/backends/pytorch/include/ATen/ops/var_ops.h new file mode 100644 index 000000000..b61b7ba2f --- /dev/null +++ b/backends/pytorch/include/ATen/ops/var_ops.h @@ -0,0 +1,122 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API var { + using schema = at::Tensor (const at::Tensor &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::var"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "var(Tensor self, bool unbiased=True) -> Tensor"; + static at::Tensor call(const at::Tensor & self, bool unbiased); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, bool unbiased); +}; + +struct TORCH_API var_dim { + using schema = at::Tensor (const at::Tensor &, at::OptionalIntArrayRef, bool, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::var"; + static constexpr const char* overload_name = "dim"; + static constexpr const char* schema_str = "var.dim(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool unbiased, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool unbiased, bool keepdim); +}; + +struct TORCH_API var_correction { + using schema = at::Tensor (const at::Tensor &, at::OptionalIntArrayRef, const ::std::optional &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::var"; + static constexpr const char* overload_name = "correction"; + static constexpr const char* schema_str = "var.correction(Tensor self, int[1]? dim=None, *, Scalar? correction=None, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::OptionalIntArrayRef dim, const ::std::optional & correction, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, const ::std::optional & correction, bool keepdim); +}; + +struct TORCH_API var_out { + using schema = at::Tensor & (const at::Tensor &, at::OptionalIntArrayRef, bool, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::var"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "var.out(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::OptionalIntArrayRef dim, bool unbiased, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, bool unbiased, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API var_correction_out { + using schema = at::Tensor & (const at::Tensor &, at::OptionalIntArrayRef, const ::std::optional &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::var"; + static constexpr const char* overload_name = "correction_out"; + static constexpr const char* schema_str = "var.correction_out(Tensor self, int[1]? dim=None, *, Scalar? correction=None, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::OptionalIntArrayRef dim, const ::std::optional & correction, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::OptionalIntArrayRef dim, const ::std::optional & correction, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API var_names_dim { + using schema = at::Tensor (const at::Tensor &, at::DimnameList, bool, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::var"; + static constexpr const char* overload_name = "names_dim"; + static constexpr const char* schema_str = "var.names_dim(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::DimnameList dim, bool unbiased, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, bool unbiased, bool keepdim); +}; + +struct TORCH_API var_names_out { + using schema = at::Tensor & (const at::Tensor &, at::DimnameList, bool, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::var"; + static constexpr const char* overload_name = "names_out"; + static constexpr const char* schema_str = "var.names_out(Tensor self, Dimname[1] dim, bool unbiased=True, bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::DimnameList dim, bool unbiased, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, bool unbiased, bool keepdim, at::Tensor & out); +}; + +struct TORCH_API var_correction_names { + using schema = at::Tensor (const at::Tensor &, at::DimnameList, const ::std::optional &, bool); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::var"; + static constexpr const char* overload_name = "correction_names"; + static constexpr const char* schema_str = "var.correction_names(Tensor self, Dimname[1] dim, *, Scalar? correction=None, bool keepdim=False) -> Tensor"; + static at::Tensor call(const at::Tensor & self, at::DimnameList dim, const ::std::optional & correction, bool keepdim); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, const ::std::optional & correction, bool keepdim); +}; + +struct TORCH_API var_correction_names_out { + using schema = at::Tensor & (const at::Tensor &, at::DimnameList, const ::std::optional &, bool, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::var"; + static constexpr const char* overload_name = "correction_names_out"; + static constexpr const char* schema_str = "var.correction_names_out(Tensor self, Dimname[1] dim, *, Scalar? correction=None, bool keepdim=False, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::DimnameList dim, const ::std::optional & correction, bool keepdim, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::DimnameList dim, const ::std::optional & correction, bool keepdim, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/vdot_ops.h b/backends/pytorch/include/ATen/ops/vdot_ops.h new file mode 100644 index 000000000..8ced44d2a --- /dev/null +++ b/backends/pytorch/include/ATen/ops/vdot_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API vdot { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::vdot"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "vdot(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API vdot_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::vdot"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "vdot.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/view_as_ops.h b/backends/pytorch/include/ATen/ops/view_as_ops.h new file mode 100644 index 000000000..b087a8e9d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/view_as_ops.h @@ -0,0 +1,34 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API view_as { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::view_as"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "view_as(Tensor(a) self, Tensor other) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/view_ops.h b/backends/pytorch/include/ATen/ops/view_ops.h new file mode 100644 index 000000000..36de89dbd --- /dev/null +++ b/backends/pytorch/include/ATen/ops/view_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API view { + using schema = at::Tensor (const at::Tensor &, c10::SymIntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::view"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "view(Tensor(a) self, SymInt[] size) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, c10::SymIntArrayRef size); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, c10::SymIntArrayRef size); +}; + +struct TORCH_API view_dtype { + using schema = at::Tensor (const at::Tensor &, at::ScalarType); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::view"; + static constexpr const char* overload_name = "dtype"; + static constexpr const char* schema_str = "view.dtype(Tensor(a) self, ScalarType dtype) -> Tensor(a)"; + static at::Tensor call(const at::Tensor & self, at::ScalarType dtype); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::ScalarType dtype); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/vsplit_ops.h b/backends/pytorch/include/ATen/ops/vsplit_ops.h new file mode 100644 index 000000000..81f818d75 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/vsplit_ops.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API vsplit_int { + using schema = ::std::vector (const at::Tensor &, int64_t); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::vsplit"; + static constexpr const char* overload_name = "int"; + static constexpr const char* schema_str = "vsplit.int(Tensor(a -> *) self, int sections) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, int64_t sections); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, int64_t sections); +}; + +struct TORCH_API vsplit_array { + using schema = ::std::vector (const at::Tensor &, at::IntArrayRef); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::vsplit"; + static constexpr const char* overload_name = "array"; + static constexpr const char* schema_str = "vsplit.array(Tensor(a -> *) self, int[] indices) -> Tensor(a)[]"; + static ::std::vector call(const at::Tensor & self, at::IntArrayRef indices); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::IntArrayRef indices); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/where_ops.h b/backends/pytorch/include/ATen/ops/where_ops.h new file mode 100644 index 000000000..4fdef5d58 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/where_ops.h @@ -0,0 +1,89 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API where_self { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::where"; + static constexpr const char* overload_name = "self"; + static constexpr const char* schema_str = "where.self(Tensor condition, Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & condition, const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & condition, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API where_self_out { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::where"; + static constexpr const char* overload_name = "self_out"; + static constexpr const char* schema_str = "where.self_out(Tensor condition, Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & condition, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & condition, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API where_ScalarSelf { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::where"; + static constexpr const char* overload_name = "ScalarSelf"; + static constexpr const char* schema_str = "where.ScalarSelf(Tensor condition, Scalar self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & condition, const at::Scalar & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & condition, const at::Scalar & self, const at::Tensor & other); +}; + +struct TORCH_API where_ScalarOther { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::where"; + static constexpr const char* overload_name = "ScalarOther"; + static constexpr const char* schema_str = "where.ScalarOther(Tensor condition, Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & condition, const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & condition, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API where_Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::where"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "where.Scalar(Tensor condition, Scalar self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & condition, const at::Scalar & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & condition, const at::Scalar & self, const at::Scalar & other); +}; + +struct TORCH_API where { + using schema = ::std::vector (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::where"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "where(Tensor condition) -> Tensor[]"; + static ::std::vector call(const at::Tensor & condition); + static ::std::vector redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & condition); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/xlogy_ops.h b/backends/pytorch/include/ATen/ops/xlogy_ops.h new file mode 100644 index 000000000..de881978d --- /dev/null +++ b/backends/pytorch/include/ATen/ops/xlogy_ops.h @@ -0,0 +1,111 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API xlogy_Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::xlogy"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "xlogy.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API xlogy_Scalar_Self { + using schema = at::Tensor (const at::Scalar &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::xlogy"; + static constexpr const char* overload_name = "Scalar_Self"; + static constexpr const char* schema_str = "xlogy.Scalar_Self(Scalar self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Scalar & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other); +}; + +struct TORCH_API xlogy_Scalar_Other { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::xlogy"; + static constexpr const char* overload_name = "Scalar_Other"; + static constexpr const char* schema_str = "xlogy.Scalar_Other(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API xlogy__Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::xlogy_"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "xlogy_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API xlogy__Scalar_Other { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::xlogy_"; + static constexpr const char* overload_name = "Scalar_Other"; + static constexpr const char* schema_str = "xlogy_.Scalar_Other(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API xlogy_OutTensor { + using schema = at::Tensor & (const at::Tensor &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::xlogy"; + static constexpr const char* overload_name = "OutTensor"; + static constexpr const char* schema_str = "xlogy.OutTensor(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API xlogy_OutScalar_Self { + using schema = at::Tensor & (const at::Scalar &, const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::xlogy"; + static constexpr const char* overload_name = "OutScalar_Self"; + static constexpr const char* schema_str = "xlogy.OutScalar_Self(Scalar self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Scalar & self, const at::Tensor & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Scalar & self, const at::Tensor & other, at::Tensor & out); +}; + +struct TORCH_API xlogy_OutScalar_Other { + using schema = at::Tensor & (const at::Tensor &, const at::Scalar &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::xlogy"; + static constexpr const char* overload_name = "OutScalar_Other"; + static constexpr const char* schema_str = "xlogy.OutScalar_Other(Tensor self, Scalar other, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, const at::Scalar & other, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other, at::Tensor & out); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/xor_ops.h b/backends/pytorch/include/ATen/ops/xor_ops.h new file mode 100644 index 000000000..33cde2fe7 --- /dev/null +++ b/backends/pytorch/include/ATen/ops/xor_ops.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API __xor___Scalar { + using schema = at::Tensor (const at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__xor__"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "__xor__.Scalar(Tensor self, Scalar other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Scalar & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API __xor___Tensor { + using schema = at::Tensor (const at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__xor__"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "__xor__.Tensor(Tensor self, Tensor other) -> Tensor"; + static at::Tensor call(const at::Tensor & self, const at::Tensor & other); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, const at::Tensor & other); +}; + +struct TORCH_API __ixor___Scalar { + using schema = at::Tensor & (at::Tensor &, const at::Scalar &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__ixor__"; + static constexpr const char* overload_name = "Scalar"; + static constexpr const char* schema_str = "__ixor__.Scalar(Tensor(a!) self, Scalar other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Scalar & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Scalar & other); +}; + +struct TORCH_API __ixor___Tensor { + using schema = at::Tensor & (at::Tensor &, const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::__ixor__"; + static constexpr const char* overload_name = "Tensor"; + static constexpr const char* schema_str = "__ixor__.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self, const at::Tensor & other); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self, const at::Tensor & other); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/ops/zero_ops.h b/backends/pytorch/include/ATen/ops/zero_ops.h new file mode 100644 index 000000000..a06946acc --- /dev/null +++ b/backends/pytorch/include/ATen/ops/zero_ops.h @@ -0,0 +1,56 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// @generated by torchgen/gen.py from Operator.h + +#include +#include +#include + +// Forward declarations of any types needed in the operator signatures. +// We can't directly include these classes because it will cause circular include dependencies. +// This file is included by TensorBody.h, which defines the Tensor class. +#include + +namespace at { +namespace _ops { + + +struct TORCH_API zero_ { + using schema = at::Tensor & (at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::zero_"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "zero_(Tensor(a!) self) -> Tensor(a!)"; + static at::Tensor & call(at::Tensor & self); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, at::Tensor & self); +}; + +struct TORCH_API zero_out { + using schema = at::Tensor & (const at::Tensor &, at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::zero"; + static constexpr const char* overload_name = "out"; + static constexpr const char* schema_str = "zero.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)"; + static at::Tensor & call(const at::Tensor & self, at::Tensor & out); + static at::Tensor & redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self, at::Tensor & out); +}; + +struct TORCH_API zero { + using schema = at::Tensor (const at::Tensor &); + using ptr_schema = schema*; + // See Note [static constexpr char* members for windows NVCC] + static constexpr const char* name = "aten::zero"; + static constexpr const char* overload_name = ""; + static constexpr const char* schema_str = "zero(Tensor self) -> Tensor"; + static at::Tensor call(const at::Tensor & self); + static at::Tensor redispatch(c10::DispatchKeySet dispatchKeySet, const at::Tensor & self); +}; + +}} // namespace at::_ops + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/ATen/record_function.h b/backends/pytorch/include/ATen/record_function.h new file mode 100644 index 000000000..c0ca1a69e --- /dev/null +++ b/backends/pytorch/include/ATen/record_function.h @@ -0,0 +1,799 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include + +namespace c10 { +class TORCH_API OperatorHandle; +} + +namespace at { + +// Function name to record NCCL metadata +extern TORCH_API const std::string kParamCommsCallName; + +// Kind of record function scope; +enum class C10_API_ENUM RecordScope : uint8_t { + // c10/ATen ops, autograd nodes + FUNCTION = 0, + // Functions/nodes called from the autograd + BACKWARD_FUNCTION, + // TorchScript functions, methods + TORCHSCRIPT_FUNCTION, + // Kernel Function dtype Tag + KERNEL_FUNCTION_DTYPE, + // Torchbind custom class, + CUSTOM_CLASS, + // Generic Build Feature + BUILD_FEATURE, + // Kernel Function dtype Tag + LITE_INTERPRETER, + // User defined scope (e.g. with record_function()) + USER_SCOPE, + // Scopes for static runtime, a specialized TorchScript interpreter + STATIC_RUNTIME_OP, + STATIC_RUNTIME_MODEL, + NUM_SCOPES, // must be the last in the list +}; + +} // namespace at + +namespace std { +template <> +struct hash { + size_t operator()(const at::RecordScope& sc) const { + return static_cast(sc); + } +}; +} // namespace std + +namespace at { + +struct TORCH_API StringView { + StringView() : StringView(nullptr) {} + explicit StringView(const char* str_ptr) + : owned_str_ptr_(nullptr), str_ptr_(str_ptr) {} + explicit StringView(std::string str) + : owned_str_ptr_(std::make_shared(std::move(str))), + str_ptr_(owned_str_ptr_->c_str()) {} + + const char* str() const { + return str_ptr_; + } + + friend std::ostream& operator<<(std::ostream& os, const StringView& dt) { + os << dt.str(); + return os; + } + + friend bool operator==(const StringView& lhs, const StringView& rhs) { + return strcmp(lhs.str(), rhs.str()) == 0; + } + + friend bool operator!=(const StringView& lhs, const StringView& rhs) { + return !(lhs == rhs); + } + + private: + std::shared_ptr owned_str_ptr_; + const char* str_ptr_; +}; + +// Soft limit on the number of callbacks to use; +constexpr std::size_t kSoftLimitCallbacks = 4; + +// An abstract base class for various observer contexts that can be attached to +// the RecordFunction. +struct ObserverContext { + virtual ~ObserverContext() = default; + + protected: + ObserverContext() = default; +}; + +typedef c10::SmallVector CallbackHandles; +typedef c10::SmallVector, kSoftLimitCallbacks> + ObserverContextList; +typedef uint64_t RecordFunctionHandle; +struct RecordFunction; + +// +// PyTorch callbacks/observers API: +// + +/** + * RecordFunctionCallback represents a pair of callbacks to be used with + * RecordFunction, members: + * start, end - the callbacks to run when entering and exiting the scope; + * optionally, the start callback may return an ObserverContext which will + * be passed to the end callback, use appropriate constructor accordingly. + * needs_inputs - whether the callbacks need the inputs passed from the + * observed function/range; NOTE: passing the inputs incurs an additional + * overhead; sampling_probability - if not 1.0, then the callback is + * probabilistically sampled to run; NOTE: start and end callbacks always run as + * a pair and are sampled together; scopes - types of scopes to execute the + * callbacks on (see RecordScope); passing empty set means the callbacks will be + * executed for all possible scope types should_run - optional function that + * returns whether this callback should run; overwrites the effect of setting + * sampling_probability + */ +class TORCH_API RecordFunctionCallback { + public: + using StartCallback = + std::unique_ptr (*)(const RecordFunction&); + using EndCallback = void (*)(const RecordFunction&, ObserverContext*); + + // This interface supports observers that require passing an ObserverContext + // between start and end callbacks. + explicit RecordFunctionCallback( + StartCallback start, + EndCallback end = nullptr) + : start_(start), end_(end) { + scopes_.fill(true); + } + + RecordFunctionCallback& needsInputs(bool needs_inputs) { + needs_inputs_ = needs_inputs; + return *this; + } + + RecordFunctionCallback& needsOutputs(bool needs_outputs) { + needs_outputs_ = needs_outputs; + return *this; + } + + RecordFunctionCallback& needsIds(bool needs_ids) { + needs_ids_ = needs_ids; + return *this; + } + + RecordFunctionCallback& samplingProb(double sampling_prob) { + TORCH_CHECK( + sampling_prob >= 0.0 && sampling_prob <= 1.0, + "Invalid sampling probability"); + sampling_prob_ = sampling_prob; + return *this; + } + + RecordFunctionCallback& scopes( + const std::unordered_set>& scopes) { + if (!scopes.empty()) { + scopes_.fill(false); + for (auto sc : scopes) { + scopes_[static_cast(sc)] = true; + } + } else { + scopes_.fill(true); + } + return *this; + } + + bool needsInputs() const { + return needs_inputs_; + } + + bool needsOutputs() const { + return needs_outputs_; + } + + bool needsIds() const { + return needs_ids_; + } + + double samplingProb() const { + return sampling_prob_; + } + + bool checkScope(RecordScope sc) const { + return scopes_[(size_t)sc]; + } + + StartCallback start() const { + return start_; + } + + EndCallback end() const { + return end_; + } + + private: + StartCallback start_; + EndCallback end_; + double sampling_prob_ = 1.0; + std::array(RecordScope::NUM_SCOPES)> scopes_ = {}; + bool needs_inputs_ = false; + bool needs_outputs_ = false; + bool needs_ids_ = false; +}; + +// Notes: +// - two types of callbacks are provided: thread local and global +// - thread local callbacks are added/removed only for the given thread +// and are stored locally for each thread and separately from the list +// of the global callbacks +// - global callbacks are stored in a single per process list and are +// invoked by every RecordFunction, in addition to the thread local +// callbacks specific to the given thread +// - we allow the added callbacks to be sampled, by specifying a sampling +// probability for each callback pair, if the start callback is +// not picked to run, the corresponding end callback won't be called +// - a typical use case for the global callbacks is passive monitoring +// in the background (e.g. fleet-wide monitoring), without focusing on +// the specific piece of code +// - in contrast, thread local callbacks are enabled locally, on demand, +// for the specific piece of code (range) and are not sampled +// - a typical use case for thread local callbacks is profiler and code +// execution tracer +// - note, thread local callbacks are automatically propagated with +// ThreadLocalState across JIT continuations and async tasks (at::launch) + +typedef uint64_t CallbackHandle; + +constexpr CallbackHandle INVALID_CALLBACK_HANDLE{0}; + +// It is unnecessary to use atomic operations for enabling +// thread-local function callbacks. Moreover, it prevents saving to +// ThreadLocalState because std::atomic is non-copyable. +struct RecordFunctionCallbacksEntry { + RecordFunctionCallbacksEntry(RecordFunctionCallback cb, CallbackHandle h) + : callback_(cb), handle_(h) {} + + RecordFunctionCallback callback_; + bool enabled_{true}; + CallbackHandle handle_; +}; + +// Holds pairs (callbacks, unique_id) +using RecordFunctionCallbacks = std::vector; + +// Generated by the callback managers to determine which functions to run. +struct StepCallbacks { + StepCallbacks() = default; + StepCallbacks(uint64_t thread_id, RecordScope scope) + : thread_id_{thread_id}, scope_{scope} {} + + bool empty() const { + return callbacks_.empty(); + } + + struct StartEndPair { + RecordFunctionCallback::StartCallback start_; + RecordFunctionCallback::EndCallback end_; + }; + + using StartEndPairs = c10::SmallVector; + + StartEndPairs callbacks_; + uint64_t thread_id_{0}; + RecordScope scope_{RecordScope::FUNCTION}; + bool needs_inputs_{false}; + bool needs_outputs_{false}; + bool needs_ids_{false}; +}; + +struct TORCH_API RecordFunction { + // Default constructor is used with before function called afterwards: + // scope - record scope that this function tracks + // pre_sampled - whether this RecordFunction was already pre-sampled with + // kLowProb probability + explicit RecordFunction(RecordScope scope = RecordScope::FUNCTION); + explicit RecordFunction(StepCallbacks&& step_callbacks); + + using schema_ref_t = std::reference_wrapper; + using FunctionDescriptor = std::variant; + + void before( + FunctionDescriptor fn, + c10::ArrayRef args, + int64_t current_sequence_nr = -1) { + if (!isActive()) { + return; + } + inputs_ = args; + before(fn, current_sequence_nr); + } + + void before( + FunctionDescriptor fn, + c10::ArrayRef args, + const std::unordered_map* kwargs, + int64_t current_sequence_nr = -1) { + if (!isActive()) { + return; + } + kwinputs_ = *kwargs; + before(fn, args, current_sequence_nr); + } + + void before( + FunctionDescriptor fn, + const std::unordered_map* kwargs, + int64_t current_sequence_nr = -1) { + if (!isActive()) { + return; + } + kwinputs_ = *kwargs; + before(fn, current_sequence_nr); + } + + void before( + FunctionDescriptor fn, + const std::vector* args, + int64_t current_sequence_nr = -1) { + before( + fn, + c10::ArrayRef(args->data(), args->size()), + current_sequence_nr); + } + + void before( + FunctionDescriptor fn, + const std::vector* args, + const std::unordered_map* kwargs, + int64_t current_sequence_nr = -1) { + if (!isActive()) { + return; + } + kwinputs_ = *kwargs; + before(std::move(fn), args, current_sequence_nr); + } + + // Destructor calls end callbacks + virtual ~RecordFunction(); + + RecordFunction(const RecordFunction&) = delete; + RecordFunction& operator=(const RecordFunction&) = delete; + RecordFunction(RecordFunction&&) = delete; + RecordFunction& operator=(RecordFunction&&) = delete; + + const char* name() const; + const char* overload_name() const; + + int64_t seqNr() const { + return sequence_nr_; + } + + c10::ArrayRef inputs() const { +#ifndef NDEBUG + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + inputs_valid_, "Called inputs() outside RecordFunction start callback"); +#endif + return inputs_; + } + + std::unordered_map kwinputs() const { +#ifndef NDEBUG + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + inputs_valid_, + "Called kwinputs() outside RecordFunction start callback"); +#endif + return kwinputs_; + } + + const std::vector& outputs() const { + return outputs_; + } + + void setOutputs(std::vector&& outputs) { + outputs_ = std::move(outputs); + } + + void setOutputs(c10::ArrayRef outputs) { + outputs_ = outputs.vec(); + } + + size_t num_inputs() const; + size_t num_outputs() const; + + // Retrieves the thread_id that this RecordFunction ran start callbacks with. + // Useful for writing thread safe end callbacks that may be potentially + // executed in a different thread (async ops) + uint64_t threadId() const { + return step_callbacks_.thread_id_; + } + + // For backward functions - thread id of the corresponding forward function, + // or zero otherwise; + // used alongside with sequence number to correlate backward functions with + // the forward ones + uint64_t forwardThreadId() const { + return fwd_thread_id_; + } + + void setForwardThreadId(uint64_t thread_id) { + fwd_thread_id_ = thread_id; + } + + RecordScope scope() const { + return step_callbacks_.scope_; + } + + // Returns logical thread_id for the current thread + static uint64_t currentThreadId(); + + // Internal functions, do not use directly; + // used in python's context manager + + // before functions initialize RecordFunction members and call + // start callbacks + void before(FunctionDescriptor schema, int64_t sequence_nr = -1); + + // Sets node ID for distributed profiling + static void setDefaultNodeId(int64_t defaultNodeId); + // Gets node ID for distributed profiling + static int64_t getDefaultNodeId(); + + // Calls end callbacks. After end(), accessors will no longer provide useful + // results. + void end(); + + // Internal-only, used only force async event for distributed events + // profiling. + void _setAsync(); + + // Returns whether this RecordFunction corresponds to an async event or not. + bool isAsync() const; + + // Returns whether this RecordFunction corresponds to NCCL metadata collection + // or not. + bool isNcclMeta() const { + return is_nccl_meta_; + } + + // Internal-only, used to denote out variant used for Static Runtime execution + void _setStaticRuntimeOutVariant(); + bool isStaticRuntimeOutVariant() const; + + RecordFunctionHandle handle() const { + return handle_; + } + + std::optional operator_name() const; + + // This method returns a copy of the FunctionSchema and can be expensive. + std::optional operator_schema() const; + + void setHandle(RecordFunctionHandle handle) { + handle_ = handle; + } + + // Whether this RecordFunction runs any callbacks. + bool isActive() const { + return !step_callbacks_.empty(); + } + + bool needsInputs() const { + return step_callbacks_.needs_inputs_; + } + + bool needsOutputs() const { + return step_callbacks_.needs_outputs_; + } + + int64_t debugHandle() const { + return debug_handle_; + } + + void setDebugHandle(int64_t debug_handle) { + debug_handle_ = debug_handle; + } + + void invalidateInputs() { +#ifndef NDEBUG + inputs_valid_ = false; +#endif + } + + private: + void runStartCallbacks(); + + StepCallbacks step_callbacks_; + + // In cases when RecordFunction might be active but we chose not to + // use the observers (e.g. operator is not observed), this boolean + // flag is used to check whether the start callbacks were called + bool called_start_callbacks_ = false; + +#ifndef NDEBUG + bool inputs_valid_ = false; +#endif + + // Stores various ObserverContext objects with event metadata for callbacks. + ObserverContextList ctx_; + + std::variant fn_; + + int64_t sequence_nr_ = -1; + c10::ArrayRef inputs_; + std::unordered_map kwinputs_; + std::vector outputs_; + + // For backward functions - thread id of the forward function + uint64_t fwd_thread_id_ = 0; + + // Unique id for this RecordFunction, used in callbacks to track start + // and end of ranges + RecordFunctionHandle handle_{0}; + + // Whether this record_function corresponds to an async event or not. Async + // events can complete in different threads or follow a future-like pattern + // of use. + bool is_async_{false}; + + // Debug handles are used for lazy annotation of module hierarchy + // and callstack. + // This is specifically is useful for mobile runtime, where generated + // debug handles can be lazily symbolicated using debug information + int64_t debug_handle_{-1}; + + // Whether this RecordFunction is used for an out variant run with + // Static Runtime + bool is_static_runtime_out_variant_{false}; + + // Whether this RecordFunction is used for NCCL metadata collection + bool is_nccl_meta_{false}; +}; + +TORCH_API StepCallbacks getStepCallbacks(RecordScope scope); + +TORCH_API std::optional getStepCallbacksUnlessEmpty( + RecordScope scope); + +namespace detail { +template +void record_function_with_scope( + RecordFunction& guard, + RecordFunction::FunctionDescriptor fn, + const Inputs& inputs, + Args&&... args) { + if (guard.needsInputs()) { + guard.before( + fn, + c10::ArrayRef(inputs.data(), inputs.size()), + std::forward(args)...); + } else { + guard.before(fn, std::forward(args)...); + } +} + +template +void record_function_with_scope_and_debug_handle( + RecordFunction& guard, + RecordFunction::FunctionDescriptor fn, + int64_t debug_handle, + const Inputs& inputs, + Args&&... args) { + guard.setDebugHandle(debug_handle); + if (guard.needsInputs()) { + guard.before( + fn, + c10::ArrayRef(inputs.data(), inputs.size()), + std::forward(args)...); + } else { + guard.before(fn, std::forward(args)...); + } +} + +template +void record_function_with_scope( + RecordFunction& guard, + RecordFunction::FunctionDescriptor fn, + c10::ArrayRef inputs, + Args&&... args) { + return record_function_with_scope, Args...>( + guard, fn, inputs, std::forward(args)...); +} + +template +void record_function_with_scope_and_debug_handle( + RecordFunction& guard, + RecordFunction::FunctionDescriptor fn, + int64_t debug_handle, + c10::ArrayRef inputs, + Args&&... args) { + return record_function_with_scope_and_debug_handle< + c10::ArrayRef, + Args...>(guard, fn, debug_handle, inputs, std::forward(args)...); +} + +} // namespace detail + +// optional argument - function's seq_no +#define RECORD_FUNCTION_WITH_SCOPE(scope, fn, inputs, ...) \ + at::RecordFunction guard(scope); \ + if (guard.isActive()) { \ + ::at::detail::record_function_with_scope( \ + guard, fn, inputs, ##__VA_ARGS__); \ + } + +#define RECORD_FUNCTION_WITH_SCOPE_INPUTS_OUTPUTS( \ + scope, fn, inputs, outputs, ...) \ + at::RecordFunction guard(scope); \ + if (guard.isActive()) { \ + if (guard.needsInputs()) { \ + guard.before(fn, inputs, ##__VA_ARGS__); \ + } else { \ + guard.before(fn, ##__VA_ARGS__); \ + } \ + if (guard.needsOutputs()) { \ + guard.setOutputs(outputs); \ + } \ + } + +#define RECORD_FUNCTION(fn, inputs, ...) \ + RECORD_FUNCTION_WITH_SCOPE( \ + at::RecordScope::FUNCTION, fn, inputs, ##__VA_ARGS__) + +#define RECORD_TORCHSCRIPT_FUNCTION(mn, inputs) \ + RECORD_FUNCTION_WITH_SCOPE(at::RecordScope::TORCHSCRIPT_FUNCTION, mn, inputs) + +#define RECORD_FUNCTION_WITH_INPUTS_OUTPUTS(fn, inputs, outputs, ...) \ + RECORD_FUNCTION_WITH_SCOPE_INPUTS_OUTPUTS( \ + at::RecordScope::FUNCTION, fn, inputs, outputs, ##__VA_ARGS__) + +// Custom user scopes in C++; similar to Python's 'with record_function("..."):' +#define RECORD_USER_SCOPE(fn) \ + RECORD_FUNCTION_WITH_SCOPE( \ + at::RecordScope::USER_SCOPE, fn, c10::ArrayRef{}) + +// RECORD_USER_SCOPE with inputs +#define RECORD_USER_SCOPE_WITH_INPUTS(fn, inputs) \ + RECORD_FUNCTION_WITH_SCOPE(at::RecordScope::USER_SCOPE, fn, inputs) + +#define RECORD_USER_SCOPE_WITH_KWARGS_ONLY(fn, kwargs) \ + RECORD_FUNCTION_WITH_SCOPE( \ + at::RecordScope::USER_SCOPE, \ + fn, \ + c10::ArrayRef{}, \ + kwargs) + +// Helper macro to pass in debug handle that is used to +// post process events +#define RECORD_WITH_SCOPE_DEBUG_HANDLE_AND_INPUTS( \ + scope, fn, debug_handle, inputs, ...) \ + at::RecordFunction guard(scope); \ + if (guard.isActive()) { \ + ::at::detail::record_function_with_scope_and_debug_handle( \ + guard, fn, debug_handle, inputs, ##__VA_ARGS__); \ + } + +// Helper macros to record LITE INTERPRETER scope events with debug handles +#define RECORD_EDGE_SCOPE_WITH_DEBUG_HANDLE_AND_INPUTS( \ + fn, debug_handle, inputs) \ + RECORD_WITH_SCOPE_DEBUG_HANDLE_AND_INPUTS( \ + at::RecordScope::LITE_INTERPRETER, fn, debug_handle, inputs) + +// Bookend to the RECORD_FUNCTION macros. Use this after the kernel +// launch to let the profiler bind the outputs to the op that produced +// them. Note that guard is declared by RECORD_FUNCTION so this macro +// needs to be called from the same scope as RECORD_FUNCTION +#define RECORD_OUTPUTS(outputs) \ + if (guard.needsOutputs()) { \ + guard.setOutputs( \ + std::vector(outputs.begin(), outputs.end())); \ + } + +/** + * addThreadLocalCallback adds a thread local callback to run with + * RecordFunction, returns handle to use with removeThreadLocalCallback + */ +TORCH_API CallbackHandle addThreadLocalCallback(RecordFunctionCallback cb); + +/** + * hasThreadLocalCallbacks returns whether there're callbacks registered + * with addThreadLocalCallback + */ +TORCH_API bool hasThreadLocalCallbacks(); + +/** + * clearThreadLocalCallbacks removes all thread local callbacks + */ +TORCH_API void clearThreadLocalCallbacks(); + +/** + * addGlobalCallback adds a global callback to run with RecordFunction: + * + * only during the program initialization + */ +TORCH_API CallbackHandle addGlobalCallback(RecordFunctionCallback cb); + +/** + * removeCallback removes a callback given the handle returned by + * addThreadLocalCallback or addGlobalCallback; + * + * no other code can run simultaneously + */ +TORCH_API void removeCallback(CallbackHandle handle); + +/** + * Prevent the given callback from executing. If handle is invalid, + * does nothing. + */ +TORCH_API void disableCallback(CallbackHandle handle); + +/** + * Allow the given callback, previously disabled with disableCallback, to + * execute again. If handle is invalid, does nothing. + */ +TORCH_API void reenableCallback(CallbackHandle handle); + +/** + * hasGlobalCallbacks returns whether there're global callbacks + * registered with pushGlobalCallback + */ +TORCH_API bool hasGlobalCallbacks(); + +/** + * clearGlobalCallbacks removes all global callbacks + */ +TORCH_API void clearGlobalCallbacks(); + +// for both thread local and global callbacks +TORCH_API bool hasCallbacks(); +TORCH_API void clearCallbacks(); + +/** + * enableRecordFunction enables RecordFunction thread locally + */ +TORCH_API void enableRecordFunction(bool enable = true); + +/** + * isRecordFunctionEnabled returns whether RecordFunction + * is enabled thread locally + */ +TORCH_API bool isRecordFunctionEnabled(); + +class TORCH_API RecordFunctionGuard { + public: + explicit RecordFunctionGuard(bool is_enabled = true) + : prev_value_(isRecordFunctionEnabled()) { + enableRecordFunction(is_enabled); + } + + RecordFunctionGuard(RecordFunctionGuard&& other) = delete; + RecordFunctionGuard(const RecordFunctionGuard&) = delete; + RecordFunctionGuard& operator=(const RecordFunctionGuard&) = delete; + RecordFunctionGuard& operator=(RecordFunctionGuard&&) = delete; + virtual ~RecordFunctionGuard() { + enableRecordFunction(prev_value_); + } + + private: + bool prev_value_ = false; +}; + +class TORCH_API DisableRecordFunctionGuard : public RecordFunctionGuard { + public: + DisableRecordFunctionGuard() : RecordFunctionGuard(false) {} + ~DisableRecordFunctionGuard() override = default; +}; + +struct TORCH_API RecordFunctionTLS { + // Thread local vector of callbacks, holds pairs (callbacks, unique_id); + // must be sorted in increasing handles order + RecordFunctionCallbacks sorted_tls_callbacks_; + + bool tls_record_function_enabled_ = true; +}; + +TORCH_API const RecordFunctionTLS& get_record_function_tls_(); + +TORCH_API void set_record_function_tls_(const RecordFunctionTLS& tls); + +TORCH_API void set_record_function_seed_for_testing(uint32_t seed); + +} // namespace at + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/Allocator.h b/backends/pytorch/include/c10/core/Allocator.h new file mode 100644 index 000000000..b66f075ec --- /dev/null +++ b/backends/pytorch/include/c10/core/Allocator.h @@ -0,0 +1,455 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +using CaptureId_t = unsigned long long; +// first is set if the instance is created by CUDAGraph::capture_begin. +// second is set if the instance is created by at::cuda::graph_pool_handle. +using MempoolId_t = std::pair; + +struct MempoolIdHash { + std::size_t operator()(const MempoolId_t& mempool_id) const noexcept { + return mempool_id.first != 0 ? mempool_id.first : mempool_id.second; + } +}; + +// A DataPtr is a unique pointer (with an attached deleter and some +// context for the deleter) to some memory, which also records what +// device is for its data. +// +// nullptr DataPtrs can still have a nontrivial device; this allows +// us to treat zero-size allocations uniformly with non-zero allocations. +// +class C10_API DataPtr { + private: + c10::detail::UniqueVoidPtr ptr_; + Device device_; + + public: + // Choice of CPU here is arbitrary; if there's an "undefined" device + // we could use that too + DataPtr() : device_(DeviceType::CPU) {} + DataPtr(void* data, Device device) : ptr_(data), device_(device) {} + DataPtr(void* data, void* ctx, DeleterFnPtr ctx_deleter, Device device) + : ptr_(data, ctx, ctx_deleter), device_(device) {} + void* operator->() const { + return ptr_.get(); + } + C10_ALWAYS_INLINE bool /* success */ unsafe_reset_data_and_ctx( + void* new_data_and_ctx) { + return ptr_.unsafe_reset_data_and_ctx(new_data_and_ctx); + } + void clear() { + ptr_.clear(); + } + void* get() const { + return ptr_.get(); + } + void* mutable_get() { + return ptr_.get(); + } + void* get_context() const { + return ptr_.get_context(); + } + void* release_context() { + return ptr_.release_context(); + } + std::unique_ptr&& move_context() { + return ptr_.move_context(); + } + operator bool() const { + return static_cast(ptr_); + } + template + T* cast_context(DeleterFnPtr expected_deleter) const { + return ptr_.cast_context(expected_deleter); + } + DeleterFnPtr get_deleter() const { + return ptr_.get_deleter(); + } + /** + * Compare the deleter in a DataPtr to expected_deleter. + * If it matches, replace the deleter with new_deleter + * and return true; otherwise, does nothing and returns + * false. + * + * In general, it is not safe to unconditionally set the + * deleter on a DataPtr, because you don't know what + * the deleter is, and thus will have a hard time properly + * disposing of the deleter without storing the original + * deleter (this is difficult to do, because DeleterFnPtr + * is not a closure, and because the context on DataPtr is + * only a single word, you generally don't have enough + * space to store both the original deleter and its context). + * However, in some cases, you know /exactly/ what the deleter + * is, and you have a new deleter that manually wraps + * the old one. In this case, you can safely swap the deleter + * after asserting that the deleters line up. + * + * What are the requirements on new_deleter? It must still + * properly dispose of the void* pointer passed in as its argument, + * where void* is whatever the context of the original deleter + * is. So in general, you expect the new deleter to look something + * like this: + * + * [](void* ptr) { + * some_new_stuff(ptr); + * get_orig_allocator()->raw_deleter(ptr); + * } + * + * Note that it won't work to close over the original + * allocator; you don't have enough space to do that! Also, + * it's unsafe to assume that the passed in pointer in + * question is the memory pointer in question; it might not + * be; be sure to read the source code of the Allocator + * in question to confirm this. + */ + [[nodiscard]] bool compare_exchange_deleter( + DeleterFnPtr expected_deleter, + DeleterFnPtr new_deleter) { + return ptr_.compare_exchange_deleter(expected_deleter, new_deleter); + } + Device device() const { + return device_; + } + // Unsafely mutates the device on a DataPtr. Under normal use, + // you should never actually need to call this function. + // We need this for the implementation of the hack detailed + // in Note [Masquerading as CUDA] + void unsafe_set_device(Device device) { + device_ = device; + } +}; + +// NB: Device is NOT tested for here; a CUDA nullptr is as much a nullptr as a +// CPU nullptr + +inline bool operator==(const DataPtr& dp, std::nullptr_t) noexcept { + return !dp; +} +inline bool operator==(std::nullptr_t, const DataPtr& dp) noexcept { + return !dp; +} +inline bool operator!=(const DataPtr& dp, std::nullptr_t) noexcept { + return dp; +} +inline bool operator!=(std::nullptr_t, const DataPtr& dp) noexcept { + return dp; +} + +// Note [raw_allocate/raw_deallocate and Thrust] +// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +// Thrust's support for custom allocators requires us to write something +// like this: +// +// class ThrustAllocator { +// char* allocate(size_t); +// void deallocate(char*, size_t); +// }; +// +// This is not good for our unique_ptr based allocator interface, as +// there is no way to get to the context when we free. +// +// However, in some cases the context is exactly the same as +// the data pointer. In this case, we can support the "raw" +// allocate and deallocate interface. This is what +// raw_deleter signifies. By default, it returns a nullptr, which means that +// the raw interface is not implemented. Be sure to implement it whenever +// possible, or the raw interface will incorrectly reported as unsupported, +// when it is actually possible. + +// NOLINTNEXTLINE(cppcoreguidelines-special-member-functions) +struct C10_API Allocator { + virtual ~Allocator() = default; + + virtual DataPtr allocate(size_t n) = 0; + + // Clones an allocation that came from this allocator. + // + // To perform the copy, this function calls `copy_data`, which + // must be implemented by derived classes. + // + // Note that this explicitly ignores any context that may have been + // attached to the input data. + // + // Requires: input data was allocated by the same allocator. + DataPtr clone(const void* data, std::size_t n); + + // Checks if DataPtr has a simple context, not wrapped with any out of the + // ordinary contexts. + virtual bool is_simple_data_ptr(const DataPtr& data_ptr) const; + + // If this returns a non nullptr, it means that allocate() + // is guaranteed to return a unique_ptr with this deleter attached; + // it means the rawAllocate and rawDeallocate APIs are safe to use. + // This function MUST always return the same BoundDeleter. + virtual DeleterFnPtr raw_deleter() const { + return nullptr; + } + void* raw_allocate(size_t n) { + auto dptr = allocate(n); + AT_ASSERT(dptr.get() == dptr.get_context()); + return dptr.release_context(); + } + void raw_deallocate(void* ptr) { + auto d = raw_deleter(); + AT_ASSERT(d); + d(ptr); + } + + // Copies data from one allocation to another. + // Pure virtual, so derived classes must define behavior. + // Derived class implementation can simply call `default_copy_data` + // to use `std::memcpy`. + // + // Requires: src and dest were allocated by this allocator + // Requires: src and dest both have length >= count + virtual void copy_data(void* dest, const void* src, std::size_t count) + const = 0; + + protected: + // Uses `std::memcpy` to copy data. + // Child classes can use this as `copy_data` when an alternative copy + // API is not needed. + void default_copy_data(void* dest, const void* src, std::size_t count) const; +}; + +// This context is used to generate DataPtr which have arbitrary +// std::function deleters associated with them. In some user facing +// functions, we give a (user-friendly) interface for constructing +// tensors from external data which take an arbitrary std::function +// deleter. Grep for InefficientStdFunctionContext to find these +// occurrences. +// +// This context is inefficient because we have to do a dynamic +// allocation InefficientStdFunctionContext, on top of the dynamic +// allocation which is implied by std::function itself. +struct C10_API InefficientStdFunctionContext { + void* ptr_{nullptr}; + std::function deleter_; + InefficientStdFunctionContext(void* ptr, std::function deleter) + : ptr_(ptr), deleter_(std::move(deleter)) {} + InefficientStdFunctionContext(const InefficientStdFunctionContext&) = delete; + InefficientStdFunctionContext(InefficientStdFunctionContext&& rhs) noexcept + : ptr_(std::exchange(rhs.ptr_, nullptr)), + deleter_(std::move(rhs.deleter_)) {} + InefficientStdFunctionContext& operator=( + const InefficientStdFunctionContext&) = delete; + // NOLINTNEXTLINE(*-noexcept-move-*) + InefficientStdFunctionContext& operator=( + InefficientStdFunctionContext&& rhs) { + this->~InefficientStdFunctionContext(); + ptr_ = std::exchange(rhs.ptr_, nullptr); + deleter_ = std::move(rhs.deleter_); + return *this; + } + ~InefficientStdFunctionContext() { + if (deleter_) { + deleter_(ptr_); + } + } + static DataPtr makeDataPtr( + void* ptr, + std::function deleter, + Device device); +}; + +/** Set the allocator for DeviceType `t`. The passed in allocator pointer is + * expected to have static lifetime; this function does NOT take ownership + * of the raw pointer. (The reason for this is to prevent existing pointers + * to an allocator of a particular device from being invalidated when + * SetAllocator is called.) + * + * Also note that this is not thread-safe, and we assume this function will + * only be called during initialization. + * + * The 'priority' flag is introduced when we want to overwrite the default + * allocator, since the allocators are set statically. The default priority + * is 0, which means the lowest. Only higher or equal priority can overwrite + * existing ones. + */ +C10_API void SetAllocator(DeviceType t, Allocator* alloc, uint8_t priority = 0); +C10_API Allocator* GetAllocator(const DeviceType& t); + +template +struct AllocatorRegisterer { + explicit AllocatorRegisterer(Allocator* alloc) { + SetAllocator(t, alloc); + } +}; + +#define REGISTER_ALLOCATOR(t, f) \ + namespace { \ + static c10::AllocatorRegisterer g_allocator_d(f); \ + } + +// An interface for reporting thread local memory usage +// per device +struct C10_API MemoryReportingInfoBase : public c10::DebugInfoBase { + /** + * alloc_size corresponds to the size of the ptr. + * + * total_allocated corresponds to total allocated memory. + * + * total_reserved corresponds to total size of memory pool, both used and + * unused, if applicable. + */ + virtual void reportMemoryUsage( + void* ptr, + int64_t alloc_size, + size_t total_allocated, + size_t total_reserved, + Device device) = 0; + + virtual void reportOutOfMemory( + int64_t alloc_size, + size_t total_allocated, + size_t total_reserved, + Device device); + + virtual bool memoryProfilingEnabled() const = 0; +}; + +C10_API bool memoryProfilingEnabled(); +C10_API void reportMemoryUsageToProfiler( + void* ptr, + int64_t alloc_size, + size_t total_allocated, + size_t total_reserved, + Device device); + +C10_API void reportOutOfMemoryToProfiler( + int64_t alloc_size, + size_t total_allocated, + size_t total_reserved, + Device device); + +// used to hold traceback information in allocators +// NOLINTNEXTLINE(cppcoreguidelines-special-member-functions) +struct GatheredContext { + virtual ~GatheredContext() = default; +}; + +namespace CachingAllocator { +struct Stat { + void increase(size_t amount) { + current += static_cast(amount); + peak = std::max(current, peak); + allocated += static_cast(amount); + } + + void decrease(size_t amount) { + current -= static_cast(amount); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + current >= 0, + "Negative tracked stat in device allocator (likely logic error)."); + freed += static_cast(amount); + } + + void reset_accumulated() { + allocated = 0; + freed = 0; + } + + void reset_peak() { + peak = current; + } + + int64_t current = 0; + int64_t peak = 0; + int64_t allocated = 0; + int64_t freed = 0; +}; + +enum struct StatType : uint64_t { + AGGREGATE = 0, + SMALL_POOL = 1, + LARGE_POOL = 2, + NUM_TYPES = 3 // remember to update this whenever a new stat type is added +}; + +using StatArray = std::array(StatType::NUM_TYPES)>; +using StatTypes = std::array(StatType::NUM_TYPES)>; + +template +void for_each_selected_stat_type(const StatTypes& stat_types, Func f) { + for (const auto stat_type : c10::irange(stat_types.size())) { + if (stat_types[stat_type]) { + f(stat_type); + } + } +} + +// Structure for keeping timing information +struct DurationStat { + void increase(int64_t amount) { + total += amount; + count += 1; + max = std::max(amount, max); + if (min == 0) { + min = amount; + } else { + min = std::min(amount, min); + } + } + + void reset_accumulated() { + total = 0; + count = 0; + } + + void reset_peak() { + min = 0; + max = 0; + } + + int64_t total = 0; + int64_t max = 0; + int64_t min = 0; + int64_t count = 0; +}; + +// Size pretty-printer +inline std::string format_size(uint64_t size) { + std::ostringstream os; + os.precision(2); + os << std::fixed; + if (size <= 1024) { + os << size << " bytes"; + } else if (size <= 1048576) { + os << (static_cast(size) / 1024.0); + os << " KiB"; + } else if (size <= 1073741824ULL) { + os << static_cast(size) / 1048576.0; + os << " MiB"; + } else { + os << static_cast(size) / 1073741824.0; + os << " GiB"; + } + return os.str(); +} + +} // namespace CachingAllocator +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/AutogradState.h b/backends/pytorch/include/c10/core/AutogradState.h new file mode 100644 index 000000000..9d596b01d --- /dev/null +++ b/backends/pytorch/include/c10/core/AutogradState.h @@ -0,0 +1,90 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +namespace c10 { + +// Structure used to pack all the thread local boolean +// flags used by autograd +struct C10_API AutogradState { + static AutogradState& get_tls_state(); + static void set_tls_state(AutogradState state); + + AutogradState( + bool grad_mode, + bool inference_mode, + bool fw_grad_mode, + bool multithreading_enabled) + : graph_exec_group_(std::nullopt), + grad_mode_(grad_mode), + inference_mode_(inference_mode), + fw_grad_mode_(fw_grad_mode), + multithreading_enabled_(multithreading_enabled), + view_replay_enabled_(false) {} + + void set_grad_mode(bool enabled) { + grad_mode_ = enabled; + } + + void set_fw_grad_mode(bool enabled) { + fw_grad_mode_ = enabled; + } + + void set_inference_mode(bool enabled) { + inference_mode_ = enabled; + } + + void set_multithreading_enabled(bool multithreading_enabled) { + multithreading_enabled_ = multithreading_enabled; + } + + void set_view_replay_enabled(bool view_replay_enabled) { + view_replay_enabled_ = view_replay_enabled; + } + + void set_graph_exec_group(std::optional group) { + graph_exec_group_ = std::move(group); + } + + bool get_grad_mode() const { + return grad_mode_; + } + + bool get_fw_grad_mode() const { + return fw_grad_mode_; + } + + bool get_inference_mode() const { + return inference_mode_; + } + + bool get_multithreading_enabled() const { + return multithreading_enabled_; + } + + bool get_view_replay_enabled() const { + return view_replay_enabled_; + } + + const std::optional& get_graph_exec_group() const { + return graph_exec_group_; + } + + private: + std::optional graph_exec_group_; + bool grad_mode_ : 1; + bool inference_mode_ : 1; + bool fw_grad_mode_ : 1; + bool multithreading_enabled_ : 1; + // NOLINTNEXTLINE(cppcoreguidelines-use-default-member-init) + bool view_replay_enabled_ : 1; +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/Backend.h b/backends/pytorch/include/c10/core/Backend.h new file mode 100644 index 000000000..d26c0089a --- /dev/null +++ b/backends/pytorch/include/c10/core/Backend.h @@ -0,0 +1,414 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include + +#include + +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum") + +namespace c10 { + +/** + * This legacy enum class defines the set of backends supported by old school, + * code generated Type-based ATen. A "backend" in this sense roughly + * corresponds to the cartesian product of (device type, layout), but restricted + * only to combinations which we actually have kernels for. Backend does NOT + * include dtype. + * + * The reason we are sunsetting this enum class is because it doesn't allow for + * open registration; e.g., if you want to add SparseXLA, you'd have to + * edit this enum; you wouldn't be able to do it out of tree. DispatchKey is + * the replacement for Backend which supports open registration. + * + * NB: The concept of 'Backend' here disagrees with the notion of backend + * exposed to users in torch.backends. Backend here is something like "CPU" + * or "SparseCUDA"; backend in torch.backends is something like "MKL" or + * "CUDNN". + */ +enum class Backend { + CPU, + CUDA, + HIP, + VE, + FPGA, + IPU, + XPU, + SparseCPU, + SparseCUDA, + SparseCsrCPU, + SparseCsrCUDA, + SparseCsrMPS, + SparseMPS, + SparseHIP, + SparseVE, + SparseXPU, + SparsePrivateUse1, + SparseCsrHIP, + SparseCsrVE, + SparseCsrXPU, + SparseCsrPrivateUse1, + MAIA, + XLA, + Vulkan, + Metal, + Meta, + QuantizedCPU, + QuantizedCUDA, + QuantizedXPU, + QuantizedPrivateUse1, + Undefined, + MkldnnCPU, + MPS, + HPU, + Lazy, + MTIA, + PrivateUse1, + NumOptions +}; + +inline Backend dispatchKeyToBackend(DispatchKey t) { + if (t == DispatchKey::CPU || t == DispatchKey::AutogradCPU) { + return Backend::CPU; + } else if (t == DispatchKey::CUDA || t == DispatchKey::AutogradCUDA) { + return Backend::CUDA; + } else if (t == DispatchKey::HIP) { + return Backend::HIP; + } else if (t == DispatchKey::VE) { + return Backend::VE; + } else if (t == DispatchKey::FPGA) { + return Backend::FPGA; + } else if (t == DispatchKey::MAIA || t == DispatchKey::AutogradMAIA) { + return Backend::MAIA; + } else if (t == DispatchKey::XLA || t == DispatchKey::AutogradXLA) { + return Backend::XLA; + } else if (t == DispatchKey::Lazy || t == DispatchKey::AutogradLazy) { + return Backend::Lazy; + } else if (t == DispatchKey::MPS || t == DispatchKey::AutogradMPS) { + return Backend::MPS; + } else if (t == DispatchKey::Vulkan) { + return Backend::Vulkan; + } else if (t == DispatchKey::Metal) { + return Backend::Metal; + } else if (t == DispatchKey::Meta) { + return Backend::Meta; + } else if (t == DispatchKey::SparseCPU) { + return Backend::SparseCPU; + } else if (t == DispatchKey::SparseCUDA) { + return Backend::SparseCUDA; + } else if (t == DispatchKey::SparseMPS) { + return Backend::SparseMPS; + } else if (t == DispatchKey::SparseCsrMPS) { + return Backend::SparseCsrMPS; + } else if (t == DispatchKey::SparseHIP) { + return Backend::SparseHIP; + } else if (t == DispatchKey::SparseVE) { + return Backend::SparseVE; + } else if (t == DispatchKey::SparsePrivateUse1) { + return Backend::SparsePrivateUse1; + } else if (t == DispatchKey::SparseCsrCPU) { + return Backend::SparseCsrCPU; + } else if (t == DispatchKey::SparseCsrCUDA) { + return Backend::SparseCsrCUDA; + } else if (t == DispatchKey::SparseCsrHIP) { + return Backend::SparseCsrHIP; + } else if (t == DispatchKey::SparseCsrVE) { + return Backend::SparseCsrVE; + } else if (t == DispatchKey::SparseCsrPrivateUse1) { + return Backend::SparseCsrPrivateUse1; + } else if (t == DispatchKey::MkldnnCPU) { + return Backend::MkldnnCPU; + } else if (t == DispatchKey::QuantizedCPU) { + return Backend::QuantizedCPU; + } else if (t == DispatchKey::QuantizedCUDA) { + return Backend::QuantizedCUDA; + } else if (t == DispatchKey::IPU || t == DispatchKey::AutogradIPU) { + return Backend::IPU; + } else if (t == DispatchKey::XPU || t == DispatchKey::AutogradXPU) { + return Backend::XPU; + } else if (t == DispatchKey::SparseXPU) { + return Backend::SparseXPU; + } else if (t == DispatchKey::SparseCsrXPU) { + return Backend::SparseCsrXPU; + } else if (t == DispatchKey::QuantizedXPU) { + return Backend::QuantizedXPU; + } else if (t == DispatchKey::QuantizedPrivateUse1) { + return Backend::QuantizedPrivateUse1; + } else if (t == DispatchKey::HPU || t == DispatchKey::AutogradHPU) { + return Backend::HPU; + } else if (t == DispatchKey::MTIA || t == DispatchKey::AutogradMTIA) { + return Backend::MTIA; + } else if ( + t == DispatchKey::PrivateUse1 || t == DispatchKey::AutogradPrivateUse1) { + return Backend::PrivateUse1; + } else if (t == DispatchKey::Undefined) { + return Backend::Undefined; + } else { + TORCH_CHECK(false, "Unrecognized tensor type ID: ", t); + } +} + +inline DispatchKey backendToDispatchKey(Backend b) { + switch (b) { + case Backend::CPU: + return DispatchKey::CPU; + case Backend::CUDA: + return DispatchKey::CUDA; + case Backend::HIP: + return DispatchKey::HIP; + case Backend::VE: + return DispatchKey::VE; + case Backend::FPGA: + return DispatchKey::FPGA; + case Backend::MAIA: + return DispatchKey::MAIA; + case Backend::XLA: + return DispatchKey::XLA; + case Backend::Lazy: + return DispatchKey::Lazy; + case Backend::IPU: + return DispatchKey::IPU; + case Backend::XPU: + return DispatchKey::XPU; + case Backend::SparseXPU: + return DispatchKey::SparseXPU; + case Backend::SparseCsrXPU: + return DispatchKey::SparseCsrXPU; + case Backend::SparseCPU: + return DispatchKey::SparseCPU; + case Backend::SparseCUDA: + return DispatchKey::SparseCUDA; + case Backend::SparseMPS: + return DispatchKey::SparseMPS; + case Backend::SparseCsrMPS: + return DispatchKey::SparseCsrMPS; + case Backend::SparseHIP: + return DispatchKey::SparseHIP; + case Backend::SparseVE: + return DispatchKey::SparseVE; + case Backend::SparsePrivateUse1: + return DispatchKey::SparsePrivateUse1; + case Backend::SparseCsrCPU: + return DispatchKey::SparseCsrCPU; + case Backend::SparseCsrCUDA: + return DispatchKey::SparseCsrCUDA; + case Backend::SparseCsrHIP: + return DispatchKey::SparseCsrHIP; + case Backend::SparseCsrVE: + return DispatchKey::SparseCsrVE; + case Backend::SparseCsrPrivateUse1: + return DispatchKey::SparseCsrPrivateUse1; + case Backend::MkldnnCPU: + return DispatchKey::MkldnnCPU; + case Backend::Vulkan: + return DispatchKey::Vulkan; + case Backend::Metal: + return DispatchKey::Metal; + case Backend::Meta: + return DispatchKey::Meta; + case Backend::QuantizedCPU: + return DispatchKey::QuantizedCPU; + case Backend::QuantizedCUDA: + return DispatchKey::QuantizedCUDA; + case Backend::QuantizedPrivateUse1: + return DispatchKey::QuantizedPrivateUse1; + case Backend::Undefined: + return DispatchKey::Undefined; + case Backend::MPS: + return DispatchKey::MPS; + case Backend::HPU: + return DispatchKey::HPU; + case Backend::MTIA: + return DispatchKey::MTIA; + case Backend::PrivateUse1: + return DispatchKey::PrivateUse1; + default: + TORCH_CHECK(false, "Unknown backend"); + } +} + +inline DeviceType backendToDeviceType(Backend b) { + switch (b) { + case Backend::CPU: + case Backend::MkldnnCPU: + case Backend::SparseCPU: + case Backend::SparseCsrCPU: + case Backend::QuantizedCPU: + return DeviceType::CPU; + case Backend::CUDA: + case Backend::SparseCUDA: + case Backend::QuantizedCUDA: + case Backend::SparseCsrCUDA: + return DeviceType::CUDA; + case Backend::HIP: + return DeviceType::HIP; + case Backend::VE: + return DeviceType::VE; + case Backend::FPGA: + return DeviceType::FPGA; + case Backend::MAIA: + return DeviceType::MAIA; + case Backend::XLA: + return DeviceType::XLA; + case Backend::Lazy: + return DeviceType::Lazy; + case Backend::SparseHIP: + return DeviceType::HIP; + case Backend::SparseVE: + return DeviceType::VE; + case Backend::SparseCsrHIP: + return DeviceType::HIP; + case Backend::SparseCsrVE: + return DeviceType::VE; + case Backend::IPU: + return DeviceType::IPU; + case Backend::XPU: + case Backend::SparseXPU: + case Backend::SparseCsrXPU: + case Backend::QuantizedXPU: + return DeviceType::XPU; + case Backend::Vulkan: + return DeviceType::Vulkan; + case Backend::Metal: + return DeviceType::Metal; + case Backend::Meta: + return DeviceType::Meta; + case Backend::MPS: + case Backend::SparseMPS: + case Backend::SparseCsrMPS: + return DeviceType::MPS; + case Backend::HPU: + return DeviceType::HPU; + case Backend::MTIA: + return DeviceType::MTIA; + case Backend::PrivateUse1: + case Backend::SparsePrivateUse1: + case Backend::SparseCsrPrivateUse1: + case Backend::QuantizedPrivateUse1: + return DeviceType::PrivateUse1; + case Backend::Undefined: + TORCH_CHECK(false, "Undefined backend is not a valid device type"); + default: + TORCH_CHECK(false, "Unknown backend"); + } +} + +inline const char* toString(Backend b) { + switch (b) { + case Backend::CPU: + return "CPU"; + case Backend::CUDA: + return "CUDA"; + case Backend::HIP: + return "HIP"; + case Backend::VE: + return "VE"; + case Backend::FPGA: + return "FPGA"; + case Backend::XPU: + return "XPU"; + case Backend::IPU: + return "IPU"; + case Backend::MAIA: + return "MAIA"; + case Backend::XLA: + return "XLA"; + case Backend::Lazy: + return "Lazy"; + case Backend::MPS: + return "MPS"; + case Backend::SparseCPU: + return "SparseCPU"; + case Backend::SparseCUDA: + return "SparseCUDA"; + case Backend::SparseMPS: + return "SparseMPS"; + case Backend::SparseCsrMPS: + return "SparseCsrMPS"; + case Backend::SparseHIP: + return "SparseHIP"; + case Backend::SparseVE: + return "SparseVE"; + case Backend::SparseXPU: + return "SparseXPU"; + case Backend::SparsePrivateUse1: + return "SparsePrivateUse1"; + case Backend::SparseCsrCPU: + return "SparseCsrCPU"; + case Backend::SparseCsrCUDA: + return "SparseCsrCUDA"; + case Backend::SparseCsrHIP: + return "SparseCsrHIP"; + case Backend::SparseCsrVE: + return "SparseCsrVE"; + case Backend::SparseCsrXPU: + return "SparseCsrXPU"; + case Backend::SparseCsrPrivateUse1: + return "SparseCsrPrivateUse1"; + case Backend::MkldnnCPU: + return "MkldnnCPU"; + case Backend::Vulkan: + return "Vulkan"; + case Backend::Metal: + return "Metal"; + case Backend::Meta: + return "Meta"; + case Backend::QuantizedCPU: + return "QuantizedCPU"; + case Backend::QuantizedCUDA: + return "QuantizedCUDA"; + case Backend::QuantizedXPU: + return "QuantizedXPU"; + case Backend::QuantizedPrivateUse1: + return "QuantizedPrivateUse1"; + case Backend::HPU: + return "HPU"; + case Backend::MTIA: + return "MTIA"; + case Backend::PrivateUse1: + return "PrivateUseOne"; + default: + return "UNKNOWN_BACKEND"; + } +} + +inline bool isSparse(Backend b) { + switch (b) { + case Backend::SparseXPU: + case Backend::SparseCPU: + case Backend::SparseCUDA: + case Backend::SparseMPS: + case Backend::SparseHIP: + case Backend::SparseVE: + case Backend::SparsePrivateUse1: + return true; + default: + return false; + } +} + +inline bool isSparseCsr(Backend b) { + switch (b) { + case Backend::SparseCsrXPU: + case Backend::SparseCsrCPU: + case Backend::SparseCsrCUDA: + case Backend::SparseCsrHIP: + case Backend::SparseCsrVE: + case Backend::SparseCsrPrivateUse1: + return true; + default: + return false; + } +} + +} // namespace c10 + +C10_DIAGNOSTIC_POP() + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/DefaultDtype.h b/backends/pytorch/include/c10/core/DefaultDtype.h new file mode 100644 index 000000000..240c173ca --- /dev/null +++ b/backends/pytorch/include/c10/core/DefaultDtype.h @@ -0,0 +1,20 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace caffe2 { +class TypeMeta; +} // namespace caffe2 + +namespace c10 { +C10_API void set_default_dtype(caffe2::TypeMeta dtype); +C10_API const caffe2::TypeMeta get_default_dtype(); +C10_API ScalarType get_default_dtype_as_scalartype(); +C10_API const caffe2::TypeMeta get_default_complex_dtype(); +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/Device.h b/backends/pytorch/include/c10/core/Device.h new file mode 100644 index 000000000..d3380f434 --- /dev/null +++ b/backends/pytorch/include/c10/core/Device.h @@ -0,0 +1,221 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +#include +#include +#include +#include +#include + +namespace c10 { + +/// An index representing a specific device; e.g., the 1 in GPU 1. +/// A DeviceIndex is not independently meaningful without knowing +/// the DeviceType it is associated; try to use Device rather than +/// DeviceIndex directly. +using DeviceIndex = int8_t; + +/// Represents a compute device on which a tensor is located. A device is +/// uniquely identified by a type, which specifies the type of machine it is +/// (e.g. CPU or CUDA GPU), and a device index or ordinal, which identifies the +/// specific compute device when there is more than one of a certain type. The +/// device index is optional, and in its defaulted state represents (abstractly) +/// "the current device". Further, there are two constraints on the value of the +/// device index, if one is explicitly stored: +/// 1. A negative index represents the current device, a non-negative index +/// represents a specific, concrete device, +/// 2. When the device type is CPU, the device index must be zero. +struct C10_API Device final { + using Type = DeviceType; + + /// Constructs a new `Device` from a `DeviceType` and an optional device + /// index. + /* implicit */ Device(DeviceType type, DeviceIndex index = -1) + : type_(type), index_(index) { + validate(); + } + + /// Constructs a `Device` from a string description, for convenience. + /// The string supplied must follow the following schema: + /// `(cpu|cuda)[:]` + /// where `cpu` or `cuda` specifies the device type, and + /// `:` optionally specifies a device index. + /* implicit */ Device(const std::string& device_string); + + /// Returns true if the type and index of this `Device` matches that of + /// `other`. + bool operator==(const Device& other) const noexcept { + return this->type_ == other.type_ && this->index_ == other.index_; + } + + /// Returns true if the type or index of this `Device` differs from that of + /// `other`. + bool operator!=(const Device& other) const noexcept { + return !(*this == other); + } + + /// Sets the device index. + void set_index(DeviceIndex index) { + index_ = index; + } + + /// Returns the type of device this is. + DeviceType type() const noexcept { + return type_; + } + + /// Returns the optional index. + DeviceIndex index() const noexcept { + return index_; + } + + /// Returns true if the device has a non-default index. + bool has_index() const noexcept { + return index_ != -1; + } + + /// Return true if the device is of CUDA type. + bool is_cuda() const noexcept { + return type_ == DeviceType::CUDA; + } + + /// Return true if the device is of PrivateUse1 type. + bool is_privateuseone() const noexcept { + return type_ == DeviceType::PrivateUse1; + } + + /// Return true if the device is of MPS type. + bool is_mps() const noexcept { + return type_ == DeviceType::MPS; + } + + /// Return true if the device is of HIP type. + bool is_hip() const noexcept { + return type_ == DeviceType::HIP; + } + + /// Return true if the device is of VE type. + bool is_ve() const noexcept { + return type_ == DeviceType::VE; + } + + /// Return true if the device is of XPU type. + bool is_xpu() const noexcept { + return type_ == DeviceType::XPU; + } + + /// Return true if the device is of IPU type. + bool is_ipu() const noexcept { + return type_ == DeviceType::IPU; + } + + /// Return true if the device is of XLA type. + bool is_xla() const noexcept { + return type_ == DeviceType::XLA; + } + + /// Return true if the device is of MTIA type. + bool is_mtia() const noexcept { + return type_ == DeviceType::MTIA; + } + + /// Return true if the device is of HPU type. + bool is_hpu() const noexcept { + return type_ == DeviceType::HPU; + } + + /// Return true if the device is of Lazy type. + bool is_lazy() const noexcept { + return type_ == DeviceType::Lazy; + } + + /// Return true if the device is of Vulkan type. + bool is_vulkan() const noexcept { + return type_ == DeviceType::Vulkan; + } + + /// Return true if the device is of Metal type. + bool is_metal() const noexcept { + return type_ == DeviceType::Metal; + } + + /// Return true if the device is of MAIA type. + bool is_maia() const noexcept { + return type_ == DeviceType::MAIA; + } + + /// Return true if the device is of META type. + bool is_meta() const noexcept { + return type_ == DeviceType::Meta; + } + + /// Return true if the device is of CPU type. + bool is_cpu() const noexcept { + return type_ == DeviceType::CPU; + } + + /// Return true if the device supports arbitrary strides. + bool supports_as_strided() const noexcept { + return type_ != DeviceType::IPU && type_ != DeviceType::XLA && + type_ != DeviceType::Lazy; + } + + /// Same string as returned from operator<<. + std::string str() const; + + private: + DeviceType type_; + DeviceIndex index_ = -1; + void validate() { + // Removing these checks in release builds noticeably improves + // performance in micro-benchmarks. + // This is safe to do, because backends that use the DeviceIndex + // have a later check when we actually try to switch to that device. + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + index_ >= -1, + "Device index must be -1 or non-negative, got ", + static_cast(index_)); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + !is_cpu() || index_ <= 0, + "CPU device index must be -1 or zero, got ", + static_cast(index_)); + } +}; + +C10_API std::ostream& operator<<(std::ostream& stream, const Device& device); + +} // namespace c10 + +namespace std { +template <> +struct hash { + size_t operator()(c10::Device d) const noexcept { + // Are you here because this static assert failed? Make sure you ensure + // that the bitmasking code below is updated accordingly! + static_assert(sizeof(c10::DeviceType) == 1, "DeviceType is not 8-bit"); + static_assert(sizeof(c10::DeviceIndex) == 1, "DeviceIndex is not 8-bit"); + // Note [Hazard when concatenating signed integers] + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // We must first convert to a same-sized unsigned type, before promoting to + // the result type, to prevent sign extension when any of the values is -1. + // If sign extension occurs, you'll clobber all of the values in the MSB + // half of the resulting integer. + // + // Technically, by C/C++ integer promotion rules, we only need one of the + // uint32_t casts to the result type, but we put in both for explicitness's + // sake. + uint32_t bits = static_cast(static_cast(d.type())) + << 16 | + static_cast(static_cast(d.index())); + return std::hash{}(bits); + } +}; +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/DeviceCapability.h b/backends/pytorch/include/c10/core/DeviceCapability.h new file mode 100644 index 000000000..854772812 --- /dev/null +++ b/backends/pytorch/include/c10/core/DeviceCapability.h @@ -0,0 +1,81 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +namespace c10 { + +constexpr size_t NUMBER_OF_DEVICE_CAPABILITIES = NumScalarTypes; + +// Generate bitfields for each scalar type +#define DEFINE_SCALAR_TYPE(_1, n) unsigned int has_##n : 1; + +// Generate enum indices for each scalar type +#define DEFINE_SCALAR_ENUM(_1, name) kIndex_##name, + +enum ScalarTypeIndex { + AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(DEFINE_SCALAR_ENUM) +}; + +/** + * @brief DeviceCapability represents the the common capabilities that all + * devices should support. + * + * This struct provides a compact way to represent the common capabilities that + * all devices should support. Includes the following capabilities: + * - Supported data types + * + * Purpose + * - Enable device-specific optimizations based on supported capabilities + * + * Contract + * + * Supported data types: + * - Each bitfield represents support for one device capability + * - Bit value 1 means the capability is supported, 0 means not supported + * - The struct is initialized with all capabilities enabled by default + * + * @note Adding New Capabilities + * + * 1. Define the new capability in the `DeviceCapability` struct + * 2. Update the support of the new capability in each accelerator + * implementation + * 3. Add the new capability to the returned PyObject Dictionary + */ +struct C10_API DeviceCapability { + union { + struct { + AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(DEFINE_SCALAR_TYPE) + } supported_scalar_types; + uint64_t capability_bits; // Allow direct bit manipulation + } capability_data; + + // Default constructor with all capabilities enabled. + DeviceCapability() { + capability_data.capability_bits = + ((1ULL << NUMBER_OF_DEVICE_CAPABILITIES) - 1); + } + + // Iterate supported ScalarTypes without allocating a vector + template + void forEachSupportedScalarType(F&& visitor) const { +#define VISIT_SCALAR_TYPE(_1, n) \ + if (capability_data.supported_scalar_types.has_##n) { \ + visitor(ScalarType::n); \ + } + + AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(VISIT_SCALAR_TYPE) + +#undef VISIT_SCALAR_TYPE + } +}; + +#undef DEFINE_SCALAR_ENUM +#undef DEFINE_SCALAR_TYPE +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/DeviceGuard.h b/backends/pytorch/include/c10/core/DeviceGuard.h new file mode 100644 index 000000000..389ac29d1 --- /dev/null +++ b/backends/pytorch/include/c10/core/DeviceGuard.h @@ -0,0 +1,207 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include + +namespace c10 { + +/// RAII guard that sets a certain default device in its constructor, and +/// changes it back to the device that was originally active upon destruction. +/// +/// The device is always reset to the one that was active at the time of +/// construction of the guard. Even if you `set_device` after construction, the +/// destructor will still reset the device to the one that was active at +/// construction time. +/// +/// This device guard does NOT have an uninitialized state; it is guaranteed +/// to reset a device on exit. If you are in a situation where you *might* +/// want to setup a guard (i.e., are looking for the moral equivalent +/// of std::optional), see OptionalDeviceGuard. +class DeviceGuard { + public: + /// No default constructor; see Note [Omitted default constructor from RAII] + explicit DeviceGuard() = delete; + + /// Set the current device to the passed Device. + explicit DeviceGuard(Device device) : guard_(device) {} + + /// This constructor is for testing only. + explicit DeviceGuard( + Device device, + const impl::DeviceGuardImplInterface* impl) + : guard_(device, impl) {} + + ~DeviceGuard() = default; + + /// Copy is disallowed + DeviceGuard(const DeviceGuard&) = delete; + DeviceGuard& operator=(const DeviceGuard&) = delete; + + /// Move is disallowed, as DeviceGuard does not have an uninitialized state, + /// which is required for moves on types with nontrivial destructors. + DeviceGuard(DeviceGuard&& other) = delete; + DeviceGuard& operator=(DeviceGuard&& other) = delete; + + /// Sets the device to the given one. The specified device must be consistent + /// with the device type originally specified during guard construction. + /// + /// TODO: The consistency check here is inconsistent with StreamGuard's + /// behavior with set_stream, where a stream on a different device than + /// the original one isn't an error; we just reset the stream and then + /// switch devices. + void reset_device(at::Device device) { + guard_.reset_device(device); + } + + /// This method is for testing only. + void reset_device( + at::Device device, + const impl::DeviceGuardImplInterface* impl) { + guard_.reset_device(device, impl); + } + + /// Sets the device index to the given one. The device type is inferred + /// from the original device type the guard was constructed with. + void set_index(DeviceIndex index) { + guard_.set_index(index); + } + + /// Returns the device that was set at the time the guard was constructed. + Device original_device() const { + return guard_.original_device(); + } + + /// Returns the most recent device that was set using this device guard, + /// either from construction, or via set_device. + Device current_device() const { + return guard_.current_device(); + } + + private: + impl::InlineDeviceGuard guard_; +}; + +/** + * A OptionalDeviceGuard is an RAII class that sets a device to some value on + * initialization, and resets the device to its original value on destruction. + * Morally, a OptionalDeviceGuard is equivalent to std::optional, + * but with extra constructors and methods as appropriate. + * + * Besides its obvious use (optionally applying a DeviceGuard), + * OptionalDeviceGuard is often also used for the following idiom: + * + * OptionalDeviceGuard g; + * for (const auto& t : tensors) { + * g.set_device(t.device()); + * do_something_with(t); + * } + * + * This usage is marginally more efficient than constructing a DeviceGuard every + * iteration of the for loop, as it avoids an unnecessary device reset. + * + * Unlike DeviceGuard, a OptionalDeviceGuard may be uninitialized. This occurs + * when you use the nullary constructor, or pass a nullopt to the constructor. + * Uninitialized OptionalDeviceGuards do *nothing*; they do not know what the + * original device was and they do not reset on destruction. This is why + * original_device() and current_device() return std::optional rather + * than Device (as they do in DeviceGuard), and also is why we didn't just + * provide OptionalDeviceGuard by default and hide DeviceGuard from users. + * + * The semantics of an OptionalDeviceGuard are exactly explained by thinking + * of it as an std::optional. In particular, an initialized + * OptionalDeviceGuard doesn't restore device to its value at construction; it + * restores device to its value *at initialization*. So if you have the + * program: + * + * setDevice(1); + * OptionalDeviceGuard g; + * setDevice(2); + * g.reset_device(Device(DeviceType::CUDA, 3)); // initializes! + * + * On destruction, g will reset device to 2, rather than 1. + * + * An uninitialized OptionalDeviceGuard is distinct from a (initialized) + * DeviceGuard whose original_device_ and current_device_ match, since the + * DeviceGuard will still reset the device to original_device_. + */ +class OptionalDeviceGuard { + public: + /// Create an uninitialized guard. Set the guard later using reset_device. + explicit OptionalDeviceGuard() = default; + + /// Initialize the guard, setting the current device to the passed Device. + explicit OptionalDeviceGuard(Device device) : guard_(device) {} + + /// Initialize the guard if a Device is passed; otherwise leave the + /// guard uninitialized. + explicit OptionalDeviceGuard(std::optional device) : guard_(device) {} + + /// Constructor for testing only. + explicit OptionalDeviceGuard( + Device device, + const impl::DeviceGuardImplInterface* impl) + : guard_(device, impl) {} + + ~OptionalDeviceGuard() = default; + /// Copy is disallowed + OptionalDeviceGuard(const OptionalDeviceGuard&) = delete; + OptionalDeviceGuard& operator=(const OptionalDeviceGuard&) = delete; + + /// Move is disallowed + /// See Note [Explicit initialization of optional fields] + /// and // Note [Move construction for RAII guards is tricky] + /// for rationale. + OptionalDeviceGuard(OptionalDeviceGuard&& other) = delete; + OptionalDeviceGuard& operator=(OptionalDeviceGuard&& other) = delete; + + /// Sets the device to the given one. The specified device must be consistent + /// with the device type originally specified during guard construction. + void reset_device(at::Device device) { + guard_.reset_device(device); + } + + /// For testing only + void reset_device( + at::Device device, + const impl::DeviceGuardImplInterface* impl) { + guard_.reset_device(device, impl); + } + + /// Returns the device that was set at the time the guard was constructed. + std::optional original_device() const { + return guard_.original_device(); + } + + /// Returns the most recent device that was set using this device guard, + /// either from construction, or via reset_device. + std::optional current_device() const { + return guard_.current_device(); + } + + private: + impl::InlineOptionalDeviceGuard guard_; +}; + +// Note [Whither the DeviceGuard boilerplate] +// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +// Design note: in principle, we could avoid these wrappers using: +// +// using DeviceGuard = impl::InlineDeviceGuard; +// using OptionalDeviceGuard = +// impl::InlineOptionalDeviceGuard; +// +// But the error messages are worse, and our users can't just look at the +// header file to find out what's going on. Furthermore, for specializations +// like CUDAStreamGuard, it can be profitable to replace some interfaces with +// refined types (e.g., return CUDAStream instead of Stream). So, we eat +// the boilerplate and write out the API explicitly. + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/DeviceType.h b/backends/pytorch/include/c10/core/DeviceType.h new file mode 100644 index 000000000..3847b5e26 --- /dev/null +++ b/backends/pytorch/include/c10/core/DeviceType.h @@ -0,0 +1,35 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include + +// If you modified DeviceType in caffe2/proto/caffe2.proto, please also sync +// your changes into torch/headeronly/core/DeviceType.h. +#include + +#include +#include + +namespace c10 { + +C10_API std::string DeviceTypeName(DeviceType d, bool lower_case = false); + +C10_API bool isValidDeviceType(DeviceType d); + +C10_API std::ostream& operator<<(std::ostream& stream, DeviceType type); + +C10_API void register_privateuse1_backend(const std::string& backend_name); +C10_API std::string get_privateuse1_backend(bool lower_case = true); + +C10_API bool is_privateuse1_backend_registered(); + +} // namespace c10 + +namespace torch { +// NOLINTNEXTLINE(misc-unused-using-decls) +using c10::DeviceType; +} // namespace torch + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/DispatchKey.h b/backends/pytorch/include/c10/core/DispatchKey.h new file mode 100644 index 000000000..2aa647574 --- /dev/null +++ b/backends/pytorch/include/c10/core/DispatchKey.h @@ -0,0 +1,750 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +// Semantically, each value of BackendComponent identifies a "backend" for our +// dispatch. Some functionalities that we may dispatch to are allowed to +// register different handlers for each backend. The BackendComponent is then +// used to figure out which backend implementation to dispatch to. + +// In implementation terms, the backend component identifies a specific "bit" in +// a DispatchKeySet. The bits in the DispatchKeySet are split between the bottom +// ~12 "BackendComponent" bits, while the remaining upper bits are assigned to +// functionalities. When we encounter a functionality bit that is known to be +// customizable per-backend, then we also look at the lower BackendComponent +// bits and take the highest bit to determine which backend's implementation to +// use. + +// WARNING! If you add a new backend component to the end of this list, +// make sure you register it before Meta. +// Meta must be at the end so that meta key in tls triggers meta kernels. +// (But you shouldn't: private use keys should have higher precedence than all +// built-in keys) + +// If you add a new (non-privateuse) backend here, +// make sure to add an Autograd fallthrough kernel +// in aten/src/ATen/core/VariableFallbackKernel.cpp + +#define C10_FORALL_BACKEND_COMPONENTS(_, extra) \ + _(CPU, extra) \ + _(CUDA, extra) \ + _(HIP, extra) \ + _(XLA, extra) \ + _(MPS, extra) \ + _(IPU, extra) \ + _(XPU, extra) \ + _(HPU, extra) \ + _(VE, extra) \ + _(Lazy, extra) \ + _(MTIA, extra) \ + _(MAIA, extra) \ + _(PrivateUse1, extra) \ + _(PrivateUse2, extra) \ + _(PrivateUse3, extra) \ + _(Meta, extra) + +// WARNING! If we add a new per-backend functionality key that has higher +// priority than Autograd, then make sure you update EndOfRuntimeBackendKeys + +#define C10_FORALL_FUNCTIONALITY_KEYS(_) \ + _(Dense, ) \ + _(Quantized, Quantized) \ + _(Sparse, Sparse) \ + _(SparseCsr, SparseCsr) \ + _(NestedTensor, NestedTensor) \ + _(AutogradFunctionality, Autograd) + +enum class BackendComponent : uint8_t { + + // A "backend" is colloquially used to refer to handlers for dispatch + // which actually implement the numerics of an operation in question. + // + // Due to the nature of the enum, these backends are specified in + // an ordered way, but for most backends this order is not semantically + // meaningful (e.g., it's valid to reorder these backends without changing + // semantics). The only situation when backend ordering is meaningful + // is when the backend participates in multiple dispatch with another + // backend; e.g., CPU and CUDA (cuda must have higher priority). + + // These keys don't correspond to individual kernels. + // Instead, they represent the backends that are allowed to override specific + // pieces of functionality: + // - dense kernels (e.g. DispatchKey::CPU) + // - sparse kernels (e.g. DispatchKey::SparseCPU) + // - quantized kernels (e.g. DispatchKey::QuantizedCPU) + // - autograd kernels (e.g. DispatchKey::AutogradCPU) + // We reserve space in the runtime operator table for this full cross product + // of + // [backends in this enum] x [keys below that are explicitly marked as having + // per-backend functionality] + // + // A meta tensor is a tensor without any data associated with it. (They + // have also colloquially been referred to as tensors on the "null" device). + // A meta tensor can be used to dry run operators without actually doing any + // computation, e.g., add on two meta tensors would give you another meta + // tensor with the output shape and dtype, but wouldn't actually add anything. + + InvalidBit = 0, +#define DEFINE_BACKEND_COMPONENT(n, _) n##Bit, + C10_FORALL_BACKEND_COMPONENTS(DEFINE_BACKEND_COMPONENT, unused) +#undef DEFINE_BACKEND_COMPONENT + + // Define an alias to represent end of backend dispatch keys. + // If you add new backend keys after PrivateUse3, please also update it here. + EndOfBackendKeys = MetaBit, +}; + +// Semantically, a dispatch key identifies a possible "level" in our +// dispatch, for which a handler may be registered. Each handler corresponds +// to a type of functionality. +// +// In implementation terms, the dispatch key identifies a specific "bit" in a +// DispatchKeySet. Higher bit indexes get handled by dispatching first (because +// we "count leading zeros" when we extract the highest priority dispatch +// key.) +// +// Note [DispatchKey Classification] +// This enum actually contains several types of keys, which are explained +// in more detail further down: +// (1) non-customizable backends (e.g. FPGA) +// (2) non-customizable functionalities (e.g. Functionalize) +// (3) functionalized that are customizable per backend (e.g. Dense, Sparse, +// AutogradFunctionality) (4) per-backend instances of customizable +// functionalities (e.g. CPU, SparseCPU, AutogradCPU) (5) alias keys (e.g. +// CompositeImplicitAutograd) +// +// Of the categories above, it's important to note: +// (a) which keys are assigned individual bits in a DispatchKeySet +// (b) which keys are assigned individual slots in the runtime operator table +// ("Runtime keys") +// +// (1), (2) and (3) all get their own dedicated bits in the DispatchKeySet. +// (1), (2) and (4) all get their own dedicated slots in the runtime operator +// table. + +// See Note [DispatchKeySet Internal Representation] for more details. +// +// NOTE: Keep the list in sync with `DispatchKey` in torchgen/model.py +enum class DispatchKey : uint16_t { + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~ UNDEFINED ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ // + // This is not a "real" functionality, but it exists to give us a "nullopt" + // element we can return for cases when a DispatchKeySet contains no elements. + // You can think a more semantically accurate definition of DispatchKey is: + // + // using DispatchKey = std::optional + // + // and Undefined == nullopt. We didn't actually represent + // it this way because std::optional would take two + // words, when DispatchKey fits in eight bits. + + Undefined = 0, + + // Define an alias for Undefined to represent CatchAll (long term + // this will get eliminated, but for now it's convenient) + CatchAll = Undefined, + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~ Functionality Keys ~~~~~~~~~~~~~~~~~~~~~~ // + // Every value in the enum (up to EndOfFunctionalityKeys) + // corresponds to an individual "functionality" that can be dispatched to. + // This is represented in the DispatchKeySet by assigning each of these enum + // values + // to each of the remaining (64 - len(BackendComponent)) bits. + // + // Most of these functionalities have a single handler assigned to them, + // making them "runtime keys". + // That map to a single slot in the runtime operator table. + // + // A few functionalities are allowed to be customizable per backend. + // See [Note: Per-Backend Functionality Dispatch Keys] for details. + + // See [Note: Per-Backend Functionality Dispatch Keys] + Dense, + + // Below are non-extensible backends. + // These are backends that currently don't have their own overrides for + // Autograd/Sparse/Quantized kernels, + // and we therefore don't waste space in the runtime operator table allocating + // space for them. + // If any of these backends ever need to customize, e.g., Autograd, then we'll + // need to add a DispatchKey::*Bit for them. + + // TODO: put this in BackendComponents + FPGA, // Xilinx support lives out of tree at + // https://gitlab.com/pytorch-complex/vitis_kernels + + Vulkan, // TODO: put this in BackendComponents + Metal, // TODO: put this in BackendComponents + + // See [Note: Per-Backend Functionality Dispatch Keys] + Quantized, + + // This backend is to support custom RNGs; it lets you go + // to a different kernel if you pass in a generator that is not a + // traditional CPUGeneratorImpl/CUDAGeneratorImpl. To make use of this + // key: + // 1) set it as a second parameter of at::Generator constructor call in + // the user-defined PRNG class. + // 2) use it as a dispatch key while registering custom kernels + // (templatized kernels specialized for user-defined PRNG class) + // intended for out of tree use; tested by aten/src/ATen/test/rng_test.cpp + CustomRNGKeyId, + + // TODO: Make Mkldnn a functionality key, so we can give it Meta + // support + // Here are backends which specify more specialized operators + // based on the layout of the tensor. Note that the sparse backends + // are one case where ordering matters: sparse multi-dispatches with + // the corresponding dense tensors, and must be handled before them. + MkldnnCPU, // registered at build/aten/src/ATen/RegisterMkldnnCPU.cpp + // NB: not to be confused with MKLDNN, which is Caffe2 only + + // See [Note: Per-Backend Functionality Dispatch Keys] + Sparse, + + SparseCsr, + + NestedTensor, + + // In some situations, it is not immediately obvious what the correct + // backend for function is, because the function in question doesn't + // have any "tensor" arguments. In this case, a BackendSelect function + // can be registered to implement the custom determination of the + // correct backend. + BackendSelect, + + Python, + + // Out-of-core key for Fake Tensor in torchdistx. + // See https://pytorch.org/torchdistx/latest/fake_tensor.html + // TODO: delete this in favor of Python-implemented fake tensor + Fake, + // See Note [Out-of-tree vmap+grad prototype]. The purpose of this key + // is to insert code after the "autograd subsystem" runs, so this key should + // be directly after ADInplaceOrView and all of the autograd keys. + FuncTorchDynamicLayerBackMode, + + // Alias and mutation removal. + // If some backends want to opt into only alias removal or only mutation + // removal, + // we can consider adding separate keys dedicated to those individual passes. + // See Note [Functionalization Pass In Core] for details. + Functionalize, + + // The named dispatch key is set for any tensors with named dimensions. + // Although we have a dispatch key for named tensors, for historical reasons, + // this dispatch key doesn't do any of the substantive functionality for named + // tensor (though, hypothetically, it could!) At the moment, it's just + // responsible for letting us give good error messages when operations + // don't support named tensors. + // + // NB: If you ever consider moving named tensor functionality into + // this dispatch key, note that it might be necessary add another dispatch + // key that triggers before composite operators, in case a composite operator + // has named dimension propagation that doesn't match that of its + // constituent parts. + // TODO: delete this once torchdim lands in functorch + Named, + + // The Conjugate dispatch key is set for any tensors that need to perform + // conjugation + // This is implemented at a dispatch level right before any backends run + Conjugate, + + // The Negative dispatch key is set for any tensors that need to perform + // negation + // This is implemented at a dispatch level right before any backends run + Negative, + + ZeroTensor, // registered at build/aten/src/ATen/RegisterZeroTensor.cpp + + // Note [ADInplaceOrView key] + // ADInplaceOrView key is used by inplace or view ops to register a kernel + // that does additional setup for future autograd computation. + // + // 1. For inplace ops this kernel does version bump + // 2. For view ops this kernel does `as_view` setup where we properly setup + // DifferentiableViewMeta on the view tensors. + // + // For other ops it's fallthrough kernel since there's no extra + // work to do. + // + // Note [Dream: skip VariableType kernel when requires_grad=false] + // + // In an ideal world where we can skip VariableType kernel for inputs + // with requires_grad=false, instead of a fallthrough kernel, we'll + // register a kernel shown below to all functional ops as well: + // torch::Tensor my_functional_op(...) { + // { + // // Note for every op in VariableType, you need to go through + // // `AutoDispatchBelowADInplaceOrView` guard exactly once to add the + // // key to TLS excluded set. If you don't go through it at all, + // // inplace/view ops called through `at::` inside your backend + // // kernel will dispatch to ADInplaceOrView kernels and do a lot + // // of extra work. + // at::AutoDispatchBelowADInplaceOrView guard; + // at::redispatch::my_functional_op(...); + // } + // } + // But this work is currently blocked since it adds an extra dispatch + // for all ops and it's non-trivial overhead at model level(a few percents). + // Thus our current approach takes advantage of the fact every kernel go + // through VariableType kernel first and pulls the + // `at::AutoDispatchBelowADInplaceOrView` guard of functional ops + // up to the `VariableType` kernel. Thus we only add the extra dispatch + // to view/inplace ops to minimize its perf impact to real models. + ADInplaceOrView, + // Note [Alias Dispatch Key : Autograd] + // All backends are oblivious to autograd; autograd is handled as a + // layer which happens on top of all backends. It inspects the autograd + // metadata of all inputs, determines what autograd metadata should be + // constructed by the output, and otherwise defers to the backend to + // actually do the numeric computation. Autograd contains + // the bulk of this logic. + + // Autograd is now an alias dispatch key which by default maps to all + // backend-specific autograd keys. + // Backend-specific allow backends to override the default kernel registered + // to Autograd key as needed. + // For example, XLA wants to define autograd for einsum directly. + // Registering a custom autograd implementation at the XLA key won't work + // because we process Autograd before XLA. This key has higher priority and + // gets processed first. You generally should NOT redispatch after handling + // autograd here (since that would result in execution of the Autograd + // operator, which you're trying to skip). In AutogradXLA implementations, + // you are responsible for handling autograd yourself, or deferring to other + // operators which support autograd. + + // Currently we only have backend-specific autograd keys for CPU/CUDA/XLA and + // reserved user-defined backends. All other in-tree backends share the + // AutogradOther key. We can add specific autograd key for those backends + // upon request. + AutogradOther, + + // See [Note: Per-Backend Functionality Dispatch Keys] + AutogradFunctionality, + + // NestedTensor is an example of something that isn't a "real backend" + // (because it mostly consists of redispatching kernels) + // but it would like to override autograd functionality in C++. + // We can handle cases like this by adding an extra functionality key + // exclusively for handling autograd for NestedTensor. + // lives out of tree at + // https://github.com/pytorch/nestedtensor + AutogradNestedTensor, + + Tracer, + + // TODO: make Autocast a functionality key + // Autocasting precedes VariableTypeId, to ensure casts are autograd-exposed + // and inputs are saved for backward in the post-autocast type. + AutocastCPU, + AutocastMTIA, + AutocastMAIA, + AutocastXPU, + AutocastIPU, + AutocastHPU, + AutocastXLA, + // AutocastXLA is only being used for TPUs. XLA GPUs continue to use + // AutocastCUDA. + AutocastMPS, + AutocastCUDA, + AutocastPrivateUse1, + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~ WRAPPERS ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ // + // There are a number of alternative modes which may want to handle before + // autograd; for example, error checking, tracing, profiling or vmap. They + // go here. + + FuncTorchBatched, // See Note [Out-of-tree vmap+grad prototype] + + // Dispatch key for BatchedTensorImpl wrapping a nested tensor. + BatchedNestedTensor, + + FuncTorchVmapMode, // See Note [Out-of-tree vmap+grad prototype] + + // This is the dispatch key for BatchedTensorImpl, which is used to implement + // batching rules for vmap. + Batched, + + // When we are inside a vmap, all tensors dispatch on this key. + // See Note: [DispatchKey::VmapMode usage] for more details. + VmapMode, + + FuncTorchGradWrapper, // See Note [Out-of-tree vmap+grad prototype] + + // Out-of-core key for Deferred Module Initialization in torchdistx. + // See https://pytorch.org/torchdistx/latest/deferred_init.html + DeferredInit, + + // Used by Python key logic to know the set of tls on entry to the dispatcher + // This kernel assumes it is the top-most non-functorch-related DispatchKey. + // If you add a key above, make sure to update the fallback implementation for + // this. + PythonTLSSnapshot, + + // This key should be at the very top of the dispatcher + FuncTorchDynamicLayerFrontMode, // See Note [Out-of-tree vmap+grad prototype] + + // TESTING: This is intended to be a generic testing tensor type id. + // Don't use it for anything real; its only acceptable use is within a single + // process test. Use it by creating a TensorImpl with this DispatchKey, and + // then registering operators to operate on this type id. See + // aten/src/ATen/core/dispatch/backend_fallback_test.cpp for a usage example. + TESTING_ONLY_GenericWrapper, + + // TESTING: This is intended to be a generic testing tensor type id. + // Don't use it for anything real; its only acceptable use is within a ingle + // process test. Use it by toggling the mode on and off via + // TESTING_ONLY_tls_generic_mode_set_enabled and then registering operators + // to operate on this type id. See + // aten/src/ATen/core/dispatch/backend_fallback_test.cpp + // for a usage example + TESTING_ONLY_GenericMode, + + // This key is used for pre-dispatch tracing in make_fx. + // It has lower priority than the PythonDispatcher key + // because we use the PythonDispatcher to intercept the key from python, + // and avoid having to implement it in C++. + PreDispatch, + + // This is a bypass that allows you to skip running the C++ dispatcher + // entirely + PythonDispatcher, + + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ FIN ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ // + EndOfFunctionalityKeys, // End of functionality keys. + +// ~~~~~~~~~~~~~~ "Dense" Per-Backend Dispatch keys ~~~~~~~~~~~~~~~~~~~~ // +// Here are backends which you think of as traditionally specifying +// how to implement operations on some device. + +#define DEFINE_PER_BACKEND_KEYS_FOR_BACKEND(n, prefix) prefix##n, + +#define DEFINE_PER_BACKEND_KEYS(fullname, prefix) \ + StartOf##fullname##Backends, \ + C10_FORALL_BACKEND_COMPONENTS( \ + DEFINE_PER_BACKEND_KEYS_FOR_BACKEND, prefix) \ + EndOf##fullname##Backends = prefix##Meta, + + C10_FORALL_FUNCTIONALITY_KEYS(DEFINE_PER_BACKEND_KEYS) + +#undef DEFINE_PER_BACKEND_KEYS +#undef DEFINE_PER_BACKEND_KEYS_FOR_BACKEND + + EndOfRuntimeBackendKeys = EndOfAutogradFunctionalityBackends, + + // ~~~~~~~~~~~~~~~~~~~~~~ Alias Dispatch Keys ~~~~~~~~~~~~~~~~~~~~~~~~~~ // + // Note [Alias Dispatch Keys] + // Alias dispatch keys are synthetic dispatch keys which map to multiple + // runtime dispatch keys. Alisa keys have precedence, but they are always + // lower precedence than runtime keys. You can register a kernel to an + // alias key, the kernel might be populated to the mapped runtime keys + // during dispatch table computation. + // If a runtime dispatch key has multiple kernels from alias keys, which + // kernel wins is done based on the precedence of alias keys (but runtime + // keys always have precedence over alias keys). + // Alias keys won't be directly called during runtime. + + // See Note [Alias Dispatch Key : Autograd] + Autograd, + CompositeImplicitAutograd, // registered at + // build/aten/src/ATen/RegisterCompositeImplicitAutograd.cpp + + // Note: The alias keyset for FuncTorchBatchedDecomposition is disjoint from + // all + // other alias keysets + // and so precedence order doesn't matter + FuncTorchBatchedDecomposition, // registered at + // build/aten/src/ATen/RegisterFuncTorchBatchedDecomposition.cpp + // Note: The alias keyset for CompositeImplicitAutogradNestedTensor is + // disjoint from all other alias keysets + CompositeImplicitAutogradNestedTensor, // registered at + // build/aten/src/ATen/RegisterCompositeImplicitAutogradNestedTensor.cpp + CompositeExplicitAutograd, // registered at + // build/aten/src/ATen/RegisterCompositeExplicitAutograd.cpp + // See Note [CompositeExplicitAutogradNonFunctional Key] + CompositeExplicitAutogradNonFunctional, // registered at + // build/aten/src/ATen/RegisterCompositeExplicitAutograd.cpp + + // Define an alias key to represent end of alias dispatch keys. + // If you add new alias keys after Autograd, please also update it here. + StartOfAliasKeys = Autograd, + EndOfAliasKeys = CompositeExplicitAutogradNonFunctional, // + + // ~~~~~~~~~~~~~~~~~~~~~~~~~ BC ALIASES ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ // + // The aliases exist for backwards compatibility reasons, they shouldn't + // be used + CPUTensorId = CPU, + CUDATensorId = CUDA, + DefaultBackend = CompositeExplicitAutograd, + PrivateUse1_PreAutograd = AutogradPrivateUse1, + PrivateUse2_PreAutograd = AutogradPrivateUse2, + PrivateUse3_PreAutograd = AutogradPrivateUse3, + Autocast = AutocastCUDA, +}; + +// Note [Private use DispatchKey] +// ~~~~~~~~~~~~~~~~~~~~~~~~~~~ +// Private use tensor IDs are preallocated tensor type IDs for use in user +// applications. Similar to private use fields in HTTP, they can be used +// by end users for experimental or private applications, without needing +// to "standardize" the tensor ID (which would be done by submitting a PR +// to PyTorch to add your type ID). +// +// Private use tensor IDs are appropriate to use if you want to experiment +// with adding a new tensor type (without having to patch PyTorch first) or +// have a private, non-distributed application that needs to make use of a +// new tensor type. Private use tensor IDs are NOT appropriate to use for +// libraries intended to be distributed to further users: please contact +// the PyTorch developers to get a type ID registered in this case. +// +// We provide two classes of private user tensor id: regular DispatchKeys +// and Autograd DispatchKeys. DispatchKeys serve the role of ordinary "backend" +// DispatchKeys; if you were adding support for a new type of accelerator, you +// would use a backend DispatchKey, and ideally automatically reuse +// AutogradOther definitions already defined in PyTorch. AutogradPrivateUse +// DispatchKeys serve as "wrapper" DispatchKeys: they are only necessary for +// tensors that compose multiple internal tensors, and for cases when the +// built-in autograd formulas for operators are not appropriate. + +static_assert( + (static_cast(BackendComponent::EndOfBackendKeys) + + static_cast(DispatchKey::EndOfFunctionalityKeys)) <= 64, + "The BackendComponent and DispatchKey enums (below EndOfFunctionalityKeys)" + " both map to backend and functionality bits" + " into a 64-bit bitmask; you must have less than 64 total entries between them"); + +// Check if a DispatchKey is an alias mapping to other runtime keys. +constexpr bool isAliasDispatchKey(DispatchKey k) { + return k >= DispatchKey::StartOfAliasKeys && k <= DispatchKey::EndOfAliasKeys; +} + +// [Note: Per-Backend Functionality Dispatch Keys] +// Check if a DispatchKey is a per-backend functionality key +// Any functionalities that can be customized per-backend should be added here. +// These keys correspond to functionalities that can be customized individually +// per backend. While they only take up one bit in the `DispatchKeySet` bitset, +// they map to (# backends) slots in the operator table. +// Each of these keys also has a separate set of "runtime keys" in the dispatch +// key enum, per backend, which *do* map to the individual operator table slots. +// For example, the "Sparse" key maps to an individual bit in the +// DispatchKeySet, while `SparseCPU`, `SparseCUDA`, etc all map to individual +// slots in the runtime operator table. + +constexpr bool isPerBackendFunctionalityKey(DispatchKey k) { + if (k == DispatchKey::Dense || k == DispatchKey::Quantized || + k == DispatchKey::Sparse || k == DispatchKey::SparseCsr || + k == DispatchKey::AutogradFunctionality || + k == DispatchKey::NestedTensor) { + return true; + } else { + return false; + } +} + +// Note that this includes Undefined in the total count. +// BUT EndOfFunctionalityKeys is its own (placeholder) key. +// e.g. Undefined=0, Dense=1, Sparse=2, EndOfFunctionalityKeys=3. +// In the above example, there are 3 total functionality keys. +constexpr uint8_t num_functionality_keys = + static_cast(DispatchKey::EndOfFunctionalityKeys); + +constexpr uint8_t num_backends = + static_cast(BackendComponent::EndOfBackendKeys); + +// Note [No More Than 16 Backends] +// Search for this note to find places in the code where the "no more than 16 +// backends" invariant is baked in. +static_assert( + static_cast(BackendComponent::EndOfBackendKeys) <= 16, + "BackendComponent currently only supports <= 16 backends. If we really need to extend this, \ +there are a few places where this invariant is baked in"); + +constexpr uint8_t numPerBackendFunctionalityKeys() { + uint8_t count = 0; + for (uint8_t k = 0; k <= num_functionality_keys; ++k) { + if (isPerBackendFunctionalityKey(static_cast(k))) + ++count; + } + return count; +} + +#if defined(C10_MOBILE_TRIM_DISPATCH_KEYS) +// See [Note: Trimmed Mobile Dispatch Keys] +constexpr uint16_t num_runtime_entries = 8; +#else +constexpr uint16_t num_runtime_entries = num_functionality_keys + + (numPerBackendFunctionalityKeys() * (num_backends - 1)); +#endif + +// See Note [No More Than 16 Backends] +constexpr uint16_t full_backend_mask = + (static_cast(1) << num_backends) - 1; + +C10_API const char* toString(DispatchKey /*t*/); +C10_API const char* toString(BackendComponent /*t*/); +C10_API std::ostream& operator<<(std::ostream& /*str*/, DispatchKey /*rhs*/); +C10_API std::ostream& operator<<( + std::ostream& /*str*/, + BackendComponent /*rhs*/); + +C10_API DispatchKey getAutogradKeyFromBackend(BackendComponent k); + +// Parses a string into a dispatch key. +// If the string cannot be correctly parsed, throws an exception. +C10_API c10::DispatchKey parseDispatchKey(const std::string& k); + +// These are some convenience identifiers for dispatch keys which are +// shorter to type than their long counterparts. Note that some of these +// dispatch keys directly correspond to DeviceType; and most APIs that +// accept DispatchKey also accept DeviceType; e.g., +// torch::dispatch(torch::kCPU, ...) is also valid. +constexpr DispatchKey kAutograd = DispatchKey::Autograd; + +// See Note [The Ordering of Per-Backend Dispatch Keys Matters!] +// This function relies on the invariant that the dispatch keys between +// StartOfDenseBackends and EndOfRuntimeBackendKeys are ordered by backend +// in the same order as `BackendComponent`. +constexpr BackendComponent toBackendComponent(DispatchKey k) { + if (k >= DispatchKey::StartOfDenseBackends && + k <= DispatchKey::EndOfDenseBackends) { + return static_cast( + static_cast(k) - + static_cast(DispatchKey::StartOfDenseBackends)); + } else if ( + k >= DispatchKey::StartOfQuantizedBackends && + k <= DispatchKey::EndOfQuantizedBackends) { + return static_cast( + static_cast(k) - + static_cast(DispatchKey::StartOfQuantizedBackends)); + } else if ( + k >= DispatchKey::StartOfSparseBackends && + k <= DispatchKey::EndOfSparseBackends) { + return static_cast( + static_cast(k) - + static_cast(DispatchKey::StartOfSparseBackends)); + } else if ( + k >= DispatchKey::StartOfSparseCsrBackends && + k <= DispatchKey::EndOfSparseCsrBackends) { + return static_cast( + static_cast(k) - + static_cast(DispatchKey::StartOfSparseCsrBackends)); + } else if ( + k >= DispatchKey::StartOfNestedTensorBackends && + k <= DispatchKey::EndOfNestedTensorBackends) { + return static_cast( + static_cast(k) - + static_cast(DispatchKey::StartOfNestedTensorBackends)); + } else if ( + k >= DispatchKey::StartOfAutogradFunctionalityBackends && + k <= DispatchKey::EndOfAutogradFunctionalityBackends) { + return static_cast( + static_cast(k) - + static_cast( + DispatchKey::StartOfAutogradFunctionalityBackends)); + } else { + return BackendComponent::InvalidBit; + } +} + +constexpr DispatchKey toFunctionalityKey(DispatchKey k) { + if (k <= DispatchKey::EndOfFunctionalityKeys) { + return k; + } else if (k <= DispatchKey::EndOfDenseBackends) { + return DispatchKey::Dense; + } else if (k <= DispatchKey::EndOfQuantizedBackends) { + return DispatchKey::Quantized; + } else if (k <= DispatchKey::EndOfSparseBackends) { + return DispatchKey::Sparse; + } else if (k <= DispatchKey::EndOfSparseCsrBackends) { + return DispatchKey::SparseCsr; + } else if (k <= DispatchKey::EndOfNestedTensorBackends) { + return DispatchKey::NestedTensor; + } else if (k <= DispatchKey::EndOfAutogradFunctionalityBackends) { + return DispatchKey::AutogradFunctionality; + } else { + return DispatchKey::Undefined; + } +} + +BackendComponent toBackendComponent(DeviceType device_type); + +// Given (DispatchKey::Dense, BackendComponent::CUDABit), returns +// DispatchKey::CUDA. +// See Note [The Ordering of Per-Backend Dispatch Keys Matters!] +// This function relies on the invariant that the dispatch keys between +// StartOfDenseBackends and EndOfRuntimeBackendKeys are ordered by backend +// in the same order as `BackendComponent`. +constexpr DispatchKey toRuntimePerBackendFunctionalityKey( + DispatchKey functionality_k, + BackendComponent backend_k) { + if (functionality_k == DispatchKey::Dense) { + return static_cast( + static_cast(DispatchKey::StartOfDenseBackends) + + static_cast(backend_k)); + } + if (functionality_k == DispatchKey::Sparse) { + return static_cast( + static_cast(DispatchKey::StartOfSparseBackends) + + static_cast(backend_k)); + } + if (functionality_k == DispatchKey::SparseCsr) { + return static_cast( + static_cast(DispatchKey::StartOfSparseCsrBackends) + + static_cast(backend_k)); + } + if (functionality_k == DispatchKey::Quantized) { + return static_cast( + static_cast(DispatchKey::StartOfQuantizedBackends) + + static_cast(backend_k)); + } + if (functionality_k == DispatchKey::NestedTensor) { + return static_cast( + static_cast(DispatchKey::StartOfNestedTensorBackends) + + static_cast(backend_k)); + } + if (functionality_k == DispatchKey::AutogradFunctionality) { + return static_cast( + static_cast( + DispatchKey::StartOfAutogradFunctionalityBackends) + + static_cast(backend_k)); + } + return DispatchKey::Undefined; +} + +} // namespace c10 + +namespace torch { +// Expose the constant, but not the TYPE (DispatchKey is an implementation +// detail!) +// NOLINTNEXTLINE(misc-unused-using-decls) +using c10::kAutograd; +} // namespace torch + +// NB: You really shouldn't use this instance; this enum is guaranteed +// to be pretty small so a regular array should be acceptable. +namespace std { +template <> +struct hash { + typedef size_t result_type; + typedef c10::DispatchKey argument_type; + + size_t operator()(c10::DispatchKey x) const { + return static_cast(x); + } +}; +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/DispatchKeySet.h b/backends/pytorch/include/c10/core/DispatchKeySet.h new file mode 100644 index 000000000..ec3aff4e0 --- /dev/null +++ b/backends/pytorch/include/c10/core/DispatchKeySet.h @@ -0,0 +1,977 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum") + +namespace c10 { + +struct FunctionalityOffsetAndMask { + // empty constructor shouldn't be used; only needed to initialize + // the array before populating it. + FunctionalityOffsetAndMask() = default; + FunctionalityOffsetAndMask(uint16_t offset, uint16_t mask) + : offset(offset), mask(mask) {} + // This needs to big enough to cover the size of the operator table. + uint16_t offset{}; + // See Note [No More Than 16 Backends] + // This mask needs to be big enough to mask all of the backend bits. + // We probably don't ever want to have more than 16 backend bits, so uint16_t + // should be enough. + uint16_t mask{}; +}; +static_assert( + c10::num_runtime_entries < 65536, + "The dispatcher currently only supports up to 2^16 runtime entries"); + +C10_API std::array +initializeFunctionalityOffsetsAndMasks(); + +C10_ALWAYS_INLINE static const std:: + array& + offsetsAndMasks() { + static auto offsets_and_masks_ = initializeFunctionalityOffsetsAndMasks(); + return offsets_and_masks_; +} + +// A representation of a set of DispatchKeys. A DispatchKeySet contains both +// "functionality" bits and "backend bits", and every tensor holds its own +// DispatchKeySet. The Dispatcher implements multiple dispatch by grabbing the +// keyset on every input tensor, or’ing them together, and dispatching to a +// specific piece of functionality. The functionality bits are *ordered*. When +// multiple functionality bits are set, we use the highest priority +// functionality. Similarly, multiple backend bits can theoretically be set if +// you call an operator with multiple tensors from difference devices (e.g. CPU +// and CUDA), although support for mixed device dispatch is limited (the only +// kernels that gracefully handle mixed device inputs for now are cuda kernels +// that take in a scalar cpu tensor). + +// A representation of a set of DispatchKeys. A tensor may have multiple +// tensor type ids, e.g., a Variable tensor can also be a CPU tensor; the +// DispatchKeySet specifies what type ids apply. The internal representation is +// as a 64-bit bit set (this means only 64 tensor type ids are supported). +// +// As mentioned above, DispatchKeys are ordered; thus, we can ask questions like +// "what is the highest priority DispatchKey in the set"? (The set itself is +// not ordered; two sets with the same ids will always have the ids ordered in +// the same way.) +// +// Note [DispatchKeySet Internal Representation] +// Internally, dispatch keys are packed into 64-bit DispatchKeySet objects +// that get passed around at runtime. +// However, there isn't necessarily a 1-to-1 mapping between bits in the keyset +// and individual dispatch keys. +// +// First: why do we have this distinction, and why not map every dispatch key +// directly to a bit? This is mostly because we have several types of +// functionalities that different backends would like to customize. For example, +// we have: +// - "Dense": CPU, CUDA, XLA, ... (~12 keys) +// - "Sparse": SparseCPU, SparseCUDA, ... +// - "SparseCsr": SparseCsrCPU, SparseCsrCUDA, ... +// - "Quantized": QuantizedCPU, QuantizedCUDA, QuantizedXLA, ... +// - "Autograd": AutogradCPU, AutogradCUDA, Autograd XLA, ... +// The problem is that total number of keys grows quadratically with [# +// backends] x [# functionalities], making it very difficult to map each key +// directly to a bit in a bitset without dramatically increasing the size of the +// bitset over time. +// +// The two enums (BackendComponent and DispatchKey) can be divided roughly into +// 5 categories. +// +// (1) "Building block" keys +// (a) backends: Everything in the BackendComponent enum (e.g. CPUBit, +// CUDABit) (b) functionalities: (per-backend) functionality-bit DispatchKeys +// (e.g. AutogradFunctionality, SparseCsr, Sparse, Dense) +// (2) "Runtime" keys +// (a) "non-customizable backends" (e.g. FPGA) +// (b) "non-customizable functionalities" (e.g. Functionalize) +// (c) "per-backend instances of customizable functionalities" (e.g. CPU, +// SparseCPU, AutogradCPU) +// (3) "Alias" DispatchKeys (see Note [Alias Dispatch Keys]) +// +// (1) Building block keys always correspond to individual bits in a +// DispatchKeySet. They can also be combined in a DispatchKeySet to form actual +// runtime keys. e.g. +// auto dense_cpu_ks = DispatchKeySet({DispatchKey::CPUBit, +// DispatchKey::Dense}); +// // The keyset has the runtime dense-cpu key. +// dense_cpu_ks.has(DispatchKey::CPU); +// // And it contains the building block keys too. +// dense_cpu_ks.has(DispatchKey::CPUBit); +// dense_cpu_ks.has(DispatchKey::Dense); +// +// Not every backend and not every functionality counts as a "building block +// key". This is mostly to give us more levers to pull in the design space. +// Backend keys and functionality keys that count as "building blocks" will +// contribute to a full cross product of functionality that can be overridden. +// +// For example, right now we have at least 12 "backend" building +// blocks (CPU, CUDA, XLA, ...) and at least 5 "functionality" +// building blocks (Dense, Sparse, SparseCsr, Quantized, +// AutogradFunctionality, ...). These keys together allow every +// dispatcher operator to be customized in up to 12*4 different +// ways. Each of those requires a slot in the operator table of every +// dispatcher operator. Not every piece of functionality necessarily +// needs to be customizable per-backend, and not every backend +// necessarily needs to be able to customize every type of +// functionality. +// +// +// (2) Every runtime key corresponds directly to a slot in an operator's runtime +// dispatch table, and you can directly register kernels to a runtime dispatch +// key. +// +// For per-backend functionalities like "Dense" or "AutogradFunctionality", +// you can think of the corresponding runtime dispatch keys as "instances" of +// that functionality, per backend. E.g. "CPU", "CUDA", "XLA", etc. are all +// runtime instances of the "Dense" building block key. + +// (2a) and (2b) are represented identically in the DispatchKeySet logic: +// - backend-agnostic functionalities (e.g. FuncTorchBatched) are NOT +// customizable per backend. +// In order to do so, we'd need to promote it to a per-backend functionality +// "building block" key. +// - non-customizable backends (e.g. FPGA) can NOT customize existing +// functionality like Sparse, Autograd, etc. +// In order to do so, we'd need to promote it to a backend "building block" +// key. +// +// In both cases, these keys directly correspond to runtime slots in the +// operator table. +// +// +// (3) "Alias" keys +// See Note [Alias Dispatch Keys] +// +// Final note: for anyone making future changes to the Dispatcher + +// DispatchKeySet internals, there's a closed PR with a basic +// python-implementation of the Dispatcher that might be useful in quickly +// testing out and validating changes. See it at +// https://github.com/pytorch/pytorch/pull/68743 + +// An undefined tensor is one with an empty tensor type set. +class DispatchKeySet final { + public: + enum Full { FULL }; + enum FullAfter { FULL_AFTER }; + enum Raw { RAW }; + + // NB: default constructor representation as zero is MANDATORY as + // use of DispatchKeySet in TLS requires this. + constexpr DispatchKeySet() = default; + + constexpr DispatchKeySet(Full /*unused*/) + : repr_((1ULL << (num_backends + num_functionality_keys - 1)) - 1) {} + + constexpr DispatchKeySet(FullAfter /*unused*/, DispatchKey t) + // LSB after t are OK, but not t itself. + // "functionalities" have a notion of ordering (e.g. Autograd > Sparse > + // Quantized > Dense). But backends don't really have an ordering. + // Therefore, we're enforcing that FullAfter can only be used on + // "functionality" keys. + : repr_( + (1ULL + << (num_backends + static_cast(toFunctionalityKey(t)) - + 1)) - + 1) { + *this = add(DispatchKey::PythonDispatcher); + } + + // Public version of DispatchKeySet(uint64_t) API; external users + // must be explicit when they do this! + constexpr DispatchKeySet(Raw /*unused*/, uint64_t x) : repr_(x) {} + + constexpr explicit DispatchKeySet(BackendComponent k) { + if (k == BackendComponent::InvalidBit) { + repr_ = 0; + } else { + repr_ = 1ULL << (static_cast(k) - 1); + } + } + + constexpr explicit DispatchKeySet(DispatchKey k) { + // NOLINTNEXTLINE(bugprone-branch-clone) + if (k == DispatchKey::Undefined) { + // Case 1: handle Undefined specifically + repr_ = 0; + } else if (k <= DispatchKey::EndOfFunctionalityKeys) { + // Case 2: handle "functionality-only" keys + // These keys have a functionality bit set, but no backend bits + // These can technically be either: + // - valid runtime keys (e.g. DispatchKey::AutogradOther, + // DispatchKey::FuncTorchBatched, etc) + // - "building block" keys that aren't actual runtime keys (e.g. + // DispatchKey::Dense or Sparse) + uint64_t functionality_val = 1ULL + << (num_backends + static_cast(k) - 1); + repr_ = functionality_val; + } else if (k <= DispatchKey::EndOfRuntimeBackendKeys) { + // Case 3: "runtime" keys that have a functionality bit AND a backend bit. + // First compute which bit to flip for the functionality. + auto functionality_k = toFunctionalityKey(k); + // The - 1 is because Undefined is technically a "functionality" that + // doesn't show up in the bitset. So e.g. Dense is technically the second + // functionality, but the lowest functionality bit. + uint64_t functionality_val = 1ULL + << (num_backends + static_cast(functionality_k) - 1); + + // then compute which bit to flip for the backend + // Case 4a: handle the runtime instances of "per-backend functionality" + // keys For example, given DispatchKey::CPU, we should set: + // - the Dense functionality bit + // - the CPUBit backend bit + // first compute which bit to flip for the backend + auto backend_k = toBackendComponent(k); + uint64_t backend_val = backend_k == BackendComponent::InvalidBit + ? 0 + : 1ULL << (static_cast(backend_k) - 1); + repr_ = functionality_val + backend_val; + } else { + // At this point, we should have covered every case except for alias keys. + // Technically it would be possible to add alias dispatch keys to a + // DispatchKeySet, but the semantics are a little confusing and this + // currently isn't needed anywhere. + repr_ = 0; + } + } + + constexpr uint64_t keys_to_repr(std::initializer_list ks) { + uint64_t repr = 0; + for (auto k : ks) { + repr |= DispatchKeySet(k).repr_; + } + return repr; + } + + constexpr uint64_t backend_bits_to_repr( + std::initializer_list ks) { + uint64_t repr = 0; + for (auto k : ks) { + repr |= DispatchKeySet(k).repr_; + } + return repr; + } + + explicit constexpr DispatchKeySet(std::initializer_list ks) + : repr_(keys_to_repr(ks)) {} + + explicit constexpr DispatchKeySet(std::initializer_list ks) + // Note: for some reason, putting this logic directly in the constructor + // appears to fail to compile on CUDA 10.1. + // See an example internal failure at + // https://www.internalfb.com/intern/skycastle/run/76561193669136035/artifact/actionlog.76561193742069401.stderr + : repr_(backend_bits_to_repr(ks)) {} + + // Test if a DispatchKey is in the set + inline bool has(DispatchKey t) const { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(t != DispatchKey::Undefined); + return has_all(DispatchKeySet(t)); + } + constexpr bool has_backend(BackendComponent t) const { + return has_all(DispatchKeySet(t)); + } + + // Test if a DispatchKey is in the set + // Given a DispatchKeySet of functionality keys and (potentially) backend + // keys, tests if all of them are in the current set. + constexpr bool has_all(DispatchKeySet ks) const { + return static_cast((repr_ & ks.repr_) == ks.repr_); + } + + // Given a DispatchKeySet of functionality keys and (potentially) backend + // keys, tests if any of them are in the current set. This could technically + // be pretty easily implemented using has(). It is strictly a perf + // optimization though. There are many places in the code base where we want + // to test for multiple functionality keys together. HOWEVER, runtime + // per-backend functionality keys aren't allowed to be used with this + // function, because you can end up with weird results. e.g. + // DispatchKeySet(DispatchKey::AutogradCPU).has_any(DispatchKeySet(DispatchKey::CPU)) + // would return true. + inline bool has_any(DispatchKeySet ks) const { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + // Either there are no backend bits in the input keyset + ((ks.repr_ & full_backend_mask) == 0) || + // or there are no per-backend-functionality bits + // See [Note: Per-Backend Functionality Dispatch Keys] + ((ks & + DispatchKeySet({ + DispatchKey::Dense, + DispatchKey::Quantized, + DispatchKey::Sparse, + DispatchKey::SparseCsr, + DispatchKey::AutogradFunctionality, + }) + .repr_) == 0)); + return static_cast((repr_ & ks.repr_) != 0); + } + // Test if DispatchKeySet is a superset of ks. + bool isSupersetOf(DispatchKeySet ks) const { + return (repr_ & ks.repr_) == ks.repr_; + } + // Perform set union + constexpr DispatchKeySet operator|(DispatchKeySet other) const { + return DispatchKeySet(repr_ | other.repr_); + } + // Perform set intersection + constexpr DispatchKeySet operator&(DispatchKeySet other) const { + return DispatchKeySet(repr_ & other.repr_); + } + // Compute the set difference self - other, + // but ONLY for the functionality keys. + // Any backend bits set on self will remain unchanged. + // See Note [Removing keys from DispatchKeySet Only Affects Functionality + // Keys] + constexpr DispatchKeySet operator-(DispatchKeySet other) const { + return DispatchKeySet(repr_ & (full_backend_mask | ~other.repr_)); + } + + // Compute self ^ other + constexpr DispatchKeySet operator^(DispatchKeySet other) const { + return DispatchKeySet(repr_ ^ other.repr_); + } + bool operator==(DispatchKeySet other) const { + return repr_ == other.repr_; + } + bool operator!=(DispatchKeySet other) const { + return repr_ != other.repr_; + } + // Add a DispatchKey to the DispatchKey set. Does NOT mutate, + // returns the extended DispatchKeySet! + [[nodiscard]] constexpr DispatchKeySet add(DispatchKey t) const { + return *this | DispatchKeySet(t); + } + [[nodiscard]] constexpr DispatchKeySet add(DispatchKeySet ks) const { + return *this | ks; + } + + // Remove a DispatchKey from the DispatchKey set. + // This is generally not an operation you should be doing + // (it's used to implement the printing overload, operator<<) + // + // Note [Removing keys from DispatchKeySet Only Affects Functionality Keys] + // Only functionality bits are allowed to be removed from a keyset. + // For now, we're only allowing removal of "functionality bits" from the + // keyset, which is specifically needed by the fallthrough key calculation + // logic. Why is removing backend bits problematic? Consider this example: + // + // DispatchKeySet([DispatchKey.CPU, DispatchKey.AutogradCUDA, + // DispatchKey.CUDA]).remove(DispatchKey.AutogradCUDA) + // DispatchKeySet([DispatchKey.CPU, + // DispatchKey.AutogradCUDA]).remove(DispatchKey.AutogradCUDA) + // + // What do we want to happen? + // Technically, we'd like it to be true that after removal, + // the first keyset still has the CUDA dispatch key while the second doesn't. + // Unfortunately there's no way to represent that, because the two keysets are + // represented the same way internally: functionality bits: Autograd, Dense + // backend bits: CPU, CUDA + // + // Instead, remove(DispatchKey.AutogradCPU) will only remove the "Autograd" + // bit from the bitset. + [[nodiscard]] constexpr DispatchKeySet remove(DispatchKey t) const { + return DispatchKeySet( + repr_ & ~(DispatchKeySet(t).repr_ & ~full_backend_mask)); + } + // You're allowed to remove a backend bit from a DispatchKeySet, + // but you have to be explicit about it (remove_backend() instead of + // remove()). + constexpr DispatchKeySet remove_backend(BackendComponent b) const { + return DispatchKeySet(repr_ & ~(DispatchKeySet(b).repr_)); + } + // Is the set empty? (AKA undefined tensor) + bool empty() const { + return repr_ == 0; + } + uint64_t raw_repr() const { + return repr_; + } + + static DispatchKeySet from_raw_repr(uint64_t x) { + return DispatchKeySet(RAW, x); + } + + DispatchKey highestFunctionalityKey() const { + auto functionality_idx = indexOfHighestBit(); + // This means that none of the functionality bits were set. + if (functionality_idx < num_backends) + return DispatchKey::Undefined; + // The first num_backend bits in the keyset don't correspond to real + // dispatch keys. + return static_cast(functionality_idx - num_backends); + } + + // This is similar like toBackendComponent(DispatchKey), but less restrictive. + // toBackendComponent() errors out if the key that it was passed has no + // backend bits, which is useful for error checking. We need a version of that + // here that can also handle "fake" backends like FPGA, because they need to + // map to the AutogradOther key. For those backends, we return + // BackendComponent::InvalidBit. + BackendComponent highestBackendKey() const { + // mask to mask out functionality bits + auto backend_idx = + DispatchKeySet(repr_ & full_backend_mask).indexOfHighestBit(); + // all zeros across the backend bits means that no backend bits are set. + if (backend_idx == 0) + return BackendComponent::InvalidBit; + return static_cast(backend_idx); + } + + // returns the DispatchKey of highest priority in the set. + DispatchKey highestPriorityTypeId() const { + auto functionality_k = highestFunctionalityKey(); + if (isPerBackendFunctionalityKey(functionality_k)) { + return toRuntimePerBackendFunctionalityKey( + functionality_k, highestBackendKey()); + } + return functionality_k; + } + + // Returns the index of the most-significant bit in the keyset. + // This is used to as part of the calculation into the operator table to get: + // - the highest "functionality" bit in the keyset. + // - the highest "backend" bit in the keyset. + uint8_t indexOfHighestBit() const { + return 64 - llvm::countLeadingZeros(repr_); + } + +#if defined(C10_MOBILE_TRIM_DISPATCH_KEYS) + // [Note: Trimmed Mobile Dispatch Keys] + /** + * The method below maps the dispatch key in the enum DispatchKey to an + * integer index in the dispatchTable_ array in OperatorEntry. The array + * is trimmed for mobile to reduce peak memory usage since it's + * unnecessary to reserve additional space for dispatch keys that will + * never be used on mobile. + */ + int getDispatchTableIndexForDispatchKeySet() const { + auto dk = highestPriorityTypeId(); + switch (dk) { + case DispatchKey::Undefined: + return 0; + case DispatchKey::CPU: + return 1; + case DispatchKey::QuantizedCPU: + return 2; + case DispatchKey::SparseCPU: + return 3; + case DispatchKey::BackendSelect: + return 4; + case DispatchKey::ADInplaceOrView: + return 5; + case DispatchKey::AutogradOther: + return 6; + case DispatchKey::AutogradCPU: + return 7; + default: + return -1; + } + } +#else + // returns the index in the operator table of highest priority key in the the + // keyset Note that we could in theory implement this using + // highestPriorityTypeId(), but this code is very hotpath and we can do it + // faster without it. + int getDispatchTableIndexForDispatchKeySet() const { + auto functionality_idx = + DispatchKeySet(repr_ >> num_backends).indexOfHighestBit(); + auto offset_and_mask = offsetsAndMasks()[functionality_idx]; + // Mask the functionality bits out first, then right-shift by 1. + // right-shifting by 1 because everything is zero-indexed. + // E.g. 000001 (CPU) should give us an offset of 0, 000010 (CUDA) should + // give us an offset of 1, etc. + auto backend_idx = + DispatchKeySet((repr_ & offset_and_mask.mask) >> 1).indexOfHighestBit(); + return offset_and_mask.offset + backend_idx; + } +#endif + + // returns the "index" of the highest priority backend in the keyset. + // This is pretty similar to getBackendKey(), but: + // - It's hotpath code (part of the runtime bitset calculation) + // - I's returns an integer index, not an enum value + // - Everything is shifted to the right by 1. + // BackendComponent::InvalidBit is technically the lowest enum value, + // but it isn't included in the runtime table. So CPUBit = 1, CUDABit = 2, + // etc. + uint64_t getBackendIndex() const { + return DispatchKeySet((repr_ & full_backend_mask) >> 1).indexOfHighestBit(); + } + + private: + constexpr DispatchKeySet(uint64_t repr) : repr_(repr) {} + uint64_t repr_ = 0; + + public: + // STL iterator for DispatchKeySet. Iterates through all runtime DispatchKeys + // in the set. The iterator is only invalidated by the destruction of the + // underlying DispatchKeySet as the iterator stores a pointer to the raw + // representation of the DispatchKeySet. Note: When we encounter a per-backend + // functionality (e.g. Dense or Sparse), we will iterate through EVERY backend + // in the keyset, for that functionality. For example, if the next + // functionality key to iterate over is Autograd, and the backend bits in the + // keyset correspond to [BackendComponent::CPUBit, BackendComponent::CUDABit], + // then the next two keys we return will be DispatchKey::AutogradCPU, + // DispatchKey::AutogradCUDA (CPU first because it has lower precedence than + // CUDA in DispatchKey.h). + class iterator { + public: + using self_type = iterator; + using iterator_category = std::input_iterator_tag; + using value_type = DispatchKey; + using difference_type = ptrdiff_t; + using reference = value_type&; + using pointer = value_type*; + // final mask value should mask out the entire keyset + static const uint8_t end_iter_mask_val = + num_backends + num_functionality_keys; + // final key value should be the last DispatchKey + static const uint8_t end_iter_key_val = num_functionality_keys; + + // current_dispatchkey_idx_ will iterate through all functionality bits. + // current_backendcomponent_idx_ will iterate through all backend bits. + explicit iterator( + const uint64_t* data_ptr, + uint8_t next_functionality = num_backends, + uint8_t next_backend = 0) + : data_ptr_(data_ptr), + next_functionality_(next_functionality), + next_backend_(next_backend), + // These are in an invalid state at construction time, and set by the + // first increment call + current_dispatchkey_idx_(end_iter_key_val), + current_backendcomponent_idx_(end_iter_key_val) { + // Go to the first key in the set + TORCH_INTERNAL_ASSERT( + next_functionality_ >= num_backends, + "num_backends=", + static_cast(num_backends), + "next_functionality_=", + static_cast(next_functionality_)); + ++(*this); + } + + C10_API self_type& operator++(); + + self_type operator++(int) { + self_type previous_iterator = *this; + ++(*this); + return previous_iterator; + } + + bool operator==(const self_type& rhs) const { + return next_functionality_ == rhs.next_functionality_ && + current_dispatchkey_idx_ == rhs.current_dispatchkey_idx_ && + next_backend_ == rhs.next_backend_ && + current_backendcomponent_idx_ == rhs.current_backendcomponent_idx_; + } + bool operator!=(const self_type& rhs) const { + return next_functionality_ != rhs.next_functionality_ || + current_dispatchkey_idx_ != rhs.current_dispatchkey_idx_ || + next_backend_ != rhs.next_backend_ || + current_backendcomponent_idx_ != rhs.current_backendcomponent_idx_; + } + DispatchKey operator*() const { + auto functionality_key = + static_cast(current_dispatchkey_idx_); + if (isPerBackendFunctionalityKey(functionality_key)) { + auto next_key = toRuntimePerBackendFunctionalityKey( + functionality_key, + static_cast(current_backendcomponent_idx_)); + // We expect all of the Dense, Sparse, Quantized, and Autograd keys to + // be ordered the same way with respect to their backends + TORCH_INTERNAL_ASSERT( + toBackendComponent(next_key) == + static_cast(current_backendcomponent_idx_), + "Tried to map functionality key ", + toString(functionality_key), + " and backend bit ", + toString( + static_cast(current_backendcomponent_idx_)), + " to a runtime key, but ended up with ", + toString(next_key), + ". This can happen if the order of the backend dispatch keys in DispatchKey.h isn't consistent.", + " Please double check that enum for inconsistencies."); + return next_key; + } else { + return functionality_key; + } + } + + private: + const uint64_t* data_ptr_; + uint8_t next_functionality_; + uint8_t next_backend_; + uint8_t current_dispatchkey_idx_; + uint8_t current_backendcomponent_idx_; + }; + + public: + // Returns iterator to the first key in the set. If no keys are in the + // set, then will return the end iterator. + iterator begin() const { + return iterator(&repr_); + } + + // We do not need to iterate beyond EndOfFunctionalityKeys so we will treat + // this as the end iterator. + iterator end() const { + return iterator(&repr_, iterator::end_iter_mask_val); + } +}; + +C10_API std::string toString(DispatchKeySet /*ts*/); +C10_API std::ostream& operator<<(std::ostream& /*os*/, DispatchKeySet /*ts*/); + +inline int getDispatchTableIndexForDispatchKey(DispatchKey k) { + return DispatchKeySet(k).getDispatchTableIndexForDispatchKeySet(); +} + +// Alias key DispatchKey::Autograd maps to +// (autograd_dispatch_keyset x full_backend_mask) +// NB: keys in this set also get associated with CompositeImplicitAutograd +// +// Note [autograd_dispatch_keyset Does Not Include Backend Bits] +// We don't want to include any backend bits (BackendComponent::CPUBit, etc) +// directly in autograd_dispatch_keyset. +// Why? keysets like autograd_dispatch_keyset are commonly used to remove +// autograd keys from a DispatchKeySet throughout the code base. However, you +// are only allowed to remove functionality bits from a keyset, not backend +// bits. See Note [Removing keys from DispatchKeySet Only Affects Functionality +// Keys] for details. To be consistent and avoid confusion, we're explicitly +// setting up autograd_dispatch_keyset to not have any backend bits. +constexpr DispatchKeySet autograd_dispatch_keyset = DispatchKeySet({ + DispatchKey::AutogradFunctionality, + DispatchKey::AutogradOther, + DispatchKey::AutogradNestedTensor, +}); + +constexpr DispatchKeySet autocast_dispatch_keyset = DispatchKeySet({ + DispatchKey::AutocastCPU, + DispatchKey::AutocastMPS, + DispatchKey::AutocastCUDA, + DispatchKey::AutocastXPU, + DispatchKey::AutocastIPU, + DispatchKey::AutocastHPU, + DispatchKey::AutocastXLA, + DispatchKey::AutocastPrivateUse1, + DispatchKey::AutocastMTIA, + DispatchKey::AutocastMAIA, +}); + +// See Note [TLS Initialization] +constexpr DispatchKeySet default_included_set = DispatchKeySet({ + DispatchKey::BackendSelect, + DispatchKey::ADInplaceOrView, +}); + +constexpr DispatchKeySet default_excluded_set = DispatchKeySet({ + DispatchKey::AutocastCPU, + DispatchKey::AutocastMPS, + DispatchKey::AutocastCUDA, + DispatchKey::AutocastXPU, + DispatchKey::AutocastIPU, + DispatchKey::AutocastHPU, + DispatchKey::AutocastXLA, + DispatchKey::AutocastPrivateUse1, + DispatchKey::AutocastMTIA, + DispatchKey::AutocastMAIA, +}); + +constexpr DispatchKeySet autograd_dispatch_keyset_with_ADInplaceOrView = + autograd_dispatch_keyset | DispatchKeySet(DispatchKey::ADInplaceOrView); + +constexpr DispatchKeySet python_ks = DispatchKeySet({ + DispatchKey::Python, + DispatchKey::PythonTLSSnapshot, +}); + +constexpr DispatchKeySet sparse_ks = DispatchKeySet(DispatchKey::Sparse); + +constexpr DispatchKeySet sparse_csr_ks = DispatchKeySet(DispatchKey::SparseCsr); + +constexpr DispatchKeySet mkldnn_ks = DispatchKeySet(DispatchKey::MkldnnCPU); + +// backend dispatch keys that map to DispatchKey::AutogradOther +// NB: keys in this set also get associated with CompositeImplicitAutograd +constexpr DispatchKeySet autogradother_backends = + DispatchKeySet( + // HIP and VE aren't in this list: they now have their own backend bits + // which means that they can now have their own Autograd keys. + // Technically, HIP will now redispatch to its own custom AutogradHIP + // slot in the runtime table. + {DispatchKey::FPGA, + DispatchKey::Vulkan, + DispatchKey::Metal, + DispatchKey::CustomRNGKeyId, + DispatchKey::MkldnnCPU, + // Sparse and Quantized backends also live here. + DispatchKey::Sparse, + DispatchKey::SparseCsr, + DispatchKey::Quantized}) + // Including the backend bits because this keyset is used during op + // registration, which requires looping over all runtime autogradother + // backend keys. + | DispatchKeySet(DispatchKeySet::RAW, full_backend_mask); + +// The set of dispatch keys that come after autograd +// n.b. this relies on the fact that AutogradOther is currently the lowest +// Autograd key +constexpr DispatchKeySet after_autograd_keyset = + DispatchKeySet(DispatchKeySet::FULL_AFTER, c10::DispatchKey::AutogradOther); + +// The set of dispatch keys that come after ADInplaceOrView +constexpr DispatchKeySet after_ADInplaceOrView_keyset = DispatchKeySet( + DispatchKeySet::FULL_AFTER, + c10::DispatchKey::ADInplaceOrView); + +// The set of dispatch keys that come after Functionalize +constexpr DispatchKeySet after_func_keyset = + DispatchKeySet(DispatchKeySet::FULL_AFTER, c10::DispatchKey::Functionalize) + .remove( + // NOTE: we also need to remove ADInplaceOrView from the keyset when + // redispatching after the func kernels. This is because we're not + // calling the same op; we originally called an inplace op, and now + // we aren't. The original key calculation figured out which keys + // were Fallthrough based on the inplace op. That means that it did + // not include the ADInPlaceOrView kernel as a fallthrough key. + // However, we WANT the ADInPlaceOrView kernel to be ignored now + // that we're calling an out-of-place op. Re-invoking + // Dispatcher::call would re-run the Fallthrough key calculation and + // get us that, But at::redispatch is more performant. We can get + // away with it by explicitly removing the key here. + c10::DispatchKey::ADInplaceOrView); + +constexpr DispatchKeySet backend_bitset_mask = + DispatchKeySet(DispatchKeySet::RAW, (1ULL << num_backends) - 1); + +constexpr auto inplace_or_view_ks = + DispatchKeySet(DispatchKey::ADInplaceOrView); +constexpr auto autograd_cpu_ks = DispatchKeySet(DispatchKey::AutogradCPU); +constexpr auto autograd_ipu_ks = DispatchKeySet(DispatchKey::AutogradIPU); +constexpr auto autograd_mtia_ks = DispatchKeySet(DispatchKey::AutogradMTIA); +constexpr auto autograd_maia_ks = DispatchKeySet(DispatchKey::AutogradMAIA); +constexpr auto autograd_xpu_ks = DispatchKeySet(DispatchKey::AutogradXPU); +constexpr auto autograd_cuda_ks = DispatchKeySet(DispatchKey::AutogradCUDA); +constexpr auto autograd_xla_ks = DispatchKeySet(DispatchKey::AutogradXLA); +constexpr auto autograd_lazy_ks = DispatchKeySet(DispatchKey::AutogradLazy); +constexpr auto autograd_meta_ks = DispatchKeySet(DispatchKey::AutogradMeta); +constexpr auto autograd_mps_ks = DispatchKeySet(DispatchKey::AutogradMPS); +constexpr auto autograd_hpu_ks = DispatchKeySet(DispatchKey::AutogradHPU); +constexpr auto autograd_privateuse1_ks = + DispatchKeySet(DispatchKey::AutogradPrivateUse1); +constexpr auto autograd_privateuse2_ks = + DispatchKeySet(DispatchKey::AutogradPrivateUse2); +constexpr auto autograd_privateuse3_ks = + DispatchKeySet(DispatchKey::AutogradPrivateUse3); +constexpr auto autograd_other_ks = DispatchKeySet(DispatchKey::AutogradOther); +constexpr auto autograd_nested = + DispatchKeySet(DispatchKey::AutogradNestedTensor); +// keyset corresponding to functorch keys that have their own dedicated +// TensorImpl subclass. +constexpr auto functorch_transforms_ks = DispatchKeySet( + {DispatchKey::FuncTorchBatched, + DispatchKey::FuncTorchVmapMode, + DispatchKey::Batched, + DispatchKey::VmapMode, + DispatchKey::FuncTorchGradWrapper}); + +constexpr auto functorch_batched_ks = + DispatchKeySet({DispatchKey::FuncTorchBatched}); + +// This keyset has: +// (1) the functionality bits corresponding to backends (dense, sparse, +// quantized) (2) all of the backend bits set +constexpr DispatchKeySet backend_functionality_keys = + DispatchKeySet({ + DispatchKey::Dense, + DispatchKey::Quantized, + DispatchKey::Sparse, + DispatchKey::SparseCsr, + }) | + DispatchKeySet(DispatchKeySet::RAW, full_backend_mask); + +struct OpTableOffsetAndMask { + uint16_t offset; + uint16_t backend_mask; +}; + +static_assert( + num_backends <= 16, + "Right now we expect the number of backends not to exceed 16. In the (unlikely) event" + " that this changes, the size of OpTableOffsetAndMask::backend_mask needs to be increased too."); + +// true if t is a backend dispatch key +C10_API bool isBackendDispatchKey(DispatchKey t); + +// Resolve alias dispatch key to DispatchKeySet if applicable +C10_API DispatchKeySet getRuntimeDispatchKeySet(DispatchKey t); + +// Resolve alias dispatch key to DispatchKeySet if applicable, +// and check if k is a part of that set +C10_API bool runtimeDispatchKeySetHas(DispatchKey t, DispatchKey k); + +// Returns a DispatchKeySet of all backend keys mapped to Autograd dispatch key +// t, DispatchKeySet is empty if t is not alias of DispatchKey::Autograd. +C10_API DispatchKeySet getBackendKeySetFromAutograd(DispatchKey t); + +// Returns a DispatchKeySet of autograd related keys mapped to backend. +// for a given backend key, use the associated autograd key. +// for non-backend keys, use AutogradOther as a default. +// Note: it's convenient and fast to return a default here rather than (say) +// returning an std::optional, or throwing. But it makes callers +// responsible for either a) enforcing the invariant that only backend keys +// be passed as arguments, or b) interpreting our return value carefully. +inline DispatchKeySet getAutogradRelatedKeySetFromBackend(BackendComponent t) { + switch (t) { + case BackendComponent::CPUBit: + return inplace_or_view_ks | autograd_cpu_ks; + case BackendComponent::IPUBit: + return inplace_or_view_ks | autograd_ipu_ks; + case BackendComponent::MTIABit: + return inplace_or_view_ks | autograd_mtia_ks; + case BackendComponent::MAIABit: + return inplace_or_view_ks | autograd_maia_ks; + case BackendComponent::XPUBit: + return inplace_or_view_ks | autograd_xpu_ks; + case BackendComponent::CUDABit: + return inplace_or_view_ks | autograd_cuda_ks; + case BackendComponent::XLABit: + return inplace_or_view_ks | autograd_xla_ks; + case BackendComponent::LazyBit: + return inplace_or_view_ks | autograd_lazy_ks; + case BackendComponent::MetaBit: + return inplace_or_view_ks | autograd_meta_ks; + case BackendComponent::MPSBit: + return inplace_or_view_ks | autograd_mps_ks; + case BackendComponent::HPUBit: + return inplace_or_view_ks | autograd_hpu_ks; + case BackendComponent::PrivateUse1Bit: + return inplace_or_view_ks | autograd_privateuse1_ks; + case BackendComponent::PrivateUse2Bit: + return inplace_or_view_ks | autograd_privateuse2_ks; + case BackendComponent::PrivateUse3Bit: + return inplace_or_view_ks | autograd_privateuse3_ks; + default: + return inplace_or_view_ks | autograd_other_ks; + } +} + +// Returns a DispatchKeySet of autocast related keys mapped to backend. +inline DispatchKeySet getAutocastRelatedKeySetFromBackend(BackendComponent t) { + constexpr auto autocast_cpu_ks = DispatchKeySet(DispatchKey::AutocastCPU); + constexpr auto autocast_mtia_ks = DispatchKeySet(DispatchKey::AutocastMTIA); + constexpr auto autocast_maia_ks = DispatchKeySet(DispatchKey::AutocastMAIA); + constexpr auto autocast_xpu_ks = DispatchKeySet(DispatchKey::AutocastXPU); + constexpr auto autocast_ipu_ks = DispatchKeySet(DispatchKey::AutocastIPU); + constexpr auto autocast_hpu_ks = DispatchKeySet(DispatchKey::AutocastHPU); + constexpr auto autocast_cuda_ks = DispatchKeySet(DispatchKey::AutocastCUDA); + constexpr auto autocast_xla_ks = DispatchKeySet(DispatchKey::AutocastXLA); + constexpr auto autocast_privateuse1_ks = + DispatchKeySet(DispatchKey::AutocastPrivateUse1); + constexpr auto autocast_mps_ks = DispatchKeySet(DispatchKey::AutocastMPS); + switch (t) { + case BackendComponent::CPUBit: + return autocast_cpu_ks; + case BackendComponent::MTIABit: + return autocast_mtia_ks; + case BackendComponent::MAIABit: + return autocast_maia_ks; + case BackendComponent::XPUBit: + return autocast_xpu_ks; + case BackendComponent::IPUBit: + return autocast_ipu_ks; + case BackendComponent::HPUBit: + return autocast_hpu_ks; + case BackendComponent::CUDABit: + return autocast_cuda_ks; + case BackendComponent::XLABit: + return autocast_xla_ks; + case BackendComponent::PrivateUse1Bit: + return autocast_privateuse1_ks; + case BackendComponent::MPSBit: + return autocast_mps_ks; + default: + return DispatchKeySet(); + } +} + +// returns the "backend" DispatchKey of highest priority in the set. +// This is basically like highestBackendKey(), except that we have some +// "functionality" bits that correspond to backends (Sparse, Quantized) +inline DispatchKey highestPriorityBackendTypeId(DispatchKeySet ks) { + return (ks & backend_functionality_keys).highestPriorityTypeId(); +} + +// This API exists because we have a use case for checking +// getRuntimeDispatchKeySet(alias).has(DispatchKey::Undefined) +// in OperatorEntry.cpp but we disallow it in has() API. +C10_API bool isIncludedInAlias(DispatchKey k, DispatchKey alias); + +// Historically, every tensor only had a single DispatchKey, and it was always +// something like CPU, and there wasn't any of this business where TLS +// could cause the DispatchKey of a tensor to change. But we still have some +// legacy code that is still using DispatchKey for things like instanceof +// checks; if at all possible, refactor the code to stop using DispatchKey in +// those cases. +inline DispatchKey legacyExtractDispatchKey(DispatchKeySet s) { + // NB: If you add any extra keys that can be stored in TensorImpl on + // top of existing "backend" keys like CPU/CUDA, you need to add it + // here. At the moment, autograd keys and ADInplaceOrView key need this + // treatment; + return (s - autograd_dispatch_keyset_with_ADInplaceOrView - + autocast_dispatch_keyset - + DispatchKeySet( + {DispatchKey::Functionalize, + DispatchKey::PythonTLSSnapshot, + DispatchKey::FuncTorchGradWrapper, + DispatchKey::FuncTorchVmapMode, + DispatchKey::FuncTorchBatched, + DispatchKey::Python})) + .highestPriorityTypeId(); +} + +template +using is_not_DispatchKeySet = std::negation>; + +// Given a function type, constructs a function_traits type that drops the first +// parameter type if the first parameter is of type DispatchKeySet. NB: +// DispatchKeySet is currently explicitly hidden from JIT (mainly to avoid +// pushing unnecessary arguments on the stack - see Note [ Plumbing Keys Through +// the Dispatcher] for details). If at any point in the future we need to expose +// this type to JIT, revisit the usage of this type alias. +template +using remove_DispatchKeySet_arg_from_func = guts::make_function_traits_t< + typename guts::infer_function_traits_t::return_type, + typename std::conditional_t< + std::is_same_v< + DispatchKeySet, + typename guts::typelist::head_with_default_t< + void, + typename guts::infer_function_traits_t< + FuncType>::parameter_types>>, + guts::typelist::drop_if_nonempty_t< + typename guts::infer_function_traits_t::parameter_types, + 1>, + typename guts::infer_function_traits_t::parameter_types>>; +} // namespace c10 + +C10_DIAGNOSTIC_POP() + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/Event.h b/backends/pytorch/include/c10/core/Event.h new file mode 100644 index 000000000..aed1a213b --- /dev/null +++ b/backends/pytorch/include/c10/core/Event.h @@ -0,0 +1,142 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include + +namespace c10 { + +/** + * A backend-generic movable, not copyable, not thread-safe event. + * + * The design of this event follows that of CUDA and HIP events. These events + * are recorded and waited on by streams and can be rerecorded to, + * each rerecording essentially creating a new version of the event. + * For example, if (in CPU time), stream X is asked to record E, + * stream Y waits on E, and stream X is asked to record E again, then Y will + * wait for X to finish the first call to record and not the second, because + * it's waiting on the first version of event E, not the second. + * Querying an event only returns the status of its most recent version. + * + * Backend-generic events are implemented by this class and + * impl::InlineEvent. In addition to these events there are also + * some backend-specific events, like ATen's CUDAEvent. Each of these + * classes has its own use. + * + * impl::InlineEvent<...> or a backend-specific event should be + * preferred when the backend is known at compile time and known to + * be compiled. Backend-specific events may have additional functionality. + * + * This Event should be used if a particular backend may not be available, + * or the backend required is not known at compile time. + * + * These generic events are built on top of DeviceGuardImpls, analogous + * to DeviceGuard and InlineDeviceGuard. The name "DeviceGuardImpls," + * is no longer entirely accurate, as these classes implement the + * backend-specific logic for a generic backend interface. + * + * See DeviceGuardImplInterface.h for a list of all supported flags. + */ + +struct Event final { + // Constructors + Event() = delete; + Event( + const DeviceType _device_type, + const EventFlag _flag = EventFlag::PYTORCH_DEFAULT) + : impl_{_device_type, _flag} {} + + // Copy constructor and copy assignment operator (deleted) + Event(const Event&) = delete; + Event& operator=(const Event&) = delete; + + // Move constructor and move assignment operator + Event(Event&&) noexcept = default; + Event& operator=(Event&&) noexcept = default; + + // Destructor + ~Event() = default; + + // Getters + Device device() const noexcept { + return Device(device_type(), device_index()); + } + DeviceType device_type() const noexcept { + return impl_.device_type(); + } + DeviceIndex device_index() const noexcept { + return impl_.device_index(); + } + EventFlag flag() const noexcept { + return impl_.flag(); + } + bool was_marked_for_recording() const noexcept { + return impl_.was_marked_for_recording(); + } + + /** + * Calls record() if and only if record() has never been called for this + * event. Note: because Event is not thread-safe recordOnce() may call + * record() multiple times if called from multiple threads. + */ + void recordOnce(const Stream& stream) { + impl_.recordOnce(stream); + } + + /** + * Increments the event's version and enqueues a job with this version + * in the stream's work queue. When the stream process that job + * it notifies all streams waiting on / blocked by that version of the + * event to continue and marks that version as recorded. + * */ + void record(const Stream& stream) { + impl_.record(stream); + } + + /** + * Does nothing if the event has not been scheduled to be recorded. + * If the event was previously enqueued to be recorded, a command + * to wait for the version of the event that exists at the time of this call + * is inserted in the stream's work queue. + * When the stream reaches this command it will stop processing + * additional commands until that version of the event is marked as recorded. + */ + void block(const Stream& stream) const { + impl_.block(stream); + } + + /** + * Returns true if (and only if) + * (1) the event has never been scheduled to be recorded + * (2) the current version is marked as recorded. + * Returns false otherwise. + */ + bool query() const { + return impl_.query(); + } + + double elapsedTime(const Event& event) const { + return impl_.elapsedTime(event.impl_); + } + + void* eventId() const { + return impl_.eventId(); + } + + void synchronize() const { + impl_.synchronize(); + } + + private: + impl::InlineEvent impl_; +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/GeneratorImpl.h b/backends/pytorch/include/c10/core/GeneratorImpl.h new file mode 100644 index 000000000..7d7aac924 --- /dev/null +++ b/backends/pytorch/include/c10/core/GeneratorImpl.h @@ -0,0 +1,116 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include +#include +#include +#include +#include +#include + +/** + * Note [Generator] + * ~~~~~~~~~~~~~~~~ + * A Pseudo Random Number Generator (PRNG) is an engine that uses an algorithm + * to generate a seemingly random sequence of numbers, that may be later be used + * in creating a random distribution. Such an engine almost always maintains a + * state and requires a seed to start off the creation of random numbers. Often + * times, users have found it beneficial to be able to explicitly create, + * retain, and destroy PRNG states and also be able to have control over the + * seed value. + * + * A Generator in ATen gives users the ability to read, write and modify a PRNG + * engine. For instance, it does so by letting users seed a PRNG engine, fork + * the state of the engine, etc. + * + * By default, there is one generator per device, and a device's generator is + * lazily created. A user can use the torch.Generator() api to create their own + * generator. Currently torch.Generator() can only create a CPUGeneratorImpl. + */ + +/** + * Note [Acquire lock when using random generators] + * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + * Generator and its derived classes are NOT thread-safe. Please note that most + * of the places where we have inserted locking for generators are historically + * based, and we haven't actually checked that everything is truly thread safe + * (and it probably isn't). Please use the public mutex_ when using any methods + * from these classes, except for the read-only methods. You can learn about the + * usage by looking into the unittests (aten/src/ATen/cpu_generator_test.cpp) + * and other places where we have used lock_guard. + * + * TODO: Look into changing the threading semantics of Generators in ATen (e.g., + * making them non-thread safe and instead making the generator state + * splittable, to accommodate forks into other threads). + */ + +namespace c10 { + +// The default seed is selected to be a large number +// with good distribution of 0s and 1s in bit representation +constexpr uint64_t default_rng_seed_val = 67280421310721; + +struct C10_API GeneratorImpl : public c10::intrusive_ptr_target { + // Constructors + GeneratorImpl(Device device_in, DispatchKeySet key_set); + + // Delete all copy and move assignment in favor of clone() + // method + GeneratorImpl(const GeneratorImpl& other) = delete; + GeneratorImpl(GeneratorImpl&& other) = delete; + GeneratorImpl& operator=(const GeneratorImpl& other) = delete; + GeneratorImpl& operator=(GeneratorImpl&& other) = delete; + + ~GeneratorImpl() override = default; + c10::intrusive_ptr clone() const; + + // Common methods for all generators + virtual void set_current_seed(uint64_t seed) = 0; + virtual void set_offset(uint64_t offset) = 0; + virtual uint64_t get_offset() const = 0; + virtual uint64_t current_seed() const = 0; + virtual uint64_t seed() = 0; + virtual void set_state(const c10::TensorImpl& new_state) = 0; + virtual c10::intrusive_ptr get_state() const = 0; + virtual void graphsafe_set_state( + const c10::intrusive_ptr& new_state); + virtual c10::intrusive_ptr graphsafe_get_state() const; + Device device() const; + + // See Note [Acquire lock when using random generators] + std::mutex mutex_; + + DispatchKeySet key_set() const { + return key_set_; + } + + inline void set_pyobj(PyObject* pyobj) noexcept { + pyobj_ = pyobj; + } + + inline PyObject* pyobj() const noexcept { + return pyobj_; + } + + protected: + Device device_; + DispatchKeySet key_set_; + PyObject* pyobj_ = nullptr; + + virtual GeneratorImpl* clone_impl() const = 0; +}; + +namespace detail { + +C10_API uint64_t getNonDeterministicRandom(bool is_cuda = false); + +} // namespace detail + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/InferenceMode.h b/backends/pytorch/include/c10/core/InferenceMode.h new file mode 100644 index 000000000..8da25b542 --- /dev/null +++ b/backends/pytorch/include/c10/core/InferenceMode.h @@ -0,0 +1,96 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include + +namespace c10 { + +// A RAII, thread local (!) guard that enables or disables inference mode upon +// construction, and sets it back to the original value upon destruction. +struct C10_API InferenceMode { + // Note [Expected TLS state in InferenceMode]: + // InferenceMode: ADInplaceOrView not in + // raw_local_dispatch_key_set.included(), + // Autograd in raw_local_dispatch_key_set.excluded() + // GradMode is disabled. + // NormalMode: ADInplaceOrView in raw_local_dispatch_key_set.included(), + // Autograd not in raw_local_dispatch_key_set.excluded() + // GradMode is enabled by default unless toggled manually + // through other APIs, e.g. NoGradGuard. + // + // Invariant: + // - ADInplaceOrView is never in the excluded set + // - Autograd is never in the included set + // - Setting InferenceMode will set GradMode accordingly, but not vice versa. + // + // 1. Why do we put ADInplaceOrView in included set outside InferenceMode? + // + // Inplace update to inference tensor outside InferenceMode is not + // allowed. See Note [Inplace update inference tensor] for more details. + // Without going through ADInplaceOrView kernel, we cannot throw error + // for `inference_tensor.add_(1)` case. + // + // 2. Why not put ADInplaceOrView in the excluded set inside InferenceMode? + // + // For example: + // torch::Tensor a = torch::ones({1, 2, 3}).set_requires_grad(true); + // torch::Tensor k = a + 2; + // { + // c10::InferenceMode guard(true); + // k.add_(2); + // } + // `k.add_(2)` still need to go through ADInplaceOrView kernel so that it's + // prepared for future autograd. + // + // 3. Why does setting InferenceMode also set GradMode? + // + // This is required since InferenceMode is a faster and more restrictive + // version of NoGradGuard. All runtime checks using GradMode::is_enabled() + // are applicable to InferenceMode as well, e.g. + // `tensorTypeInCurrentExecutionContext` in interpreter.cpp. + InferenceMode(bool enabled = true) + : prev_mode(AutogradState::get_tls_state()), + prev_keyset(c10::impl::tls_local_dispatch_key_set()) { + // Enabling inference mode means disabling grad modes + // And disabling inference mode means enabling grad modes + AutogradState::set_tls_state(AutogradState( + /* grad_mode */ !enabled, + /* inference_mode */ enabled, + /* fw_grad_mode */ !enabled, + /* multithreading_enabled*/ !enabled)); + DispatchKeySet included = enabled + ? prev_keyset.included_.remove(c10::DispatchKey::ADInplaceOrView) + : prev_keyset.included_.add(c10::DispatchKey::ADInplaceOrView); + DispatchKeySet excluded = enabled + ? (prev_keyset.excluded_ | c10::autograd_dispatch_keyset) + : (prev_keyset.excluded_ - c10::autograd_dispatch_keyset); + c10::impl::PODLocalDispatchKeySet cur_keyset{}; + cur_keyset.set_included(included); + cur_keyset.set_excluded(excluded); + c10::impl::_force_tls_local_dispatch_key_set(cur_keyset); + } + + InferenceMode(const InferenceMode&) = delete; + InferenceMode(InferenceMode&&) = delete; + InferenceMode& operator=(const InferenceMode&) = delete; + InferenceMode& operator=(InferenceMode&&) = delete; + + ~InferenceMode() { + AutogradState::set_tls_state(prev_mode); + c10::impl::_force_tls_local_dispatch_key_set(prev_keyset); + } + static bool is_enabled(); + + private: + AutogradState prev_mode; + c10::impl::LocalDispatchKeySet prev_keyset; +}; +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/Layout.h b/backends/pytorch/include/c10/core/Layout.h new file mode 100644 index 000000000..194e1863c --- /dev/null +++ b/backends/pytorch/include/c10/core/Layout.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include + +namespace c10 { + +inline Layout layout_from_backend(Backend backend) { + C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum") + switch (backend) { + case Backend::SparseCPU: + case Backend::SparseCUDA: + case Backend::SparseMPS: + case Backend::SparseHIP: + case Backend::SparseVE: + case Backend::SparseXPU: + case Backend::SparsePrivateUse1: + return Layout::Sparse; + case Backend::MkldnnCPU: + return Layout::Mkldnn; + case Backend::SparseCsrCPU: + case Backend::SparseCsrCUDA: + case Backend::SparseCsrMPS: + case Backend::SparseCsrHIP: + case Backend::SparseCsrVE: + case Backend::SparseCsrXPU: + TORCH_CHECK( + false, + "Cannot map Backend SparseCsr(CPU|CUDA|HIP|VE|XPU|MPS) to a unique layout."); + default: + return Layout::Strided; + } + C10_DIAGNOSTIC_POP() +} + +inline std::ostream& operator<<(std::ostream& stream, at::Layout layout) { + switch (layout) { + case at::kStrided: + return stream << "Strided"; + case at::kSparse: + return stream << "Sparse"; + case at::kSparseCsr: + return stream << "SparseCsr"; + case at::kSparseCsc: + return stream << "SparseCsc"; + case at::kSparseBsr: + return stream << "SparseBsr"; + case at::kSparseBsc: + return stream << "SparseBsc"; + case at::kMkldnn: + return stream << "Mkldnn"; + case at::kJagged: + return stream << "Jagged"; + case Layout::NumOptions: + default: + TORCH_CHECK(false, "Unknown layout"); + } +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/MemoryFormat.h b/backends/pytorch/include/c10/core/MemoryFormat.h new file mode 100644 index 000000000..63cdb7579 --- /dev/null +++ b/backends/pytorch/include/c10/core/MemoryFormat.h @@ -0,0 +1,268 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include + +#include +#include + +namespace c10 { + +// If you are seeing this, it means that this call site was not checked if +// the memory format could be preserved, and it was switched to old default +// behaviour of contiguous +#define LEGACY_CONTIGUOUS_MEMORY_FORMAT c10::get_contiguous_memory_format() + +inline std::ostream& operator<<( + std::ostream& stream, + at::MemoryFormat memory_format) { + switch (memory_format) { + case MemoryFormat::Preserve: + return stream << "Preserve"; + case MemoryFormat::Contiguous: + return stream << "Contiguous"; + case MemoryFormat::ChannelsLast: + return stream << "ChannelsLast"; + case MemoryFormat::ChannelsLast3d: + return stream << "ChannelsLast3d"; + case MemoryFormat::NumOptions: + default: + TORCH_CHECK(false, "Unknown memory format ", memory_format); + } +} + +// Note: Hardcoded the channel last stride indices here to get better +// performance +template +inline std::vector get_channels_last_strides_2d(ArrayRef sizes) { + std::vector strides(sizes.size()); + switch (sizes.size()) { + case 4: + strides[1] = 1; + strides[3] = sizes[1]; + strides[2] = strides[3] * sizes[3]; + strides[0] = strides[2] * sizes[2]; + return strides; + case 3: + strides[0] = 1; + strides[2] = sizes[0]; + strides[1] = strides[2] * sizes[2]; + return strides; + default: + TORCH_INTERNAL_ASSERT( + false, "ChannelsLast2d doesn't support size ", sizes.size()); + } +} + +inline std::vector get_channels_last_strides_2d(IntArrayRef sizes) { + return get_channels_last_strides_2d(sizes); +} + +template +std::vector get_channels_last_strides_3d(ArrayRef sizes) { + std::vector strides(sizes.size()); + switch (sizes.size()) { + case 5: + strides[1] = 1; + strides[4] = sizes[1]; + strides[3] = strides[4] * sizes[4]; + strides[2] = strides[3] * sizes[3]; + strides[0] = strides[2] * sizes[2]; + return strides; + case 4: + strides[0] = 1; + strides[3] = sizes[0]; + strides[2] = strides[3] * sizes[3]; + strides[1] = strides[2] * sizes[2]; + return strides; + default: + TORCH_INTERNAL_ASSERT( + false, "ChannelsLast3d doesn't support size ", sizes.size()); + } +} + +inline std::vector get_channels_last_strides_3d(IntArrayRef sizes) { + return get_channels_last_strides_3d(sizes); +} + +// NOTE: +// Below are Helper functions for is_channels_last_strides_xd. +// 1. Please do not combine these helper functions, each helper function handles +// exactly one case of sizes + memory_format, by doing this, the strides indices +// will be a constant array and we can access it using constant index number, +// the compiler will fully unroll the loop on strides indices to gain a better +// performance. +// 2. No error check in helper function, caller ensures the correctness of the +// input +// 3. All helper functions have similar comments, only 1st helper function is +// commented here. +template +inline bool is_channels_last_strides_2d_s4( + const ArrayRef sizes, + const ArrayRef strides) { + T min = 0; + // special case for trivial C dimension. default to NCHW + if (strides[1] == 0) { + return false; + } + // loop strides indices + for (auto& d : {1, 3, 2, 0}) { + if (sizes[d] == 0) { + return false; + } + if (strides[d] < min) { + return false; + } + // Fallback to NCHW as default layout for ambiguous cases + // This is the flaw of implicit memory_format from strides. + // N111 tensor with identical strides for size 1 dimension; + // Two cases could lead us here: + // a. N111 contiguous Tensor ([N,1,1,1]@[1,1,1,1]) + // b. N11W contiguous Tensor sliced on the W-dimension. + // ([N,1,1,1]@[W,W,W,W]) + if (d == 0 && min == strides[1]) { + return false; + } + // This is necessary to: + // 1. distinguish the memory_format of N1H1; + // [H, 1, 1, 1] channels_last stride + // [H, H, 1, 1] contiguous stride + // 2. permutation of 1C1W: + // [1, C, 1, H]@[HC, H, H, 1] transpose(1, 3) + // [1, H, 1, C]@[HC, 1, H, H] shouldn't be identified as channels_last + min = strides[d]; + if (sizes[d] > 1) { + min *= sizes[d]; + } + } + return true; +} + +template +inline bool is_channels_last_strides_3d_s5( + const ArrayRef sizes, + const ArrayRef strides) { + T min = 0; + if (strides[1] == 0) { + return false; + } + for (auto& d : {1, 4, 3, 2, 0}) { + if (sizes[d] == 0) { + return false; + } + if (strides[d] < min) { + return false; + } + if (d == 0 && min == strides[1]) { + return false; + } + min = strides[d]; + if (sizes[d] > 1) { + min *= sizes[d]; + } + } + return true; +} + +// Note [Ambiguous is_channels_last_strides_xd] +// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +// The flaw of carrying memory_format implicitly through strides is very hard +// to WAR properly. issue #24090 +// Without the history of permutation, we can't infer the memory_format of a +// tensor from the snapshot of its size & stride +// e.g. +// +// 1. We can NOT specify the memory_format of N111 tensor through strides in a +// meaningful way; +// +// 2. Two path that ended up with identical size/stride +// N11W contiguous tensor sliced at w-dimension becomes [N,1,1,1]@[W,W,W,W] +// NC11 channels_last tensor sliced at c-dimension becomes [N,1,1,1]@[C,C,C,C] +// So if we see a tensor [N,1,1,1]@[X,X,X,X], there's no way for us to infer +// the memory_format of the original tensor. +// +// Due to the limitations, our temporary WAR `is_channels_last_strides` does the +// best effort to infer whether the original memory_format of a tensor is +// at::MemoryFormat::ChannelsLast. The two objectives of this function (ordered +// by their importance): +// 1. Ensure that normal shape manipulation does not accidentally change the +// MemoryFormat of an existing tensor. +// 2. Allows user to mark MemoryFormat::ChannelsLast to tensors; +// +// The function does so via checking strides of the tensor, including strides of +// size-1 dimensions. Although conventionally PyTorch implies no restriction on +// trivial stride (stride for size-1 dimension). +// +// Note that this approach is a compromise. We did not solve the problem +// completely. Many cases we will not be able to infer the correct memory +// format. +// The implementation of `is_channels_last_strides` is to serve the objectives: +// MemoryFormat::ChannelsLast has to be explicitly opted-in (no accidental +// conversion); Best effort to maintain the ChannelsLast flag. +// +// Due to the fact that this is not a bulletproof solution, through testing +// (aten/src/ATen/test/memory_format_test.cpp) +// a. we ensure that the common tasks are supported; +// a. we identify corner cases where the implementation compromises on. +// +// By the time accumulated permutation is enabled to replace implicit +// memory_format through strides, we should be updating our tests and fix the +// issues in our tests. +// +// We use Channels Last 2d as an example above. +// This is a general problem for all the is_channels_last_strides_xd +// implementation. Please check the helper functions +// (is_channels_last_strides_*d_s*) for more details. + +template +inline bool is_channels_last_strides_2d( + const ArrayRef sizes, + const ArrayRef strides) { + switch (sizes.size()) { + case 4: + return is_channels_last_strides_2d_s4(sizes, strides); + // NOLINTNEXTLINE(bugprone-branch-clone) + case 3: + // TODO dim == 3 case will be enabled once it is fully tested + return false; + default: + return false; + } +} + +template +inline bool is_channels_last_strides_3d( + const ArrayRef sizes, + const ArrayRef strides) { + switch (sizes.size()) { + case 5: + return is_channels_last_strides_3d_s5(sizes, strides); + // NOLINTNEXTLINE(bugprone-branch-clone) + case 4: + // TODO dim == 4 case will be enabled once it is fully tested + return false; + default: + return false; + } +} + +inline bool is_channels_last_strides_2d( + const IntArrayRef sizes, + const IntArrayRef strides) { + return is_channels_last_strides_2d(sizes, strides); +} + +inline bool is_channels_last_strides_3d( + const IntArrayRef sizes, + const IntArrayRef strides) { + return is_channels_last_strides_3d(sizes, strides); +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/OptionalRef.h b/backends/pytorch/include/c10/core/OptionalRef.h new file mode 100644 index 000000000..f1199e194 --- /dev/null +++ b/backends/pytorch/include/c10/core/OptionalRef.h @@ -0,0 +1,36 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +namespace c10 { + +template +class OptionalRef { + public: + OptionalRef() : data_(nullptr) {} + OptionalRef(const T* data) : data_(data) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(data_); + } + OptionalRef(const T& data) : data_(&data) {} + + bool has_value() const { + return data_ != nullptr; + } + + const T& get() const { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(data_); + return *data_; + } + + operator bool() const { + return has_value(); + } + + private: + const T* data_; +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/QScheme.h b/backends/pytorch/include/c10/core/QScheme.h new file mode 100644 index 000000000..f557affb1 --- /dev/null +++ b/backends/pytorch/include/c10/core/QScheme.h @@ -0,0 +1,60 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include + +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum") + +namespace c10 { + +/** + * QScheme is an enum that specifies the type of quantization. This has a one + * to one correspondence with Quantizer + * Please refer to ATen/quantized/Quantizer.h to see the Quantizers classes. + * Keep this file in sync with torch/nn/_qscheme.py + */ +enum class QScheme : uint8_t { + PER_TENSOR_AFFINE = 0, + PER_CHANNEL_AFFINE = 1, + PER_TENSOR_SYMMETRIC = 2, + PER_CHANNEL_SYMMETRIC = 3, + PER_CHANNEL_AFFINE_FLOAT_QPARAMS = 4, + COMPILE_TIME_NUM_QSCHEMES = 5, +}; + +constexpr auto kPerTensorAffine = QScheme::PER_TENSOR_AFFINE; +constexpr auto kPerChannelAffine = QScheme::PER_CHANNEL_AFFINE; +constexpr auto kPerTensorSymmetric = QScheme::PER_TENSOR_SYMMETRIC; +constexpr auto kPerChannelSymmetric = QScheme::PER_CHANNEL_SYMMETRIC; +constexpr auto kPerChannelAffineFloatQParams = + QScheme::PER_CHANNEL_AFFINE_FLOAT_QPARAMS; +constexpr int COMPILE_TIME_NUM_QSCHEMES = + static_cast(QScheme::COMPILE_TIME_NUM_QSCHEMES); + +inline std::string toString(QScheme qscheme) { + switch (qscheme) { + case kPerTensorAffine: + return "per_tensor_affine"; + case kPerChannelAffine: + return "per_channel_affine"; + case kPerTensorSymmetric: + return "per_tensor_symmetric"; + case kPerChannelSymmetric: + return "per_channel_symmetric"; + case kPerChannelAffineFloatQParams: + return "per_channel_affine_float_qparams"; + default: + TORCH_CHECK(false, "Unrecognized qscheme: ", static_cast(qscheme)); + } +} + +} // namespace c10 + +C10_DIAGNOSTIC_POP() + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/SafePyObject.h b/backends/pytorch/include/c10/core/SafePyObject.h new file mode 100644 index 000000000..bf8eee0e0 --- /dev/null +++ b/backends/pytorch/include/c10/core/SafePyObject.h @@ -0,0 +1,125 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include + +namespace c10 { + +// This is an safe owning holder for a PyObject, akin to pybind11's +// py::object, with two major differences: +// +// - It is in c10/core; i.e., you can use this type in contexts where +// you do not have a libpython dependency +// +// - It is multi-interpreter safe (ala torchdeploy); when you fetch +// the underlying PyObject* you are required to specify what the current +// interpreter context is and we will check that you match it. +// +// It is INVALID to store a reference to a Tensor object in this way; +// you should just use TensorImpl directly in that case! +struct C10_API SafePyObject { + // Steals a reference to data + SafePyObject(PyObject* data, c10::impl::PyInterpreter* pyinterpreter) + : data_(data), pyinterpreter_(pyinterpreter) {} + SafePyObject(SafePyObject&& other) noexcept + : data_(std::exchange(other.data_, nullptr)), + pyinterpreter_(other.pyinterpreter_) {} + // For now it's not used, so we just disallow it. + SafePyObject& operator=(SafePyObject&&) = delete; + + SafePyObject(SafePyObject const& other) + : data_(other.data_), pyinterpreter_(other.pyinterpreter_) { + if (data_ != nullptr) { + (*pyinterpreter_)->incref(data_); + } + } + + SafePyObject& operator=(SafePyObject const& other) { + if (this == &other) { + return *this; // Handle self-assignment + } + if (other.data_ != nullptr) { + (*other.pyinterpreter_)->incref(other.data_); + } + if (data_ != nullptr) { + (*pyinterpreter_)->decref(data_); + } + data_ = other.data_; + pyinterpreter_ = other.pyinterpreter_; + return *this; + } + + ~SafePyObject() { + if (data_ != nullptr) { + (*pyinterpreter_)->decref(data_); + } + } + + c10::impl::PyInterpreter& pyinterpreter() const { + return *pyinterpreter_; + } + PyObject* ptr(const c10::impl::PyInterpreter* /*interpreter*/) const; + + // stop tracking the current object, and return it + PyObject* release() { + auto rv = data_; + data_ = nullptr; + return rv; + } + + private: + PyObject* data_; + c10::impl::PyInterpreter* pyinterpreter_; +}; + +// A newtype wrapper around SafePyObject for type safety when a python object +// represents a specific type. Note that `T` is only used as a tag and isn't +// actually used for any true purpose. +template +struct SafePyObjectT : private SafePyObject { + SafePyObjectT(PyObject* data, c10::impl::PyInterpreter* pyinterpreter) + : SafePyObject(data, pyinterpreter) {} + ~SafePyObjectT() = default; + SafePyObjectT(SafePyObjectT&& other) noexcept : SafePyObject(other) {} + SafePyObjectT(SafePyObjectT const&) = delete; + SafePyObjectT& operator=(SafePyObjectT const&) = delete; + SafePyObjectT& operator=(SafePyObjectT&&) = delete; + + using SafePyObject::ptr; + using SafePyObject::pyinterpreter; + using SafePyObject::release; +}; + +// Like SafePyObject, but non-owning. Good for references to global PyObjects +// that will be leaked on interpreter exit. You get a copy constructor/assign +// this way. +struct C10_API SafePyHandle { + SafePyHandle() : data_(nullptr), pyinterpreter_(nullptr) {} + SafePyHandle(PyObject* data, c10::impl::PyInterpreter* pyinterpreter) + : data_(data), pyinterpreter_(pyinterpreter) {} + + c10::impl::PyInterpreter& pyinterpreter() const { + return *pyinterpreter_; + } + PyObject* ptr(const c10::impl::PyInterpreter* /*interpreter*/) const; + void reset() { + data_ = nullptr; + pyinterpreter_ = nullptr; + } + operator bool() { + return data_; + } + + private: + PyObject* data_; + c10::impl::PyInterpreter* pyinterpreter_; +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/Scalar.h b/backends/pytorch/include/c10/core/Scalar.h new file mode 100644 index 000000000..863a993ed --- /dev/null +++ b/backends/pytorch/include/c10/core/Scalar.h @@ -0,0 +1,471 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +/** + * Scalar represents a 0-dimensional tensor which contains a single element. + * Unlike a tensor, numeric literals (in C++) are implicitly convertible to + * Scalar (which is why, for example, we provide both add(Tensor) and + * add(Scalar) overloads for many operations). It may also be used in + * circumstances where you statically know a tensor is 0-dim and single size, + * but don't know its type. + */ +class C10_API Scalar { + public: + Scalar() : Scalar(int64_t(0)) {} + + void destroy() { + if (Tag::HAS_si == tag || Tag::HAS_sd == tag || Tag::HAS_sb == tag) { + raw::intrusive_ptr::decref(v.p); + v.p = nullptr; + } + } + + ~Scalar() { + destroy(); + } + +#define DEFINE_IMPLICIT_CTOR(type, name) \ + Scalar(type vv) : Scalar(vv, true) {} + + AT_FORALL_SCALAR_TYPES_AND3(Half, BFloat16, ComplexHalf, DEFINE_IMPLICIT_CTOR) + AT_FORALL_COMPLEX_TYPES(DEFINE_IMPLICIT_CTOR) + AT_FORALL_FLOAT8_TYPES(DEFINE_IMPLICIT_CTOR) + + // Helper constructors to allow Scalar creation from long and long long types + // As std::is_same_v is false(except Android), one needs to + // provide a constructor from either long or long long in addition to one from + // int64_t +#if defined(__APPLE__) || defined(__MACOSX) + static_assert( + std::is_same_v, + "int64_t is the same as long long on MacOS"); + Scalar(long vv) : Scalar(vv, true) {} +#endif +#if defined(_MSC_VER) + static_assert( + std::is_same_v, + "int64_t is the same as long long on Windows"); + Scalar(long vv) : Scalar(vv, true) {} +#endif +#if defined(__linux__) && !defined(__ANDROID__) + static_assert( + sizeof(void*) != 8 || std::is_same_v, + "int64_t is the same as long on 64 bit Linux"); +#if LONG_MAX != INT_MAX + Scalar(long long vv) : Scalar(vv, true) {} +#endif /* not 32-bit system */ +#endif + + Scalar(uint16_t vv) : Scalar(vv, true) {} + Scalar(uint32_t vv) : Scalar(vv, true) {} + Scalar(uint64_t vv) { + if (vv > static_cast(INT64_MAX)) { + tag = Tag::HAS_u; + v.u = vv; + } else { + tag = Tag::HAS_i; + // NB: no need to use convert, we've already tested convertibility + v.i = static_cast(vv); + } + } + +#undef DEFINE_IMPLICIT_CTOR + + // Value* is both implicitly convertible to SymbolicVariable and bool which + // causes ambiguity error. Specialized constructor for bool resolves this + // problem. + template < + typename T, + typename std::enable_if_t, bool>* = nullptr> + Scalar(T vv) : tag(Tag::HAS_b) { + v.i = convert(vv); + } + + template < + typename T, + typename std::enable_if_t, bool>* = + nullptr> + Scalar(T vv) : tag(Tag::HAS_sb) { + v.i = convert(vv); + } + +#define DEFINE_ACCESSOR(type, name) \ + type to##name() const { \ + if (Tag::HAS_d == tag) { \ + return checked_convert(v.d, #type); \ + } else if (Tag::HAS_z == tag) { \ + return checked_convert>(v.z, #type); \ + } else if (Tag::HAS_sd == tag) { \ + return checked_convert( \ + toSymFloat().guard_float(__FILE__, __LINE__), #type); \ + } \ + if (Tag::HAS_b == tag) { \ + return checked_convert(v.i, #type); \ + } else if (Tag::HAS_i == tag) { \ + return checked_convert(v.i, #type); \ + } else if (Tag::HAS_u == tag) { \ + return checked_convert(v.u, #type); \ + } else if (Tag::HAS_si == tag) { \ + return checked_convert( \ + toSymInt().guard_int(__FILE__, __LINE__), #type); \ + } else if (Tag::HAS_sb == tag) { \ + return checked_convert( \ + toSymBool().guard_bool(__FILE__, __LINE__), #type); \ + } \ + TORCH_CHECK(false) \ + } + + // TODO: Support ComplexHalf accessor + AT_FORALL_SCALAR_TYPES_WITH_COMPLEX(DEFINE_ACCESSOR) + DEFINE_ACCESSOR(uint16_t, UInt16) + DEFINE_ACCESSOR(uint32_t, UInt32) + DEFINE_ACCESSOR(uint64_t, UInt64) + +#undef DEFINE_ACCESSOR + + SymInt toSymInt() const { + if (Tag::HAS_si == tag) { + return c10::SymInt(intrusive_ptr::reclaim_copy( + static_cast(v.p))); + } else { + return toLong(); + } + } + + SymFloat toSymFloat() const { + if (Tag::HAS_sd == tag) { + return c10::SymFloat(intrusive_ptr::reclaim_copy( + static_cast(v.p))); + } else { + return toDouble(); + } + } + + SymBool toSymBool() const { + if (Tag::HAS_sb == tag) { + return c10::SymBool(intrusive_ptr::reclaim_copy( + static_cast(v.p))); + } else { + return toBool(); + } + } + + // also support scalar.to(); + // Deleted for unsupported types, but specialized below for supported types + template + T to() const = delete; + + // audit uses of data_ptr + const void* data_ptr() const { + TORCH_INTERNAL_ASSERT(!isSymbolic()); + return static_cast(&v); + } + + bool isFloatingPoint() const { + return Tag::HAS_d == tag || Tag::HAS_sd == tag; + } + + [[deprecated( + "isIntegral is deprecated. Please use the overload with 'includeBool' parameter instead.")]] bool + isIntegral() const { + return Tag::HAS_i == tag || Tag::HAS_si == tag || Tag::HAS_u == tag; + } + + bool isIntegral(bool includeBool) const { + return Tag::HAS_i == tag || Tag::HAS_si == tag || Tag::HAS_u == tag || + (includeBool && isBoolean()); + } + + // See Note [Meaning of HAS_u] + bool isUnsigned() const { + return Tag::HAS_u == tag || (Tag::HAS_i == tag && v.i >= 0); + } + + bool isComplex() const { + return Tag::HAS_z == tag; + } + bool isBoolean() const { + return Tag::HAS_b == tag || Tag::HAS_sb == tag; + } + + // you probably don't actually want these; they're mostly for testing + bool isSymInt() const { + return Tag::HAS_si == tag; + } + bool isSymFloat() const { + return Tag::HAS_sd == tag; + } + bool isSymBool() const { + return Tag::HAS_sb == tag; + } + + bool isSymbolic() const { + return Tag::HAS_si == tag || Tag::HAS_sd == tag || Tag::HAS_sb == tag; + } + + C10_ALWAYS_INLINE Scalar& operator=(Scalar&& other) noexcept { + if (&other == this) { + return *this; + } + + destroy(); + moveFrom(std::move(other)); + return *this; + } + + C10_ALWAYS_INLINE Scalar& operator=(const Scalar& other) { + if (&other == this) { + return *this; + } + + *this = Scalar(other); + return *this; + } + + Scalar operator-() const; + Scalar conj() const; + Scalar log() const; + + template < + typename T, + typename std::enable_if_t::value, int> = 0> + bool equal(T num) const { + if (isComplex()) { + TORCH_INTERNAL_ASSERT(!isSymbolic()); + auto val = v.z; + return (val.real() == num) && (val.imag() == T()); + } else if (isFloatingPoint()) { + return toDouble() == num; + } else if (tag == Tag::HAS_i) { + if (overflows(v.i, /* strict_unsigned */ true)) { + return false; + } else { + return static_cast(v.i) == num; + } + } else if (tag == Tag::HAS_u) { + if (overflows(v.u, /* strict_unsigned */ true)) { + return false; + } else { + return static_cast(v.u) == num; + } + } else if (tag == Tag::HAS_si) { + TORCH_INTERNAL_ASSERT(false, "NYI SymInt equality"); + } else if (isBoolean()) { + // boolean scalar does not equal to a non boolean value + TORCH_INTERNAL_ASSERT(!isSymbolic()); + return false; + } else { + TORCH_INTERNAL_ASSERT(false); + } + } + + template < + typename T, + typename std::enable_if_t::value, int> = 0> + bool equal(T num) const { + if (isComplex()) { + TORCH_INTERNAL_ASSERT(!isSymbolic()); + return v.z == num; + } else if (isFloatingPoint()) { + return (toDouble() == num.real()) && (num.imag() == T()); + } else if (tag == Tag::HAS_i) { + if (overflows(v.i, /* strict_unsigned */ true)) { + return false; + } else { + return static_cast(v.i) == num.real() && num.imag() == T(); + } + } else if (tag == Tag::HAS_u) { + if (overflows(v.u, /* strict_unsigned */ true)) { + return false; + } else { + return static_cast(v.u) == num.real() && num.imag() == T(); + } + } else if (tag == Tag::HAS_si) { + TORCH_INTERNAL_ASSERT(false, "NYI SymInt equality"); + } else if (isBoolean()) { + // boolean scalar does not equal to a non boolean value + TORCH_INTERNAL_ASSERT(!isSymbolic()); + return false; + } else { + TORCH_INTERNAL_ASSERT(false); + } + } + + bool equal(bool num) const { + if (isBoolean()) { + TORCH_INTERNAL_ASSERT(!isSymbolic()); + return static_cast(v.i) == num; + } else { + return false; + } + } + + ScalarType type() const { + if (isComplex()) { + return ScalarType::ComplexDouble; + } else if (isFloatingPoint()) { + return ScalarType::Double; + } else if (isIntegral(/*includeBool=*/false)) { + // Represent all integers as long, UNLESS it is unsigned and therefore + // unrepresentable as long + if (Tag::HAS_u == tag) { + return ScalarType::UInt64; + } + return ScalarType::Long; + } else if (isBoolean()) { + return ScalarType::Bool; + } else { + TORCH_CHECK(false, "Unknown scalar type."); + } + } + + Scalar(Scalar&& rhs) noexcept : tag(rhs.tag) { + moveFrom(std::move(rhs)); + } + + Scalar(const Scalar& rhs) : tag(rhs.tag), v(rhs.v) { + if (isSymbolic()) { + c10::raw::intrusive_ptr::incref(v.p); + } + } + + Scalar(c10::SymInt si) { + if (auto m = si.maybe_as_int()) { + tag = Tag::HAS_i; + v.i = *m; + } else { + tag = Tag::HAS_si; + v.p = std::move(si).release(); + } + } + + Scalar(c10::SymFloat sd) { + if (sd.is_symbolic()) { + tag = Tag::HAS_sd; + v.p = std::move(sd).release(); + } else { + tag = Tag::HAS_d; + v.d = sd.as_float_unchecked(); + } + } + + Scalar(c10::SymBool sb) { + if (auto m = sb.maybe_as_bool()) { + tag = Tag::HAS_b; + v.i = *m; + } else { + tag = Tag::HAS_sb; + v.p = std::move(sb).release(); + } + } + + // We can't set v in the initializer list using the + // syntax v{ .member = ... } because it doesn't work on MSVC + private: + enum class Tag { HAS_d, HAS_i, HAS_u, HAS_z, HAS_b, HAS_sd, HAS_si, HAS_sb }; + + // Note [Meaning of HAS_u] + // ~~~~~~~~~~~~~~~~~~~~~~~ + // HAS_u is a bit special. On its face, it just means that we + // are holding an unsigned integer. However, we generally don't + // distinguish between different bit sizes in Scalar (e.g., we represent + // float as double), instead, it represents a mathematical notion + // of some quantity (integral versus floating point). So actually, + // HAS_u is used solely to represent unsigned integers that could + // not be represented as a signed integer. That means only uint64_t + // potentially can get this tag; smaller types like uint8_t fits into a + // regular int and so for BC reasons we keep as an int. + + // NB: assumes that self has already been cleared + // NOLINTNEXTLINE(cppcoreguidelines-rvalue-reference-param-not-moved) + C10_ALWAYS_INLINE void moveFrom(Scalar&& rhs) noexcept { + v = rhs.v; + tag = rhs.tag; + if (rhs.tag == Tag::HAS_si || rhs.tag == Tag::HAS_sd || + rhs.tag == Tag::HAS_sb) { + // Move out of scalar + rhs.tag = Tag::HAS_i; + rhs.v.i = 0; + } + } + + Tag tag; + + union v_t { + double d{}; + int64_t i; + // See Note [Meaning of HAS_u] + uint64_t u; + c10::complex z; + c10::intrusive_ptr_target* p; + // NOLINTNEXTLINE(modernize-use-equals-default) + v_t() {} // default constructor + } v; + + template < + typename T, + typename std::enable_if_t< + std::is_integral_v && !std::is_same_v, + bool>* = nullptr> + Scalar(T vv, bool /*unused*/) : tag(Tag::HAS_i) { + v.i = convert(vv); + } + + template < + typename T, + typename std::enable_if_t< + !std::is_integral_v && !c10::is_complex::value, + bool>* = nullptr> + Scalar(T vv, bool /*unused*/) : tag(Tag::HAS_d) { + v.d = convert(vv); + } + + template < + typename T, + typename std::enable_if_t::value, bool>* = nullptr> + Scalar(T vv, bool /*unused*/) : tag(Tag::HAS_z) { + v.z = convert(vv); + } +}; + +using OptionalScalarRef = c10::OptionalRef; + +// define the scalar.to() specializations +#define DEFINE_TO(T, name) \ + template <> \ + inline T Scalar::to() const { \ + return to##name(); \ + } +AT_FORALL_SCALAR_TYPES_WITH_COMPLEX(DEFINE_TO) +DEFINE_TO(uint16_t, UInt16) +DEFINE_TO(uint32_t, UInt32) +DEFINE_TO(uint64_t, UInt64) +#undef DEFINE_TO + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/ScalarType.h b/backends/pytorch/include/c10/core/ScalarType.h new file mode 100644 index 000000000..b678a2263 --- /dev/null +++ b/backends/pytorch/include/c10/core/ScalarType.h @@ -0,0 +1,285 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include + +#include + +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum") +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default") + +namespace c10 { + +// See [dtype Macros note] in torch/headeronly/core/ScalarType.h +// regarding macros. + +#define DEFINE_CONSTANT(_, name) \ + constexpr ScalarType k##name = ScalarType::name; + +// NOLINTNEXTLINE(clang-diagnostic-unused-const-variable) +AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(DEFINE_CONSTANT) +#undef DEFINE_CONSTANT + +inline size_t elementSize(ScalarType t) { +#define CASE_ELEMENTSIZE_CASE(ctype, name) \ + case ScalarType::name: \ + return sizeof(ctype); + + switch (t) { + AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(CASE_ELEMENTSIZE_CASE) + default: + TORCH_CHECK(false, "Unknown ScalarType"); + } +#undef CASE_ELEMENTSIZE_CASE +} + +inline bool isIntegralType(ScalarType t, bool includeBool) { + bool isIntegral = + (t == ScalarType::Byte || t == ScalarType::Char || t == ScalarType::Int || + t == ScalarType::Long || t == ScalarType::Short || + t == ScalarType::UInt16 || t == ScalarType::UInt32 || + t == ScalarType::UInt64); + + return isIntegral || (includeBool && t == ScalarType::Bool); +} + +[[deprecated( + "isIntegralType is deprecated. Please use the overload with 'includeBool' parameter instead.")]] inline bool +isIntegralType(ScalarType t) { + return isIntegralType(t, /*includeBool=*/false); +} + +inline bool isFloat8Type(ScalarType t) { + return t == ScalarType::Float8_e5m2 || t == ScalarType::Float8_e5m2fnuz || + t == ScalarType::Float8_e4m3fn || t == ScalarType::Float8_e4m3fnuz || + t == ScalarType::Float8_e8m0fnu; +} + +inline bool isReducedFloatingType(ScalarType t) { + return t == ScalarType::Half || t == ScalarType::BFloat16 || + isFloat8Type(t) || t == ScalarType::Float4_e2m1fn_x2; +} + +inline bool isFloatingType(ScalarType t) { + return t == ScalarType::Double || t == ScalarType::Float || + isReducedFloatingType(t); +} + +inline bool isComplexType(ScalarType t) { + return ( + t == ScalarType::ComplexHalf || t == ScalarType::ComplexFloat || + t == ScalarType::ComplexDouble); +} + +inline bool isBitsType(ScalarType t) { + return t == ScalarType::Bits1x8 || t == ScalarType::Bits2x4 || + t == ScalarType::Bits4x2 || t == ScalarType::Bits8 || + t == ScalarType::Bits16; +} + +inline bool isBarebonesUnsignedType(ScalarType t) { + return t == ScalarType::UInt1 || t == ScalarType::UInt2 || + t == ScalarType::UInt3 || t == ScalarType::UInt4 || + t == ScalarType::UInt5 || t == ScalarType::UInt6 || + t == ScalarType::UInt7 || t == ScalarType::UInt16 || + t == ScalarType::UInt32 || t == ScalarType::UInt64; +} + +inline ScalarType toQIntType(ScalarType t) { + switch (t) { + case ScalarType::Byte: + return ScalarType::QUInt8; + case ScalarType::Char: + return ScalarType::QInt8; + case ScalarType::Int: + return ScalarType::QInt32; + default: + return t; + } +} + +inline bool isSignedType(ScalarType t) { +#define CASE_ISSIGNED(name) \ + case ScalarType::name: \ + return std::numeric_limits< \ + ::c10::impl::ScalarTypeToCPPTypeT>::is_signed; + + // TODO(#146647): If we expect to have numeric_limits for everything, + // let's just have a big macro for the whole thing. + // If we're hardcoding it, let's just use the macro and a "true"/"false" + // below? + switch (t) { + case ScalarType::QInt8: + case ScalarType::QUInt8: + case ScalarType::QInt32: + case ScalarType::QUInt4x2: + case ScalarType::QUInt2x4: + TORCH_CHECK(false, "isSignedType not supported for quantized types"); + case ScalarType::Bits1x8: + case ScalarType::Bits2x4: + case ScalarType::Bits4x2: + case ScalarType::Bits8: + case ScalarType::Bits16: + TORCH_CHECK(false, "Bits types are undefined"); + CASE_ISSIGNED(UInt16); + CASE_ISSIGNED(UInt32); + CASE_ISSIGNED(UInt64); + CASE_ISSIGNED(BFloat16); + CASE_ISSIGNED(Float8_e5m2); + CASE_ISSIGNED(Float8_e5m2fnuz); + CASE_ISSIGNED(Float8_e4m3fn); + CASE_ISSIGNED(Float8_e4m3fnuz); + CASE_ISSIGNED(Float8_e8m0fnu); + CASE_ISSIGNED(Byte); + CASE_ISSIGNED(Char); + CASE_ISSIGNED(Short); + CASE_ISSIGNED(Int); + CASE_ISSIGNED(Long); + CASE_ISSIGNED(Half); + CASE_ISSIGNED(Float); + CASE_ISSIGNED(Double); + CASE_ISSIGNED(ComplexHalf); + CASE_ISSIGNED(ComplexFloat); + CASE_ISSIGNED(ComplexDouble); + CASE_ISSIGNED(Bool); + case ScalarType::Int1: + case ScalarType::Int2: + case ScalarType::Int3: + case ScalarType::Int4: + case ScalarType::Int5: + case ScalarType::Int6: + case ScalarType::Int7: + case ScalarType::Float4_e2m1fn_x2: + return true; + case ScalarType::UInt1: + case ScalarType::UInt2: + case ScalarType::UInt3: + case ScalarType::UInt4: + case ScalarType::UInt5: + case ScalarType::UInt6: + case ScalarType::UInt7: + return false; + case ScalarType::Undefined: + case ScalarType::NumOptions: + break; + // Do not add default here, but rather define behavior of every new entry + // here. `-Wswitch-enum` would raise a warning in those cases. + // TODO: get PyTorch to adopt exhaustive switches by default with a way to + // opt specific switches to being non-exhaustive. + // Exhaustive: + // `-Wswitch-enum`, `-Wswitch-default`, `-Wno-covered-switch-default` + // Non-Exhaustive: + // `-Wno-switch-enum`, `-Wswitch-default`, `-Wcovered-switch-default` + } + TORCH_CHECK(false, "Unknown ScalarType ", t); +#undef CASE_ISSIGNED +} + +inline bool isUnderlying(ScalarType type, ScalarType qtype) { + return type == toUnderlying(qtype); +} + +inline ScalarType toRealValueType(ScalarType t) { + switch (t) { + case ScalarType::ComplexHalf: + return ScalarType::Half; + case ScalarType::ComplexFloat: + return ScalarType::Float; + case ScalarType::ComplexDouble: + return ScalarType::Double; + default: + return t; + } +} + +inline ScalarType toComplexType(ScalarType t) { + switch (t) { + case ScalarType::BFloat16: + // BFloat16 has range equivalent to Float, + // so we map it to ComplexFloat. + return ScalarType::ComplexFloat; + case ScalarType::Half: + return ScalarType::ComplexHalf; + case ScalarType::Float: + return ScalarType::ComplexFloat; + case ScalarType::Double: + return ScalarType::ComplexDouble; + case ScalarType::ComplexHalf: + return ScalarType::ComplexHalf; + case ScalarType::ComplexFloat: + return ScalarType::ComplexFloat; + case ScalarType::ComplexDouble: + return ScalarType::ComplexDouble; + default: + TORCH_CHECK(false, "Unknown Complex ScalarType for ", t); + } +} + +// see tensor_attributes.rst for detailed explanation and examples +// of casting rules. +inline bool canCast(const ScalarType from, const ScalarType to) { + // We disallow complex -> non complex, e.g., float_tensor *= complex is + // disallowed. + if (isComplexType(from) && !isComplexType(to)) { + return false; + } + // We disallow float -> integral, e.g., int_tensor *= float is disallowed. + if (isFloatingType(from) && isIntegralType(to, false)) { + return false; + } + + // Treat bool as a distinct "category," to be consistent with type promotion + // rules (e.g. `bool_tensor + 5 -> int64_tensor`). If `5` was in the same + // category as `bool_tensor`, we would not promote. Differing categories + // implies `bool_tensor += 5` is disallowed. + // + // NB: numpy distinguishes "unsigned" as a category to get the desired + // `bool_tensor + 5 -> int64_tensor` behavior. We don't, because: + // * We don't want the performance hit of checking the runtime sign of + // Scalars. + // * `uint8_tensor + 5 -> int64_tensor` would be undesirable. + if (from != ScalarType::Bool && to == ScalarType::Bool) { + return false; + } + return true; +} + +C10_API ScalarType promoteTypes(ScalarType a, ScalarType b); + +// Returns a pair of strings representing the names for each dtype. +// The returned pair is (name, legacy_name_if_applicable) +C10_API std::pair getDtypeNames( + c10::ScalarType scalarType); + +// Returns a map of string name to dtype. +C10_API const std::unordered_map& getStringToDtypeMap(); + +} // namespace c10 + +C10_DIAGNOSTIC_POP() + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/ScalarTypeToTypeMeta.h b/backends/pytorch/include/c10/core/ScalarTypeToTypeMeta.h new file mode 100644 index 000000000..d952b0dd2 --- /dev/null +++ b/backends/pytorch/include/c10/core/ScalarTypeToTypeMeta.h @@ -0,0 +1,62 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +// these just expose TypeMeta/ScalarType bridge functions in c10 +// TODO move to typeid.h (or codemod away) when TypeMeta et al +// are moved from caffe2 to c10 (see note at top of typeid.h) + +namespace c10 { + +/** + * convert ScalarType enum values to TypeMeta handles + */ +inline caffe2::TypeMeta scalarTypeToTypeMeta(ScalarType scalar_type) { + return caffe2::TypeMeta::fromScalarType(scalar_type); +} + +/** + * convert TypeMeta handles to ScalarType enum values + */ +inline ScalarType typeMetaToScalarType(caffe2::TypeMeta dtype) { + return dtype.toScalarType(); +} + +/** + * typeMetaToScalarType(), lifted to optional + */ +inline std::optional optTypeMetaToScalarType( + std::optional type_meta) { + if (!type_meta.has_value()) { + return std::nullopt; + } + return type_meta->toScalarType(); +} + +/** + * convenience: equality across TypeMeta/ScalarType conversion + */ +inline bool operator==(ScalarType t, caffe2::TypeMeta m) { + return m.isScalarType(t); +} + +inline bool operator==(caffe2::TypeMeta m, ScalarType t) { + return t == m; +} + +inline bool operator!=(ScalarType t, caffe2::TypeMeta m) { + return !(t == m); +} + +inline bool operator!=(caffe2::TypeMeta m, ScalarType t) { + return !(t == m); +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/Storage.h b/backends/pytorch/include/c10/core/Storage.h new file mode 100644 index 000000000..203eec24c --- /dev/null +++ b/backends/pytorch/include/c10/core/Storage.h @@ -0,0 +1,293 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +struct Storage; + +C10_API bool isSharedStorageAlias( + const Storage& storage0, + const Storage& storage1); + +struct C10_API Storage { + public: + struct use_byte_size_t {}; + struct unsafe_borrow_t { + explicit unsafe_borrow_t() = default; + }; + + Storage() = default; + Storage(c10::intrusive_ptr ptr) + : storage_impl_(std::move(ptr)) {} + + // Allocates memory buffer using given allocator and creates a storage with it + Storage( + use_byte_size_t /*use_byte_size*/, + const SymInt& size_bytes, + Allocator* allocator = nullptr, + bool resizable = false) + : storage_impl_(c10::make_intrusive( + StorageImpl::use_byte_size_t(), + size_bytes, + allocator, + resizable)) {} + + // Creates storage with pre-allocated memory buffer. Allocator is given for + // potential future reallocations, however it can be nullptr if the storage + // is non-resizable + Storage( + use_byte_size_t /*use_byte_size*/, + size_t size_bytes, + at::DataPtr data_ptr, + at::Allocator* allocator = nullptr, + bool resizable = false) + : storage_impl_(c10::make_intrusive( + StorageImpl::use_byte_size_t(), + size_bytes, + std::move(data_ptr), + allocator, + resizable)) {} + + // Creates storage with pre-allocated memory buffer. Allocator is given for + // potential future reallocations, however it can be nullptr if the storage + // is non-resizable + Storage( + use_byte_size_t /*use_byte_size*/, + SymInt size_bytes, + at::DataPtr data_ptr, + at::Allocator* allocator = nullptr, + bool resizable = false) + : storage_impl_(c10::make_intrusive( + StorageImpl::use_byte_size_t(), + std::move(size_bytes), + std::move(data_ptr), + allocator, + resizable)) {} + + protected: + explicit Storage(unsafe_borrow_t /*unused*/, const Storage& rhs) + : storage_impl_(c10::intrusive_ptr::reclaim( + rhs.storage_impl_.get())) {} + + friend MaybeOwnedTraits; + + public: + // Legacy constructor for partially initialized (dtype or memory) storages + // that can be temporarily created with Caffe2 APIs. See the note on top of + // TensorImpl.h for details. + static Storage create_legacy(at::Device device) { + auto allocator = GetAllocator(device.type()); + return Storage(c10::make_intrusive( + StorageImpl::use_byte_size_t(), + 0, + allocator->allocate(0), // materialize a non-default Device. + allocator, + true)); + } + + // Mimic create_legacy, but without requiring a newly-created StorageImpl. + void reset_legacy() { + TORCH_CHECK(resizable() && allocator()); + set_nbytes(0); + set_data_ptr_noswap(allocator()->allocate(0)); + } + + // TODO: remove later + void set_nbytes(size_t size_bytes) const { + storage_impl_->set_nbytes(size_bytes); + } + + void set_nbytes(c10::SymInt size_bytes) const { + storage_impl_->set_nbytes(std::move(size_bytes)); + } + + bool resizable() const { + return storage_impl_->resizable(); + } + + size_t nbytes() const { + return storage_impl_->nbytes(); + } + + SymInt sym_nbytes() const { + return storage_impl_->sym_nbytes(); + } + // get() use here is to get const-correctness + + const void* data() const { + return storage_impl_->data(); + } + + void* mutable_data() const { + return storage_impl_->mutable_data(); + } + + at::DataPtr& mutable_data_ptr() const { + return storage_impl_->mutable_data_ptr(); + } + + const at::DataPtr& data_ptr() const { + return storage_impl_->data_ptr(); + } + + // Returns the previous data_ptr + at::DataPtr set_data_ptr(at::DataPtr&& data_ptr) const { + return storage_impl_->set_data_ptr(std::move(data_ptr)); + } + + void set_data_ptr_noswap(at::DataPtr&& data_ptr) const { + storage_impl_->set_data_ptr_noswap(std::move(data_ptr)); + } + + DeviceType device_type() const { + return storage_impl_->device_type(); + } + + at::Allocator* allocator() const { + return storage_impl_->allocator(); + } + + at::Device device() const { + return storage_impl_->device(); + } + + StorageImpl* unsafeReleaseStorageImpl() { + return storage_impl_.release(); + } + + StorageImpl* unsafeGetStorageImpl() const noexcept { + return storage_impl_.get(); + } + + c10::weak_intrusive_ptr getWeakStorageImpl() const { + return c10::weak_intrusive_ptr(storage_impl_); + } + + operator bool() const { + return storage_impl_; + } + + size_t use_count() const { + return storage_impl_.use_count(); + } + + inline bool unique() const { + return storage_impl_.unique(); + } + + bool is_alias_of(const Storage& other) const { + return ( + storage_impl_ == other.storage_impl_ || + isSharedStorageAlias(*this, other)); + } + + void UniqueStorageShareExternalPointer( + void* src, + size_t capacity, + DeleterFnPtr d = nullptr) { + if (!storage_impl_.unique()) { + TORCH_CHECK( + false, + "UniqueStorageShareExternalPointer can only be called when use_count == 1"); + } + storage_impl_->UniqueStorageShareExternalPointer(src, capacity, d); + } + + void UniqueStorageShareExternalPointer( + at::DataPtr&& data_ptr, + size_t capacity) { + if (!storage_impl_.unique()) { + TORCH_CHECK( + false, + "UniqueStorageShareExternalPointer can only be called when use_count == 1"); + } + storage_impl_->UniqueStorageShareExternalPointer( + std::move(data_ptr), capacity); + } + + protected: + c10::intrusive_ptr storage_impl_; +}; + +template <> +struct MaybeOwnedTraits { + using owned_type = c10::Storage; + using borrow_type = c10::Storage; + + static borrow_type createBorrow(const owned_type& from) { + return borrow_type(borrow_type::unsafe_borrow_t{}, from); + } + + static void assignBorrow(borrow_type& lhs, const borrow_type& rhs) { + lhs.unsafeReleaseStorageImpl(); + lhs = borrow_type(borrow_type::unsafe_borrow_t{}, rhs); + } + + static void destroyBorrow(borrow_type& toDestroy) { + toDestroy.unsafeReleaseStorageImpl(); // "leak" it, but it was already +0. + } + + static const owned_type& referenceFromBorrow(const borrow_type& borrow) { + return borrow; + } + + static const owned_type* pointerFromBorrow(const borrow_type& borrow) { + return &borrow; + } + + static bool debugBorrowIsValid(const borrow_type& /*borrow*/) { + return true; + } +}; + +template <> +struct ExclusivelyOwnedTraits { + using repr_type = c10::Storage; + using pointer_type = c10::Storage*; + using const_pointer_type = const c10::Storage*; + + static repr_type nullRepr() { + return c10::Storage(); + } + + template + static repr_type createInPlace(Args&&... args) { + return c10::Storage(std::forward(args)...); + } + + static repr_type moveToRepr(c10::Storage&& x) { + return std::move(x); + } + + static c10::Storage take(c10::Storage& x) { + return std::move(x); + } + + static pointer_type getImpl(repr_type& x) { + return &x; + } + + static const_pointer_type getImpl(const repr_type& x) { + return &x; + } +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/StorageImpl.h b/backends/pytorch/include/c10/core/StorageImpl.h new file mode 100644 index 000000000..2acfa4077 --- /dev/null +++ b/backends/pytorch/include/c10/core/StorageImpl.h @@ -0,0 +1,398 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +[[noreturn]] C10_API void throwNullDataPtrError(); +C10_API void warnDeprecatedDataPtr(); + +// Used in StorageImpl to store extra metadata. +// Currently used only for storing a custom error message +// used when throwing an exception when data_ptr is accessed. +struct C10_API StorageExtraMeta { + std::optional custom_data_ptr_error_msg_ = std::nullopt; +}; + +// A storage represents the underlying backing data buffer for a +// tensor. This concept was inherited from the original Torch7 +// codebase; we'd kind of like to get rid of the concept +// (see https://github.com/pytorch/pytorch/issues/14797) but +// it's hard work and no one has gotten around to doing it. +// +// NB: storage is supposed to uniquely own a data pointer; e.g., +// two non-null data pointers alias if and only if they are from +// the same storage. Technically you can violate this invariant +// (e.g., you can create a non-owning StorageImpl with at::from_blob) +// but a lot of things won't work correctly, including: +// +// - An ordinary deleter on such a storage is wrong, because normal deleters +// assume unique ownership, but if you have two storages at the same data, +// that implies there is some sort of shared ownership. So your deleter would +// have to actually be internally doing some sort of refcount thing +// - Deepcopy in Python side relies on storage equality and not data pointer +// equality; so if there are two separate storages pointing to the same data, +// the data will actually get duplicated in that case (one data ptr before, +// two data ptrs after) +// - Version counts won't work correctly, because we do all VC tracking at the +// level of storages (unless you explicitly disconnect the VC with detach); +// mutation because data pointers are the same are totally untracked +struct C10_API StorageImpl : public c10::intrusive_ptr_target { + public: + struct use_byte_size_t {}; + + StorageImpl( + use_byte_size_t /*use_byte_size*/, + SymInt size_bytes, + at::DataPtr data_ptr, + at::Allocator* allocator, + bool resizable) + : data_ptr_(std::move(data_ptr)), + size_bytes_(std::move(size_bytes)), + size_bytes_is_heap_allocated_(size_bytes_.is_heap_allocated()), + resizable_(resizable), + received_cuda_(false), + allocator_(allocator) { + if (resizable) { + TORCH_INTERNAL_ASSERT( + allocator_, "For resizable storage, allocator must be provided"); + } + refresh_has_data_ptr_check(); + } + + StorageImpl( + use_byte_size_t /*use_byte_size*/, + const SymInt& size_bytes, + at::Allocator* allocator, + bool resizable) + : StorageImpl( + use_byte_size_t(), + size_bytes, + size_bytes.is_heap_allocated() + ? allocator->allocate(0) + : allocator->allocate(size_bytes.as_int_unchecked()), + allocator, + resizable) {} + + StorageImpl& operator=(StorageImpl&& other) = delete; + StorageImpl& operator=(const StorageImpl&) = delete; + StorageImpl() = delete; + StorageImpl(StorageImpl&& other) = delete; + StorageImpl(const StorageImpl&) = delete; + ~StorageImpl() override = default; + + void reset() { + data_ptr_.clear(); + size_bytes_ = 0; + size_bytes_is_heap_allocated_ = false; + } + + // Destructor doesn't call release_resources because it's + // unnecessary; don't forget to change that if needed! + void release_resources() override { + data_ptr_.clear(); + } + + void incref_pyobject() const noexcept override final; + + void decref_pyobject() const noexcept override final; + + bool try_incref_pyobject() const noexcept override final; + + size_t nbytes() const { + // OK to do this instead of maybe_as_int as nbytes is guaranteed positive + TORCH_CHECK(!size_bytes_is_heap_allocated_); + return size_bytes_.as_int_unchecked(); + } + + SymInt sym_nbytes() const { + return size_bytes_; + } + + // TODO: remove later + void set_nbytes(size_t size_bytes) { + size_bytes_ = static_cast(size_bytes); + size_bytes_is_heap_allocated_ = false; + } + + void unsafe_set_nbytes(size_t size_bytes) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!size_bytes_is_heap_allocated_); + size_bytes_.unsafe_set_data(size_bytes); + } + + void set_nbytes(c10::SymInt size_bytes) { + size_bytes_ = std::move(size_bytes); + } + + bool resizable() const { + return resizable_; + } + + const at::DataPtr& data_ptr() const { + if (C10_UNLIKELY(throw_on_immutable_data_ptr_)) { + throw_data_ptr_access_error(); + } + return data_ptr_; + } + + at::DataPtr& mutable_data_ptr() { + if (C10_UNLIKELY(has_mutable_data_ptr_check_)) { + if (throw_on_immutable_data_ptr_) { + throw_data_ptr_access_error(); + } + if (throw_on_mutable_data_ptr_) { + throwNullDataPtrError(); + } + if (warn_deprecated_on_mutable_data_ptr_) { + warnDeprecatedDataPtr(); + } + maybe_materialize_cow(); + } + return data_ptr_; + } + + // Returns the data_ptr. Bypasses all checks. + at::DataPtr& _mutable_data_ptr_no_checks() { + return data_ptr_; + } + + // Returns the previous data_ptr + at::DataPtr set_data_ptr(at::DataPtr&& data_ptr) { + // We need to materialize the old COW DataPtr because it is + // being returned as mutable. + maybe_materialize_cow(); + return set_data_ptr_no_materialize_cow(std::move(data_ptr)); + } + + void set_data_ptr_noswap(at::DataPtr&& data_ptr) { + data_ptr_ = std::move(data_ptr); + refresh_has_data_ptr_check(); + } + + const void* data() const { + if (C10_UNLIKELY(throw_on_immutable_data_ptr_)) { + throw_data_ptr_access_error(); + } + return data_ptr_.get(); + } + + void* mutable_data() { + if (C10_UNLIKELY(has_mutable_data_ptr_check_)) { + if (throw_on_immutable_data_ptr_) { + throw_data_ptr_access_error(); + } + if (throw_on_mutable_data_ptr_) { + throwNullDataPtrError(); + } + if (warn_deprecated_on_mutable_data_ptr_) { + warnDeprecatedDataPtr(); + } + maybe_materialize_cow(); + } + return data_ptr_.mutable_get(); + } + + at::DeviceType device_type() const { + return data_ptr_.device().type(); + } + + at::Allocator* allocator() { + return allocator_; + } + + const at::Allocator* allocator() const { + return allocator_; + } + + // You generally shouldn't use this method, but it is occasionally + // useful if you want to override how a tensor will be reallocated, + // after it was already allocated (and its initial allocator was + // set) + void set_allocator(at::Allocator* allocator) { + allocator_ = allocator; + } + + Device device() const { + return data_ptr_.device(); + } + + void set_resizable(bool resizable) { + if (resizable) { + // We need an allocator to be resizable + AT_ASSERT(allocator_); + } + resizable_ = resizable; + } + + /** + * Can only be called when use_count is 1 + */ + void UniqueStorageShareExternalPointer( + void* src, + size_t size_bytes, + DeleterFnPtr d = nullptr) { + UniqueStorageShareExternalPointer( + at::DataPtr(src, src, d, data_ptr_.device()), size_bytes); + } + + /** + * Can only be called when use_count is 1 + */ + void UniqueStorageShareExternalPointer( + at::DataPtr&& data_ptr, + size_t size_bytes) { + data_ptr_ = std::move(data_ptr); + size_bytes_ = static_cast(size_bytes); + size_bytes_is_heap_allocated_ = false; + allocator_ = nullptr; + resizable_ = false; + } + + // This method can be used only after storage construction and cannot be used + // to modify storage status + void set_received_cuda(bool received_cuda) { + received_cuda_ = received_cuda; + } + + bool received_cuda() { + return received_cuda_; + } + + impl::PyObjectSlot* pyobj_slot() { + return &pyobj_slot_; + } + + const impl::PyObjectSlot* pyobj_slot() const { + return &pyobj_slot_; + } + + StorageExtraMeta& get_extra_meta() { + if (!extra_meta_) { + extra_meta_ = std::make_unique(); + } + return *extra_meta_; + } + + [[noreturn]] void throw_data_ptr_access_error() const; + + void release_data_and_set_meta_custom_data_ptr_error_msg_( + std::optional s) { + throw_on_immutable_data_ptr_ = true; + get_extra_meta().custom_data_ptr_error_msg_ = std::move(s); + refresh_has_data_ptr_check(); + } + + void set_throw_on_mutable_data_ptr() { + throw_on_mutable_data_ptr_ = true; + refresh_has_data_ptr_check(); + } + + void set_warn_deprecated_on_mutable_data_ptr() { + warn_deprecated_on_mutable_data_ptr_ = true; + refresh_has_data_ptr_check(); + } + + protected: + // materialize_cow_storage needs to call set_data_ptr_no_materlize_cow + friend void c10::impl::cow::materialize_cow_storage(StorageImpl& storage); + + // Returns the previous data_ptr. If the old data_ptr was COW, + // this avoids materializing it + at::DataPtr set_data_ptr_no_materialize_cow(at::DataPtr&& data_ptr) { + at::DataPtr old_data_ptr(std::move(data_ptr_)); + data_ptr_ = std::move(data_ptr); + refresh_has_data_ptr_check(); + return old_data_ptr; + } + + private: + void refresh_has_data_ptr_check() { + has_mutable_data_ptr_check_ = is_cow() || throw_on_mutable_data_ptr_ || + warn_deprecated_on_mutable_data_ptr_ || throw_on_immutable_data_ptr_; + } + + inline bool is_cow() const { + return c10::impl::cow::is_cow_data_ptr(data_ptr_); + } + + // Triggers a copy if this is a copy-on-write tensor. + void maybe_materialize_cow() { + if (is_cow()) { + impl::cow::materialize_cow_storage(*this); + } + } + + DataPtr data_ptr_; + SymInt size_bytes_; + bool size_bytes_is_heap_allocated_; + bool resizable_; + // Identifies that Storage was received from another process and doesn't have + // local to process cuda memory allocation + bool received_cuda_; + // All special checks in data/data_ptr calls are guarded behind this single + // boolean. This is for performance: .data/.data_ptr calls are commonly in the + // hot-path. + bool has_mutable_data_ptr_check_ = false; + // If we should throw when mutable_data_ptr() or mutable_data() is called. + bool throw_on_mutable_data_ptr_ = false; + // If we should throw when data_ptr() or data() is called. + bool throw_on_immutable_data_ptr_ = false; + // If we warn when mutable_data_ptr() or mutable_data() is called. + bool warn_deprecated_on_mutable_data_ptr_ = false; + Allocator* allocator_; + impl::PyObjectSlot pyobj_slot_; + std::unique_ptr extra_meta_ = nullptr; +}; + +// Declare StorageImpl create function pointer types. +using StorageImplCreateHelper = intrusive_ptr (*)( + StorageImpl::use_byte_size_t, + SymInt size_bytes, + DataPtr data_ptr, + Allocator* allocator, + bool resizable); + +C10_API void SetStorageImplCreate(DeviceType t, StorageImplCreateHelper fptr); + +C10_API StorageImplCreateHelper GetStorageImplCreate(DeviceType t); + +C10_API c10::intrusive_ptr make_storage_impl( + c10::StorageImpl::use_byte_size_t use_byte_size, + c10::SymInt size_bytes, + c10::DataPtr data_ptr, + c10::Allocator* allocator, + bool resizable, + std::optional device_opt); + +namespace detail { + +#ifndef C10_MOBILE +template +struct TargetTraits< + T, + std::enable_if_t< + std::is_base_of_v>>> { + static constexpr bool can_have_pyobject = true; +}; +#endif + +} // namespace detail + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/Stream.h b/backends/pytorch/include/c10/core/Stream.h new file mode 100644 index 000000000..4d3a50984 --- /dev/null +++ b/backends/pytorch/include/c10/core/Stream.h @@ -0,0 +1,182 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +/// An index representing a specific stream. A StreamId is not independently +/// meaningful without knowing the Device it is associated with; try to +/// use Stream rather than StreamId directly. +/// +/// StreamIds are opaque; they are assigned by some DeviceType-specific +/// numbering system which is not visible to the user. HOWEVER, we +/// guarantee that StreamId 0 is always a valid stream, and corresponds +/// to some sort of "default" stream. +using StreamId = int64_t; + +struct C10_API StreamData3 { + StreamId stream_id; + DeviceIndex device_index; + DeviceType device_type; +}; + +// NB: I decided not to call the above StreamIndex to avoid confusion with +// DeviceIndex. This way, you access device index with index(), and stream id +// with id() + +/** + * A stream is a software mechanism used to synchronize launched kernels + * without requiring explicit synchronizations between kernels. The basic + * model is that every kernel launch is associated with a stream: every + * kernel on the same stream is implicitly synchronized so that if I launch + * kernels A and B on the same stream, A is guaranteed to finish before B + * launches. If I want B to run concurrently with A, I must schedule + * it on a different stream. + * + * The Stream class is a backend agnostic value class representing a stream + * which I may schedule a kernel on. Every stream is associated with a device, + * which is recorded in stream, which is used to avoid confusion about which + * device a stream refers to. + * + * Streams are explicitly thread-safe, in the sense that it is OK to pass + * a Stream from one thread to another, and kernels queued from two different + * threads will still get serialized appropriately. (Of course, the + * time when the kernels get queued is undetermined unless you synchronize + * host side ;) + * + * Stream does NOT have a default constructor. Streams are for expert + * users; if you want to use Streams, we're going to assume you know + * how to deal with C++ template error messages if you try to + * resize() a vector of Streams. + * + * Known instances of streams in backends: + * + * - cudaStream_t (CUDA) + * - hipStream_t (HIP) + * - cl_command_queue (OpenCL) (NB: Caffe2's existing OpenCL integration + * does NOT support command queues.) + * + * Because this class is device agnostic, it cannot provide backend-specific + * functionality (e.g., get the cudaStream_t of a CUDA stream.) There are + * wrapper classes which provide this functionality, e.g., CUDAStream. + */ +class C10_API Stream final { + private: + Device device_; + StreamId id_; + + public: + enum Unsafe { UNSAFE }; + enum Default { DEFAULT }; + + /// Unsafely construct a stream from a Device and a StreamId. In + /// general, only specific implementations of streams for a + /// backend should manufacture Stream directly in this way; other users + /// should use the provided APIs to get a stream. In particular, + /// we don't require backends to give any guarantees about non-zero + /// StreamIds; they are welcome to allocate in whatever way they like. + explicit Stream(Unsafe /*unused*/, Device device, StreamId id) + : device_(device), id_(id) {} + + /// Construct the default stream of a Device. The default stream is + /// NOT the same as the current stream; default stream is a fixed stream + /// that never changes, whereas the current stream may be changed by + /// StreamGuard. + explicit Stream(Default /*unused*/, Device device) + : device_(device), id_(0) {} + + bool operator==(const Stream& other) const noexcept { + return this->device_ == other.device_ && this->id_ == other.id_; + } + bool operator!=(const Stream& other) const noexcept { + return !(*this == other); + } + + Device device() const noexcept { + return device_; + } + DeviceType device_type() const noexcept { + return device_.type(); + } + DeviceIndex device_index() const noexcept { + return device_.index(); + } + StreamId id() const noexcept { + return id_; + } + + // Enqueues a wait instruction in the stream's work queue. + // This instruction is a no-op unless the event is marked + // for recording. In that case the stream stops processing + // until the event is recorded. + template + void wait(const T& event) const { + event.block(*this); + } + + // Return whether all asynchronous work previously enqueued on this stream + // has completed running on the device. + bool query() const; + + // Wait (by blocking the calling thread) until all asynchronous work enqueued + // on this stream has completed running on the device. + void synchronize() const; + + // The purpose of this function is to more conveniently permit binding + // of Stream to and from Python. Without packing, I have to setup a whole + // class with two fields (device and stream id); with packing I can just + // store a single uint64_t. + // + // The particular way we pack streams into a uint64_t is considered an + // implementation detail and should not be relied upon. + uint64_t hash() const noexcept { + // Concat these together into a 64-bit integer + uint64_t bits = static_cast(device_type()) << 56 | + static_cast(device_index()) << 48 | + // Remove the sign extension part of the 64-bit address because + // the id might be used to hold a pointer. + (static_cast(id()) & ((1ull << 48) - 1)); + return bits; + } + + struct StreamData3 pack3() const { + return {id(), device_index(), device_type()}; + } + + static Stream unpack3( + StreamId stream_id, + DeviceIndex device_index, + DeviceType device_type) { + TORCH_CHECK(isValidDeviceType(device_type)); + return Stream(UNSAFE, Device(device_type, device_index), stream_id); + } + + // I decided NOT to provide setters on this class, because really, + // why would you change the device of a stream? Just construct + // it correctly from the beginning dude. +}; + +C10_API std::ostream& operator<<(std::ostream& stream, const Stream& s); + +} // namespace c10 + +namespace std { +template <> +struct hash { + size_t operator()(c10::Stream s) const noexcept { + return std::hash{}(s.hash()); + } +}; +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/StreamGuard.h b/backends/pytorch/include/c10/core/StreamGuard.h new file mode 100644 index 000000000..003816d62 --- /dev/null +++ b/backends/pytorch/include/c10/core/StreamGuard.h @@ -0,0 +1,178 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include + +namespace c10 { + +/** + * A StreamGuard is an RAII class that changes the current device + * to the device corresponding to some stream, and changes the + * default stream on that device to be this stream. + * + * Use of StreamGuard is HIGHLY discouraged in operator definitions. In + * a single operator, you probably don't know enough about the global + * state of the world to profitably decide how to set streams. Let + * the caller handle this appropriately, and just use the current stream + * in your operator code. + * + * This StreamGuard does NOT have an uninitialized state; it is guaranteed + * to reset the stream and device on exit. If you are in a situation + * where you *might* want to setup a stream guard, see OptionalStreamGuard. + */ +struct StreamGuard { + /// No default constructor, see Note [Omitted default constructor from RAII] + explicit StreamGuard() = delete; + ~StreamGuard() = default; + + /// Set the current device to the device associated with the passed stream, + /// and set the current stream on that device to the passed stream. + explicit StreamGuard(Stream stream) : guard_(stream) {} + + /// Copy is disallowed + StreamGuard(const StreamGuard&) = delete; + StreamGuard& operator=(const StreamGuard&) = delete; + + /// Move is disallowed, as StreamGuard does not have an uninitialized state, + /// which is required for moves on types with nontrivial destructors. + StreamGuard(StreamGuard&& other) = delete; + StreamGuard& operator=(StreamGuard&& other) = delete; + + /// Resets the currently set stream to the original stream and + /// the currently set device to the original device. Then, + /// set the current device to the device associated with the passed stream, + /// and set the current stream on that device to the passed stream. + /// + /// NOTE: this implementation may skip some stream/device setting if + /// it can prove that it is unnecessary. + /// + /// WARNING: reset_stream does NOT preserve previously set streams on + /// different devices. If you need to set streams on multiple devices + /// on , use MultiStreamGuard instead. + void reset_stream(Stream stream) { + guard_.reset_stream(stream); + } + + /// Returns the stream that was set at the time the guard was constructed. + Stream original_stream() const { + return guard_.original_stream(); + } + + /// Returns the most recent stream that was set using this device guard, + /// either from construction, or via set_stream. + Stream current_stream() const { + return guard_.current_stream(); + } + + /// Returns the most recent device that was set using this device guard, + /// either from construction, or via set_device/reset_device/set_index. + Device current_device() const { + return guard_.current_device(); + } + + /// Returns the device that was set at the most recent reset_stream(), + /// or otherwise the device at construction time. + Device original_device() const { + return guard_.original_device(); + } + + private: + c10::impl::InlineStreamGuard guard_; +}; + +/** + * An OptionalStreamGuard is an RAII class that sets a device to some value on + * initialization, and resets the device to its original value on destruction. + * See OptionalDeviceGuard for more guidance on how to use this class. + */ +struct OptionalStreamGuard { + /// Create an uninitialized guard. + explicit OptionalStreamGuard() = default; + + /// Set the current device to the device associated with the passed stream, + /// and set the current stream on that device to the passed stream. + explicit OptionalStreamGuard(Stream stream) : guard_(stream) {} + + /// Set the current device to the device associated with the passed stream, + /// and set the current stream on that device to the passed stream, + /// if the passed stream is not nullopt. + explicit OptionalStreamGuard(std::optional stream_opt) + : guard_(stream_opt) {} + + /// Copy is disallowed + OptionalStreamGuard(const OptionalStreamGuard&) = delete; + OptionalStreamGuard& operator=(const OptionalStreamGuard&) = delete; + + // See Note [Move construction for RAII guards is tricky] + OptionalStreamGuard(OptionalStreamGuard&& other) = delete; + + // See Note [Move assignment for RAII guards is tricky] + OptionalStreamGuard& operator=(OptionalStreamGuard&& other) = delete; + ~OptionalStreamGuard() = default; + + /// Resets the currently set stream to the original stream and + /// the currently set device to the original device. Then, + /// set the current device to the device associated with the passed stream, + /// and set the current stream on that device to the passed stream. + /// Initializes the guard if it was not previously initialized. + void reset_stream(Stream stream) { + guard_.reset_stream(stream); + } + + /// Returns the stream that was set at the time the guard was most recently + /// initialized, or nullopt if the guard is uninitialized. + std::optional original_stream() const { + return guard_.original_stream(); + } + + /// Returns the most recent stream that was set using this stream guard, + /// either from construction, or via reset_stream, if the guard is + /// initialized, or nullopt if the guard is uninitialized. + std::optional current_stream() const { + return guard_.current_stream(); + } + + /// Restore the original device and stream, resetting this guard to + /// uninitialized state. + void reset() { + guard_.reset(); + } + + private: + c10::impl::InlineOptionalStreamGuard guard_; +}; + +/** + * A MultiStreamGuard is an RAII class that sets the current streams of a set of + * devices all at once, and resets them to their original values on destruction. + */ +struct MultiStreamGuard { + /// Set the current streams to the passed streams on each of their respective + /// devices. + explicit MultiStreamGuard(ArrayRef streams) : guard_(streams) {} + + /// Copy is disallowed + MultiStreamGuard(const MultiStreamGuard&) = delete; + MultiStreamGuard& operator=(const MultiStreamGuard&) = delete; + + // See Note [Move construction for RAII guards is tricky] + MultiStreamGuard(MultiStreamGuard&& other) = delete; + + // See Note [Move assignment for RAII guards is tricky] + MultiStreamGuard& operator=(MultiStreamGuard&& other) = delete; + ~MultiStreamGuard() = default; + + private: + c10::impl::InlineMultiStreamGuard guard_; +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/SymBool.h b/backends/pytorch/include/c10/core/SymBool.h new file mode 100644 index 000000000..d12fa75fb --- /dev/null +++ b/backends/pytorch/include/c10/core/SymBool.h @@ -0,0 +1,184 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) + +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +class SymInt; + +class C10_API SymBool { + public: + /*implicit*/ SymBool(bool b) : data_(b) {} + SymBool(SymNode ptr) : data_(false), ptr_(std::move(ptr)) { + TORCH_CHECK(ptr_->is_bool()); + } + SymBool() : data_(false) {} + + SymNodeImpl* toSymNodeImplUnowned() const { + return ptr_.get(); + } + + SymNodeImpl* release() && { + return std::move(ptr_).release(); + } + + // Only valid if is_heap_allocated() + SymNode toSymNodeImpl() const; + + // Guaranteed to return a SymNode, wrapping using base if necessary + SymNode wrap_node(const SymNode& base) const; + + bool expect_bool() const { + std::optional c = maybe_as_bool(); + TORCH_CHECK(c.has_value()); + return *c; + } + + SymBool sym_and(const SymBool& /*sci*/) const; + SymBool sym_or(const SymBool& /*sci*/) const; + SymBool sym_not() const; + + SymBool operator&(const SymBool& other) const { + return sym_and(other); + } + SymBool operator|(const SymBool& other) const { + return sym_or(other); + } + SymBool operator||(const SymBool& other) const { + return sym_or(other); + } + SymBool operator~() const { + return sym_not(); + } + + // Insert a guard for the bool to be its concrete value, and then return + // that value. Note that C++ comparison operations default to returning + // bool, so it's not so common to have to call this + bool guard_bool(const char* file, int64_t line) const; + bool expect_true(const char* file, int64_t line) const; + bool guard_size_oblivious(const char* file, int64_t line) const; + bool statically_known_true(const char* file, int64_t line) const; + bool guard_or_false(const char* file, int64_t line) const; + bool guard_or_true(const char* file, int64_t line) const; + + bool has_hint() const; + + bool as_bool_unchecked() const { + return data_; + } + + std::optional maybe_as_bool() const { + if (!is_heap_allocated()) { + return data_; + } + return toSymNodeImplUnowned()->constant_bool(); + } + + // Convert SymBool to SymInt (0 or 1) + // This is the C++ equivalent of Python's cast_symbool_to_symint_guardless + SymInt toSymInt() const; + + bool is_heap_allocated() const { + return ptr_; + } + + private: + // TODO: optimize to union + bool data_; + SymNode ptr_; +}; + +C10_API std::ostream& operator<<(std::ostream& os, const SymBool& s); + +#define TORCH_SYM_CHECK(cond, ...) \ + TORCH_CHECK((cond).expect_true(__FILE__, __LINE__), __VA_ARGS__) +#define TORCH_SYM_INTERNAL_ASSERT(cond, ...) \ + TORCH_INTERNAL_ASSERT((cond).expect_true(__FILE__, __LINE__), __VA_ARGS__) +#define TORCH_MAYBE_SYM_CHECK(cond, ...) \ + if constexpr (std::is_same_v, SymBool>) { \ + TORCH_CHECK((cond).expect_true(__FILE__, __LINE__), __VA_ARGS__) \ + } else { \ + TORCH_CHECK((cond), __VA_ARGS__) \ + } + +inline bool guard_size_oblivious( + bool b, + const char* file [[maybe_unused]], + int64_t line [[maybe_unused]]) { + return b; +} + +inline bool guard_size_oblivious( + const c10::SymBool& b, + const char* file, + int64_t line) { + return b.guard_size_oblivious(file, line); +} + +inline bool guard_or_false( + bool b, + const char* file [[maybe_unused]], + int64_t line [[maybe_unused]]) { + return b; +} + +inline bool guard_or_false( + const c10::SymBool& b, + const char* file, + int64_t line) { + return b.guard_or_false(file, line); +} + +inline bool statically_known_true( + bool b, + const char* file [[maybe_unused]], + int64_t line [[maybe_unused]]) { + return b; +} + +inline bool statically_known_true( + const c10::SymBool& b, + const char* file, + int64_t line) { + return b.statically_known_true(file, line); +} + +inline bool guard_or_true( + bool b, + const char* file [[maybe_unused]], + int64_t line [[maybe_unused]]) { + return b; +} + +inline bool guard_or_true( + const c10::SymBool& b, + const char* file, + int64_t line) { + return b.guard_or_true(file, line); +} + +#define TORCH_GUARD_SIZE_OBLIVIOUS(cond) \ + c10::guard_size_oblivious((cond), __FILE__, __LINE__) + +#define TORCH_STATICALLY_KNOWN_TRUE(cond) \ + c10::statically_known_true((cond), __FILE__, __LINE__) + +#define TORCH_GUARD_OR_FALSE(cond) \ + c10::guard_or_false((cond), __FILE__, __LINE__) + +#define TORCH_GUARD_OR_TRUE(cond) c10::guard_or_true((cond), __FILE__, __LINE__) + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/SymFloat.h b/backends/pytorch/include/c10/core/SymFloat.h new file mode 100644 index 000000000..332726ba4 --- /dev/null +++ b/backends/pytorch/include/c10/core/SymFloat.h @@ -0,0 +1,123 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include + +namespace c10 { + +// NB: this is actually double precision; we're using the Python naming here +class C10_API SymFloat { + public: + /*implicit*/ SymFloat(double d) : data_(d) {} + SymFloat(SymNode ptr) + : data_(std::numeric_limits::quiet_NaN()), ptr_(std::move(ptr)) { + TORCH_CHECK(ptr_->is_float()); + } + SymFloat() : data_(0.0) {} + + SymNodeImpl* toSymNodeImplUnowned() const { + return ptr_.get(); + } + + SymNodeImpl* release() && { + return std::move(ptr_).release(); + } + + // Only valid if is_symbolic() + SymNode toSymNodeImpl() const; + + // Guaranteed to return a SymNode, wrapping using base if necessary + SymNode wrap_node(const SymNode& base) const; + + double expect_float() const { + TORCH_CHECK(!is_symbolic()); + return data_; + } + + SymFloat operator+(const SymFloat& /*sci*/) const; + SymFloat operator-(const SymFloat& /*sci*/) const; + SymFloat operator*(const SymFloat& /*sci*/) const; + SymFloat operator/(const SymFloat& /*sci*/) const; + + SymBool sym_eq(const SymFloat& /*sci*/) const; + SymBool sym_ne(const SymFloat& /*sci*/) const; + SymBool sym_lt(const SymFloat& /*sci*/) const; + SymBool sym_le(const SymFloat& /*sci*/) const; + SymBool sym_gt(const SymFloat& /*sci*/) const; + SymBool sym_ge(const SymFloat& /*sci*/) const; + + bool operator==(const SymFloat& o) const { + return sym_eq(o).guard_bool(__FILE__, __LINE__); + } + bool operator!=(const SymFloat& o) const { + return sym_ne(o).guard_bool(__FILE__, __LINE__); + } + bool operator<(const SymFloat& o) const { + return sym_lt(o).guard_bool(__FILE__, __LINE__); + } + bool operator<=(const SymFloat& o) const { + return sym_le(o).guard_bool(__FILE__, __LINE__); + } + bool operator>(const SymFloat& o) const { + return sym_gt(o).guard_bool(__FILE__, __LINE__); + } + bool operator>=(const SymFloat& o) const { + return sym_ge(o).guard_bool(__FILE__, __LINE__); + } + + SymFloat min(const SymFloat& sci) const; + SymFloat max(const SymFloat& sci) const; + + // Need guidance on where to put this code + SymFloat sqrt() const; + + // Insert a guard for the float to be its concrete value, and then return + // that value. This operation always works, even if the float is symbolic, + // so long as we know what the underlying value is. Don't blindly put this + // everywhere; you can cause overspecialization of PyTorch programs with + // this method. + // + // It should be called as guard_float(__FILE__, __LINE__). The file and line + // number can be used to diagnose overspecialization. + double guard_float(const char* file, int64_t line) const; + + bool has_hint() const; + + // N.B. It's important to keep this definition in the header + // as we expect if checks to be folded for mobile builds + // where `is_symbolic` is always false + C10_ALWAYS_INLINE bool is_symbolic() const { + return ptr_; + } + + // UNSAFELY coerce this SymFloat into a double. You MUST have + // established that this is a non-symbolic by some other means, + // typically by having tested is_symbolic(). You will get garbage + // from this function if is_symbolic() + double as_float_unchecked() const { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!is_symbolic()); + return data_; + } + + private: + // TODO: optimize to union + double data_; + SymNode ptr_; +}; + +C10_API std::ostream& operator<<(std::ostream& os, const SymFloat& s); +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/SymInt.h b/backends/pytorch/include/c10/core/SymInt.h new file mode 100644 index 000000000..f9fa7f645 --- /dev/null +++ b/backends/pytorch/include/c10/core/SymInt.h @@ -0,0 +1,586 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +class SymFloat; + +// SymInt represents either a regular int64_t, or a symbolic integer +// (represented in a type erased way as SymNode). The intention is for SymInt +// to represent symbolic sizes that arise when doing shape computation in +// operator kernels. This allows for tracing through programs without baking in +// concrete sizes into kernel calls. +// +// SymInt has an API equivalent to int64_t. In particular, it is a value type. +// Internally, SymInt is represented in a clever packed way, so that it only +// occupies one word of space; but morally, it is a union between an int64_t +// and an intrusive pointer to SymNodeImpl. +// +// Invariant: the referenced SymNodeImpl is guaranteed to be a SymNode where +// is_int() returns true + +class C10_API SymInt { + public: + enum Unchecked { + UNCHECKED, + }; + + /*implicit*/ SymInt(int64_t d) : data_(d) { + if (is_heap_allocated()) { + // Large negative number, heap allocate it + promote_to_negative(); + } + } + SymInt() : data_(0) {} + SymInt(SymNode n); + + // unchecked c-tor accepting raw `data_` + // One appropriate use for this is when you are constructing a symint + // in a situation where you know it is non-negative (or, if it is negative, + // the negative value is -1; i.e., not user controlled) + SymInt(Unchecked /*unused*/, int64_t d) : data_(d) {} + + // TODO: these implementations are not optimal because they allocate a + // temporary and then use the move constructor/assignment + SymInt(const SymInt& s) : data_(0) { + if (s.is_heap_allocated()) { + *this = SymInt(s.toSymNode()); + } else { + data_ = s.data_; + } + } + SymInt(SymInt&& s) noexcept : data_(s.data_) { + s.data_ = 0; + } + + SymInt& operator=(const SymInt& s) { + if (this != &s) { + if (s.is_heap_allocated()) { + *this = SymInt(s.toSymNode()); + } else { + data_ = s.data_; + } + } + return *this; + } + SymInt& operator=(SymInt&& s) noexcept { + if (this != &s) { + release_(); // release the current SymNode if any + data_ = s.data_; + if (s.is_heap_allocated()) + s.data_ = 0; + }; + return *this; + } + + SymNodeImpl* toSymNodeImplUnowned() const { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(is_heap_allocated()); + uint64_t unextended_bits = static_cast(data_) & ~MASK; + uint64_t sign_bit_mask = 1ULL << (62 - 1); + // https://stackoverflow.com/questions/42534749/signed-extension-from-24-bit-to-32-bit-in-c + uint64_t extended_bits = (unextended_bits ^ sign_bit_mask) - sign_bit_mask; + return static_cast( + // NOLINTNEXTLINE(performance-no-int-to-ptr, bugprone*) + reinterpret_cast(static_cast(extended_bits))); + } + + void release_() { + if (is_heap_allocated()) { + SymNode::reclaim(toSymNodeImplUnowned()); // steal + } + } + + SymNodeImpl* release() && { +#ifndef C10_MOBILE + TORCH_INTERNAL_ASSERT(is_heap_allocated()); + auto* r = toSymNodeImplUnowned(); + data_ = 0; // transfer ownership + return r; +#else + TORCH_INTERNAL_ASSERT(false); +#endif + } + + // Only valid if is_heap_allocated() + SymNode toSymNode() const; + + // Guaranteed to return a SymNode, wrapping using base if necessary + SymNode wrap_node(const SymNode& base) const; + + ~SymInt() { + release_(); + } + + // Require the int to be non-symbolic, and if it is symbolic raise an + // error. This is safe to use for C++ code that doesn't work for symbolic + // shapes, and you don't have time to fix it immediately, as if we + // try to trigger the path in C++ you'll appropriately get an error + int64_t expect_int() const { + if (auto r = maybe_as_int()) { + return *r; + } + TORCH_CHECK_ALWAYS_SHOW_CPP_STACKTRACE( + false, "when unpacking SymInt, expected int but got ", *this); + } + + // Test if we have a hint for this int (e.g., guard_int would work). + // Most of the time this is true; it is only false when you have + // an unbacked SymInt. + bool has_hint() const; + + // Insert a guard for the int to be its concrete value, and then return + // that value. This operation always works, even if the int is symbolic, + // so long as we know what the underlying value is (e.g., this won't work + // if you call it on the size of nonzero output). Don't blindly put this + // everywhere; you can cause overspecialization of PyTorch programs with + // this method. + // + // It should be called as guard_int(__FILE__, __LINE__). The file and line + // number can be used to diagnose overspecialization. + int64_t guard_int(const char* file, int64_t line) const; + + // Distinguish actual symbolic values from constants stored on the heap + bool is_symbolic() const { + return is_heap_allocated() && + !toSymNodeImplUnowned()->constant_int().has_value(); + } + + // N.B. It's important to keep this definition in the header + // as we expect if checks to be folded for mobile builds + // where `is_heap_allocated` is always false and optimize dead code paths + C10_ALWAYS_INLINE bool is_heap_allocated() const { +#ifdef C10_MOBILE + return false; +#else + return !check_range(data_); +#endif + } + + SymInt operator+(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymInt(*ma + *mb); + } + } + return operator_add_slow_path(sci); + } + + SymInt operator-(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymInt(*ma - *mb); + } + } + return operator_sub_slow_path(sci); + } + + SymInt operator*(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymInt(*ma * *mb); + } + } + return operator_mul_slow_path(sci); + } + + SymInt operator/(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymInt(*ma / *mb); + } + } + return operator_div_slow_path(sci); + } + + SymInt operator%(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymInt(*ma % *mb); + } + } + return operator_mod_slow_path(sci); + } + + void operator*=(const SymInt& sci) { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + *this = SymInt(*ma * *mb); + return; + } + } + operator_imul_slow_path(sci); + } + + void operator+=(const SymInt& sci) { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + *this = SymInt(*ma + *mb); + return; + } + } + operator_iadd_slow_path(sci); + } + + void operator/=(const SymInt& sci) { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + *this = SymInt(*ma / *mb); + return; + } + } + operator_idiv_slow_path(sci); + } + + SymInt clone() const; + + SymBool sym_eq(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymBool(*ma == *mb); + } + } + return sym_eq_slow_path(sci); + } + + SymBool sym_ne(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymBool(*ma != *mb); + } + } + return sym_ne_slow_path(sci); + } + + SymBool sym_lt(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymBool(*ma < *mb); + } + } + return sym_lt_slow_path(sci); + } + + SymBool sym_le(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymBool(*ma <= *mb); + } + } + return sym_le_slow_path(sci); + } + + SymBool sym_gt(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymBool(*ma > *mb); + } + } + return sym_gt_slow_path(sci); + } + + SymBool sym_ge(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymBool(*ma >= *mb); + } + } + return sym_ge_slow_path(sci); + } + + bool operator==(const SymInt& o) const { + return sym_eq(o).guard_bool(__FILE__, __LINE__); + } + bool operator!=(const SymInt& o) const { + return sym_ne(o).guard_bool(__FILE__, __LINE__); + } + bool operator<(const SymInt& o) const { + return sym_lt(o).guard_bool(__FILE__, __LINE__); + } + bool operator<=(const SymInt& o) const { + return sym_le(o).guard_bool(__FILE__, __LINE__); + } + bool operator>(const SymInt& o) const { + return sym_gt(o).guard_bool(__FILE__, __LINE__); + } + bool operator>=(const SymInt& o) const { + return sym_ge(o).guard_bool(__FILE__, __LINE__); + } + + SymInt min(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymInt(std::min(*ma, *mb)); + } + } + return min_slow_path(sci); + } + + SymInt max(const SymInt& sci) const { + if (auto ma = maybe_as_int()) { + if (auto mb = sci.maybe_as_int()) { + return SymInt(std::max(*ma, *mb)); + } + } + return max_slow_path(sci); + } + + // If both are symbolic, this checks if + // they share the same node. + // If both are not symbolic this just checks normal equality. + bool is_same(const SymInt& other) const; + + operator SymFloat() const; + + void unsafe_set_data(size_t nbytes) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!is_heap_allocated()); + data_ = static_cast(nbytes); + } + + // Don't use this. Prefer maybe_as_int instead + int64_t as_int_unchecked() const { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!is_heap_allocated()); + return data_; + } + + std::optional maybe_as_int() const { + if (!is_heap_allocated()) { + return data_; + } + return maybe_as_int_slow_path(); + } + + // Return whether the integer is directly coercible to a SymInt + // without requiring heap allocation. You don't need to use this + // to check if you can pass an integer to SymInt; this is guaranteed + // to work (it just might heap allocate!) + static bool check_range(int64_t i) { + return i > MAX_UNREPRESENTABLE_INT; + } + + // Return the min representable integer as a SymInt without + // heap allocation. For quantities that count bytes (or larger), + // this is still much larger than you need, so you may consider + // using this as a more efficient version of MIN_INT + static constexpr int64_t min_representable_int() { + return MAX_UNREPRESENTABLE_INT + 1; + } + + private: + void promote_to_negative(); + SymInt operator_add_slow_path(const SymInt& sci) const; + SymInt operator_sub_slow_path(const SymInt& sci) const; + SymInt operator_mul_slow_path(const SymInt& sci) const; + SymInt operator_div_slow_path(const SymInt& sci) const; + SymInt operator_mod_slow_path(const SymInt& sci) const; + void operator_imul_slow_path(const SymInt& sci); + void operator_iadd_slow_path(const SymInt& sci); + void operator_idiv_slow_path(const SymInt& sci); + SymBool sym_eq_slow_path(const SymInt& sci) const; + SymBool sym_ne_slow_path(const SymInt& sci) const; + SymBool sym_lt_slow_path(const SymInt& sci) const; + SymBool sym_le_slow_path(const SymInt& sci) const; + SymBool sym_gt_slow_path(const SymInt& sci) const; + SymBool sym_ge_slow_path(const SymInt& sci) const; + + SymInt min_slow_path(const SymInt& sci) const; + SymInt max_slow_path(const SymInt& sci) const; + + std::optional maybe_as_int_slow_path() const; + + // Constraints on the internal representation: + // + // - Should represent positive and small negative ints + // - No conversion necessary for operations on ints + // - Must represent valid 64-bit pointers + // - Is symbolic test should be FAST (two arithmetic instructions is too + // much). + // This code being a hotpath is based on Strobelight profiles of + // is_heap_allocated(). FB only: https://fburl.com/strobelight/5l50ncxd + // (you will need to change the time window). + // + // So, the scheme is to reserve large negative numbers (assuming + // two's complement): + // + // - 0b0.... means we are a positive int + // - 0b11... means we are a small negative int + // - 0b10... means we are are a pointer. This means that + // [-2^63, -2^62-1] are not representable as ints. + // We don't actually need all of this space as on x86_64 + // as the top 16bits aren't used for anything + static constexpr uint64_t MASK = 1ULL << 63 | 1ULL << 62 | 1ULL << 61; + static constexpr uint64_t IS_SYM = 1ULL << 63 | 1ULL << 61; + // We must manually translate the bit pattern test into a greater + // than test because compiler doesn't figure it out: + // https://godbolt.org/z/356aferaW + static constexpr int64_t MAX_UNREPRESENTABLE_INT = + -1LL & static_cast(~(1ULL << 62)); + int64_t data_; +}; + +/// Sum of a list of SymInt; accumulates into the c10::SymInt expression +template < + typename C, + typename std::enable_if_t< + std::is_same_v, + int> = 0> +inline c10::SymInt multiply_integers(const C& container) { + return std::accumulate( + container.begin(), + container.end(), + c10::SymInt(1), + [](const c10::SymInt& a, const c10::SymInt& b) { return a * b; }); +} + +template < + typename Iter, + typename = std::enable_if_t::value_type, + c10::SymInt>>> +inline c10::SymInt multiply_integers(Iter begin, Iter end) { + return std::accumulate( + begin, + end, + c10::SymInt(1), + [](const c10::SymInt& a, const c10::SymInt& b) { return a * b; }); +} + +#define DECLARE_SYMINT_OP_INTONLY(scalar_t, RetTy) \ + C10_API RetTy operator%(const SymInt& a, scalar_t b); \ + C10_API RetTy operator%(scalar_t a, const SymInt& b); + +#define DECLARE_SYMINT_OP(scalar_t, RetTy) \ + C10_API RetTy operator+(const SymInt& a, scalar_t b); \ + C10_API RetTy operator-(const SymInt& a, scalar_t b); \ + C10_API RetTy operator*(const SymInt& a, scalar_t b); \ + C10_API RetTy operator/(const SymInt& a, scalar_t b); \ + C10_API RetTy operator+(scalar_t a, const SymInt& b); \ + C10_API RetTy operator-(scalar_t a, const SymInt& b); \ + C10_API RetTy operator*(scalar_t a, const SymInt& b); \ + C10_API RetTy operator/(scalar_t a, const SymInt& b); \ + C10_API bool operator==(const SymInt& a, scalar_t b); \ + C10_API bool operator!=(const SymInt& a, scalar_t b); \ + C10_API bool operator<(const SymInt& a, scalar_t b); \ + C10_API bool operator<=(const SymInt& a, scalar_t b); \ + C10_API bool operator>(const SymInt& a, scalar_t b); \ + C10_API bool operator>=(const SymInt& a, scalar_t b); \ + C10_API bool operator==(scalar_t a, const SymInt& b); \ + C10_API bool operator!=(scalar_t a, const SymInt& b); \ + C10_API bool operator<(scalar_t a, const SymInt& b); \ + C10_API bool operator<=(scalar_t a, const SymInt& b); \ + C10_API bool operator>(scalar_t a, const SymInt& b); \ + C10_API bool operator>=(scalar_t a, const SymInt& b); + +DECLARE_SYMINT_OP_INTONLY(int64_t, SymInt) +DECLARE_SYMINT_OP_INTONLY(int32_t, SymInt) +DECLARE_SYMINT_OP_INTONLY(uint64_t, SymInt) +DECLARE_SYMINT_OP_INTONLY(uint32_t, SymInt) +DECLARE_SYMINT_OP(int64_t, SymInt) +DECLARE_SYMINT_OP(int32_t, SymInt) // make sure constants work +DECLARE_SYMINT_OP(uint64_t, SymInt) +DECLARE_SYMINT_OP(uint32_t, SymInt) +DECLARE_SYMINT_OP(double, SymFloat) +DECLARE_SYMINT_OP(float, SymFloat) // just for completeness + +// On OSX size_t is different than uint64_t so we have to +// define it separately +#if defined(__APPLE__) +DECLARE_SYMINT_OP_INTONLY(size_t, SymInt) +DECLARE_SYMINT_OP(size_t, SymInt) +#endif + +#undef DECLARE_SYMINT_OP + +C10_API std::ostream& operator<<(std::ostream& os, const SymInt& s); +C10_API SymInt operator-(const SymInt& s); + +inline bool sym_eq(int64_t a, int64_t b) { + return a == b; +} + +inline SymBool sym_eq(const SymInt& a, const SymInt& b) { + return a.sym_eq(b); +} + +inline bool sym_ne(int64_t a, int64_t b) { + return a != b; +} + +inline SymBool sym_ne(const SymInt& a, const SymInt& b) { + return a.sym_ne(b); +} + +inline bool sym_lt(int64_t a, int64_t b) { + return a < b; +} + +inline SymBool sym_lt(const SymInt& a, const SymInt& b) { + return a.sym_lt(b); +} + +inline bool sym_le(int64_t a, int64_t b) { + return a <= b; +} + +inline SymBool sym_le(const SymInt& a, const SymInt& b) { + return a.sym_le(b); +} + +inline bool sym_gt(int64_t a, int64_t b) { + return a > b; +} + +inline SymBool sym_gt(const SymInt& a, const SymInt& b) { + return a.sym_gt(b); +} + +inline bool sym_ge(int64_t a, int64_t b) { + return a >= b; +} + +inline SymBool sym_ge(const SymInt& a, const SymInt& b) { + return a.sym_ge(b); +} + +} // namespace c10 + +#include + +namespace std { + +template <> +class numeric_limits { + public: + static constexpr bool is_specialized = true; + + static constexpr int64_t max() noexcept { + return std::numeric_limits::max(); + } + + static constexpr int64_t min() noexcept { + return std::numeric_limits::min(); + } + + static constexpr bool is_signed = true; + static constexpr bool is_integer = true; +}; + +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/SymIntArrayRef.h b/backends/pytorch/include/c10/core/SymIntArrayRef.h new file mode 100644 index 000000000..b63753b18 --- /dev/null +++ b/backends/pytorch/include/c10/core/SymIntArrayRef.h @@ -0,0 +1,113 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { +using SymIntArrayRef = ArrayRef; + +inline at::IntArrayRef asIntArrayRefUnchecked(c10::SymIntArrayRef ar) { + return IntArrayRef(reinterpret_cast(ar.data()), ar.size()); +} + +// TODO: a SymIntArrayRef containing a heap allocated large negative integer +// can actually technically be converted to an IntArrayRef... but not with +// the non-owning API we have here. We can't reinterpet cast; we have to +// allocate another buffer and write the integers into it. If you need it, +// we can do it. But I don't think you need it. + +inline std::optional asIntArrayRefSlowOpt( + c10::SymIntArrayRef ar) { + for (const c10::SymInt& sci : ar) { + if (sci.is_heap_allocated()) { + return std::nullopt; + } + } + + return {asIntArrayRefUnchecked(ar)}; +} + +inline at::IntArrayRef asIntArrayRefSlow( + c10::SymIntArrayRef ar, + const char* file, + int64_t line) { + for (const c10::SymInt& sci : ar) { + TORCH_CHECK( + !sci.is_heap_allocated(), + file, + ":", + line, + ": SymIntArrayRef expected to contain only concrete integers"); + } + return asIntArrayRefUnchecked(ar); +} + +// Even slower than asIntArrayRefSlow, as it forces an allocation for a +// destination int, BUT it is able to force specialization (it never errors) +inline c10::DimVector asIntArrayRefSlowAlloc( + c10::SymIntArrayRef ar, + const char* file, + int64_t line) { + c10::DimVector res(ar.size(), 0); + for (const auto i : c10::irange(ar.size())) { + res[i] = ar[i].guard_int(file, line); + } + return res; +} + +#define C10_AS_INTARRAYREF_SLOW(a) c10::asIntArrayRefSlow(a, __FILE__, __LINE__) +#define C10_AS_INTARRAYREF_SLOW_ALLOC(a) \ + c10::asIntArrayRefSlowAlloc(a, __FILE__, __LINE__) + +// Prefer using a more semantic constructor, like +// fromIntArrayRefKnownNonNegative +inline SymIntArrayRef fromIntArrayRefUnchecked(IntArrayRef array_ref) { + return SymIntArrayRef( + reinterpret_cast(array_ref.data()), array_ref.size()); +} + +inline SymIntArrayRef fromIntArrayRefKnownNonNegative(IntArrayRef array_ref) { + return fromIntArrayRefUnchecked(array_ref); +} + +inline SymIntArrayRef fromIntArrayRefSlow(IntArrayRef array_ref) { + for (long i : array_ref) { + TORCH_CHECK( + SymInt::check_range(i), + "IntArrayRef contains an int that cannot be represented as a SymInt: ", + i); + } + return SymIntArrayRef( + reinterpret_cast(array_ref.data()), array_ref.size()); +} + +inline c10::SymBool sym_equals(SymIntArrayRef LHS, SymIntArrayRef RHS) { + if (LHS.size() != RHS.size()) { + return c10::SymBool(false); + } + + c10::SymBool result = sym_eq(LHS.size(), RHS.size()); + for (size_t i = 0; i < RHS.size(); ++i) { + c10::SymBool equals = sym_eq(LHS[i], RHS[i]); + std::optional equals_bool = equals.maybe_as_bool(); + + if (equals_bool.has_value() && !*equals_bool) { + // Early return if element comparison is known to be false + return equals; + } + result = result.sym_and(equals); + } + return result; +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/SymNodeImpl.h b/backends/pytorch/include/c10/core/SymNodeImpl.h new file mode 100644 index 000000000..a4257684e --- /dev/null +++ b/backends/pytorch/include/c10/core/SymNodeImpl.h @@ -0,0 +1,261 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) + +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include + +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-parameter") + +namespace c10 { + +class SymNodeImpl; +using SymNode = c10::intrusive_ptr; + +// When you add a method, you also need to edit +// torch/csrc/jit/python/init.cpp +// torch/csrc/utils/python_symnode.h +// c10/core/ConstantSymNodeImpl.h +class C10_API SymNodeImpl : public c10::intrusive_ptr_target { + public: + ~SymNodeImpl() override = default; + + template + c10::intrusive_ptr dyn_cast() const { + return c10::intrusive_ptr::reclaim_copy(dynamic_cast(this)); + } + + // these could be pure virtual when we implement LTC versions + virtual bool is_int() { + TORCH_CHECK(false, "NYI"); + } + virtual bool is_bool() { + TORCH_CHECK(false, "NYI"); + } + virtual bool is_float() { + TORCH_CHECK(false, "NYI"); + } + virtual bool is_nested_int() const { + return false; + } + virtual SymNode add(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode sub(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode mul(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + // NB: legacy, prefer float_truediv or int_truediv + virtual SymNode truediv(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode float_truediv(const SymNode& other) { + return truediv(other); + } + virtual SymNode int_truediv(const SymNode& other) { + return truediv(other); + } + // NB: legacy, prefer float_pow or pow_by_natural + virtual SymNode pow(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode float_pow(const SymNode& other) { + return pow(other); + } + virtual SymNode pow_by_natural(const SymNode& other) { + return pow(other); + } + // NB: legacy, prefer int_floordiv + virtual SymNode floordiv(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode int_floordiv(const SymNode& other) { + return floordiv(other); + } + virtual SymNode mod(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode eq(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode ne(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode gt(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode lt(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode le(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode ge(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode ceil() { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode floor() { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode neg() { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode sym_min(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode sym_max(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode sym_or(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode sym_and(const SymNode& other) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode sym_not() { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode sym_ite(const SymNode& then_val, const SymNode& else_val) { + TORCH_CHECK(false, "NYI"); + } + // NB: self is ignored here, only the arguments are used + virtual SymNode is_contiguous( + ArrayRef sizes, + ArrayRef strides) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode is_channels_last_contiguous_2d( + ArrayRef sizes, + ArrayRef strides) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode is_channels_last_contiguous_3d( + ArrayRef sizes, + ArrayRef strides) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode is_channels_last_strides_2d( + ArrayRef sizes, + ArrayRef strides) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode is_channels_last_strides_3d( + ArrayRef sizes, + ArrayRef strides) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode is_non_overlapping_and_dense( + ArrayRef sizes, + ArrayRef strides) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode clone() { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode sym_float() { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode wrap_int(int64_t num) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode wrap_float(double num) { + TORCH_CHECK(false, "NYI"); + } + virtual SymNode wrap_bool(bool num) { + TORCH_CHECK(false, "NYI"); + } + virtual int64_t guard_int(const char* file, int64_t line) { + TORCH_CHECK(false, "NYI"); + } + virtual bool guard_bool(const char* file, int64_t line) { + TORCH_CHECK(false, "NYI"); + } + virtual double guard_float(const char* file, int64_t line) { + TORCH_CHECK(false, "NYI"); + } + virtual bool guard_size_oblivious(const char* file, int64_t line) { + // No improvement for unbacked SymBools by default, replace this + // with a better implementation! + return guard_bool(file, line); + } + virtual bool guard_or_false(const char* file, int64_t line) { + // Note: PT2 primarily uses PythonSymNodeImpl for this functionality. + // XLA is currently the main consumer of this fallback path since it uses + // ahead-of-time compilation and cannot depend on Python runtime. + return guard_bool(file, line); + } + virtual bool statically_known_true(const char* file, int64_t line) { + // Note: PT2 primarily uses PythonSymNodeImpl for this functionality. + // XLA is currently the main consumer of this fallback path since it uses + // ahead-of-time compilation and cannot depend on Python runtime. + return guard_bool(file, line); + } + virtual bool guard_or_true(const char* file, int64_t line) { + // Note: PT2 primarily uses PythonSymNodeImpl for this functionality. + // XLA is currently the main consumer of this fallback path since it uses + // ahead-of-time compilation and cannot depend on Python runtime. + return guard_bool(file, line); + } + virtual bool expect_true(const char* file, int64_t line) { + // No improvement for unbacked SymBools by default, replace this + // with a better implementation! + return guard_bool(file, line); + } + virtual int64_t int_() { + TORCH_CHECK(false, "NYI"); + } + virtual bool bool_() { + TORCH_CHECK(false, "NYI"); + } + virtual bool has_hint() { + TORCH_CHECK(false, "NYI"); + } + virtual std::string str() { + TORCH_CHECK(false, "NYI"); + } + virtual std::string _graph_repr() { + return str(); + } + virtual std::optional nested_int() { + return std::nullopt; + } + virtual std::optional nested_int_coeff() { + return std::nullopt; + } + virtual std::optional constant_int() { + return std::nullopt; + } + virtual std::optional constant_bool() { + return std::nullopt; + } + virtual std::optional maybe_as_int() { + return std::nullopt; + } + virtual bool is_constant() { + return false; + } + virtual bool is_symbolic() { + return true; + } + std::ostream& operator<<(std::ostream& os) { + os << str(); + return os; + } +}; + +} // namespace c10 +C10_DIAGNOSTIC_POP() + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/SymbolicShapeMeta.h b/backends/pytorch/include/c10/core/SymbolicShapeMeta.h new file mode 100644 index 000000000..411c81a98 --- /dev/null +++ b/backends/pytorch/include/c10/core/SymbolicShapeMeta.h @@ -0,0 +1,234 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include + +namespace c10 { + +class C10_API SymbolicShapeMeta { + public: + // Basic metadata from which other quantities are derived + SymDimVector sizes_ = {0}; + SymDimVector strides_ = {1}; + SymInt storage_offset_ = 0; + + bool strides_valid_ = true; // e.g. for sparse where there are no strides + + SymbolicShapeMeta() = default; + ~SymbolicShapeMeta() = default; + SymbolicShapeMeta(const SymbolicShapeMeta& other); + SymbolicShapeMeta(SymbolicShapeMeta&& other) = delete; + SymbolicShapeMeta& operator=(const SymbolicShapeMeta& other) = delete; + SymbolicShapeMeta& operator=(SymbolicShapeMeta&& other) = delete; + + void refresh_numel() { + // Non-const, don't need to hold mutables_ lock + available_.fetch_and(~numel_avail); + numel_ = 1; + } + + void refresh_contiguous() { + // Non-const, don't need to hold mutables_ lock + available_.fetch_and(numel_avail); + is_contiguous_ = false; + is_channels_last_contiguous_ = false; + is_channels_last_3d_contiguous_ = false; + is_channels_last_ = false; + is_channels_last_3d_ = false; + is_non_overlapping_and_dense_ = false; + } + + int64_t dim() const { + return static_cast(sizes_.size()); + } + + // Accessors for derived quantities, computed lazily on first access + + bool has_numel() const { + return available_.load() & numel_avail; + } + bool has_is_contiguous() const { + return available_.load() & is_contiguous_avail; + } + bool has_is_channels_last_contiguous() const { + return available_.load() & is_channels_last_contiguous_avail; + } + bool has_is_channels_last_3d_contiguous() const { + return available_.load() & is_channels_last_3d_contiguous_avail; + } + bool has_is_channels_last() const { + return available_.load() & is_channels_last_avail; + } + bool has_is_channels_last_3d() const { + return available_.load() & is_channels_last_3d_avail; + } + bool has_is_non_overlapping_and_dense() const { + return available_.load() & is_non_overlapping_and_dense_avail; + } + + // Accessors to cached derived properties + // DO NOT call with mutables_ lock held + const SymInt& numel() const { + if (C10_UNLIKELY(!has_numel())) { + init_numel(); + } + return numel_; + } + + const SymBool& is_contiguous(at::MemoryFormat memory_format) const { + if (memory_format == at::MemoryFormat::ChannelsLast) { + return this->is_channels_last_contiguous(); + } else if (memory_format == at::MemoryFormat::ChannelsLast3d) { + return this->is_channels_last_3d_contiguous(); + } + return this->is_contiguous(); + } + + const SymBool& is_contiguous() const { + if (C10_UNLIKELY(!has_is_contiguous())) { + init_is_contiguous(); + } + return is_contiguous_; + } + + const SymBool& is_channels_last_contiguous() const { + if (C10_UNLIKELY(!has_is_channels_last_contiguous())) { + init_is_channels_last_contiguous(); + } + return is_channels_last_contiguous_; + } + + const SymBool& is_channels_last_3d_contiguous() const { + if (C10_UNLIKELY(!has_is_channels_last_3d_contiguous())) { + init_is_channels_last_3d_contiguous(); + } + return is_channels_last_3d_contiguous_; + } + + const SymBool& is_channels_last() const { + if (C10_UNLIKELY(!has_is_channels_last())) { + init_is_channels_last(); + } + return is_channels_last_; + } + + const SymBool& is_channels_last_3d() const { + if (C10_UNLIKELY(!has_is_channels_last_3d())) { + init_is_channels_last_3d(); + } + return is_channels_last_3d_; + } + + const SymBool& is_non_overlapping_and_dense() const { + if (C10_UNLIKELY(!has_is_non_overlapping_and_dense())) { + init_is_non_overlapping_and_dense(); + } + return is_non_overlapping_and_dense_; + } + + // Assumptions so we can short-circuit computation + // NOTE: Don't need to lock mutables_ since these aren't const + void assume_contiguous(SymBool val = true) { + is_contiguous_ = std::move(val); + available_.fetch_or(is_contiguous_avail); + } + void assume_channels_last_contiguous(SymBool val = true) { + is_contiguous_ = std::move(val); + available_.fetch_or(is_channels_last_contiguous_avail); + } + void assume_channels_last_3d_contiguous(SymBool val = true) { + is_channels_last_3d_contiguous_ = std::move(val); + available_.fetch_or(is_channels_last_3d_contiguous_avail); + } + void assume_channels_last(SymBool val = true) { + is_channels_last_ = std::move(val); + available_.fetch_or(is_channels_last_avail); + } + void assume_channels_last_3d(SymBool val = true) { + is_channels_last_3d_ = std::move(val); + available_.fetch_or(is_channels_last_3d_avail); + } + void assume_non_overlapping_and_dense(SymBool val = true) { + is_non_overlapping_and_dense_ = std::move(val); + available_.fetch_or(is_non_overlapping_and_dense_avail); + } + + private: + SymBool compute_contiguous() const; + SymBool compute_channels_last_contiguous_2d() const; + SymBool compute_channels_last_contiguous_3d() const; + SymBool compute_strides_like_channels_last_2d() const; + SymBool compute_strides_like_channels_last_3d() const; + SymBool compute_non_overlapping_and_dense() const; + + // These are little wrappers over the real compute_ functions that + // can make use of other contiguity fields to short circuit. + // They need to be implemented separately for SymBool, as SymBool does + // not short circuit. + // TODO: should the SymBool cases avoid the short circuit? Need to reason + // if its correct, and reason if the simpler expressions are better for + // analysis (maybe not!) + + SymBool compute_channels_last_contiguous_3d_dim5() const; + SymBool compute_channels_last_2d_dim5() const; + SymBool compute_channels_last_3d_dim5() const; + SymBool compute_is_non_overlapping_and_dense_dim4() const; + SymBool compute_is_non_overlapping_and_dense_dim5() const; + SymBool compute_is_non_overlapping_and_dense_anydim() const; + + void init_numel() const; + void init_is_contiguous() const; + void init_is_channels_last_contiguous() const; + void init_is_channels_last_3d_contiguous() const; + void init_is_channels_last() const; + void init_is_channels_last_3d() const; + void init_is_non_overlapping_and_dense() const; + + // NOTE: These only set if !has_foo() + void set_numel(SymInt val) const; + void set_is_contiguous(SymBool val) const; + void set_is_channels_last_contiguous(SymBool val) const; + void set_is_channels_last_3d_contiguous(SymBool val) const; + void set_is_channels_last(SymBool val) const; + void set_is_channels_last_3d(SymBool val) const; + void set_is_non_overlapping_and_dense(SymBool val) const; + + // Lazily initialized variables, with the corresponding available_ flag + // indicating whether the value has been initialized + mutable std::atomic available_{0}; + + enum avail { + numel_avail = 1 << 0, + is_contiguous_avail = 1 << 1, + is_channels_last_contiguous_avail = 1 << 2, + is_channels_last_3d_contiguous_avail = 1 << 3, + is_channels_last_avail = 1 << 4, + is_channels_last_3d_avail = 1 << 5, + is_non_overlapping_and_dense_avail = 1 << 6, + }; + + // Mutex to prevent races when initializing the variable from const accessors + mutable std::mutex mutables_; + mutable SymInt numel_ = 1; + mutable SymBool is_contiguous_{true}; + mutable SymBool is_channels_last_contiguous_{false}; + mutable SymBool is_channels_last_3d_contiguous_{false}; + mutable SymBool is_channels_last_{false}; + mutable SymBool is_channels_last_3d_{false}; + mutable SymBool is_non_overlapping_and_dense_{true}; +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/TensorImpl.h b/backends/pytorch/include/c10/core/TensorImpl.h new file mode 100644 index 000000000..03faea3fb --- /dev/null +++ b/backends/pytorch/include/c10/core/TensorImpl.h @@ -0,0 +1,3333 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +// A global boolean variable to control whether we free memory when a Tensor +// is shrunk to a smaller size. As a result, a Tensor is always going to +// keep the memory allocated for its maximum capacity reshaped to so far. +// +// This parameter is respected "upper-case" methods which call Resize() +// (e.g., CopyFrom, ResizeLike); it is NOT respected by Tensor::resize_ +// or ShrinkTo, both of which guarantee to never to free memory. +C10_DECLARE_bool(caffe2_keep_on_shrink); + +// Since we can have high variance in blob memory allocated across different +// inputs in the same run, we will shrink the blob only if the memory gain +// is larger than this flag in bytes. This only applies to functions which +// respect caffe2_keep_on_shrink. +C10_DECLARE_int64(caffe2_max_keep_on_shrink_memory); + +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-default") + +namespace at { +class Tensor; +class TensorBase; +} // namespace at + +namespace c10 { + +/** + * A utility function to convert vector to vector. + */ +inline std::vector ToVectorint64_t(const ArrayRef& src) { + return std::vector(src.begin(), src.end()); +} + +/** + * Return product of all dimensions starting from k + */ +inline int64_t size_from_dim_(int k, IntArrayRef dims) { + int64_t r = 1; + for (const auto i : c10::irange(k, dims.size())) { + r *= dims[i]; + } + return r; +} + +// Product of all dims up to k (not including dims[k]) +inline int64_t size_to_dim_(int k, IntArrayRef dims) { + TORCH_CHECK(k >= 0 && static_cast(k) <= dims.size()); + int64_t r = 1; + for (const auto i : c10::irange(k)) { + r *= dims[i]; + } + return r; +} + +// Product of all dims between k and l (not including dims[k] and dims[l]) +inline int64_t size_between_dim_(int k, int l, IntArrayRef dims) { + TORCH_CHECK((unsigned)l < dims.size() && (unsigned)k < dims.size()); + int64_t r = 1; + if (k < l) { + for (int i = k + 1; i < l; ++i) { + r *= dims[i]; + } + } else { + for (int i = l + 1; i < k; ++i) { + r *= dims[i]; + } + } + return r; +} + +// Wrap around axis_index if it is negative, s.t., -1 is the last dim +inline int canonical_axis_index_(int axis_index, int ndims) { + TORCH_CHECK(axis_index >= -ndims); + TORCH_CHECK(axis_index < ndims); + if (axis_index < 0) { + return axis_index + ndims; + } + return axis_index; +} + +using PlacementDtor = void (*)(void*, size_t); + +/* + * A Context that will call extra placement deleter during + * deconstruction. + * + * Accept a already constructed DataPtr and store it as member + * during destruction, we'll call extra deleter on the underlying + * data pointer before the DataPtr is destructed. + * `data_ptr_` owns the memory. + */ +struct C10_API PlacementDeleteContext { + DataPtr data_ptr_; + PlacementDtor placement_dtor_; + size_t size_; + + PlacementDeleteContext( + DataPtr&& data_ptr, + PlacementDtor placement_dtor, + size_t size) + : data_ptr_(std::move(data_ptr)), + placement_dtor_(placement_dtor), + size_(size) {} + + PlacementDeleteContext(PlacementDeleteContext&&) noexcept = delete; + PlacementDeleteContext(const PlacementDeleteContext&) = delete; + PlacementDeleteContext& operator=(const PlacementDeleteContext&) = delete; + PlacementDeleteContext& operator=(PlacementDeleteContext&&) = delete; + static DataPtr makeDataPtr( + DataPtr&& data_ptr, + PlacementDtor placement_dtor, + size_t size, + Device device); + ~PlacementDeleteContext() { + placement_dtor_(data_ptr_.get(), size_); + // original memory will be freed when data_ptr_ is destructed + } +}; + +struct C10_API AutogradMetaInterface { + virtual void set_requires_grad( + bool requires_grad, + at::TensorImpl* self_impl) = 0; + virtual bool requires_grad() const = 0; + virtual at::Tensor& mutable_grad() = 0; + virtual const at::Tensor& grad() const = 0; + virtual const at::Tensor& fw_grad(uint64_t level, const at::TensorBase& self) + const = 0; + virtual void set_fw_grad( + const at::TensorBase& new_grad, + const at::TensorBase& self, + uint64_t level, + bool is_inplace_op) = 0; + virtual ~AutogradMetaInterface(); +}; + +namespace impl { + +// Unfortunately, the definition of AutogradMeta lives in a separate +// compilation unit than TensorImpl (libtorch.so versus libc10.so) +// which means that we cannot construct an AutogradMeta from TensorImpl, +// not even from the cpp file. So we have to indirect it through a factory +// function which will be initialized when we load libtorch.so. + +struct C10_API AutogradMetaFactory { + virtual ~AutogradMetaFactory() = default; + virtual std::unique_ptr make() const = 0; + // This method is the dumbest method. But I don't have access + // to Tensor (not TensorImpl) which is undefined in this header. + virtual const at::Tensor& undefined_tensor() const = 0; +}; + +C10_API void SetAutogradMetaFactory(AutogradMetaFactory* factory); +C10_API AutogradMetaFactory* GetAutogradMetaFactory(); + +struct C10_API AutogradMetaFactoryRegisterer{ + explicit AutogradMetaFactoryRegisterer(AutogradMetaFactory * factory){ + SetAutogradMetaFactory(factory); +} // namespace impl +}; // namespace c10 + +} // namespace impl + +struct C10_API NamedTensorMetaInterface { + virtual ~NamedTensorMetaInterface() = default; + virtual std::unique_ptr clone() const { + TORCH_INTERNAL_ASSERT( + false, "Not implemented: NamedTensorMetaInterface::clone"); + } + virtual int64_t slow_dim() const { + TORCH_INTERNAL_ASSERT( + false, "Not implemented: NamedTensorMetaInterface::slow_dim"); + } +}; + +// For ease of copy pasting +#if 0 +is_contiguous +is_channels_last_contiguous +is_channels_last_3d_contiguous +is_channels_last +is_channels_last_3d +is_non_overlapping_and_dense +#endif + +/** + * This structure is intended to hold additional metadata of the specific device + * backend. + **/ +struct C10_API BackendMeta : intrusive_ptr_target { + ~BackendMeta() override = default; + virtual intrusive_ptr clone( + const intrusive_ptr& ptr) const { + return ptr; + } +}; + +struct C10_API ExtraMeta { + std::unique_ptr symbolic_shape_meta_ = nullptr; + std::unique_ptr named_tensor_meta_ = nullptr; + intrusive_ptr backend_meta_ = nullptr; + std::optional custom_data_ptr_error_msg_ = std::nullopt; + std::optional custom_storage_error_msg_ = std::nullopt; + + ExtraMeta() = default; + ~ExtraMeta() = default; + ExtraMeta(const ExtraMeta& other) { + if (other.symbolic_shape_meta_) { + symbolic_shape_meta_ = + std::make_unique(*other.symbolic_shape_meta_); + } + if (other.named_tensor_meta_) { + named_tensor_meta_ = other.named_tensor_meta_->clone(); + } + if (other.backend_meta_) { + backend_meta_ = other.backend_meta_->clone(other.backend_meta_); + } + if (other.custom_data_ptr_error_msg_) { + custom_data_ptr_error_msg_ = other.custom_data_ptr_error_msg_; + } + if (other.custom_storage_error_msg_) { + custom_storage_error_msg_ = other.custom_storage_error_msg_; + } + } + ExtraMeta& operator=(const ExtraMeta& other) = delete; + ExtraMeta(ExtraMeta&& other) = delete; + ExtraMeta& operator=(ExtraMeta&& other) = delete; + + ExtraMeta( + std::unique_ptr symbolic_shape_meta, + std::unique_ptr named_tensor_meta, + intrusive_ptr backend_meta, + std::optional custom_data_ptr_error_msg = std::nullopt, + std::optional custom_storage_access_error_msg = std::nullopt) + : symbolic_shape_meta_(std::move(symbolic_shape_meta)), + named_tensor_meta_(std::move(named_tensor_meta)), + backend_meta_(std::move(backend_meta)), + custom_data_ptr_error_msg_(std::move(custom_data_ptr_error_msg)), + custom_storage_error_msg_(std::move(custom_storage_access_error_msg)) {} + + std::unique_ptr clone() const { + return std::make_unique(*this); + } +}; + +// NOTE [ Version Counter Sharing ] +// +// Every Tensor has a version counter. Version counters are incremented whenever +// the data or size of a tensor changes through in-place Variable operations. +// Version counters are used to detect modifications to saved variables which +// would result in incorrect gradient calculations. Version counters may be +// shared between Variables: +// +// 1. A view shares the version counter of the base Variable, +// 2. `x.detach()` shares the version counter of `x`, +// 3. Unpacked saved variables share the version counter of the source. +// +// Version counters are not shared in these scenarios: +// +// 1. When we replace a `Variable`'s underlying `Tensor` by calling +// `set_data(...)`, +// 2. `x.data` does not share the version counter of `x`. (See discussion at +// https://github.com/pytorch/pytorch/issues/5396) +// +// Question: Why do we put the version counter in TensorImpl instead of +// AutogradMeta? +// +// Answer: After the Variable/Tensor merge, a tensor will not have AutogradMeta +// when its `requires_grad_` is false, but when we use this tensor in the +// forward pass of a function that requires saving this tensor for backward, we +// need to keep track of this tensor's version to make sure it's always valid in +// the autograd graph. +// +// To achieve this goal, we put the version counter in TensorImpl instead of +// AutogradMeta, and have it always be available. This allows us to have the +// optimization of not carrying AutogradMeta when a tensor doesn't require +// gradient. +// +// A hypothetical alternative way to achieve this goal is to initialize +// AutogradMeta and create the version counter for the non-requires-grad tensor +// only when it's saved for backward. However, since saving a tensor for +// backward happens in the forward pass, and our invariant is that forward pass +// needs to be thread-safe, lazy-initializing AutogradMeta when saving a tensor +// can introduce race conditions when we are running the forward pass in +// multi-thread scenarios, thus making the forward pass not thread-safe anymore, +// which breaks the invariant. +struct C10_API VariableVersion { + private: + struct VersionCounter : intrusive_ptr_target { + VersionCounter(uint32_t version) : version_(version) {} + std::atomic version_; + }; + c10::intrusive_ptr version_counter_; + + public: + // Note [Disabled VariableVersion] + // VariableVersion struct has an intrusive_ptr pointing VersionCounter struct + // with an atomic variable. Thus `VariableVersion(/*version=*/0)` is not as + // cheap as we expected. In some cases constructing a VariableVersion with + // version 0 is not necessary so we add a cheap constructor which + // doesn't allocate the intrusive_ptr. + // Example use cases are: + // - Inference tensors don't track version counter, so they'll just always + // have disabled VariableVersion. + // - In SavedVariable class we override version_counter_ inside its + // constructor + // so that we can use the cheap constructor there. + enum Disabled { DISABLED }; + // It's okay to return true even for inference tensor which + // doesn't have version counter enabled. + // We want to be permissive here since in many cases (e.g. make_variable) + // we can std::move a TensorImpl if there's no other uses which saves us + // an additional TensorImpl allocation. + bool unique() const { + return version_counter_ ? 1 == version_counter_.use_count() : true; + } + // NOTE: As of C++11 and 14, default-constructing a std::atomic variable + // leaves it in a persistently undefined state. See + // https://cplusplus.github.io/LWG/issue2334. + VariableVersion(uint32_t version) + : version_counter_(c10::make_intrusive(version)) {} + VariableVersion(Disabled /*unused*/ = DISABLED) {} + + bool enabled() const { + return version_counter_; + } + + // Note [Inplace update inference tensor] + // 1. Inplace update to inference tensor is forbidden in normal mode. + // For example: + // inference_tensor.copy_(normal_tensor_requires_grad) + // This inplace makes inference_tensor have requires_grad=True and + // have a grad_fn. This is bad because views of `inference_tensor` + // created in InferenceMode won't be able to know the grad_fn since + // their ViewMeta were not recorded. To match NoGradMode behavior + // that "inplace update to a view created in NoGradMode raise an error", + // we just ban inplace update to inference tensor since we can't tell + // if an inference tensor is a view created in InferenceMode. + // + // Note that views of normal tensor created in InferenceMode has proper + // ViewMeta so that they're aware of the grad_fn correctly. + // + // 2. Inplace update to inference tensor in inference tensor doesn't bump + // version counter. + // * It either doesn't call bump() by skipping ADInplaceOrView kernel, + // - e.g. inference_tensor.add_(1) + // * or bump() is a no-op for inference tensor. + // - e.g. inference_tensor.add_(normal_tensor) + void bump() { + // TODO: Replace the link to the documentation once it's available. + TORCH_CHECK( + version_counter_ || InferenceMode::is_enabled(), + "Inplace update to inference tensor outside InferenceMode is not allowed." + "You can make a clone to get a normal tensor before doing inplace update." + "See https://github.com/pytorch/rfcs/pull/17 for more details."); + if (version_counter_) { + ++version_counter_->version_; + } + } + + void set_version(int64_t i) { + TORCH_CHECK( + version_counter_, + "Tried to call torch.autograd._unsafe_set_version() on a tensor " + "that does not have a version counter. Was it created in inference mode?"); + TORCH_CHECK(i >= 0, "Cannot set a version_counter to a value below 0: ", i); + version_counter_->version_ = i; + } + + // Inference tensor doesn't have version counter so it shouldn't be + // accessed. + uint32_t current_version() const { + TORCH_CHECK( + version_counter_, "Inference tensors do not track version counter."); + return version_counter_->version_; + } +}; + +// Forward declaration of TensorImpl needed for forward declaration of +// C10_TensorImpl_Size_Check_Dummy_Class +struct C10_API TensorImpl; + +/** + * NOTE: Some TensorImpl methods are small and not overridden in the + * PyTorch codebase itself, but may theoretically need to be + * overridden by third-party TensorImpl subclasses. This macro allows + * users that need maximum performance and don't need these extension + * points to disable them with a build-time flag. (In particular, + * XLA's XLATensorImpl currently overrides these methods, so we can't + * enable this flag by default.) + */ +#ifdef C10_DISABLE_TENSORIMPL_EXTENSIBILITY +#define TENSORIMPL_MAYBE_VIRTUAL +#else +#define TENSORIMPL_MAYBE_VIRTUAL virtual +#endif + +/** + * The low-level representation of a tensor, which contains a pointer + * to a storage (which contains the actual data) and metadata (e.g., sizes and + * strides) describing this particular view of the data as a tensor. + * + * Some basic characteristics about our in-memory representation of + * tensors: + * + * - It contains a pointer to a storage struct (Storage/StorageImpl) + * which contains the pointer to the actual data and records the + * data type and device of the view. This allows multiple tensors + * to alias the same underlying data, which allows to efficiently + * implement differing *views* on a tensor. + * + * - The tensor struct itself records view-specific metadata about + * the tensor, e.g., sizes, strides and offset into storage. + * Each view of a storage can have a different size or offset. + * + * - This class is intrusively refcounted. It is refcounted so that + * we can support prompt deallocation of large tensors; it is + * intrusively refcounted so that we can still perform reference + * counted operations on raw pointers, which is often more convenient + * when passing tensors across language boundaries. + * + * - For backwards-compatibility reasons, a tensor may be in an + * uninitialized state. A tensor may be uninitialized in the following + * two ways: + * + * - A tensor may be DTYPE UNINITIALIZED. A tensor of this + * form has an uninitialized dtype. This situation most + * frequently arises when a user writes Tensor x(CPU). The dtype + * is subsequently initialized when mutable_data() is + * invoked for the first time. + * + * - A tensor may be STORAGE UNINITIALIZED. A tensor of this form + * has non-zero size, but has a storage with a null data pointer. + * This situation most frequently arises when a user calls + * Resize() or FreeMemory(). This is because Caffe2 historically + * does lazy allocation: allocation of data doesn't occur until + * mutable_data() is invoked. A tensor with zero size is + * always storage initialized, because no allocation is necessary + * in this case. + * + * All combinations of these two uninitialized states are possible. + * Consider the following transcript in idiomatic Caffe2 API: + * + * Tensor x(CPU); // x is storage-initialized, dtype-UNINITIALIZED + * x.Resize(4); // x is storage-UNINITIALIZED, dtype-UNINITIALIZED + * x.mutable_data(); // x is storage-initialized, dtype-initialized + * x.FreeMemory(); // x is storage-UNINITIALIZED, dtype-initialized. + * + * All other fields on tensor are always initialized. In particular, + * size is always valid. (Historically, a tensor declared as Tensor x(CPU) + * also had uninitialized size, encoded as numel == -1, but we have now + * decided to default to zero size, resulting in numel == 0). + * + * Uninitialized storages MUST be uniquely owned, to keep our model + * simple. Thus, we will reject operations which could cause an + * uninitialized storage to become shared (or a shared storage to + * become uninitialized, e.g., from FreeMemory). + * + * In practice, tensors which are storage-UNINITIALIZED and + * dtype-UNINITIALIZED are *extremely* ephemeral: essentially, + * after you do a Resize(), you basically always call mutable_data() + * immediately afterwards. Most functions are not designed to + * work if given a storage-UNINITIALIZED, dtype-UNINITIALIZED tensor. + * + * We intend to eliminate all uninitialized states, so that every + * tensor is fully initialized in all fields. Please do not write new code + * that depends on these uninitialized states. + */ +struct C10_API TensorImpl : public c10::intrusive_ptr_target { + TensorImpl() = delete; + ~TensorImpl() override; + // Note [Enum ImplType] + // This enum is temporary. In the followup refactor we should + // think about how to specialize TensorImpl creation for view + // tensors. Currently we only special case its key_set_ but + // there's also potential to share version_counter_ directly + // without creating first and then override in as_view. + enum ImplType { VIEW }; + + /** + * Construct a 1-dim 0-size tensor backed by the given storage. + */ + TensorImpl( + Storage&& storage, + DispatchKeySet /*key_set*/, + const caffe2::TypeMeta data_type); + + // See Note [Enum ImplType] + TensorImpl( + ImplType /*unused*/, + Storage&& storage, + DispatchKeySet /*key_set*/, + const caffe2::TypeMeta data_type); + + /** + * Construct a 1-dim 0 size tensor that doesn't have a storage. + */ + TensorImpl( + DispatchKeySet /*key_set*/, + const caffe2::TypeMeta data_type, + std::optional device_opt); + + // Legacy constructors so I don't have to go update call sites. + // TODO: When Variable is added, delete these constructors + TensorImpl( + Storage&& storage, + DispatchKey dispatch_key, + const caffe2::TypeMeta data_type) + : TensorImpl( + std::move(storage), + DispatchKeySet(dispatch_key), + data_type) {} + TensorImpl( + DispatchKey dispatch_key, + const caffe2::TypeMeta data_type, + std::optional device_opt) + : TensorImpl(DispatchKeySet(dispatch_key), data_type, device_opt) {} + + private: + // This constructor is private, because the data_type is redundant with + // storage. Still, we pass it in separately because it's easier to write + // the initializer list if we're not worried about storage being moved out + // from under us. + TensorImpl( + Storage&& storage, + DispatchKeySet /*key_set*/, + const caffe2::TypeMeta data_type, + std::optional /*device_opt*/); + + public: + TensorImpl(const TensorImpl&) = delete; + TensorImpl& operator=(const TensorImpl&) = delete; + TensorImpl(TensorImpl&&) = delete; + TensorImpl& operator=(TensorImpl&&) = delete; + + /** + * Release (decref) storage, and any other external allocations. This + * override is for `intrusive_ptr_target` and is used to implement weak + * tensors. + */ + void release_resources() override; + + public: + /** + * Return the DispatchKeySet corresponding to this Tensor, specifying + * all of the DispatchKeys that this Tensor identifies as. This is the + * information used to dispatch operations on this tensor. + */ + DispatchKeySet key_set() const { + return key_set_; + } + + private: + [[noreturn]] void throw_cannot_call_with_symbolic(const char* meth) const; + + // NOTE: The general recipe for customizable methods is that the fastpath + // function (e.g., sizes()) does an unlikely policy test, and if doesn't + // trigger, it does the fast path implementation with no checks and going + // directly to on-TensorImpl fields. In particular, you never need to + // check ExtraMeta if the policy doesn't trigger, as non-trivial ExtraMeta + // implies the policy will always match. + // + // The default implementations of methods are "safe": they do extra tests + // to make sure the internal state is consistent no matter if you are + // doing symbolic shapes or not. If you don't want the tests, directly + // override the custom method (e.g., custom_sizes()) to do your preferred + // behavior. + + public: + /** + * Return a reference to the sizes of this tensor. This reference remains + * valid as long as the tensor is live and not resized. + */ + IntArrayRef sizes() const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomSizes))) { + return sizes_custom(); + } + return sizes_and_strides_.sizes_arrayref(); + } + + SymIntArrayRef sym_sizes() const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomSizes))) { + return sym_sizes_custom(); + } + // Sizes guaranteed to be non-negative, so unchecked cast is OK + return c10::fromIntArrayRefKnownNonNegative( + sizes_and_strides_.sizes_arrayref()); + } + + IntArrayRef sizes_default() const { + if (C10_UNLIKELY(has_symbolic_sizes_strides_)) { + throw_cannot_call_with_symbolic("sizes"); + } + return sizes_and_strides_.sizes_arrayref(); + } + + SymIntArrayRef sym_sizes_default() const { + if (has_symbolic_sizes_strides_) { + return symbolic_shape_meta().sizes_; + } else { + // Sizes guaranteed to be non-negative, so unchecked cast is OK + return c10::fromIntArrayRefKnownNonNegative(sizes_default()); + } + } + + template + ArrayRef generic_sizes() { + static_assert( + std::is_same_v || std::is_same_v, + "Only supports int64_t and c10::SymInt."); + + if constexpr (std::is_same_v) { + return sizes(); + } else { + return sym_sizes(); + } + } + + template + ArrayRef generic_strides() { + static_assert( + std::is_same_v || std::is_same_v, + "Only supports int64_t and c10::SymInt."); + + if constexpr (std::is_same_v) { + return strides(); + } else { + return sym_strides(); + } + } + + template + T generic_storage_offset() { + static_assert( + std::is_same_v || std::is_same_v, + "Only supports int64_t and c10::SymInt."); + + if constexpr (std::is_same_v) { + return storage_offset(); + } else { + return sym_storage_offset(); + } + } + + /** + * The number of elements in a tensor. + * + * WARNING: Previously, if you were using the Caffe2 API, you could + * test numel() == -1 to see if a tensor was uninitialized. This + * is no longer true; numel always accurately reports the product + * of sizes of a tensor. + */ + int64_t numel() const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomSizes))) { + return numel_custom(); + } + return numel_; + } + + c10::SymInt sym_numel() const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomSizes))) { + return sym_numel_custom(); + } + return c10::SymInt(SymInt::UNCHECKED, numel_); + } + + int64_t numel_default() const { + if (C10_UNLIKELY(has_symbolic_sizes_strides_)) { + throw_cannot_call_with_symbolic("numel"); + } + return numel_; + } + + c10::SymInt sym_numel_default() const { + if (has_symbolic_sizes_strides_) { + return symbolic_shape_meta().numel(); + } else { + return c10::SymInt(SymInt::UNCHECKED, numel_); + } + } + + /** + * Return the number of dimensions of this tensor. Note that 0-dimension + * represents a Tensor that is a Scalar, e.g., one that has a single element. + */ + int64_t dim() const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomSizes))) { + return dim_custom(); + } + return static_cast(sizes_and_strides_.size()); + } + + int64_t dim_default() const { + if (has_symbolic_sizes_strides_) { + return static_cast(symbolic_shape_meta().sizes_.size()); + } else { + return static_cast(sizes_and_strides_.size()); + } + } + + /** + * Return the offset in number of elements into the storage that this + * tensor points to. Most tensors have storage_offset() == 0, but, + * for example, an index into a tensor will have a non-zero storage_offset(). + * + * WARNING: This is NOT computed in bytes. + */ + int64_t storage_offset() const { + // TODO: maybe this should be toggled by strides + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomSizes))) { + return storage_offset_custom(); + } + return storage_offset_; + } + + c10::SymInt sym_storage_offset() const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomSizes))) { + return sym_storage_offset_custom(); + } + return c10::SymInt(SymInt::UNCHECKED, storage_offset_); + } + + int64_t storage_offset_default() const { + if (C10_UNLIKELY(has_symbolic_sizes_strides_)) { + throw_cannot_call_with_symbolic("storage_offset"); + } + return storage_offset_; + } + + c10::SymInt sym_storage_offset_default() const { + if (has_symbolic_sizes_strides_) { + return symbolic_shape_meta().storage_offset_; + } else { + return c10::SymInt(SymInt::UNCHECKED, storage_offset_); + } + } + + /** + * Return a reference to the strides of this tensor. This reference remains + * valid as long as the tensor is live and not restrided. + */ + IntArrayRef strides() const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomStrides))) { + return strides_custom(); + } + return sizes_and_strides_.strides_arrayref(); + } + + c10::SymIntArrayRef sym_strides() const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomStrides))) { + return sym_strides_custom(); + } + return c10::fromIntArrayRefKnownNonNegative(strides_default()); + } + + IntArrayRef strides_default() const { + if (C10_UNLIKELY(has_symbolic_sizes_strides_)) { + throw_cannot_call_with_symbolic("strides"); + } + return sizes_and_strides_.strides_arrayref(); + } + + c10::SymIntArrayRef sym_strides_default() const { + if (has_symbolic_sizes_strides_) { + return symbolic_shape_meta().strides_; + } else { + return c10::fromIntArrayRefKnownNonNegative(strides_default()); + } + } + + c10::SymBool sym_is_contiguous( + at::MemoryFormat memory_format = at::MemoryFormat::Contiguous) const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomStrides))) { + return sym_is_contiguous_custom(memory_format); + } + return sym_is_contiguous_default(memory_format); + } + + template + T is_contiguous_default_impl(at::MemoryFormat memory_format) const { + if (!has_symbolic_sizes_strides_) { + if (memory_format == at::MemoryFormat::ChannelsLast) { + return is_channels_last_contiguous_; + } else if (memory_format == at::MemoryFormat::ChannelsLast3d) { + return is_channels_last_3d_contiguous_; + } + return is_contiguous_; + } + + // Handle dynamic shapes. + const auto& symbolic = symbolic_shape_meta().is_contiguous(memory_format); + + if constexpr (std::is_same_v) { + return symbolic.guard_bool(__FILE__, __LINE__); + } else { + return symbolic; + } + } + + bool is_contiguous_default(at::MemoryFormat memory_format) const { + return is_contiguous_default_impl(memory_format); + } + + c10::SymBool sym_is_contiguous_default(at::MemoryFormat memory_format) const { + return is_contiguous_default_impl(memory_format); + } + + /** + * Whether or not a tensor is laid out in contiguous memory. + * + * Tensors with non-trivial strides are not contiguous. See + * compute_contiguous() for the exact definition of whether or not + * a tensor is contiguous or not. + */ + bool is_contiguous( + at::MemoryFormat memory_format = at::MemoryFormat::Contiguous) const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomStrides))) { + return is_contiguous_custom(memory_format); + } + return is_contiguous_default(memory_format); + } + + bool is_strides_like_default(at::MemoryFormat memory_format) const { + if (has_symbolic_sizes_strides_) { + if (memory_format == at::MemoryFormat::ChannelsLast) { + return symbolic_shape_meta().is_channels_last().guard_bool( + __FILE__, __LINE__); + } else if (memory_format == at::MemoryFormat::ChannelsLast3d) { + return symbolic_shape_meta().is_channels_last_3d().guard_bool( + __FILE__, __LINE__); + } else { + return false; + } + } + + if (memory_format == at::MemoryFormat::ChannelsLast) { + return is_channels_last_; + } else if (memory_format == at::MemoryFormat::ChannelsLast3d) { + return is_channels_last_3d_; + } else { + return false; + } + } + + SymBool sym_is_non_overlapping_and_dense_default() const { + if (has_symbolic_sizes_strides_) { + return symbolic_shape_meta().is_non_overlapping_and_dense(); + } else { + return is_non_overlapping_and_dense_; + } + } + + bool is_non_overlapping_and_dense_default() const { + if (has_symbolic_sizes_strides_) { + return sym_is_non_overlapping_and_dense_default().guard_bool( + __FILE__, __LINE__); + } else { + return is_non_overlapping_and_dense_; + } + } + + // NB: these dim accessor functions don't have _default(), as you can use + // sizes_default/strides_default + /** + * Return the size of a tensor at some dimension, wrapping the dimension if + * necessary. + * + * NOTE: if you know wrapping is unnecessary, do sizes()[d] instead; it will + * be faster + */ + int64_t size(int64_t d) const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomSizes))) { + return size_custom(d); + } + d = maybe_wrap_dim(d, dim(), /*wrap_scalar=*/false); + return sizes_and_strides_.size_at_unchecked(d); + } + + c10::SymInt sym_size(int64_t d) const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomSizes))) { + return sym_size_custom(d); + } + d = maybe_wrap_dim(d, dim(), /*wrap_scalar=*/false); + const auto sizes = this->sym_sizes(); + return sizes[d]; + } + + /** + * Return the stride of a tensor at some dimension, wrapping the dimension + * if necessary. + * + * NOTE: if you know wrapping is unnecessary, do sizes()[d] instead; it will + * be faster + */ + int64_t stride(int64_t d) const { + d = maybe_wrap_dim(d, dim(), false); + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomStrides))) { + // TODO: provide stride_custom, symmetrically with size_custom. + // There is presently no user for it; only NestedTensor is using + // size_custom overrideability + return strides_custom()[d]; // unchecked (maybe_wrap_dim enforces bounds) + } + // Intentionally don't call default, which also handles symbolic + return sizes_and_strides_.stride_at_unchecked(d); + } + + enum class SizesStridesPolicy : uint8_t { + // Default behavior, e.g., dense tensor. + // + // Can override: nothing + Default = 0, + // Customizable strides behavior, e.g., sparse tensor, + // mkldnn tensor. + // + // Can override: strides(), is_contiguous() + CustomStrides = 1, + // Customizable sizes behavior, e.g., nested tensor + // + // Can override: strides(), is_contiguous(), sizes(), dim(), numel() + CustomSizes = 2 + }; + + protected: + inline bool matches_policy(SizesStridesPolicy policy) const { + return sizes_strides_policy_ >= static_cast(policy); + } + + inline bool matches_custom(SizesStridesPolicy policy) const { + return custom_sizes_strides_ >= static_cast(policy); + } + + inline bool matches_python_custom(SizesStridesPolicy policy) const { + auto r = python_custom_sizes_strides_ >= static_cast(policy); + if (r) { + TORCH_INTERNAL_ASSERT(is_python_dispatch()) + } + return r; + } + + /** + * Customization points for the functions above. sizes_strides_policy_ + * must be set to enable these. + * + * NB: dim is overridable separately from sizes because it is possible + * for a tensor to have rank, but not well defined sizes. + */ + // sizes_strides_policy_ >= CustomStrides + + virtual bool is_strides_like_custom(at::MemoryFormat memory_format) const; + + virtual c10::SymBool sym_is_non_overlapping_and_dense_custom() const; + + bool is_non_overlapping_and_dense_custom() const { + return sym_is_non_overlapping_and_dense_custom().guard_bool( + __FILE__, __LINE__); + } + + virtual c10::SymBool sym_is_contiguous_custom( + at::MemoryFormat memory_format) const; + + bool is_contiguous_custom(at::MemoryFormat memory_format) const { + return sym_is_contiguous_custom(memory_format) + .guard_bool(__FILE__, __LINE__); + } + + // sizes_strides_policy_ >= CustomSizes + // Currently this method only exists to be overwritten by subclasses such as + // NestedTensorImpl. + virtual int64_t size_custom(int64_t d) const { + // TODO: We could add support to Python dispatch here. + // TODO: We could call into aten::size.int instead of + // sizes_custom()[d] and enable use of the dispatcher. + d = maybe_wrap_dim(d, dim(), /*wrap_scalar=*/false); + return sizes_custom()[d]; // unchecked (maybe_wrap_dim enforces bounds) + } + + virtual c10::SymInt sym_size_custom(int64_t d) const { + // TODO: We could add support to Python dispatch here. + // TODO: We could call into aten::size.int instead of + // sym_sizes_custom()[d] and enable use of the dispatcher. + d = maybe_wrap_dim(d, dim(), /*wrap_scalar=*/false); + return sym_sizes_custom()[d]; // unchecked (maybe_wrap_dim enforces bounds) + } + + virtual IntArrayRef sizes_custom() const; + virtual IntArrayRef strides_custom() const; + virtual int64_t numel_custom() const; + virtual int64_t storage_offset_custom() const; + virtual int64_t dim_custom() const; + virtual Device device_custom() const; + virtual Layout layout_custom() const; + + virtual c10::SymIntArrayRef sym_sizes_custom() const; + virtual c10::SymIntArrayRef sym_strides_custom() const; + virtual c10::SymInt sym_numel_custom() const; + virtual c10::SymInt sym_storage_offset_custom() const; + + public: +/** + * True if this tensor has storage. See storage() for details. + */ +#ifdef DEBUG + // Allow subclasses to check that their storage_ is never getting set in debug + // builds. + virtual +#else + TENSORIMPL_MAYBE_VIRTUAL +#endif + bool + has_storage() const +// NOTE: we devirtualize this because it arguably shouldn't be an +// error just to ask subclasses if they have storage. +// This used to throw for most subclasses, but OpaqueTensorImpl +// wanted it to successfully return false, so we went ahead and made +// it a non-error. +#ifdef C10_DISABLE_TENSORIMPL_EXTENSIBILITY + { + return storage_; + } +#else + ; +#endif + + /** + * Return the underlying storage of a Tensor. Multiple tensors may share + * a single storage. A Storage is an impoverished, Tensor-like class + * which supports far less operations than Tensor. + * + * Avoid using this method if possible; try to use only Tensor APIs to perform + * operations. + */ + TENSORIMPL_MAYBE_VIRTUAL const Storage& storage() const { + if (C10_UNLIKELY(storage_access_should_throw_)) { + throw_storage_access_error(); + } + return storage_; + } + + /** + * Return the underlying storage, unsafely assuming this is a basic strided + * tensor. In cases where `storage` access would throw, this returns a + * default-constructed Storage. + */ + inline const Storage& unsafe_storage() const { + return storage_; + } + + bool unique_version() const { + return version_counter_.unique(); + } + + protected: + virtual Layout layout_impl() const { + TORCH_CHECK( + false, "layout_impl is only implemented for TensorImpl subclasses."); + } + + public: + // Whether a tensor is sparse COO or not. + bool is_sparse() const { + // NB: This method is not virtual and avoid dispatches for performance + // reasons. + return key_set_.has_all(c10::sparse_ks); + } + + // Whether a tensor is sparse CSR or not. + bool is_sparse_csr() const { + return layout() == kSparseCsr; + } + + // Whether a tensor is sparse CSR/CSC/BSR/BSC or not. + bool is_sparse_compressed() const { + return key_set_.has_all(c10::sparse_csr_ks); + } + + bool is_quantized() const { + // NB: This method is not virtual and avoid dispatches for performance + // reasons. + constexpr auto quantized_ks = DispatchKeySet(DispatchKey::Quantized); + return key_set_.has_all(quantized_ks); + } + + bool is_meta() const { + // NB: This method is not virtual and avoid dispatches for performance + // reasons. + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_meta(); + } + return device_opt_.has_value() && device_opt_->type() == kMeta; + } + + bool is_cpu() const { + // NB: This method is not virtual and avoid dispatches for performance + // reasons. + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_cpu(); + } + // Note: we cannot rely on dispatch keys to determine the device type + // of a tensor, because "wrapper" tensors (like FunctionalTensorWrapper) + // don't include backend dispatch keys. + return device_opt_.has_value() && device_opt_->type() == kCPU; + } + + bool is_cuda() const { + // NB: This method is not virtual and avoid dispatches for performance + // reasons. + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_cuda(); + } + return device_opt_.has_value() && device_opt_->type() == kCUDA; + } + + bool is_xpu() const { + // NB: This method is not virtual and avoid dispatches for performance + // reasons. + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_xpu(); + } + return device_opt_.has_value() && device_opt_->type() == kXPU; + } + + bool is_ipu() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_ipu(); + } + return device_opt_.has_value() && device_opt_->type() == kIPU; + } + + bool is_xla() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_xla(); + } + return device_opt_.has_value() && device_opt_->type() == kXLA; + } + + bool is_mtia() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_mtia(); + } + return device_opt_.has_value() && device_opt_->type() == kMTIA; + } + + bool is_hpu() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_hpu(); + } + return device_opt_.has_value() && device_opt_->type() == kHPU; + } + + bool is_lazy() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_lazy(); + } + return device_opt_.has_value() && device_opt_->type() == kLazy; + } + + bool is_hip() const { + // NB: This method is not virtual and avoid dispatches for performance + // reasons. + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_hip(); + } + return device_opt_.has_value() && device_opt_->type() == kHIP; + } + + bool is_ve() const { + // NB: This method is not virtual and avoid dispatches for performance + // reasons. + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_ve(); + } + return device_opt_.has_value() && device_opt_->type() == kVE; + } + + bool is_privateuseone() const { + // NB: This method is not virtual and avoid dispatches for performance + // reasons. + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_privateuseone(); + } + return device_opt_.has_value() && device_opt_->type() == kPrivateUse1; + } + + bool is_mkldnn() const { + return key_set_.has_all(c10::mkldnn_ks); + } + + bool is_vulkan() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_vulkan(); + } + return device_opt_.has_value() && device_opt_->type() == kVulkan; + } + + bool is_metal() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_metal(); + } + return device_opt_.has_value() && device_opt_->type() == kMetal; + } + + bool is_mps() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_mps(); + } + return device_opt_.has_value() && device_opt_->type() == kMPS; + } + + bool is_maia() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom().is_maia(); + } + return device_opt_.has_value() && device_opt_->type() == kMAIA; + } + + bool is_nested() const { + return key_set_.has(DispatchKey::NestedTensor); + } + + // TODO: remove this once we don't automatically enabled Autograd dispatch + // keys + // in TensorImpl constructor. + // DON'T USE THIS API!! It's only created for testing purpose in + // file aten/src/ATen/core/boxing/impl/test_helpers.h + void remove_autograd_key() { + key_set_ = key_set_ - autograd_dispatch_keyset; + } + + // Inference tensor doesn't have autograd or ADInplaceOrView key. + // Invariant: + // Inference tensor has version_counter_.enabled() == false + bool is_inference() { + bool no_ADInplaceOrView = !key_set_.has_any(c10::inplace_or_view_ks); + bool no_Autograd = !key_set_.has_any(c10::autograd_dispatch_keyset); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + no_ADInplaceOrView == no_Autograd, + "ADInplaceOrView and Autograd keys must be on/off at the same time."); + return no_ADInplaceOrView && no_Autograd; + } + + DeviceIndex get_device() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom().index(); + } + return device_default().index(); + } + + Device device() const { + if (C10_UNLIKELY(device_policy_)) { + return device_custom(); + } + return device_default(); + } + + protected: + c10::Device device_default() const { + TORCH_CHECK(device_opt_.has_value(), "tensor does not have a device"); + // See NOTE [std::optional operator usage in CUDA] + return *device_opt_; + } + + public: + Layout layout() const { + if (C10_UNLIKELY(layout_policy_)) { + return layout_custom(); + } + + // NB: This method is not virtual and avoid dispatches for perf. + // strided is also the most common layout type, so we check for + // strided case first. + // This keyset must also be kept in sync with the logic in + // is_sparse() / is_sparse_csr() / is_mkldnn() + constexpr auto sparse_and_sparsecsr_and_mkldnn_ks = + c10::sparse_ks | c10::sparse_csr_ks | c10::mkldnn_ks; + if (!key_set_.has_any(sparse_and_sparsecsr_and_mkldnn_ks)) { + return kStrided; + } else if (is_sparse()) { + return kSparse; + } else if (is_sparse_compressed()) { + // Typically, the tensor dispatch keys define the tensor layout + // uniquely. This allows using non-virtual layout method for + // better performance. However, when tensor's layout depends, + // say, on tensor attributes, one must use this execution path + // where the corresponding tensor impl class overwrites virtual + // layout_impl() method. + // + // TODO: implement layout() as native function/method so that + // __torch_dispatch__ users will be able to redefine the + // layout() method. + return layout_impl(); + } else { + TORCH_INTERNAL_ASSERT( + is_mkldnn(), "There is an error in the layout calculation logic."); + return kMkldnn; + } + } + + /** + * True if a tensor was auto-wrapped from a C++ or Python number. + * For example, when you write 't + 2', 2 is auto-wrapped into a Tensor + * with `is_wrapped_number_` set to true. + * + * Wrapped numbers do not participate in the result type computation for + * mixed-type operations if there are any Tensors that are not wrapped + * numbers. This is useful, because we want 't + 2' to work with + * any type of tensor, not just LongTensor (which is what integers + * in Python represent). + * + * Otherwise, they behave like their non-wrapped equivalents. + * See [Result type computation] in TensorIterator.h. + * + * Why did we opt for wrapped numbers, as opposed to just having + * an extra function add(Tensor, Scalar)? This helps greatly reduce + * the amount of code we have to write for add, when actually + * a Tensor-Scalar addition is really just a Tensor-Tensor + * addition when the RHS is 0-dim (except for promotion behavior.) + */ + bool is_wrapped_number() const { + return is_wrapped_number_; + } + + /** + * Set whether or not a tensor was auto-wrapped from a C++ or Python + * number. You probably don't want to call this, unless you are + * writing binding code. + */ + void set_wrapped_number(bool value) { + TORCH_INTERNAL_ASSERT(dim() == 0); + is_wrapped_number_ = value; + } + + /** + * Returns true if Tensor supports as_strided and as_strided_backward. + * This is used in autograd to perform inplace update on view Tensors. + * See Note [View + Inplace update for base tensor] and + * [View + Inplace update for view tensor] for details. + * Note this method only returns true for XLA backend, where it + * simulates strided Tensor to support most view ops, but it cannot + * fully support general `as_strided` case. + * It can be expanded as needed in the future, e.g sparse Tensor. + */ + inline bool support_as_strided() const { + if (is_nested()) { + return false; + } + if (key_set_.has(DispatchKey::Functionalize)) { + return false; + } + return device().supports_as_strided(); + } + + // ~~~~~ Autograd API ~~~~~ + // Some methods below are defined in TensorImpl.cpp because Tensor is an + // incomplete type. + + /** + * Set whether or not a tensor requires gradient. + */ + void set_requires_grad(bool requires_grad); + + /** + * True if a tensor requires gradient. Tensors which require gradient + * have history tracked for any operations performed on them, so that + * we can automatically differentiate back to them. A tensor that + * requires gradient and has no history is a "leaf" tensor, which we + * accumulate gradients into. + */ + bool requires_grad() const; + + /** + * Return a mutable reference to the gradient. This is conventionally + * used as `t.grad() = x` to set a gradient to a completely new tensor. + */ + at::Tensor& mutable_grad(); + + /** + * Return the accumulated gradient of a tensor. This gradient is written + * into when performing backwards, when this tensor is a leaf tensor. + */ + const at::Tensor& grad() const; + + /** + * Whether or not the imaginary part of the tensor should be negated + */ + inline bool is_conj() const { + constexpr auto conjugate_ks = DispatchKeySet(DispatchKey::Conjugate); + return key_set_.has_all(conjugate_ks); + } + + /** + * Set whether or not to take the conjugate of the tensor (flip the imaginary + * bit). + */ + void _set_conj(bool value) { + if (value) { + key_set_ = key_set_.add(DispatchKey::Conjugate); + TORCH_INTERNAL_ASSERT(isComplexType(typeMetaToScalarType(dtype()))); + } else { + key_set_ = key_set_.remove(DispatchKey::Conjugate); + } + } + + /** + * XXX: do not use, private api! + * Update the backend component related keys to the backend component + * corresponding to this device. + */ + void _change_backend_component_keys(c10::Device device); + + /** + * Whether or not the tensor is a zerotensor + */ + inline bool _is_zerotensor() const { + constexpr auto zerotensor_ks = DispatchKeySet(DispatchKey::ZeroTensor); + return key_set_.has_all(zerotensor_ks); + } + + /** + Set whether or not the tensor is a zero tensor + */ + void _set_zero(bool value) { + if (value) { + TORCH_INTERNAL_ASSERT( + false, + "Please call `torch._efficientzerotensor` if you want to create a tensor with no storage."); + } else { + key_set_ = key_set_.remove(DispatchKey::ZeroTensor); + } + } + + /** + * Whether or not the tensor should be negated + */ + inline bool is_neg() const { + constexpr auto negative_ks = DispatchKeySet(DispatchKey::Negative); + return key_set_.has_all(negative_ks); + } + + /** + * Set whether or not to take the conjugate of the tensor (flip the imaginary + * bit). + */ + void _set_neg(bool value) { + if (value) { + key_set_ = key_set_.add(DispatchKey::Negative); + } else { + key_set_ = key_set_.remove(DispatchKey::Negative); + } + } + + /** + * Return the accumulated gradient of a tensor. This gradient is computed + * using forward mode AD. + * + * This is an internal API that should never be used by end users. + * + * The API is as follows: + * - "level" allows to specify the level of forward AD nesting for which the + * gradient should be returned. Note that since levels are not fully + * supported yet, this argument should be 0. See documentation for + * torch::autograd::enter_dual_level for more details about forward AD + * nesting. + * - "self" should represent the Tensor whose forward grad is accessed. It + * is required when dealing with view. + */ + const at::Tensor& _fw_grad(uint64_t level, const at::TensorBase& self) const; + + /** + * Sets the forward gradient for this Tensor. + * The given Tensor might not be used directly and its content will be copied. + * + * This is an internal API that should never be used by end users. + * + * The API is as follows: + * - "new_grad" is a Tensor containing the new value of the gradient that + * should be set + * - "self" should represent the Tensor whose forward grad is accessed. It + * is required when dealing with view. + * - "level" allows to specify the level of forward AD nesting for which the + * gradient should be set. Note that since levels are not fully supported + * yet, this argument should be 0. See documentation for + * torch::autograd::enter_dual_level for more details about forward AD + * nesting. + * - "is_inplace_op" is a boolean flag that tells if this gradient was + * generated by an inplace operation or an out of place one. This allows + * better error checking. + */ + void _set_fw_grad( + const at::TensorBase& new_grad, + const at::TensorBase& self, + uint64_t level, + bool is_inplace_op); + + /** + * Return a typed data pointer to the actual data which this tensor refers to. + * This checks that the requested type (from the template parameter) matches + * the internal type of the tensor. + * + * It is invalid to call data() on a dtype-uninitialized tensor, even if + * the size is 0. + * + * WARNING: If a tensor is not contiguous, you MUST use strides when + * performing index calculations to determine the location of elements in + * the tensor. We recommend using 'TensorAccessor' to handle this computation + * for you; this class is available from 'Tensor'. + */ + template + const T* data_dtype_initialized() const { + return data_dtype_initialized_impl( + [this] { return static_cast(storage_.data()); }); + } + + /** + * Return a mutable typed data pointer to the actual data which this + * tensor refers to. This checks that the requested type (from the + * template parameter) matches the internal type of the tensor. + * + * It is invalid to call data() on a dtype-uninitialized tensor, even if + * the size is 0. + * + * WARNING: If a tensor is not contiguous, you MUST use strides when + * performing index calculations to determine the location of elements in + * the tensor. We recommend using 'TensorAccessor' to handle this computation + * for you; this class is available from 'Tensor'. + */ + template + T* mutable_data_dtype_initialized() { + return data_dtype_initialized_impl( + [this] { return static_cast(storage_.mutable_data()); }); + } + + private: + // Shared implementation of data_dtype_initialized() and + // mutable_data_dtype_initialized(). + template + T* data_dtype_initialized_impl(const Func& get_data) const { + TORCH_CHECK( + data_type_.Match>(), + "Tensor type mismatch, caller expects elements to be ", + caffe2::TypeMeta::TypeName>(), + ", while tensor contains ", + data_type_.name(), + ". "); + return data_ptr_impl_impl(get_data); + } + + public: + /** + * More efficient helper for Tensor::data_ptr(). Like data(), but + * does not do a type check. Unlike the untemplated data(), does + * check has_storage() and storage_initialized(). + */ + template + inline const T* data_ptr_impl() const { + return data_ptr_impl_impl( + [this] { return static_cast(storage_.data()); }); + } + + /** + * More efficient helper for Tensor::data_ptr(). Like data(), but + * does not do a type check. Unlike the untemplated data(), does + * check has_storage() and storage_initialized(). + */ + template + inline T* mutable_data_ptr_impl() { + return data_ptr_impl_impl( + [this] { return static_cast(storage_.mutable_data()); }); + } + + private: + // Shared implementation of mutable_data_ptr_impl() and the future + // mutable_data_ptr_impl(). + template + __ubsan_ignore_pointer_overflow__ T* data_ptr_impl_impl( + const Func& get_data) const { + if (C10_UNLIKELY(!has_storage())) { + throw_data_ptr_access_error(); + } + TORCH_CHECK( + storage_initialized(), + "The tensor has a non-zero number of elements, but its data is not allocated yet.\n" + "If you're using torch.compile/export/fx, it is likely that we are erroneously " + "tracing into a custom kernel. To fix this, please wrap the custom kernel into " + "an opaque custom op. Please see the following for details: " + "https://pytorch.org/tutorials/advanced/custom_ops_landing_page.html\n" + "If you're using Caffe2, Caffe2 uses a lazy allocation, so you will need to call " + "mutable_data() or raw_mutable_data() to actually allocate memory."); + // Caller does the type check. + // Note: storage_offset_ can be non-null even for zero-elements tensors + // (for example if created as `torch.empty(5)[10:]`) that triggers + // applying non-zero offset to null pointer in UBSan + return get_data() + storage_offset_; + } + + public: + /** + * Return a const void* data pointer to the actual data which this + * tensor refers to. + * + * It is invalid to call data() on a dtype-uninitialized tensor, even if the + * size is 0. + * + * WARNING: The data pointed to by this tensor may not contiguous; do NOT + * assume that itemsize() * numel() is sufficient to compute the bytes that + * can be validly read from this tensor. + */ + inline const void* data() const { + return data_impl( + [this] { return static_cast(storage_.data()); }); + } + + /** + * Return a void* data pointer to the actual data which this tensor refers to. + * + * It is invalid to call mutable_data() on a dtype-uninitialized + * tensor, even if the size is 0. + * + * WARNING: The data pointed to by this tensor may not contiguous; do NOT + * assume that itemsize() * numel() is sufficient to compute the bytes that + * can be validly read from this tensor. + */ + inline void* mutable_data() { + return data_impl( + [this] { return static_cast(storage_.mutable_data()); }); + } + + private: + /// Shared implementation of data() and mutable_data(). + /// + /// get_data must return a byte-addressed pointer, e.g. char*, + /// std::byte const*, etc. + template + Void* data_impl(const Func& get_data) const { + if (C10_UNLIKELY(!has_storage())) { + throw_data_ptr_access_error(); + } + TORCH_CHECK( + dtype_initialized(), + "Cannot access data pointer of Tensor that doesn't have initialized dtype " + "(e.g., caffe2::Tensor x(CPU), prior to calling mutable_data() on x)"); + auto* data = get_data(); + static_assert( + sizeof(*data) == 1, "get_data must return a byte-addressed pointer."); + // Computing an offset into an empty tensor would be UB, since an empty + // tensor's storage will be nullptr, and adding a nonzero offset to nullptr + // is UB. So we skip the offset computation in this case. + if (is_empty()) { + return nullptr; + } + return data + data_type_.itemsize() * storage_offset_; + } + + public: + /** + * Returns the TypeMeta of a tensor, which describes what data type + * it is (e.g., int, float, ...) + */ + const caffe2::TypeMeta dtype() const { + return data_type_; + } + + /** + * Return the size of a single element of this tensor in bytes. + */ + size_t itemsize() const { + TORCH_CHECK( + dtype_initialized(), + "Cannot report itemsize of Tensor that doesn't have initialized dtype " + "(e.g., caffe2::Tensor x(CPU), prior to calling mutable_data() on x)"); + return data_type_.itemsize(); + } + + void set_backend_meta(intrusive_ptr backend_meta) { + get_extra_meta().backend_meta_ = std::move(backend_meta); + } + + c10::BackendMeta* get_backend_meta() { + if (!extra_meta_) { + return nullptr; + } + return extra_meta_->backend_meta_.get(); + } + + intrusive_ptr get_backend_meta_intrusive_ptr() const { + if (!extra_meta_) { + return nullptr; + } + return extra_meta_->backend_meta_; + } + + void release_storage_and_set_meta_custom_data_ptr_error_msg_( + std::optional s) { + storage_ = {}; + set_storage_access_should_throw(); + get_extra_meta().custom_data_ptr_error_msg_ = s; + get_extra_meta().custom_storage_error_msg_ = std::move(s); + } + + protected: + /** + * Returns the human-readable name of the actual type of this object (e.g., + * TensorImpl, BatchedTensorImpl, etc.). Used for error messages. + */ + virtual const char* tensorimpl_type_name() const { + return "TensorImpl"; + } + + private: + [[noreturn]] void throw_storage_access_error() const; + [[noreturn]] void throw_data_ptr_access_error() const; + + ExtraMeta& get_extra_meta() { + if (!extra_meta_) { + extra_meta_ = std::make_unique(); + } + return *extra_meta_; + } + + c10::SymbolicShapeMeta& symbolic_shape_meta() { + TORCH_INTERNAL_ASSERT(extra_meta_ && extra_meta_->symbolic_shape_meta_); + return *extra_meta_->symbolic_shape_meta_; + } + + const c10::SymbolicShapeMeta& symbolic_shape_meta() const { + TORCH_INTERNAL_ASSERT(extra_meta_ && extra_meta_->symbolic_shape_meta_); + return *extra_meta_->symbolic_shape_meta_; + } + + public: + /** + * True if a tensor has no elements (e.g., numel() == 0). + */ + inline bool is_empty() const { + return numel() == 0; + } + + // if we are going to use sym sizes, we should be setting sym strides at the + // same time, otherwise it's very easy to misuse this API + void set_sizes_and_strides( + c10::SymIntArrayRef sizes, + c10::SymIntArrayRef strides, + std::optional storage_offset = std::nullopt); + // This is renamed to avoid breaking overload BC + void generic_set_sizes_contiguous(c10::SymIntArrayRef sizes); + void generic_set_sizes_contiguous(c10::IntArrayRef sizes) { + set_sizes_contiguous(sizes); + } + + /** + * Change the size at some dimension. This DOES NOT update strides; + * thus, most changes to size will not preserve contiguity. You probably + * also want to call set_stride() when you call this. + * + * TODO: This should be jettisoned in favor of `set_sizes_and_strides`, + * which is harder to misuse. + */ + virtual void set_size(int64_t dim, int64_t new_size) { + TORCH_CHECK( + allow_tensor_metadata_change(), + "set_size ", + err_msg_tensor_metadata_change_not_allowed); + TORCH_CHECK( + !matches_policy(SizesStridesPolicy::CustomSizes), + "set_size() called on tensor with dynamic shapes or customized size behavior") + sizes_and_strides_.size_at(dim) = new_size; + refresh_numel(); + refresh_contiguous(); + } + + /** + * Change the stride at some dimension. + * + * TODO: This should be jettisoned in favor of `set_sizes_and_strides`, + * which is harder to misuse. + */ + virtual void set_stride(int64_t dim, int64_t new_stride) { + TORCH_CHECK( + allow_tensor_metadata_change(), + "set_stride ", + err_msg_tensor_metadata_change_not_allowed); + TORCH_CHECK( + !has_symbolic_sizes_strides_, + "set_stride() called on tensor with symbolic shape") + sizes_and_strides_.stride_at_unchecked(dim) = new_stride; + refresh_contiguous(); + } + + /** + * Set the offset into the storage of this tensor. + * + * WARNING: This does NOT check if the tensor is in bounds for the new + * location at the storage; the caller is responsible for checking this + * (and resizing if necessary.) + */ + virtual void set_storage_offset(int64_t storage_offset) { + TORCH_CHECK( + allow_tensor_metadata_change(), + "set_storage_offset ", + err_msg_tensor_metadata_change_not_allowed); + // TODO: this should probably consult policy + TORCH_CHECK( + !has_symbolic_sizes_strides_, + "set_storage_offset() called on tensor with symbolic shape") + storage_offset_ = storage_offset; + } + + /** + * Like set_sizes_and_strides but assumes contiguous strides. + * + * WARNING: This function does not check if the requested + * sizes/strides are in bounds for the storage that is allocated; + * this is the responsibility of the caller + */ + void set_sizes_contiguous(IntArrayRef new_size) { + TORCH_CHECK( + allow_tensor_metadata_change(), + "set_sizes_contiguous ", + err_msg_tensor_metadata_change_not_allowed); + TORCH_CHECK( + !matches_policy(SizesStridesPolicy::CustomStrides), + "tried to directly modify sizes for customized tensor"); + sizes_and_strides_.set_sizes(new_size); + + refresh_numel(); + empty_tensor_restride( + MemoryFormat::Contiguous); // calls refresh_contiguous() + } + + C10_ALWAYS_INLINE const impl::SizesAndStrides& sizes_and_strides() { + return sizes_and_strides_; + } + + /** + * Set the sizes and strides of a tensor. + * + * WARNING: This function does not check if the requested + * sizes/strides are in bounds for the storage that is allocated; + * this is the responsibility of the caller + */ + void set_sizes_and_strides( + IntArrayRef new_size, + IntArrayRef new_stride, + std::optional storage_offset = std::nullopt) { + TORCH_CHECK( + allow_tensor_metadata_change(), + "set_sizes_and_strides ", + err_msg_tensor_metadata_change_not_allowed); + TORCH_CHECK( + !has_symbolic_sizes_strides_, + "set_sizes_and_strides() called on tensor with symbolic shape") + TORCH_CHECK( + new_size.size() == new_stride.size(), + "dimensionality of sizes (", + new_size.size(), + ") must match dimensionality of strides (", + new_stride.size(), + ")"); + const auto new_dim = new_size.size(); + bool overflowed = false; + sizes_and_strides_.set_sizes(new_size); + + if (new_dim > 0) { + for (size_t dim = new_dim - 1;; dim--) { + if (new_stride[dim] >= 0) { + sizes_and_strides_.stride_at_unchecked(dim) = new_stride[dim]; + } else { + // XXX: This behavior is surprising and may need to be removed to + // support negative strides. Some pytorch functions rely on it: + // for example, torch.cat (run TestTorch.test_cat_empty). + if (dim == new_dim - 1) { + sizes_and_strides_.stride_at_unchecked(dim) = 1; + } else { + // Keep stride monotonically increasing to match NumPy. + overflowed |= c10::mul_overflows( + sizes_and_strides_.stride_at_unchecked(dim + 1), + std::max( + sizes_and_strides_.size_at_unchecked(dim + 1), 1), + std::addressof(sizes_and_strides_.stride_at_unchecked(dim))); + } + } + if (dim == 0) + break; + } + TORCH_CHECK(!overflowed, "Stride calculation overflowed"); + } + + refresh_numel(); + refresh_contiguous(); + + if (storage_offset.has_value()) { + storage_offset_ = *storage_offset; + } + } + + /** + * Set whether a tensor allows changes to its metadata (e.g. sizes / strides / + * storage / storage_offset). See NOTE [ Metadata Change for a Detached Tensor + * ] for details. + */ + void set_allow_tensor_metadata_change(bool value [[maybe_unused]]) { + // TODO: at some point, we should kill this field completely. + allow_tensor_metadata_change_ = true; + } + + /** + * True if a tensor allows changes to its metadata (e.g. sizes / strides / + * storage / storage_offset). See NOTE [ Metadata Change for a Detached Tensor + * ] for details. + */ + bool allow_tensor_metadata_change() const { + return allow_tensor_metadata_change_; + } + + /** + * Set the pointer to autograd metadata. + */ + void set_autograd_meta( + std::unique_ptr autograd_meta); + + /** + * Return the pointer to autograd metadata. May return nullptr if the + * tensor does not track gradients. + */ + c10::AutogradMetaInterface* autograd_meta() const; + + /** + * Set the pointer to named tensor metadata. + */ + void set_named_tensor_meta( + std::unique_ptr named_tensor_meta) { + TORCH_WARN_ONCE( + "Named tensors and all their associated APIs are an experimental feature ", + "and subject to change. Please do not use them for anything important ", + "until they are released as stable."); +#ifdef DEBUG + if (named_tensor_meta) { + TORCH_INTERNAL_ASSERT(named_tensor_meta->slow_dim() == dim()); + } +#endif + if (named_tensor_meta) { + get_extra_meta().named_tensor_meta_ = std::move(named_tensor_meta); + key_set_ = key_set_.add(DispatchKey::Named); + } else { + if (extra_meta_) { + extra_meta_->named_tensor_meta_ = nullptr; + } + key_set_ = key_set_.remove(DispatchKey::Named); + } + } + + void set_python_dispatch(bool k) { + if (k) { + key_set_ = key_set_.add(c10::python_ks); + } else { + key_set_ = key_set_ - c10::python_ks; + } + } + + bool is_python_dispatch() const { + return key_set_.has_all(c10::python_ks); + } + + /** + * Return the pointer to named tensor metadata. + */ + const c10::NamedTensorMetaInterface* named_tensor_meta() const { + if (!extra_meta_) { + return nullptr; + } + return extra_meta_->named_tensor_meta_.get(); + } + + c10::NamedTensorMetaInterface* named_tensor_meta() { + if (!extra_meta_) { + return nullptr; + } + return extra_meta_->named_tensor_meta_.get(); + } + + bool has_named_tensor_meta() const { + if (!extra_meta_) { + return false; + } + return extra_meta_->named_tensor_meta_ != nullptr; + } + + // NOTE [ TensorImpl Shallow-Copying ] + // + // TensorImpl shallow-copying is used when we want to have two Variables share + // the same tensor metadata (e.g. sizes / strides / storage pointer / + // storage_offset), but each with a different autograd history. Example call + // sites: + // + // 1. `var_detached = var.detach()` uses `shallow_copy_and_detach()` to create + // `var_detached` that shares the same tensor metadata with `var`, but with a + // completely new autograd history. + // 2. `var.set_data(tensor)` uses `shallow_copy_from()` to copy tensor + // metadata from `tensor` into `var`, while keeping `var`'s original + // AutogradMeta. + // + // Functions that shallow-copy a TensorImpl (such as + // `shallow_copy_and_detach()` / `shallow_copy_from()` / + // `copy_tensor_metadata()`) copy the tensor metadata fields (e.g. sizes / + // strides / storage pointer / storage_offset) by value. However, the + // following fields are not copied: + // + // 1. the AutogradMeta pointer, because it is unique for each Variable. + // 2. the version counter, because the destination TensorImpl's version + // counter is either set to the passed-in `version_counter` (in + // `shallow_copy_and_detach()` and `copy_tensor_metadata()`), or it is kept + // intact (in `shallow_copy_from()`). See NOTE [ Version Counter Sharing ] for + // details. + // + // In `shallow_copy_and_detach()` and `copy_tensor_metadata()`, the passed-in + // `allow_tensor_metadata_change` determines whether the TensorImpl + // shallow-copy allows changes to its metadata (e.g. sizes / strides / storage + // / storage_offset). See NOTE [ Metadata Change for a Detached Tensor ] for + // details. + // + // In `shallow_copy_from()`, we don't check the destination TensorImpl's + // `allow_tensor_metadata_change_`, because `shallow_copy_from()` is used for + // implementing functions such as `var.set_data(tensor)`, which changes + // `var`'s tensor metadata and expects its `allow_tensor_metadata_change_` to + // be ignored. + + /** + * One TensorImpl can be copied to another TensorImpl if they have the same + * DispatchKeySet. The only two special cases (for legacy reason) are: + * CPU is compatible with CUDA and SparseCPU is + * compatible with SparseCUDA. + */ + inline bool has_compatible_shallow_copy_type(DispatchKeySet from) { + auto is_dense = [](DispatchKeySet ts) { + constexpr auto dense_backends = DispatchKeySet( + {BackendComponent::CPUBit, + BackendComponent::CUDABit, + BackendComponent::MPSBit, + BackendComponent::HIPBit, + BackendComponent::XPUBit, + BackendComponent::HPUBit, + BackendComponent::MTIABit}); + constexpr auto dense_k = DispatchKeySet(DispatchKey::Dense); + return ts.has_any(dense_k) && ts.has_any(dense_backends); + }; + auto is_sparse = [](DispatchKeySet ts) { + constexpr auto sparse_backends = DispatchKeySet( + {BackendComponent::CPUBit, + BackendComponent::CUDABit, + BackendComponent::MPSBit, + BackendComponent::HIPBit, + BackendComponent::XPUBit}); + constexpr auto sparse_k = DispatchKeySet(DispatchKey::Sparse); + return ts.has_any(sparse_k) && ts.has_any(sparse_backends); + }; + auto is_sparse_compressed = [](DispatchKeySet ts) { + constexpr auto sparse_compressed_k = + DispatchKeySet(DispatchKey::SparseCsr); + return ts.has_any(sparse_compressed_k); + }; + return (key_set_ == from) || (is_dense(key_set_) && is_dense(from)) || + (is_sparse(key_set_) && is_sparse(from)) || + (is_sparse_compressed(key_set_) && is_sparse_compressed(from)); + ; + } + + private: + template + c10::intrusive_ptr shallow_copy_and_detach_core( + VariableVersion&& version_counter, + bool allow_tensor_metadata_change) const; + + public: + /** + * Return a TensorImpl that is a shallow-copy of this TensorImpl. + * + * For usage of `version_counter` and `allow_tensor_metadata_change`, + * see NOTE [ TensorImpl Shallow-Copying ]. + */ + virtual c10::intrusive_ptr shallow_copy_and_detach( + const c10::VariableVersion& version_counter, + bool allow_tensor_metadata_change) const; + + /** + * Return a TensorImpl that is a shallow-copy of this TensorImpl. + * + * For usage of `version_counter` and `allow_tensor_metadata_change`, + * see NOTE [ TensorImpl Shallow-Copying ]. + */ + virtual c10::intrusive_ptr shallow_copy_and_detach( + c10::VariableVersion&& version_counter, + bool allow_tensor_metadata_change) const; + + /** + * Shallow-copies data from another TensorImpl into this TensorImpl. + * + * For why this function doesn't check this TensorImpl's + * `allow_tensor_metadata_change_`, see NOTE [ TensorImpl Shallow-Copying ]. + */ + virtual void shallow_copy_from(const c10::intrusive_ptr& impl) { + copy_tensor_metadata( + /*src_impl=*/impl.get(), + /*dest_impl=*/this, + /*version_counter=*/version_counter(), + /*allow_tensor_metadata_change=*/allow_tensor_metadata_change()); + } + + // Inference tensor doesn't have version counter, + // set_version_counter is no-op for them. + void set_version_counter(const c10::VariableVersion& version_counter) { + TORCH_CHECK( + !(is_inference() && version_counter.enabled()), + "Cannot set version_counter for inference tensor"); + version_counter_ = version_counter; + } + + void set_version_counter(c10::VariableVersion&& version_counter) { + TORCH_CHECK( + !(is_inference() && version_counter.enabled()), + "Cannot set version_counter for inference tensor"); + version_counter_ = std::move(version_counter); + } + + const c10::VariableVersion& version_counter() const noexcept { + return version_counter_; + } + + void bump_version() { + version_counter_.bump(); + } + + impl::PyObjectSlot* pyobj_slot() { + return &pyobj_slot_; + } + + const impl::PyObjectSlot* pyobj_slot() const { + return &pyobj_slot_; + } + + void incref_pyobject() const noexcept override final; + + void decref_pyobject() const noexcept override final; + + bool try_incref_pyobject() const noexcept override final; + + private: + // See NOTE [std::optional operator usage in CUDA] + // We probably don't want to expose this publicly until + // the note is addressed. + std::optional device_opt() const { + return device_opt_; + } + + public: + /** + * The device type of a Tensor, e.g., DeviceType::CPU or DeviceType::CUDA. + */ + DeviceType device_type() const { + // TODO: A useful internal assert would be to show that device_opt_ is null + // only if you are an undefined tensor + TORCH_CHECK( + device_opt_.has_value(), + "device_type cannot be run on undefined Tensor"); + // See NOTE [std::optional operator usage in CUDA] + return (*device_opt_).type(); + } + + /** + * @brief Extends the outer-most dimension of this tensor by num elements, + * preserving the existing data. + * + * The underlying data may be reallocated in order to accommodate the new + * elements, in which case this tensors' capacity is grown at a factor of + * growthPct. This ensures that Extend runs on an amortized O(1) time + * complexity. + * + * This op is auto-asynchronous if the underlying device (CUDA) supports it. + */ + void Extend(int64_t num, float growthPct); + + /** + * @brief Reserve space for the underlying tensor. + * + * This must be called after Resize(), since we only specify the first + * dimension This does not copy over the old data to the newly allocated space + */ + void ReserveSpace(int64_t outer_dim); + + /** + * @brief Resizes a tensor. + * + * Resize takes in a vector of ints specifying the dimensions of the tensor. + * You can pass in an empty vector to specify that it is a scalar (i.e. + * containing one single item). + * + * The underlying storage may be deleted after calling Resize: if the new + * shape leads to a different number of items in the tensor, the old memory + * is deleted and new memory will be allocated next time you call + * mutable_data(). However, if the shape is different but the total number of + * items is the same, the underlying storage is kept. + * + * This method respects caffe2_keep_on_shrink. Consult the internal logic + * of this method to see exactly under what circumstances this flag matters. + */ + template + void Resize(Ts... dim_source) { + bool size_changed = SetDims(dim_source...); + if (size_changed) { + HandleResize(); + } + } + + template + void Resize(const std::vector& dim_source) { + Resize(ArrayRef(dim_source)); + } + + /** + * Resizes the tensor without touching underlying storage. + * This requires the total size of the tensor to remains constant. + */ + void Reshape(const std::vector& dims); + + /** + * Release whatever memory the tensor was holding but keep size and type + * information. Subsequent call to mutable_data will trigger new memory + * allocation. + */ + void FreeMemory(); + + /** + * @brief Shares the data with another tensor. + * + * To share data between two tensors, the sizes of the two tensors must be + * equal already. The reason we do not implicitly do a Resize to make the two + * tensors have the same shape is that we want to allow tensors of different + * shapes but the same number of items to still be able to share data. This + * allows one to e.g. have a n-dimensional Tensor and a flattened version + * sharing the same underlying storage. + * + * The source tensor should already have its data allocated. + */ + // To be deprecated + void ShareData(const TensorImpl& src); + + void ShareExternalPointer( + DataPtr&& data_ptr, + const caffe2::TypeMeta data_type, + size_t size_bytes); + + /** + * Returns a mutable raw pointer of the underlying storage. Since we will need + * to know the type of the data for allocation, a TypeMeta object is passed in + * to specify the necessary information. This is conceptually equivalent of + * calling mutable_data() where the TypeMeta parameter meta is derived from + * the type T. This function differs from mutable_data() in the sense that + * the type T can be specified during runtime via the TypeMeta object. + * + * If the existing data does not match the desired type, it will be deleted + * and a new storage will be created. + */ + inline void* raw_mutable_data(const caffe2::TypeMeta& meta) { + // For 0-size tensors it's fine to return any pointer (including nullptr) + if (data_type_ == meta && storage_initialized()) { + return static_cast( + static_cast(storage_.mutable_data()) + + storage_offset_ * meta.itemsize()); + } else { + bool had_special_dtor = data_type_.placementDelete() != nullptr; + storage_offset_ = 0; + data_type_ = meta; + // NB: device is not changed + + // We can reuse the existing buffer if the current data does not have + // a special destructor and the new data doesn't have a special + // constructor. + if (numel_ == 0 || + (meta.placementNew() == nullptr && !had_special_dtor && + (storage_.nbytes() >= (numel_ * data_type_.itemsize())))) { + TORCH_INTERNAL_ASSERT( + storage_offset_ == 0); // because we just reallocated + return storage_.mutable_data(); + } + Allocator* allocator = storage_.allocator(); + // Storage might have nullptr allocator in rare cases, for example, if + // an external memory segment has been wrapped with Tensor and we don't + // know how to reallocate it. However, in order to preserve legacy C2 + // behavior, we allow reallocating the memory using default allocator. + if (allocator == nullptr) { + allocator = GetAllocator(storage_.device_type()); + } + if (meta.placementNew()) { + // For types that need placement new, we will call it, as well as + // making sure that when the data is freed, it calls the right + // destruction procedure. + auto size = numel_; + auto dtor = data_type_.placementDelete(); + auto data_ptr = allocator->allocate(numel_ * data_type_.itemsize()); + storage_.set_data_ptr_noswap(PlacementDeleteContext::makeDataPtr( + std::move(data_ptr), dtor, size, storage_.device())); + data_type_.placementNew()(storage_.mutable_data(), numel_); + } else { + // For fundamental type, new and delete is easier. + storage_.set_data_ptr_noswap( + allocator->allocate(numel_ * data_type_.itemsize())); + } + storage_.set_nbytes(numel_ * data_type_.itemsize()); + TORCH_INTERNAL_ASSERT( + storage_offset_ == 0); // because we just reallocated + device_opt_ = storage_.device(); + return storage_.mutable_data(); + } + } + + /** + * Returns a typed pointer of the underlying storage. + * + * For fundamental types, we reuse possible existing storage if there + * is sufficient capacity. + */ + template + inline T* mutable_data() { + if (storage_initialized() && data_type_.Match()) { + return static_cast(storage_.mutable_data()) + storage_offset_; + } + // Check it here statically - otherwise TypeMeta would throw the runtime + // error in attempt to invoke TypeMeta::ctor() + static_assert( + std::is_default_constructible_v, + "Tensor can't hold non-default-constructable types"); + return static_cast(raw_mutable_data(caffe2::TypeMeta::Make())); + } + + /** + * True if a tensor is storage initialized. A tensor may become + * storage UNINITIALIZED after a Resize() or FreeMemory() + */ + bool storage_initialized() const { + TORCH_CHECK( + has_storage(), + "cannot call storage_initialized on tensor that does not have storage"); + return storage_.data() || numel_ == 0; + } + + /** + * True if a tensor is dtype initialized. A tensor allocated with + * Caffe2-style constructors is dtype uninitialized until the + * first time mutable_data() is called. + */ + bool dtype_initialized() const noexcept { + return data_type_ != caffe2::TypeMeta(); + } + + void set_storage_keep_dtype(at::Storage storage) { + TORCH_CHECK( + allow_tensor_metadata_change(), + "set_storage ", + err_msg_tensor_metadata_change_not_allowed); + storage_ = std::move(storage); + device_opt_ = storage_.device(); + } + + void set_storage_and_dtype( + at::Storage storage, + const caffe2::TypeMeta data_type) { + set_storage_keep_dtype(std::move(storage)); + data_type_ = data_type; + } + + void empty_tensor_restride_symint(MemoryFormat memory_format); + + /** + * Set the strides of the tensor to match memory_format + * + * WARNING: This function doesn't rearrange data and assumes tensor is a + * memory contiguous + */ + void empty_tensor_restride(MemoryFormat memory_format) { + if (has_symbolic_sizes_strides_) { + empty_tensor_restride_symint(memory_format); + return; + } +#ifdef DEBUG + TORCH_INTERNAL_ASSERT( + compute_numel() == numel_, + "If you are seeing this error, that means empty_tensor_restride was " + "called before setting correct numel"); +#endif + switch (memory_format) { + case MemoryFormat::Contiguous: { + // dim_ is a virtual call, don't repeat it + const auto dim_ = dim(); + sizes_and_strides_.resize(dim_); + if (dim_ > 0) { + bool overflowed = false; + const auto last_idx = dim_ - 1; + sizes_and_strides_.stride_at_unchecked(last_idx) = 1; + for (auto i = last_idx - 1; i >= 0; --i) { + overflowed |= c10::mul_overflows( + sizes_and_strides_.stride_at_unchecked(i + 1), + std::max( + sizes_and_strides_.size_at_unchecked(i + 1), 1), + std::addressof(sizes_and_strides_.stride_at_unchecked(i))); + } + TORCH_CHECK(!overflowed, "Stride calculation overflowed"); + } + break; + } + case MemoryFormat::ChannelsLast: { + TORCH_CHECK( + dim() == 4, "required rank 4 tensor to use channels_last format"); + set_sizes_and_strides(sizes(), get_channels_last_strides_2d(sizes())); + break; + } + case MemoryFormat::ChannelsLast3d: { + TORCH_CHECK( + dim() == 5, + "required rank 5 tensor to use channels_last_3d format"); + set_sizes_and_strides(sizes(), get_channels_last_strides_3d(sizes())); + break; + } + case MemoryFormat::Preserve: + TORCH_CHECK(false, "unsupported memory format ", memory_format); + // Cleaning warning messages, no need to break as TORCH_CHECK(false) + // terminates flow. + // break; + case MemoryFormat::NumOptions: + TORCH_INTERNAL_ASSERT(false, "invalid memory format ", memory_format); + } + // recompute contiguous flag, as currently NHWC/NCHW flags are not mutually + // exclusive see #24090 + refresh_contiguous(); + } + + bool is_strides_like(at::MemoryFormat memory_format) const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomStrides))) { + return is_strides_like_custom(memory_format); + } + return is_strides_like_default(memory_format); + } + + bool is_strides_like_channels_last() const { + return is_strides_like(at::MemoryFormat::ChannelsLast); + } + + bool is_strides_like_channels_last_3d() const { + return is_strides_like(at::MemoryFormat::ChannelsLast3d); + } + + bool is_non_overlapping_and_dense_or_false() const { + return sym_is_non_overlapping_and_dense().guard_or_false( + __FILE__, __LINE__); + } + + bool is_non_overlapping_and_dense() const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomStrides))) { + return is_non_overlapping_and_dense_custom(); + } + return is_non_overlapping_and_dense_default(); + } + + SymBool sym_is_non_overlapping_and_dense() const { + if (C10_UNLIKELY(matches_policy(SizesStridesPolicy::CustomStrides))) { + return sym_is_non_overlapping_and_dense_custom(); + } + return sym_is_non_overlapping_and_dense_default(); + } + + // if this returns true, then it is guaranteed that this tensor has symbolic + // sizes/strides + bool has_symbolic_sizes_strides() const { + return has_symbolic_sizes_strides_; + } + + private: + void HandleResize(); + + // The Caffe2 Resize() method supports being called both as Resize({2,2}) as + // well as variadic with Resize(2, 2). These overloads provide all of the + // supported calling configurations, while being overloads (and not templates) + // so that implicit conversions still work. + // + // SetDims on ArrayRef is internally implemented as a template, so we can + // handle both ArrayRefs of different types (there are some uses of + // Resize in Caffe2 which pass in int, not int64_t.) + + template < + typename T, + typename = typename std::enable_if_t>> + bool SetDimsTemplate(ArrayRef src) { + TORCH_CHECK( + !has_symbolic_sizes_strides_, + "SetDims() called on tensor with symbolic shape") + + auto old_numel = numel_; + sizes_and_strides_.resize(src.size()); + int64_t new_numel = 1; + for (const auto i : c10::irange(src.size())) { + new_numel *= src[i]; + sizes_and_strides_.size_at_unchecked(i) = src[i]; + } + numel_ = new_numel; + empty_tensor_restride(MemoryFormat::Contiguous); + return numel_ != old_numel; + } + + bool SetDims(ArrayRef s) { + return SetDimsTemplate(s); + } + + bool SetDims(ArrayRef s) { + return SetDimsTemplate(s); + } + + bool SetDims(ArrayRef s) { + return SetDimsTemplate(s); + } + + bool SetDims() { + return SetDims(IntArrayRef{}); + } + + bool SetDims(const int64_t d0) { + return SetDims(IntArrayRef{d0}); + } + + bool SetDims(const int64_t d0, const int64_t d1) { + return SetDims(IntArrayRef{d0, d1}); + } + + bool SetDims(const int64_t d0, const int64_t d1, const int64_t d2) { + return SetDims(IntArrayRef{d0, d1, d2}); + } + + bool SetDims( + const int64_t d0, + const int64_t d1, + const int64_t d2, + const int64_t d3) { + return SetDims(IntArrayRef{d0, d1, d2, d3}); + } + + /** + * Compute the number of elements based on the sizes of a tensor. + */ + // NB: This is ONLY called when sizes_and_strides_ is used directly; if + // we are virtualizing, then numel calls are virtualized as well, and this + // should never get called + int64_t compute_numel() const { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!has_symbolic_sizes_strides_); +#if C10_HAS_BUILTIN_OVERFLOW() && !defined(C10_MOBILE) + // Use overflow checks if supported by the compiler + return safe_compute_numel(); +#else + return c10::multiply_integers(sizes_and_strides_.sizes_arrayref()); +#endif + } + + /** + * Compute the number of elements based on the sizes of a + * tensor. Catches integer overflow that may occur when a tensor + * using a sparse layout has multiple dimensions with large sizes. + */ + int64_t safe_compute_numel() const { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!has_symbolic_sizes_strides_); + uint64_t n = 1; + bool overflows = + c10::safe_multiplies_u64(sizes_and_strides_.sizes_arrayref(), &n); + constexpr auto numel_max = std::min( + static_cast(std::numeric_limits::max()), + static_cast(std::numeric_limits::max())); + + overflows |= (n > numel_max); + TORCH_CHECK(!overflows, "numel: integer multiplication overflow"); + return static_cast(n); + } + + /** + * Compute whether or not a tensor is contiguous based on the sizes and + * strides of a tensor. + */ + bool compute_contiguous() const; + + bool compute_channels_last_contiguous_2d() const; + + bool compute_channels_last_contiguous_3d() const; + + bool compute_strides_like_channels_last_2d() const; + + bool compute_strides_like_channels_last_3d() const; + + bool compute_non_overlapping_and_dense() const; + + protected: + /** + * Recompute the cached numel of a tensor. Call this if you modify + * sizes. + * + * For tensors with sparse layouts, use safe_refresh_numel() instead + * because it will catch integer overflow that may occur for tensors + * with sparse layouts and large dimensions. + * + * NB: We may uselessly recompute cached numel even in situations where + * it is completely never used (e.g., if CustomSizes for Python). However, + * we still must keep it up to date in case the Python overload + * returns None (in which case we will consult the field here). This also + * implies that sizes/strides will never be complete garbage; in the + * very worst case scenario, it will reflect a 1-dim zero size tensor. + */ + void refresh_numel() { + if (has_symbolic_sizes_strides_) { + symbolic_shape_meta().refresh_numel(); + } else { + numel_ = compute_numel(); + } + } + + /** + * Recompute the cached numel of a tensor. Call this if you modify + * sizes. Use only for tensors with sparse layouts because only + * sparse tensor are likely to have sizes that may lead to integer + * overflow when computing numel. + */ + void safe_refresh_numel() { + if (has_symbolic_sizes_strides_) { + // NB: sym numel is done with symbolic integers, which handle overflow + // checking + symbolic_shape_meta().refresh_numel(); + } else { + numel_ = safe_compute_numel(); + } + } + + private: + void _set_is_contiguous(bool b) { + is_contiguous_ = b; + } + + void _set_is_channels_last_contiguous(bool b) { + is_channels_last_contiguous_ = b; + } + + void _set_is_channels_last_3d_contiguous(bool b) { + is_channels_last_3d_contiguous_ = b; + } + + void _set_is_channels_last(bool b) { + is_channels_last_ = b; + } + + void _set_is_channels_last_3d(bool b) { + is_channels_last_3d_ = b; + } + + void _set_is_non_overlapping_and_dense(bool b) { + is_non_overlapping_and_dense_ = b; + } + + // These are little wrappers over the real compute_ functions that + // can make use of other contiguity fields to short circuit. + + bool compute_is_non_overlapping_and_dense_dim4() { + return is_contiguous_ || is_channels_last_contiguous_ || + compute_non_overlapping_and_dense(); + } + + bool compute_channels_last_contiguous_3d_dim5() { + return !is_channels_last_contiguous_ && + compute_channels_last_contiguous_3d(); + } + + bool compute_channels_last_2d_dim5() { + return !is_channels_last_3d_contiguous_ && + compute_strides_like_channels_last_2d(); + } + + bool compute_channels_last_3d_dim5() { + return !is_channels_last_ && compute_strides_like_channels_last_3d(); + } + + bool compute_is_non_overlapping_and_dense_dim5() { + return is_contiguous_ || is_channels_last_contiguous_ || + is_channels_last_3d_contiguous_ || compute_non_overlapping_and_dense(); + } + + bool compute_is_non_overlapping_and_dense_anydim() { + return is_contiguous_ || compute_non_overlapping_and_dense(); + } + + void _refresh_contiguous() { + // Note: + // Dim 0, 1, 2 will never be a channels last 2d/3d format + // Dim 3+ is possibly be a channels last 2d format (Dim 4 only at this + // point) Dim 4+ is possibly be a channels last 3d format (Dim 5 only at + // this point) + switch (dim()) { + case 4: { + _set_is_contiguous(compute_contiguous()); + _set_is_channels_last_contiguous(compute_channels_last_contiguous_2d()); + _set_is_channels_last_3d_contiguous(false); + _set_is_channels_last(compute_strides_like_channels_last_2d()); + _set_is_channels_last_3d(false); + _set_is_non_overlapping_and_dense( + compute_is_non_overlapping_and_dense_dim4()); + break; + } + case 5: { + _set_is_contiguous(compute_contiguous()); + _set_is_channels_last_contiguous(compute_channels_last_contiguous_2d()); + _set_is_channels_last_3d_contiguous( + compute_channels_last_contiguous_3d_dim5()); + _set_is_channels_last(compute_channels_last_2d_dim5()); + _set_is_channels_last_3d(compute_channels_last_3d_dim5()); + _set_is_non_overlapping_and_dense( + compute_is_non_overlapping_and_dense_dim5()); + break; + } + default: + // is_channels_last_ and is_channels_last_3d_ are suggested + // memory_format. Being channels_last_contiguous doesn't necessarily + // mean the tensor is strided like channels_last: for strides on channel + // dimension could suggest desired memory_layout, but it doesn't affect + // memory storage + _set_is_contiguous(compute_contiguous()); + _set_is_channels_last_contiguous(false); + _set_is_channels_last_3d_contiguous(false); + _set_is_channels_last(false); + _set_is_channels_last_3d(false); + _set_is_non_overlapping_and_dense( + compute_is_non_overlapping_and_dense_anydim()); + break; + } + } + + protected: + /** + * Recompute the cached contiguity of a tensor. Call this if you modify sizes + * or strides. + */ + void refresh_contiguous() { + if (has_symbolic_sizes_strides_) { + symbolic_shape_meta().refresh_contiguous(); + } else { + _refresh_contiguous(); + } + } + + /** + * Copy the tensor metadata fields (e.g. sizes / strides / storage pointer / + * storage_offset) from one TensorImpl to another TensorImpl. + * + * For usage of `version_counter` and `allow_tensor_metadata_change`, see NOTE + * [ TensorImpl Shallow-Copying ]. + */ + static void copy_tensor_metadata( + const TensorImpl* src_impl, + TensorImpl* dest_impl, + const c10::VariableVersion& version_counter, + bool allow_tensor_metadata_change); + + /** + * Copy the tensor metadata fields (e.g. sizes / strides / storage pointer / + * storage_offset) from one TensorImpl to another TensorImpl. + * + * For usage of `version_counter` and `allow_tensor_metadata_change`, see NOTE + * [ TensorImpl Shallow-Copying ]. + */ + static void copy_tensor_metadata( + const TensorImpl* src_impl, + TensorImpl* dest_impl, + c10::VariableVersion&& version_counter, + bool allow_tensor_metadata_change); + + private: + static void copy_tensor_metadata_except_version_counter( + const TensorImpl* src_impl, + TensorImpl* dest_impl, + bool allow_tensor_metadata_change); + + protected: + // Error message to show when the user tries to change tensor metadata on + // Tensor created from .data or .detach(). + // + // See NOTE [ Metadata Change for a Detached Tensor ] for details. + static const char* const err_msg_tensor_metadata_change_not_allowed; + + static void copy_generic_tensor_metadata( + const TensorImpl* src_impl, + TensorImpl* dest_impl); + + public: + void set_storage_access_should_throw() { + storage_access_should_throw_ = true; + } + + public: + void set_custom_sizes_strides(SizesStridesPolicy policy) { + custom_sizes_strides_ = static_cast(policy); + refresh_sizes_strides_policy(); + } + + void set_python_custom_sizes_strides(SizesStridesPolicy policy) { + python_custom_sizes_strides_ = static_cast(policy); + refresh_sizes_strides_policy(); + } + + void set_custom_device(bool custom_device) { + custom_device_ = custom_device; + refresh_device_policy(); + } + + void set_custom_layout(bool custom_layout) { + custom_layout_ = custom_layout; + refresh_layout_policy(); + } + + void set_python_custom_device(bool custom_device) { + python_custom_device_ = custom_device; + refresh_device_policy(); + } + + void set_python_custom_layout(bool custom_layout) { + python_custom_layout_ = custom_layout; + refresh_layout_policy(); + } + + protected: + void refresh_sizes_strides_policy() { + if (has_symbolic_sizes_strides_) { + sizes_strides_policy_ = + static_cast(SizesStridesPolicy::CustomSizes); + } else { + sizes_strides_policy_ = + std::max(custom_sizes_strides_, python_custom_sizes_strides_); + } + } + + void refresh_device_policy() { + device_policy_ = custom_device_ || python_custom_device_; + } + + void refresh_layout_policy() { + layout_policy_ = custom_layout_ || python_custom_layout_; + } + + protected: + Storage storage_; + + private: + // This pointer points to an AutogradMeta struct that stores autograd-specific + // fields (such as grad_ / grad_fn_ / grad_accumulator_). This pointer always + // has unique ownership (meaning only one TensorImpl can own it at a time). + // + // autograd_meta_ can be nullptr, as an optimization. When this occurs, it is + // equivalent to having an autograd_meta_ pointing to a default constructed + // AutogradMeta; intuitively, tensors which don't require grad will have this + // field set to null. + // + // This means accessors on autograd_meta_ have to be careful to test if they + // got a nullptr, and handle default behavior appropriately in that case. + // + // Note that we don't enforce the invariant that if the AutogradMeta is + // default constructed, it is nullptr (to do this, we'd have to continuously + // check if an AutogradMeta became, by mutation, equal to the default + // constructed form. (This might be useful, but it seems rare enough that + // a requires_grad=True variable will turn back into the requires_grad=False + // version.) So there are three representable states: + // + // 1. autograd_meta_ == nullptr + // 2. autograd_meta_ is default constructed (semantically, same as (1)) + // 3. autograd_meta_ has nontrivial information content + // + std::unique_ptr autograd_meta_ = nullptr; + + protected: + std::unique_ptr extra_meta_ = nullptr; + + c10::VariableVersion version_counter_; + + impl::PyObjectSlot pyobj_slot_; + + c10::impl::SizesAndStrides sizes_and_strides_; + + int64_t storage_offset_ = 0; + // If sizes and strides are empty, the numel is 1!! However, most of the + // time, we will immediately set sizes to {0} and reset numel to 0. + // (Can't do that in the default initializers, because there's no way to + // spell "allocate a one-element array" for strides_). + int64_t numel_ = 1; + + // INVARIANT: When storage is non-null, this type meta must + // agree with the type meta in storage + caffe2::TypeMeta data_type_; + + // NOTE [std::optional operator usage in CUDA] + // Our optional definition doesn't compile in .cu file if `value()` or + // `operator->` are used. Instead, we always use `operator*`. + // See https://github.com/pytorch/pytorch/issues/18496 for more info. + // If this is too burdensome to maintain, we can just + // manually implement this with an additional bool. + + // INVARIANT: When storage is non-null, this Device must + // agree with the type meta in storage. + // + // INVARIANT: device_opt_ is only nullopt for undefined tensors + // (which do not have a device.) + std::optional device_opt_; + + // default member initializers for bit-fields only available with -std=c++2a + // or -std=gnu++2a + inline void init_bitfields() { + is_contiguous_ = true; + is_channels_last_ = false; + is_channels_last_contiguous_ = false; + is_channels_last_3d_ = false; + is_channels_last_3d_contiguous_ = false; + is_non_overlapping_and_dense_ = true; + is_wrapped_number_ = false; + allow_tensor_metadata_change_ = true; + reserved_ = false; + sizes_strides_policy_ = static_cast(SizesStridesPolicy::Default); + custom_sizes_strides_ = static_cast(SizesStridesPolicy::Default); + python_custom_sizes_strides_ = + static_cast(SizesStridesPolicy::Default); + python_custom_device_ = false; + python_custom_layout_ = false; + custom_device_ = false; + custom_layout_ = false; + device_policy_ = false; + layout_policy_ = false; + storage_access_should_throw_ = false; + has_symbolic_sizes_strides_ = false; + } + + // Tensor is contiguous + bool is_contiguous_ : 1; + + // Tensor is a subclass that does not permit storage access. + bool storage_access_should_throw_ : 1; + + // Tensor is stored in the channels last 2d memory format, when dimensions + // order is (N)CHW and C-strides < W-strides < H-strides (< N-strides) + // (If size of any dimension is equal to 1, this dimension strides value + // is not taken into account). + bool is_channels_last_ : 1; + + // Channels last contiguous tensor is channel last tensor which occupies + // contiguous memory block. + bool is_channels_last_contiguous_ : 1; + + // Tensor is stored in the channels last 3d memory format, when dimensions + // order is (N)CDHW and C-strides < W-strides < H-strides < D - strides (< + // N-strides) (If size of any dimension is equal to 1, this dimension strides + // value is not taken into account). + bool is_channels_last_3d_ : 1; + + // Channels last 3d contiguous tensor is channel last 3d tensor which occupies + // contiguous memory block. + bool is_channels_last_3d_contiguous_ : 1; + + // Dense tensor is the tensor that store values in a contiguous block of + // memory. Non-overlapping tensor is the tensor in which elements occupy + // individual non-repetitive memory. + bool is_non_overlapping_and_dense_ : 1; + + bool is_wrapped_number_ : 1; + + // NOTE [ Metadata Change for a Detached Tensor ] + // + // Normally, a user is allowed to change the tensor metadata + // (e.g. sizes / strides / storage / storage_offset) of a tensor. + // However, if the tensor is created by `t1_detached = t1.data` in Python + // or `t1_detached = t1.detach()` in Python/C++, those changes to the + // tensor metadata of `t1_detached` will not be propagated back to the + // original tensor `t1`. In order to make such changes explicitly illegal, + // we created the `allow_tensor_metadata_change_` flag, to prevent users + // from changing metadata of the detached tensor and expecting the original + // tensor to also be updated. + // + // NOTE: For a full list of tensor metadata fields, please see + // `copy_tensor_metadata()` in TensorImpl and its subclasses to find + // which fields are copied by value. + bool allow_tensor_metadata_change_ : 1; + + // we decide to keep reserved_ and it will + // live in Tensor after the split + // The logic is that if Extend() or ReserveSpace() were ever called, + // then subsequent Resize()s will not free up Storage. + bool reserved_ : 1; + + // Call _custom() virtual methods for + // strides()/is_contiguous()/sizes()/dim()/numel() + // This is a combination of sizes_strides_custom_dispatch_ + // and has_symbolic_sizes_strides_ + uint8_t sizes_strides_policy_ : 2; + + // Whether or not sizes_and_strides_ contains a symbolic value. + bool has_symbolic_sizes_strides_ : 1; + + // Call _custom() virtual method for + // strides()/is_contiguous()/sizes()/dim()/numel() + uint8_t custom_sizes_strides_ : 2; + + // Combo of custom_ and python_custom_ + bool device_policy_ : 1; + bool layout_policy_ : 1; + + // Call _custom() virtual method for device() + bool custom_device_ : 1; + + // Call _custom() virtual method for layout() + bool custom_layout_ : 1; + + // Call into Python for + // strides()/is_contiguous()/sizes()/dim()/numel() + uint8_t python_custom_sizes_strides_ : 2; + + // Call into Python for device() + bool python_custom_device_ : 1; + + // Call into Python for layout() + bool python_custom_layout_ : 1; + + // The set of DispatchKeys which describe this tensor. NB: this + // does NOT include Autograd (historically, it did, but + // not anymore!) + // + // INVARIANT: extra_meta_->named_tensor_meta_ != nullptr <==> + // key_set_.has(DispatchKey::Named) + DispatchKeySet key_set_; + + private: + // C10_TensorImpl_Size_Check_Dummy_Class needs to be friends with + // TensorImpl so it can inspect the size of private fields + template < + size_t cplusplus, + size_t clang_ver_major, + size_t gcc_ver, + size_t gcc_ver_minor, + size_t nvcc, + size_t cuda_version, + size_t cuda_version_major, + size_t ptr_size> + friend class C10_TensorImpl_Size_Check_Dummy_Class; +}; + +namespace detail { + +#ifndef C10_MOBILE +template +struct TargetTraits< + T, + std::enable_if_t>>> { + static constexpr bool can_have_pyobject = true; +}; +#endif + +} // namespace detail + +// Note [TensorImpl size constraints] +// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +// Changed the size of TensorImpl? If the size went down, good for +// you! Adjust the documentation below and the expected size. +// Did it go up? Read on... +// +// Struct size matters. In some production systems at Facebook, we have +// 400M live tensors during a training run. Do the math: every 64-bit +// word you add to Tensor is an extra 3.2 gigabytes in RAM. +// +// If you are a Facebook employee, you can check if the run in question +// has tipped you over the point using the command here: +// https://fburl.com/q5enpv98 +// +// For reference, we OOMed at 160 bytes (20 words) per TensorImpl. +// This is not counting overhead from strides out-of-line allocation and +// StorageImpl space and this is from before we inlined sizes and strides +// directly into TensorImpl as SmallVectors. +// +// Our memory usage on 32-bit systems is suboptimal, but we're not checking +// for it at the moment (to help avoid rage inducing cycles when the +// 32-bit number is wrong). +// +// Current breakdown: +// +// vtable pointer +// strong refcount TODO: pack these into one word +// weak refcount +// storage pointer +// autograd metadata pointer +// named tensor metadata pointer +// version counter pointer +// PyObjectSlot +// SizesAndStrides size/pointer +// SizesAndStrides sizes (pre-allocated 0) +// SizesAndStrides sizes (pre-allocated 1) +// SizesAndStrides sizes (pre-allocated 2) +// SizesAndStrides sizes (pre-allocated 3) +// SizesAndStrides sizes (pre-allocated 4) +// SizesAndStrides strides (pre-allocated 0) +// SizesAndStrides strides (pre-allocated 1) +// SizesAndStrides strides (pre-allocated 2) +// SizesAndStrides strides (pre-allocated 3) +// SizesAndStrides strides (pre-allocated 4) +// storage offset +// numel +// data type, device, is_contiguous, storage_access_should_throw_, bitfields +// DispatchKeySet +// + +// Various preprocessor macros we use to check that the +// TensorImpl size hasn't changed unexpectedly. We undef +// these later. +#ifndef __NVCC__ +#define C10_NVCC 0 +#else +#define C10_NVCC __NVCC__ +#endif + +#ifndef __CUDA_VER_MAJOR__ +#define C10_CUDA_VERSION_MAJOR 0 +#else +#define C10_CUDA_VERSION_MAJOR __CUDA_VER_MAJOR__ +#endif + +#ifndef CUDA_VERSION +#define C10_CUDA_VERSION 0 +#else +#define C10_CUDA_VERSION CUDA_VERSION +#endif + +#ifndef __clang_major__ +#define C10_CLANG_MAJOR_VERSION 0 +#else +#define C10_CLANG_MAJOR_VERSION __clang_major__ +#endif + +#ifndef __GNUC__ +#define C10_GCC_VERSION 0 +#else +#define C10_GCC_VERSION __GNUC__ +#endif + +#ifndef __GNUC_MINOR__ +#define C10_GCC_VERSION_MINOR 0 +#else +#define C10_GCC_VERSION_MINOR __GNUC_MINOR__ +#endif + +// We use a templatized class to both contain the logic of checking the sizes +// as well as to provide compile-time information that might be useful in +// figuring out why sizes may have changed. +// All the compile time information is given by the template fields that are +// always printed by the compiler when the static_assert fails. +template < + size_t cplusplus = __cplusplus, + size_t clang_ver_major = C10_CLANG_MAJOR_VERSION, + size_t gcc_ver = C10_GCC_VERSION, + size_t gcc_ver_minor = C10_GCC_VERSION_MINOR, + size_t nvcc = C10_NVCC, + size_t cuda_version = C10_CUDA_VERSION, + size_t cuda_version_major = C10_CUDA_VERSION_MAJOR, + size_t ptr_size = sizeof(void*)> +class C10_TensorImpl_Size_Check_Dummy_Class : private TensorImpl { + // Names of (non-bitfield) fields in TensorImpl; used to provide + // compile-time info about fields whose size changes unexpectedly. + enum class FieldNameEnum { + storage_, + autograd_meta_, + extra_meta_, + version_counter_, + pyobj_slot_, + sizes_and_strides_, + storage_offset_, + numel_, + data_type_, + device_opt_, + key_set_, + TOTAL_SIZE + }; + + // Provides compile-time equality check that reveals what numbers + // were used and on which quantity + template + constexpr static bool are_equal() { + static_assert( + Actual == Expected, + "Actual and Expected sizes of a field did not match!"); + return true; + } + + // Provides compile-time <= check that reveals what numbers + // were used and on which quantity + template + constexpr static bool is_le() { + static_assert( + Actual <= Expected, + "Actual and Expected sizes of a field did not match!"); + return true; + } + + public: + // Compile-time check that TensorImpl field sizes are as expected + // + // Observed total sizes and associated versions + // If you find a flag that predicts when unique_ptr has 16 bytes + // on 64-bit systems or when sizes_and_strides_ is 84 vs 88 bytes + // on 32-bit systems you get a cookie! + // Length | LLVM | GCC | C++ | CUDA + // 192 | ? | 11.2 | 201703 | 11040 + // 208 | ? | 11.2 | 201703 | 11040 + // 208 | ? | 11.2 | 201402 | 11040 + // 192 | ? | 11.2 | 201402 | 11040 + // 160 | 12 | 4.2 | 201703 | 0 + // + // To keep things clean, we split on systems here. + +#if UINTPTR_MAX == 0xFFFFFFFF + // This is a 32-bit system + static constexpr bool check_sizes() { + constexpr size_t tsize = 20 * sizeof(int64_t); + + // clang-format off + are_equal(); + are_equal(); + are_equal(); + are_equal(); + are_equal(); + is_le(); + are_equal(); + are_equal(); + are_equal(); + are_equal(); + are_equal(); + is_le(); + // clang-format on + + return true; + } +#else + // This is a 64-bit system + static constexpr bool check_sizes() { + constexpr size_t tsize = 26 * sizeof(int64_t); + + // clang-format off + are_equal(); + // On some systems involving NVCC the size of unique_ptr is 16 bytes. We haven't + // figured out how to detect those via macro preprocessors yet, so we use <= + // comparisons for the relevant fields. + is_le(); + is_le(); + are_equal(); + are_equal(); + are_equal(); + are_equal(); + are_equal(); + are_equal(); + are_equal(); + are_equal(); + is_le(); + // clang-format on + + return true; + } +#endif +}; + +// We use a class to encapsulate size-checking logic with +// templates to capture sizes and flags. We call this within +// a static assert to prove there is no run-time behaviour. +// Since the methods we call return either true or fail their +// own static_asserts, we should never see the error messages +// below. We have to provide it though for c++ <17. +static_assert( + C10_TensorImpl_Size_Check_Dummy_Class<>::check_sizes(), + "You should not see this message."); + +// Clean up after ourselves +#undef C10_NVCC +#undef C10_CUDA_VERSION_MAJOR +#undef C10_CUDA_VERSION +#undef C10_CLANG_MAJOR_VERSION +#undef C10_GCC_VERSION +#undef C10_GCC_VERSION_MINOR + +} // namespace c10 + +C10_DIAGNOSTIC_POP() + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/TensorOptions.h b/backends/pytorch/include/c10/core/TensorOptions.h new file mode 100644 index 000000000..7add8edc4 --- /dev/null +++ b/backends/pytorch/include/c10/core/TensorOptions.h @@ -0,0 +1,791 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include + +#include +#include +#include +#include +#include + +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum") + +namespace c10 { + +inline ScalarType dtype_or_default(std::optional dtype) { + return dtype.value_or(get_default_dtype_as_scalartype()); +} + +inline caffe2::TypeMeta dtype_or_default( + std::optional dtype) { + return dtype.value_or(get_default_dtype()); +} + +inline Layout layout_or_default(std::optional layout) { + return layout.value_or(kStrided); +} + +inline Device device_or_default(std::optional device) { + return device.value_or(Device(kCPU)); +} + +inline bool pinned_memory_or_default(std::optional pinned_memory) { + return pinned_memory.value_or(false); +} + +/// A class to encapsulate construction axes of an Tensor. TensorOptions was +/// designed to support the Python style API for specifying construction options +/// on factory functions, e.g., +/// +/// torch.zeros(2, 3, dtype=torch.int32) +/// +/// Because C++ doesn't natively support keyword arguments, there must be +/// another way of specifying keyword-like arguments. TensorOptions is a +/// builder class which can be used to construct this "dictionary" of keyword +/// arguments: functions which support TensorOptions conventionally take this +/// argument optionally as their last argument. +/// +/// WARNING: In PyTorch, there are `torch::` variants of factory functions, +/// e.g., torch::zeros for at::zeros. These return Variables (while the +/// stock ATen functions return plain Tensors). If you mix these functions +/// up, you WILL BE SAD. +/// +/// Rather than use the constructor of this class directly, you should prefer to +/// use the constructor functions, and then chain setter methods on top of them. +/// +/// at::device(at::kCUDA).dtype(kInt) +/// at::dtype(at::kInt) +/// +/// Additionally, anywhere a TensorOptions is expected, you can directly +/// pass at::kCUDA / at::kInt, and it will implicitly convert to a +/// TensorOptions. +/// +/// Here are some recommended ways to create a 2x2 tensor of zeros +/// with certain properties. These all *implicitly* make use of +/// TensorOptions, even if they don't mention the class explicitly: +/// +/// at::zeros({2,2}, at::kCUDA); +/// at::zeros({2,2}, at::kLong); +/// at::zeros({2,2}, at::device(at::kCUDA).dtype(at::kLong())); +/// at::zeros({2,2}, at::device({at::kCUDA, 1})); // place on device 1 +/// at::zeros({2,2}, at::requires_grad()); +/// + +/// NOTE [ TensorOptions Constructors ] +/// +/// TensorOptions is like a dictionary with entries from the set: +/// {requires_grad, device, dtype, layout}, where each entry may be +/// unspecified (i.e., is optional). It is used to specify the properties of +/// tensors in many places both in C++ internal and API, e.g., tensor factory +/// methods like `at::empty({10}, options)`, tensor conversions like +/// `tensor.to(...)`, etc. +/// +/// To provide a simple API that is consistent with Python, where one can do +/// `torch.empty(sizes, X)` with `X` being a `torch.device`, `torch.dtype`, or a +/// `torch.layout`, we want TensorOptions to be implicitly convertible from +/// `ScalarType dtype`, `Layout layout` and `Device device`. Therefore, we have +/// three implicit constructors from each of these three types. +/// +/// This is sufficient for `ScalarType` and `Layout` as they are simple Enum +/// classes. However, `Device` is an ordinary class with implicit constructors +/// `Device(DeviceType, DeviceIndex = -1)` and `Device(std::string)` to be +/// consistent with Python API, where strings are treated as equivalent with a +/// `torch.device` object (e.g., "cuda:1" can be passed to everywhere a +/// `torch.device("cuda:1")` is accepted). To support the syntax +/// `at::empty({10}, {kCUDA, 1})` and `tensor.to(kCUDA)`, we need to make sure +/// that `TensorOptions` is implicitly constructible with any arguments that a +/// `Device` can constructed from. So we have, +/// +/// /* implicit */ TensorOptions(T&& device) : TensorOptions() { +/// this->set_device(device); +/// } +/// +/// template ::value>> +/// /* implicit */ TensorOptions(Args&&... args) +/// : TensorOptions(Device(std::forward(args)...)) {} +/// +/// +/// But this will be problematic. Consider this: `TensorOptions({kCUDA, 1})`. +/// Compiler will complain about ambiguity between the copy constructor and the +/// `Device` constructor because `{kCUDA, 1}` can be converted to both a +/// `TensorOption` and a `Device`. +/// +/// To get around this, we templatize the `Device` constructor. Since overload +/// resolution is done before template resolution, our problem is solved. + +DispatchKey computeDispatchKey( + std::optional dtype, + std::optional layout, + std::optional device); + +struct C10_API TensorOptions { + TensorOptions() + : requires_grad_(false), + pinned_memory_(false), + has_device_(false), + has_dtype_(false), + has_layout_(false), + has_requires_grad_(false), + has_pinned_memory_(false), + has_memory_format_(false) {} + + /// Constructs a `TensorOptions` object with the given layout. + /* implicit */ TensorOptions(Layout layout) : TensorOptions() { + this->set_layout(layout); + } + + /// Constructs a `TensorOptions` object with the given device. + /// See NOTE [ TensorOptions Constructors ] on why this is templatized. + template < + typename T, + typename = std::enable_if_t, Device>>> + /* implicit */ TensorOptions(T&& device) : TensorOptions() { + this->set_device(std::forward(device)); + } + + /// Constructs a `TensorOptions` object from arguments allowed in `Device` + /// constructors. + /// + /// See NOTE [ TensorOptions Constructors ]. + /// + /// NB: Ideally we only allow implicit constructors here. But there is no easy + /// way to detect them. So we have this one that allows explicit + /// constructors too. + template < + typename... Args, + typename = std::enable_if_t>> + /* implicit */ TensorOptions(Args&&... args) + : TensorOptions(Device(std::forward(args)...)) {} + + /// Constructs a `TensorOptions` object with the given dtype. + /* implicit */ TensorOptions(caffe2::TypeMeta dtype) : TensorOptions() { + this->set_dtype(dtype); + } + + /// legacy constructor to support ScalarType + /* implicit */ TensorOptions(ScalarType dtype) : TensorOptions() { + this->set_dtype(dtype); + } + + /// Constructs a `TensorOptions` object with the given memory format. + /* implicit */ TensorOptions(MemoryFormat memory_format) : TensorOptions() { + set_memory_format(memory_format); + } + + /// Return a copy of `TensorOptions` with `device` set to the given one, or + /// cleared if `device` is `nullopt`. + [[nodiscard]] TensorOptions device( + std::optional device) const noexcept { + TensorOptions r = *this; + r.set_device(device); + return r; + } + + /// Return a copy of `TensorOptions` with `device` set to the given one. + /// (This overload ensures that variadic template std::optional constructor + /// for Device work correctly.) + template + [[nodiscard]] TensorOptions device(Args&&... args) const noexcept { + return device( + std::optional(std::in_place, std::forward(args)...)); + } + + /// Return a copy of `TensorOptions`, but with device set to CUDA, and the + /// device index set to the given one. + /// + /// TODO: This function encourages bad behavior (assuming CUDA is + /// the only device that matters). Get rid of it / rename it. + [[nodiscard]] TensorOptions device_index( + c10::DeviceIndex device_index) const noexcept { + return device(Device::Type::CUDA, device_index); + } + + /// Return a copy of `TensorOptions` with `dtype` set to the given one. + [[nodiscard]] TensorOptions dtype( + std::optional dtype) const noexcept { + TensorOptions r = *this; + r.set_dtype(dtype); + return r; + } + + // legacy function to support ScalarType + [[nodiscard]] TensorOptions dtype( + std::optional dtype) const noexcept { + TensorOptions r = *this; + r.set_dtype(dtype); + return r; + } + + // Since dtype is taken... + template + TensorOptions& dtype() { + dtype_ = caffe2::TypeMeta::Make(); + has_dtype_ = true; + return *this; + } + + /// Sets the layout of the `TensorOptions`. + [[nodiscard]] TensorOptions layout( + std::optional layout) const noexcept { + TensorOptions r = *this; + r.set_layout(layout); + return r; + } + + /// Sets the `requires_grad` property of the `TensorOptions`. + [[nodiscard]] TensorOptions requires_grad( + std::optional requires_grad) const noexcept { + TensorOptions r = *this; + r.set_requires_grad(requires_grad); + return r; + } + + /// Sets the `pinned_memory` property on the `TensorOptions`. + [[nodiscard]] TensorOptions pinned_memory( + std::optional pinned_memory) const noexcept { + TensorOptions r = *this; + r.set_pinned_memory(pinned_memory); + return r; + } + + /// Sets the `memory_format` property on `TensorOptions`. + [[nodiscard]] TensorOptions memory_format( + std::optional memory_format) const noexcept { + TensorOptions r = *this; + r.set_memory_format(memory_format); + return r; + } + + /// Returns the device of the `TensorOptions`. + Device device() const noexcept { + return device_or_default(device_opt()); + } + + /// Returns whether the device is specified. + bool has_device() const noexcept { + return has_device_; + } + + /// Returns the device of the `TensorOptions`, or `std::nullopt` if + /// device is not specified. + std::optional device_opt() const noexcept { + return has_device_ ? std::make_optional(device_) : std::nullopt; + } + + /// Returns the device index of the `TensorOptions`. + c10::DeviceIndex device_index() const noexcept { + return device().index(); + } + + /// Returns the dtype of the `TensorOptions`. + caffe2::TypeMeta dtype() const noexcept { + return dtype_or_default(dtype_opt()); + } + + /// Returns whether the dtype is specified. + bool has_dtype() const noexcept { + return has_dtype_; + } + + /// Returns the dtype of the `TensorOptions`, or `std::nullopt` if + /// device is not specified. + std::optional dtype_opt() const noexcept { + return has_dtype_ ? std::make_optional(dtype_) : std::nullopt; + } + + /// Returns the layout of the `TensorOptions`. + Layout layout() const noexcept { + return layout_or_default(layout_opt()); + } + + /// Returns whether the layout is specified. + bool has_layout() const noexcept { + return has_layout_; + } + + /// Returns the layout of the `TensorOptions`, or `std::nullopt` if + /// layout is not specified. + std::optional layout_opt() const noexcept { + return has_layout_ ? std::make_optional(layout_) : std::nullopt; + } + + /// Returns the `requires_grad` property of the `TensorOptions`. + bool requires_grad() const noexcept { + return has_requires_grad_ ? requires_grad_ : false; + } + + /// Returns whether the `requires_grad` is specified. + bool has_requires_grad() const noexcept { + return has_requires_grad_; + } + + /// Returns the `requires_grad` property of the `TensorOptions`, or + /// `std::nullopt` if `requires_grad` is not specified. + std::optional requires_grad_opt() const noexcept { + return has_requires_grad_ ? std::make_optional(requires_grad_) + : std::nullopt; + } + + /// Returns the `pinned_memory` property of the `TensorOptions`. + bool pinned_memory() const noexcept { + return pinned_memory_or_default(pinned_memory_opt()); + } + + /// Returns whether the `pinned_memory` is specified. + bool has_pinned_memory() const noexcept { + return has_pinned_memory_; + } + + /// Returns if the layout is sparse + bool is_sparse() const { + return layout_ == c10::Layout::Sparse; + } + + /// Returns if the layout is sparse CSR, deprecated, use + /// is_sparse_compressed() instead + bool is_sparse_csr() const { + return layout_ == c10::Layout::SparseCsr; + } + + bool is_sparse_compressed() const { + return layout_ == c10::Layout::SparseCsr || + layout_ == c10::Layout::SparseCsc || + layout_ == c10::Layout::SparseBsr || layout_ == c10::Layout::SparseBsc; + } + + // For compatibility with legacy tensor.type() comparisons + bool type_equal(const TensorOptions& other) const { + return computeDispatchKey() == other.computeDispatchKey() && + typeMetaToScalarType(dtype_) == typeMetaToScalarType(other.dtype()); + } + + /// Returns the `pinned_memory` property of the `TensorOptions`, or + /// `std::nullopt` if `pinned_memory` is not specified. + std::optional pinned_memory_opt() const noexcept { + return has_pinned_memory_ ? std::make_optional(pinned_memory_) + : std::nullopt; + } + + /// Returns whether the `memory_layout` is specified + bool has_memory_format() const noexcept { + return has_memory_format_; + } + + // NB: memory_format() getter is PURPOSELY not defined, as the default + // behavior of memory_format varies from function to function. + + /// Returns the `memory_layout` property of `TensorOptions, or + /// `std::nullopt` if `memory_format` is not specified. + std::optional memory_format_opt() const noexcept { + return has_memory_format_ ? std::make_optional(memory_format_) + : std::nullopt; + } + + // Resolves the ATen backend specified by the current construction axes. + // TODO: Deprecate this + Backend backend() const { + return at::dispatchKeyToBackend(computeDispatchKey()); + } + + /// Return the right-biased merge of two TensorOptions. This has the + /// effect of overwriting settings from self with specified options + /// of options. + /// + /// NB: This merging operation does NOT respect device merges. + /// For example, if you device({kCUDA, 1}).merge_in(kCUDA) + /// you will get kCUDA in the end! Functions like Tensor.new_empty + /// ensure the right device is selected anyway by way of a + /// device guard. + /// + TensorOptions merge_in(TensorOptions options) const noexcept { + TensorOptions merged = *this; + if (options.has_device()) + merged.set_device(options.device_opt()); + if (options.has_dtype()) + merged.set_dtype(options.dtype_opt()); + if (options.has_layout()) + merged.set_layout(options.layout_opt()); + // NB: requires grad is right biased; not a logical AND/OR! + if (options.has_requires_grad()) + merged.set_requires_grad(options.requires_grad_opt()); + if (options.has_pinned_memory()) + merged.set_pinned_memory(options.pinned_memory_opt()); + if (options.has_memory_format()) + merged.set_memory_format(options.memory_format_opt()); + return merged; + } + + // TODO remove after TensorOptions rationalization + TensorOptions merge_memory_format( + std::optional optional_memory_format) const noexcept { + TensorOptions merged = *this; + if (optional_memory_format.has_value()) { + merged.set_memory_format(optional_memory_format); + } + return merged; + } + + // INVARIANT: computeDispatchKey returns only the subset of dispatch keys for + // which dispatchKeyToBackend is injective, if it is defined at all (for + // the most part, this just means that this function never returns an + // Autograd key) + DispatchKey computeDispatchKey() const { + return c10::computeDispatchKey( + optTypeMetaToScalarType(dtype_opt()), layout_opt(), device_opt()); + } + + private: + // These methods are currently private because I'm not sure if it's wise + // to actually publish them. They are methods because I need them in + // the constructor and the functional API implementation. + // + // If you really, really need it, you can make these public, but check if you + // couldn't just do what you need with the functional API. Similarly, these + // methods are not chainable, because if you wanted chaining, you probably + // want to use the functional API instead. (It's probably OK to make + // these chainable, because these functions are all explicitly annotated + // with a ref-qualifier, the trailing &, that makes them illegal to call + // on temporaries.) + + /// Mutably set the device of `TensorOptions`. + void set_device(std::optional device) & noexcept { + if (device) { + device_ = *device; + has_device_ = true; + } else { + has_device_ = false; + } + } + + /// Mutably set the dtype of `TensorOptions`. + void set_dtype(std::optional dtype) & noexcept { + if (dtype) { + dtype_ = *dtype; + has_dtype_ = true; + } else { + has_dtype_ = false; + } + } + + // legacy function to support ScalarType + void set_dtype(std::optional dtype) & noexcept { + if (dtype) { + dtype_ = scalarTypeToTypeMeta(*dtype); + has_dtype_ = true; + } else { + has_dtype_ = false; + } + } + + /// Mutably set the layout of `TensorOptions`. + void set_layout(std::optional layout) & noexcept { + if (layout) { + layout_ = *layout; + has_layout_ = true; + } else { + has_layout_ = false; + } + } + + /// Mutably set the `requires_grad` property of `TensorOptions`. + void set_requires_grad(std::optional requires_grad) & noexcept { + if (requires_grad) { + requires_grad_ = *requires_grad; + has_requires_grad_ = true; + } else { + has_requires_grad_ = false; + } + } + + /// Mutably set the `pinned_memory` property of `TensorOptions`. + void set_pinned_memory(std::optional pinned_memory) & noexcept { + if (pinned_memory) { + pinned_memory_ = *pinned_memory; + has_pinned_memory_ = true; + } else { + has_pinned_memory_ = false; + } + } + + /// Mutably set the `memory_Format` property of `TensorOptions`. + void set_memory_format(std::optional memory_format) & noexcept { + if (memory_format) { + memory_format_ = *memory_format; + has_memory_format_ = true; + } else { + has_memory_format_ = false; + } + } + + // WARNING: If you edit TensorOptions to add more options, you + // may need to adjust the implementation of Tensor::options. + // The criteria for whether or not Tensor::options must be adjusted + // is whether or not the new option you added should preserved + // by functions such as empty_like(); if it should be preserved, + // you must adjust options(). + // + // TODO: MemoryFormat is not implemented in this way + + // NB: We didn't use std::optional here, because then we can't pack + // the has_***_ boolean fields. + + Device device_ = at::kCPU; // 16-bit + caffe2::TypeMeta dtype_ = caffe2::TypeMeta::Make(); // 16-bit + Layout layout_ = at::kStrided; // 8-bit + MemoryFormat memory_format_ = MemoryFormat::Contiguous; // 8-bit + + // Bitmask required here to get this to fit inside 32 bits (or even 64 bits, + // for that matter) + + bool requires_grad_ : 1; + bool pinned_memory_ : 1; + + bool has_device_ : 1; + bool has_dtype_ : 1; + bool has_layout_ : 1; + bool has_requires_grad_ : 1; + bool has_pinned_memory_ : 1; + bool has_memory_format_ : 1; +}; + +// We should aspire to fit in one machine-size word; but a size greater than two +// words is too much. (We are doing terribly on 32-bit archs, where we require +// three machine size words to store tensor options. Eek!) +static_assert( + sizeof(TensorOptions) <= sizeof(int64_t) * 2, + "TensorOptions must fit in 128-bits"); + +/// Convenience function that returns a `TensorOptions` object with the `dtype` +/// set to the given one. +inline TensorOptions dtype(caffe2::TypeMeta dtype) { + return TensorOptions().dtype(dtype); +} + +// legacy function to support ScalarType +inline TensorOptions dtype(ScalarType dtype) { + return TensorOptions().dtype(scalarTypeToTypeMeta(dtype)); +} + +/// Convenience function that returns a `TensorOptions` object with the `layout` +/// set to the given one. +inline TensorOptions layout(Layout layout) { + return TensorOptions().layout(layout); +} + +/// Convenience function that returns a `TensorOptions` object with the `device` +/// set to the given one. +inline TensorOptions device(Device device) { + return TensorOptions().device(device); +} + +/// Convenience function that returns a `TensorOptions` object with the +/// `device` set to CUDA and the `device_index` set to the given one. +inline TensorOptions device_index(c10::DeviceIndex device_index) { + return TensorOptions().device_index(device_index); +} + +/// Convenience function that returns a `TensorOptions` object with the +/// `requires_grad` set to the given one. +inline TensorOptions requires_grad(bool requires_grad = true) { + return TensorOptions().requires_grad(requires_grad); +} + +/// Convenience function that returns a `TensorOptions` object with the +/// `memory_format` set to the given one. +inline TensorOptions memory_format(MemoryFormat memory_format) { + return TensorOptions().memory_format(memory_format); +} + +C10_API std::ostream& operator<<( + std::ostream& stream, + const TensorOptions& options); + +template +inline TensorOptions dtype() { + return dtype(caffe2::TypeMeta::Make()); +} + +inline std::string toString(const TensorOptions& options) { + std::ostringstream stream; + stream << options; + return stream.str(); +} + +// This is intended to be a centralized location by which we can determine +// what an appropriate DispatchKey for a tensor is. +inline DispatchKey computeDispatchKey( + std::optional dtype, + std::optional layout, + std::optional device) { + const auto layout_ = layout_or_default(layout); + const auto device_ = device_or_default(device); + switch (layout_) { + case Layout::Jagged: + case Layout::Strided: { + const auto dtype_ = dtype_or_default(dtype); + switch (device_.type()) { +#define DO_CASE(device, _) \ + case c10::DeviceType::device: { \ + if (isQIntType(dtype_)) { \ + return DispatchKey::Quantized##device; \ + } \ + return DispatchKey::device; \ + } + C10_FORALL_BACKEND_DEVICE_TYPES(DO_CASE, unused) +#undef DO_CASE + case c10::DeviceType::FPGA: + return DispatchKey::FPGA; + case c10::DeviceType::MAIA: + return DispatchKey::MAIA; + case c10::DeviceType::Vulkan: + return DispatchKey::Vulkan; + case c10::DeviceType::Metal: + return DispatchKey::Metal; + case c10::DeviceType::MKLDNN: + case c10::DeviceType::OPENGL: + case c10::DeviceType::OPENCL: + case c10::DeviceType::IDEEP: + TORCH_INTERNAL_ASSERT( + 0, + "This is a grandfathered Caffe2 device type ", + device_.type(), + ", it shouldn't ever convert to a DispatchKey. File a bug describing what you were doing if you think this is in error."); + default: + TORCH_CHECK_NOT_IMPLEMENTED( + false, + "Unsupported device type for dense layout: ", + device_.type()); + } + } + case Layout::Sparse: + switch (device_.type()) { +#define DO_CASE(device, _) \ + case c10::DeviceType::device: { \ + return DispatchKey::Sparse##device; \ + } + C10_FORALL_BACKEND_DEVICE_TYPES(DO_CASE, unused) +#undef DO_CASE + default: + TORCH_CHECK_NOT_IMPLEMENTED( + false, + "Unsupported device type for sparse layout: ", + device_.type()); + } + case Layout::Mkldnn: + switch (device_.type()) { + case c10::DeviceType::CPU: + return DispatchKey::MkldnnCPU; + default: + TORCH_CHECK_NOT_IMPLEMENTED( + false, + "Unsupported device type for mkldnn layout: ", + device_.type()); + } + case Layout::SparseCsr: + case Layout::SparseCsc: + case Layout::SparseBsr: + case Layout::SparseBsc: + switch (device_.type()) { +#define DO_CASE(device, _) \ + case c10::DeviceType::device: { \ + return DispatchKey::SparseCsr##device; \ + } + C10_FORALL_BACKEND_DEVICE_TYPES(DO_CASE, unused) +#undef DO_CASE + default: + TORCH_CHECK_NOT_IMPLEMENTED( + false, + "Unsupported device type for ", + layout_, + " layout: ", + device_.type()); + } + default: + TORCH_CHECK(false, "Unsupported layout: ", layout_); + } +} + +inline Layout dispatchKeyToLayout(DispatchKey dispatch_key) { + switch (dispatch_key) { +#define DO_CASE(bc, _) case DispatchKey::Sparse##bc: + C10_FORALL_BACKEND_COMPONENTS(DO_CASE, unused) +#undef DO_CASE + return Layout::Sparse; +#define DO_CASE(bc, _) case DispatchKey::SparseCsr##bc: + C10_FORALL_BACKEND_COMPONENTS(DO_CASE, unused) +#undef DO_CASE + TORCH_CHECK( + false, "Cannot map DispatchKey ", dispatch_key, " to a unique layout."); + case DispatchKey::MkldnnCPU: + return Layout::Mkldnn; + default: + return Layout::Strided; + } +} + +inline c10::DeviceType dispatchKeyToDeviceType(DispatchKey dispatch_key) { + switch (dispatch_key) { + // stuff that's real +#define DO_CASE(suffix, prefix) \ + case DispatchKey::prefix##suffix: \ + return c10::DeviceType::suffix; +#define DO_CASES(_, prefix) C10_FORALL_BACKEND_DEVICE_TYPES(DO_CASE, prefix) + C10_FORALL_FUNCTIONALITY_KEYS(DO_CASES) +#undef DO_CASES +#undef DO_CASE + + case DispatchKey::MkldnnCPU: + return c10::DeviceType::CPU; + case DispatchKey::Vulkan: + return c10::DeviceType::Vulkan; + + case DispatchKey::MAIA: + return c10::DeviceType::MAIA; + default: + TORCH_CHECK( + false, + "DispatchKey ", + dispatch_key, + " doesn't correspond to a device"); + } +} + +inline TensorOptions dispatchKeyToTensorOptions(DispatchKey dispatch_key) { + return TensorOptions() + .layout(dispatchKeyToLayout(dispatch_key)) + .device(dispatchKeyToDeviceType(dispatch_key)); +} + +namespace detail { +inline bool backend_supports_empty_operator(const TensorOptions& options) { + // Quantized backends don't support at::empty(). + // They have separate operators like at::empty_quantized() that take in + // extra information about how to quantize the tensor. + return !isQIntType(typeMetaToScalarType(options.dtype())); +} + +} // namespace detail + +} // namespace c10 + +C10_DIAGNOSTIC_POP() + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/UndefinedTensorImpl.h b/backends/pytorch/include/c10/core/UndefinedTensorImpl.h new file mode 100644 index 000000000..3a8381e88 --- /dev/null +++ b/backends/pytorch/include/c10/core/UndefinedTensorImpl.h @@ -0,0 +1,54 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include + +namespace c10 { + +struct C10_API UndefinedTensorImpl final : public TensorImpl { + public: + // Without this, we get: + // error: identifier "at::UndefinedTensorImpl::_singleton" is undefined in + // device code + // (ostensibly because the constexpr tricks MSVC into trying to compile this + // function for device as well). +#ifdef _WIN32 + static inline TensorImpl* singleton() { + return &getInstance(); + } +#else + static constexpr inline TensorImpl* singleton() { + return &_singleton; + } +#endif + +#ifdef DEBUG + bool has_storage() const override; +#endif + void set_storage_offset(int64_t offset) override; + + protected: + c10::SymBool sym_is_contiguous_custom(MemoryFormat format) const override; + IntArrayRef strides_custom() const override; + SymIntArrayRef sym_strides_custom() const override; + + private: + UndefinedTensorImpl(); +#ifdef _WIN32 + static UndefinedTensorImpl& getInstance(); +#else + static UndefinedTensorImpl _singleton; +#endif + const char* tensorimpl_type_name() const override; +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/WrapDimMinimal.h b/backends/pytorch/include/c10/core/WrapDimMinimal.h new file mode 100644 index 000000000..02570ae84 --- /dev/null +++ b/backends/pytorch/include/c10/core/WrapDimMinimal.h @@ -0,0 +1,53 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include + +namespace c10 { + +namespace detail { +// This template can only be specialized at int64_t and c10::SymInt; +// you'll get linker errors otherwise +template +C10_API T maybe_wrap_dim_slow(T dim, T dim_post_expr, bool wrap_scalar); +} // namespace detail + +template +T _maybe_wrap_dim(T dim, T dim_post_expr, bool wrap_scalar = true) { + // Inline the fast paths + if (C10_LIKELY(dim_post_expr * -1 <= dim && dim < dim_post_expr)) { + // For SymInts, we want an explicit control flow to trigger a guard, so we + // may as well branch too. + if (dim < 0) { + return dim + dim_post_expr; + } + return dim; + } + // Check edge-cases out-of-line (wrapping scalars and out-of-bounds errors) + return c10::detail::maybe_wrap_dim_slow( + std::move(dim), std::move(dim_post_expr), wrap_scalar); +} + +inline int64_t maybe_wrap_dim( + int64_t dim, + int64_t dim_post_expr, + bool wrap_scalar = true) { + return _maybe_wrap_dim(dim, dim_post_expr, wrap_scalar); +} + +inline c10::SymInt maybe_wrap_dim( + c10::SymInt dim, + c10::SymInt dim_post_expr, + bool wrap_scalar = true) { + return _maybe_wrap_dim(std::move(dim), std::move(dim_post_expr), wrap_scalar); +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/alignment.h b/backends/pytorch/include/c10/core/alignment.h new file mode 100644 index 000000000..4ef01f7bf --- /dev/null +++ b/backends/pytorch/include/c10/core/alignment.h @@ -0,0 +1,35 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace c10 { + +#ifdef C10_MOBILE +// Use 16-byte alignment on mobile +// - ARM NEON AArch32 and AArch64 +// - x86[-64] < AVX +constexpr size_t gAlignment = 16; +#else +// Use 64-byte alignment should be enough for computation up to AVX512. +constexpr size_t gAlignment = 64; +#endif + +constexpr size_t gPagesize = 4096; +// since the default thp pagesize is 2MB, enable thp only +// for buffers of size 2MB or larger to avoid memory bloating +constexpr size_t gAlloc_threshold_thp = static_cast(2) * 1024 * 1024; + +// Cache line size used to avoid false sharing between threads. Falls back to 64 +// bytes if C++17 feature is unavailable. +#ifdef __cpp_lib_hardware_interference_size +using std::hardware_destructive_interference_size; +#else +constexpr std::size_t hardware_destructive_interference_size = 64; +#endif +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/COW.h b/backends/pytorch/include/c10/core/impl/COW.h new file mode 100644 index 000000000..1ef394e6e --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/COW.h @@ -0,0 +1,37 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace c10 { +struct StorageImpl; +class DataPtr; +} // namespace c10 + +namespace c10::impl::cow { + +// Creates a Copy-on-write (COW) clone of the given storage. This will also +// convert the given storage into a COW storage if it is not COW already. +// +// Converting the storage into a COW storage will not be successful if the +// storage's DataPtr has some context (`DataPtr::get_context()`) which is not +// equal to the data pointer (`DataPtr::get()`). In this case, a nullptr is +// returned. +C10_API c10::intrusive_ptr lazy_clone_storage( + StorageImpl& storage); + +// Check if a storage has a simple DataPtr with no abnormal context +C10_API bool has_simple_data_ptr(const c10::StorageImpl& storage); + +// Check if a DataPtr is COW +C10_API bool is_cow_data_ptr(const c10::DataPtr& data_ptr); + +// Eagerly copies a COW storage's data, turning it into a non-COW storage. +C10_API void materialize_cow_storage(StorageImpl& storage); + +} // namespace c10::impl::cow + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/COWDeleter.h b/backends/pytorch/include/c10/core/impl/COWDeleter.h new file mode 100644 index 000000000..90a618003 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/COWDeleter.h @@ -0,0 +1,71 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include +#include +#include +#include +#include + +namespace c10::impl::cow { + +// A COWDeleterContext object is used as the `ctx` argument for DataPtr +// to implement a Copy-on-write (COW) DataPtr. +class C10_API COWDeleterContext { + public: + // Creates an instance, holding the pair of data and original + // deleter. + // + // Note that the deleter will only be called in our destructor if + // the last reference to this goes away without getting + // materialized. + explicit COWDeleterContext(std::unique_ptr data); + + // Increments the current refcount. + void increment_refcount(); + + // See README.md in this directory to understand the locking + // strategy. + + // Represents a reference to the context. + // + // This is returned by decrement_refcount to allow the caller to + // copy the data under the shared lock. + using NotLastReference = std::shared_lock; + + // Represents the last reference to the context. + // + // This will be returned by decrement_refcount when it is the last + // reference remaining and after any pending copies have completed. + using LastReference = std::unique_ptr; + + // Decrements the refcount, returning a handle indicating what to + // do with it. + std::variant decrement_refcount(); + + private: + // The destructor is hidden, this should only ever be used within + // UniqueVoidPtr using cow::delete_context as the deleter. + ~COWDeleterContext(); + + std::shared_mutex mutex_; + std::unique_ptr data_; + std::atomic refcount_ = 1; +}; + +// `cow_deleter` is used as the `ctx_deleter` for DataPtr to implement a COW +// DataPtr. +// +// Warning: This should only be called on a pointer to a COWDeleterContext that +// was allocated on the heap with `new`, because when the refcount reaches 0, +// the context is deleted with `delete`. +C10_API void cow_deleter(void* ctx); + +} // namespace c10::impl::cow + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/DeviceGuardImplInterface.h b/backends/pytorch/include/c10/core/impl/DeviceGuardImplInterface.h new file mode 100644 index 000000000..f8b12a993 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/DeviceGuardImplInterface.h @@ -0,0 +1,417 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include + +// Just for C10_ANONYMOUS_VARIABLE +#include +#include + +#include +#include + +namespace c10 { + +// Forward declaration +class DataPtr; + +/** + * Note [Flags defining the behavior of events] + * + * PYTORCH_DEFAULT and BACKEND_DEFAULT are valid for all backends. The + * BACKEND_DEFAULT is what a particular backend would select if no + * flags were given. PYTORCH_DEFAULT is the PyTorch's framework default + * choice for events on that backend, which may not be the same. + * + * The mapping of PYTORCH_DEFAULT and BACKEND_DEFAULT is done by each + * backend implementation. + */ +enum class EventFlag { + // Disable timing + PYTORCH_DEFAULT, + // Enable timing + BACKEND_DEFAULT, + // FOR TESTING ONLY + INVALID +}; + +namespace impl { + +/** + * DeviceGuardImplInterface represents the virtual interface which provides + * functionality to provide an RAII class for device and stream switching, + * via DeviceGuard. Every distinct device type, e.g., CUDA and HIP, is + * expected to implement and register an implementation of this interface. + * All classes which inherit from DeviceGuardImplInterface should be declared + * 'final'. + * + * This class exists because we provide a unified interface for performing + * device guards via DeviceGuard, but we cannot assume that we have actually + * compiled against the, e.g., CUDA library, which actually implements + * this guard functionality. In this case, a dynamic dispatch is required + * to cross the library boundary. + * + * If possible, you should directly use implementations of this interface; + * those uses will be devirtualized. + */ +struct C10_API DeviceGuardImplInterface { + DeviceGuardImplInterface() = default; + DeviceGuardImplInterface(const DeviceGuardImplInterface&) = default; + DeviceGuardImplInterface& operator=(const DeviceGuardImplInterface&) = + default; + DeviceGuardImplInterface(DeviceGuardImplInterface&&) noexcept = default; + DeviceGuardImplInterface& operator=(DeviceGuardImplInterface&&) noexcept = + default; + + /** + * Return the type of device managed by this guard implementation. + */ + virtual DeviceType type() const = 0; + + /** + * Set the current device to Device, and return the previous Device. + */ + virtual Device exchangeDevice(Device) const = 0; + // NB: Implementations of exchangeDevice can be a bit boilerplatey. You might + // consider replacing exchangeDevice with a non-virtual function with a baked + // in implementation; however, note that this will triple the number of + // virtual calls (when you implement exchangeDevice in a final subclass, + // the compiler gets to devirtualize everything; it won't do that if you don't + // define it in the subclass!) A common way to solve this problem is to use + // some sort of CRTP; however, we can template DeviceGuardImplInterface since + // we really *do* need it to be virtual. A little boilerplate seems easiest + // to explain. (Another way around this problem is to provide inline + // functions that provide the default implementations, but this seems a little + // hard to explain. In any case, we're only going to have on order of ten + // implementations of this anyway.) + + /** + * Get the current device. + */ + virtual Device getDevice() const = 0; + + /** + * Set the current device to Device. + */ + virtual void setDevice(Device) const = 0; + + /** + * Set the current device to Device, without checking for errors + * (so, e.g., this can be called from a destructor). + */ + virtual void uncheckedSetDevice(Device) const noexcept = 0; + + /** + * Get the current stream for a given device. + */ + virtual Stream getStream(Device) const = 0; + + /** + * Get the default stream for a given device. + */ + virtual Stream getDefaultStream(Device /*unused*/) const { + TORCH_CHECK(false, "Backend doesn't support acquiring a default stream.") + } + + /** + * Get a stream from the global pool for a given device. + */ + virtual Stream getStreamFromGlobalPool( + Device /*unused*/, + bool isHighPriority = false) const { + (void)isHighPriority; // Suppress unused variable warning + TORCH_CHECK(false, "Backend doesn't support acquiring a stream from pool.") + } + + /** + * Return a new stream for a given device and priority. The stream will be + * copied and shared around, device backend should be able to correctly handle + * the lifetime of the stream. + */ + virtual Stream getNewStream(Device /*unused*/, int priority = 0) const { + (void)priority; + TORCH_CHECK(false, "Backend doesn't support create a new Stream.") + } + + /** + * Set a stream to be the thread local current stream for its device. + * Return the previous stream for that device. You are NOT required + * to set the current device to match the device of this stream. + */ + virtual Stream exchangeStream(Stream) const = 0; + + /** + * Destroys the given event. + */ + virtual void destroyEvent(void* /*event*/, const DeviceIndex /*device_index*/) + const noexcept {} + + /** + * Increments the event's version and enqueues a job with this version + * in the stream's work queue. When the stream process that job + * it notifies all streams waiting on / blocked by that version of the + * event to continue and marks that version as recorded. + * */ + virtual void record( + void** /*event*/, + const Stream& /*stream*/, + const DeviceIndex /*device_index*/, + const c10::EventFlag /*flag*/) const { + TORCH_CHECK(false, "Backend doesn't support events."); + } + + /** + * Does nothing if the event has not been scheduled to be recorded. + * If the event was previously enqueued to be recorded, a command + * to wait for the version of the event that exists at the time of this call + * is inserted in the stream's work queue. + * When the stream reaches this command it will stop processing + * additional commands until that version of the event is marked as recorded. + */ + virtual void block(void* /*event*/, const Stream& /*stream*/) const { + TORCH_CHECK(false, "Backend doesn't support events."); + } + + /** + * Returns true if (and only if) + * (1) the event has never been scheduled to be recorded + * (2) the current version is marked as recorded. + * Returns false otherwise. + */ + virtual bool queryEvent(void* /*event*/) const { + TORCH_CHECK(false, "Backend doesn't support events."); + } + + /** + * Get the number of devices. WARNING: This is REQUIRED to not raise + * an exception. If there is some sort of problem, e.g., driver error, + * you should report that there are zero available devices. + */ + virtual DeviceIndex deviceCount() const noexcept = 0; + + /** + * Get the following capabilities of the current device: + * (1) Data type support + * Returns DeviceCapability object. + */ + virtual DeviceCapability getDeviceCapability(Device /*unused*/) const { + TORCH_CHECK(false, "Backend doesn't support getting device capabilities."); + } + + /** + * Return true if all the work previously enqueued on the stream for + * asynchronous execution has completed running on the device. + */ + virtual bool queryStream(const Stream& /*stream*/) const { + TORCH_CHECK(false, "Backend doesn't support querying streams."); + } + + /** + * Wait (by blocking the calling thread) until all the work previously + * enqueued on the stream has completed running on the device. + */ + virtual void synchronizeStream(const Stream& /*stream*/) const { + TORCH_CHECK(false, "Backend doesn't support synchronizing streams."); + } + + /** + * Wait (by blocking the calling thread) until all the work previously + * recorded on the event has completed running on the device. + */ + virtual void synchronizeEvent(void* /*event*/) const { + TORCH_CHECK(false, "Backend doesn't support synchronizing events."); + } + + /** + * Wait (by blocking the calling thread) until all the work previously + * enqueued on the device has been completed. + */ + virtual void synchronizeDevice(const DeviceIndex /*device_index*/) const { + TORCH_CHECK( + false, "Backend doesn't support synchronizing all streams on device."); + } + + /** + * Ensure the caching allocator (if any) is aware that the given DataPtr is + * being used on the given stream, and that it should thus avoid recycling the + * DataPtr until all work on that stream is done. + */ + virtual void recordDataPtrOnStream( + const c10::DataPtr& /*unused*/, + const Stream& /*unused*/) const {} + + /** + * Fetch the elapsed time between two recorded events. + */ + virtual double elapsedTime( + void* /*event1*/, + void* /*event2*/, + const DeviceIndex /*device_index*/) const { + TORCH_CHECK(false, "Backend doesn't support elapsedTime."); + } + + /** + * Intended use of this class is to leak the DeviceGuardImpl at program end. + * So you better not call the destructor, buster! + */ + virtual ~DeviceGuardImplInterface() = default; +}; + +// A no-op device guard impl that doesn't do anything interesting. Useful +// for devices that don't actually have a concept of device index. Prominent +// examples are CPU and Meta. +template +struct NoOpDeviceGuardImpl : public DeviceGuardImplInterface { + NoOpDeviceGuardImpl() = default; + DeviceType type() const override { + return D; + } + Device exchangeDevice(Device /*unused*/) const override { + return Device(D, -1); // no-op + } + Device getDevice() const override { + return Device(D, -1); + } + void setDevice(Device /*unused*/) const override { + // no-op + } + void uncheckedSetDevice(Device /*unused*/) const noexcept override { + // no-op + } + Stream getStream(Device /*unused*/) const noexcept override { + // no-op + return Stream(Stream::DEFAULT, Device(D, -1)); + } + + Stream getNewStream(Device /*unused*/, int priority = 0) const override { + // no-op + (void)priority; + return Stream(Stream::DEFAULT, Device(D, -1)); + } + + // NB: These do NOT set the current device + Stream exchangeStream(Stream /*unused*/) const noexcept override { + // no-op + return Stream(Stream::DEFAULT, Device(D, -1)); + } + DeviceIndex deviceCount() const noexcept override { + return 1; + } + + DeviceCapability getDeviceCapability(Device /*unused*/) const override { + DeviceCapability cap; + if constexpr (D == DeviceType::Meta) { + cap.capability_data.capability_bits = 0; + // Meta only supports basic types for shape inference + // Byte, Char, Short, Int, Long, Float, Double, + // Bool, ComplexFloat, ComplexDouble + cap.capability_data.capability_bits = (1ULL << kIndex_Byte) | + (1ULL << kIndex_Char) | (1ULL << kIndex_Short) | + (1ULL << kIndex_Int) | (1ULL << kIndex_Long) | + (1ULL << kIndex_Float) | (1ULL << kIndex_Double) | + (1ULL << kIndex_ComplexFloat) | (1ULL << kIndex_ComplexDouble) | + (1ULL << kIndex_Bool); + } + return cap; + } + + // Event-related functions + void record( + void** /*event*/, + const Stream& /*stream*/, + const DeviceIndex /*device_index*/, + const EventFlag /*flag*/) const override { + TORCH_CHECK(false, D, " backend doesn't support events."); + } + void block(void* /*event*/, const Stream& /*stream*/) const override { + TORCH_CHECK(false, D, " backend doesn't support events.") + } + bool queryEvent(void* /*event*/) const override { + TORCH_CHECK(false, D, " backend doesn't support events.") + } + void destroyEvent(void* /*event*/, const DeviceIndex /*device_index*/) + const noexcept override {} + + // Stream-related functions + bool queryStream(const Stream& /*stream*/) const override { + return true; + } + void synchronizeStream(const Stream& /*stream*/) const override { + // Don't wait for anything. + } +}; + +// The registry is NON-owning. Each stored pointer is std::atomic so +// that under all interleavings of registry calls the structure is +// race-free. This doesn't cost us anything on reads in X86. (An +// unsynchronized implementation probably is OK too, but I didn't want +// to prove that we never read from device_guard_impl_registry at the +// same time some registration is occurring. Shiver.) +// +// I'd like this registry to be valid even at program destruction time +// (in case someone uses a DeviceGuard in a destructor to do some cleanup +// in the CUDA API.) Since there are no direct accesses of the underlying +// owning objects which I can use to enforce initialization order (unlike +// in a Meyer singleton), it implies that you must *leak* objects when +// putting them in the registry. This is done by deleting the destructor +// on DeviceGuardImplInterface. +extern C10_API std::array< + std::atomic, + static_cast(DeviceType::COMPILE_TIME_MAX_DEVICE_TYPES)> + device_guard_impl_registry; + +// I can't conveniently use c10/util/Registry.h for the following reason: +// c10/util/Registry.h gives me a slow way of Create'ing a object of some +// interface from the registry, but no way of quickly accessing an already +// created object. I'll be banging on getDeviceGuardImpl every time we do a +// DeviceGuard, so I really don't want to be doing an unordered_map lookup. +// Better if the registration mechanism directly drops its implementation +// into device_guard_impl_registry. + +class C10_API DeviceGuardImplRegistrar { + public: + DeviceGuardImplRegistrar( + DeviceType /*type*/, + const DeviceGuardImplInterface* /*impl*/); +}; + +#define C10_REGISTER_GUARD_IMPL(DevType, DeviceGuardImpl) \ + static ::c10::impl::DeviceGuardImplRegistrar C10_ANONYMOUS_VARIABLE( \ + g_##DeviceType)(::c10::DeviceType::DevType, new DeviceGuardImpl()); + +inline const DeviceGuardImplInterface* getDeviceGuardImpl(DeviceType type) { + // Two adjacent int16_t fields DeviceType and DeviceIndex has field access + // miscompiled on NVCC. To workaround this issue, we apply a mask to the + // DeviceType. First check if the DeviceType is 16-bit. + // FB employees can see + // https://fb.workplace.com/groups/llvm.gcc/permalink/4053565044692080/ + // for more details + static_assert(sizeof(DeviceType) == 1, "DeviceType is not 8-bit"); + auto p = device_guard_impl_registry[static_cast(type) & 0xFF].load(); + + // This seems to be the first place where you make use of a device + // when you pass devices to factory functions. Give a nicer error + // message in this case. + TORCH_CHECK(p, "PyTorch is not linked with support for ", type, " devices"); + return p; +} + +void C10_API +registerDeviceGuard(DeviceType type, const DeviceGuardImplInterface* impl); + +inline bool hasDeviceGuardImpl(DeviceType type) { + return device_guard_impl_registry[static_cast(type)].load(); +} + +void C10_API ensureCUDADeviceGuardSet(); + +} // namespace impl +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/HermeticPyObjectTLS.h b/backends/pytorch/include/c10/core/impl/HermeticPyObjectTLS.h new file mode 100644 index 000000000..032b90a20 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/HermeticPyObjectTLS.h @@ -0,0 +1,67 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace c10::impl { + +// This TLS controls whether or not we permanently associate PyObject +// with Tensor the first time it is allocated. When hermetic PyObject +// TLS is enabled (state is true), we DO NOT save PyObjects to Tensor, +// meaning you get a distinct PyObject whenever you execute the code in +// question. +struct C10_API HermeticPyObjectTLS { + static void set_state(bool state); + static bool get_state() { + // Hypothetical fastpath if torchdeploy/multipy // codespell:ignore multipy + // isn't used. Per + // https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p2055r0.pdf + // this qualifies relaxed access because it is a single-location data + // structure (only the boolean here). + // + // Forgetting about data races for a moment, is there a logical race? + // + // - Boolean only ever transitions from false to true. So the + // critical situation is when one interpreter is already running + // when a second interpreter switches haveState from false to true. + // + // - The first interpreter is indifferent whether or not it sees + // hasState true/false; obviously false works (this is what the + // interpreter was previously using; more directly, the interpreter + // calls into itself as the handler, so being hermetic is not + // required), and true simply means serviced python operator calls will + // be hermetic; in these cases it is expected to be functionally + // equivalent. + // + // - The second interpreter MUST see hasState true (as its requests will + // be forwarded to the first interpreter), but it is assumed that there + // is a synchronization between the interpreter initialization, and + // when we actually perform operations, so it is guaranteed to see + // hasState true. + // + // QED. + // + // This fastpath is currently disabled so that we can more easily test that + // hermetic mode works correctly even on stock build of PyTorch. + if (false && !haveState_.load(std::memory_order_relaxed)) + return false; + return get_tls_state(); + } + // Call this from the multipy/torchdeploy // codespell:ignore multipy + // top level + static void init_state(); + + private: + // This only flipped once from false to true during + // torchdeploy/multipy initialization, // codespell:ignore multipy + // and never again. + static std::atomic haveState_; + static bool get_tls_state(); +}; + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/InlineDeviceGuard.h b/backends/pytorch/include/c10/core/impl/InlineDeviceGuard.h new file mode 100644 index 000000000..34d6dff97 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/InlineDeviceGuard.h @@ -0,0 +1,438 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +// This file provides implementations of InlineDeviceGuard and +// InlineOptionalDeviceGuard. + +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10::impl { + +/** + * A DeviceGuard is an RAII class that sets a device to some value + * on construction, and resets the device to its original value on + * destruction. + * + * InlineDeviceGuard is a helper class for implementing DeviceGuards. + * It is templated over a DeviceGuardImpl (anything that implements + * DeviceGuardImplInterface). There are two primary ways to instantiate + * InlineDeviceGuard: + * + * - With a concrete implementation of DeviceGuardImpl, e.g., CUDAGuardImpl. + * This is the best way to use InlineDeviceGuard, as all calls are + * devirtualized, giving you code as efficient as straight line + * calls to cudaGetDevice/cudaSetDevice. + * + * - With VirtualGuardImpl, which does a virtual dispatch to a DeviceGuardImpl + * retrieved from a DeviceType registry. We have explicitly instantiated + * InlineDeviceGuard this way as c10::DeviceGuard. + * + * If you are in a hurry, you can use InlineDeviceGuard directly: + * + * using CUDAGuard = impl::InlineDeviceGuard; + * + * However, you can provide a better user experience if you explicitly write a + * wrapper class that itself contains the template instantiation: + * + * class CUDAGuard { + * public: + * // ... the API ... + * private: + * impl::InlineDeviceGuard guard_; + * } + * + * The wrapper class provides a good place to write documentation, and helps + * avoid weird template instantiation errors when a user incorrectly uses the + * class. + * + * If you need to test this class, consider instantiating it with FakeGuardImpl. + */ +template +class InlineDeviceGuard { + public: + // Note [Omitted default constructor from RAII] + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // In principle, we could add a default constructor to + // DeviceGuard which reads the current device and promises to + // restore to that device on exit. However, most cases where you + // would have written this, you probably meant to actually just + // use DeviceGuard (since you don't actually need the + // restore to happen if you don't ever actually set the device). + // We remove the constructor here to encourage you to think about + // what you actually want to happen. + explicit InlineDeviceGuard() = delete; + + /// Set the current device to the passed Device. + explicit InlineDeviceGuard(Device device) + : impl_(device.type()), + original_device_( + device.index() == -1 ? impl_.getDevice() + : impl_.exchangeDevice(device)), + current_device_(device.index() == -1 ? original_device_ : device) {} + + /// Set the current device index to the passed DeviceIndex. (The + /// device type is inferred from the template parameter T). + template < + typename U = T, + typename = + typename std::enable_if_t>> + explicit InlineDeviceGuard(DeviceIndex device_index) + : InlineDeviceGuard(Device(U::static_type, device_index)) {} + + /// Construct an InlineDeviceGuard using VirtualGuardImpl with an explicit + /// DeviceGuardImplInterface pointer. + template < + typename U = T, + typename = typename std::enable_if_t>> + explicit InlineDeviceGuard( + Device device, + const DeviceGuardImplInterface* impl) + : impl_( + VirtualGuardImpl(impl ? impl : getDeviceGuardImpl(device.type()))), + original_device_( + device.index() == -1 ? impl_.getDevice() + : impl_.exchangeDevice(device)), + current_device_(device.index() == -1 ? original_device_ : device) {} + + /// Copy is disallowed + InlineDeviceGuard(const InlineDeviceGuard&) = delete; + InlineDeviceGuard& operator=(const InlineDeviceGuard&) = delete; + + /// Move is disallowed, as DeviceGuard does not have an uninitialized state, + /// which is required for moves on types with nontrivial destructors. + InlineDeviceGuard(InlineDeviceGuard&& other) = delete; + InlineDeviceGuard& operator=(InlineDeviceGuard&& other) = delete; + + ~InlineDeviceGuard() { + impl_.uncheckedSetDevice(original_device_); + } + + /// Sets the device to the given one. + template < + typename U = T, + typename std::enable_if_t, int> = 0> + void set_device(at::Device device) { + AT_ASSERT( + (U::static_type == DeviceType::HIP && device.is_cuda()) || + device.type() == U::static_type); + auto index = device.index(); + if (index == -1) + return; + impl_.setDevice(device); + current_device_ = device; + } + + /// Resets the currently set device to its original device, and then sets the + /// current device to the passed device. This is effectively equivalent to + /// set_device when a guard supports only a single device type. + template + typename std::enable_if_t> reset_device( + at::Device device) { + set_device(device); + } + + /// Resets the currently set device to its original device, and then sets the + /// current device to the passed device (for a possibly different device + /// type). + /// + /// This method is named reset_device to highlight the fact that previous + /// device settings from this guard are NOT preserved, even if the device + /// has a different device type. For example: + /// + /// // CUDA device is 0 + /// DeviceGuard g(Device(kCUDA, 1)); + /// g.reset_device(Device(kHIP, 2)); + /// // CUDA device is 0 (!!) + /// + /// NOTE: this implementation may skip some device setting if it can prove + /// that it is unnecessary. + /// + /// Optional argument is for testing only. + template + typename std::enable_if_t> reset_device( + at::Device device, + const impl::DeviceGuardImplInterface* impl = nullptr) { + auto index = device.index(); + if (index == -1) + return; + if (device.type() == original_device_.type()) { + AT_ASSERT(impl == nullptr || impl->type() == device.type()); + impl_.setDevice(device); + current_device_ = device; + } else { + // Destruct and reconstruct the DeviceGuard in place + impl_.setDevice(original_device_); + impl_ = !impl ? VirtualGuardImpl(device.type()) : VirtualGuardImpl(impl); + original_device_ = impl_.exchangeDevice(device); + current_device_ = device; + } + } + + /// Sets the device index to the given one. The device type is inferred + /// from the original device type. + void set_index(DeviceIndex index) { + reset_device(Device(original_device_.type(), index)); + } + + /// Returns the device that was set at the time the most recent + /// reset_device(), or otherwise the device at construction time. + Device original_device() const { + return original_device_; + } + + /// Returns the most recent device that was set using this device guard, + /// either from construction, or via set_device/reset_device/set_index. + Device current_device() const { + return current_device_; + } + + protected: + T impl_; + + private: + Device original_device_; + Device current_device_; +}; + +/** + * A OptionalDeviceGuard is an RAII class that sets a device to some value on + * initialization, and resets the device to its original value on destruction. + * + * InlineOptionalDeviceGuard is a helper class for implementing + * OptionalDeviceGuards. See guidance in InlineDeviceGuard on how to + * use this. See OptionalDeviceGuard for user-oriented usage notes. + */ +template +class InlineOptionalDeviceGuard { + public: + // Note [Explicit initialization of optional fields] + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // Explicit initialization of optional fields + // required to workaround an nvcc bug; see + // https://github.com/pytorch/pytorch/issues/12117 + + /// Creates an uninitialized OptionalDeviceGuard. + explicit InlineOptionalDeviceGuard() + : guard_() // See Note [Explicit initialization of optional fields] + {} + ~InlineOptionalDeviceGuard() = default; + + /// Set the current device to the passed Device, if it is not nullopt. + explicit InlineOptionalDeviceGuard(std::optional device_opt) + : guard_() { // See Note [Explicit initialization of optional fields] + if (device_opt.has_value()) { + guard_.emplace(device_opt.value()); + } + } + + /// Set the current device to the passed DeviceIndex, if it is not nullopt. + template < + typename U = T, + typename = + typename std::enable_if_t>> + explicit InlineOptionalDeviceGuard( + std::optional device_index_opt) + : guard_() { // See Note [Explicit initialization of optional fields] + if (device_index_opt.has_value()) { + guard_.emplace(device_index_opt.value()); + } + } + + /// All constructors of DeviceGuard are valid for OptionalDeviceGuard + /// and result in initialized OptionalDeviceGuard. + template + explicit InlineOptionalDeviceGuard(Args&&... args) + : guard_(std::in_place, std::forward(args)...) {} + + // TODO: Consider reading Tensor and TensorList constructors here, when + // Tensor moves to c10. (These are only valid on OptionalDeviceGuard, + // because a Tensor may be undefined, in which case we need an uninitialized + // tensor guard.) + + // Note [Move construction for RAII guards is tricky] + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // In principle, move construction is useful for terminating + // the lifetime of a `OptionalDeviceGuard` early; for example: + // + // // current device is d0 + // OptionalDeviceGuard g1(d1); + // // current device is d1 + // { + // OptionalDeviceGuard g2(std::move(g1)); + // } + // // current device is d0!! + // + // However, it's difficult to implement the move constructor + // in a way that works in all situations. For example, consider + // the following example: + // + // OptionalDeviceGuard g1(d1); + // { + // OptionalDeviceGuard g2(d2); + // { + // OptionalDeviceGuard g3(std::move(g1)); // !!! + // } + // } + // + // What should the current device be while g3 in scope... and what + // should it be after it goes out of scope? What about g2? + // There don't seem to be satisfactory answers for these questions. + // + // It's in principle possible to raise an error when this occurs + // by doing some extra thread-local bookkeeping. But why bother? + // Just don't provide the constructor. + InlineOptionalDeviceGuard(const InlineOptionalDeviceGuard& other) = delete; + InlineOptionalDeviceGuard(InlineOptionalDeviceGuard&& other) = delete; + + // Note [Move assignment for RAII guards is tricky] + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // Move assignment is deleted, because you need to know which guard was + // defined "first", as that guard's original_device_ wins--with the current + // representation, we have no way of telling which is the case. (Move + // construction does not have this problem, as one guard is always + // uninitialized.) + // + // We can make this clear by way of a pair of examples: + // + // Example 1: + // + // // initial device is n0 + // { + // CUDAGuard g1(n1); + // { + // CUDAGuard g2(n2); + // // current device should be n2 + // g1 = std::move(g2); + // // current device should still be n2 + // } + // // current device should still be n2 + // } + // // current device should be n0 + // + // Example 2 (flip the order of the two guards): + // + // // initial device is n0 + // { + // CUDAGuard g2(n2); + // { + // CUDAGuard g1(n1); + // // current device should be n1 + // g1 = std::move(g2); + // // current device should be n2 + // } + // // current device should be n0 (since g2 has been vacated) + // } + // + // In both examples, we need g1 to restore to n0 after move assignment. + // However, in example 1, this is determined by the restore value of g1 + // (prior to the move). In example 2, however, it is determined by the the + // restore value of g2(!!). We don't know which one should win, without having + // a way of telling which guard was allocated first. + // + // We could solve this with an extra thread-local variable. But no one is + // actually using move-assignment. So just get rid of it. + InlineOptionalDeviceGuard& operator=(const InlineOptionalDeviceGuard& other) = + delete; + InlineOptionalDeviceGuard& operator=(InlineOptionalDeviceGuard&& other) = + delete; + + /// Sets the device to the given one. Initializes OptionalDeviceGuard if it + /// is not already initialized. + template < + typename U = T, + typename = + typename std::enable_if_t>> + void set_device(at::Device device) { + if (!guard_.has_value()) { + guard_.emplace(device); + } else { + guard_->set_device(device); + } + } + + /// Resets the currently set device to its original device, and then sets the + /// current device to the passed device (for a possibly different device + /// type). Initializes OptionalDeviceGuard if it is not already initialized. + /// + /// See notes on why this is called reset_device on InlineDeviceGuard. + /// + /// Optional argument is for testing only. + template < + typename U = T, + typename = typename std::enable_if_t>> + void reset_device( + at::Device device, + const DeviceGuardImplInterface* impl = nullptr) { + if (!guard_.has_value()) { + guard_.emplace(device, impl); + } else { + guard_->reset_device(device, impl); + } + } + + /// Resets the currently set device to its original device, and then sets the + /// current device to the passed device. Initializes the guard if it is + /// not already initialized. This is effectively equivalent to set_device + /// when a guard supports only a single device type. + template < + typename U = T, + typename = + typename std::enable_if_t>> + void reset_device(at::Device device) { + if (!guard_.has_value()) { + guard_.emplace(device); + } else { + guard_->reset_device(device); + } + } + + /// Sets the device index to the given one. The device type is statically + /// known. + template < + typename U = T, + typename = + typename std::enable_if_t>> + void set_index(DeviceIndex index) { + if (!guard_.has_value()) { + guard_.emplace(index); + } else { + guard_->set_index(index); + } + } + + /// Returns the device that was set immediately prior to initialization of + /// the, guard, or nullopt if the guard is uninitialized. + std::optional original_device() const { + return guard_.has_value() ? std::make_optional(guard_->original_device()) + : std::nullopt; + } + + /// Returns the most recent device that was set using this device guard, + /// either from construction, or via set_device, if the guard is initialized, + /// or nullopt if the guard is uninitialized. + std::optional current_device() const { + return guard_.has_value() ? std::make_optional(guard_->current_device()) + : std::nullopt; + } + + /// Restore the original device, resetting this guard to uninitialized state. + void reset() { + guard_.reset(); + } + + private: + std::optional> guard_; +}; + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/InlineEvent.h b/backends/pytorch/include/c10/core/impl/InlineEvent.h new file mode 100644 index 000000000..15d4083da --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/InlineEvent.h @@ -0,0 +1,152 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include + +namespace c10::impl { + +template +struct InlineEvent final { + InlineEvent() = delete; + InlineEvent( + const DeviceType _device_type, + const EventFlag _flag = EventFlag::PYTORCH_DEFAULT) + : backend_{_device_type}, device_type_{_device_type}, flag_{_flag} {} + + // Copy constructor and copy assignment operator (deleted) + InlineEvent(const InlineEvent&) = delete; + InlineEvent& operator=(const InlineEvent&) = delete; + + // Move constructor and move assignment operator + InlineEvent(InlineEvent&& other) noexcept + : event_(other.event_), + backend_(std::move(other.backend_)), + device_type_(other.device_type_), + device_index_(other.device_index_), + flag_(other.flag_), + was_marked_for_recording_(other.was_marked_for_recording_) { + other.event_ = nullptr; + } + InlineEvent& operator=(InlineEvent&& other) noexcept { + swap(other); + return *this; + } + + void swap(InlineEvent& other) noexcept { + std::swap(event_, other.event_); + std::swap(backend_, other.backend_); + std::swap(device_type_, other.device_type_); + std::swap(device_index_, other.device_index_); + std::swap(flag_, other.flag_); + std::swap(was_marked_for_recording_, other.was_marked_for_recording_); + } + + ~InlineEvent() noexcept { + if (event_) + backend_.destroyEvent(event_, device_index_); + } + + DeviceType device_type() const noexcept { + return device_type_; + } + DeviceIndex device_index() const noexcept { + return device_index_; + } + EventFlag flag() const noexcept { + return flag_; + } + bool was_marked_for_recording() const noexcept { + return was_marked_for_recording_; + } + + void recordOnce(const Stream& stream) { + if (!was_marked_for_recording_) + record(stream); + } + + void record(const Stream& stream) { + TORCH_CHECK( + stream.device_type() == device_type_, + "Event device type ", + DeviceTypeName(device_type_), + " does not match recording stream's device type ", + DeviceTypeName(stream.device_type()), + "."); + + backend_.record(&event_, stream, device_index_, flag_); + was_marked_for_recording_ = true; + device_index_ = stream.device_index(); + } + + void block(const Stream& stream) const { + if (!was_marked_for_recording_) + return; + + TORCH_CHECK( + stream.device_type() == device_type_, + "Event device type ", + DeviceTypeName(device_type_), + " does not match blocking stream's device type ", + DeviceTypeName(stream.device_type()), + "."); + + backend_.block(event_, stream); + } + + bool query() const { + if (!was_marked_for_recording_) + return true; + return backend_.queryEvent(event_); + } + + void* eventId() const { + return event_; + } + + double elapsedTime(const InlineEvent& other) const { + TORCH_CHECK( + other.device_type() == device_type_, + "Event device type ", + DeviceTypeName(device_type_), + " does not match other's device type ", + DeviceTypeName(other.device_type()), + "."); + TORCH_CHECK_VALUE( + (flag_ == EventFlag::BACKEND_DEFAULT) && + (other.flag_ == EventFlag::BACKEND_DEFAULT), + "Both events must be created with argument 'enable_timing=True'."); + TORCH_CHECK_VALUE( + was_marked_for_recording() && other.was_marked_for_recording(), + "Both events must be recorded before calculating elapsed time."); + // elapsedTime in MPS can wait event to be completed if event is not ready, + // which is a little different from CUDA + TORCH_CHECK( + (query() && other.query()) || device_type_ == DeviceType::MPS, + "Both events must be completed before calculating elapsed time."); + + return backend_.elapsedTime(event_, other.event_, device_index_); + } + + void synchronize() const { + if (!was_marked_for_recording_) + return; + backend_.synchronizeEvent(event_); + } + + private: + void* event_ = nullptr; + T backend_; + DeviceType device_type_; + DeviceIndex device_index_ = -1; + EventFlag flag_ = EventFlag::PYTORCH_DEFAULT; + bool was_marked_for_recording_ = false; +}; + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/InlineStreamGuard.h b/backends/pytorch/include/c10/core/impl/InlineStreamGuard.h new file mode 100644 index 000000000..7ce87a9a8 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/InlineStreamGuard.h @@ -0,0 +1,265 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +namespace c10::impl { + +/** + * A StreamGuard is an RAII class that changes the current device + * to the device corresponding to some stream, and changes the + * default stream on that device to be this stream. + * + * InlineStreamGuard is a helper class for implementing StreamGuards. + * See InlineDeviceGuard for guidance on how to use this class. + */ +template +class InlineStreamGuard : private InlineDeviceGuard { + public: + /// No default constructor, see Note [Omitted default constructor from RAII] + explicit InlineStreamGuard() = delete; + + /// Set the current device to the device associated with the passed stream, + /// and set the current stream on that device to the passed stream. + explicit InlineStreamGuard(Stream stream) + : InlineDeviceGuard(stream.device()), + original_stream_of_original_device_( + this->impl_.getStream(original_device())), + original_stream_of_current_device_(this->impl_.exchangeStream(stream)), + current_stream_(stream) {} + + /// This constructor exists purely for testing + template < + typename U = T, + typename = typename std::enable_if_t>> + explicit InlineStreamGuard( + Stream stream, + const DeviceGuardImplInterface* impl) + : InlineDeviceGuard( + stream.device(), + impl ? impl : getDeviceGuardImpl(stream.device_type())), + original_stream_of_original_device_( + this->impl_.getStream(original_device())), + original_stream_of_current_device_(this->impl_.exchangeStream(stream)), + current_stream_(stream) {} + + /// Copy is disallowed + InlineStreamGuard(const InlineStreamGuard&) = delete; + InlineStreamGuard& operator=(const InlineStreamGuard&) = delete; + + /// Move is disallowed, as StreamGuard does not have an uninitialized state, + /// which is required for moves on types with nontrivial destructors. + InlineStreamGuard(InlineStreamGuard&& other) = delete; + InlineStreamGuard& operator=(InlineStreamGuard&& other) = delete; + + ~InlineStreamGuard() { + this->impl_.exchangeStream(original_stream_of_current_device_); + } + + /// Resets the currently set stream to the original stream and + /// the currently set device to the original device. Then, + /// set the current device to the device associated with the passed stream, + /// and set the current stream on that device to the passed stream. + /// + /// NOTE: this implementation may skip some stream/device setting if + /// it can prove that it is unnecessary. + /// + /// WARNING: reset_stream does NOT preserve previously set streams on + /// different devices. If you need to set streams on multiple devices + /// use MultiStreamGuard instead. + void reset_stream(Stream stream) { + // TODO: make a version that takes an impl argument. Unfortunately, + // that will require SFINAE because impl is only valid for the + // VirtualGuardImpl specialization. + if (stream.device() == this->current_device()) { + this->impl_.exchangeStream(stream); + current_stream_ = stream; + } else { + // Destruct and reconstruct the StreamGuard in-place + this->impl_.exchangeStream(original_stream_of_current_device_); + this->reset_device(stream.device()); + original_stream_of_current_device_ = this->impl_.exchangeStream(stream); + current_stream_ = stream; + } + } + + // It's not clear if set_device should also reset the current stream + // if the device is unchanged; therefore, we don't provide it. + // The situation is somewhat clearer with reset_device, but it's still + // a pretty weird thing to do, so haven't added this either. + + /// Returns the stream of the original device prior to this guard. Subtly, + /// the stream returned here is the original stream of the *original* + /// device; i.e., it's the stream that your computation *would* have + /// been put on, if it hadn't been for this meddling stream guard. + /// This is usually what you want. + Stream original_stream() const { + return original_stream_of_original_device_; + } + + /// Returns the most recent stream that was set using this device guard, + /// either from construction, or via set_stream. + Stream current_stream() const { + return current_stream_; + } + + /// Returns the most recent device that was set using this device guard, + /// either from construction, or via set_device/reset_device/set_index. + Device current_device() const { + return InlineDeviceGuard::current_device(); + } + + /// Returns the device that was set at the most recent reset_stream(), + /// or otherwise the device at construction time. + Device original_device() const { + return InlineDeviceGuard::original_device(); + } + + private: + Stream + original_stream_of_original_device_; // what the user probably cares about + Stream original_stream_of_current_device_; // what we need to restore + Stream current_stream_; +}; + +/** + * An OptionalStreamGuard is an RAII class that sets a device to some value on + * initialization, and resets the device to its original value on destruction. + * See InlineOptionalDeviceGuard for more guidance on how to use this class. + */ +template +class InlineOptionalStreamGuard { + public: + /// Creates an uninitialized stream guard. + explicit InlineOptionalStreamGuard() + : guard_() // See Note [Explicit initialization of optional fields] + {} + ~InlineOptionalStreamGuard() = default; + + /// Set the current device to the device associated with the passed stream, + /// and set the current stream on that device to the passed stream, + /// if the passed stream is not nullopt. + explicit InlineOptionalStreamGuard(std::optional stream_opt) + : guard_() { + if (stream_opt.has_value()) { + guard_.emplace(stream_opt.value()); + } + } + + /// All constructors of StreamGuard are valid for OptionalStreamGuard + template + explicit InlineOptionalStreamGuard(Args&&... args) + : guard_(std::in_place, std::forward(args)...) {} + + InlineOptionalStreamGuard(const InlineOptionalStreamGuard& other) = delete; + InlineOptionalStreamGuard& operator=(const InlineOptionalStreamGuard& other) = + delete; + // See Note [Move construction for RAII guards is tricky] + InlineOptionalStreamGuard(InlineOptionalStreamGuard&& other) = delete; + + // See Note [Move assignment for RAII guards is tricky] + InlineOptionalStreamGuard& operator=(InlineOptionalStreamGuard&& other) = + delete; + + /// Resets the currently set stream to the original stream and + /// the currently set device to the original device. Then, + /// set the current device to the device associated with the passed stream, + /// and set the current stream on that device to the passed stream. + /// Initializes the OptionalStreamGuard if it was not previously initialized. + void reset_stream(Stream stream) { + if (guard_.has_value()) { + guard_->reset_stream(stream); + } else { + guard_.emplace(stream); + } + } + + /// Returns the stream that was set at the time the guard was most recently + /// initialized, or nullopt if the guard is uninitialized. + std::optional original_stream() const { + return guard_.has_value() ? std::make_optional(guard_->original_stream()) + : std::nullopt; + } + + /// Returns the most recent stream that was set using this stream guard, + /// either from construction, or via reset_stream, if the guard is + /// initialized, or nullopt if the guard is uninitialized. + std::optional current_stream() const { + return guard_.has_value() ? std::make_optional(guard_->current_stream()) + : std::nullopt; + } + + /// Restore the original device and stream, resetting this guard to + /// uninitialized state. + void reset() { + guard_.reset(); + } + + private: + std::optional> guard_; +}; + +template +class InlineMultiStreamGuard { + public: + /// Calls `set_stream` on each of the streams in the list. + /// This may be useful if you need to set different streams + /// for different devices. + explicit InlineMultiStreamGuard(ArrayRef streams) { + if (!streams.empty()) { + impl_.emplace(getDeviceTypeOfStreams(streams)); + original_streams_.reserve(streams.size()); + for (const Stream& s : streams) { + original_streams_.emplace_back(this->impl_->exchangeStream(s)); + } + } + } + + /// Copy is disallowed + InlineMultiStreamGuard(const InlineMultiStreamGuard&) = delete; + InlineMultiStreamGuard& operator=(const InlineMultiStreamGuard&) = delete; + + /// Move is disallowed, as StreamGuard does not have an uninitialized state, + /// which is required for moves on types with nontrivial destructors. + InlineMultiStreamGuard(InlineMultiStreamGuard&& other) = delete; + InlineMultiStreamGuard& operator=(InlineMultiStreamGuard&& other) = delete; + + ~InlineMultiStreamGuard() noexcept { + if (this->impl_.has_value()) { + for (const Stream& s : original_streams_) { + this->impl_->exchangeStream(s); + } + } + } + + protected: + std::optional impl_; + + private: + /// The original streams that were active on all devices. + std::vector original_streams_; + + static DeviceType getDeviceTypeOfStreams(ArrayRef streams) { + TORCH_INTERNAL_ASSERT(!streams.empty()); + DeviceType type = streams[0].device_type(); + for (const auto idx : c10::irange(1, streams.size())) { + TORCH_CHECK_VALUE( + streams[idx].device_type() == type, + "Streams have a mix of device types: stream 0 is on ", + streams[0].device(), + " while stream ", + idx, + " is on device ", + streams[idx].device()); + } + return type; + } +}; + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/LocalDispatchKeySet.h b/backends/pytorch/include/c10/core/impl/LocalDispatchKeySet.h new file mode 100644 index 000000000..123a288a0 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/LocalDispatchKeySet.h @@ -0,0 +1,174 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +// TLS management for DispatchKeySet (the "local" DispatchKeySet(s)) +// +// This manages two thread-local DispatchKeySets: +// +// - The included type set, which adds a tensor type for consideration +// in dispatch. (For example, you might add Profiling to +// the included type set to turn on profiling on all tensor operations.) +// +// - The excluded type set, which disqualifies a tensor type from dispatch. +// (For example, after redispatching on variable, we disqualify +// Autograd so we don't attempt to handle variable again.) +// (Exclusion wins over inclusion.) +// +// NB: Originally, I implemented the excluded type set as storing the inverted +// set, but TLS is defined to be zero-initialized, so this doesn't actually work +// (if it's inverted, you want the set to be -1 initialized). + +namespace c10::impl { + +// POD version of LocalDispatchKeySet. Declared here just so that +// we can put it in the guards. +// This struct encapsulates special handling for TLS initialization +// in set_included()/included() API so that they reflect the truth. +// If you want to create PODLocalDispatchKeySet with non-zero state, +// use set_included() instead of default constructor. +struct C10_API PODLocalDispatchKeySet { + uint64_t included_; + uint64_t excluded_; + + // See Note [TLS Initialization] + DispatchKeySet included() const { + return DispatchKeySet(DispatchKeySet::RAW, included_) ^ + c10::default_included_set; + } + DispatchKeySet excluded() const { + return DispatchKeySet(DispatchKeySet::RAW, excluded_) ^ + c10::default_excluded_set; + } + + void set_included(DispatchKeySet x) { + included_ = (x ^ c10::default_included_set).raw_repr(); + } + void set_excluded(DispatchKeySet x) { + excluded_ = (x ^ c10::default_excluded_set).raw_repr(); + } +}; +static_assert( + std::is_trivial_v, + "PODLocalDispatchKeySet must be a POD type."); + +struct C10_API LocalDispatchKeySet { + /* implicit */ LocalDispatchKeySet(PODLocalDispatchKeySet x) + : included_(x.included()), excluded_(x.excluded()) {} + DispatchKeySet included_; + DispatchKeySet excluded_; +}; + +// thread_local variables cannot be C10_API on Windows. +// Inlining this seems to break AutoDispatchBelowAutograd on Android. +#if defined(_MSC_VER) || defined(C10_ANDROID) || defined(C10_IPHONE) +C10_API LocalDispatchKeySet tls_local_dispatch_key_set(); +#else // defined(_MSC_VER) || defined(C10_ANDROID) || defined(C10_IPHONE) +extern C10_API thread_local PODLocalDispatchKeySet raw_local_dispatch_key_set; + +inline C10_API LocalDispatchKeySet tls_local_dispatch_key_set() { + // Don't let people fiddle with the thread_local directly just + // because they include this header. + return raw_local_dispatch_key_set; +} +#endif // defined(_MSC_VER) || defined(C10_ANDROID) || defined(C10_IPHONE) + +// Internal, use ThreadLocalStateGuard +C10_API void _force_tls_local_dispatch_key_set(LocalDispatchKeySet key_set); + +// RAII API for manipulating the thread-local dispatch state. + +class C10_API IncludeDispatchKeyGuard { + public: + IncludeDispatchKeyGuard(DispatchKeySet /*include*/); + IncludeDispatchKeyGuard(DispatchKey k) + : IncludeDispatchKeyGuard(DispatchKeySet(k)) {} + IncludeDispatchKeyGuard(const IncludeDispatchKeyGuard&) = delete; + IncludeDispatchKeyGuard operator=(const IncludeDispatchKeyGuard&) = delete; + IncludeDispatchKeyGuard(IncludeDispatchKeyGuard&&) = delete; + IncludeDispatchKeyGuard operator=(IncludeDispatchKeyGuard&&) = delete; + ~IncludeDispatchKeyGuard(); + + private: + // A little micro-optimization to save us from tls_get_addr call + // on destruction + PODLocalDispatchKeySet* tls_; + DispatchKeySet include_; +}; + +class C10_API ExcludeDispatchKeyGuard { + public: + ExcludeDispatchKeyGuard(DispatchKeySet /*exclude*/); + ExcludeDispatchKeyGuard(DispatchKey k) + : ExcludeDispatchKeyGuard(DispatchKeySet(k)) {} + ExcludeDispatchKeyGuard(const ExcludeDispatchKeyGuard&) = delete; + ExcludeDispatchKeyGuard operator=(const ExcludeDispatchKeyGuard&) = delete; + ExcludeDispatchKeyGuard(ExcludeDispatchKeyGuard&&) = delete; + ExcludeDispatchKeyGuard operator=(ExcludeDispatchKeyGuard&&) = delete; + ~ExcludeDispatchKeyGuard(); + + private: + // A little micro-optimization to save us from tls_get_addr call + // on destruction + PODLocalDispatchKeySet* tls_; + DispatchKeySet exclude_; +}; + +struct C10_API ForceDispatchKeyGuard { + public: + ForceDispatchKeyGuard() + : saved_keyset_(c10::impl::tls_local_dispatch_key_set()) {} + ForceDispatchKeyGuard(c10::impl::LocalDispatchKeySet key_set) + : ForceDispatchKeyGuard() { + c10::impl::_force_tls_local_dispatch_key_set(key_set); + } + ForceDispatchKeyGuard( + c10::DispatchKeySet include, + c10::DispatchKeySet exclude) + : ForceDispatchKeyGuard() { + auto updated_set = saved_keyset_; + updated_set.included_ = include; + updated_set.excluded_ = exclude; + c10::impl::_force_tls_local_dispatch_key_set(updated_set); + } + + ForceDispatchKeyGuard(ForceDispatchKeyGuard&&) noexcept = delete; + ForceDispatchKeyGuard(const ForceDispatchKeyGuard&) = delete; + ForceDispatchKeyGuard& operator=(const ForceDispatchKeyGuard&) = delete; + ForceDispatchKeyGuard& operator=(ForceDispatchKeyGuard&&) = delete; + ~ForceDispatchKeyGuard() { + c10::impl::_force_tls_local_dispatch_key_set(saved_keyset_); + } + + private: + c10::impl::LocalDispatchKeySet saved_keyset_; +}; + +// Non-RAII API for manipulating the thread-local dispatch state. +// Please prefer the RAII API. The non-RAII API may be useful when +// the included/excluded state of a given DispatchKey must span +// many calls from the Python to the C++, so you cannot conveniently +// use an RAII guard. +// +// Example use case: a Python context manager that includes a certain +// DispatchKey, to ensure ops running under the context manager dispatch +// through that DispatchKey's registered overrides. +// +// The non-RAII API is less efficient than the RAII guards because both the +// getter and setter will do a tls_getaddr lookup (the RAII struct only needs +// one!) + +C10_API bool tls_is_dispatch_key_excluded(DispatchKey x); +C10_API void tls_set_dispatch_key_excluded(DispatchKey x, bool desired_state); +C10_API bool tls_is_dispatch_key_included(DispatchKey x); +C10_API void tls_set_dispatch_key_included(DispatchKey x, bool desired_state); +C10_API bool tls_is_dispatch_keyset_excluded(DispatchKeySet ks); +C10_API bool tls_is_dispatch_keyset_included(DispatchKeySet ks); + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/PyInterpreter.h b/backends/pytorch/include/c10/core/impl/PyInterpreter.h new file mode 100644 index 000000000..ce74e9b90 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/PyInterpreter.h @@ -0,0 +1,257 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +// Forward declarations + +namespace c10 { +struct IValue; +class OperatorHandle; +struct TensorImpl; +namespace impl { +struct PyObjectSlot; +} // namespace impl +} // namespace c10 + +namespace torch::jit { +using Stack = std::vector; +} + +// Actual implementation + +namespace c10::impl { + +struct C10_API PyInterpreter; + +// Note [Python interpreter tag] +// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +// Traditionally, PyTorch is layered such that our Python library +// (libtorch_python) references our pure C++ library (libtorch) as the +// natural order of things. However, sometimes this natural order is +// subverted: C++ objects refer to Python objects (for example, we +// store a PyObject* pointer on TensorImpl so that converting from a +// C++ Tensor to a Python Tensor is just a memory dereference). +// +// These unusual orderings must be treated with care. To start, you need to +// virtualize the destructor so that the PyObject can be decref'ed on +// destruction (because the C++ object itself doesn't know anything about +// Python--remember, layering!). This process itself is fraught, since +// acquiring the GIL could lead to deadlocks if someone is blocking on you +// while holding the GIL. Furthermore, if the C++ objects outlive the +// interpreter (which can happen if you stash them in a static global +// variable defined in libtorch), you may attempt to decref the object when +// the Python interpreter has already been shutdown. +// +// BUT WAIT, IT GETS WORSE. With torchdeploy, there may be multiple Python +// interpreters in a single process. If a C++ object is accessible from +// multiple interpreters, we must take care not to accidentally pass a +// PyObject from one interpreter with another interpreter. +// +// To prevent these mixups, we introduce a PyInterpreter "tag" (object with +// a vtable), which specifies a specific Python interpreter. +// +// - Any given object can be associated with AT MOST one Python interpreter. +// We represent the interpreter tag as a memory address to an instance of +// a virtual class that is allocated once per interpreter (this is so that +// we can request the interpreter to perform operations for us, if +// necessary). +// +// - It can be recorded with a PyObject (PyInterpreterObject) so that +// we know what interpreter the object is associated with, and we can +// raise an error if you try to use the PyObject from the wrong +// interpreter context. +// +// - It contains a vtable that can be used to perform various Python +// operations from ordinary C++ code that ordinarily wouldn't be accessible +// from libtorch. +// +// A simple use case is when a C++ object must be associated with a PyObject. +// However, for TensorImpl, we lazily allocate a PyObject the first time the +// object passes into Python. The invariants for this situation are more +// subtle: +// +// - A given TensorImpl's interpreter tag can only go from uninitialized to +// tagged; once tagged, this is a quiescent state (once tagged to an +// interpreter, ALWAYS tagged to that interpreter) +// +// - A thread may mutate the PyObject field of a TensorImpl if and only if it +// holds the GIL for the interpreter tagged on the TensorImpl. (If the +// TensorImpl is not tagged, it must first atomically claim its tag before it +// can validly write) +// +// WARNING: This class has to be written very carefully, because it may be +// possible for a Tensor to have a reference an interpreter corresponding to +// a shared library that has ALREADY BEEN UNLOADED. This makes blindly calling +// virtual methods very dangerous, because the vtable may be garbage at that +// point (on a good day, you might get "pure virtual method called"). +// +// The idea to solve this problem is we always leak PyInterpreters (so they +// always stay live even after dlclose), and make sure we can disarm their +// virtual methods by indirecting through a separate PyInterpreterVTable +// object. This can be replaced with a no-op vtable from libc10.so, which +// is guaranteed to stick around until the bitter end. +// +// NB: The downside with representing PyInterpreter tags as full objects is that +// it takes an extra word on TensorImpl. If tags were instead just integer +// indices, on 64-bit architectures we could pack the tag and PyObject together +// into a single atomic word. On 32-bit architectures we could simply say that +// only one Python interpreter is supported (erroring if a nontrivial +// interpreter tag is attempted to be set). +// +// The difficulty with this scheme is we need to maintain an out-of-line table +// to get at the PyInterpreters so that we can do virtual method calls on them, +// and registration/deregistration to this table must be done in a thread safe +// manner. This can be easily done if the number of possible PyInterpreters is +// small enough (e.g., 8-bit integer) by simply preallocating an array of +// sufficient size to hold all possible interpreters. Surely 128 threads is +// more than enough for anyone! +// +// I didn't decide to do this technique at the moment, because the extra word +// added by the PyInterpreter tag takes us to 24 words, which means that we +// still fit inside three eight word cache lines. If you need to penny pinch +// another word consider doing this! + +struct C10_API PyInterpreterVTable { + virtual ~PyInterpreterVTable() = default; + + // Report the name of this interpreter + virtual std::string name() const = 0; + + // Run Py_INCREF on a PyObject. + virtual void incref(PyObject* pyobj) const = 0; + // Run Py_DECREF on a PyObject. We DO NOT assume the GIL is held on call. + virtual void decref(PyObject* pyobj) const = 0; + // Run PyUnstable_TryIncRef on a PyObject if it's not NULL. + virtual bool try_incref(const c10::impl::PyObjectSlot& pyobj_slot) const = 0; + // Run Py_REFCNT on a PyObject. + virtual size_t refcnt(PyObject* pyobj) const = 0; + + // Perform a detach by deferring to the __torch_dispatch__ implementation of + // detach, which will also arrange for the PyObject to get copied in this + // situation + virtual c10::intrusive_ptr detach( + const TensorImpl* self) const = 0; + + // Invoke the Python boxed fallback dispatch to go back into Python + virtual void dispatch(const c10::OperatorHandle& op, torch::jit::Stack* stack) + const = 0; + + virtual void reportErrorCallback(PyObject* callback, DispatchKey key) + const = 0; + + // This is only invoked in the multipy/torchdeploy // codespell:ignore multipy + // situation from pythonOpRegistrationTrampoline; this lets us get to the + // Python interpreter to actually find the appropriate Python op registration + // entry to call. + virtual void python_op_registration_trampoline( + const c10::OperatorHandle& op, + c10::DispatchKey, + c10::DispatchKeySet keyset, + torch::jit::Stack* stack, + bool with_keyset, + bool with_op) const = 0; + + virtual void throw_abstract_impl_not_imported_error( + std::string opname, + const char* pymodule, + const char* context) const = 0; + + // Invoke the Python dispatcher to handle this call + virtual void python_dispatcher( + const c10::OperatorHandle& op, + c10::DispatchKeySet, + torch::jit::Stack* stack) const = 0; + + virtual bool is_contiguous(const TensorImpl* self, at::MemoryFormat) + const = 0; + virtual c10::SymBool sym_is_contiguous( + const TensorImpl* self, + at::MemoryFormat) const = 0; + virtual bool is_strides_like(const TensorImpl* self, at::MemoryFormat) + const = 0; + virtual bool is_non_overlapping_and_dense(const TensorImpl* self) const = 0; + virtual c10::Device device(const TensorImpl* self) const = 0; + virtual int64_t dim(const TensorImpl* self) const = 0; + virtual c10::IntArrayRef strides(const TensorImpl* self) const = 0; + virtual c10::IntArrayRef sizes(const TensorImpl* self) const = 0; + virtual c10::SymIntArrayRef sym_sizes(const TensorImpl* self) const = 0; + virtual c10::Layout layout(const TensorImpl* self) const = 0; + virtual int64_t numel(const TensorImpl* self) const = 0; + virtual c10::SymInt sym_numel(const TensorImpl* self) const = 0; + virtual c10::SymIntArrayRef sym_strides(const TensorImpl* self) const = 0; + virtual c10::SymInt sym_storage_offset(const TensorImpl* self) const = 0; + + virtual void trace_gpu_event_creation( + c10::DeviceType device_type, + uintptr_t event) const = 0; + virtual void trace_gpu_event_deletion( + c10::DeviceType device_type, + uintptr_t event) const = 0; + virtual void trace_gpu_event_record( + c10::DeviceType device_type, + uintptr_t event, + uintptr_t stream) const = 0; + virtual void trace_gpu_event_wait( + c10::DeviceType device_type, + uintptr_t event, + uintptr_t stream) const = 0; + virtual void trace_gpu_memory_allocation( + c10::DeviceType device_type, + uintptr_t ptr) const = 0; + virtual void trace_gpu_memory_deallocation( + c10::DeviceType device_type, + uintptr_t ptr) const = 0; + virtual void trace_gpu_stream_creation( + c10::DeviceType device_type, + uintptr_t stream) const = 0; + virtual void trace_gpu_device_synchronization( + c10::DeviceType device_type) const = 0; + virtual void trace_gpu_stream_synchronization( + c10::DeviceType device_type, + uintptr_t stream) const = 0; + virtual void trace_gpu_event_synchronization( + c10::DeviceType device_type, + uintptr_t event) const = 0; + + virtual void reset_backward_hooks(const TensorImpl* self) const = 0; +}; + +struct C10_API PyInterpreter { + const PyInterpreterVTable* vtable_; + + PyInterpreter(const PyInterpreterVTable* vtable) : vtable_(vtable) {} + + const PyInterpreterVTable& operator*() const noexcept { + return *vtable_; + } + const PyInterpreterVTable* operator->() const noexcept { + return vtable_; + } + + // Disarm this PyInterpreter, making all of its methods noops. + // The vtable pointer is not an atomic at the moment, which means + // a disarm() invocation that is concurrent with active destructors + // is not thread safe and will trigger TSAN. My hope is that this + // situations doesn't ever actually happen; tensor destruction should + // quiesce when a dlclose happens, and any long lived tensors whose + // destructors would be disarmed here only begin the destruction process + // on process shutdown (long after the dlclose has occurred). + void disarm() noexcept; +}; + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/PyInterpreterHooks.h b/backends/pytorch/include/c10/core/impl/PyInterpreterHooks.h new file mode 100644 index 000000000..acd200356 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/PyInterpreterHooks.h @@ -0,0 +1,45 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include + +namespace c10::impl { + +// Minimal interface for PyInterpreter hooks +struct C10_API PyInterpreterHooksInterface { + virtual ~PyInterpreterHooksInterface() = default; + + // Get the PyInterpreter instance + // Stub implementation throws error when Python is not available + virtual PyInterpreter* getPyInterpreter() const { + TORCH_CHECK( + false, + "PyTorch was compiled without Python support. " + "Cannot access Python interpreter from C++."); + } +}; + +struct C10_API PyInterpreterHooksArgs{}; + +C10_DECLARE_REGISTRY( + PyInterpreterHooksRegistry, + PyInterpreterHooksInterface, + PyInterpreterHooksArgs); + +#define REGISTER_PYTHON_HOOKS(clsname) \ + C10_REGISTER_CLASS(PyInterpreterHooksRegistry, clsname, clsname) + +// Get the global PyInterpreter hooks instance +C10_API const PyInterpreterHooksInterface& getPyInterpreterHooks(); + +// Helper function to get the global interpreter +C10_API PyInterpreter* getGlobalPyInterpreter(); + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/PyObjectSlot.h b/backends/pytorch/include/c10/core/impl/PyObjectSlot.h new file mode 100644 index 000000000..8ba0688f6 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/PyObjectSlot.h @@ -0,0 +1,70 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include + +#include + +namespace torch::utils { +class PyObjectPreservation; +} + +namespace c10::impl { + +struct C10_API PyObjectSlot { + public: + PyObjectSlot() : pyobj_interpreter_(nullptr), pyobj_(nullptr) {} + + // Query the PyObject interpreter. This may return null if there is no + // interpreter. + PyInterpreter* pyobj_interpreter() const { + return pyobj_interpreter_.load(std::memory_order_acquire); + } + + PyInterpreter& load_pyobj_interpreter() const { + auto interpreter = pyobj_interpreter_.load(std::memory_order_acquire); + TORCH_INTERNAL_ASSERT( + interpreter, "cannot access PyObject for Tensor - no interpreter set"); + return *interpreter; + } + + PyObject* load_pyobj() const { + return pyobj_.load(std::memory_order_acquire); + } + + void store_pyobj(PyObject* obj) { + pyobj_.store(obj, std::memory_order_release); + } + + bool has_unique_reference() const { + PyObject* pyobj = load_pyobj(); + return pyobj != nullptr && load_pyobj_interpreter()->refcnt(pyobj) == 1; + } + + void clear() { + pyobj_.store(nullptr, std::memory_order_relaxed); + pyobj_interpreter_.store(nullptr, std::memory_order_relaxed); + } + + private: + // This is now always the global interpreter if the PyObject is set. + // Maybe we can remove this field some day... + std::atomic pyobj_interpreter_; + + // The PyObject representing this Tensor or nullptr. Ownership is managed + // by intrusive_ptr. By the time the PyObjectSlot is destroyed, this + // reference is already dead. + std::atomic pyobj_; + + friend class torch::utils::PyObjectPreservation; +}; + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/SizesAndStrides.h b/backends/pytorch/include/c10/core/impl/SizesAndStrides.h new file mode 100644 index 000000000..da3a9a0c4 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/SizesAndStrides.h @@ -0,0 +1,336 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include +#include +#include + +#define C10_SIZES_AND_STRIDES_MAX_INLINE_SIZE 5 + +namespace c10::impl { + +// Packed container for TensorImpl sizes and strides. +// This design improves on the previous approach of using a pair of +// c10::SmallVector by specializing for the operations we +// actually use and enforcing that the number of sizes is the same as +// the number of strides. The memory layout is as follows: +// +// 1 size_t for the size +// 5 eightbytes of inline sizes and 5 eightbytes of inline strides, OR pointer +// to out-of-line array +class C10_API SizesAndStrides { + public: + // TODO: different iterator types for sizes & strides to prevent + // mixing the two accidentally. + using sizes_iterator = int64_t*; + using sizes_const_iterator = const int64_t*; + using strides_iterator = int64_t*; + using strides_const_iterator = const int64_t*; + + SizesAndStrides() { + size_at_unchecked(0) = 0; + stride_at_unchecked(0) = 1; + } + + ~SizesAndStrides() { + if (C10_UNLIKELY(!isInline())) { + // NOLINTNEXTLINE(cppcoreguidelines-no-malloc) + free(outOfLineStorage_); + } + } + + SizesAndStrides(const SizesAndStrides& rhs) : size_(rhs.size_) { + if (C10_LIKELY(rhs.isInline())) { + copyDataInline(rhs); + } else { + allocateOutOfLineStorage(size_); + copyDataOutline(rhs); + } + } + + bool operator==(const SizesAndStrides& other) const { + if (size_ != other.size_) { + return false; + } + return !( + isInline() + ? std::memcmp( + inlineStorage_, other.inlineStorage_, sizeof(inlineStorage_)) + : std::memcmp( + outOfLineStorage_, + other.outOfLineStorage_, + storageBytes(size_))); + } + + bool operator!=(const SizesAndStrides& other) const { + return !(*this == other); + } + + SizesAndStrides& operator=(const SizesAndStrides& rhs) { + if (this == &rhs) { + return *this; + } + if (C10_LIKELY(rhs.isInline())) { + if (C10_UNLIKELY(!isInline())) { + // NOLINTNEXTLINE(cppcoreguidelines-no-malloc) + free(outOfLineStorage_); + } + copyDataInline(rhs); + } else { + if (isInline()) { + allocateOutOfLineStorage(rhs.size_); + } else { + resizeOutOfLineStorage(rhs.size_); + } + copyDataOutline(rhs); + } + size_ = rhs.size_; + return *this; + } + + // Move from rhs. rhs.size() == 0 afterwards. + SizesAndStrides(SizesAndStrides&& rhs) noexcept : size_(rhs.size_) { + if (C10_LIKELY(isInline())) { + memcpy(inlineStorage_, rhs.inlineStorage_, sizeof(inlineStorage_)); + } else { + outOfLineStorage_ = rhs.outOfLineStorage_; + rhs.outOfLineStorage_ = nullptr; + } + + rhs.size_ = 0; + } + + // Move from rhs. rhs.size() == 0 afterwards. + SizesAndStrides& operator=(SizesAndStrides&& rhs) noexcept { + if (this == &rhs) { + return *this; + } + if (C10_LIKELY(rhs.isInline())) { + if (C10_UNLIKELY(!isInline())) { + // NOLINTNEXTLINE(cppcoreguidelines-no-malloc) + free(outOfLineStorage_); + } + copyDataInline(rhs); + } else { + // They're outline. We're going to steal their vector. + if (!isInline()) { + // NOLINTNEXTLINE(cppcoreguidelines-no-malloc) + free(outOfLineStorage_); + } + outOfLineStorage_ = rhs.outOfLineStorage_; + rhs.outOfLineStorage_ = nullptr; + } + size_ = rhs.size_; + rhs.size_ = 0; + + return *this; + } + + size_t size() const noexcept { + return size_; + } + + const int64_t* sizes_data() const noexcept { + if (C10_LIKELY(isInline())) { + return &inlineStorage_[0]; + } else { + return &outOfLineStorage_[0]; + } + } + + int64_t* sizes_data() noexcept { + if (C10_LIKELY(isInline())) { + return &inlineStorage_[0]; + } else { + return &outOfLineStorage_[0]; + } + } + + sizes_const_iterator sizes_begin() const noexcept { + return sizes_data(); + } + + sizes_iterator sizes_begin() noexcept { + return sizes_data(); + } + + sizes_const_iterator sizes_end() const noexcept { + return sizes_begin() + size(); + } + + sizes_iterator sizes_end() noexcept { + return sizes_begin() + size(); + } + + IntArrayRef sizes_arrayref() const noexcept { + return IntArrayRef{sizes_data(), size()}; + } + + void set_sizes(IntArrayRef newSizes) { + resize(newSizes.size()); + std::copy(newSizes.begin(), newSizes.end(), sizes_begin()); + } + + void set_strides(IntArrayRef strides) { + TORCH_INTERNAL_ASSERT(strides.size() == size()); + std::copy(strides.begin(), strides.end(), strides_begin()); + } + + const int64_t* strides_data() const noexcept { + if (C10_LIKELY(isInline())) { + return &inlineStorage_[C10_SIZES_AND_STRIDES_MAX_INLINE_SIZE]; + } else { + return &outOfLineStorage_[size()]; + } + } + + int64_t* strides_data() noexcept { + if (C10_LIKELY(isInline())) { + return &inlineStorage_[C10_SIZES_AND_STRIDES_MAX_INLINE_SIZE]; + } else { + return &outOfLineStorage_[size()]; + } + } + + strides_const_iterator strides_begin() const noexcept { + if (C10_LIKELY(isInline())) { + return &inlineStorage_[C10_SIZES_AND_STRIDES_MAX_INLINE_SIZE]; + } else { + return &outOfLineStorage_[size()]; + } + } + + strides_iterator strides_begin() noexcept { + if (C10_LIKELY(isInline())) { + return &inlineStorage_[C10_SIZES_AND_STRIDES_MAX_INLINE_SIZE]; + } else { + return &outOfLineStorage_[size()]; + } + } + + strides_const_iterator strides_end() const noexcept { + return strides_begin() + size(); + } + + strides_iterator strides_end() noexcept { + return strides_begin() + size(); + } + + IntArrayRef strides_arrayref() const noexcept { + return IntArrayRef{strides_data(), size()}; + } + + // Size accessors. + int64_t size_at(size_t idx) const noexcept { + assert(idx < size()); + return sizes_data()[idx]; + } + + int64_t& size_at(size_t idx) noexcept { + assert(idx < size()); + return sizes_data()[idx]; + } + + int64_t size_at_unchecked(size_t idx) const noexcept { + return sizes_data()[idx]; + } + + int64_t& size_at_unchecked(size_t idx) noexcept { + return sizes_data()[idx]; + } + + // Size accessors. + int64_t stride_at(size_t idx) const noexcept { + assert(idx < size()); + return strides_data()[idx]; + } + + int64_t& stride_at(size_t idx) noexcept { + assert(idx < size()); + return strides_data()[idx]; + } + + int64_t stride_at_unchecked(size_t idx) const noexcept { + return strides_data()[idx]; + } + + int64_t& stride_at_unchecked(size_t idx) noexcept { + return strides_data()[idx]; + } + + void resize(size_t newSize) { + const auto oldSize = size(); + if (newSize == oldSize) { + return; + } + if (C10_LIKELY( + newSize <= C10_SIZES_AND_STRIDES_MAX_INLINE_SIZE && isInline())) { + if (oldSize < newSize) { + const auto bytesToZero = + (newSize - oldSize) * sizeof(inlineStorage_[0]); + memset(&inlineStorage_[oldSize], 0, bytesToZero); + memset( + &inlineStorage_[C10_SIZES_AND_STRIDES_MAX_INLINE_SIZE + oldSize], + 0, + bytesToZero); + } + size_ = newSize; + } else { + resizeSlowPath(newSize, oldSize); + } + } + + void resizeSlowPath(size_t newSize, size_t oldSize); + + private: + bool isInline() const noexcept { + return size_ <= C10_SIZES_AND_STRIDES_MAX_INLINE_SIZE; + } + + void copyDataInline(const SizesAndStrides& rhs) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(rhs.isInline()); + memcpy(inlineStorage_, rhs.inlineStorage_, sizeof(inlineStorage_)); + } + + static size_t storageBytes(size_t size) noexcept { + return size * 2 * sizeof(int64_t); + } + + void allocateOutOfLineStorage(size_t size) { + // NOLINTNEXTLINE(cppcoreguidelines-no-malloc) + outOfLineStorage_ = static_cast(malloc(storageBytes(size))); + TORCH_CHECK( + outOfLineStorage_, + "Could not allocate memory for Tensor SizesAndStrides!"); + } + + void resizeOutOfLineStorage(size_t newSize) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!isInline()); + outOfLineStorage_ = static_cast( + // NOLINTNEXTLINE(cppcoreguidelines-no-malloc) + realloc(outOfLineStorage_, storageBytes(newSize))); + TORCH_CHECK( + outOfLineStorage_, + "Could not allocate memory for Tensor SizesAndStrides!"); + } + + void copyDataOutline(const SizesAndStrides& rhs) noexcept { + memcpy(outOfLineStorage_, rhs.outOfLineStorage_, storageBytes(rhs.size_)); + } + + size_t size_{1}; + union { + int64_t* outOfLineStorage_; + // NOLINTNEXTLINE(*c-array*) + int64_t inlineStorage_[C10_SIZES_AND_STRIDES_MAX_INLINE_SIZE * 2]{}; + }; +}; + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/TorchDispatchModeTLS.h b/backends/pytorch/include/c10/core/impl/TorchDispatchModeTLS.h new file mode 100644 index 000000000..002bf4283 --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/TorchDispatchModeTLS.h @@ -0,0 +1,72 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace c10::impl { + +enum class TorchDispatchModeKey : int8_t { + FAKE, + PROXY, + FUNCTIONAL, + NUM_MODE_KEYS +}; + +using PyObject_TorchDispatchMode = SafePyObjectT; + +struct C10_API TorchDispatchModeTLS { + // This API is NOT invariant safe. + // It must not take in an infra mode that uses TorchDispatchModeKey + // If you're pushing an infra mode onto the stack, we expect + // you to use set_mode + static void push_non_infra_mode_onto_stack( + std::shared_ptr mode); + // Pops the top mode of the stack, + // giving precedence to user modes before attempting to pop + // any infra modes + static const std::shared_ptr pop_stack(); + // Returns the highest-priority infra mode on the stack, + // along with its mode key. + static const std:: + tuple, TorchDispatchModeKey> + pop_highest_infra_mode(); + + static const std::shared_ptr& get_stack_at( + int64_t idx); + static int64_t stack_len(); + + static const std::optional> + get_mode(TorchDispatchModeKey mode_key); + static const std::optional> + unset_mode(TorchDispatchModeKey mode_key); + static void set_mode( + const std::shared_ptr& mode, + TorchDispatchModeKey mode_key); + + static const TorchDispatchModeTLS& get_state(); + static void set_state(TorchDispatchModeTLS state); + + static bool any_modes_set(bool skip_infra_modes = false); + + private: + std::vector> stack_; + // Users are allowed to push multiple ProxyTorchDispatchMode objects onto the + // stack + // However, we only allow a single FakeTensorMode onto the stack at a time + // (Pushing additional FakeTensorModes onto the stack is a no-op) + std::array< + std::optional>, + static_cast(TorchDispatchModeKey::NUM_MODE_KEYS)> + infra_modes_; +}; + +C10_API bool dispatch_mode_enabled(); + +C10_API std::string to_string(TorchDispatchModeKey mode_key); + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/core/impl/VirtualGuardImpl.h b/backends/pytorch/include/c10/core/impl/VirtualGuardImpl.h new file mode 100644 index 000000000..16b1970bf --- /dev/null +++ b/backends/pytorch/include/c10/core/impl/VirtualGuardImpl.h @@ -0,0 +1,117 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include + +namespace c10::impl { + +/** + * An implementation of DeviceGuardImplInterface which delegates + * to virtual dispatch on the DeviceGuardImpl registry. + */ +class VirtualGuardImpl final : public DeviceGuardImplInterface { + public: + VirtualGuardImpl(DeviceType device_type) + : impl_(getDeviceGuardImpl(device_type)) {} + // This constructor exists purely for testing + VirtualGuardImpl(const DeviceGuardImplInterface* impl) : impl_(impl) {} + + // Copying and moving is OK! + VirtualGuardImpl(const VirtualGuardImpl&) = default; + VirtualGuardImpl& operator=(const VirtualGuardImpl&) = default; + VirtualGuardImpl(VirtualGuardImpl&&) noexcept = default; + VirtualGuardImpl& operator=(VirtualGuardImpl&&) noexcept = default; + ~VirtualGuardImpl() override = default; + + DeviceType type() const override { + return impl_->type(); + } + Device exchangeDevice(Device d) const override { + return impl_->exchangeDevice(d); + } + Device getDevice() const override { + return impl_->getDevice(); + } + void setDevice(Device d) const override { + impl_->setDevice(d); + } + void uncheckedSetDevice(Device d) const noexcept override { + impl_->uncheckedSetDevice(d); + } + Stream getStream(Device d) const override { + return impl_->getStream(d); + } + Stream getNewStream(Device d, int priority = 0) const override { + return impl_->getNewStream(d, priority); + } + Stream getDefaultStream(Device d) const override { + return impl_->getDefaultStream(d); + } + Stream getStreamFromGlobalPool(Device d, bool isHighPriority = false) + const override { + return impl_->getStreamFromGlobalPool(d, isHighPriority); + } + Stream exchangeStream(Stream s) const override { + return impl_->exchangeStream(s); + } + DeviceIndex deviceCount() const noexcept override { + return impl_->deviceCount(); + } + + DeviceCapability getDeviceCapability(Device d) const override { + return impl_->getDeviceCapability(d); + } + + // Event functions + void record( + void** event, + const Stream& stream, + const DeviceIndex device_index, + const EventFlag flag) const override { + impl_->record(event, stream, device_index, flag); + } + void block(void* event, const Stream& stream) const override { + impl_->block(event, stream); + } + bool queryEvent(void* event) const override { + return impl_->queryEvent(event); + } + void destroyEvent(void* event, const DeviceIndex device_index) + const noexcept override { + impl_->destroyEvent(event, device_index); + } + + bool queryStream(const Stream& stream) const override { + return impl_->queryStream(stream); + } + void synchronizeStream(const Stream& stream) const override { + impl_->synchronizeStream(stream); + } + + void recordDataPtrOnStream(const c10::DataPtr& data_ptr, const Stream& stream) + const override { + impl_->recordDataPtrOnStream(data_ptr, stream); + } + + double elapsedTime(void* event1, void* event2, const DeviceIndex device_index) + const override { + return impl_->elapsedTime(event1, event2, device_index); + } + + void synchronizeEvent(void* event) const override { + impl_->synchronizeEvent(event); + } + + void synchronizeDevice(const DeviceIndex device_index) const override { + impl_->synchronizeDevice(device_index); + } + + private: + const DeviceGuardImplInterface* impl_ = nullptr; +}; + +} // namespace c10::impl + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/macros/Export.h b/backends/pytorch/include/c10/macros/Export.h new file mode 100644 index 000000000..dfc4378c4 --- /dev/null +++ b/backends/pytorch/include/c10/macros/Export.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/macros/Macros.h b/backends/pytorch/include/c10/macros/Macros.h new file mode 100644 index 000000000..02fdbd4df --- /dev/null +++ b/backends/pytorch/include/c10/macros/Macros.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/AlignOf.h b/backends/pytorch/include/c10/util/AlignOf.h new file mode 100644 index 000000000..ce9fe9096 --- /dev/null +++ b/backends/pytorch/include/c10/util/AlignOf.h @@ -0,0 +1,181 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +//===--- AlignOf.h - Portable calculation of type alignment -----*- C++ -*-===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file defines the AlignedCharArray and AlignedCharArrayUnion classes. +// +//===----------------------------------------------------------------------===// + +// ATen: modified from llvm::AlignOf +// replaced LLVM_ALIGNAS with alignas + +#pragma once + +#include + +namespace c10 { + +/// \struct AlignedCharArray +/// \brief Helper for building an aligned character array type. +/// +/// This template is used to explicitly build up a collection of aligned +/// character array types. We have to build these up using a macro and explicit +/// specialization to cope with MSVC (at least till 2015) where only an +/// integer literal can be used to specify an alignment constraint. Once built +/// up here, we can then begin to indirect between these using normal C++ +/// template parameters. + +// MSVC requires special handling here. +#ifndef _MSC_VER + +template +struct AlignedCharArray { + // NOLINTNEXTLINE(*c-arrays) + alignas(Alignment) char buffer[Size]; +}; + +#else // _MSC_VER + +/// \brief Create a type with an aligned char buffer. +template +struct AlignedCharArray; + +// We provide special variations of this template for the most common +// alignments because __declspec(align(...)) doesn't actually work when it is +// a member of a by-value function argument in MSVC, even if the alignment +// request is something reasonably like 8-byte or 16-byte. Note that we can't +// even include the declspec with the union that forces the alignment because +// MSVC warns on the existence of the declspec despite the union member forcing +// proper alignment. + +template +struct AlignedCharArray<1, Size> { + union { + char aligned; + char buffer[Size]; + }; +}; + +template +struct AlignedCharArray<2, Size> { + union { + short aligned; + char buffer[Size]; + }; +}; + +template +struct AlignedCharArray<4, Size> { + union { + int aligned; + char buffer[Size]; + }; +}; + +template +struct AlignedCharArray<8, Size> { + union { + double aligned; + char buffer[Size]; + }; +}; + +// The rest of these are provided with a __declspec(align(...)) and we simply +// can't pass them by-value as function arguments on MSVC. + +#define AT_ALIGNEDCHARARRAY_TEMPLATE_ALIGNMENT(x) \ + template \ + struct AlignedCharArray { \ + __declspec(align(x)) char buffer[Size]; \ + }; + +AT_ALIGNEDCHARARRAY_TEMPLATE_ALIGNMENT(16) +AT_ALIGNEDCHARARRAY_TEMPLATE_ALIGNMENT(32) +AT_ALIGNEDCHARARRAY_TEMPLATE_ALIGNMENT(64) +AT_ALIGNEDCHARARRAY_TEMPLATE_ALIGNMENT(128) + +#undef AT_ALIGNEDCHARARRAY_TEMPLATE_ALIGNMENT + +#endif // _MSC_VER + +namespace detail { +template < + typename T1, + typename T2 = char, + typename T3 = char, + typename T4 = char, + typename T5 = char, + typename T6 = char, + typename T7 = char, + typename T8 = char, + typename T9 = char, + typename T10 = char> +class AlignerImpl { + T1 t1; + T2 t2; + T3 t3; + T4 t4; + T5 t5; + T6 t6; + T7 t7; + T8 t8; + T9 t9; + T10 t10; + + public: + AlignerImpl() = delete; +}; + +template < + typename T1, + typename T2 = char, + typename T3 = char, + typename T4 = char, + typename T5 = char, + typename T6 = char, + typename T7 = char, + typename T8 = char, + typename T9 = char, + typename T10 = char> +union SizerImpl { + // NOLINTNEXTLINE(*c-arrays) + char arr1[sizeof(T1)], arr2[sizeof(T2)], arr3[sizeof(T3)], arr4[sizeof(T4)], + arr5[sizeof(T5)], arr6[sizeof(T6)], arr7[sizeof(T7)], arr8[sizeof(T8)], + arr9[sizeof(T9)], arr10[sizeof(T10)]; +}; +} // end namespace detail + +/// \brief This union template exposes a suitably aligned and sized character +/// array member which can hold elements of any of up to ten types. +/// +/// These types may be arrays, structs, or any other types. The goal is to +/// expose a char array buffer member which can be used as suitable storage for +/// a placement new of any of these types. Support for more than ten types can +/// be added at the cost of more boilerplate. +template < + typename T1, + typename T2 = char, + typename T3 = char, + typename T4 = char, + typename T5 = char, + typename T6 = char, + typename T7 = char, + typename T8 = char, + typename T9 = char, + typename T10 = char> +struct AlignedCharArrayUnion + : AlignedCharArray< + alignof(detail::AlignerImpl), + sizeof(::c10::detail:: + SizerImpl)> {}; +} // end namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/ArrayRef.h b/backends/pytorch/include/c10/util/ArrayRef.h new file mode 100644 index 000000000..9da524e96 --- /dev/null +++ b/backends/pytorch/include/c10/util/ArrayRef.h @@ -0,0 +1,326 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +//===--- ArrayRef.h - Array Reference Wrapper -------------------*- C++ -*-===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// + +// ATen: modified from llvm::ArrayRef. +// removed llvm-specific functionality +// removed some implicit const -> non-const conversions that rely on +// complicated std::enable_if meta-programming +// removed a bunch of slice variants for simplicity... + +#pragma once + +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { +/// ArrayRef - Represent a constant reference to an array (0 or more elements +/// consecutively in memory), i.e. a start pointer and a length. It allows +/// various APIs to take consecutive elements easily and conveniently. +/// +/// This class does not own the underlying data, it is expected to be used in +/// situations where the data resides in some other buffer, whose lifetime +/// extends past that of the ArrayRef. For this reason, it is not in general +/// safe to store an ArrayRef. +/// +/// This is intended to be trivially copyable, so it should be passed by +/// value. +/// +/// NOTE: We have refactored out the headeronly parts of the ArrayRef struct +/// into HeaderOnlyArrayRef. As adding `virtual` would change the performance of +/// the underlying constexpr calls, we rely on apparent-type dispatch for +/// inheritance. This should be fine because their memory format is the same, +/// and it is never incorrect for ArrayRef to call HeaderOnlyArrayRef methods. +/// However, you should prefer to use ArrayRef when possible, because its use +/// of TORCH_CHECK will lead to better user-facing error messages. +template +// ArrayRef cannot be derived from. Normally, we would use `final` +// specifier to force this constraint at compile time. However, Intel +// compiler does not recognize ArrayRef as a class template (which is +// required in the definition of at::TensorAccessor, for instance) +// when `final` specifier is used. So, we cannot define ArrayRef as +// final because of the Intel compiler issue. +class ArrayRef : public HeaderOnlyArrayRef { + public: + /// @name Constructors, all inherited from HeaderOnlyArrayRef except for + /// SmallVector. As inherited constructors won't work with class template + /// argument deduction (CTAD) until C++23, we add deduction guides after + /// the class definition to enable CTAD. + /// @{ + + using HeaderOnlyArrayRef::HeaderOnlyArrayRef; + + /// Construct an ArrayRef from a SmallVector. This is templated in order to + /// avoid instantiating SmallVectorTemplateCommon whenever we + /// copy-construct an ArrayRef. + /// NOTE: this is the only constructor that is not inherited from + /// HeaderOnlyArrayRef. + template + /* implicit */ ArrayRef(const SmallVectorTemplateCommon& Vec) + : HeaderOnlyArrayRef(Vec.data(), Vec.size()) {} + + /// @} + /// @name Simple Operations, mostly inherited from HeaderOnlyArrayRef + /// @{ + + /// front - Get the first element. + /// We deviate from HeaderOnlyArrayRef by using TORCH_CHECK instead of + /// STD_TORCH_CHECK + constexpr const T& front() const { + TORCH_CHECK( + !this->empty(), "ArrayRef: attempted to access front() of empty list"); + return this->Data[0]; + } + + /// back - Get the last element. + /// We deviate from HeaderOnlyArrayRef by using TORCH_CHECK instead of + /// STD_TORCH_CHECK + constexpr const T& back() const { + TORCH_CHECK( + !this->empty(), "ArrayRef: attempted to access back() of empty list"); + return this->Data[this->Length - 1]; + } + + /// slice(n, m) - Take M elements of the array starting at element N + /// We deviate from HeaderOnlyArrayRef by using TORCH_CHECK instead of + /// STD_TORCH_CHECK + constexpr ArrayRef slice(size_t N, size_t M) const { + TORCH_CHECK( + N + M <= this->size(), + "ArrayRef: invalid slice, N = ", + N, + "; M = ", + M, + "; size = ", + this->size()); + return ArrayRef(this->data() + N, M); + } + + /// slice(n) - Chop off the first N elements of the array. + /// We deviate from HeaderOnlyArrayRef by using TORCH_CHECK instead of + /// STD_TORCH_CHECK + constexpr ArrayRef slice(size_t N) const { + TORCH_CHECK( + N <= this->size(), + "ArrayRef: invalid slice, N = ", + N, + "; size = ", + this->size()); + return slice(N, this->size() - N); // should this slice be this->slice? + } + + /// @} + /// @name Operator Overloads + /// @{ + + /// Vector compatibility + /// We deviate from HeaderOnlyArrayRef by using TORCH_CHECK instead of + /// STD_TORCH_CHECK + constexpr const T& at(size_t Index) const { + TORCH_CHECK( + Index < this->Length, + "ArrayRef: invalid index Index = ", + Index, + "; Length = ", + this->Length); + return this->Data[Index]; + } + + /// Disallow accidental assignment from a temporary. + /// + /// The declaration here is extra complicated so that "arrayRef = {}" + /// continues to select the move assignment operator. + template + std::enable_if_t, ArrayRef>& operator=( + // NOLINTNEXTLINE(cppcoreguidelines-missing-std-forward) + U&& Temporary) = delete; + + /// Disallow accidental assignment from a temporary. + /// + /// The declaration here is extra complicated so that "arrayRef = {}" + /// continues to select the move assignment operator. + template + std::enable_if_t, ArrayRef>& operator=( + std::initializer_list) = delete; + + /// @} +}; + +/// Deduction guides for ArrayRef to support CTAD with inherited constructors +/// These mirror the constructors inherited from HeaderOnlyArrayRef +/// @{ + +// Single element constructor +template +ArrayRef(const T&) -> ArrayRef; + +// Pointer and length constructor +template +ArrayRef(const T*, size_t) -> ArrayRef; + +// Range constructor (begin, end) +template +ArrayRef(const T*, const T*) -> ArrayRef; + +// Generic container constructor (anything with .data() and .size()) +template +ArrayRef(const Container&) -> ArrayRef< + std::remove_pointer_t().data())>>; + +// std::vector constructor +template +ArrayRef(const std::vector&) -> ArrayRef; + +// std::array constructor +template +ArrayRef(const std::array&) -> ArrayRef; + +// C array constructor +template +ArrayRef(const T (&)[N]) -> ArrayRef; + +// std::initializer_list constructor +template +ArrayRef(const std::initializer_list&) -> ArrayRef; + +/// @} + +template +std::ostream& operator<<(std::ostream& out, ArrayRef list) { + int i = 0; + out << '['; + for (const auto& e : list) { + if (i++ > 0) + out << ", "; + out << e; + } + out << ']'; + return out; +} + +/// @name ArrayRef Convenience constructors +/// @{ + +/// Construct an ArrayRef from a single element. +template +ArrayRef makeArrayRef(const T& OneElt) { + return OneElt; +} + +/// Construct an ArrayRef from a pointer and length. +template +ArrayRef makeArrayRef(const T* data, size_t length) { + return ArrayRef(data, length); +} + +/// Construct an ArrayRef from a range. +template +ArrayRef makeArrayRef(const T* begin, const T* end) { + return ArrayRef(begin, end); +} + +/// Construct an ArrayRef from a SmallVector. +template +ArrayRef makeArrayRef(const SmallVectorImpl& Vec) { + return Vec; +} + +/// Construct an ArrayRef from a SmallVector. +template +ArrayRef makeArrayRef(const SmallVector& Vec) { + return Vec; +} + +/// Construct an ArrayRef from a std::vector. +template +ArrayRef makeArrayRef(const std::vector& Vec) { + return Vec; +} + +/// Construct an ArrayRef from a std::array. +template +ArrayRef makeArrayRef(const std::array& Arr) { + return Arr; +} + +/// Construct an ArrayRef from an ArrayRef (no-op) (const) +template +ArrayRef makeArrayRef(const ArrayRef& Vec) { + return Vec; +} + +/// Construct an ArrayRef from an ArrayRef (no-op) +template +ArrayRef& makeArrayRef(ArrayRef& Vec) { + return Vec; +} + +/// Construct an ArrayRef from a C array. +template +// NOLINTNEXTLINE(*c-arrays*) +ArrayRef makeArrayRef(const T (&Arr)[N]) { + return ArrayRef(Arr); +} + +// WARNING: Template instantiation will NOT be willing to do an implicit +// conversions to get you to an c10::ArrayRef, which is why we need so +// many overloads. + +template +bool operator==(c10::ArrayRef a1, c10::ArrayRef a2) { + return a1.equals(a2); +} + +template +bool operator!=(c10::ArrayRef a1, c10::ArrayRef a2) { + return !a1.equals(a2); +} + +template +bool operator==(const std::vector& a1, c10::ArrayRef a2) { + return c10::ArrayRef(a1).equals(a2); +} + +template +bool operator!=(const std::vector& a1, c10::ArrayRef a2) { + return !c10::ArrayRef(a1).equals(a2); +} + +template +bool operator==(c10::ArrayRef a1, const std::vector& a2) { + return a1.equals(c10::ArrayRef(a2)); +} + +template +bool operator!=(c10::ArrayRef a1, const std::vector& a2) { + return !a1.equals(c10::ArrayRef(a2)); +} + +using IntArrayRef = ArrayRef; + +using IntList [[deprecated( + "This alias is deprecated because it doesn't make ownership semantics obvious. Use IntArrayRef instead!")]] = + ArrayRef; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/BFloat16.h b/backends/pytorch/include/c10/util/BFloat16.h new file mode 100644 index 000000000..90ca6b677 --- /dev/null +++ b/backends/pytorch/include/c10/util/BFloat16.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Backtrace.h b/backends/pytorch/include/c10/util/Backtrace.h new file mode 100644 index 000000000..0a9e8d2c2 --- /dev/null +++ b/backends/pytorch/include/c10/util/Backtrace.h @@ -0,0 +1,36 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#ifndef C10_UTIL_BACKTRACE_H_ +#define C10_UTIL_BACKTRACE_H_ + +#include +#include +#include +#include + +#include +#include + +namespace c10 { + +// Symbolizing the backtrace can be expensive; pass it around as a lazy string +// so it is symbolized only if actually needed. +using Backtrace = std::shared_ptr>; + +// DEPRECATED: Prefer get_lazy_backtrace(). +C10_API std::string get_backtrace( + size_t frames_to_skip = 0, + size_t maximum_number_of_frames = 64, + bool skip_python_frames = true); + +C10_API Backtrace get_lazy_backtrace( + size_t frames_to_skip = 0, + size_t maximum_number_of_frames = 64, + bool skip_python_frames = true); + +} // namespace c10 + +#endif // C10_UTIL_BACKTRACE_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/C++17.h b/backends/pytorch/include/c10/util/C++17.h new file mode 100644 index 000000000..f9e010daa --- /dev/null +++ b/backends/pytorch/include/c10/util/C++17.h @@ -0,0 +1,75 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once +#ifndef C10_UTIL_CPP17_H_ +#define C10_UTIL_CPP17_H_ + +#include +#include +#include +#include +#include + +#if !defined(__clang__) && !defined(_MSC_VER) && defined(__GNUC__) && \ + __GNUC__ < 9 +#error \ + "You're trying to build PyTorch with a too old version of GCC. We need GCC 9 or later." +#endif + +#if defined(__clang__) && __clang_major__ < 9 +#error \ + "You're trying to build PyTorch with a too old version of Clang. We need Clang 9 or later." +#endif + +#if (defined(_MSC_VER) && (!defined(_MSVC_LANG) || _MSVC_LANG < 201703L)) || \ + (!defined(_MSC_VER) && __cplusplus < 201703L) +#error You need C++17 to compile PyTorch +#endif + +#if defined(_WIN32) && (defined(min) || defined(max)) +#error Macro clash with min and max -- define NOMINMAX when compiling your program on Windows +#endif + +/* + * This header adds some polyfills with C++17 functionality + */ + +namespace c10 { + +namespace guts { + +#if defined(__HIP__) + +// Implementation from http://en.cppreference.com/w/cpp/utility/apply (but +// modified) +// TODO This is an incomplete implementation of std::apply, not working for +// member functions. +namespace detail { +template +C10_HOST_DEVICE constexpr auto apply_impl( + F&& f, + Tuple&& t, + std::index_sequence) { + return std::forward(f)(std::get(std::forward(t))...); +} +} // namespace detail + +template +C10_HOST_DEVICE constexpr auto apply(F&& f, Tuple&& t) { + return detail::apply_impl( + std::forward(f), + std::forward(t), + std::make_index_sequence< + std::tuple_size>::value>{}); +} + +#endif + +} // namespace guts + +} // namespace c10 + +#endif // C10_UTIL_CPP17_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/ConstexprCrc.h b/backends/pytorch/include/c10/util/ConstexprCrc.h new file mode 100644 index 000000000..56dd979ce --- /dev/null +++ b/backends/pytorch/include/c10/util/ConstexprCrc.h @@ -0,0 +1,137 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include + +namespace c10::util { + +namespace detail { +// NOLINTNEXTLINE(*c-arrays*) +constexpr uint64_t crc64_table[] = { + 0x0000000000000000, 0x7ad870c830358979, 0xf5b0e190606b12f2, + 0x8f689158505e9b8b, 0xc038e5739841b68f, 0xbae095bba8743ff6, + 0x358804e3f82aa47d, 0x4f50742bc81f2d04, 0xab28ecb46814fe75, + 0xd1f09c7c5821770c, 0x5e980d24087fec87, 0x24407dec384a65fe, + 0x6b1009c7f05548fa, 0x11c8790fc060c183, 0x9ea0e857903e5a08, + 0xe478989fa00bd371, 0x7d08ff3b88be6f81, 0x07d08ff3b88be6f8, + 0x88b81eabe8d57d73, 0xf2606e63d8e0f40a, 0xbd301a4810ffd90e, + 0xc7e86a8020ca5077, 0x4880fbd87094cbfc, 0x32588b1040a14285, + 0xd620138fe0aa91f4, 0xacf86347d09f188d, 0x2390f21f80c18306, + 0x594882d7b0f40a7f, 0x1618f6fc78eb277b, 0x6cc0863448deae02, + 0xe3a8176c18803589, 0x997067a428b5bcf0, 0xfa11fe77117cdf02, + 0x80c98ebf2149567b, 0x0fa11fe77117cdf0, 0x75796f2f41224489, + 0x3a291b04893d698d, 0x40f16bccb908e0f4, 0xcf99fa94e9567b7f, + 0xb5418a5cd963f206, 0x513912c379682177, 0x2be1620b495da80e, + 0xa489f35319033385, 0xde51839b2936bafc, 0x9101f7b0e12997f8, + 0xebd98778d11c1e81, 0x64b116208142850a, 0x1e6966e8b1770c73, + 0x8719014c99c2b083, 0xfdc17184a9f739fa, 0x72a9e0dcf9a9a271, + 0x08719014c99c2b08, 0x4721e43f0183060c, 0x3df994f731b68f75, + 0xb29105af61e814fe, 0xc849756751dd9d87, 0x2c31edf8f1d64ef6, + 0x56e99d30c1e3c78f, 0xd9810c6891bd5c04, 0xa3597ca0a188d57d, + 0xec09088b6997f879, 0x96d1784359a27100, 0x19b9e91b09fcea8b, + 0x636199d339c963f2, 0xdf7adabd7a6e2d6f, 0xa5a2aa754a5ba416, + 0x2aca3b2d1a053f9d, 0x50124be52a30b6e4, 0x1f423fcee22f9be0, + 0x659a4f06d21a1299, 0xeaf2de5e82448912, 0x902aae96b271006b, + 0x74523609127ad31a, 0x0e8a46c1224f5a63, 0x81e2d7997211c1e8, + 0xfb3aa75142244891, 0xb46ad37a8a3b6595, 0xceb2a3b2ba0eecec, + 0x41da32eaea507767, 0x3b024222da65fe1e, 0xa2722586f2d042ee, + 0xd8aa554ec2e5cb97, 0x57c2c41692bb501c, 0x2d1ab4dea28ed965, + 0x624ac0f56a91f461, 0x1892b03d5aa47d18, 0x97fa21650afae693, + 0xed2251ad3acf6fea, 0x095ac9329ac4bc9b, 0x7382b9faaaf135e2, + 0xfcea28a2faafae69, 0x8632586aca9a2710, 0xc9622c4102850a14, + 0xb3ba5c8932b0836d, 0x3cd2cdd162ee18e6, 0x460abd1952db919f, + 0x256b24ca6b12f26d, 0x5fb354025b277b14, 0xd0dbc55a0b79e09f, + 0xaa03b5923b4c69e6, 0xe553c1b9f35344e2, 0x9f8bb171c366cd9b, + 0x10e3202993385610, 0x6a3b50e1a30ddf69, 0x8e43c87e03060c18, + 0xf49bb8b633338561, 0x7bf329ee636d1eea, 0x012b592653589793, + 0x4e7b2d0d9b47ba97, 0x34a35dc5ab7233ee, 0xbbcbcc9dfb2ca865, + 0xc113bc55cb19211c, 0x5863dbf1e3ac9dec, 0x22bbab39d3991495, + 0xadd33a6183c78f1e, 0xd70b4aa9b3f20667, 0x985b3e827bed2b63, + 0xe2834e4a4bd8a21a, 0x6debdf121b863991, 0x1733afda2bb3b0e8, + 0xf34b37458bb86399, 0x8993478dbb8deae0, 0x06fbd6d5ebd3716b, + 0x7c23a61ddbe6f812, 0x3373d23613f9d516, 0x49aba2fe23cc5c6f, + 0xc6c333a67392c7e4, 0xbc1b436e43a74e9d, 0x95ac9329ac4bc9b5, + 0xef74e3e19c7e40cc, 0x601c72b9cc20db47, 0x1ac40271fc15523e, + 0x5594765a340a7f3a, 0x2f4c0692043ff643, 0xa02497ca54616dc8, + 0xdafce7026454e4b1, 0x3e847f9dc45f37c0, 0x445c0f55f46abeb9, + 0xcb349e0da4342532, 0xb1eceec59401ac4b, 0xfebc9aee5c1e814f, + 0x8464ea266c2b0836, 0x0b0c7b7e3c7593bd, 0x71d40bb60c401ac4, + 0xe8a46c1224f5a634, 0x927c1cda14c02f4d, 0x1d148d82449eb4c6, + 0x67ccfd4a74ab3dbf, 0x289c8961bcb410bb, 0x5244f9a98c8199c2, + 0xdd2c68f1dcdf0249, 0xa7f41839ecea8b30, 0x438c80a64ce15841, + 0x3954f06e7cd4d138, 0xb63c61362c8a4ab3, 0xcce411fe1cbfc3ca, + 0x83b465d5d4a0eece, 0xf96c151de49567b7, 0x76048445b4cbfc3c, + 0x0cdcf48d84fe7545, 0x6fbd6d5ebd3716b7, 0x15651d968d029fce, + 0x9a0d8ccedd5c0445, 0xe0d5fc06ed698d3c, 0xaf85882d2576a038, + 0xd55df8e515432941, 0x5a3569bd451db2ca, 0x20ed197575283bb3, + 0xc49581ead523e8c2, 0xbe4df122e51661bb, 0x3125607ab548fa30, + 0x4bfd10b2857d7349, 0x04ad64994d625e4d, 0x7e7514517d57d734, + 0xf11d85092d094cbf, 0x8bc5f5c11d3cc5c6, 0x12b5926535897936, + 0x686de2ad05bcf04f, 0xe70573f555e26bc4, 0x9ddd033d65d7e2bd, + 0xd28d7716adc8cfb9, 0xa85507de9dfd46c0, 0x273d9686cda3dd4b, + 0x5de5e64efd965432, 0xb99d7ed15d9d8743, 0xc3450e196da80e3a, + 0x4c2d9f413df695b1, 0x36f5ef890dc31cc8, 0x79a59ba2c5dc31cc, + 0x037deb6af5e9b8b5, 0x8c157a32a5b7233e, 0xf6cd0afa9582aa47, + 0x4ad64994d625e4da, 0x300e395ce6106da3, 0xbf66a804b64ef628, + 0xc5bed8cc867b7f51, 0x8aeeace74e645255, 0xf036dc2f7e51db2c, + 0x7f5e4d772e0f40a7, 0x05863dbf1e3ac9de, 0xe1fea520be311aaf, + 0x9b26d5e88e0493d6, 0x144e44b0de5a085d, 0x6e963478ee6f8124, + 0x21c640532670ac20, 0x5b1e309b16452559, 0xd476a1c3461bbed2, + 0xaeaed10b762e37ab, 0x37deb6af5e9b8b5b, 0x4d06c6676eae0222, + 0xc26e573f3ef099a9, 0xb8b627f70ec510d0, 0xf7e653dcc6da3dd4, + 0x8d3e2314f6efb4ad, 0x0256b24ca6b12f26, 0x788ec2849684a65f, + 0x9cf65a1b368f752e, 0xe62e2ad306bafc57, 0x6946bb8b56e467dc, + 0x139ecb4366d1eea5, 0x5ccebf68aecec3a1, 0x2616cfa09efb4ad8, + 0xa97e5ef8cea5d153, 0xd3a62e30fe90582a, 0xb0c7b7e3c7593bd8, + 0xca1fc72bf76cb2a1, 0x45775673a732292a, 0x3faf26bb9707a053, + 0x70ff52905f188d57, 0x0a2722586f2d042e, 0x854fb3003f739fa5, + 0xff97c3c80f4616dc, 0x1bef5b57af4dc5ad, 0x61372b9f9f784cd4, + 0xee5fbac7cf26d75f, 0x9487ca0fff135e26, 0xdbd7be24370c7322, + 0xa10fceec0739fa5b, 0x2e675fb4576761d0, 0x54bf2f7c6752e8a9, + 0xcdcf48d84fe75459, 0xb71738107fd2dd20, 0x387fa9482f8c46ab, + 0x42a7d9801fb9cfd2, 0x0df7adabd7a6e2d6, 0x772fdd63e7936baf, + 0xf8474c3bb7cdf024, 0x829f3cf387f8795d, 0x66e7a46c27f3aa2c, + 0x1c3fd4a417c62355, 0x935745fc4798b8de, 0xe98f353477ad31a7, + 0xa6df411fbfb21ca3, 0xdc0731d78f8795da, 0x536fa08fdfd90e51, + 0x29b7d047efec8728, +}; + +inline constexpr uint64_t crc64impl( + uint64_t accumulator, + const char* data, + size_t size) { + for (size_t i = 0; i < size; ++i) { + accumulator = + crc64_table[(accumulator ^ data[i]) & 0xFF] ^ (accumulator >> 8); + } + return accumulator; +} +} // namespace detail + +struct crc64_t final : IdWrapper { + constexpr crc64_t(uint64_t checksum) : IdWrapper(checksum) {} + constexpr uint64_t checksum() const { + return this->underlyingId(); + } +}; + +// CRC64 with Jones coefficients and an init value of 0. +inline constexpr crc64_t crc64(const char* str, size_t size) { + return crc64_t{detail::crc64impl(0, str, size)}; +} + +inline constexpr crc64_t crc64(std::string_view str) { + return crc64(str.data(), str.size()); +} +} // namespace c10::util + +// Allow usage of crc64_t in std::unordered_set +C10_DEFINE_HASH_FOR_IDWRAPPER(c10::util::crc64_t) + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Deprecated.h b/backends/pytorch/include/c10/util/Deprecated.h new file mode 100644 index 000000000..ccd1ac504 --- /dev/null +++ b/backends/pytorch/include/c10/util/Deprecated.h @@ -0,0 +1,7 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/DimVector.h b/backends/pytorch/include/c10/util/DimVector.h new file mode 100644 index 000000000..682b8f364 --- /dev/null +++ b/backends/pytorch/include/c10/util/DimVector.h @@ -0,0 +1,22 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include + +namespace c10 { + +constexpr size_t kDimVectorStaticSize = C10_SIZES_AND_STRIDES_MAX_INLINE_SIZE; + +/// A container for sizes or strides +using DimVector = SmallVector; +using SymDimVector = SmallVector; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Exception.h b/backends/pytorch/include/c10/util/Exception.h new file mode 100644 index 000000000..c6b4a7fa2 --- /dev/null +++ b/backends/pytorch/include/c10/util/Exception.h @@ -0,0 +1,875 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#ifndef C10_UTIL_EXCEPTION_H_ +#define C10_UTIL_EXCEPTION_H_ + +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include + +#if defined(_MSC_VER) && _MSC_VER <= 1900 +#define __func__ __FUNCTION__ +#endif + +namespace c10 { + +/// The primary ATen error class. +/// Provides a complete error message with source location information via +/// `what()`, and a more concise message via `what_without_backtrace()`. +/// Don't throw this directly; use TORCH_CHECK/TORCH_INTERNAL_ASSERT instead. +/// +/// NB: c10::Error is handled specially by the default torch to suppress the +/// backtrace, see torch/csrc/Exceptions.h +class C10_API Error : public std::exception { + private: + // The actual error message. + std::string msg_; + + // Context for the message (in order of decreasing specificity). Context will + // be automatically formatted appropriately, so it is not necessary to add + // extra leading/trailing newlines to strings inside this vector + std::vector context_; + + // The C++ backtrace at the point when this exception was raised. This + // may be empty if there is no valid backtrace. (We don't use optional + // here to reduce the dependencies this file has.) + Backtrace backtrace_; + + // These two are derived fields from msg_stack_ and backtrace_, but we need + // fields for the strings so that we can return a const char* (as the + // signature of std::exception requires). Currently, the invariant + // is that these fields are ALWAYS populated consistently with respect + // to msg_stack_ and backtrace_. + mutable OptimisticLazy what_; + std::string what_without_backtrace_; + + // This is a little debugging trick: you can stash a relevant pointer + // in caller, and then when you catch the exception, you can compare + // against pointers you have on hand to get more information about + // where the exception came from. In Caffe2, this is used to figure + // out which operator raised an exception. + const void* caller_; + + public: + // PyTorch-style Error constructor. NB: the implementation of this + // is actually in Logging.cpp + Error(SourceLocation source_location, std::string msg); + + // Caffe2-style error message + Error( + const char* file, + const uint32_t line, + const char* condition, + const std::string& msg, + Backtrace backtrace, + const void* caller = nullptr); + + // Base constructor + Error( + std::string msg, + Backtrace backtrace = nullptr, + const void* caller = nullptr); + + // Add some new context to the message stack. The last added context + // will be formatted at the end of the context list upon printing. + // WARNING: This method is O(n) in the size of the stack, so don't go + // wild adding a ridiculous amount of context to error messages. + void add_context(std::string msg); + + const std::string& msg() const { + return msg_; + } + + const std::vector& context() const { + return context_; + } + + const Backtrace& backtrace() const; + + /// Returns the complete error message, including the source location. + /// The returned pointer is invalidated if you call add_context() on + /// this object. + const char* what() const noexcept override; + + const void* caller() const noexcept { + return caller_; + } + + /// Returns only the error message string, without source location. + /// The returned pointer is invalidated if you call add_context() on + /// this object. + virtual const char* what_without_backtrace() const noexcept { + return what_without_backtrace_.c_str(); + } + + private: + void refresh_what(); + std::string compute_what(bool include_backtrace) const; +}; + +class C10_API Warning { + public: + class C10_API UserWarning{}; + class C10_API DeprecationWarning{}; + + using warning_variant_t = std::variant; + + Warning( + warning_variant_t type, + const SourceLocation& source_location, + std::string msg, + bool verbatim); + + Warning( + warning_variant_t type, + SourceLocation source_location, + const char* msg, + bool verbatim); + + Warning( + warning_variant_t type, + SourceLocation source_location, + ::c10::detail::CompileTimeEmptyString msg, + bool verbatim); + + // Getters for members + warning_variant_t type() const; + const SourceLocation& source_location() const; + const std::string& msg() const; + bool verbatim() const; + + private: + // The type of warning + warning_variant_t type_; + + // Where the warning happened. + SourceLocation source_location_; + + // The actual warning message. + std::string msg_; + + // See note: [Verbatim Warnings] + bool verbatim_; +}; + +using UserWarning = Warning::UserWarning; +using DeprecationWarning = Warning::DeprecationWarning; + +// Issue a warning with a given message. Dispatched to the current +// warning handler. +void C10_API warn(const Warning& warning); + +class C10_API WarningHandler { + public: + virtual ~WarningHandler() = default; + /// The default warning handler. Prints the message to stderr. + virtual void process(const Warning& warning); +}; + +namespace WarningUtils { + +// Note: [Verbatim Warnings] +// Warnings originating in C++ code can appear out-of-place to Python users: +// a user runs a line in Python, but the warning references a line in C++. +// Some parts of PyTorch, like the JIT, are cognizant of this mismatch +// and take care to map warnings back to the user's program, but most +// of PyTorch simply throws a context-free warning. To allow warning +// handlers to add context where appropriate, warn takes the +// "verbatim" flag. When this is false a warning handler might append +// the C++ warning to a Python warning message that relates the warning +// back to the user's program. Callers who have already accounted for +// context in their warnings should set verbatim to true so their warnings +// appear without modification. + +/// Sets the global warning handler. This is not thread-safe, so it should +/// generally be called once during initialization or while holding the GIL +/// for programs that use python. +/// User is responsible for keeping the WarningHandler alive until +/// it is not needed. +C10_API void set_warning_handler(WarningHandler* handler) noexcept(true); +/// Gets the global warning handler. +C10_API WarningHandler* get_warning_handler() noexcept(true); + +class C10_API WarningHandlerGuard { + WarningHandler* prev_handler_; + + public: + WarningHandlerGuard(WarningHandler* new_handler) + : prev_handler_(c10::WarningUtils::get_warning_handler()) { + c10::WarningUtils::set_warning_handler(new_handler); + } + WarningHandlerGuard(WarningHandlerGuard&& other) = delete; + WarningHandlerGuard(const WarningHandlerGuard&) = delete; + WarningHandlerGuard& operator=(const WarningHandlerGuard&) = delete; + WarningHandlerGuard& operator=(WarningHandlerGuard&&) = delete; + ~WarningHandlerGuard() { + c10::WarningUtils::set_warning_handler(prev_handler_); + } +}; + +/// The TORCH_WARN_ONCE macro is difficult to test for. Use +/// setWarnAlways(true) to turn it into TORCH_WARN, which can be +/// tested for more easily. +C10_API void set_warnAlways(bool /*setting*/) noexcept(true); +C10_API bool get_warnAlways() noexcept(true); + +// A RAII guard that sets warn_always (not thread-local) on +// construction, and sets it back to the original value upon destruction. +struct C10_API WarnAlways { + public: + explicit WarnAlways(bool setting = true); + ~WarnAlways(); + + private: + bool prev_setting; +}; + +} // namespace WarningUtils + +// Like Error, but we always report the C++ backtrace, instead of only +// reporting when TORCH_SHOW_CPP_STACKTRACES +class C10_API ErrorAlwaysShowCppStacktrace : public Error { + using Error::Error; + const char* what_without_backtrace() const noexcept override { + return what(); + } +}; + +// Used in ATen for out-of-bound indices that can reasonably only be detected +// lazily inside a kernel (See: advanced indexing). These turn into +// IndexError when they cross to Python. +class C10_API IndexError : public Error { + using Error::Error; +}; + +// Used in ATen for invalid values. These turn into +// ValueError when they cross to Python. +class C10_API ValueError : public Error { + using Error::Error; +}; + +// Used in ATen for invalid types. These turn into +// TypeError when they cross to Python. +class C10_API TypeError : public Error { + using Error::Error; +}; + +// Used in ATen for functionality that is not implemented. These turn into +// NotImplementedError when they cross to Python. +class C10_API NotImplementedError : public Error { + using Error::Error; +}; + +// Used in ATen for buffer-related errors, e.g. trying to create a DLPack of +// an unsupported device. These turn into BufferError when they cross to +// Python. +class C10_API BufferError : public Error { + using Error::Error; +}; + +// Used in ATen for non finite indices. These turn into +// ExitException when they cross to Python. +class C10_API EnforceFiniteError : public Error { + using Error::Error; +}; + +// Used in Onnxifi backend lowering. These turn into +// ExitException when they cross to Python. +class C10_API OnnxfiBackendSystemError : public Error { + using Error::Error; +}; + +// Used for numerical errors from the linalg module. These +// turn into LinAlgError when they cross into Python. +class C10_API LinAlgError : public Error { + using Error::Error; +}; + +class C10_API OutOfMemoryError : public Error { + using Error::Error; +}; + +// Used for handling syntactic errors in input arguments. +// These turn into SyntaxError when the cross into Python. +class C10_API SyntaxError : public Error { + using Error::Error; +}; + +// Raised when accelerator API call hits an error. +// These turn into AcceleratorError when the cross into Python +class C10_API AcceleratorError : public Error { + int32_t error_code; + + public: + AcceleratorError(SourceLocation loc, int32_t code, const std::string& msg) + : Error(loc, msg), error_code(code) {} + int32_t get_error_code() const { + return error_code; + } +}; + +// Base error type for all distributed errors. +// These turn into DistError when they cross into Python. +class C10_API DistError : public Error { + using Error::Error; +}; + +// Used for collective communication library errors from the distributed module. +// These turn into DistBackendError when they cross into Python. +class C10_API DistBackendError : public DistError { + using DistError::DistError; +}; + +// Used for errors originating from the store. +// These turn into DistStoreError when they cross into Python. +class C10_API DistStoreError : public DistError { + using DistError::DistError; +}; + +// Used for errors originating from the TCP/IP stack and not from collective +// libraries. These turn into DistNetworkError when they cross into Python. +class C10_API DistNetworkError : public DistError { + using DistError::DistError; +}; + +// Raised when a queue is empty and a non-blocking pop is called. +// Translated to torch.distributed.QueueEmptyError in Python +class C10_API DistQueueEmptyError : public DistStoreError { + using DistStoreError::DistStoreError; +}; + +// A utility function to return an exception std::string by prepending its +// exception type before its what() content +C10_API std::string GetExceptionString(const std::exception& e); + +} // namespace c10 + +// Private helper macro for implementing TORCH_INTERNAL_ASSERT and TORCH_CHECK +// +// Note: In the debug build With MSVC, __LINE__ might be of long type (a.k.a +// int32_t), which is different from the definition of `SourceLocation` that +// requires unsigned int (a.k.a uint32_t) and may cause a compile error with the +// message: error C2397: conversion from 'long' to 'uint32_t' requires a +// narrowing conversion Here the static cast is used to pass the build. if this +// is used inside a lambda the __func__ macro expands to operator(), which isn't +// very useful, but hard to fix in a macro so suppressing the warning. +#define C10_THROW_ERROR(err_type, msg) \ + throw ::c10::err_type( \ + {__func__, __FILE__, static_cast(__LINE__)}, msg) + +#define C10_BUILD_ERROR(err_type, msg) \ + ::c10::err_type({__func__, __FILE__, static_cast(__LINE__)}, msg) + +// Private helper macro for workaround MSVC misexpansion of nested macro +// invocations involving __VA_ARGS__. See +// https://stackoverflow.com/questions/5134523/msvc-doesnt-expand-va-args-correctly +#define C10_EXPAND_MSVC_WORKAROUND(x) x + +#include + +// ---------------------------------------------------------------------------- +// Error reporting macros +// ---------------------------------------------------------------------------- + +#ifdef STRIP_ERROR_MESSAGES +#define TORCH_RETHROW(e, ...) \ + do { \ + (void)e; /* Suppress unused variable warning */ \ + throw; \ + } while (false) +#else +#define TORCH_RETHROW(e, ...) \ + do { \ + e.add_context(::c10::str(__VA_ARGS__)); \ + throw; \ + } while (false) +#endif + +// A utility macro to provide assert()-like functionality; that is, enforcement +// of internal invariants in code. It supports an arbitrary number of extra +// arguments (evaluated only on failure), which will be printed in the assert +// failure message using operator<< (this is useful to print some variables +// which may be useful for debugging.) +// +// Usage: +// TORCH_INTERNAL_ASSERT(should_be_true); +// TORCH_INTERNAL_ASSERT(x == 0, "x = ", x); +// +// Assuming no bugs in PyTorch, the conditions tested by this macro should +// always be true; e.g., it should be possible to disable all of these +// conditions without changing observable user behavior. If you would like to +// do error reporting for user input, please use TORCH_CHECK instead. +// +// NOTE: It is SAFE to use this macro in production code; on failure, this +// simply raises an exception, it does NOT unceremoniously quit the process +// (unlike assert()). +// +#ifdef STRIP_ERROR_MESSAGES +#define TORCH_INTERNAL_ASSERT(cond, ...) \ + if (C10_UNLIKELY_OR_CONST(!(cond))) { \ + ::c10::detail::torchCheckFail( \ + __func__, \ + __FILE__, \ + static_cast(__LINE__), \ + #cond " INTERNAL ASSERT FAILED at " C10_STRINGIZE(__FILE__)); \ + } +#else +// It would be nice if we could build a combined string literal out of +// the TORCH_INTERNAL_ASSERT prefix and a user-provided string literal +// as the first argument, but there doesn't seem to be any good way to +// do that while still supporting having a first argument that isn't a +// string literal. +#define TORCH_INTERNAL_ASSERT(cond, ...) \ + if (C10_UNLIKELY_OR_CONST(!(cond))) { \ + ::c10::detail::torchInternalAssertFail( \ + __func__, \ + __FILE__, \ + static_cast(__LINE__), \ + #cond \ + " INTERNAL ASSERT FAILED at " C10_STRINGIZE(__FILE__) ":" C10_STRINGIZE( \ + __LINE__) ", please report a bug to PyTorch. ", \ + c10::str(__VA_ARGS__)); \ + } +#endif + +// A utility macro to make it easier to test for error conditions from user +// input. Like TORCH_INTERNAL_ASSERT, it supports an arbitrary number of extra +// arguments (evaluated only on failure), which will be printed in the error +// message using operator<< (e.g., you can pass any object which has +// operator<< defined. Most objects in PyTorch have these definitions!) +// +// Usage: +// TORCH_CHECK(should_be_true); // A default error message will be provided +// // in this case; but we recommend writing an +// // explicit error message, as it is more +// // user friendly. +// TORCH_CHECK(x == 0, "Expected x to be 0, but got ", x); +// +// On failure, this macro will raise an exception. If this exception propagates +// to Python, it will convert into a Python RuntimeError. +// +// NOTE: It is SAFE to use this macro in production code; on failure, this +// simply raises an exception, it does NOT unceremoniously quit the process +// (unlike CHECK() from glog.) +// +#define TORCH_CHECK_WITH(error_t, cond, ...) \ + TORCH_CHECK_WITH_MSG(error_t, cond, "", __VA_ARGS__) + +#ifdef STRIP_ERROR_MESSAGES +#define TORCH_CHECK_MSG(cond, type, ...) \ + (#cond #type " CHECK FAILED at " C10_STRINGIZE(__FILE__)) +#define TORCH_CHECK_WITH_MSG(error_t, cond, type, ...) \ + if (C10_UNLIKELY_OR_CONST(!(cond))) { \ + C10_THROW_ERROR(Error, TORCH_CHECK_MSG(cond, type, __VA_ARGS__)); \ + } +#else + +namespace c10::detail { +template +auto torchCheckMsgImpl(const char* /*msg*/, const Args&... args) { + return ::c10::str(args...); +} +inline C10_API const char* torchCheckMsgImpl(const char* msg) { + return msg; +} +// If there is just 1 user-provided C-string argument, use it. +inline C10_API const char* torchCheckMsgImpl( + const char* /*msg*/, + const char* args) { + return args; +} +} // namespace c10::detail + +#define TORCH_CHECK_MSG(cond, type, ...) \ + (::c10::detail::torchCheckMsgImpl( \ + "Expected " #cond \ + " to be true, but got false. " \ + "(Could this error message be improved? If so, " \ + "please report an enhancement request to PyTorch.)", \ + ##__VA_ARGS__)) +#define TORCH_CHECK_WITH_MSG(error_t, cond, type, ...) \ + if (C10_UNLIKELY_OR_CONST(!(cond))) { \ + C10_THROW_ERROR(error_t, TORCH_CHECK_MSG(cond, type, __VA_ARGS__)); \ + } +#endif + +namespace c10::detail { + +[[noreturn]] C10_API void torchCheckFail( + const char* func, + const char* file, + uint32_t line, + const std::string& msg); +[[noreturn]] C10_API void torchCheckFail( + const char* func, + const char* file, + uint32_t line, + const char* msg); + +// The c10::str() call that creates userMsg can have 1 of 3 return +// types depending on the number and types of arguments passed to +// TORCH_INTERNAL_ASSERT. 0 arguments will get a +// CompileTimeEmptyString, 1 const char * will be passed straight +// through, and anything else will get converted to std::string. +[[noreturn]] C10_API void torchInternalAssertFail( + const char* func, + const char* file, + uint32_t line, + const char* condMsg, + const char* userMsg); +[[noreturn]] inline C10_API void torchInternalAssertFail( + const char* func, + const char* file, + uint32_t line, + const char* condMsg, + ::c10::detail::CompileTimeEmptyString /*userMsg*/) { + torchCheckFail(func, file, line, condMsg); +} +[[noreturn]] C10_API void torchInternalAssertFail( + const char* func, + const char* file, + uint32_t line, + const char* condMsg, + const std::string& userMsg); + +} // namespace c10::detail + +#ifdef STANDALONE_TORCH_HEADER + +// TORCH_CHECK throws std::runtime_error instead of c10::Error which is +// useful when certain headers are used in a libtorch-independent way, +// e.g. when Vectorized is used in AOTInductor generated code. +#ifdef STRIP_ERROR_MESSAGES +#define TORCH_CHECK(cond, ...) \ + if (C10_UNLIKELY_OR_CONST(!(cond))) { \ + throw std::runtime_error(TORCH_CHECK_MSG( \ + cond, \ + "", \ + __func__, \ + ", ", \ + __FILE__, \ + ":", \ + __LINE__, \ + ", ", \ + __VA_ARGS__)); \ + } +#else +#define TORCH_CHECK(cond, ...) \ + if (C10_UNLIKELY_OR_CONST(!(cond))) { \ + throw std::runtime_error(TORCH_CHECK_MSG( \ + cond, \ + "", \ + __func__, \ + ", ", \ + __FILE__, \ + ":", \ + __LINE__, \ + ", ", \ + ##__VA_ARGS__)); \ + } +#endif + +#else + +#ifdef STRIP_ERROR_MESSAGES +#define TORCH_CHECK(cond, ...) \ + if (C10_UNLIKELY_OR_CONST(!(cond))) { \ + ::c10::detail::torchCheckFail( \ + __func__, \ + __FILE__, \ + static_cast(__LINE__), \ + TORCH_CHECK_MSG(cond, "", __VA_ARGS__)); \ + } +#else +#define TORCH_CHECK(cond, ...) \ + if (C10_UNLIKELY_OR_CONST(!(cond))) { \ + ::c10::detail::torchCheckFail( \ + __func__, \ + __FILE__, \ + static_cast(__LINE__), \ + TORCH_CHECK_MSG(cond, "", ##__VA_ARGS__)); \ + } +#endif + +#endif + +// An utility macro that does what `TORCH_CHECK` does if compiled in the host +// code, otherwise does nothing. Supposed to be used in the code shared between +// host and device code as an alternative for `TORCH_CHECK`. +#if defined(__CUDACC__) || defined(__HIPCC__) +#define TORCH_CHECK_IF_NOT_ON_CUDA(cond, ...) +#else +#define TORCH_CHECK_IF_NOT_ON_CUDA(cond, ...) TORCH_CHECK(cond, ##__VA_ARGS__) +#endif + +// Debug only version of TORCH_INTERNAL_ASSERT. This macro only checks in debug +// build, and does nothing in release build. It is appropriate to use +// in situations where you want to add an assert to a hotpath, but it is +// too expensive to run this assert on production builds. +#ifdef NDEBUG +// Optimized version - generates no code. +#define TORCH_INTERNAL_ASSERT_DEBUG_ONLY(...) \ + while (false) \ + C10_EXPAND_MSVC_WORKAROUND(TORCH_INTERNAL_ASSERT(__VA_ARGS__)) +#else +#define TORCH_INTERNAL_ASSERT_DEBUG_ONLY(...) \ + C10_EXPAND_MSVC_WORKAROUND(TORCH_INTERNAL_ASSERT(__VA_ARGS__)) +#endif + +// TODO: We're going to get a lot of similar looking string literals +// this way; check if this actually affects binary size. + +// Like TORCH_CHECK, but raises LinAlgError instead of Error. +#define TORCH_CHECK_LINALG(cond, ...) \ + TORCH_CHECK_WITH_MSG(LinAlgError, cond, "LINALG", __VA_ARGS__) + +// Like TORCH_CHECK, but raises IndexErrors instead of Errors. +#define TORCH_CHECK_INDEX(cond, ...) \ + TORCH_CHECK_WITH_MSG(IndexError, cond, "INDEX", __VA_ARGS__) + +// Like TORCH_CHECK, but raises ValueErrors instead of Errors. +#define TORCH_CHECK_VALUE(cond, ...) \ + TORCH_CHECK_WITH_MSG(ValueError, cond, "VALUE", __VA_ARGS__) + +// Like TORCH_CHECK, but raises TypeErrors instead of Errors. +#define TORCH_CHECK_TYPE(cond, ...) \ + TORCH_CHECK_WITH_MSG(TypeError, cond, "TYPE", __VA_ARGS__) + +// Like TORCH_CHECK, but raises NotImplementedErrors instead of Errors. +#define TORCH_CHECK_NOT_IMPLEMENTED(cond, ...) \ + TORCH_CHECK_WITH_MSG(NotImplementedError, cond, "TYPE", __VA_ARGS__) + +// Like TORCH_CHECK, but raises BufferError instead of Errors. +#define TORCH_CHECK_BUFFER(cond, ...) \ + TORCH_CHECK_WITH_MSG(BufferError, cond, "TYPE", __VA_ARGS__) + +#define TORCH_CHECK_ALWAYS_SHOW_CPP_STACKTRACE(cond, ...) \ + TORCH_CHECK_WITH_MSG( \ + ErrorAlwaysShowCppStacktrace, cond, "TYPE", ##__VA_ARGS__) + +#ifdef STRIP_ERROR_MESSAGES +#define WARNING_MESSAGE_STRING(...) \ + ::c10::detail::CompileTimeEmptyString {} +#else +#define WARNING_MESSAGE_STRING(...) ::c10::str(__VA_ARGS__) +#endif + +// Report a warning to the user. Accepts an arbitrary number of extra +// arguments which are concatenated into the warning message using operator<< +// +#ifdef DISABLE_WARN +#define _TORCH_WARN_WITH(...) ((void)0); +#else +#define _TORCH_WARN_WITH(warning_t, ...) \ + ::c10::warn(::c10::Warning( \ + warning_t(), \ + {__func__, __FILE__, static_cast(__LINE__)}, \ + WARNING_MESSAGE_STRING(__VA_ARGS__), \ + false)); +#endif + +#define TORCH_WARN(...) _TORCH_WARN_WITH(::c10::UserWarning, __VA_ARGS__); + +#define TORCH_WARN_DEPRECATION(...) \ + _TORCH_WARN_WITH(::c10::DeprecationWarning, __VA_ARGS__); + +// Report a warning to the user only once. Accepts an arbitrary number of extra +// arguments which are concatenated into the warning message using operator<< +// +#define _TORCH_WARN_ONCE(...) \ + [[maybe_unused]] static const auto C10_ANONYMOUS_VARIABLE( \ + torch_warn_once_) = [&] { \ + TORCH_WARN(__VA_ARGS__); \ + return true; \ + }() + +#ifdef DISABLE_WARN +#define TORCH_WARN_ONCE(...) ((void)0); +#else +#define TORCH_WARN_ONCE(...) \ + if (::c10::WarningUtils::get_warnAlways()) { \ + TORCH_WARN(__VA_ARGS__); \ + } else { \ + _TORCH_WARN_ONCE(__VA_ARGS__); \ + } +#endif + +// Report an error with a specific argument +// NOTE: using the argument name in TORCH_CHECK's message is preferred +#define TORCH_CHECK_ARG(cond, argN, ...) \ + TORCH_CHECK(cond, "invalid argument ", argN, ": ", __VA_ARGS__) + +#ifndef FATAL_IF +#ifdef C10_USE_GLOG +#define FATAL_IF(condition) \ + condition ? (void)0 \ + : ::c10::LoggerVoidify() & \ + ::c10::MessageLogger(__FILE__, __LINE__, ::google::GLOG_FATAL) \ + .stream() +#else +#define FATAL_IF(condition) \ + condition ? (void)0 \ + : ::c10::LoggerVoidify() & \ + ::c10::MessageLogger(__FILE__, __LINE__, ::c10::GLOG_FATAL).stream() +#endif +#endif + +#ifndef NON_FATAL_IF +#ifdef C10_USE_GLOG +#define NON_FATAL_IF(condition) \ + condition ? (void)0 \ + : ::c10::LoggerVoidify() & \ + ::c10::MessageLogger( \ + __FILE__, __LINE__, ::google::GLOG_FATAL, false) \ + .stream() +#else +#define NON_FATAL_IF(condition) \ + condition ? (void)0 \ + : ::c10::LoggerVoidify() & \ + ::c10::MessageLogger(__FILE__, __LINE__, ::c10::GLOG_FATAL, false) \ + .stream() +#endif +#endif + +// Binary comparison check macros +#define TORCH_CHECK_OP(val1, val2, op) \ + NON_FATAL_IF(((val1)op(val2))) \ + << "Check failed: " #val1 " " #op " " #val2 " (" << (val1) << " vs. " \ + << (val2) << "). " + +#define TORCH_DCHECK_OP(val1, val2, op) \ + FATAL_IF(((val1)op(val2))) << "Check failed: " #val1 " " #op " " #val2 " (" \ + << (val1) << " vs. " << (val2) << "). " + +#define TORCH_CHECK_EQ(val1, val2) TORCH_CHECK_OP(val1, val2, ==) +#define TORCH_CHECK_NE(val1, val2) TORCH_CHECK_OP(val1, val2, !=) +#define TORCH_CHECK_LE(val1, val2) TORCH_CHECK_OP(val1, val2, <=) +#define TORCH_CHECK_LT(val1, val2) TORCH_CHECK_OP(val1, val2, <) +#define TORCH_CHECK_GE(val1, val2) TORCH_CHECK_OP(val1, val2, >=) +#define TORCH_CHECK_GT(val1, val2) TORCH_CHECK_OP(val1, val2, >) + +// Debug versions of TORCH_CHECK_OP macros +#ifndef NDEBUG +#define TORCH_DCHECK_EQ(val1, val2) TORCH_DCHECK_OP(val1, val2, ==) +#define TORCH_DCHECK_NE(val1, val2) TORCH_DCHECK_OP(val1, val2, !=) +#define TORCH_DCHECK_LE(val1, val2) TORCH_DCHECK_OP(val1, val2, <=) +#define TORCH_DCHECK_LT(val1, val2) TORCH_DCHECK_OP(val1, val2, <) +#define TORCH_DCHECK_GE(val1, val2) TORCH_DCHECK_OP(val1, val2, >=) +#define TORCH_DCHECK_GT(val1, val2) TORCH_DCHECK_OP(val1, val2, >) +#else // !NDEBUG +// Optimized versions - generate no code +#define TORCH_DCHECK_EQ(val1, val2) \ + while (false) \ + TORCH_DCHECK_OP(val1, val2, ==) +#define TORCH_DCHECK_NE(val1, val2) \ + while (false) \ + TORCH_DCHECK_OP(val1, val2, !=) +#define TORCH_DCHECK_LE(val1, val2) \ + while (false) \ + TORCH_DCHECK_OP(val1, val2, <=) +#define TORCH_DCHECK_LT(val1, val2) \ + while (false) \ + TORCH_DCHECK_OP(val1, val2, <) +#define TORCH_DCHECK_GE(val1, val2) \ + while (false) \ + TORCH_DCHECK_OP(val1, val2, >=) +#define TORCH_DCHECK_GT(val1, val2) \ + while (false) \ + TORCH_DCHECK_OP(val1, val2, >) +#endif // NDEBUG + +// Null pointer check macro +#define TORCH_CHECK_NOTNULL(val) \ + ::c10::CheckNotNull(__FILE__, __LINE__, #val, (val), false) + +#ifndef NDEBUG +#define TORCH_DCHECK_NOTNULL(val) \ + ::c10::CheckNotNull(__FILE__, __LINE__, #val, (val), true) +#else // !NDEBUG +#define TORCH_DCHECK_NOTNULL(val) \ + while (false) \ + TORCH_CHECK_NOTNULL(val) +#endif // NDEBUG + +// ---------------------------------------------------------------------------- +// Deprecated macros +// ---------------------------------------------------------------------------- + +namespace c10::detail { + +/* +// Deprecation disabled until we fix sites in our codebase +[[deprecated("AT_ERROR(msg) is deprecated, use TORCH_CHECK(false, msg) +instead.")]] +*/ +inline void deprecated_AT_ERROR() {} + +/* +// Deprecation disabled until we fix sites in our codebase +[[deprecated("AT_ASSERT is deprecated, if you mean to indicate an +internal invariant failure, use " \ + "TORCH_INTERNAL_ASSERT instead; if you mean to do user +error checking, use " \ "TORCH_CHECK. See +https://github.com/pytorch/pytorch/issues/20287 for more details.")]] +*/ +inline void deprecated_AT_ASSERT() {} + +/* +// Deprecation disabled until we fix sites in our codebase +[[deprecated("AT_ASSERTM is deprecated, if you mean to indicate an +internal invariant failure, use " \ + "TORCH_INTERNAL_ASSERT instead; if you mean to do user +error checking, use " \ "TORCH_CHECK. See +https://github.com/pytorch/pytorch/issues/20287 for more details.")]] +*/ +inline void deprecated_AT_ASSERTM() {} + +} // namespace c10::detail + +// Deprecated alias; this alias was deprecated because people kept mistakenly +// using it for user error checking. Use TORCH_INTERNAL_ASSERT or TORCH_CHECK +// instead. See https://github.com/pytorch/pytorch/issues/20287 for more +// details. +#define AT_ASSERT(...) \ + do { \ + ::c10::detail::deprecated_AT_ASSERT(); \ + C10_EXPAND_MSVC_WORKAROUND(TORCH_INTERNAL_ASSERT(__VA_ARGS__)); \ + } while (false) + +// Deprecated alias, like AT_ASSERT. The new TORCH_INTERNAL_ASSERT macro +// supports both 0-ary and variadic calls, so having a separate +// message-accepting macro is not necessary. +// +// NB: we MUST include cond explicitly here, as MSVC will miscompile the macro +// expansion, shunting all of __VA_ARGS__ to cond. An alternate workaround +// can be seen at +// https://stackoverflow.com/questions/5134523/msvc-doesnt-expand-va-args-correctly +#define AT_ASSERTM(cond, ...) \ + do { \ + ::c10::detail::deprecated_AT_ASSERTM(); \ + C10_EXPAND_MSVC_WORKAROUND(TORCH_INTERNAL_ASSERT(cond, __VA_ARGS__)); \ + } while (false) + +// Deprecated alias; this alias was deprecated because it represents extra API +// surface that makes it hard for people to understand what macro to use. +// Use TORCH_CHECK(false, ...) or TORCH_INTERNAL_ASSERT(false, ...) to +// unconditionally fail at a line of code. +#define AT_ERROR(...) \ + do { \ + ::c10::detail::deprecated_AT_ERROR(); \ + C10_EXPAND_MSVC_WORKAROUND(TORCH_CHECK(false, ::c10::str(__VA_ARGS__))); \ + } while (false) + +#endif // C10_UTIL_EXCEPTION_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/ExclusivelyOwned.h b/backends/pytorch/include/c10/util/ExclusivelyOwned.h new file mode 100644 index 000000000..24cdba8d3 --- /dev/null +++ b/backends/pytorch/include/c10/util/ExclusivelyOwned.h @@ -0,0 +1,145 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include + +namespace c10 { + +// See example implementation in TensorBase.h and TensorBody.h. +// Synopsis: +// +// repr_type -- type to use to store an owned T in ExclusivelyOwned. +// +// pointer_type -- pointer-esque type to return from +// ExclusivelyOwned's get() and operator*() methods. +// +// const_pointer_type -- similar to pointer_type, used for the const methods. +// +// static repr_type nullRepr() -- return a null instance of repr_type. +// +// template +// static repr_type createInPlace(Args&&... args) -- used by the in-place +// ExclusivelyOwned constructor. +// +// static repr_type moveToRepr(T&& x) -- move the given x into an +// instance of repr_type. used by the ExclusivelyOwned(T&&) +// constructor. +// +// static void destroyOwned(repr_type x) -- free memory for a +// known-exclusively-owned instance of x. Replaces calling repr_type's +// destructor. Being able to implement this more efficiently than +// repr_type's destructor is the main reason to use ExclusivelyOwned +// for a type. +// +// static T take(repr_type&) -- move out of the given repr_type into an owned T. +// +// static pointer_type getImpl(const repr_type&) -- return a pointer +// to the given repr_type. May take repr_type by value if that is more +// efficient. +template +struct ExclusivelyOwnedTraits; + +/// ExclusivelyOwned is a smart-pointer-like wrapper around an +/// exclusively-owned instance of some type T that normally has +/// mandatory reference counting (currently just Tensor). If you have +/// an isolated piece of code that knows that it has sole ownership of +/// an object of one of these types (i.e., because you created it +/// directly or using a factory function) and that object will not +/// escape from that isolated piece of code, then moving the object +/// into an ExclusivelyOwned will avoid an atomic reference count +/// decrement at destruction time. +/// +/// If you directly create the Tensor in the first +/// place, you can use the in_place constructor of ExclusivelyOwned to +/// additionally avoid doing any stores to initialize the refcount & +/// weakcount. +template +class ExclusivelyOwned { + using EOT = ExclusivelyOwnedTraits; + typename ExclusivelyOwnedTraits::repr_type repr_; + + public: + ExclusivelyOwned() : repr_(EOT::nullRepr()) {} + + explicit ExclusivelyOwned(T&& t) : repr_(EOT::moveToRepr(std::move(t))) {} + + template + explicit ExclusivelyOwned(std::in_place_t /*unused*/, Args&&... args) + : repr_(EOT::createInPlace(std::forward(args)...)) {} + + ExclusivelyOwned(const ExclusivelyOwned&) = delete; + + ExclusivelyOwned(ExclusivelyOwned&& rhs) noexcept + : repr_(std::move(rhs.repr_)) { + rhs.repr_ = EOT::nullRepr(); + } + + ExclusivelyOwned& operator=(const ExclusivelyOwned&) = delete; + + ExclusivelyOwned& operator=(ExclusivelyOwned&& rhs) noexcept { + EOT::destroyOwned(repr_); + repr_ = std::move(rhs.repr_); + rhs.repr_ = EOT::nullRepr(); + return *this; + } + + ExclusivelyOwned& operator=(T&& rhs) noexcept { + EOT::destroyOwned(repr_); + repr_ = EOT::moveToRepr(std::move(rhs)); + return *this; + } + + ~ExclusivelyOwned() { + EOT::destroyOwned(repr_); + // Don't bother to call the destructor of repr_, since we already + // did specialized destruction for the exclusively-owned case in + // destroyOwned! + } + + // We don't provide this because it would require us to be able to + // differentiate an owned-but-empty T from a lack of T. This is + // particularly problematic for Tensor, which wants to use an + // undefined Tensor as its null state. + explicit operator bool() const noexcept = delete; + + operator T() && { + return take(); + } + + // NOTE: the equivalent operation on MaybeOwned is a moving + // operator*. For ExclusivelyOwned, take() and operator*() may well + // have different return types, so they are different functions. + T take() && { + return EOT::take(repr_); + } + + typename EOT::const_pointer_type operator->() const { + return get(); + } + + typename EOT::const_pointer_type get() const { + return EOT::getImpl(repr_); + } + + typename EOT::pointer_type operator->() { + return get(); + } + + typename EOT::pointer_type get() { + return EOT::getImpl(repr_); + } + + std::remove_pointer_t& operator*() const { + return *get(); + } + + std::remove_pointer_t& operator*() { + return *get(); + } +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/ExclusivelyOwnedTensorTraits.h b/backends/pytorch/include/c10/util/ExclusivelyOwnedTensorTraits.h new file mode 100644 index 000000000..5b3a76fe9 --- /dev/null +++ b/backends/pytorch/include/c10/util/ExclusivelyOwnedTensorTraits.h @@ -0,0 +1,80 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include + +namespace c10 { +// Shared ExclusivelyOwnedTraits implementation between caffe2::Tensor and +// at::TensorBase. +template +struct ExclusivelyOwnedTensorTraits { + using repr_type = TensorType; + using pointer_type = TensorType*; + using const_pointer_type = const TensorType*; + + static repr_type nullRepr() { + return TensorType(); + } + + template + static repr_type createInPlace(Args&&... args) { + return TensorType(std::forward(args)...); + } + + static repr_type moveToRepr(TensorType&& x) { + return std::move(x); + } + + static void destroyOwned(TensorType& x) { + TensorImpl* const toDestroy = x.unsafeReleaseTensorImpl(); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + toDestroy != nullptr, "Tensor somehow got null TensorImpl?"); + // May be 0 because UndefinedTensorImpl doesn't get its refcount + // incremented. + const bool isUndefined = toDestroy == UndefinedTensorImpl::singleton(); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + toDestroy->refcount() == 1 || + (toDestroy->refcount() == 0 && isUndefined), + "ExclusivelyOwned destroyed with isUndefined ", + isUndefined, + " and refcount ", + toDestroy->refcount(), + ", expected 1 or, if isUndefined, 0!"); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + toDestroy->weakcount() == 1 || + (toDestroy->weakcount() == 0 && + toDestroy == UndefinedTensorImpl::singleton()), + "ExclusivelyOwned destroyed with isUndefined ", + isUndefined, + " and weakcount ", + toDestroy->weakcount(), + ", expected 1 or, if isUndefined, 0!"); + if (!isUndefined) { +#ifndef NDEBUG + // Needed to pass the debug assertions in ~intrusive_ptr_target. + toDestroy->combined_refcount_.store(0, std::memory_order_relaxed); +#endif + delete toDestroy; + } + } + + static TensorType take(TensorType& x) { + return std::move(x); + } + + static pointer_type getImpl(repr_type& x) { + return &x; + } + + static const_pointer_type getImpl(const repr_type& x) { + return &x; + } +}; +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Flags.h b/backends/pytorch/include/c10/util/Flags.h new file mode 100644 index 000000000..c2485bfde --- /dev/null +++ b/backends/pytorch/include/c10/util/Flags.h @@ -0,0 +1,247 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#ifndef C10_UTIL_FLAGS_H_ +#define C10_UTIL_FLAGS_H_ + +/* Commandline flags support for C10. + * + * This is a portable commandline flags tool for c10, so we can optionally + * choose to use gflags or a lightweight custom implementation if gflags is + * not possible on a certain platform. If you have gflags installed, set the + * macro C10_USE_GFLAGS will seamlessly route everything to gflags. + * + * To define a flag foo of type bool default to true, do the following in the + * *global* namespace: + * C10_DEFINE_bool(foo, true, "An example."); + * + * To use it in another .cc file, you can use C10_DECLARE_* as follows: + * C10_DECLARE_bool(foo); + * + * In both cases, you can then access the flag via FLAGS_foo. + * + * It is recommended that you build with gflags. To learn more about the flags + * usage, refer to the gflags page here: + * + * https://gflags.github.io/gflags/ + * + * Note about Python users / devs: gflags is initiated from a C++ function + * ParseCommandLineFlags, and is usually done in native binaries in the main + * function. As Python does not have a modifiable main function, it is usually + * difficult to change the flags after Python starts. Hence, it is recommended + * that one sets the default value of the flags to one that's acceptable in + * general - that will allow Python to run without wrong flags. + */ + +#include +#include + +#include + +namespace c10 { +/** + * Sets the usage message when a commandline tool is called with "--help". + */ +C10_API void SetUsageMessage(const std::string& str); + +/** + * Returns the usage message for the commandline tool set by SetUsageMessage. + */ +C10_API const char* UsageMessage(); + +/** + * Parses the commandline flags. + * + * This command parses all the commandline arguments passed in via pargc + * and argv. Once it is finished, partc and argv will contain the remaining + * commandline args that c10 does not deal with. Note that following + * convention, argv[0] contains the binary name and is not parsed. + */ +C10_API bool ParseCommandLineFlags(int* pargc, char*** pargv); + +/** + * Checks if the commandline flags has already been passed. + */ +C10_API bool CommandLineFlagsHasBeenParsed(); + +} // namespace c10 + +//////////////////////////////////////////////////////////////////////////////// +// Below are gflags and non-gflags specific implementations. +// In general, they define the following macros for one to declare (use +// C10_DECLARE) or define (use C10_DEFINE) flags: +// C10_{DECLARE,DEFINE}_{int,int64,double,bool,string} +//////////////////////////////////////////////////////////////////////////////// + +#ifdef C10_USE_GFLAGS + +//////////////////////////////////////////////////////////////////////////////// +// Begin gflags section: most functions are basically rerouted to gflags. +//////////////////////////////////////////////////////////////////////////////// +#include + +// C10 uses hidden visibility by default. However, in gflags, it only uses +// export on Windows platform (with dllexport) but not on linux/mac (with +// default visibility). As a result, to ensure that we are always exporting +// global variables, we will redefine the GFLAGS_DLL_DEFINE_FLAG macro if we +// are building C10 as a shared library. +// This has to be done after the inclusion of gflags, because some early +// versions of gflags.h (e.g. 2.0 on ubuntu 14.04) directly defines the +// macros, so we need to do definition after gflags is done. +#ifdef GFLAGS_DLL_DEFINE_FLAG +#undef GFLAGS_DLL_DEFINE_FLAG +#endif // GFLAGS_DLL_DEFINE_FLAG +#ifdef GFLAGS_DLL_DECLARE_FLAG +#undef GFLAGS_DLL_DECLARE_FLAG +#endif // GFLAGS_DLL_DECLARE_FLAG +#define GFLAGS_DLL_DEFINE_FLAG C10_EXPORT +#define GFLAGS_DLL_DECLARE_FLAG C10_IMPORT + +// gflags before 2.0 uses namespace google and after 2.1 uses namespace gflags. +// Using GFLAGS_GFLAGS_H_ to capture this change. +#ifndef GFLAGS_GFLAGS_H_ +namespace gflags = google; +#endif // GFLAGS_GFLAGS_H_ + +// Motivation about the gflags wrapper: +// (1) We would need to make sure that the gflags version and the non-gflags +// version of C10 are going to expose the same flags abstraction. One should +// explicitly use FLAGS_flag_name to access the flags. +// (2) For flag names, it is recommended to start with c10_ to distinguish it +// from regular gflags flags. For example, do +// C10_DEFINE_BOOL(c10_my_flag, true, "An example"); +// to allow one to use FLAGS_c10_my_flag. +// (3) Gflags has a design issue that does not properly expose the global flags, +// if one builds the library with -fvisibility=hidden. The current gflags (as of +// Aug 2018) only deals with the Windows case using dllexport, and not the Linux +// counterparts. As a result, we will explicitly use C10_EXPORT to export the +// flags defined in C10. This is done via a global reference, so the flag +// itself is not duplicated - under the hood it is the same global gflags flag. +#define C10_GFLAGS_DEF_WRAPPER(type, real_type, name, default_value, help_str) \ + DEFINE_##type(name, default_value, help_str); + +#define C10_DEFINE_int(name, default_value, help_str) \ + C10_GFLAGS_DEF_WRAPPER(int32, gflags::int32, name, default_value, help_str) +#define C10_DEFINE_int32(name, default_value, help_str) \ + C10_DEFINE_int(name, default_value, help_str) +#define C10_DEFINE_int64(name, default_value, help_str) \ + C10_GFLAGS_DEF_WRAPPER(int64, gflags::int64, name, default_value, help_str) +#define C10_DEFINE_double(name, default_value, help_str) \ + C10_GFLAGS_DEF_WRAPPER(double, double, name, default_value, help_str) +#define C10_DEFINE_bool(name, default_value, help_str) \ + C10_GFLAGS_DEF_WRAPPER(bool, bool, name, default_value, help_str) +#define C10_DEFINE_string(name, default_value, help_str) \ + C10_GFLAGS_DEF_WRAPPER(string, ::fLS::clstring, name, default_value, help_str) + +// DECLARE_typed_var should be used in header files and in the global namespace. +#define C10_GFLAGS_DECLARE_WRAPPER(type, real_type, name) DECLARE_##type(name); + +#define C10_DECLARE_int(name) \ + C10_GFLAGS_DECLARE_WRAPPER(int32, gflags::int32, name) +#define C10_DECLARE_int32(name) C10_DECLARE_int(name) +#define C10_DECLARE_int64(name) \ + C10_GFLAGS_DECLARE_WRAPPER(int64, gflags::int64, name) +#define C10_DECLARE_double(name) \ + C10_GFLAGS_DECLARE_WRAPPER(double, double, name) +#define C10_DECLARE_bool(name) C10_GFLAGS_DECLARE_WRAPPER(bool, bool, name) +#define C10_DECLARE_string(name) \ + C10_GFLAGS_DECLARE_WRAPPER(string, ::fLS::clstring, name) + +#define TORCH_DECLARE_int(name) C10_DECLARE_int(name) +#define TORCH_DECLARE_int32(name) C10_DECLARE_int32(name) +#define TORCH_DECLARE_int64(name) C10_DECLARE_int64(name) +#define TORCH_DECLARE_double(name) C10_DECLARE_double(name) +#define TORCH_DECLARE_bool(name) C10_DECLARE_bool(name) +#define TORCH_DECLARE_string(name) C10_DECLARE_string(name) + +//////////////////////////////////////////////////////////////////////////////// +// End gflags section. +//////////////////////////////////////////////////////////////////////////////// + +#else // C10_USE_GFLAGS + +//////////////////////////////////////////////////////////////////////////////// +// Begin non-gflags section: providing equivalent functionality. +//////////////////////////////////////////////////////////////////////////////// + +namespace c10 { + +class C10_API C10FlagParser { + public: + bool success() { + return success_; + } + + protected: + template + bool Parse(const std::string& content, T* value); + bool success_{false}; +}; + +C10_DECLARE_REGISTRY(C10FlagsRegistry, C10FlagParser, const std::string&); + +} // namespace c10 + +// The macros are defined outside the c10 namespace. In your code, you should +// write the C10_DEFINE_* and C10_DECLARE_* macros outside any namespace +// as well. + +#define C10_DEFINE_typed_var(type, name, default_value, help_str) \ + C10_EXPORT type FLAGS_##name = default_value; \ + namespace c10 { \ + namespace { \ + class C10FlagParser_##name : public C10FlagParser { \ + public: \ + explicit C10FlagParser_##name(const std::string& content) { \ + success_ = C10FlagParser::Parse(content, &FLAGS_##name); \ + } \ + }; \ + RegistererC10FlagsRegistry g_C10FlagsRegistry_##name( \ + #name, \ + C10FlagsRegistry(), \ + RegistererC10FlagsRegistry::DefaultCreator, \ + "(" #type ", default " #default_value ") " help_str); \ + } \ + } + +#define C10_DEFINE_int(name, default_value, help_str) \ + C10_DEFINE_typed_var(int, name, default_value, help_str) +#define C10_DEFINE_int32(name, default_value, help_str) \ + C10_DEFINE_int(name, default_value, help_str) +#define C10_DEFINE_int64(name, default_value, help_str) \ + C10_DEFINE_typed_var(int64_t, name, default_value, help_str) +#define C10_DEFINE_double(name, default_value, help_str) \ + C10_DEFINE_typed_var(double, name, default_value, help_str) +#define C10_DEFINE_bool(name, default_value, help_str) \ + C10_DEFINE_typed_var(bool, name, default_value, help_str) +#define C10_DEFINE_string(name, default_value, help_str) \ + C10_DEFINE_typed_var(std::string, name, default_value, help_str) + +// DECLARE_typed_var should be used in header files and in the global namespace. +#define C10_DECLARE_typed_var(type, name) C10_API extern type FLAGS_##name + +#define C10_DECLARE_int(name) C10_DECLARE_typed_var(int, name) +#define C10_DECLARE_int32(name) C10_DECLARE_int(name) +#define C10_DECLARE_int64(name) C10_DECLARE_typed_var(int64_t, name) +#define C10_DECLARE_double(name) C10_DECLARE_typed_var(double, name) +#define C10_DECLARE_bool(name) C10_DECLARE_typed_var(bool, name) +#define C10_DECLARE_string(name) C10_DECLARE_typed_var(std::string, name) + +#define TORCH_DECLARE_typed_var(type, name) TORCH_API extern type FLAGS_##name + +#define TORCH_DECLARE_int(name) TORCH_DECLARE_typed_var(int, name) +#define TORCH_DECLARE_int32(name) TORCH_DECLARE_int(name) +#define TORCH_DECLARE_int64(name) TORCH_DECLARE_typed_var(int64_t, name) +#define TORCH_DECLARE_double(name) TORCH_DECLARE_typed_var(double, name) +#define TORCH_DECLARE_bool(name) TORCH_DECLARE_typed_var(bool, name) +#define TORCH_DECLARE_string(name) TORCH_DECLARE_typed_var(std::string, name) + +//////////////////////////////////////////////////////////////////////////////// +// End non-gflags section. +//////////////////////////////////////////////////////////////////////////////// + +#endif // C10_USE_GFLAGS + +#endif // C10_UTIL_FLAGS_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Float4_e2m1fn_x2.h b/backends/pytorch/include/c10/util/Float4_e2m1fn_x2.h new file mode 100644 index 000000000..fd690e5aa --- /dev/null +++ b/backends/pytorch/include/c10/util/Float4_e2m1fn_x2.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Float8_e4m3fn.h b/backends/pytorch/include/c10/util/Float8_e4m3fn.h new file mode 100644 index 000000000..ed07b9551 --- /dev/null +++ b/backends/pytorch/include/c10/util/Float8_e4m3fn.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Float8_e4m3fnuz.h b/backends/pytorch/include/c10/util/Float8_e4m3fnuz.h new file mode 100644 index 000000000..30481a624 --- /dev/null +++ b/backends/pytorch/include/c10/util/Float8_e4m3fnuz.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Float8_e5m2.h b/backends/pytorch/include/c10/util/Float8_e5m2.h new file mode 100644 index 000000000..f4e0802e2 --- /dev/null +++ b/backends/pytorch/include/c10/util/Float8_e5m2.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Float8_e5m2fnuz.h b/backends/pytorch/include/c10/util/Float8_e5m2fnuz.h new file mode 100644 index 000000000..f3e8c2509 --- /dev/null +++ b/backends/pytorch/include/c10/util/Float8_e5m2fnuz.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Float8_e8m0fnu.h b/backends/pytorch/include/c10/util/Float8_e8m0fnu.h new file mode 100644 index 000000000..030b23d64 --- /dev/null +++ b/backends/pytorch/include/c10/util/Float8_e8m0fnu.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/FunctionRef.h b/backends/pytorch/include/c10/util/FunctionRef.h new file mode 100644 index 000000000..342824b5b --- /dev/null +++ b/backends/pytorch/include/c10/util/FunctionRef.h @@ -0,0 +1,80 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +//===- llvm/ADT/STLExtras.h - Useful STL related functions ------*- C++ -*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===----------------------------------------------------------------------===// +// +// This file contains some templates that are useful if you are working with the +// STL at all. +// +// No library is required when using these functions. +// +//===----------------------------------------------------------------------===// + +// c10: modified from llvm::function_ref +// c10: added more SFINAE to enable use in overloaded functions + +#pragma once + +#include +#include +#include + +namespace c10 { + +/// An efficient, type-erasing, non-owning reference to a callable. This is +/// intended for use as the type of a function parameter that is not used +/// after the function in question returns. +/// +/// This class does not own the callable, so it is not in general safe to store +/// a function_ref. +template +class function_ref; + +template +class function_ref { + Ret (*callback)(intptr_t callable, Params... params) = nullptr; + intptr_t callable{}; + + template + static Ret callback_fn(intptr_t callable, Params... params) { + return (*reinterpret_cast(callable))( + std::forward(params)...); + } + + public: + function_ref() = default; + function_ref(std::nullptr_t) {} + + template + function_ref( + // NOLINTNEXTLINE(cppcoreguidelines-missing-std-forward) + Callable&& callable, + std::enable_if_t, + function_ref>>* /*unused*/ + = nullptr, + std::enable_if_t, + Ret>>* /*unused*/ + = nullptr) + : callback(callback_fn>), + callable(reinterpret_cast(&callable)) {} + + Ret operator()(Params... params) const { + return callback(callable, std::forward(params)...); + } + + operator bool() const { + return callback; + } +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Half.h b/backends/pytorch/include/c10/util/Half.h new file mode 100644 index 000000000..0a3d44626 --- /dev/null +++ b/backends/pytorch/include/c10/util/Half.h @@ -0,0 +1,13 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +// need to keep the following for BC because the APIs in here were exposed +// before migrating Half to torch/headeronly +#if (defined(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_AVX512)) && \ + !defined(__APPLE__) +#include +#endif + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/IdWrapper.h b/backends/pytorch/include/c10/util/IdWrapper.h new file mode 100644 index 000000000..b985cd3e5 --- /dev/null +++ b/backends/pytorch/include/c10/util/IdWrapper.h @@ -0,0 +1,82 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +namespace c10 { + +/** + * This template simplifies generation of simple classes that wrap an id + * in a typesafe way. Namely, you can use it to create a very lightweight + * type that only offers equality comparators and hashing. Example: + * + * struct MyIdType final : IdWrapper { + * constexpr explicit MyIdType(uint32_t id): IdWrapper(id) {} + * }; + * + * Then in the global top level namespace: + * + * C10_DEFINE_HASH_FOR_IDWRAPPER(MyIdType); + * + * That's it - equality operators and hash functions are automatically defined + * for you, given the underlying type supports it. + */ +template +class IdWrapper { + public: + using underlying_type = UnderlyingType; + using concrete_type = ConcreteType; + + protected: + constexpr explicit IdWrapper(underlying_type id) noexcept( + noexcept(underlying_type(std::declval()))) + : id_(id) {} + + constexpr underlying_type underlyingId() const + noexcept(noexcept(underlying_type(std::declval()))) { + return id_; + } + + private: + friend size_t hash_value(const concrete_type& v) { + return std::hash()(v.id_); + } + + // TODO Making operator== noexcept if underlying type is noexcept equality + // comparable doesn't work with GCC 4.8. + // Fix this once we don't need GCC 4.8 anymore. + friend constexpr bool operator==( + const concrete_type& lhs, + const concrete_type& rhs) noexcept { + return lhs.id_ == rhs.id_; + } + + // TODO Making operator!= noexcept if operator== is noexcept doesn't work with + // GCC 4.8. + // Fix this once we don't need GCC 4.8 anymore. + friend constexpr bool operator!=( + const concrete_type& lhs, + const concrete_type& rhs) noexcept { + return !(lhs == rhs); + } + + underlying_type id_; +}; + +} // namespace c10 + +#define C10_DEFINE_HASH_FOR_IDWRAPPER(ClassName) \ + namespace std { \ + template <> \ + struct hash { \ + size_t operator()(ClassName x) const { \ + return hash_value(x); \ + } \ + }; \ + } + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Lazy.h b/backends/pytorch/include/c10/util/Lazy.h new file mode 100644 index 000000000..204fc205e --- /dev/null +++ b/backends/pytorch/include/c10/util/Lazy.h @@ -0,0 +1,125 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +namespace c10 { + +/** + * Thread-safe lazy value with opportunistic concurrency: on concurrent first + * access, the factory may be called by multiple threads, but only one result is + * stored and its reference returned to all the callers. + * + * Value is heap-allocated; this optimizes for the case in which the value is + * never actually computed. + */ +template +class OptimisticLazy { + public: + OptimisticLazy() = default; + OptimisticLazy(const OptimisticLazy& other) { + if (T* value = other.value_.load(std::memory_order_acquire)) { + value_ = new T(*value); + } + } + OptimisticLazy(OptimisticLazy&& other) noexcept + : value_(other.value_.exchange(nullptr, std::memory_order_acq_rel)) {} + ~OptimisticLazy() { + reset(); + } + + template + T& ensure(const Factory& factory) { + if (T* value = value_.load(std::memory_order_acquire)) { + return *value; + } + T* value = new T(factory()); + T* old = nullptr; + if (!value_.compare_exchange_strong( + old, value, std::memory_order_release, std::memory_order_acquire)) { + delete value; + value = old; + } + return *value; + } + + // The following methods are not thread-safe: they should not be called + // concurrently with any other method. + + OptimisticLazy& operator=(const OptimisticLazy& other) { + *this = OptimisticLazy{other}; + return *this; + } + + OptimisticLazy& operator=(OptimisticLazy&& other) noexcept { + if (this != &other) { + reset(); + value_.store( + other.value_.exchange(nullptr, std::memory_order_acquire), + std::memory_order_release); + } + return *this; + } + + void reset() { + if (T* old = value_.load(std::memory_order_relaxed)) { + value_.store(nullptr, std::memory_order_relaxed); + delete old; + } + } + + private: + std::atomic value_{nullptr}; +}; + +/** + * Interface for a value that is computed on first access. + */ +template +class LazyValue { + public: + virtual ~LazyValue() = default; + + virtual const T& get() const = 0; +}; + +/** + * Convenience thread-safe LazyValue implementation with opportunistic + * concurrency. + */ +template +class OptimisticLazyValue : public LazyValue { + public: + const T& get() const override { + return value_.ensure([this] { return compute(); }); + } + + private: + virtual T compute() const = 0; + + mutable OptimisticLazy value_; +}; + +/** + * Convenience immutable (thus thread-safe) LazyValue implementation for cases + * in which the value is not actually lazy. + */ +template +class PrecomputedLazyValue : public LazyValue { + public: + PrecomputedLazyValue(T value) : value_(std::move(value)) {} + + const T& get() const override { + return value_; + } + + private: + T value_; +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Logging.h b/backends/pytorch/include/c10/util/Logging.h new file mode 100644 index 000000000..49420110e --- /dev/null +++ b/backends/pytorch/include/c10/util/Logging.h @@ -0,0 +1,378 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#ifndef C10_UTIL_LOGGING_H_ +#define C10_UTIL_LOGGING_H_ + +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include + +// CAFFE2_LOG_THRESHOLD is a compile time flag that would allow us to turn off +// logging at compile time so no logging message below that level is produced +// at all. The value should be between INT_MIN and CAFFE_FATAL. +#ifndef CAFFE2_LOG_THRESHOLD +// If we have not defined the compile time log threshold, we keep all the +// log cases. +#define CAFFE2_LOG_THRESHOLD INT_MIN +#endif // CAFFE2_LOG_THRESHOLD + +// Below are different implementations for glog and non-glog cases. +#ifdef C10_USE_GLOG +#include +#else // !C10_USE_GLOG +#include +#endif // C10_USE_GLOG + +C10_DECLARE_int(caffe2_log_level); +C10_DECLARE_bool(caffe2_use_fatal_for_enforce); + +// Some versions of GLOG support less-spammy version of LOG_EVERY_MS. If it's +// not available - just short-circuit to the always working one one. +// We define the C10_ name to avoid confusing other files +#ifdef LOG_EVERY_MS +#define C10_LOG_EVERY_MS(severity, ms) LOG_EVERY_MS(severity, ms) +#else +#define C10_LOG_EVERY_MS(severity, ms) LOG(severity) +#endif + +// Same for LOG_FIRST_N +#ifdef LOG_FIRST_N +#define C10_LOG_FIRST_N(severity, n) LOG_FIRST_N(severity, n) +#else +#define C10_LOG_FIRST_N(severity, n) LOG(severity) +#endif + +// Same for LOG_EVERY_N +#ifdef LOG_EVERY_N +#define C10_LOG_EVERY_N(severity, n) LOG_EVERY_N(severity, n) +#else +#define C10_LOG_EVERY_N(severity, n) LOG(severity) +#endif + +namespace c10 { + +#if !defined(C10_NODEPRECATED) +using std::string; +#endif + +// Functions that we use for initialization. +C10_API bool InitCaffeLogging(int* argc, char** argv); +C10_API void UpdateLoggingLevelsFromFlags(); + +[[noreturn]] C10_API void ThrowEnforceNotMet( + const char* file, + const int line, + const char* condition, + const std::string& msg, + const void* caller = nullptr); + +[[noreturn]] C10_API void ThrowEnforceNotMet( + const char* file, + const int line, + const char* condition, + const char* msg, + const void* caller = nullptr); + +[[noreturn]] inline void ThrowEnforceNotMet( + const char* file, + const int line, + const char* condition, + detail::CompileTimeEmptyString /*msg*/, + const void* caller = nullptr) { + ThrowEnforceNotMet(file, line, condition, "", caller); +} + +[[noreturn]] C10_API void ThrowEnforceFiniteNotMet( + const char* file, + const int line, + const char* condition, + const std::string& msg, + const void* caller = nullptr); + +[[noreturn]] C10_API void ThrowEnforceFiniteNotMet( + const char* file, + const int line, + const char* condition, + const char* msg, + const void* caller = nullptr); + +[[noreturn]] inline void ThrowEnforceFiniteNotMet( + const char* file, + const int line, + const char* condition, + detail::CompileTimeEmptyString /*msg*/, + const void* caller = nullptr) { + ThrowEnforceFiniteNotMet(file, line, condition, "", caller); +} + +constexpr bool IsUsingGoogleLogging() { +#ifdef C10_USE_GLOG + return true; +#else + return false; +#endif +} + +/** + * A utility to allow one to show log info to stderr after the program starts. + * + * This is similar to calling GLOG's --logtostderr, or setting caffe2_log_level + * to smaller than INFO. You are recommended to only use this in a few sparse + * cases, such as when you want to write a tutorial or something. Normally, use + * the commandline flags to set the log level. + */ +C10_API void ShowLogInfoToStderr(); + +C10_API void SetStackTraceFetcher(std::function<::c10::Backtrace()> fetcher); + +/** + * Convenience function for non-lazy stack trace fetchers. The Backtrace + * overload should be preferred when stringifying the backtrace is expensive. + */ +C10_API void SetStackTraceFetcher(std::function fetcher); + +using EnforceNotMet = ::c10::Error; + +#define CAFFE_ENFORCE(condition, ...) \ + do { \ + if (C10_UNLIKELY(!(condition))) { \ + ::c10::ThrowEnforceNotMet( \ + __FILE__, __LINE__, #condition, ::c10::str(__VA_ARGS__)); \ + } \ + } while (false) + +#define CAFFE_ENFORCE_FINITE(condition, ...) \ + do { \ + if (C10_UNLIKELY(!(condition))) { \ + ::c10::ThrowEnforceFiniteNotMet( \ + __FILE__, __LINE__, #condition, ::c10::str(__VA_ARGS__)); \ + } \ + } while (false) + +#define CAFFE_ENFORCE_WITH_CALLER(condition, ...) \ + do { \ + if (C10_UNLIKELY(!(condition))) { \ + ::c10::ThrowEnforceNotMet( \ + __FILE__, __LINE__, #condition, ::c10::str(__VA_ARGS__), this); \ + } \ + } while (false) + +#define CAFFE_THROW(...) \ + ::c10::ThrowEnforceNotMet(__FILE__, __LINE__, "", ::c10::str(__VA_ARGS__)) + +/** + * Rich logging messages + * + * CAFFE_ENFORCE_THAT can be used with one of the "checker functions" that + * capture input argument values and add it to the exception message. E.g. + * `CAFFE_ENFORCE_THAT(Equals(foo(x), bar(y)), "Optional additional message")` + * would evaluate both foo and bar only once and if the results are not equal - + * include them in the exception message. + * + * Some of the basic checker functions like Equals or Greater are already + * defined below. Other header might define customized checkers by adding + * functions to caffe2::enforce_detail namespace. For example: + * + * namespace caffe2 { namespace enforce_detail { + * inline EnforceFailMessage IsVector(const vector& shape) { + * if (shape.size() == 1) { return EnforceOK(); } + * return c10::str("Shape ", shape, " is not a vector"); + * } + * }} + * + * With further usages like `CAFFE_ENFORCE_THAT(IsVector(Input(0).dims()))` + * + * Convenient wrappers for binary operations like CAFFE_ENFORCE_EQ are provided + * too. Please use them instead of TORCH_CHECK_EQ and friends for failures in + * user-provided input. + */ + +namespace enforce_detail { + +template +std::string enforceFailMsgImpl(const T1& x, const T2& y) { + return c10::str(x, " vs ", y); +} + +template +std::string enforceFailMsgImpl(const T1& x, const T2& y, const Args&... args) { + return c10::str(x, " vs ", y, ". ", args...); +} + +template +void enforceThatImpl( + Pred p, + const T1& lhs, + const T2& rhs, + const char* file, + int line, + const char* expr, + const void* caller, + GetFailMsgFunc getFailMsg) { + if (C10_UNLIKELY(!(p(lhs, rhs)))) { + ::c10::ThrowEnforceNotMet(file, line, expr, getFailMsg(lhs, rhs), caller); + } +} + +#define CAFFE_ENFORCE_THAT_IMPL(op, lhs, rhs, expr, ...) \ + ::c10::enforce_detail::enforceThatImpl( \ + op, \ + (lhs), \ + (rhs), \ + __FILE__, \ + __LINE__, \ + expr, \ + nullptr, \ + [&](const auto& arg1, const auto& arg2) { \ + return ::c10::enforce_detail::enforceFailMsgImpl( \ + arg1, arg2, ##__VA_ARGS__); \ + }) + +#define CAFFE_ENFORCE_THAT_IMPL_WITH_CALLER(op, lhs, rhs, expr, ...) \ + ::c10::enforce_detail::enforceThatImpl( \ + op, \ + (lhs), \ + (rhs), \ + __FILE__, \ + __LINE__, \ + expr, \ + this, \ + [&](const auto& arg1, const auto& arg2) { \ + return ::c10::enforce_detail::enforceFailMsgImpl( \ + arg1, arg2, ##__VA_ARGS__); \ + }) + +} // namespace enforce_detail + +#define CAFFE_ENFORCE_THAT(cmp, op, lhs, rhs, ...) \ + CAFFE_ENFORCE_THAT_IMPL(cmp, lhs, rhs, #lhs " " #op " " #rhs, ##__VA_ARGS__) + +#define CAFFE_ENFORCE_BINARY_OP(cmp, op, x, y, ...) \ + CAFFE_ENFORCE_THAT_IMPL(cmp, x, y, #x " " #op " " #y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_EQ(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP(std::equal_to(), ==, x, y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_NE(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP(std::not_equal_to(), !=, x, y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_LE(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP(std::less_equal(), <=, x, y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_LT(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP(std::less(), <, x, y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_GE(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP(std::greater_equal(), >=, x, y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_GT(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP(std::greater(), >, x, y, ##__VA_ARGS__) + +#define CAFFE_ENFORCE_BINARY_OP_WITH_CALLER(cmp, op, x, y, ...) \ + CAFFE_ENFORCE_THAT_IMPL_WITH_CALLER( \ + cmp, x, y, #x " " #op " " #y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_EQ_WITH_CALLER(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP_WITH_CALLER( \ + std::equal_to(), ==, x, y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_NE_WITH_CALLER(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP_WITH_CALLER( \ + std::not_equal_to(), !=, x, y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_LE_WITH_CALLER(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP_WITH_CALLER( \ + std::less_equal(), <=, x, y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_LT_WITH_CALLER(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP_WITH_CALLER(std::less(), <, x, y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_GE_WITH_CALLER(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP_WITH_CALLER( \ + std::greater_equal(), >=, x, y, ##__VA_ARGS__) +#define CAFFE_ENFORCE_GT_WITH_CALLER(x, y, ...) \ + CAFFE_ENFORCE_BINARY_OP_WITH_CALLER( \ + std::greater(), >, x, y, ##__VA_ARGS__) + +struct IValue; +class C10_API EventSampledHandler { + public: + virtual void log( + std::string_view model_id, + const std::vector& args) = 0; + virtual ~EventSampledHandler() = default; +}; + +#define C10_LOG_EVENT_SAMPLED(event, ...) \ + static const std::unique_ptr<::c10::EventSampledHandler>& \ + _##event##EventSampledHandler = ::c10::GetEventSampledHandler(#event); \ + if (_##event##EventSampledHandler) { \ + _##event##EventSampledHandler->log(__VA_ARGS__); \ + } + +// Must be called in the main thread before any other threads are spawned. +C10_API void InitEventSampledHandlers( + std::vector>> /*handlers*/); +C10_API const std::unique_ptr& GetEventSampledHandler( + std::string_view /*event*/); + +/** + * Very lightweight logging for the first time API usage. It's beneficial for + * tracking of individual functionality usage in larger applications. + * + * In order to ensure light-weightedness of logging, we utilize static variable + * trick - LogAPIUsage will be invoked only once and further invocations will + * just do an atomic check. + * + * Example: + * // Logs caller info with an arbitrary text event, if there is a usage. + * C10_LOG_API_USAGE_ONCE("my_api"); + */ +#define C10_LOG_API_USAGE_ONCE(...) \ + [[maybe_unused]] static bool C10_ANONYMOUS_VARIABLE(logFlag) = \ + ::c10::detail::LogAPIUsageFakeReturn(__VA_ARGS__); + +// API usage logging capabilities +C10_API void SetAPIUsageLogger(std::function logger); +C10_API void LogAPIUsage(const std::string& context); + +C10_API void SetAPIUsageMetadataLogger( + std::function& metadata_map)> logger); +C10_API void LogAPIUsageMetadata( + const std::string& context, + const std::map& metadata_map); + +// PyTorch ddp usage logging capabilities +// DDPLoggingData holds data that can be logged in applications +// for analysis and debugging. Data structure is defined in +// c10 directory so that it can be easily imported by both c10 +// and torch files. +struct DDPLoggingData { + // logging fields that are string types. + std::map strs_map; + // logging fields that are int64_t types. + std::map ints_map; +}; + +C10_API void SetPyTorchDDPUsageLogger( + std::function logger); +C10_API void LogPyTorchDDPUsage(const DDPLoggingData& ddpData); + +namespace detail { +// Return value is needed to do the static variable initialization trick +C10_API bool LogAPIUsageFakeReturn(const std::string& context); +} // namespace detail + +// Initializes the c10 logger. +C10_API void initLogging(); + +// Sets the rank, which will be included in log messages +C10_API void SetGlobalRank(int64_t rank); + +} // namespace c10 + +#endif // C10_UTIL_LOGGING_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/MaybeOwned.h b/backends/pytorch/include/c10/util/MaybeOwned.h new file mode 100644 index 000000000..61e6ed82f --- /dev/null +++ b/backends/pytorch/include/c10/util/MaybeOwned.h @@ -0,0 +1,242 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include + +#include +#include +#include + +namespace c10 { + +/// MaybeOwnedTraits describes how to borrow from T. Here is how we +/// can implement borrowing from an arbitrary type T using a raw +/// pointer to const: +template +struct MaybeOwnedTraitsGenericImpl { + using owned_type = T; + using borrow_type = const T*; + + static borrow_type createBorrow(const owned_type& from) { + return &from; + } + + static void assignBorrow(borrow_type& lhs, borrow_type rhs) { + lhs = rhs; + } + + static void destroyBorrow(borrow_type& /*toDestroy*/) {} + + static const owned_type& referenceFromBorrow(const borrow_type& borrow) { + return *borrow; + } + + static const owned_type* pointerFromBorrow(const borrow_type& borrow) { + return borrow; + } + + static bool debugBorrowIsValid(const borrow_type& borrow) { + return borrow != nullptr; + } +}; + +/// It is possible to eliminate the extra layer of indirection for +/// borrows for some types that we control. For examples, see +/// intrusive_ptr.h and TensorBody.h. + +template +struct MaybeOwnedTraits; + +// Explicitly enable MaybeOwned>, rather than allowing +// MaybeOwned to be used for any type right away. +template +struct MaybeOwnedTraits> + : public MaybeOwnedTraitsGenericImpl> {}; + +/// A smart pointer around either a borrowed or owned T. When +/// constructed with borrowed(), the caller MUST ensure that the +/// borrowed-from argument outlives this MaybeOwned. Compare to +/// Rust's std::borrow::Cow +/// (https://doc.rust-lang.org/std/borrow/enum.Cow.html), but note +/// that it is probably not suitable for general use because C++ has +/// no borrow checking. Included here to support +/// Tensor::expect_contiguous. +template +class MaybeOwned final { + using borrow_type = typename MaybeOwnedTraits::borrow_type; + using owned_type = typename MaybeOwnedTraits::owned_type; + + bool isBorrowed_; + union { + borrow_type borrow_; + owned_type own_; + }; + + /// Don't use this; use borrowed() instead. + explicit MaybeOwned(const owned_type& t) + : isBorrowed_(true), borrow_(MaybeOwnedTraits::createBorrow(t)) {} + + /// Don't use this; use owned() instead. + explicit MaybeOwned(T&& t) noexcept(std::is_nothrow_move_constructible_v) + : isBorrowed_(false), own_(std::move(t)) {} + + /// Don't use this; use owned() instead. + template + explicit MaybeOwned(std::in_place_t /*unused*/, Args&&... args) + : isBorrowed_(false), own_(std::forward(args)...) {} + + public: + explicit MaybeOwned() : isBorrowed_(true), borrow_() {} + + // Copying a borrow yields another borrow of the original, as with a + // T*. Copying an owned T yields another owned T for safety: no + // chains of borrowing by default! (Note you could get that behavior + // with MaybeOwned::borrowed(*rhs) if you wanted it.) + MaybeOwned(const MaybeOwned& rhs) : isBorrowed_(rhs.isBorrowed_) { + if (C10_LIKELY(rhs.isBorrowed_)) { + MaybeOwnedTraits::assignBorrow(borrow_, rhs.borrow_); + } else { + new (&own_) T(rhs.own_); + } + } + + MaybeOwned& operator=(const MaybeOwned& rhs) { + if (this == &rhs) { + return *this; + } + if (C10_UNLIKELY(!isBorrowed_)) { + if (rhs.isBorrowed_) { + own_.~T(); + MaybeOwnedTraits::assignBorrow(borrow_, rhs.borrow_); + isBorrowed_ = true; + } else { + own_ = rhs.own_; + } + } else { + if (C10_LIKELY(rhs.isBorrowed_)) { + MaybeOwnedTraits::assignBorrow(borrow_, rhs.borrow_); + } else { + MaybeOwnedTraits::destroyBorrow(borrow_); + new (&own_) T(rhs.own_); + isBorrowed_ = false; + } + } + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(isBorrowed_ == rhs.isBorrowed_); + return *this; + } + + MaybeOwned(MaybeOwned&& rhs) noexcept( + // NOLINTNEXTLINE(*-noexcept-move-*) + std::is_nothrow_move_constructible_v && + std::is_nothrow_move_assignable_v) + : isBorrowed_(rhs.isBorrowed_) { + if (C10_LIKELY(rhs.isBorrowed_)) { + MaybeOwnedTraits::assignBorrow(borrow_, rhs.borrow_); + } else { + new (&own_) T(std::move(rhs.own_)); + } + } + + MaybeOwned& operator=(MaybeOwned&& rhs) noexcept( + std::is_nothrow_move_assignable_v && + std::is_nothrow_move_assignable_v && + std::is_nothrow_move_constructible_v && + // NOLINTNEXTLINE(*-noexcept-move-*) + std::is_nothrow_destructible_v && + std::is_nothrow_destructible_v) { + if (this == &rhs) { + return *this; + } + if (C10_UNLIKELY(!isBorrowed_)) { + if (rhs.isBorrowed_) { + own_.~T(); + MaybeOwnedTraits::assignBorrow(borrow_, rhs.borrow_); + isBorrowed_ = true; + } else { + own_ = std::move(rhs.own_); + } + } else { + if (C10_LIKELY(rhs.isBorrowed_)) { + MaybeOwnedTraits::assignBorrow(borrow_, rhs.borrow_); + } else { + MaybeOwnedTraits::destroyBorrow(borrow_); + new (&own_) T(std::move(rhs.own_)); + isBorrowed_ = false; + } + } + return *this; + } + + static MaybeOwned borrowed(const T& t) { + return MaybeOwned(t); + } + + static MaybeOwned owned(T&& t) noexcept( + std::is_nothrow_move_constructible_v) { + return MaybeOwned(std::move(t)); + } + + template + static MaybeOwned owned(std::in_place_t /*unused*/, Args&&... args) { + return MaybeOwned(std::in_place, std::forward(args)...); + } + + ~MaybeOwned() noexcept( + // NOLINTNEXTLINE(*-noexcept-destructor) + std::is_nothrow_destructible_v && + std::is_nothrow_destructible_v) { + if (C10_UNLIKELY(!isBorrowed_)) { + own_.~T(); + } else { + MaybeOwnedTraits::destroyBorrow(borrow_); + } + } + + // This is an implementation detail! You should know what you're doing + // if you are testing this. If you just want to guarantee ownership move + // this into a T + bool unsafeIsBorrowed() const { + return isBorrowed_; + } + + const T& operator*() const& { + if (isBorrowed_) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + MaybeOwnedTraits::debugBorrowIsValid(borrow_)); + } + return C10_LIKELY(isBorrowed_) + ? MaybeOwnedTraits::referenceFromBorrow(borrow_) + : own_; + } + + const T* operator->() const { + if (isBorrowed_) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + MaybeOwnedTraits::debugBorrowIsValid(borrow_)); + } + return C10_LIKELY(isBorrowed_) + ? MaybeOwnedTraits::pointerFromBorrow(borrow_) + : &own_; + } + + // If borrowed, copy the underlying T. If owned, move from + // it. borrowed/owned state remains the same, and either we + // reference the same borrow as before or we are an owned moved-from + // T. + T operator*() && { + if (isBorrowed_) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + MaybeOwnedTraits::debugBorrowIsValid(borrow_)); + return MaybeOwnedTraits::referenceFromBorrow(borrow_); + } else { + return std::move(own_); + } + } +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Metaprogramming.h b/backends/pytorch/include/c10/util/Metaprogramming.h new file mode 100644 index 000000000..55c3fb2ba --- /dev/null +++ b/backends/pytorch/include/c10/util/Metaprogramming.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Optional.h b/backends/pytorch/include/c10/util/Optional.h new file mode 100644 index 000000000..55c469736 --- /dev/null +++ b/backends/pytorch/include/c10/util/Optional.h @@ -0,0 +1,65 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#ifndef C10_UTIL_OPTIONAL_H_ +#define C10_UTIL_OPTIONAL_H_ + +#include +#include + +// Macros.h is not needed, but it does namespace shenanigans that lots +// of downstream code seems to rely on. Feel free to remove it and fix +// up builds. + +namespace c10 { + +#if !defined(FBCODE_CAFFE2) && !defined(C10_NODEPRECATED) +// NOLINTNEXTLINE(misc-unused-using-decls) +using std::bad_optional_access; +// NOLINTNEXTLINE(misc-unused-using-decls) +using std::make_optional; +// NOLINTNEXTLINE(misc-unused-using-decls) +using std::nullopt; +// NOLINTNEXTLINE(misc-unused-using-decls) +using std::nullopt_t; +// NOLINTNEXTLINE(misc-unused-using-decls) +using std::optional; +#endif + +#if !defined(FBCODE_CAFFE2) && !defined(C10_NODEPRECATED) + +namespace detail_ { +// the call to convert(b) has return type A and converts b to type A iff b +// decltype(b) is implicitly convertible to A +template +constexpr U convert(U v) { + return v; +} +} // namespace detail_ +template +[[deprecated( + "Please use std::optional::value_or instead of c10::value_or_else")]] constexpr T +value_or_else(const std::optional& v, F&& func) { + static_assert( + std::is_convertible_v, T>, + "func parameters must be a callable that returns a type convertible to the value stored in the optional"); + return v.has_value() ? *v : detail_::convert(std::forward(func)()); +} + +template +[[deprecated( + "Please use std::optional::value_or instead of c10::value_or_else")]] constexpr T +value_or_else(std::optional&& v, F&& func) { + static_assert( + std::is_convertible_v, T>, + "func parameters must be a callable that returns a type convertible to the value stored in the optional"); + return v.has_value() ? constexpr_move(std::move(v).contained_val()) + : detail_::convert(std::forward(func)()); +} + +#endif + +} // namespace c10 +#endif // C10_UTIL_OPTIONAL_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/OptionalArrayRef.h b/backends/pytorch/include/c10/util/OptionalArrayRef.h new file mode 100644 index 000000000..cd15a5f19 --- /dev/null +++ b/backends/pytorch/include/c10/util/OptionalArrayRef.h @@ -0,0 +1,242 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +// This file defines OptionalArrayRef, a class that has almost the same +// exact functionality as std::optional>, except that its +// converting constructor fixes a dangling pointer issue. +// +// The implicit converting constructor of both std::optional> and +// std::optional> can cause the underlying ArrayRef to store +// a dangling pointer. OptionalArrayRef prevents this by wrapping +// a std::optional> and fixing the constructor implementation. +// +// See https://github.com/pytorch/pytorch/issues/63645 for more on this. + +#pragma once + +#include +#include +#include +#include +#include +#include + +namespace c10 { + +template +class OptionalArrayRef final { + public: + // Constructors + + constexpr OptionalArrayRef() noexcept = default; + + constexpr OptionalArrayRef(std::nullopt_t /*unused*/) noexcept {} + + OptionalArrayRef(const OptionalArrayRef& other) = default; + + OptionalArrayRef(OptionalArrayRef&& other) noexcept = default; + + constexpr OptionalArrayRef(const std::optional>& other) noexcept + : wrapped_opt_array_ref(other) {} + + constexpr OptionalArrayRef(std::optional>&& other) noexcept + : wrapped_opt_array_ref(std::move(other)) {} + + constexpr OptionalArrayRef(const T& value) noexcept + : wrapped_opt_array_ref(value) {} + + template < + typename U = ArrayRef, + std::enable_if_t< + !std::is_same_v, OptionalArrayRef> && + !std::is_same_v, std::in_place_t> && + std::is_constructible_v, U&&> && + std::is_convertible_v> && + !std::is_convertible_v, + bool> = false> + constexpr OptionalArrayRef(U&& value) noexcept( + std::is_nothrow_constructible_v, U&&>) + : wrapped_opt_array_ref(std::forward(value)) {} + + template < + typename U = ArrayRef, + std::enable_if_t< + !std::is_same_v, OptionalArrayRef> && + !std::is_same_v, std::in_place_t> && + std::is_constructible_v, U&&> && + !std::is_convertible_v>, + bool> = false> + constexpr explicit OptionalArrayRef(U&& value) noexcept( + std::is_nothrow_constructible_v, U&&>) + : wrapped_opt_array_ref(std::forward(value)) {} + + template + constexpr explicit OptionalArrayRef( + std::in_place_t ip, + Args&&... args) noexcept + : wrapped_opt_array_ref(ip, std::forward(args)...) {} + + template + constexpr explicit OptionalArrayRef( + std::in_place_t ip, + std::initializer_list il, + Args&&... args) + : wrapped_opt_array_ref(ip, il, std::forward(args)...) {} + + constexpr OptionalArrayRef(const std::initializer_list& Vec) + : wrapped_opt_array_ref(ArrayRef(Vec)) {} + + // Destructor + + ~OptionalArrayRef() = default; + + // Assignment + + constexpr OptionalArrayRef& operator=(std::nullopt_t /*unused*/) noexcept { + wrapped_opt_array_ref = std::nullopt; + return *this; + } + + OptionalArrayRef& operator=(const OptionalArrayRef& other) = default; + + OptionalArrayRef& operator=(OptionalArrayRef&& other) noexcept = default; + + constexpr OptionalArrayRef& operator=( + const std::optional>& other) noexcept { + wrapped_opt_array_ref = other; + return *this; + } + + constexpr OptionalArrayRef& operator=( + std::optional>&& other) noexcept { + wrapped_opt_array_ref = std::move(other); + return *this; + } + + template < + typename U = ArrayRef, + typename = std::enable_if_t< + !std::is_same_v, OptionalArrayRef> && + std::is_constructible_v, U&&> && + std::is_assignable_v&, U&&>>> + constexpr OptionalArrayRef& operator=(U&& value) noexcept( + std::is_nothrow_constructible_v, U&&> && + std::is_nothrow_assignable_v&, U&&>) { + wrapped_opt_array_ref = std::forward(value); + return *this; + } + + // Observers + + constexpr ArrayRef* operator->() noexcept { + return &wrapped_opt_array_ref.value(); + } + + constexpr const ArrayRef* operator->() const noexcept { + return &wrapped_opt_array_ref.value(); + } + + constexpr ArrayRef& operator*() & noexcept { + return wrapped_opt_array_ref.value(); + } + + constexpr const ArrayRef& operator*() const& noexcept { + return wrapped_opt_array_ref.value(); + } + + constexpr ArrayRef&& operator*() && noexcept { + return std::move(wrapped_opt_array_ref.value()); + } + + constexpr const ArrayRef&& operator*() const&& noexcept { + return std::move(wrapped_opt_array_ref.value()); + } + + constexpr explicit operator bool() const noexcept { + return wrapped_opt_array_ref.has_value(); + } + + constexpr bool has_value() const noexcept { + return wrapped_opt_array_ref.has_value(); + } + + constexpr ArrayRef& value() & { + return wrapped_opt_array_ref.value(); + } + + constexpr const ArrayRef& value() const& { + // NOLINTNEXTLINE(bugprone-unchecked-optional-access) + return wrapped_opt_array_ref.value(); + } + + constexpr ArrayRef&& value() && { + return std::move(wrapped_opt_array_ref.value()); + } + + constexpr const ArrayRef&& value() const&& { + return std::move(wrapped_opt_array_ref.value()); + } + + template + constexpr std:: + enable_if_t>, ArrayRef> + value_or(U&& default_value) const& { + return wrapped_opt_array_ref.value_or(std::forward(default_value)); + } + + template + constexpr std:: + enable_if_t>, ArrayRef> + value_or(U&& default_value) && { + return wrapped_opt_array_ref.value_or(std::forward(default_value)); + } + + // Modifiers + + constexpr void swap(OptionalArrayRef& other) noexcept { + std::swap(wrapped_opt_array_ref, other.wrapped_opt_array_ref); + } + + constexpr void reset() noexcept { + wrapped_opt_array_ref.reset(); + } + + template + constexpr std:: + enable_if_t, Args&&...>, ArrayRef&> + emplace(Args&&... args) noexcept( + std::is_nothrow_constructible_v, Args&&...>) { + return wrapped_opt_array_ref.emplace(std::forward(args)...); + } + + template + constexpr ArrayRef& emplace( + std::initializer_list il, + Args&&... args) noexcept { + return wrapped_opt_array_ref.emplace(il, std::forward(args)...); + } + + private: + std::optional> wrapped_opt_array_ref; +}; + +using OptionalIntArrayRef = OptionalArrayRef; + +inline bool operator==( + const OptionalIntArrayRef& a1, + const IntArrayRef& other) { + if (!a1.has_value()) { + return false; + } + return a1.value() == other; +} + +inline bool operator==( + const c10::IntArrayRef& a1, + const c10::OptionalIntArrayRef& a2) { + return a2 == a1; +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Registry.h b/backends/pytorch/include/c10/util/Registry.h new file mode 100644 index 000000000..92d1809d8 --- /dev/null +++ b/backends/pytorch/include/c10/util/Registry.h @@ -0,0 +1,334 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#ifndef C10_UTIL_REGISTRY_H_ +#define C10_UTIL_REGISTRY_H_ + +/** + * Simple registry implementation that uses static variables to + * register object creators during program initialization time. + */ + +// NB: This Registry works poorly when you have other namespaces. +// Make all macro invocations from inside the at namespace. + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include + +namespace c10 { + +template +inline std::string KeyStrRepr(const KeyType& /*key*/) { + return "[key type printing not supported]"; +} + +template <> +inline std::string KeyStrRepr(const std::string& key) { + return key; +} + +enum RegistryPriority { + REGISTRY_FALLBACK = 1, + REGISTRY_DEFAULT = 2, + REGISTRY_PREFERRED = 3, +}; + +/** + * @brief A template class that allows one to register classes by keys. + * + * The keys are usually a std::string specifying the name, but can be anything + * that can be used in a std::map. + * + * You should most likely not use the Registry class explicitly, but use the + * helper macros below to declare specific registries as well as registering + * objects. + */ +template +class Registry { + public: + typedef std::function Creator; + + Registry(bool warning = true) : registry_(), priority_(), warning_(warning) {} + ~Registry() = default; + + void Register( + const SrcType& key, + Creator creator, + const RegistryPriority priority = REGISTRY_DEFAULT) { + std::lock_guard lock(register_mutex_); + // The if statement below is essentially the same as the following line: + // TORCH_CHECK_EQ(registry_.count(key), 0) << "Key " << key + // << " registered twice."; + // However, TORCH_CHECK_EQ depends on google logging, and since registration + // is carried out at static initialization time, we do not want to have an + // explicit dependency on glog's initialization function. + if (registry_.count(key) != 0) { + auto cur_priority = priority_[key]; + if (priority > cur_priority) { +#ifdef DEBUG + std::string warn_msg = + "Overwriting already registered item for key " + KeyStrRepr(key); + fprintf(stderr, "%s\n", warn_msg.c_str()); +#endif + registry_[key] = creator; + priority_[key] = priority; + } else if (priority == cur_priority) { + std::string err_msg = + "Key already registered with the same priority: " + KeyStrRepr(key); + fprintf(stderr, "%s\n", err_msg.c_str()); + if (terminate_) { + std::exit(1); + } else { + throw std::runtime_error(err_msg); + } + } else if (warning_) { + std::string warn_msg = + "Higher priority item already registered, skipping registration of " + + KeyStrRepr(key); + fprintf(stderr, "%s\n", warn_msg.c_str()); + } + } else { + registry_[key] = creator; + priority_[key] = priority; + } + } + + void Register( + const SrcType& key, + Creator creator, + const std::string& help_msg, + const RegistryPriority priority = REGISTRY_DEFAULT) { + Register(key, creator, priority); + help_message_[key] = help_msg; + } + + inline bool Has(const SrcType& key) { + return (registry_.count(key) != 0); + } + + ObjectPtrType Create(const SrcType& key, Args... args) { + auto it = registry_.find(key); + if (it == registry_.end()) { + // Returns nullptr if the key is not registered. + return nullptr; + } + return it->second(args...); + } + + /** + * Returns the keys currently registered as a std::vector. + */ + std::vector Keys() const { + std::vector keys; + keys.reserve(registry_.size()); + for (const auto& it : registry_) { + keys.push_back(it.first); + } + return keys; + } + + inline const std::unordered_map& HelpMessage() const { + return help_message_; + } + + const char* HelpMessage(const SrcType& key) const { + auto it = help_message_.find(key); + if (it == help_message_.end()) { + return nullptr; + } + return it->second.c_str(); + } + + // Used for testing, if terminate is unset, Registry throws instead of + // calling std::exit + void SetTerminate(bool terminate) { + terminate_ = terminate; + } + + C10_DISABLE_COPY_AND_ASSIGN(Registry); + Registry(Registry&&) = delete; + Registry& operator=(Registry&&) = delete; + + private: + std::unordered_map registry_; + std::unordered_map priority_; + bool terminate_{true}; + const bool warning_; + std::unordered_map help_message_; + std::mutex register_mutex_; +}; + +template +class Registerer { + public: + explicit Registerer( + const SrcType& key, + Registry* registry, + typename Registry::Creator creator, + const std::string& help_msg = "") { + registry->Register(key, creator, help_msg); + } + + explicit Registerer( + const SrcType& key, + const RegistryPriority priority, + Registry* registry, + typename Registry::Creator creator, + const std::string& help_msg = "") { + registry->Register(key, creator, help_msg, priority); + } + + template + static ObjectPtrType DefaultCreator(Args... args) { + return ObjectPtrType(new DerivedType(args...)); + } +}; + +/** + * C10_DECLARE_TYPED_REGISTRY is a macro that expands to a function + * declaration, as well as creating a convenient typename for its corresponding + * registerer. + */ +// Note on C10_IMPORT and C10_EXPORT below: we need to explicitly mark DECLARE +// as import and DEFINE as export, because these registry macros will be used +// in downstream shared libraries as well, and one cannot use *_API - the API +// macro will be defined on a per-shared-library basis. Semantically, when one +// declares a typed registry it is always going to be IMPORT, and when one +// defines a registry (which should happen ONLY ONCE and ONLY IN SOURCE FILE), +// the instantiation unit is always going to be exported. +// +// The only unique condition is when in the same file one does DECLARE and +// DEFINE - in Windows compilers, this generates a warning that dllimport and +// dllexport are mixed, but the warning is fine and linker will be properly +// exporting the symbol. Same thing happens in the gflags flag declaration and +// definition caes. +#define C10_DECLARE_TYPED_REGISTRY( \ + RegistryName, SrcType, ObjectType, PtrType, ...) \ + C10_API ::c10::Registry, ##__VA_ARGS__>* \ + RegistryName(); \ + typedef ::c10::Registerer, ##__VA_ARGS__> \ + Registerer##RegistryName + +#define TORCH_DECLARE_TYPED_REGISTRY( \ + RegistryName, SrcType, ObjectType, PtrType, ...) \ + TORCH_API ::c10::Registry, ##__VA_ARGS__>* \ + RegistryName(); \ + typedef ::c10::Registerer, ##__VA_ARGS__> \ + Registerer##RegistryName + +#define C10_DEFINE_TYPED_REGISTRY( \ + RegistryName, SrcType, ObjectType, PtrType, ...) \ + C10_EXPORT ::c10::Registry, ##__VA_ARGS__>* \ + RegistryName() { \ + static ::c10::Registry, ##__VA_ARGS__>* \ + registry = new ::c10:: \ + Registry, ##__VA_ARGS__>(); \ + return registry; \ + } + +#define C10_DEFINE_TYPED_REGISTRY_WITHOUT_WARNING( \ + RegistryName, SrcType, ObjectType, PtrType, ...) \ + C10_EXPORT ::c10::Registry, ##__VA_ARGS__>* \ + RegistryName() { \ + static ::c10::Registry, ##__VA_ARGS__>* \ + registry = \ + new ::c10::Registry, ##__VA_ARGS__>( \ + false); \ + return registry; \ + } + +// Note(Yangqing): The __VA_ARGS__ below allows one to specify a templated +// creator with comma in its templated arguments. +#define C10_REGISTER_TYPED_CREATOR(RegistryName, key, ...) \ + static Registerer##RegistryName C10_ANONYMOUS_VARIABLE(g_##RegistryName)( \ + key, RegistryName(), ##__VA_ARGS__); + +#define C10_REGISTER_TYPED_CREATOR_WITH_PRIORITY( \ + RegistryName, key, priority, ...) \ + static Registerer##RegistryName C10_ANONYMOUS_VARIABLE(g_##RegistryName)( \ + key, priority, RegistryName(), ##__VA_ARGS__); + +#define C10_REGISTER_TYPED_CLASS(RegistryName, key, ...) \ + static Registerer##RegistryName C10_ANONYMOUS_VARIABLE(g_##RegistryName)( \ + key, \ + RegistryName(), \ + Registerer##RegistryName::DefaultCreator<__VA_ARGS__>, \ + ::c10::demangle_type<__VA_ARGS__>()); + +#define C10_REGISTER_TYPED_CLASS_WITH_PRIORITY( \ + RegistryName, key, priority, ...) \ + static Registerer##RegistryName C10_ANONYMOUS_VARIABLE(g_##RegistryName)( \ + key, \ + priority, \ + RegistryName(), \ + Registerer##RegistryName::DefaultCreator<__VA_ARGS__>, \ + ::c10::demangle_type<__VA_ARGS__>()); + +// C10_DECLARE_REGISTRY and C10_DEFINE_REGISTRY are hard-wired to use +// std::string as the key type, because that is the most commonly used cases. +#define C10_DECLARE_REGISTRY(RegistryName, ObjectType, ...) \ + C10_DECLARE_TYPED_REGISTRY( \ + RegistryName, std::string, ObjectType, std::unique_ptr, ##__VA_ARGS__) + +#define TORCH_DECLARE_REGISTRY(RegistryName, ObjectType, ...) \ + TORCH_DECLARE_TYPED_REGISTRY( \ + RegistryName, std::string, ObjectType, std::unique_ptr, ##__VA_ARGS__) + +#define C10_DEFINE_REGISTRY(RegistryName, ObjectType, ...) \ + C10_DEFINE_TYPED_REGISTRY( \ + RegistryName, std::string, ObjectType, std::unique_ptr, ##__VA_ARGS__) + +#define C10_DEFINE_REGISTRY_WITHOUT_WARNING(RegistryName, ObjectType, ...) \ + C10_DEFINE_TYPED_REGISTRY_WITHOUT_WARNING( \ + RegistryName, std::string, ObjectType, std::unique_ptr, ##__VA_ARGS__) + +#define C10_DECLARE_SHARED_REGISTRY(RegistryName, ObjectType, ...) \ + C10_DECLARE_TYPED_REGISTRY( \ + RegistryName, std::string, ObjectType, std::shared_ptr, ##__VA_ARGS__) + +#define TORCH_DECLARE_SHARED_REGISTRY(RegistryName, ObjectType, ...) \ + TORCH_DECLARE_TYPED_REGISTRY( \ + RegistryName, std::string, ObjectType, std::shared_ptr, ##__VA_ARGS__) + +#define C10_DEFINE_SHARED_REGISTRY(RegistryName, ObjectType, ...) \ + C10_DEFINE_TYPED_REGISTRY( \ + RegistryName, std::string, ObjectType, std::shared_ptr, ##__VA_ARGS__) + +#define C10_DEFINE_SHARED_REGISTRY_WITHOUT_WARNING( \ + RegistryName, ObjectType, ...) \ + C10_DEFINE_TYPED_REGISTRY_WITHOUT_WARNING( \ + RegistryName, std::string, ObjectType, std::shared_ptr, ##__VA_ARGS__) + +// C10_REGISTER_CREATOR and C10_REGISTER_CLASS are hard-wired to use std::string +// as the key +// type, because that is the most commonly used cases. +#define C10_REGISTER_CREATOR(RegistryName, key, ...) \ + C10_REGISTER_TYPED_CREATOR(RegistryName, #key, __VA_ARGS__) + +#define C10_REGISTER_CREATOR_WITH_PRIORITY(RegistryName, key, priority, ...) \ + C10_REGISTER_TYPED_CREATOR_WITH_PRIORITY( \ + RegistryName, #key, priority, __VA_ARGS__) + +#define C10_REGISTER_CLASS(RegistryName, key, ...) \ + C10_REGISTER_TYPED_CLASS(RegistryName, #key, __VA_ARGS__) + +#define C10_REGISTER_CLASS_WITH_PRIORITY(RegistryName, key, priority, ...) \ + C10_REGISTER_TYPED_CLASS_WITH_PRIORITY( \ + RegistryName, #key, priority, __VA_ARGS__) + +} // namespace c10 + +#endif // C10_UTIL_REGISTRY_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/SmallVector.h b/backends/pytorch/include/c10/util/SmallVector.h new file mode 100644 index 000000000..b2a4dbb0f --- /dev/null +++ b/backends/pytorch/include/c10/util/SmallVector.h @@ -0,0 +1,1472 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +//===- llvm/ADT/SmallVector.h - 'Normally small' vectors --------*- C++ -*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===----------------------------------------------------------------------===// +// +// This file defines the SmallVector class. +// +//===----------------------------------------------------------------------===// + +// ATen: modified from llvm::SmallVector. +// used std::is_trivially_{copy,move}_constructible +// replaced iterator_range constructor with inline Container&& constructor +// replaced LLVM_NODISCARD, LLVM_LIKELY, and LLVM_UNLIKELY with c10 equivalents +// removed LLVM_GSL_OWNER +// added SmallVector::at +// added operator<< for std::ostream +// added C10_API to export SmallVectorBase + +#pragma once + +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +/// This is all the stuff common to all SmallVectors. +/// +/// The template parameter specifies the type which should be used to hold the +/// Size and Capacity of the SmallVector, so it can be adjusted. +/// Using 32 bit size is desirable to shrink the size of the SmallVector. +/// Using 64 bit size is desirable for cases like SmallVector, where a +/// 32 bit size would limit the vector to ~4GB. SmallVectors are used for +/// buffering bitcode output - which can exceed 4GB. +template +class C10_API SmallVectorBase { + protected: + void* BeginX; + Size_T Size = 0, Capacity; + + /// The maximum value of the Size_T used. + static constexpr size_t SizeTypeMax() { + return std::numeric_limits::max(); + } + + SmallVectorBase(void* FirstEl, size_t TotalCapacity) + : BeginX(FirstEl), Capacity(TotalCapacity) {} + + /// This is a helper for \a grow() that's out of line to reduce code + /// duplication. This function will report a fatal error if it can't grow at + /// least to \p MinSize. + void* mallocForGrow(size_t MinSize, size_t TSize, size_t& NewCapacity); + + /// This is an implementation of the grow() method which only works + /// on POD-like data types and is out of line to reduce code duplication. + /// This function will report a fatal error if it cannot increase capacity. + void grow_pod(const void* FirstEl, size_t MinSize, size_t TSize); + + public: + SmallVectorBase() = delete; + size_t size() const { + return Size; + } + size_t capacity() const { + return Capacity; + } + + [[nodiscard]] bool empty() const { + return !Size; + } + + /// Set the array size to \p N, which the current array must have enough + /// capacity for. + /// + /// This does not construct or destroy any elements in the vector. + /// + /// Clients can use this in conjunction with capacity() to write past the end + /// of the buffer when they know that more elements are available, and only + /// update the size later. This avoids the cost of value initializing elements + /// which will only be overwritten. + void set_size(size_t N) { + assert(N <= capacity()); + Size = N; + } +}; + +template +using SmallVectorSizeType = + std::conditional_t= 8, uint64_t, uint32_t>; + +/// Figure out the offset of the first element. +template +struct SmallVectorAlignmentAndSize { + // NOLINTNEXTLINE(*c-arrays*) + alignas(SmallVectorBase>) char Base[sizeof( + SmallVectorBase>)]; + // NOLINTNEXTLINE(*c-arrays*) + alignas(T) char FirstEl[sizeof(T)]; +}; + +/// This is the part of SmallVectorTemplateBase which does not depend on whether +/// the type T is a POD. The extra dummy template argument is used by ArrayRef +/// to avoid unnecessarily requiring T to be complete. +template +class SmallVectorTemplateCommon + : public SmallVectorBase> { + using Base = SmallVectorBase>; + + /// Find the address of the first element. For this pointer math to be valid + /// with small-size of 0 for T with lots of alignment, it's important that + /// SmallVectorStorage is properly-aligned even for small-size of 0. + void* getFirstEl() const { + return const_cast(reinterpret_cast( + reinterpret_cast(this) + + offsetof(SmallVectorAlignmentAndSize, FirstEl))); + } + // Space after 'FirstEl' is clobbered, do not add any instance vars after it. + + protected: + SmallVectorTemplateCommon(size_t Size) : Base(getFirstEl(), Size) {} + + void grow_pod(size_t MinSize, size_t TSize) { + Base::grow_pod(getFirstEl(), MinSize, TSize); + } + + /// Return true if this is a smallvector which has not had dynamic + /// memory allocated for it. + bool isSmall() const { + return this->BeginX == getFirstEl(); + } + + /// Put this vector in a state of being small. + void resetToSmall() { + this->BeginX = getFirstEl(); + this->Size = this->Capacity = 0; // FIXME: Setting Capacity to 0 is suspect. + } + + /// Return true if V is an internal reference to the given range. + bool isReferenceToRange(const void* V, const void* First, const void* Last) + const { + // Use std::less to avoid UB. + std::less<> LessThan; + return !LessThan(V, First) && LessThan(V, Last); + } + + /// Return true if V is an internal reference to this vector. + bool isReferenceToStorage(const void* V) const { + return isReferenceToRange(V, this->begin(), this->end()); + } + + /// Return true if First and Last form a valid (possibly empty) range in this + /// vector's storage. + bool isRangeInStorage(const void* First, const void* Last) const { + // Use std::less to avoid UB. + std::less<> LessThan; + return !LessThan(First, this->begin()) && !LessThan(Last, First) && + !LessThan(this->end(), Last); + } + + /// Return true unless Elt will be invalidated by resizing the vector to + /// NewSize. + bool isSafeToReferenceAfterResize(const void* Elt, size_t NewSize) { + // Past the end. + if (C10_LIKELY(!isReferenceToStorage(Elt))) + return true; + + // Return false if Elt will be destroyed by shrinking. + if (NewSize <= this->size()) + return Elt < this->begin() + NewSize; + + // Return false if we need to grow. + return NewSize <= this->capacity(); + } + + /// Check whether Elt will be invalidated by resizing the vector to NewSize. + void assertSafeToReferenceAfterResize(const void* Elt, size_t NewSize) { + (void)Elt; // Suppress unused variable warning + (void)NewSize; // Suppress unused variable warning + assert( + isSafeToReferenceAfterResize(Elt, NewSize) && + "Attempting to reference an element of the vector in an operation " + "that invalidates it"); + } + + /// Check whether Elt will be invalidated by increasing the size of the + /// vector by N. + void assertSafeToAdd(const void* Elt, size_t N = 1) { + this->assertSafeToReferenceAfterResize(Elt, this->size() + N); + } + + /// Check whether any part of the range will be invalidated by clearing. + void assertSafeToReferenceAfterClear(const T* From, const T* To) { + if (From == To) + return; + this->assertSafeToReferenceAfterResize(From, 0); + this->assertSafeToReferenceAfterResize(To - 1, 0); + } + template < + class ItTy, + std::enable_if_t, T*>, bool> = + false> + void assertSafeToReferenceAfterClear(ItTy /*unused*/, ItTy /*unused*/) {} + + /// Check whether any part of the range will be invalidated by growing. + void assertSafeToAddRange(const T* From, const T* To) { + if (From == To) + return; + this->assertSafeToAdd(From, To - From); + this->assertSafeToAdd(To - 1, To - From); + } + template < + class ItTy, + std::enable_if_t, T*>, bool> = + false> + void assertSafeToAddRange(ItTy /*unused*/, ItTy /*unused*/) {} + + /// Reserve enough space to add one element, and return the updated element + /// pointer in case it was a reference to the storage. + template + static const T* reserveForParamAndGetAddressImpl( + U* This, + const T& Elt, + size_t N) { + size_t NewSize = This->size() + N; + if (C10_LIKELY(NewSize <= This->capacity())) + return &Elt; + + bool ReferencesStorage = false; + int64_t Index = -1; + if constexpr (!U::TakesParamByValue) { + if (C10_UNLIKELY(This->isReferenceToStorage(&Elt))) { + ReferencesStorage = true; + Index = &Elt - This->begin(); + } + } + This->grow(NewSize); + return ReferencesStorage ? This->begin() + Index : &Elt; + } + + public: + using size_type = size_t; + using difference_type = ptrdiff_t; + using value_type = T; + using iterator = T*; + using const_iterator = const T*; + + using const_reverse_iterator = std::reverse_iterator; + using reverse_iterator = std::reverse_iterator; + + using reference = T&; + using const_reference = const T&; + using pointer = T*; + using const_pointer = const T*; + + using Base::capacity; + using Base::empty; + using Base::size; + + // forward iterator creation methods. + iterator begin() { + return (iterator)this->BeginX; + } + const_iterator begin() const { + return (const_iterator)this->BeginX; + } + iterator end() { + return begin() + size(); + } + const_iterator end() const { + return begin() + size(); + } + + // reverse iterator creation methods. + reverse_iterator rbegin() { + return reverse_iterator(end()); + } + const_reverse_iterator rbegin() const { + return const_reverse_iterator(end()); + } + reverse_iterator rend() { + return reverse_iterator(begin()); + } + const_reverse_iterator rend() const { + return const_reverse_iterator(begin()); + } + + size_type size_in_bytes() const { + return size() * sizeof(T); + } + constexpr size_type max_size() const { + return std::min(this->SizeTypeMax(), size_type(-1) / sizeof(T)); + } + + size_t capacity_in_bytes() const { + return capacity() * sizeof(T); + } + + /// Return a pointer to the vector's buffer, even if empty(). + pointer data() { + return pointer(begin()); + } + /// Return a pointer to the vector's buffer, even if empty(). + const_pointer data() const { + return const_pointer(begin()); + } + + // SmallVector::at is NOT from LLVM. + reference at(size_type idx) { + assert(idx < size()); + return begin()[idx]; + } + const_reference at(size_type idx) const { + assert(idx < size()); + return begin()[idx]; + } + reference operator[](size_type idx) { + assert(idx < size()); + return begin()[idx]; + } + const_reference operator[](size_type idx) const { + assert(idx < size()); + return begin()[idx]; + } + + reference front() { + assert(!empty()); + return begin()[0]; + } + const_reference front() const { + assert(!empty()); + return begin()[0]; + } + + reference back() { + assert(!empty()); + return end()[-1]; + } + const_reference back() const { + assert(!empty()); + return end()[-1]; + } +}; + +/// SmallVectorTemplateBase - This is where we put +/// method implementations that are designed to work with non-trivial T's. +/// +/// We approximate is_trivially_copyable with trivial move/copy construction and +/// trivial destruction. While the standard doesn't specify that you're allowed +/// copy these types with memcpy, there is no way for the type to observe this. +/// This catches the important case of std::pair, which is not +/// trivially assignable. +/// +/// XXX: if build fails here fall back to C10_IS_TRIVIALLY_COPYABLE and make a +/// note +template < + typename T, + bool = (std::is_trivially_copy_constructible_v) && + (std::is_trivially_move_constructible_v) && + std::is_trivially_destructible_v> +class SmallVectorTemplateBase : public SmallVectorTemplateCommon { + friend class SmallVectorTemplateCommon; + + protected: + static constexpr bool TakesParamByValue = false; + using ValueParamT = const T&; + + SmallVectorTemplateBase(size_t Size) : SmallVectorTemplateCommon(Size) {} + + static void destroy_range(T* S, T* E) { + while (S != E) { + --E; + E->~T(); + } + } + + /// Move the range [I, E) into the uninitialized memory starting with "Dest", + /// constructing elements as needed. + template + static void uninitialized_move(It1 I, It1 E, It2 Dest) { + std::uninitialized_copy( + std::make_move_iterator(I), std::make_move_iterator(E), Dest); + } + + /// Copy the range [I, E) onto the uninitialized memory starting with "Dest", + /// constructing elements as needed. + template + static void uninitialized_copy(It1 I, It1 E, It2 Dest) { + std::uninitialized_copy(I, E, Dest); + } + + /// Grow the allocated memory (without initializing new elements), doubling + /// the size of the allocated memory. Guarantees space for at least one more + /// element, or MinSize more elements if specified. + void grow(size_t MinSize = 0); + + /// Create a new allocation big enough for \p MinSize and pass back its size + /// in \p NewCapacity. This is the first section of \a grow(). + T* mallocForGrow(size_t MinSize, size_t& NewCapacity) { + return static_cast( + SmallVectorBase>::mallocForGrow( + MinSize, sizeof(T), NewCapacity)); + } + + /// Move existing elements over to the new allocation \p NewElts, the middle + /// section of \a grow(). + void moveElementsForGrow(T* NewElts); + + /// Transfer ownership of the allocation, finishing up \a grow(). + void takeAllocationForGrow(T* NewElts, size_t NewCapacity); + + /// Reserve enough space to add one element, and return the updated element + /// pointer in case it was a reference to the storage. + const T* reserveForParamAndGetAddress(const T& Elt, size_t N = 1) { + return this->reserveForParamAndGetAddressImpl(this, Elt, N); + } + + /// Reserve enough space to add one element, and return the updated element + /// pointer in case it was a reference to the storage. + T* reserveForParamAndGetAddress(T& Elt, size_t N = 1) { + return const_cast(this->reserveForParamAndGetAddressImpl(this, Elt, N)); + } + + static T&& forward_value_param(T&& V) { + return std::move(V); + } + static const T& forward_value_param(const T& V) { + return V; + } + + void growAndAssign(size_t NumElts, const T& Elt) { + // Grow manually in case Elt is an internal reference. + size_t NewCapacity = 0; + T* NewElts = mallocForGrow(NumElts, NewCapacity); + std::uninitialized_fill_n(NewElts, NumElts, Elt); + this->destroy_range(this->begin(), this->end()); + takeAllocationForGrow(NewElts, NewCapacity); + this->set_size(NumElts); + } + + template + T& growAndEmplaceBack(ArgTypes&&... Args) { + // Grow manually in case one of Args is an internal reference. + size_t NewCapacity = 0; + T* NewElts = mallocForGrow(0, NewCapacity); + ::new ((void*)(NewElts + this->size())) T(std::forward(Args)...); + moveElementsForGrow(NewElts); + takeAllocationForGrow(NewElts, NewCapacity); + this->set_size(this->size() + 1); + return this->back(); + } + + public: + void push_back(const T& Elt) { + const T* EltPtr = reserveForParamAndGetAddress(Elt); + ::new ((void*)this->end()) T(*EltPtr); + this->set_size(this->size() + 1); + } + + // NOLINTNEXTLINE(cppcoreguidelines-rvalue-reference-param-not-moved) + void push_back(T&& Elt) { + T* EltPtr = reserveForParamAndGetAddress(Elt); + ::new ((void*)this->end()) T(::std::move(*EltPtr)); + this->set_size(this->size() + 1); + } + + void pop_back() { + this->set_size(this->size() - 1); + this->end()->~T(); + } +}; + +// Define this out-of-line to dissuade the C++ compiler from inlining it. +template +void SmallVectorTemplateBase::grow(size_t MinSize) { + size_t NewCapacity = 0; + T* NewElts = mallocForGrow(MinSize, NewCapacity); + moveElementsForGrow(NewElts); + takeAllocationForGrow(NewElts, NewCapacity); +} + +// Define this out-of-line to dissuade the C++ compiler from inlining it. +template +void SmallVectorTemplateBase::moveElementsForGrow( + T* NewElts) { + // Move the elements over. + this->uninitialized_move(this->begin(), this->end(), NewElts); + + // Destroy the original elements. + destroy_range(this->begin(), this->end()); +} + +// Define this out-of-line to dissuade the C++ compiler from inlining it. +template +void SmallVectorTemplateBase::takeAllocationForGrow( + T* NewElts, + size_t NewCapacity) { + // If this wasn't grown from the inline copy, deallocate the old space. + if (!this->isSmall()) + free(this->begin()); + + this->BeginX = NewElts; + this->Capacity = NewCapacity; +} + +/// SmallVectorTemplateBase - This is where we put +/// method implementations that are designed to work with trivially copyable +/// T's. This allows using memcpy in place of copy/move construction and +/// skipping destruction. +template +class SmallVectorTemplateBase : public SmallVectorTemplateCommon { + friend class SmallVectorTemplateCommon; + + protected: + /// True if it's cheap enough to take parameters by value. Doing so avoids + /// overhead related to mitigations for reference invalidation. + static constexpr bool TakesParamByValue = sizeof(T) <= 2 * sizeof(void*); + + /// Either const T& or T, depending on whether it's cheap enough to take + /// parameters by value. + using ValueParamT = std::conditional_t; + + SmallVectorTemplateBase(size_t Size) : SmallVectorTemplateCommon(Size) {} + + // No need to do a destroy loop for POD's. + static void destroy_range(T* /*unused*/, T* /*unused*/) {} + + /// Move the range [I, E) onto the uninitialized memory + /// starting with "Dest", constructing elements into it as needed. + template + static void uninitialized_move(It1 I, It1 E, It2 Dest) { + // Just do a copy. + uninitialized_copy(I, E, Dest); + } + + /// Copy the range [I, E) onto the uninitialized memory + /// starting with "Dest", constructing elements into it as needed. + template + static void uninitialized_copy(It1 I, It1 E, It2 Dest) { + // Arbitrary iterator types; just use the basic implementation. + std::uninitialized_copy(I, E, Dest); + } + + /// Copy the range [I, E) onto the uninitialized memory + /// starting with "Dest", constructing elements into it as needed. + template + static void uninitialized_copy( + T1* I, + T1* E, + T2* Dest, + std::enable_if_t, T2>>* /*unused*/ + = nullptr) { + // Use memcpy for PODs iterated by pointers (which includes SmallVector + // iterators): std::uninitialized_copy optimizes to memmove, but we can + // use memcpy here. Note that I and E are iterators and thus might be + // invalid for memcpy if they are equal. + if (I != E) + memcpy(reinterpret_cast(Dest), I, (E - I) * sizeof(T)); + } + + /// Double the size of the allocated memory, guaranteeing space for at + /// least one more element or MinSize if specified. + void grow(size_t MinSize = 0) { + this->grow_pod(MinSize, sizeof(T)); + } + + /// Reserve enough space to add one element, and return the updated element + /// pointer in case it was a reference to the storage. + const T* reserveForParamAndGetAddress(const T& Elt, size_t N = 1) { + return this->reserveForParamAndGetAddressImpl(this, Elt, N); + } + + /// Reserve enough space to add one element, and return the updated element + /// pointer in case it was a reference to the storage. + T* reserveForParamAndGetAddress(T& Elt, size_t N = 1) { + return const_cast(this->reserveForParamAndGetAddressImpl(this, Elt, N)); + } + + /// Copy \p V or return a reference, depending on \a ValueParamT. + static ValueParamT forward_value_param(ValueParamT V) { + return V; + } + + void growAndAssign(size_t NumElts, T Elt) { + // Elt has been copied in case it's an internal reference, side-stepping + // reference invalidation problems without losing the realloc optimization. + this->set_size(0); + this->grow(NumElts); + std::uninitialized_fill_n(this->begin(), NumElts, Elt); + this->set_size(NumElts); + } + + template + T& growAndEmplaceBack(ArgTypes&&... Args) { + // Use push_back with a copy in case Args has an internal reference, + // side-stepping reference invalidation problems without losing the realloc + // optimization. + push_back(T(std::forward(Args)...)); + return this->back(); + } + + public: + void push_back(ValueParamT Elt) { + const T* EltPtr = reserveForParamAndGetAddress(Elt); + memcpy(reinterpret_cast(this->end()), EltPtr, sizeof(T)); + this->set_size(this->size() + 1); + } + + void pop_back() { + this->set_size(this->size() - 1); + } +}; + +/// This class consists of common code factored out of the SmallVector class to +/// reduce code duplication based on the SmallVector 'N' template parameter. +template +class SmallVectorImpl : public SmallVectorTemplateBase { + using SuperClass = SmallVectorTemplateBase; + + public: + using iterator = typename SuperClass::iterator; + using const_iterator = typename SuperClass::const_iterator; + using reference = typename SuperClass::reference; + using size_type = typename SuperClass::size_type; + + protected: + using SmallVectorTemplateBase::TakesParamByValue; + using ValueParamT = typename SuperClass::ValueParamT; + + // Default ctor - Initialize to empty. + explicit SmallVectorImpl(unsigned N) : SmallVectorTemplateBase(N) {} + + public: + SmallVectorImpl(const SmallVectorImpl&) = delete; + + ~SmallVectorImpl() { + // Subclass has already destructed this vector's elements. + // If this wasn't grown from the inline copy, deallocate the old space. + if (!this->isSmall()) + free(this->begin()); + } + + void clear() { + this->destroy_range(this->begin(), this->end()); + this->Size = 0; + } + + private: + template + void resizeImpl(size_type N) { + if (N < this->size()) { + this->pop_back_n(this->size() - N); + } else if (N > this->size()) { + this->reserve(N); + for (auto I = this->end(), E = this->begin() + N; I != E; ++I) + if (ForOverwrite) + new (&*I) T; + else + new (&*I) T(); + this->set_size(N); + } + } + + public: + void resize(size_type N) { + resizeImpl(N); + } + + /// Like resize, but \ref T is POD, the new values won't be initialized. + void resize_for_overwrite(size_type N) { + resizeImpl(N); + } + + void resize(size_type N, ValueParamT NV) { + if (N == this->size()) + return; + + if (N < this->size()) { + this->pop_back_n(this->size() - N); + return; + } + + // N > this->size(). Defer to append. + this->append(N - this->size(), NV); + } + + void reserve(size_type N) { + if (this->capacity() < N) + this->grow(N); + } + + void pop_back_n(size_type NumItems) { + assert(this->size() >= NumItems); + this->destroy_range(this->end() - NumItems, this->end()); + this->set_size(this->size() - NumItems); + } + + [[nodiscard]] T pop_back_val() { + T Result = ::std::move(this->back()); + this->pop_back(); + return Result; + } + + void swap(SmallVectorImpl& RHS) noexcept; + + /// Add the specified range to the end of the SmallVector. + template < + typename in_iter, + typename = std::enable_if_t::iterator_category, + std::input_iterator_tag>>> + void append(in_iter in_start, in_iter in_end) { + this->assertSafeToAddRange(in_start, in_end); + size_type NumInputs = std::distance(in_start, in_end); + this->reserve(this->size() + NumInputs); + this->uninitialized_copy(in_start, in_end, this->end()); + this->set_size(this->size() + NumInputs); + } + + /// Append \p NumInputs copies of \p Elt to the end. + void append(size_type NumInputs, ValueParamT Elt) { + const T* EltPtr = this->reserveForParamAndGetAddress(Elt, NumInputs); + std::uninitialized_fill_n(this->end(), NumInputs, *EltPtr); + this->set_size(this->size() + NumInputs); + } + + void append(std::initializer_list IL) { + append(IL.begin(), IL.end()); + } + + void append(const SmallVectorImpl& RHS) { + append(RHS.begin(), RHS.end()); + } + + void assign(size_type NumElts, ValueParamT Elt) { + // Note that Elt could be an internal reference. + if (NumElts > this->capacity()) { + this->growAndAssign(NumElts, Elt); + return; + } + + // Assign over existing elements. + std::fill_n(this->begin(), std::min(NumElts, this->size()), Elt); + if (NumElts > this->size()) + std::uninitialized_fill_n(this->end(), NumElts - this->size(), Elt); + else if (NumElts < this->size()) + this->destroy_range(this->begin() + NumElts, this->end()); + this->set_size(NumElts); + } + + // FIXME: Consider assigning over existing elements, rather than clearing & + // re-initializing them - for all assign(...) variants. + + template < + typename in_iter, + typename = std::enable_if_t::iterator_category, + std::input_iterator_tag>>> + void assign(in_iter in_start, in_iter in_end) { + this->assertSafeToReferenceAfterClear(in_start, in_end); + clear(); + append(in_start, in_end); + } + + void assign(std::initializer_list IL) { + clear(); + append(IL); + } + + void assign(const SmallVectorImpl& RHS) { + assign(RHS.begin(), RHS.end()); + } + + iterator erase(iterator I) { + assert( + this->isReferenceToStorage(I) && "Iterator to erase is out of bounds."); + + iterator N = I; + // Shift all elts down one. + std::move(I + 1, this->end(), I); + // Drop the last elt. + this->pop_back(); + return N; + } + + iterator erase(iterator S, iterator E) { + assert(this->isRangeInStorage(S, E) && "Range to erase is out of bounds."); + + iterator N = S; + // Shift all elts down. + iterator I = std::move(E, this->end(), S); + // Drop the last elts. + this->destroy_range(I, this->end()); + this->set_size(I - this->begin()); + return N; + } + + private: + template + iterator insert_one_impl(iterator I, ArgType&& Elt) { + // Callers ensure that ArgType is derived from T. + static_assert( + std::is_same>, T>:: + value, + "ArgType must be derived from T!"); + + if (I == this->end()) { // Important special case for empty vector. + this->push_back(::std::forward(Elt)); + return this->end() - 1; + } + + assert( + this->isReferenceToStorage(I) && + "Insertion iterator is out of bounds."); + + // Grow if necessary. + size_t Index = I - this->begin(); + std::remove_reference_t* EltPtr = + this->reserveForParamAndGetAddress(Elt); + I = this->begin() + Index; + + ::new ((void*)this->end()) T(::std::move(this->back())); + // Push everything else over. + std::move_backward(I, this->end() - 1, this->end()); + this->set_size(this->size() + 1); + + // If we just moved the element we're inserting, be sure to update + // the reference (never happens if TakesParamByValue). + static_assert( + !TakesParamByValue || std::is_same_v, + "ArgType must be 'T' when taking by value!"); + if (!TakesParamByValue && this->isReferenceToRange(EltPtr, I, this->end())) + ++EltPtr; + + *I = ::std::forward(*EltPtr); + return I; + } + + public: + iterator insert(iterator I, T&& Elt) { + return insert_one_impl(I, this->forward_value_param(std::move(Elt))); + } + + iterator insert(iterator I, const T& Elt) { + return insert_one_impl(I, this->forward_value_param(Elt)); + } + + iterator insert(iterator I, size_type NumToInsert, ValueParamT Elt) { + // Convert iterator to elt# to avoid invalidating iterator when we reserve() + size_t InsertElt = I - this->begin(); + + if (I == this->end()) { // Important special case for empty vector. + append(NumToInsert, Elt); + return this->begin() + InsertElt; + } + + assert( + this->isReferenceToStorage(I) && + "Insertion iterator is out of bounds."); + + // Ensure there is enough space, and get the (maybe updated) address of + // Elt. + const T* EltPtr = this->reserveForParamAndGetAddress(Elt, NumToInsert); + + // Uninvalidate the iterator. + I = this->begin() + InsertElt; + + // If there are more elements between the insertion point and the end of the + // range than there are being inserted, we can use a simple approach to + // insertion. Since we already reserved space, we know that this won't + // reallocate the vector. + if (size_t(this->end() - I) >= NumToInsert) { + T* OldEnd = this->end(); + append( + std::move_iterator(this->end() - NumToInsert), + std::move_iterator(this->end())); + + // Copy the existing elements that get replaced. + std::move_backward(I, OldEnd - NumToInsert, OldEnd); + + // If we just moved the element we're inserting, be sure to update + // the reference (never happens if TakesParamByValue). + if (!TakesParamByValue && I <= EltPtr && EltPtr < this->end()) + EltPtr += NumToInsert; + + std::fill_n(I, NumToInsert, *EltPtr); + return I; + } + + // Otherwise, we're inserting more elements than exist already, and we're + // not inserting at the end. + + // Move over the elements that we're about to overwrite. + T* OldEnd = this->end(); + this->set_size(this->size() + NumToInsert); + size_t NumOverwritten = OldEnd - I; + this->uninitialized_move(I, OldEnd, this->end() - NumOverwritten); + + // If we just moved the element we're inserting, be sure to update + // the reference (never happens if TakesParamByValue). + if (!TakesParamByValue && I <= EltPtr && EltPtr < this->end()) + EltPtr += NumToInsert; + + // Replace the overwritten part. + std::fill_n(I, NumOverwritten, *EltPtr); + + // Insert the non-overwritten middle part. + std::uninitialized_fill_n(OldEnd, NumToInsert - NumOverwritten, *EltPtr); + return I; + } + + template < + typename ItTy, + typename = std::enable_if_t::iterator_category, + std::input_iterator_tag>>> + iterator insert(iterator I, ItTy From, ItTy To) { + // Convert iterator to elt# to avoid invalidating iterator when we reserve() + size_t InsertElt = I - this->begin(); + + if (I == this->end()) { // Important special case for empty vector. + append(From, To); + return this->begin() + InsertElt; + } + + assert( + this->isReferenceToStorage(I) && + "Insertion iterator is out of bounds."); + + // Check that the reserve that follows doesn't invalidate the iterators. + this->assertSafeToAddRange(From, To); + + size_t NumToInsert = std::distance(From, To); + + // Ensure there is enough space. + reserve(this->size() + NumToInsert); + + // Uninvalidate the iterator. + I = this->begin() + InsertElt; + + // If there are more elements between the insertion point and the end of the + // range than there are being inserted, we can use a simple approach to + // insertion. Since we already reserved space, we know that this won't + // reallocate the vector. + if (size_t(this->end() - I) >= NumToInsert) { + T* OldEnd = this->end(); + append( + std::move_iterator(this->end() - NumToInsert), + std::move_iterator(this->end())); + + // Copy the existing elements that get replaced. + std::move_backward(I, OldEnd - NumToInsert, OldEnd); + + std::copy(From, To, I); + return I; + } + + // Otherwise, we're inserting more elements than exist already, and we're + // not inserting at the end. + + // Move over the elements that we're about to overwrite. + T* OldEnd = this->end(); + this->set_size(this->size() + NumToInsert); + size_t NumOverwritten = OldEnd - I; + this->uninitialized_move(I, OldEnd, this->end() - NumOverwritten); + + // Replace the overwritten part. + for (T* J = I; NumOverwritten > 0; --NumOverwritten) { + *J = *From; + ++J; + ++From; + } + + // Insert the non-overwritten middle part. + this->uninitialized_copy(From, To, OldEnd); + return I; + } + + void insert(iterator I, std::initializer_list IL) { + insert(I, IL.begin(), IL.end()); + } + + template + reference emplace_back(ArgTypes&&... Args) { + if (C10_UNLIKELY(this->size() >= this->capacity())) + return this->growAndEmplaceBack(std::forward(Args)...); + + ::new ((void*)this->end()) T(std::forward(Args)...); + this->set_size(this->size() + 1); + return this->back(); + } + + SmallVectorImpl& operator=(const SmallVectorImpl& RHS); + + SmallVectorImpl& operator=(SmallVectorImpl&& RHS) noexcept( + std::is_nothrow_move_constructible_v && + std::is_nothrow_destructible_v); + + bool operator==(const SmallVectorImpl& RHS) const { + if (this->size() != RHS.size()) + return false; + return std::equal(this->begin(), this->end(), RHS.begin()); + } + bool operator!=(const SmallVectorImpl& RHS) const { + return !(*this == RHS); + } + + bool operator<(const SmallVectorImpl& RHS) const { + return std::lexicographical_compare( + this->begin(), this->end(), RHS.begin(), RHS.end()); + } +}; + +template +void SmallVectorImpl::swap(SmallVectorImpl& RHS) noexcept { + if (this == &RHS) + return; + + // We can only avoid copying elements if neither vector is small. + if (!this->isSmall() && !RHS.isSmall()) { + std::swap(this->BeginX, RHS.BeginX); + std::swap(this->Size, RHS.Size); + std::swap(this->Capacity, RHS.Capacity); + return; + } + this->reserve(RHS.size()); + RHS.reserve(this->size()); + + // Swap the shared elements. + size_t NumShared = this->size(); + if (NumShared > RHS.size()) + NumShared = RHS.size(); + for (size_type i = 0; i != NumShared; ++i) + std::swap((*this)[i], RHS[i]); + + // Copy over the extra elts. + if (this->size() > RHS.size()) { + size_t EltDiff = this->size() - RHS.size(); + this->uninitialized_copy(this->begin() + NumShared, this->end(), RHS.end()); + RHS.set_size(RHS.size() + EltDiff); + this->destroy_range(this->begin() + NumShared, this->end()); + this->set_size(NumShared); + } else if (RHS.size() > this->size()) { + size_t EltDiff = RHS.size() - this->size(); + this->uninitialized_copy(RHS.begin() + NumShared, RHS.end(), this->end()); + this->set_size(this->size() + EltDiff); + this->destroy_range(RHS.begin() + NumShared, RHS.end()); + RHS.set_size(NumShared); + } +} + +template +SmallVectorImpl& SmallVectorImpl::operator=( + const SmallVectorImpl& RHS) { + // Avoid self-assignment. + if (this == &RHS) + return *this; + + // If we already have sufficient space, assign the common elements, then + // destroy any excess. + size_t RHSSize = RHS.size(); + size_t CurSize = this->size(); + if (CurSize >= RHSSize) { + // Assign common elements. + iterator NewEnd; + if (RHSSize) + NewEnd = std::copy(RHS.begin(), RHS.begin() + RHSSize, this->begin()); + else + NewEnd = this->begin(); + + // Destroy excess elements. + this->destroy_range(NewEnd, this->end()); + + // Trim. + this->set_size(RHSSize); + return *this; + } + + // If we have to grow to have enough elements, destroy the current elements. + // This allows us to avoid copying them during the grow. + // FIXME: don't do this if they're efficiently moveable. + if (this->capacity() < RHSSize) { + // Destroy current elements. + this->clear(); + CurSize = 0; + this->grow(RHSSize); + } else if (CurSize) { + // Otherwise, use assignment for the already-constructed elements. + std::copy(RHS.begin(), RHS.begin() + CurSize, this->begin()); + } + + // Copy construct the new elements in place. + this->uninitialized_copy( + RHS.begin() + CurSize, RHS.end(), this->begin() + CurSize); + + // Set end. + this->set_size(RHSSize); + return *this; +} + +template +SmallVectorImpl& SmallVectorImpl:: +operator=(SmallVectorImpl&& RHS) noexcept( + std::is_nothrow_move_constructible_v && + std::is_nothrow_destructible_v) { + // Avoid self-assignment. + if (this == &RHS) + return *this; + + // If the RHS isn't small, clear this vector and then steal its buffer. + if (!RHS.isSmall()) { + this->destroy_range(this->begin(), this->end()); + if (!this->isSmall()) + free(this->begin()); + this->BeginX = RHS.BeginX; + this->Size = RHS.Size; + this->Capacity = RHS.Capacity; + RHS.resetToSmall(); + return *this; + } + + // If we already have sufficient space, assign the common elements, then + // destroy any excess. + size_t RHSSize = RHS.size(); + size_t CurSize = this->size(); + if (CurSize >= RHSSize) { + // Assign common elements. + iterator NewEnd = this->begin(); + if (RHSSize) + NewEnd = std::move(RHS.begin(), RHS.end(), NewEnd); + + // Destroy excess elements and trim the bounds. + this->destroy_range(NewEnd, this->end()); + this->set_size(RHSSize); + + // Clear the RHS. + RHS.clear(); + + return *this; + } + + // If we have to grow to have enough elements, destroy the current elements. + // This allows us to avoid copying them during the grow. + // FIXME: this may not actually make any sense if we can efficiently move + // elements. + if (this->capacity() < RHSSize) { + // Destroy current elements. + this->clear(); + CurSize = 0; + this->grow(RHSSize); + } else if (CurSize) { + // Otherwise, use assignment for the already-constructed elements. + std::move(RHS.begin(), RHS.begin() + CurSize, this->begin()); + } + + // Move-construct the new elements in place. + this->uninitialized_move( + RHS.begin() + CurSize, RHS.end(), this->begin() + CurSize); + + // Set end. + this->set_size(RHSSize); + + RHS.clear(); + return *this; +} + +/// Storage for the SmallVector elements. This is specialized for the N=0 case +/// to avoid allocating unnecessary storage. +template +struct SmallVectorStorage { + alignas(T) char InlineElts[N * sizeof(T)]; +}; + +/// We need the storage to be properly aligned even for small-size of 0 so that +/// the pointer math in \a SmallVectorTemplateCommon::getFirstEl() is +/// well-defined. +template +struct alignas(T) SmallVectorStorage {}; + +/// Forward declaration of SmallVector so that +/// calculateSmallVectorDefaultInlinedElements can reference +/// `sizeof(SmallVector)`. +template +class /* LLVM_GSL_OWNER */ SmallVector; + +/// Helper class for calculating the default number of inline elements for +/// `SmallVector`. +/// +/// This should be migrated to a constexpr function when our minimum +/// compiler support is enough for multi-statement constexpr functions. +template +struct CalculateSmallVectorDefaultInlinedElements { + // Parameter controlling the default number of inlined elements + // for `SmallVector`. + // + // The default number of inlined elements ensures that + // 1. There is at least one inlined element. + // 2. `sizeof(SmallVector) <= kPreferredSmallVectorSizeof` unless + // it contradicts 1. + static constexpr size_t kPreferredSmallVectorSizeof = 64; + + // static_assert that sizeof(T) is not "too big". + // + // Because our policy guarantees at least one inlined element, it is possible + // for an arbitrarily large inlined element to allocate an arbitrarily large + // amount of inline storage. We generally consider it an antipattern for a + // SmallVector to allocate an excessive amount of inline storage, so we want + // to call attention to these cases and make sure that users are making an + // intentional decision if they request a lot of inline storage. + // + // We want this assertion to trigger in pathological cases, but otherwise + // not be too easy to hit. To accomplish that, the cutoff is actually somewhat + // larger than kPreferredSmallVectorSizeof (otherwise, + // `SmallVector>` would be one easy way to trip it, and that + // pattern seems useful in practice). + // + // One wrinkle is that this assertion is in theory non-portable, since + // sizeof(T) is in general platform-dependent. However, we don't expect this + // to be much of an issue, because most LLVM development happens on 64-bit + // hosts, and therefore sizeof(T) is expected to *decrease* when compiled for + // 32-bit hosts, dodging the issue. The reverse situation, where development + // happens on a 32-bit host and then fails due to sizeof(T) *increasing* on a + // 64-bit host, is expected to be very rare. + static_assert( + sizeof(T) <= 256, + "You are trying to use a default number of inlined elements for " + "`SmallVector` but `sizeof(T)` is really big! Please use an " + "explicit number of inlined elements with `SmallVector` to make " + "sure you really want that much inline storage."); + + // Discount the size of the header itself when calculating the maximum inline + // bytes. + static constexpr size_t PreferredInlineBytes = + kPreferredSmallVectorSizeof - sizeof(SmallVector); + static constexpr size_t NumElementsThatFit = PreferredInlineBytes / sizeof(T); + static constexpr size_t value = + NumElementsThatFit == 0 ? 1 : NumElementsThatFit; +}; + +/// This is a 'vector' (really, a variable-sized array), optimized +/// for the case when the array is small. It contains some number of elements +/// in-place, which allows it to avoid heap allocation when the actual number of +/// elements is below that threshold. This allows normal "small" cases to be +/// fast without losing generality for large inputs. +/// +/// \note +/// In the absence of a well-motivated choice for the number of inlined +/// elements \p N, it is recommended to use \c SmallVector (that is, +/// omitting the \p N). This will choose a default number of inlined elements +/// reasonable for allocation on the stack (for example, trying to keep \c +/// sizeof(SmallVector) around 64 bytes). +/// +/// \warning This does not attempt to be exception safe. +/// +/// \see https://llvm.org/docs/ProgrammersManual.html#llvm-adt-smallvector-h +template < + typename T, + unsigned N = CalculateSmallVectorDefaultInlinedElements::value> +class /* LLVM_GSL_OWNER */ SmallVector : public SmallVectorImpl, + SmallVectorStorage { + public: + SmallVector() : SmallVectorImpl(N) {} + + ~SmallVector() { + // Destroy the constructed elements in the vector. + this->destroy_range(this->begin(), this->end()); + } + + explicit SmallVector(size_t Size, const T& Value = T()) + : SmallVectorImpl(N) { + this->assign(Size, Value); + } + + template < + typename ItTy, + typename = std::enable_if_t::iterator_category, + std::input_iterator_tag>>> + SmallVector(ItTy S, ItTy E) : SmallVectorImpl(N) { + this->append(S, E); + } + + // note: The enable_if restricts Container to types that have a .begin() and + // .end() that return valid input iterators. + template < + typename Container, + std::enable_if_t< + std::is_convertible_v< + typename std::iterator_traits< + decltype(std::declval() + .begin())>::iterator_category, + std::input_iterator_tag> && + std::is_convertible_v< + typename std::iterator_traits< + decltype(std::declval() + .end())>::iterator_category, + std::input_iterator_tag>, + int> = 0> + explicit SmallVector(Container&& c) : SmallVectorImpl(N) { + this->append(c.begin(), c.end()); + } + + SmallVector(std::initializer_list IL) : SmallVectorImpl(N) { + this->assign(IL); + } + + SmallVector(const SmallVector& RHS) : SmallVectorImpl(N) { + if (!RHS.empty()) + SmallVectorImpl::operator=(RHS); + } + + SmallVector& operator=(const SmallVector& RHS) { + SmallVectorImpl::operator=(RHS); + return *this; + } + + SmallVector(SmallVector&& RHS) noexcept( + std::is_nothrow_move_assignable_v>) + : SmallVectorImpl(N) { + if (!RHS.empty()) + SmallVectorImpl::operator=(::std::move(RHS)); + } + + // note: The enable_if restricts Container to types that have a .begin() and + // .end() that return valid input iterators. + template < + typename Container, + std::enable_if_t< + std::is_convertible_v< + typename std::iterator_traits< + decltype(std::declval() + .begin())>::iterator_category, + std::input_iterator_tag> && + std::is_convertible_v< + typename std::iterator_traits< + decltype(std::declval() + .end())>::iterator_category, + std::input_iterator_tag>, + int> = 0> + SmallVector& operator=(const Container& RHS) { + this->assign(RHS.begin(), RHS.end()); + return *this; + } + + SmallVector(SmallVectorImpl&& RHS) noexcept( + std::is_nothrow_move_assignable_v>) + : SmallVectorImpl(N) { + if (!RHS.empty()) + SmallVectorImpl::operator=(::std::move(RHS)); + } + + SmallVector& operator=(SmallVector&& RHS) noexcept( + std::is_nothrow_move_assignable_v>) { + SmallVectorImpl::operator=(::std::move(RHS)); + return *this; + } + + SmallVector& operator=(SmallVectorImpl&& RHS) noexcept( + std::is_nothrow_move_constructible_v>) { + SmallVectorImpl::operator=(::std::move(RHS)); + return *this; + } + + // note: The enable_if restricts Container to types that have a .begin() and + // .end() that return valid input iterators. + template < + typename Container, + std::enable_if_t< + std::is_convertible_v< + typename std::iterator_traits< + decltype(std::declval() + .begin())>::iterator_category, + std::input_iterator_tag> && + std::is_convertible_v< + typename std::iterator_traits< + decltype(std::declval() + .end())>::iterator_category, + std::input_iterator_tag>, + int> = 0> + // NOLINTNEXTLINE(cppcoreguidelines-missing-std-forward) + SmallVector& operator=(Container&& C) { + this->assign(C.begin(), C.end()); + return *this; + } + + SmallVector& operator=(std::initializer_list IL) { + this->assign(IL); + return *this; + } +}; + +template +inline size_t capacity_in_bytes(const SmallVector& X) { + return X.capacity_in_bytes(); +} + +template +std::ostream& operator<<(std::ostream& out, const SmallVector& list) { + int i = 0; + out << '['; + for (auto e : list) { + if (i++ > 0) + out << ", "; + out << e; + } + out << ']'; + return out; +} + +template +using ValueTypeFromRangeType = std::remove_const_t< + std::remove_reference_t()))>>; + +/// Given a range of type R, iterate the entire range and return a +/// SmallVector with elements of the vector. This is useful, for example, +/// when you want to iterate a range and then sort the results. +template +// NOLINTNEXTLINE(cppcoreguidelines-missing-std-forward) +SmallVector, Size> to_vector(R&& Range) { + return {std::begin(Range), std::end(Range)}; +} +template +SmallVector< + ValueTypeFromRangeType, + CalculateSmallVectorDefaultInlinedElements< + ValueTypeFromRangeType>::value> +// NOLINTNEXTLINE(cppcoreguidelines-missing-std-forward) +to_vector(R&& Range) { + return {std::begin(Range), std::end(Range)}; +} + +} // end namespace c10 + +namespace std { + +/// Implement std::swap in terms of SmallVector swap. +template +inline void swap( + c10::SmallVectorImpl& LHS, + c10::SmallVectorImpl& RHS) noexcept { + LHS.swap(RHS); +} + +/// Implement std::swap in terms of SmallVector swap. +template +inline void swap( + c10::SmallVector& LHS, + c10::SmallVector& RHS) noexcept { + LHS.swap(RHS); +} + +} // end namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/StringUtil.h b/backends/pytorch/include/c10/util/StringUtil.h new file mode 100644 index 000000000..7c7790508 --- /dev/null +++ b/backends/pytorch/include/c10/util/StringUtil.h @@ -0,0 +1,267 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#ifndef C10_UTIL_STRINGUTIL_H_ +#define C10_UTIL_STRINGUTIL_H_ + +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include + +C10_CLANG_DIAGNOSTIC_PUSH() +#if C10_CLANG_HAS_WARNING("-Wshorten-64-to-32") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wshorten-64-to-32") +#endif + +namespace c10 { + +namespace detail { + +// Obtains the base name from a full path. +C10_API std::string StripBasename(const std::string& full_path); + +C10_API std::string ExcludeFileExtension(const std::string& full_path); + +struct CompileTimeEmptyString { + operator const std::string&() const { + static const std::string empty_string_literal; + return empty_string_literal; + } + operator const char*() const { + return ""; + } +}; + +template +struct CanonicalizeStrTypes { + using type = const T&; +}; + +template +// NOLINTNEXTLINE(*c-arrays*) +struct CanonicalizeStrTypes { + using type = const char*; +}; + +inline std::ostream& _str(std::ostream& ss) { + return ss; +} + +template +struct Streamable : std::false_type {}; + +template +struct Streamable() << T{})> + : std::true_type {}; + +template +inline std::ostream& _str(std::ostream& ss, const T& t) { + if constexpr (std::is_enum_v && !Streamable::value) { + // NOLINTNEXTLINE(modernize-type-traits) + return _str(ss, static_cast::type>(t)); + } else { + // NOLINTNEXTLINE(clang-analyzer-core.CallAndMessage) + ss << t; + return ss; + } +} + +template +inline std::ostream& _str(std::ostream& ss, const std::optional& t) { + if (t.has_value()) { + return _str(ss, t.value()); + } + ss << "std::nullopt"; + return ss; +} +// Overloads of _str for wide types; forces narrowing. +C10_API std::ostream& _str(std::ostream& ss, const wchar_t* wCStr); +C10_API std::ostream& _str(std::ostream& ss, const wchar_t& wChar); +C10_API std::ostream& _str(std::ostream& ss, const std::wstring& wString); + +template <> +inline std::ostream& _str( + std::ostream& ss, + const CompileTimeEmptyString& /*unused*/) { + return ss; +} + +template +inline std::ostream& _str(std::ostream& ss, const T& t, const Args&... args) { + return _str(_str(ss, t), args...); +} + +template +struct _str_wrapper final { + static std::string call(const Args&... args) { + std::ostringstream ss; + _str(ss, args...); + return ss.str(); + } +}; + +// Specializations for already-a-string types. +template <> +struct _str_wrapper final { + // return by reference to avoid the binary size of a string copy + static const std::string& call(const std::string& str) { + return str; + } +}; + +template <> +struct _str_wrapper final { + static const char* call(const char* str) { + return str; + } +}; + +// For c10::str() with an empty argument list (which is common in our assert +// macros), we don't want to pay the binary size for constructing and +// destructing a stringstream or even constructing a string. +template <> +struct _str_wrapper<> final { + static CompileTimeEmptyString call() { + return CompileTimeEmptyString(); + } +}; + +} // namespace detail + +// Convert a list of string-like arguments into a single string. +template +inline auto str(const Args&... args) { + return detail::_str_wrapper< + typename detail::CanonicalizeStrTypes::type...>::call(args...); +} + +template +inline std::string Join(const std::string& delimiter, const Container& v) { + std::stringstream s; + int cnt = static_cast(v.size()) - 1; + for (auto i = v.begin(); i != v.end(); ++i, --cnt) { + s << (*i) << (cnt ? delimiter : ""); + } + return std::move(s).str(); +} + +// Replace all occurrences of "from" substring to "to" string. +// Returns number of replacements +size_t C10_API +ReplaceAll(std::string& s, std::string_view from, std::string_view to); + +/// Represents a location in source code (for debugging). +struct C10_API SourceLocation { + const char* function; + const char* file; + uint32_t line; +}; + +std::ostream& operator<<(std::ostream& out, const SourceLocation& loc); + +// unix isprint but insensitive to locale +inline bool isPrint(char s) { + return s > 0x1f && s < 0x7f; +} + +inline void printQuotedString(std::ostream& stmt, const std::string_view str) { + stmt << '"'; + for (auto s : str) { + switch (s) { + case '\\': + stmt << "\\\\"; + break; + case '\'': + stmt << "\\'"; + break; + case '\"': + stmt << "\\\""; + break; + case '\a': + stmt << "\\a"; + break; + case '\b': + stmt << "\\b"; + break; + case '\f': + stmt << "\\f"; + break; + case '\n': + stmt << "\\n"; + break; + case '\r': + stmt << "\\r"; + break; + case '\t': + stmt << "\\t"; + break; + case '\v': + stmt << "\\v"; + break; + default: + if (isPrint(s)) { + stmt << s; + } else { + // C++ io has stateful formatting settings. Messing with + // them is probably worse than doing this manually. + // NOLINTNEXTLINE(*c-arrays*) + char buf[4] = "000"; + // NOLINTNEXTLINE(*narrowing-conversions) + buf[2] += s % 8; + s /= 8; + // NOLINTNEXTLINE(*narrowing-conversions) + buf[1] += s % 8; + s /= 8; + // NOLINTNEXTLINE(*narrowing-conversions) + buf[0] += s; + stmt << "\\" << buf; + } + break; + } + } + stmt << '"'; +} + +template +std::optional tryToNumber(const char* symbol) = delete; +template +std::optional tryToNumber(const std::string& symbol) = delete; + +/* + * Convert a string to a 64 bit integer. Trailing whitespaces are not supported. + * Similarly, integer string with trailing characters like "123abc" will be + * rejected. + */ +template <> +C10_API std::optional tryToNumber(const char* symbol); +template <> +C10_API std::optional tryToNumber(const std::string& symbol); + +/* + * Convert a string to a double. Trailing whitespaces are not supported. + * Similarly, integer string with trailing characters like "123abc" will + * be rejected. + */ +template <> +C10_API std::optional tryToNumber(const char* symbol); +template <> +C10_API std::optional tryToNumber(const std::string& symbol); + +C10_API std::vector split( + std::string_view target, + char delimiter); +} // namespace c10 + +C10_CLANG_DIAGNOSTIC_POP() + +#endif // C10_UTIL_STRINGUTIL_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/ThreadLocalDebugInfo.h b/backends/pytorch/include/c10/util/ThreadLocalDebugInfo.h new file mode 100644 index 000000000..03ba6f5b3 --- /dev/null +++ b/backends/pytorch/include/c10/util/ThreadLocalDebugInfo.h @@ -0,0 +1,90 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include + +#include +#include + +namespace c10 { + +enum class C10_API_ENUM DebugInfoKind : uint8_t { + PRODUCER_INFO = 0, + MOBILE_RUNTIME_INFO, + PROFILER_STATE, + INFERENCE_CONTEXT, // for inference usage + PARAM_COMMS_INFO, + + TEST_INFO, // used only in tests + TEST_INFO_2, // used only in tests +}; + +class C10_API DebugInfoBase { + public: + DebugInfoBase() = default; + virtual ~DebugInfoBase() = default; +}; + +// Thread local debug information is propagated across the forward +// (including async fork tasks) and backward passes and is supposed +// to be utilized by the user's code to pass extra information from +// the higher layers (e.g. model id) down to the lower levels +// (e.g. to the operator observers used for debugging, logging, +// profiling, etc) +class C10_API ThreadLocalDebugInfo { + public: + static DebugInfoBase* get(DebugInfoKind kind); + + // Get current ThreadLocalDebugInfo + static std::shared_ptr current(); + + // Internal, use DebugInfoGuard/ThreadLocalStateGuard + static void _forceCurrentDebugInfo( + std::shared_ptr info); + + // Push debug info struct of a given kind + static void _push(DebugInfoKind kind, std::shared_ptr info); + // Pop debug info, throws in case the last pushed + // debug info is not of a given kind + static std::shared_ptr _pop(DebugInfoKind kind); + // Peek debug info, throws in case the last pushed debug info is not of the + // given kind + static std::shared_ptr _peek(DebugInfoKind kind); + + private: + std::shared_ptr info_; + DebugInfoKind kind_; + std::shared_ptr parent_info_; + + friend class DebugInfoGuard; +}; + +// DebugInfoGuard is used to set debug information, +// ThreadLocalDebugInfo is semantically immutable, the values are set +// through the scope-based guard object. +// Nested DebugInfoGuard adds/overrides existing values in the scope, +// restoring the original values after exiting the scope. +// Users can access the values through the ThreadLocalDebugInfo::get() call; +class C10_API DebugInfoGuard { + public: + DebugInfoGuard(DebugInfoKind kind, std::shared_ptr info); + + explicit DebugInfoGuard(std::shared_ptr info); + + ~DebugInfoGuard(); + + DebugInfoGuard(const DebugInfoGuard&) = delete; + DebugInfoGuard(DebugInfoGuard&&) = delete; + DebugInfoGuard& operator=(const DebugInfoGuard&) = delete; + DebugInfoGuard& operator=(DebugInfoGuard&&) = delete; + + private: + bool active_ = false; + std::shared_ptr prev_info_ = nullptr; +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/Type.h b/backends/pytorch/include/c10/util/Type.h new file mode 100644 index 000000000..9f460d4bd --- /dev/null +++ b/backends/pytorch/include/c10/util/Type.h @@ -0,0 +1,35 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#ifndef C10_UTIL_TYPE_H_ +#define C10_UTIL_TYPE_H_ + +#include +#include +#ifdef __GXX_RTTI +#include +#endif // __GXX_RTTI + +#include + +namespace c10 { + +/// Utility to demangle a C++ symbol name. +C10_API std::string demangle(const char* name); + +/// Returns the printable name of the type. +template +inline const char* demangle_type() { +#ifdef __GXX_RTTI + static const auto& name = *(new std::string(demangle(typeid(T).name()))); + return name.c_str(); +#else // __GXX_RTTI + return "(RTTI disabled, cannot show name)"; +#endif // __GXX_RTTI +} + +} // namespace c10 + +#endif // C10_UTIL_TYPE_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/TypeCast.h b/backends/pytorch/include/c10/util/TypeCast.h new file mode 100644 index 000000000..1d95fd909 --- /dev/null +++ b/backends/pytorch/include/c10/util/TypeCast.h @@ -0,0 +1,215 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include + +C10_CLANG_DIAGNOSTIC_PUSH() +#if C10_CLANG_HAS_WARNING("-Wimplicit-float-conversion") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-float-conversion") +#endif +#if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion") +#endif + +namespace c10 { + +template +struct needs_real { + constexpr static bool value = + (is_complex::value && !is_complex::value); +}; + +template +struct maybe_real { + C10_HOST_DEVICE static inline src_t apply(src_t src) { + return src; + } +}; + +template +struct maybe_real { + C10_HOST_DEVICE static inline decltype(auto) apply(src_t src) { + return src.real(); + } +}; + +template +struct maybe_bool { + C10_HOST_DEVICE static inline src_t apply(src_t src) { + return src; + } +}; + +template +struct maybe_bool { + C10_HOST_DEVICE static inline decltype(auto) apply(src_t src) { + // Don't use bool operator so as to also compile for ComplexHalf. + return src.real() || src.imag(); + } +}; + +// Note: deliberately ignores undefined behavior, consistent with NumPy. +// PyTorch's type conversions can cause a variety of undefined behavior, +// including float to integral overflow and signed to unsigned integer overflow. +// Some of this undefined behavior is addressed below. +template +struct static_cast_with_inter_type { + C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline dest_t apply( + src_t src) { + constexpr bool real = needs_real::value; + auto r = maybe_real::apply(src); + return static_cast(r); + } +}; + +// Partial template specialization for casting to bool. +// Need to handle complex types separately, as we don't +// simply want to cast the real part to bool. +template +struct static_cast_with_inter_type { + C10_HOST_DEVICE static inline bool apply(src_t src) { + constexpr bool complex = needs_real::value; + return static_cast(maybe_bool::apply(src)); + } +}; + +// Partial template instantiation for casting to uint8. +// Note: Converting from negative float values to unsigned integer types is +// undefined behavior in C++, and current CPU and GPU compilers exhibit +// divergent behavior. Casting from negative float values to signed +// integer types and then to unsigned integer types is not undefined, +// however, so this cast improves the consistency of type conversions +// to uint8 across compilers. +// Further note: Type conversions across compilers still have other undefined +// and divergent behavior. +template +struct static_cast_with_inter_type { + C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline uint8_t apply( + src_t src) { + constexpr bool real = needs_real::value; + return static_cast( + static_cast(maybe_real::apply(src))); + } +}; + +template <> +struct static_cast_with_inter_type, c10::BFloat16> { + C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< + c10::Half> + apply(c10::BFloat16 src) { + return static_cast>(c10::complex{src}); + } +}; + +template <> +struct static_cast_with_inter_type, c10::Float8_e5m2> { + C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< + c10::Half> + apply(c10::Float8_e5m2 src) { + return static_cast>(c10::complex{src}); + } +}; + +template <> +struct static_cast_with_inter_type< + c10::complex, + c10::Float8_e5m2fnuz> { + C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< + c10::Half> + apply(c10::Float8_e5m2fnuz src) { + return static_cast>(c10::complex{src}); + } +}; + +template <> +struct static_cast_with_inter_type< + c10::complex, + c10::Float8_e4m3fn> { + C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< + c10::Half> + apply(c10::Float8_e4m3fn src) { + return static_cast>(c10::complex{src}); + } +}; + +template <> +struct static_cast_with_inter_type< + c10::complex, + c10::Float8_e4m3fnuz> { + C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< + c10::Half> + apply(c10::Float8_e4m3fnuz src) { + return static_cast>(c10::complex{src}); + } +}; + +// TODO(#146647): Can we make all these template specialization happen +// based off our apply macros? +template <> +struct static_cast_with_inter_type< + c10::complex, + c10::Float8_e8m0fnu> { + C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< + c10::Half> + apply(c10::Float8_e8m0fnu src) { + return static_cast>(c10::complex{src}); + } +}; + +template <> +struct static_cast_with_inter_type, c10::Half> { + C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< + c10::Half> + apply(c10::Half src) { + return static_cast>(c10::complex{src}); + } +}; + +template <> +struct static_cast_with_inter_type< + c10::complex, + c10::complex> { + C10_HOST_DEVICE __ubsan_ignore_undefined__ static inline c10::complex< + c10::Half> + apply(c10::complex src) { + return static_cast>( + static_cast>(src)); + } +}; + +template +C10_HOST_DEVICE To convert(From f) { + return static_cast_with_inter_type::apply(f); +} + +// Define separately to avoid being inlined and prevent code-size bloat +[[noreturn]] C10_API void report_overflow(const char* name); + +template +To checked_convert(From f, const char* name) { + // Converting to bool can't overflow so we exclude this case from checking. + if (!std::is_same_v && overflows(f)) { + report_overflow(name); + } + return convert(f); +} + +} // namespace c10 + +C10_CLANG_DIAGNOSTIC_POP() + +// Trigger tests for D25440771. TODO: Remove this line any time you want. + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/TypeIndex.h b/backends/pytorch/include/c10/util/TypeIndex.h new file mode 100644 index 000000000..fe2282d29 --- /dev/null +++ b/backends/pytorch/include/c10/util/TypeIndex.h @@ -0,0 +1,132 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include + +#if !defined(FBCODE_CAFFE2) && !defined(C10_NODEPRECATED) +#define C10_TYPENAME_SUPPORTS_CONSTEXPR 1 +#define C10_TYPENAME_CONSTEXPR constexpr +#endif + +namespace c10::util { + +struct type_index final : IdWrapper { + constexpr explicit type_index(uint64_t checksum) : IdWrapper(checksum) {} + + // Allow usage in std::map / std::set + // TODO Disallow this and rather use std::unordered_map/set everywhere + friend constexpr bool operator<(type_index lhs, type_index rhs) noexcept { + return lhs.underlyingId() < rhs.underlyingId(); + } + + friend std::ostream& operator<<(std::ostream& stream, type_index typeId) { + return stream << typeId.underlyingId(); + } +}; + +namespace detail { + +template +inline constexpr c10::c10_string_view fully_qualified_type_name_impl() { +#if defined(_MSC_VER) && !defined(__clang__) + constexpr std::string_view fun_sig = __FUNCSIG__; +#if defined(__NVCC__) + constexpr std::string_view prefix = + "c10::basic_string_view c10::util::detail::fully_qualified_type_name_impl<"; + constexpr std::string_view suffix = ">()"; +#else + constexpr std::string_view prefix = + "class c10::basic_string_view __cdecl c10::util::detail::fully_qualified_type_name_impl<"; + constexpr std::string_view suffix = ">(void)"; +#endif +#elif defined(__clang__) + constexpr std::string_view fun_sig = __PRETTY_FUNCTION__; + constexpr std::string_view prefix = + "c10::c10_string_view c10::util::detail::fully_qualified_type_name_impl() [T = "; + constexpr std::string_view suffix = "]"; +#elif defined(__GNUC__) + constexpr std::string_view fun_sig = __PRETTY_FUNCTION__; + constexpr std::string_view prefix = + "constexpr c10::c10_string_view c10::util::detail::fully_qualified_type_name_impl() [with T = "; + constexpr std::string_view suffix = + "; c10::c10_string_view = c10::basic_string_view]"; +#endif +#if !defined(__CUDA_ARCH__) && !defined(__CUDA_ARCH_LIST__) + static_assert(c10::starts_with( + static_cast(fun_sig), + static_cast(prefix))); + static_assert(c10::ends_with( + static_cast(fun_sig), + static_cast(suffix))); +#endif + return fun_sig.substr( + prefix.size(), fun_sig.size() - prefix.size() - suffix.size()); +} + +#if !defined(__CUDA_ARCH__) && !defined(__CUDA_ARCH_LIST__) +template +inline constexpr uint64_t type_index_impl() { +// Idea: __PRETTY_FUNCTION__ (or __FUNCSIG__ on msvc) contains a qualified name +// of this function, including its template parameter, i.e. including the +// type we want an id for. We use this name and run crc64 on it to get a type +// id. +#if defined(_MSC_VER) && !defined(__clang__) + return crc64(__FUNCSIG__, sizeof(__FUNCSIG__)).checksum(); +#elif defined(__clang__) + return crc64(__PRETTY_FUNCTION__, sizeof(__PRETTY_FUNCTION__)).checksum(); +#elif defined(__GNUC__) + return crc64(__PRETTY_FUNCTION__, sizeof(__PRETTY_FUNCTION__)).checksum(); +#endif +} +#endif + +} // namespace detail + +template +inline constexpr type_index get_type_index() { +#if !defined(__CUDA_ARCH__) && !defined(__CUDA_ARCH_LIST__) + // To enforce that this is really computed at compile time, we pass the + // type index through std::integral_constant. + return type_index{std::integral_constant< + uint64_t, + detail::type_index_impl>()>::value}; +#else + // There's nothing in theory preventing us from running this on device code + // except for nvcc throwing a compiler error if we enable it. + return (abort(), type_index(0)); +#endif +} + +#if !defined(TORCH_PEDANTIC) +// Use precomputed hashsum for std::string +// Needed to workaround ambiguity in class name resolution +// into __PRETTY_FUNCTION__ when abovementioned class is defined in inlined +// namespace. In multi-ABI C++ library, `std::string` is an alias to +// `std::__cxx11::basic_string` which depending on compiler flags can be +// resolved to `basic_string` either in `std` namespace or in +// `std::__cxx11` one (`__cxx11` is an inline namespace) +template <> +inline constexpr type_index get_type_index() { + // hashsum for std::basic_string + return type_index{4193213214807308375ULL}; +} +#endif + +template +inline constexpr std::string_view get_fully_qualified_type_name() noexcept { + return static_cast( + detail::fully_qualified_type_name_impl()); +} +} // namespace c10::util + +C10_DEFINE_HASH_FOR_IDWRAPPER(c10::util::type_index) + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/TypeList.h b/backends/pytorch/include/c10/util/TypeList.h new file mode 100644 index 000000000..7386bacca --- /dev/null +++ b/backends/pytorch/include/c10/util/TypeList.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/TypeSafeSignMath.h b/backends/pytorch/include/c10/util/TypeSafeSignMath.h new file mode 100644 index 000000000..f511333fc --- /dev/null +++ b/backends/pytorch/include/c10/util/TypeSafeSignMath.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/TypeTraits.h b/backends/pytorch/include/c10/util/TypeTraits.h new file mode 100644 index 000000000..9d49c82cb --- /dev/null +++ b/backends/pytorch/include/c10/util/TypeTraits.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/UniqueVoidPtr.h b/backends/pytorch/include/c10/util/UniqueVoidPtr.h new file mode 100644 index 000000000..dc2ba274c --- /dev/null +++ b/backends/pytorch/include/c10/util/UniqueVoidPtr.h @@ -0,0 +1,145 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once +#include +#include +#include + +#include +#include + +namespace c10 { + +using DeleterFnPtr = void (*)(void*); + +namespace detail { + +// Does not delete anything +C10_API void deleteNothing(void* /*unused*/); + +// A detail::UniqueVoidPtr is an owning smart pointer like unique_ptr, but +// with three major differences: +// +// 1) It is specialized to void +// +// 2) It is specialized for a function pointer deleter +// void(void* ctx); i.e., the deleter doesn't take a +// reference to the data, just to a context pointer +// (erased as void*). In fact, internally, this pointer +// is implemented as having an owning reference to +// context, and a non-owning reference to data; this is why +// you release_context(), not release() (the conventional +// API for release() wouldn't give you enough information +// to properly dispose of the object later.) +// +// 3) The deleter is guaranteed to be called when the unique +// pointer is destructed and the context is non-null; this is different +// from std::unique_ptr where the deleter is not called if the +// data pointer is null. +// +// Some of the methods have slightly different types than std::unique_ptr +// to reflect this. +// +class UniqueVoidPtr { + private: + // Lifetime tied to ctx_ + void* data_; + std::unique_ptr ctx_; + + public: + UniqueVoidPtr() : data_(nullptr), ctx_(nullptr, &deleteNothing) {} + explicit UniqueVoidPtr(void* data) + : data_(data), ctx_(nullptr, &deleteNothing) {} + UniqueVoidPtr(void* data, void* ctx, DeleterFnPtr ctx_deleter) + : data_(data), ctx_(ctx, ctx_deleter ? ctx_deleter : &deleteNothing) {} + void* operator->() const { + return data_; + } + void clear() { + ctx_ = nullptr; + data_ = nullptr; + } + void* get() const { + return data_; + } + + bool /* success */ unsafe_reset_data_and_ctx(void* new_data_and_ctx) { + if (C10_UNLIKELY(ctx_.get_deleter() != &deleteNothing)) { + return false; + } + // seems quicker than calling the no-op deleter when we reset + // NOLINTNEXTLINE(bugprone-unused-return-value) + ctx_.release(); + ctx_.reset(new_data_and_ctx); + data_ = new_data_and_ctx; + return true; + } + + void* get_context() const { + return ctx_.get(); + } + void* release_context() { + return ctx_.release(); + } + std::unique_ptr&& move_context() { + return std::move(ctx_); + } + [[nodiscard]] bool compare_exchange_deleter( + DeleterFnPtr expected_deleter, + DeleterFnPtr new_deleter) { + if (get_deleter() != expected_deleter) + return false; + ctx_ = std::unique_ptr(ctx_.release(), new_deleter); + return true; + } + + template + T* cast_context(DeleterFnPtr expected_deleter) const { + if (get_deleter() != expected_deleter) + return nullptr; + return static_cast(get_context()); + } + operator bool() const { + return data_ || ctx_; + } + DeleterFnPtr get_deleter() const { + return ctx_.get_deleter(); + } +}; + +// Note [How UniqueVoidPtr is implemented] +// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +// UniqueVoidPtr solves a common problem for allocators of tensor data, which +// is that the data pointer (e.g., float*) which you are interested in, is not +// the same as the context pointer (e.g., DLManagedTensor) which you need +// to actually deallocate the data. Under a conventional deleter design, you +// have to store extra context in the deleter itself so that you can actually +// delete the right thing. Implementing this with standard C++ is somewhat +// error-prone: if you use a std::unique_ptr to manage tensors, the deleter will +// not be called if the data pointer is nullptr, which can cause a leak if the +// context pointer is non-null (and the deleter is responsible for freeing both +// the data pointer and the context pointer). +// +// So, in our reimplementation of unique_ptr, which just store the context +// directly in the unique pointer, and attach the deleter to the context +// pointer itself. In simple cases, the context pointer is just the pointer +// itself. + +inline bool operator==(const UniqueVoidPtr& sp, std::nullptr_t) noexcept { + return !sp; +} +inline bool operator==(std::nullptr_t, const UniqueVoidPtr& sp) noexcept { + return !sp; +} +inline bool operator!=(const UniqueVoidPtr& sp, std::nullptr_t) noexcept { + return sp; +} +inline bool operator!=(std::nullptr_t, const UniqueVoidPtr& sp) noexcept { + return sp; +} + +} // namespace detail +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/accumulate.h b/backends/pytorch/include/c10/util/accumulate.h new file mode 100644 index 000000000..df0899a2c --- /dev/null +++ b/backends/pytorch/include/c10/util/accumulate.h @@ -0,0 +1,129 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +// Copyright 2004-present Facebook. All Rights Reserved. + +#pragma once + +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +/// Sum of a list of integers; accumulates into the int64_t datatype +template < + typename C, + std::enable_if_t, int> = 0> +inline int64_t sum_integers(const C& container) { + // std::accumulate infers return type from `init` type, so if the `init` type + // is not large enough to hold the result, computation can overflow. We use + // `int64_t` here to avoid this. + return std::accumulate( + container.begin(), container.end(), static_cast(0)); +} + +/// Sum of integer elements referred to by iterators; accumulates into the +/// int64_t datatype +template < + typename Iter, + std::enable_if_t< + std::is_integral_v::value_type>, + int> = 0> +inline int64_t sum_integers(Iter begin, Iter end) { + // std::accumulate infers return type from `init` type, so if the `init` type + // is not large enough to hold the result, computation can overflow. We use + // `int64_t` here to avoid this. + return std::accumulate(begin, end, static_cast(0)); +} + +/// Product of a list of integers; accumulates into the int64_t datatype +template < + typename C, + std::enable_if_t, int> = 0> +inline int64_t multiply_integers(const C& container) { + // std::accumulate infers return type from `init` type, so if the `init` type + // is not large enough to hold the result, computation can overflow. We use + // `int64_t` here to avoid this. + return std::accumulate( + container.begin(), + container.end(), + static_cast(1), + std::multiplies<>()); +} + +/// Product of integer elements referred to by iterators; accumulates into the +/// int64_t datatype +template < + typename Iter, + std::enable_if_t< + std::is_integral_v::value_type>, + int> = 0> +inline int64_t multiply_integers(Iter begin, Iter end) { + // std::accumulate infers return type from `init` type, so if the `init` type + // is not large enough to hold the result, computation can overflow. We use + // `int64_t` here to avoid this. + return std::accumulate( + begin, end, static_cast(1), std::multiplies<>()); +} + +/// Return product of all dimensions starting from k +/// Returns 1 if k>=dims.size() +template < + typename C, + std::enable_if_t, int> = 0> +inline int64_t numelements_from_dim(const int k, const C& dims) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(k >= 0); + + if (k > static_cast(dims.size())) { + return 1; + } else { + auto cbegin = dims.cbegin(); + std::advance(cbegin, k); + return multiply_integers(cbegin, dims.cend()); + } +} + +/// Product of all dims up to k (not including dims[k]) +/// Throws an error if k>dims.size() +template < + typename C, + std::enable_if_t, int> = 0> +inline int64_t numelements_to_dim(const int k, const C& dims) { + TORCH_INTERNAL_ASSERT(0 <= k); + TORCH_INTERNAL_ASSERT((unsigned)k <= dims.size()); + + auto cend = dims.cbegin(); + std::advance(cend, k); + return multiply_integers(dims.cbegin(), cend); +} + +/// Product of all dims between k and l (including dims[k] and excluding +/// dims[l]) k and l may be supplied in either order +template < + typename C, + std::enable_if_t, int> = 0> +inline int64_t numelements_between_dim(int k, int l, const C& dims) { + TORCH_INTERNAL_ASSERT(0 <= k); + TORCH_INTERNAL_ASSERT(0 <= l); + + if (k > l) { + std::swap(k, l); + } + + TORCH_INTERNAL_ASSERT((unsigned)l < dims.size()); + + auto cbegin = dims.cbegin(); + auto cend = dims.cbegin(); + std::advance(cbegin, k); + std::advance(cend, l); + return multiply_integers(cbegin, cend); +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/bit_cast.h b/backends/pytorch/include/c10/util/bit_cast.h new file mode 100644 index 000000000..948d03d50 --- /dev/null +++ b/backends/pytorch/include/c10/util/bit_cast.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/bits.h b/backends/pytorch/include/c10/util/bits.h new file mode 100644 index 000000000..fe5b67c45 --- /dev/null +++ b/backends/pytorch/include/c10/util/bits.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/complex.h b/backends/pytorch/include/c10/util/complex.h new file mode 100644 index 000000000..ff5ea55c5 --- /dev/null +++ b/backends/pytorch/include/c10/util/complex.h @@ -0,0 +1,83 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include + +#include +#include +#include + +// std functions +// +// The implementation of these functions also follow the design of C++20 + +namespace std { + +template +constexpr T real(const c10::complex& z) { + return z.real(); +} + +template +constexpr T imag(const c10::complex& z) { + return z.imag(); +} + +template +C10_HOST_DEVICE T abs(const c10::complex& z) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return thrust::abs(static_cast>(z)); +#else + return std::abs(static_cast>(z)); +#endif +} + +#if defined(USE_ROCM) +#define ROCm_Bug(x) +#else +#define ROCm_Bug(x) x +#endif + +template +C10_HOST_DEVICE T arg(const c10::complex& z) { + return ROCm_Bug(std)::atan2(std::imag(z), std::real(z)); +} + +#undef ROCm_Bug + +template +constexpr T norm(const c10::complex& z) { + return z.real() * z.real() + z.imag() * z.imag(); +} + +// For std::conj, there are other versions of it: +// constexpr std::complex conj( float z ); +// template< class DoubleOrInteger > +// constexpr std::complex conj( DoubleOrInteger z ); +// constexpr std::complex conj( long double z ); +// These are not implemented +// TODO(@zasdfgbnm): implement them as c10::conj +template +constexpr c10::complex conj(const c10::complex& z) { + return c10::complex(z.real(), -z.imag()); +} + +// Thrust does not have complex --> complex version of thrust::proj, +// so this function is not implemented at c10 right now. +// TODO(@zasdfgbnm): implement it by ourselves + +// There is no c10 version of std::polar, because std::polar always +// returns std::complex. Use c10::polar instead; + +} // namespace std + +#define C10_INTERNAL_INCLUDE_COMPLEX_REMAINING_H +// math functions are included in a separate file +#include // IWYU pragma: keep +// utilities for complex types +#include // IWYU pragma: keep +#undef C10_INTERNAL_INCLUDE_COMPLEX_REMAINING_H + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/complex_math.h b/backends/pytorch/include/c10/util/complex_math.h new file mode 100644 index 000000000..33da59051 --- /dev/null +++ b/backends/pytorch/include/c10/util/complex_math.h @@ -0,0 +1,411 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#if !defined(C10_INTERNAL_INCLUDE_COMPLEX_REMAINING_H) +#error \ + "c10/util/complex_math.h is not meant to be individually included. Include c10/util/complex.h instead." +#endif + +namespace c10_complex_math { + +// Exponential functions + +template +C10_HOST_DEVICE inline c10::complex exp(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::exp(static_cast>(x))); +#else + return static_cast>( + std::exp(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex log(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::log(static_cast>(x))); +#else + return static_cast>( + std::log(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex log10(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::log10(static_cast>(x))); +#else + return static_cast>( + std::log10(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex log2(const c10::complex& x) { + const c10::complex log2 = c10::complex(::log(2.0), 0.0); + return c10_complex_math::log(x) / log2; +} + +// Power functions +// +#if defined(_LIBCPP_VERSION) || \ + (defined(__GLIBCXX__) && !defined(_GLIBCXX11_USE_C99_COMPLEX)) +namespace _detail { +C10_API c10::complex sqrt(const c10::complex& in); +C10_API c10::complex sqrt(const c10::complex& in); +C10_API c10::complex acos(const c10::complex& in); +C10_API c10::complex acos(const c10::complex& in); +} // namespace _detail +#endif + +template +C10_HOST_DEVICE inline c10::complex sqrt(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::sqrt(static_cast>(x))); +#elif !( \ + defined(_LIBCPP_VERSION) || \ + (defined(__GLIBCXX__) && !defined(_GLIBCXX11_USE_C99_COMPLEX))) + return static_cast>( + std::sqrt(static_cast>(x))); +#else + return _detail::sqrt(x); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex pow( + const c10::complex& x, + const c10::complex& y) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>(thrust::pow( + static_cast>(x), static_cast>(y))); +#else + return static_cast>(std::pow( + static_cast>(x), static_cast>(y))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex pow( + const c10::complex& x, + const T& y) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::pow(static_cast>(x), y)); +#else + return static_cast>( + std::pow(static_cast>(x), y)); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex pow( + const T& x, + const c10::complex& y) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::pow(x, static_cast>(y))); +#else + return static_cast>( + std::pow(x, static_cast>(y))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex pow( + const c10::complex& x, + const c10::complex& y) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>(thrust::pow( + static_cast>(x), static_cast>(y))); +#else + return static_cast>(std::pow( + static_cast>(x), static_cast>(y))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex pow( + const c10::complex& x, + const U& y) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::pow(static_cast>(x), y)); +#else + return static_cast>( + std::pow(static_cast>(x), y)); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex pow( + const T& x, + const c10::complex& y) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::pow(x, static_cast>(y))); +#else + return static_cast>( + std::pow(x, static_cast>(y))); +#endif +} + +// Trigonometric functions + +template +C10_HOST_DEVICE inline c10::complex sin(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::sin(static_cast>(x))); +#else + return static_cast>( + std::sin(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex cos(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::cos(static_cast>(x))); +#else + return static_cast>( + std::cos(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex tan(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::tan(static_cast>(x))); +#else + return static_cast>( + std::tan(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex asin(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::asin(static_cast>(x))); +#else + return static_cast>( + std::asin(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex acos(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::acos(static_cast>(x))); +#elif !defined(_LIBCPP_VERSION) + return static_cast>( + std::acos(static_cast>(x))); +#else + return _detail::acos(x); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex atan(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::atan(static_cast>(x))); +#else + return static_cast>( + std::atan(static_cast>(x))); +#endif +} + +// Hyperbolic functions + +template +C10_HOST_DEVICE inline c10::complex sinh(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::sinh(static_cast>(x))); +#else + return static_cast>( + std::sinh(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex cosh(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::cosh(static_cast>(x))); +#else + return static_cast>( + std::cosh(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex tanh(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::tanh(static_cast>(x))); +#else + return static_cast>( + std::tanh(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex asinh(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::asinh(static_cast>(x))); +#else + return static_cast>( + std::asinh(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex acosh(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::acosh(static_cast>(x))); +#else + return static_cast>( + std::acosh(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex atanh(const c10::complex& x) { +#if defined(__CUDACC__) || defined(__HIPCC__) + return static_cast>( + thrust::atanh(static_cast>(x))); +#else + return static_cast>( + std::atanh(static_cast>(x))); +#endif +} + +template +C10_HOST_DEVICE inline c10::complex log1p(const c10::complex& z) { +#if defined(__APPLE__) || defined(__MACOSX) || defined(__CUDACC__) || \ + defined(__HIPCC__) + // For Mac, the new implementation yielded a high relative error. Falling back + // to the old version for now. + // See https://github.com/numpy/numpy/pull/22611#issuecomment-1667945354 + // For CUDA we also use this one, as thrust::log(thrust::complex) takes + // *forever* to compile + + // log1p(z) = log(1 + z) + // Let's define 1 + z = r * e ^ (i * a), then we have + // log(r * e ^ (i * a)) = log(r) + i * a + // With z = x + iy, the term r can be written as + // r = ((1 + x) ^ 2 + y ^ 2) ^ 0.5 + // = (1 + x ^ 2 + 2 * x + y ^ 2) ^ 0.5 + // So, log(r) is + // log(r) = 0.5 * log(1 + x ^ 2 + 2 * x + y ^ 2) + // = 0.5 * log1p(x * (x + 2) + y ^ 2) + // we need to use the expression only on certain condition to avoid overflow + // and underflow from `(x * (x + 2) + y ^ 2)` + T x = z.real(); + T y = z.imag(); + T zabs = std::abs(z); + T theta = std::atan2(y, x + T(1)); + if (zabs < 0.5) { + T r = x * (T(2) + x) + y * y; + if (r == 0) { // handle underflow + return {x, theta}; + } + return {T(0.5) * std::log1p(r), theta}; + } else { + T z0 = std::hypot(x + 1, y); + return {std::log(z0), theta}; + } +#else + // CPU path + // Based on https://github.com/numpy/numpy/pull/22611#issuecomment-1667945354 + c10::complex u = z + T(1); + if (u == T(1)) { + return z; + } else { + auto log_u = log(u); + if (u - T(1) == z) { + return log_u; + } + return log_u * (z / (u - T(1))); + } +#endif +} + +template +C10_HOST_DEVICE inline c10::complex expm1(const c10::complex& z) { + // expm1(z) = exp(z) - 1 + // Define z = x + i * y + // f = e ^ (x + i * y) - 1 + // = e ^ x * e ^ (i * y) - 1 + // = (e ^ x * cos(y) - 1) + i * (e ^ x * sin(y)) + // = (e ^ x - 1) * cos(y) - (1 - cos(y)) + i * e ^ x * sin(y) + // = expm1(x) * cos(y) - 2 * sin(y / 2) ^ 2 + i * e ^ x * sin(y) + T x = z.real(); + T y = z.imag(); + T a = std::sin(y / 2); + T er = std::expm1(x) * std::cos(y) - T(2) * a * a; + T ei = std::exp(x) * std::sin(y); + return {er, ei}; +} + +} // namespace c10_complex_math + +using c10_complex_math::acos; +using c10_complex_math::acosh; +using c10_complex_math::asin; +using c10_complex_math::asinh; +using c10_complex_math::atan; +using c10_complex_math::atanh; +using c10_complex_math::cos; +using c10_complex_math::cosh; +using c10_complex_math::exp; +using c10_complex_math::expm1; +using c10_complex_math::log; +using c10_complex_math::log10; +using c10_complex_math::log1p; +using c10_complex_math::log2; +using c10_complex_math::pow; +using c10_complex_math::sin; +using c10_complex_math::sinh; +using c10_complex_math::sqrt; +using c10_complex_math::tan; +using c10_complex_math::tanh; + +namespace std { + +using c10_complex_math::acos; +using c10_complex_math::acosh; +using c10_complex_math::asin; +using c10_complex_math::asinh; +using c10_complex_math::atan; +using c10_complex_math::atanh; +using c10_complex_math::cos; +using c10_complex_math::cosh; +using c10_complex_math::exp; +using c10_complex_math::expm1; +using c10_complex_math::log; +using c10_complex_math::log10; +using c10_complex_math::log1p; +using c10_complex_math::log2; +using c10_complex_math::pow; +using c10_complex_math::sin; +using c10_complex_math::sinh; +using c10_complex_math::sqrt; +using c10_complex_math::tan; +using c10_complex_math::tanh; + +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/complex_utils.h b/backends/pytorch/include/c10/util/complex_utils.h new file mode 100644 index 000000000..44152b72c --- /dev/null +++ b/backends/pytorch/include/c10/util/complex_utils.h @@ -0,0 +1,51 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#if !defined(C10_INTERNAL_INCLUDE_COMPLEX_REMAINING_H) +#error \ + "c10/util/complex_utils.h is not meant to be individually included. Include c10/util/complex.h instead." +#endif + +#include + +namespace c10 { + +template +struct is_complex : public std::false_type {}; + +template +struct is_complex> : public std::true_type {}; + +template +struct is_complex> : public std::true_type {}; + +// Extract double from std::complex; is identity otherwise +// TODO: Write in more idiomatic C++17 +template +struct scalar_value_type { + using type = T; +}; +template +struct scalar_value_type> { + using type = T; +}; +template +struct scalar_value_type> { + using type = T; +}; + +} // namespace c10 + +namespace std { + +template +class numeric_limits> : public numeric_limits {}; + +template +bool isnan(const c10::complex& v) { + return std::isnan(v.real()) || std::isnan(v.imag()); +} + +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/hash.h b/backends/pytorch/include/c10/util/hash.h new file mode 100644 index 000000000..c3fff1284 --- /dev/null +++ b/backends/pytorch/include/c10/util/hash.h @@ -0,0 +1,384 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include + +namespace c10 { + +// NOTE: hash_combine and SHA1 hashing is based on implementation from Boost +// +// Boost Software License - Version 1.0 - August 17th, 2003 +// +// Permission is hereby granted, free of charge, to any person or organization +// obtaining a copy of the software and accompanying documentation covered by +// this license (the "Software") to use, reproduce, display, distribute, +// execute, and transmit the Software, and to prepare derivative works of the +// Software, and to permit third-parties to whom the Software is furnished to +// do so, all subject to the following: +// +// The copyright notices in the Software and this entire statement, including +// the above license grant, this restriction and the following disclaimer, +// must be included in all copies of the Software, in whole or in part, and +// all derivative works of the Software, unless such copies or derivative +// works are solely in the form of machine-executable object code generated by +// a source language processor. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT +// SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE +// FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE, +// ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER +// DEALINGS IN THE SOFTWARE. + +inline size_t hash_combine(size_t seed, size_t value) { + return seed ^ (value + 0x9e3779b9 + (seed << 6u) + (seed >> 2u)); +} + +// Creates the SHA1 hash of a string. A 160-bit hash. +// Based on the implementation in Boost (see notice above). +// Note that SHA1 hashes are no longer considered cryptographically +// secure, but are the standard hash for generating unique ids. +// Usage: +// // Let 'code' be a std::string +// c10::sha1 sha1_hash{code}; +// const auto hash_code = sha1_hash.str(); +// TODO: Compare vs OpenSSL and/or CryptoPP implementations +struct sha1 { + typedef unsigned int(digest_type)[5]; + + sha1(const std::string& s = "") { + if (!s.empty()) { + reset(); + process_bytes(s.c_str(), s.size()); + } + } + + void reset() { + h_[0] = 0x67452301; + h_[1] = 0xEFCDAB89; + h_[2] = 0x98BADCFE; + h_[3] = 0x10325476; + h_[4] = 0xC3D2E1F0; + + block_byte_index_ = 0; + bit_count_low = 0; + bit_count_high = 0; + } + + std::string str() { + unsigned int digest[5]; + get_digest(digest); + + std::ostringstream buf; + for (unsigned int i : digest) { + buf << std::hex << std::setfill('0') << std::setw(8) << i; + } + + return buf.str(); + } + + private: + unsigned int left_rotate(unsigned int x, std::size_t n) { + return (x << n) ^ (x >> (32 - n)); + } + + void process_block_impl() { + unsigned int w[80]; + + for (std::size_t i = 0; i < 16; ++i) { + w[i] = (block_[i * 4 + 0] << 24); + w[i] |= (block_[i * 4 + 1] << 16); + w[i] |= (block_[i * 4 + 2] << 8); + w[i] |= (block_[i * 4 + 3]); + } + + for (std::size_t i = 16; i < 80; ++i) { + w[i] = left_rotate((w[i - 3] ^ w[i - 8] ^ w[i - 14] ^ w[i - 16]), 1); + } + + unsigned int a = h_[0]; + unsigned int b = h_[1]; + unsigned int c = h_[2]; + unsigned int d = h_[3]; + unsigned int e = h_[4]; + + for (std::size_t i = 0; i < 80; ++i) { + unsigned int f = 0; + unsigned int k = 0; + + if (i < 20) { + f = (b & c) | (~b & d); + k = 0x5A827999; + } else if (i < 40) { + f = b ^ c ^ d; + k = 0x6ED9EBA1; + } else if (i < 60) { + f = (b & c) | (b & d) | (c & d); + k = 0x8F1BBCDC; + } else { + f = b ^ c ^ d; + k = 0xCA62C1D6; + } + + unsigned temp = left_rotate(a, 5) + f + e + k + w[i]; + e = d; + d = c; + c = left_rotate(b, 30); + b = a; + a = temp; + } + + h_[0] += a; + h_[1] += b; + h_[2] += c; + h_[3] += d; + h_[4] += e; + } + + void process_byte_impl(unsigned char byte) { + block_[block_byte_index_++] = byte; + + if (block_byte_index_ == 64) { + block_byte_index_ = 0; + process_block_impl(); + } + } + + void process_byte(unsigned char byte) { + process_byte_impl(byte); + + // size_t max value = 0xFFFFFFFF + // if (bit_count_low + 8 >= 0x100000000) { // would overflow + // if (bit_count_low >= 0x100000000-8) { + if (bit_count_low < 0xFFFFFFF8) { + bit_count_low += 8; + } else { + bit_count_low = 0; + + if (bit_count_high <= 0xFFFFFFFE) { + ++bit_count_high; + } else { + TORCH_CHECK(false, "sha1 too many bytes"); + } + } + } + + void process_block(void const* bytes_begin, void const* bytes_end) { + unsigned char const* begin = static_cast(bytes_begin); + unsigned char const* end = static_cast(bytes_end); + for (; begin != end; ++begin) { + process_byte(*begin); + } + } + + void process_bytes(void const* buffer, std::size_t byte_count) { + unsigned char const* b = static_cast(buffer); + process_block(b, b + byte_count); + } + + void get_digest(digest_type& digest) { + // append the bit '1' to the message + process_byte_impl(0x80); + + // append k bits '0', where k is the minimum number >= 0 + // such that the resulting message length is congruent to 56 (mod 64) + // check if there is enough space for padding and bit_count + if (block_byte_index_ > 56) { + // finish this block + while (block_byte_index_ != 0) { + process_byte_impl(0); + } + + // one more block + while (block_byte_index_ < 56) { + process_byte_impl(0); + } + } else { + while (block_byte_index_ < 56) { + process_byte_impl(0); + } + } + + // append length of message (before pre-processing) + // as a 64-bit big-endian integer + process_byte_impl( + static_cast((bit_count_high >> 24) & 0xFF)); + process_byte_impl( + static_cast((bit_count_high >> 16) & 0xFF)); + process_byte_impl(static_cast((bit_count_high >> 8) & 0xFF)); + process_byte_impl(static_cast((bit_count_high) & 0xFF)); + process_byte_impl(static_cast((bit_count_low >> 24) & 0xFF)); + process_byte_impl(static_cast((bit_count_low >> 16) & 0xFF)); + process_byte_impl(static_cast((bit_count_low >> 8) & 0xFF)); + process_byte_impl(static_cast((bit_count_low) & 0xFF)); + + // get final digest + digest[0] = h_[0]; + digest[1] = h_[1]; + digest[2] = h_[2]; + digest[3] = h_[3]; + digest[4] = h_[4]; + } + + unsigned int h_[5]{}; + unsigned char block_[64]{}; + std::size_t block_byte_index_{}; + std::size_t bit_count_low{}; + std::size_t bit_count_high{}; +}; + +constexpr uint64_t twang_mix64(uint64_t key) noexcept { + key = (~key) + (key << 21); // key *= (1 << 21) - 1; key -= 1; + key = key ^ (key >> 24); + key = key + (key << 3) + (key << 8); // key *= 1 + (1 << 3) + (1 << 8) + key = key ^ (key >> 14); + key = key + (key << 2) + (key << 4); // key *= 1 + (1 << 2) + (1 << 4) + key = key ^ (key >> 28); + key = key + (key << 31); // key *= 1 + (1 << 31) + return key; +} + +//////////////////////////////////////////////////////////////////////////////// +// c10::hash implementation +//////////////////////////////////////////////////////////////////////////////// + +namespace _hash_detail { + +// Use template argument deduction to shorten calls to c10::hash +template +size_t simple_get_hash(const T& o); + +template +using type_if_not_enum = std::enable_if_t, V>; + +// Use SFINAE to dispatch to std::hash if possible, cast enum types to int +// automatically, and fall back to T::hash otherwise. NOTE: C++14 added support +// for hashing enum types to the standard, and some compilers implement it even +// when C++14 flags aren't specified. This is why we have to disable this +// overload if T is an enum type (and use the one below in this case). +template +auto dispatch_hash(const T& o) + -> decltype(std::hash()(o), type_if_not_enum()) { + return std::hash()(o); +} + +template +std::enable_if_t, size_t> dispatch_hash(const T& o) { + using R = std::underlying_type_t; + return std::hash()(static_cast(o)); +} + +template +auto dispatch_hash(const T& o) -> decltype(T::hash(o), size_t()) { + return T::hash(o); +} + +} // namespace _hash_detail + +// Hasher struct +template +struct hash { + size_t operator()(const T& o) const { + return _hash_detail::dispatch_hash(o); + } +}; + +// Specialization for std::tuple +template +struct hash> { + template + struct tuple_hash { + size_t operator()(const std::tuple& t) const { + return hash_combine( + _hash_detail::simple_get_hash(std::get(t)), + tuple_hash()(t)); + } + }; + + template + struct tuple_hash<0, Ts...> { + size_t operator()(const std::tuple& t) const { + return _hash_detail::simple_get_hash(std::get<0>(t)); + } + }; + + size_t operator()(const std::tuple& t) const { + return tuple_hash()(t); + } +}; + +template +struct hash> { + size_t operator()(const std::pair& pair) const { + std::tuple tuple = std::make_tuple(pair.first, pair.second); + return _hash_detail::simple_get_hash(tuple); + } +}; + +template +struct hash> { + size_t operator()(c10::ArrayRef v) const { + size_t seed = 0; + for (const auto& elem : v) { + seed = hash_combine(seed, _hash_detail::simple_get_hash(elem)); + } + return seed; + } +}; + +// Specialization for std::vector +template +struct hash> { + size_t operator()(const std::vector& v) const { + return hash>()(v); + } +}; + +namespace _hash_detail { + +template +size_t simple_get_hash(const T& o) { + return c10::hash()(o); +} + +} // namespace _hash_detail + +// Use this function to actually hash multiple things in one line. +// Dispatches to c10::hash, so it can hash containers. +// Example: +// +// static size_t hash(const MyStruct& s) { +// return get_hash(s.member1, s.member2, s.member3); +// } +template +size_t get_hash(const Types&... args) { + return c10::hash()(std::tie(args...)); +} + +// Specialization for c10::complex +template +struct hash> { + size_t operator()(const c10::complex& c) const { + return get_hash(c.real(), c.imag()); + } +}; + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/intrusive_ptr.h b/backends/pytorch/include/c10/util/intrusive_ptr.h new file mode 100644 index 000000000..148a9bf4a --- /dev/null +++ b/backends/pytorch/include/c10/util/intrusive_ptr.h @@ -0,0 +1,1278 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include + +namespace pybind11 { +template +class class_; +} + +namespace torch::utils { +class PyObjectPreservation; +} + +namespace c10 { +class intrusive_ptr_target; +namespace raw { +namespace weak_intrusive_ptr { +inline void incref(intrusive_ptr_target* self); +} +namespace intrusive_ptr { +inline void incref(intrusive_ptr_target* self); +} + +// constructor tag used by intrusive_ptr constructors +struct DontIncreaseRefcount {}; +} // namespace raw + +namespace detail { +constexpr uint64_t kImpracticallyHugeReferenceCount = 0x0FFFFFFF; +constexpr uint64_t kImpracticallyHugeWeakReferenceCount = + (kImpracticallyHugeReferenceCount << 32); +constexpr uint64_t kReferenceCountOne = 1; +constexpr uint64_t kWeakReferenceCountOne = (kReferenceCountOne << 32); +constexpr uint64_t kUniqueRef = (kReferenceCountOne | kWeakReferenceCountOne); +// Indicates whether the object has a PyObject wrapper. +constexpr uint64_t kHasPyObject = (uint64_t(1) << 63); + +template +struct intrusive_target_default_null_type final { + static constexpr TTarget* singleton() noexcept { + return nullptr; + } +}; + +template +TTarget* assign_ptr_(TTarget* rhs) { + if (FromNullType::singleton() == rhs) { + return ToNullType::singleton(); + } else { + return rhs; + } +} + +inline uint32_t refcount(uint64_t combined_refcount) { + return static_cast(combined_refcount); +} + +inline uint32_t weakcount(uint64_t combined_refcount) { + return static_cast((combined_refcount & ~kHasPyObject) >> 32); +} + +inline bool has_pyobject(uint64_t combined_refcount) { + return (combined_refcount & kHasPyObject) != 0; +} + +inline bool is_uniquely_owned(uint64_t combined_refcount) { + return (combined_refcount & ~detail::kHasPyObject) == detail::kUniqueRef; +} + +// The only requirement for refcount increment is that it happens-before +// decrement, so no additional memory ordering is needed. +inline uint64_t atomic_combined_refcount_increment( + std::atomic& combined_refcount, + uint64_t inc) { + return combined_refcount.fetch_add(inc, std::memory_order_relaxed) + inc; +} + +inline uint32_t atomic_weakcount_increment( + std::atomic& combined_refcount) { + return detail::weakcount(atomic_combined_refcount_increment( + combined_refcount, kWeakReferenceCountOne)); +} + +// The requirement is that all modifications to the managed object happen-before +// invocation of the managed object destructor, and that allocation of the +// managed object storage happens-before deallocation of the storage. +// +// To get this ordering, all non-final decrements must synchronize-with the +// final decrement. So all non-final decrements have to store-release while the +// final decrement has to load-acquire, either directly or with the help of +// fences. But it's easiest just to have all decrements be acq-rel. And it turns +// out, on modern architectures and chips, it's also fastest. +inline uint64_t atomic_combined_refcount_decrement( + std::atomic& combined_refcount, + uint64_t dec) { + return combined_refcount.fetch_sub(dec, std::memory_order_acq_rel) - dec; +} + +inline uint32_t atomic_weakcount_decrement( + std::atomic& combined_refcount) { + return detail::weakcount(atomic_combined_refcount_decrement( + combined_refcount, kWeakReferenceCountOne)); +} + +template +struct TargetTraits { + static constexpr bool can_have_pyobject = false; +}; + +} // namespace detail + +/** + * intrusive_ptr is an alternative to shared_ptr that has better + * performance because it does the refcounting intrusively + * (i.e. in a member of the object itself). + * Your class T needs to inherit from intrusive_ptr_target to allow it to be + * used in an intrusive_ptr. Your class's constructor should not allow + *`this` to escape to other threads or create an intrusive_ptr from `this`. + */ + +// Note [Stack allocated intrusive_ptr_target safety] +// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ +// A well known problem with std::enable_shared_from_this is that it +// allows you to create a std::shared_ptr from a stack allocated object, +// which is totally bogus because the object will die once you return +// from the stack. In intrusive_ptr, we can detect that this has occurred, +// because we set the refcount/weakcount of objects which inherit from +// intrusive_ptr_target to zero, *unless* we can prove that the object +// was dynamically allocated (e.g., via make_intrusive). +// +// Thus, whenever you transmute a T* into a intrusive_ptr, we check +// and make sure that the refcount isn't zero (or, a more subtle +// test for weak_intrusive_ptr, for which the refcount may validly +// be zero, but the weak refcount better not be zero), because that +// tells us if the object was allocated by us. If it wasn't, no +// intrusive_ptr for you! + +// NOLINTNEXTLINE(cppcoreguidelines-virtual-class-destructor) +class C10_API intrusive_ptr_target { + // Note [Weak references for intrusive refcounting] + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // Here's the scheme: + // + // - refcount == number of strong references to the object + // weakcount == number of weak references to the object, + // plus one more if refcount > 0 + // An invariant: refcount > 0 => weakcount > 0 + // + // - c10::StorageImpl stays live as long as there are any strong + // or weak pointers to it (weakcount > 0, since strong + // references count as a +1 to weakcount) + // + // - finalizers are called and data_ptr is deallocated when refcount == 0 + // + // - Once refcount == 0, it can never again be > 0 (the transition + // from > 0 to == 0 is monotonic) + // + // - When you access c10::StorageImpl via a weak pointer, you must + // atomically increment the use count, if it is greater than 0. + // If it is not, you must report that the storage is dead. + // + //.We use a single combined count for refcount and weakcount so that + // we can atomically operate on both at the same time for performance + // and defined behaviors. + // + // Note [PyObject preservation for Tensor and Storages] + // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + // intrusive_ptr has special support for preserving PyObject wrappers + // for TensorImpl and StorageImpl. The most significant bit (kHasPyObject) of + // the combined_refcount_ is used to indicate whether the object has a + // PyObject wrapper. + // + // - The PyObject, if it exists, holds a strong reference to the + // intrusive_ptr_target. + // + // - When the refcount goes from 1 to 2, we incref the PyObject. + // + // - When the refcount goes from 2 to 1, we decref the PyObject. + // + // In other words, the intrusive_ptr keeps the PyObject alive as long as there + // are other C++ references to the intrusive_ptr_target. + + mutable std::atomic combined_refcount_; + static_assert(sizeof(std::atomic) == 8); + static_assert(alignof(std::atomic) == 8); + static_assert(std::atomic::is_always_lock_free); + + template + friend class intrusive_ptr; + friend inline void raw::intrusive_ptr::incref(intrusive_ptr_target* self); + + template + friend class weak_intrusive_ptr; + friend inline void raw::weak_intrusive_ptr::incref( + intrusive_ptr_target* self); + + template + friend struct ExclusivelyOwnedTensorTraits; + + friend class torch::utils::PyObjectPreservation; + + protected: + // protected destructor. We never want to destruct intrusive_ptr_target* + // directly. + virtual ~intrusive_ptr_target() { +// Disable -Wterminate and -Wexceptions so we're allowed to use assertions +// (i.e. throw exceptions) in a destructor. +// We also have to disable -Wunknown-warning-option and -Wpragmas, because +// some other compilers don't know about -Wterminate or -Wexceptions and +// will show a warning about unknown warning options otherwise. +#if defined(_MSC_VER) && !defined(__clang__) +#pragma warning(push) +#pragma warning( \ + disable : 4297) // function assumed not to throw an exception but does +#else +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wpragmas" +#pragma GCC diagnostic ignored "-Wunknown-warning-option" +#pragma GCC diagnostic ignored "-Wterminate" +#pragma GCC diagnostic ignored "-Wexceptions" +#endif + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + // Second condition is there to accommodate + // unsafe_adapt_non_heap_allocated: since we are doing our own + // deallocation in that case, it is correct for each + // expected_decref to have happened (some user code tried to + // decref and thus free the object, but it didn't happen right + // away) or not (no user code tried to free the object, and + // now it's getting destroyed through whatever mechanism the + // caller of unsafe_adapt_non_heap_allocated wanted to + // use). We choose our reference count such that the count + // will not dip below kImpracticallyHugeReferenceCount regardless. + refcount() == 0 || + refcount() >= detail::kImpracticallyHugeReferenceCount, + "Tried to destruct an intrusive_ptr_target that still has intrusive_ptr to it; refcount was ", + refcount()); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + // See ~intrusive_ptr for optimization that will frequently result in 1 + // at destruction time. + weakcount() == 1 || weakcount() == 0 || + weakcount() == detail::kImpracticallyHugeReferenceCount - 1 || + weakcount() == detail::kImpracticallyHugeReferenceCount, + "Tried to destruct an intrusive_ptr_target that still has weak_intrusive_ptr to it"); +#if defined(_MSC_VER) && !defined(__clang__) +#pragma warning(pop) +#else +#pragma GCC diagnostic pop +#endif + } + + constexpr intrusive_ptr_target() noexcept : combined_refcount_(0) {} + + // intrusive_ptr_target supports copy and move: but refcount and weakcount + // don't participate (since they are intrinsic properties of the memory + // location) + intrusive_ptr_target(intrusive_ptr_target&& /*other*/) noexcept + : intrusive_ptr_target() {} + + intrusive_ptr_target& operator=(intrusive_ptr_target&& /*other*/) noexcept { + return *this; + } + + intrusive_ptr_target(const intrusive_ptr_target& /*other*/) noexcept + : intrusive_ptr_target() {} + + intrusive_ptr_target& operator=( + const intrusive_ptr_target& /*other*/) noexcept { + return *this; + } + + private: + /** + * This is called when refcount reaches zero. + * You can override this to release expensive resources. + * There might still be weak references, so your object might not get + * destructed yet, but you can assume the object isn't used anymore, + * i.e. no more calls to methods or accesses to members (we just can't + * destruct it yet because we need the weakcount accessible). + * + * If there are no weak references (i.e. your class is about to be + * destructed), this function WILL NOT be called. + */ + virtual void release_resources() {} + + /** + * These two methods are called when the refcount transitions between one + * and two and the object has a PyObject wrapper. + */ + virtual void incref_pyobject() const noexcept {} + virtual void decref_pyobject() const noexcept {} + virtual bool try_incref_pyobject() const noexcept { + return false; + } + + uint32_t refcount(std::memory_order order = std::memory_order_relaxed) const { + return detail::refcount(combined_refcount_.load(order)); + } + + uint32_t weakcount( + std::memory_order order = std::memory_order_relaxed) const { + return detail::weakcount(combined_refcount_.load(order)); + } +}; + +namespace detail { + +#ifndef C10_MOBILE +template <> +struct TargetTraits { + // A generic intrusive_ptr may actually be a TensorImpl + // or StorageImpl, so we have to allow for PyObject support. + static constexpr bool can_have_pyobject = true; +}; +#endif + +} // namespace detail + +template +class weak_intrusive_ptr; + +template < + class TTarget, + class NullType = detail::intrusive_target_default_null_type> +class intrusive_ptr final { + private: +// the following static assert would be nice to have but it requires +// the target class T to be fully defined when intrusive_ptr is instantiated +// this is a problem for classes that contain pointers to themselves +// static_assert( +// std::is_base_of_v, +// "intrusive_ptr can only be used for classes that inherit from +// intrusive_ptr_target."); +#ifndef _WIN32 + // This static_assert triggers on MSVC + // error C2131: expression did not evaluate to a constant + static_assert( + // NOLINTNEXTLINE(misc-redundant-expression) + NullType::singleton() == NullType::singleton(), + "NullType must have a constexpr singleton() method"); +#endif + static_assert( + std::is_base_of_v< + TTarget, + std::remove_pointer_t>, + "NullType::singleton() must return a element_type* pointer"); + + TTarget* target_; + + template + friend struct ExclusivelyOwnedTensorTraits; + template + friend class intrusive_ptr; + friend class weak_intrusive_ptr; + + // Make pybind11::class_ be a friend class of intrusive_ptr, so that custom + // smart holder in pybind11 could access the private constructor of + // intrusive_ptr(T*) which took the ownership of the object. This is required + // by customer holder macro PYBIND11_DECLARE_HOLDER_TYPE, where it uses + // intrusive_ptr(TTarget*) to initialize and take ownership of the object. For + // details, see + // https://pybind11.readthedocs.io/en/stable/advanced/smart_ptrs.html#custom-smart-pointers + template + friend class pybind11::class_; + + void retain_() noexcept { + if (target_ != NullType::singleton()) { + uint64_t combined = detail::atomic_combined_refcount_increment( + target_->combined_refcount_, detail::kReferenceCountOne); + uint32_t new_refcount = detail::refcount(combined); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + new_refcount != 1, + "intrusive_ptr: Cannot increase refcount after it reached zero."); + + if constexpr (detail::TargetTraits::can_have_pyobject) { + // If the refcount transitioned from 1 to 2, we need to incref the + // PyObject. In other words, we need to ensure that the PyObject stays + // alive now that we have a C++ reference to this object in addition to + // the PyObject itself. + if (detail::has_pyobject(combined) && detail::refcount(combined) == 2) { + target_->incref_pyobject(); + } + } else { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + !detail::has_pyobject(combined), + "TargetTraits indicates that type cannot have PyObject, but refcount has PyObject bit set."); + } + } + } + + void reset_() noexcept { + if (target_ != NullType::singleton()) { + reset_not_null_(target_); + } + } + + // C10_NOINLINE to keep binary size a bit smaller. We pass TTarget* here + // to avoid an extra pointer dereference in the call from reset_(). + C10_NOINLINE static void reset_not_null_(TTarget* target) noexcept { + if (detail::is_uniquely_owned( + target->combined_refcount_.load(std::memory_order_acquire))) { + // Both counts are 1, so there are no weak references and + // we are releasing the last strong reference. No other + // threads can observe the effects of this target deletion + // call (e.g. calling use_count()) without a data race. + target->combined_refcount_.store(0, std::memory_order_relaxed); + delete target; + return; + } + + auto combined_refcount = detail::atomic_combined_refcount_decrement( + target->combined_refcount_, detail::kReferenceCountOne); + uint32_t new_refcount = detail::refcount(combined_refcount); + bool has_pyobject = detail::has_pyobject(combined_refcount); + if (new_refcount == 0) { + if (detail::weakcount(combined_refcount) == 1) { + delete target; + return; + } + // See comment above about weakcount. As long as refcount>0, + // weakcount is one larger than the actual number of weak references. + // So we need to decrement it here. + release_resources_and_decrement_weakrefs_(target); + } else if constexpr (detail::TargetTraits::can_have_pyobject) { + // If the refcount transitioned from 2 to 1, we need to decref the + // PyObject. In other words, we don't want to keep the PyObject alive if + // there are no C++ references to this object other than the PyObject + // itself. + if (has_pyobject && new_refcount == 1) { + target->decref_pyobject(); + } + } else { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + !has_pyobject, + "TargetTraits indicates that type cannot have PyObject, but refcount has PyObject bit set."); + } + } + + C10_NOINLINE static void release_resources_and_decrement_weakrefs_( + TTarget* target) noexcept { + // justification for const_cast: release_resources is basically a + // destructor and a destructor always mutates the object, even for + // const objects. + const_cast*>(target)->release_resources(); + if (detail::atomic_weakcount_decrement(target->combined_refcount_) == 0) { + delete target; + } + } + + // raw pointer constructors are not public because we shouldn't make + // intrusive_ptr out of raw pointers except from inside the make_intrusive(), + // reclaim() and weak_intrusive_ptr::lock() implementations. + + // This constructor will increase the ref counter for you. + // This constructor will be used by the make_intrusive(), and also pybind11, + // which wrap the intrusive_ptr holder around the raw pointer and incref + // correspondingly (pybind11 requires raw pointer constructor to incref by + // default). + explicit intrusive_ptr(TTarget* target) + : intrusive_ptr(target, raw::DontIncreaseRefcount{}) { + if (target_ != NullType::singleton()) { + // We just created result.target_, so we know no other thread has + // access to it, so we know we needn't care about memory ordering. + // (On x86_64, a store with memory_order_relaxed generates a plain old + // `mov`, whereas an atomic increment does a lock-prefixed `add`, which is + // much more expensive: https://godbolt.org/z/eKPzj8.) + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + target_->combined_refcount_.load(std::memory_order_relaxed) == 0, + "intrusive_ptr: Newly-created target had non-zero refcounts. Does its " + "constructor do something strange like incref or create an " + "intrusive_ptr from `this`?"); + target_->combined_refcount_.store( + detail::kUniqueRef, std::memory_order_relaxed); + } + } + + public: + using element_type = TTarget; + + intrusive_ptr() noexcept + : intrusive_ptr(NullType::singleton(), raw::DontIncreaseRefcount{}) {} + + /* implicit */ intrusive_ptr(std::nullptr_t) noexcept + : intrusive_ptr(NullType::singleton(), raw::DontIncreaseRefcount{}) {} + + // This constructor will not increase the ref counter for you. + // We use the tagged dispatch mechanism to explicitly mark this constructor + // to not increase the refcount + explicit intrusive_ptr( + TTarget* target, + raw::DontIncreaseRefcount /*unused*/) noexcept + : target_(target) {} + + explicit intrusive_ptr(std::unique_ptr rhs) noexcept + : intrusive_ptr(rhs.release()) {} + + intrusive_ptr(intrusive_ptr&& rhs) noexcept : target_(rhs.target_) { + rhs.target_ = NullType::singleton(); + } + + template + // NOLINTNEXTLINE(cppcoreguidelines-rvalue-reference-param-not-moved) + /* implicit */ intrusive_ptr(intrusive_ptr&& rhs) noexcept + : target_( + detail::assign_ptr_(rhs.target_)) { + static_assert( + std::is_convertible_v, + "Type mismatch. intrusive_ptr move constructor got pointer of wrong type."); + rhs.target_ = FromNullType::singleton(); + } + + intrusive_ptr(const intrusive_ptr& rhs) : target_(rhs.target_) { + retain_(); + } + + template + /* implicit */ intrusive_ptr(const intrusive_ptr& rhs) + : target_( + detail::assign_ptr_(rhs.target_)) { + static_assert( + std::is_convertible_v, + "Type mismatch. intrusive_ptr copy constructor got pointer of wrong type."); + retain_(); + } + + ~intrusive_ptr() noexcept { + reset_(); + } + + intrusive_ptr& operator=(intrusive_ptr&& rhs) & noexcept { + // NOLINTNEXTLINE(*assign*) + return this->template operator= (std::move(rhs)); + } + + template + intrusive_ptr& operator=(intrusive_ptr&& rhs) & noexcept { + static_assert( + std::is_convertible_v, + "Type mismatch. intrusive_ptr move assignment got pointer of wrong type."); + intrusive_ptr tmp = std::move(rhs); + swap(tmp); + return *this; + } + + // Assignment is implemented using copy and swap. That's safe for self + // assignment. + // NOLINTNEXTLINE(bugprone-unhandled-self-assignment) + intrusive_ptr& operator=(const intrusive_ptr& rhs) & noexcept { + // NOLINTNEXTLINE(*assign-operator, *assignment-signature) + return this->template operator= (rhs); + } + + template + intrusive_ptr& operator=( + const intrusive_ptr& rhs) & noexcept { + static_assert( + std::is_convertible_v, + "Type mismatch. intrusive_ptr copy assignment got pointer of wrong type."); + intrusive_ptr tmp = rhs; + swap(tmp); + return *this; + } + + TTarget* get() const noexcept { + return target_; + } + + TTarget& operator*() const noexcept { + return *target_; + } + + TTarget* operator->() const noexcept { + return target_; + } + + operator bool() const noexcept { + return target_ != NullType::singleton(); + } + + void reset() noexcept { + reset_(); + target_ = NullType::singleton(); + } + + void swap(intrusive_ptr& rhs) noexcept { + std::swap(target_, rhs.target_); + } + + // We do a lot of null-pointer checks in our code, good to have this be cheap. + bool defined() const noexcept { + return target_ != NullType::singleton(); + } + + uint32_t use_count() const noexcept { + if (target_ == NullType::singleton()) { + return 0; + } + return target_->refcount(std::memory_order_relaxed); + } + + uint32_t weak_use_count() const noexcept { + if (target_ == NullType::singleton()) { + return 0; + } + return target_->weakcount(std::memory_order_relaxed); + } + + bool unique() const noexcept { + return use_count() == 1; + } + + /** + * Stronger than unique() in that it must not have any weakrefs as well. + */ + bool is_uniquely_owned() const noexcept { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(target_ != NullType::singleton()); + return detail::is_uniquely_owned( + target_->combined_refcount_.load(std::memory_order_acquire)); + } + + /** + * Returns an owning (!) pointer to the underlying object and makes the + * intrusive_ptr instance invalid. That means the refcount is not decreased. + * You *must* put the returned pointer back into a intrusive_ptr using + * intrusive_ptr::reclaim(ptr) to properly destruct it. + * This is helpful for C APIs. + */ + TTarget* release() noexcept { + // NOLINTNEXTLINE(clang-analyzer-core.uninitialized.Assign) + TTarget* result = target_; + target_ = NullType::singleton(); + return result; + } + + /** + * Takes an owning pointer to TTarget* and creates an intrusive_ptr that takes + * over ownership. That means the refcount is not increased. + * This is the counter-part to intrusive_ptr::release() and the pointer + * passed in *must* have been created using intrusive_ptr::release(). + */ + static intrusive_ptr reclaim(TTarget* owning_ptr) { + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + owning_ptr == NullType::singleton() || owning_ptr->refcount() == 0 || + owning_ptr->weakcount(), + "TTarget violates the invariant that refcount > 0 => weakcount > 0"); + return intrusive_ptr(owning_ptr, raw::DontIncreaseRefcount{}); + } + + /** + * Takes an owning pointer to TTarget* and creates an intrusive_ptr + * representing a new reference, i.e. the raw pointer retains + * ownership. + */ + static intrusive_ptr reclaim_copy(TTarget* owning_ptr) { + auto ret = reclaim(owning_ptr); + ret.retain_(); + return ret; + } + + /** + * Allocate a heap object with args and wrap it inside a intrusive_ptr and + * incref. This is a helper function to let make_intrusive() access private + * intrusive_ptr constructors. + */ + template + static intrusive_ptr make(Args&&... args) { + return intrusive_ptr(new TTarget(std::forward(args)...)); + } + + /** + * Turn a new instance of TTarget (e.g., literally allocated + * using new TTarget(...) into an intrusive_ptr. If possible, + * use intrusive_ptr::make instead which statically guarantees + * that the allocation was done properly. + * + * At the moment, the only reason this method exists is because + * pybind11 holder types expect to be able to allocate in + * this way (because pybind11 handles the new allocation itself). + */ + static intrusive_ptr unsafe_steal_from_new(TTarget* raw_ptr) { + return intrusive_ptr(raw_ptr); + } + + /** + * Turn an instance of TTarget that should not be reference counted + * (e.g., allocated into an arena with placement new) into an + * intrusive_ptr. This is gratuitously unsafe and should only be + * used if you can guarantee that the pointer will not escape and be + * refcounted as normal. + * + * `expected_decrefs` is a debugging parameter: it indicates the + * number of strong owners the intrusive_ptr_target in question is + * expected to get. In most use cases, this will likely be 1. + * + * The reason this method exists is for manually sharing + * StorageImpls across Tensors in the static runtime. It needs + * access to private intrusive_ptr members so that the refcounts can + * be initialized to custom values. + */ + static intrusive_ptr unsafe_adapt_non_heap_allocated( + TTarget* raw_ptr, + uint32_t expected_decrefs) { + intrusive_ptr result(raw_ptr, raw::DontIncreaseRefcount{}); + // kImpracticallyHugeReferenceCount is impractically huge for a reference + // count, while being in no danger of overflowing uint32_t. We actually only + // need to initialize the refcount to 2 -- we are just doing an unbalanced + // incref to prevent the non-heap-allocated target from being + // freed, and we are optimizing that incref by directly + // initializing the refcounts rather than doing an expensive + // atomic increment. The reason to use kImpracticallyHugeReferenceCount is + // to accommodate the debug assertions in ~intrusive_ptr_target. +#ifdef NDEBUG + expected_decrefs = 0; +#endif + result.target_->combined_refcount_.store( + detail::refcount( + detail::kImpracticallyHugeReferenceCount + expected_decrefs) | + detail::kImpracticallyHugeWeakReferenceCount, + std::memory_order_relaxed); + return result; + } + + /** + * Turn a **non-owning raw pointer** to an intrusive_ptr. It is + * the moral equivalent of enable_shared_from_this on a shared pointer. + * + * This method is only valid for objects that are already live. If + * you are looking for the moral equivalent of unique_ptr(T*) + * constructor, see steal_from_new. + * + * TODO: https://github.com/pytorch/pytorch/issues/56482 + */ + static intrusive_ptr unsafe_reclaim_from_nonowning(TTarget* raw_ptr) { + // See Note [Stack allocated intrusive_ptr_target safety] + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + raw_ptr == NullType::singleton() || raw_ptr->refcount() > 0, + "intrusive_ptr: Can only reclaim pointers that are owned by someone"); + auto ptr = reclaim(raw_ptr); // doesn't increase refcount + ptr.retain_(); + return ptr; + } +}; + +template < + class TTarget, + class NullType = detail::intrusive_target_default_null_type, + class... Args> +inline intrusive_ptr make_intrusive(Args&&... args) { + return intrusive_ptr::make(std::forward(args)...); +} + +template +inline void swap( + intrusive_ptr& lhs, + intrusive_ptr& rhs) noexcept { + lhs.swap(rhs); +} + +// To allow intrusive_ptr inside std::map or std::set, we need operator< +template +inline bool operator<( + const intrusive_ptr& lhs, + const intrusive_ptr& rhs) noexcept { + return lhs.get() < rhs.get(); +} + +template +inline bool operator==( + const intrusive_ptr& lhs, + const intrusive_ptr& rhs) noexcept { + return lhs.get() == rhs.get(); +} + +template +inline bool operator==( + const intrusive_ptr& lhs, + std::nullptr_t) noexcept { + return lhs.get() == nullptr; +} + +template +inline bool operator==( + std::nullptr_t, + const intrusive_ptr& rhs) noexcept { + return nullptr == rhs.get(); +} + +template +inline bool operator!=( + const intrusive_ptr& lhs, + const intrusive_ptr& rhs) noexcept { + return !operator==(lhs, rhs); +} + +template +inline bool operator!=( + const intrusive_ptr& lhs, + std::nullptr_t) noexcept { + return !operator==(lhs, nullptr); +} + +template +inline bool operator!=( + std::nullptr_t, + const intrusive_ptr& rhs) noexcept { + return !operator==(nullptr, rhs); +} +template +struct MaybeOwnedTraits> { + using owned_type = c10::intrusive_ptr; + using borrow_type = c10::intrusive_ptr; + + static borrow_type createBorrow(const owned_type& from) { + return borrow_type::reclaim(from.get()); + } + + static void assignBorrow(borrow_type& lhs, const borrow_type& rhs) { + lhs.release(); + lhs = borrow_type::reclaim(rhs.get()); + } + + static void destroyBorrow(borrow_type& toDestroy) { + toDestroy.release(); + } + + static const owned_type& referenceFromBorrow( + const borrow_type& borrow) noexcept { + return borrow; + } + + static const owned_type* pointerFromBorrow( + const borrow_type& borrow) noexcept { + return &borrow; + } + + static bool debugBorrowIsValid(const borrow_type& /*borrow*/) noexcept { + return true; + } +}; + +template < + typename TTarget, + class NullType = detail::intrusive_target_default_null_type> +class weak_intrusive_ptr final { + private: + static_assert( + std::is_base_of_v, + "intrusive_ptr can only be used for classes that inherit from intrusive_ptr_target."); +#ifndef _WIN32 + // This static_assert triggers on MSVC + // error C2131: expression did not evaluate to a constant + static_assert( + NullType::singleton() == NullType::singleton(), + "NullType must have a constexpr singleton() method"); +#endif + static_assert( + std::is_base_of_v< + TTarget, + std::remove_pointer_t>, + "NullType::singleton() must return a element_type* pointer"); + + TTarget* target_; + + template + friend class weak_intrusive_ptr; + + void retain_() { + if (target_ != NullType::singleton()) { + uint32_t new_weakcount = + detail::atomic_weakcount_increment(target_->combined_refcount_); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + new_weakcount != 1, + "weak_intrusive_ptr: Cannot increase weakcount after it reached zero."); + } + } + + void reset_() noexcept { + if (target_ != NullType::singleton() && + detail::atomic_weakcount_decrement(target_->combined_refcount_) == 0) { + // NOLINTNEXTLINE(clang-analyzer-cplusplus.NewDelete) + delete target_; + } + target_ = NullType::singleton(); + } + + constexpr explicit weak_intrusive_ptr(TTarget* target) : target_(target) {} + + public: + using element_type = TTarget; + + explicit weak_intrusive_ptr(const intrusive_ptr& ptr) + : weak_intrusive_ptr(ptr.get()) { + retain_(); + } + + weak_intrusive_ptr(weak_intrusive_ptr&& rhs) noexcept : target_(rhs.target_) { + rhs.target_ = NullType::singleton(); + } + + template + /* implicit */ weak_intrusive_ptr( + // NOLINTNEXTLINE(cppcoreguidelines-rvalue-reference-param-not-moved) + weak_intrusive_ptr&& rhs) noexcept + : target_( + detail::assign_ptr_(rhs.target_)) { + static_assert( + std::is_convertible_v, + "Type mismatch. weak_intrusive_ptr move constructor got pointer of wrong type."); + rhs.target_ = FromNullType::singleton(); + } + + weak_intrusive_ptr(const weak_intrusive_ptr& rhs) : target_(rhs.target_) { + retain_(); + } + + template + /* implicit */ weak_intrusive_ptr( + const weak_intrusive_ptr& rhs) + : target_( + detail::assign_ptr_(rhs.target_)) { + static_assert( + std::is_convertible_v, + "Type mismatch. weak_intrusive_ptr copy constructor got pointer of wrong type."); + retain_(); + } + + ~weak_intrusive_ptr() noexcept { + reset_(); + } + + weak_intrusive_ptr& operator=(weak_intrusive_ptr&& rhs) & noexcept { + // NOLINTNEXTLINE(*assign*) + return this->template operator= (std::move(rhs)); + } + + template + weak_intrusive_ptr& operator=( + weak_intrusive_ptr&& rhs) & noexcept { + static_assert( + std::is_convertible_v, + "Type mismatch. weak_intrusive_ptr move assignment got pointer of wrong type."); + weak_intrusive_ptr tmp = std::move(rhs); + swap(tmp); + return *this; + } + + weak_intrusive_ptr& operator=(const weak_intrusive_ptr& rhs) & noexcept { + if (this == &rhs) { + return *this; + } + // NOLINTNEXTLINE(*assign*) + return this->template operator= (rhs); + } + + weak_intrusive_ptr& operator=( + const intrusive_ptr& rhs) & noexcept { + weak_intrusive_ptr tmp(rhs); + swap(tmp); + return *this; + } + + template + weak_intrusive_ptr& operator=( + const weak_intrusive_ptr& rhs) & noexcept { + static_assert( + std::is_convertible_v, + "Type mismatch. weak_intrusive_ptr copy assignment got pointer of wrong type."); + weak_intrusive_ptr tmp = rhs; + swap(tmp); + return *this; + } + + void reset() noexcept { + reset_(); + } + + void swap(weak_intrusive_ptr& rhs) noexcept { + TTarget* tmp = target_; + target_ = rhs.target_; + rhs.target_ = tmp; + } + + // NB: This should ONLY be used by the std::hash implementation + // for weak_intrusive_ptr. Another way you could do this is + // friend std::hash, but this triggers two + // bugs: + // + // (1) It triggers an nvcc bug, where std::hash in a friend class + // declaration gets preprocessed into hash, which then cannot + // actually be found. The error in this case looks like: + // + // error: no template named 'hash'; did you mean 'std::hash'? + // + // (2) On OS X, std::hash is declared as a struct, not a class. + // This twings: + // + // error: class 'hash' was previously declared as a struct + // [-Werror,-Wmismatched-tags] + // + // Both of these are work-aroundable, but on the whole, I decided + // it would be simpler and easier to make work if we just expose + // an unsafe getter for target_ + // + TTarget* _unsafe_get_target() const noexcept { + return target_; + } + + uint32_t use_count() const noexcept { + if (target_ == NullType::singleton()) { + return 0; + } + return target_->refcount( + std::memory_order_relaxed); // refcount, not weakcount! + } + + uint32_t weak_use_count() const noexcept { + if (target_ == NullType::singleton()) { + return 0; + } + return target_->weakcount(std::memory_order_relaxed); + } + + bool expired() const noexcept { + return use_count() == 0; + } + + intrusive_ptr lock() const noexcept { + if (target_ == NullType::singleton()) { + return intrusive_ptr(); + } else { + bool increfed = false; + auto combined_refcount = + target_->combined_refcount_.load(std::memory_order_relaxed); + do { + if (detail::refcount(combined_refcount) == 0) { + // Object already destructed, no strong references left anymore. + // Return nullptr. + return intrusive_ptr(); + } + if constexpr (detail::TargetTraits::can_have_pyobject) { + if (detail::has_pyobject(combined_refcount) && + detail::refcount(combined_refcount) == 1 && !increfed) { + // Object has a python wrapper with no other C++ references. + // We need to to incref the Python object before we acquire a + // strong reference to the C++ object to avoid a situation + // where the Python object is deallocated concurrently. + if (!target_->try_incref_pyobject()) { + return intrusive_ptr(); + } + increfed = true; + } + } + } while (!target_->combined_refcount_.compare_exchange_weak( + combined_refcount, + combined_refcount + detail::kReferenceCountOne, + std::memory_order_acquire, + std::memory_order_relaxed)); + + if constexpr (detail::TargetTraits::can_have_pyobject) { + if (increfed && detail::refcount(combined_refcount) != 1) { + target_->decref_pyobject(); + } + } + + return intrusive_ptr( + target_, raw::DontIncreaseRefcount{}); + } + } + + /** + * Returns an owning (but still only weakly referenced) pointer to the + * underlying object and makes the weak_intrusive_ptr instance invalid. + * That means the weakcount is not decreased. + * You *must* put the returned pointer back into a weak_intrusive_ptr using + * weak_intrusive_ptr::reclaim(ptr) to properly destruct it. + * This is helpful for C APIs. + */ + TTarget* release() noexcept { + TTarget* result = target_; + target_ = NullType::singleton(); + return result; + } + + /** + * Takes an owning (but must be weakly referenced) pointer to TTarget* and + * creates a weak_intrusive_ptr that takes over ownership. + * This means that the weakcount is not increased. + * This is the counter-part to weak_intrusive_ptr::release() and the pointer + * passed in *must* have been created using weak_intrusive_ptr::release(). + */ + static weak_intrusive_ptr reclaim(TTarget* owning_weak_ptr) { + // See Note [Stack allocated intrusive_ptr_target safety] + // if refcount > 0, weakcount must be >1 for weak references to exist. + // see weak counting explanation at top of this file. + // if refcount == 0, weakcount only must be >0. + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + owning_weak_ptr == NullType::singleton() || + owning_weak_ptr->weakcount() > 1 || + (owning_weak_ptr->refcount() == 0 && + owning_weak_ptr->weakcount() > 0), + "weak_intrusive_ptr: Can only weak_intrusive_ptr::reclaim() owning pointers that were created using weak_intrusive_ptr::release()."); + return weak_intrusive_ptr(owning_weak_ptr); + } + + /** + * Takes a pointer to TTarget* (may be weak or strong) and creates a + * new weak_intrusive_ptr representing a new weak reference, i.e. + * the raw pointer retains ownership. + */ + static weak_intrusive_ptr reclaim_copy(TTarget* owning_ptr) { + auto ret = reclaim(owning_ptr); + ret.retain_(); + return ret; + } + + template + friend bool operator<( + const weak_intrusive_ptr& lhs, + const weak_intrusive_ptr& rhs) noexcept; + template + friend bool operator==( + const weak_intrusive_ptr& lhs, + const weak_intrusive_ptr& rhs) noexcept; +}; + +template +inline void swap( + weak_intrusive_ptr& lhs, + weak_intrusive_ptr& rhs) noexcept { + lhs.swap(rhs); +} + +// To allow weak_intrusive_ptr inside std::map or std::set, we need operator< +template +inline bool operator<( + const weak_intrusive_ptr& lhs, + const weak_intrusive_ptr& rhs) noexcept { + return lhs.target_ < rhs.target_; +} + +template +inline bool operator==( + const weak_intrusive_ptr& lhs, + const weak_intrusive_ptr& rhs) noexcept { + return lhs.target_ == rhs.target_; +} + +template +inline bool operator!=( + const weak_intrusive_ptr& lhs, + const weak_intrusive_ptr& rhs) noexcept { + return !operator==(lhs, rhs); +} + +// Alias for documentary purposes, to more easily distinguish +// weak raw intrusive pointers from intrusive pointers. +using weak_intrusive_ptr_target = intrusive_ptr_target; + +// This namespace provides some methods for working with +// raw pointers that subclass intrusive_ptr_target. They are not provided +// as methods on intrusive_ptr_target, because ideally you would not need these +// methods at all (use smart pointers), but if you are dealing with legacy code +// that still needs to pass around raw pointers, you may find these quite +// useful. +// +// An important usage note: some functions are only valid if you have a +// strong raw pointer to the object, while others are only valid if you +// have a weak raw pointer to the object. ONLY call intrusive_ptr namespace +// functions on strong pointers, and weak_intrusive_ptr namespace functions +// on weak pointers. If you mix it up, you may get an assert failure. +namespace raw { + +namespace intrusive_ptr { + +// WARNING: Unlike the reclaim() API, it is NOT valid to pass +// NullType::singleton to this function +inline void incref(intrusive_ptr_target* self) { + if (self) { + uint64_t combined = detail::atomic_combined_refcount_increment( + self->combined_refcount_, detail::kReferenceCountOne); + +#ifndef C10_MOBILE + if (detail::has_pyobject(combined) && detail::refcount(combined) == 2) { + self->incref_pyobject(); + } +#else + TORCH_INTERNAL_ASSERT_DEBUG_ONLY(!detail::has_pyobject(combined)); +#endif + } +} + +// WARNING: Unlike the reclaim() API, it is NOT valid to pass +// NullType::singleton to this function +inline void decref(intrusive_ptr_target* self) { + // Let it die + c10::intrusive_ptr::reclaim(self); + // NB: Caller still has 'self' pointer, but it's now invalid. + // If you want more safety, used the actual c10::intrusive_ptr class +} + +template +inline T* make_weak(T* self) { + // NB: 'this' is a strong pointer, but we return a weak pointer + auto ptr = c10::intrusive_ptr::reclaim(self); + c10::weak_intrusive_ptr wptr(ptr); + ptr.release(); + return wptr.release(); +} + +inline uint32_t use_count(intrusive_ptr_target* self) { + auto ptr = c10::intrusive_ptr::reclaim(self); + auto r = ptr.use_count(); + ptr.release(); + return r; +} + +} // namespace intrusive_ptr + +namespace weak_intrusive_ptr { + +inline void incref(weak_intrusive_ptr_target* self) { + detail::atomic_weakcount_increment(self->combined_refcount_); +} + +inline void decref(weak_intrusive_ptr_target* self) { + // Let it die + c10::weak_intrusive_ptr::reclaim(self); + // NB: You still "have" the 'self' pointer, but it's now invalid. + // If you want more safety, used the actual c10::weak_intrusive_ptr class +} + +template +inline T* lock(T* self) { + auto wptr = c10::weak_intrusive_ptr::reclaim(self); + auto ptr = wptr.lock(); + wptr.release(); + return ptr.release(); +} + +// This gives the STRONG refcount of a WEAK pointer +inline uint32_t use_count(weak_intrusive_ptr_target* self) { + auto wptr = c10::weak_intrusive_ptr::reclaim(self); + auto r = wptr.use_count(); + wptr.release(); + return r; +} + +} // namespace weak_intrusive_ptr + +} // namespace raw + +} // namespace c10 + +namespace std { +// To allow intrusive_ptr and weak_intrusive_ptr inside std::unordered_map or +// std::unordered_set, we need std::hash +template +struct hash> { + size_t operator()(const c10::intrusive_ptr& x) const { + return std::hash()(x.get()); + } +}; +template +struct hash> { + size_t operator()(const c10::weak_intrusive_ptr& x) const { + return std::hash()(x._unsafe_get_target()); + } +}; +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/irange.h b/backends/pytorch/include/c10/util/irange.h new file mode 100644 index 000000000..bc2a018db --- /dev/null +++ b/backends/pytorch/include/c10/util/irange.h @@ -0,0 +1,128 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +// Copyright 2004-present Facebook. All Rights Reserved. + +#pragma once + +#include + +#include +#include +#include +#include + +namespace c10 { + +namespace detail { + +template < + typename I, + bool one_sided = false, + std::enable_if_t, int> = 0> +struct integer_iterator { + using iterator_category = std::input_iterator_tag; + using value_type = I; + using difference_type = std::ptrdiff_t; + using pointer = I*; + using reference = I&; + + explicit constexpr integer_iterator(I val) : value(val) {} + + constexpr I operator*() const { + return value; + } + + constexpr I const* operator->() const { + return &value; + } + + constexpr integer_iterator& operator++() { + ++value; + return *this; + } + + constexpr integer_iterator operator++(int) { + const auto copy = *this; + ++*this; + return copy; + } + + constexpr bool operator==(const integer_iterator& other) const { + if constexpr (one_sided) { + // Range-for loops' end test is `begin != end`, not `begin < + // end`. To handle `c10::irange(n)` where n < 0 (which should be + // empty), we just make `begin != end` fail whenever `end` is + // negative. + return is_negative(other.value) || value == other.value; + } else { + return value == other.value; + } + // Suppress "warning: missing return statement at end of non-void function" + // which Nvidia's Robert Crovella confirms is an NVCC compiler error + // here https://stackoverflow.com/a/64561686/752843 on 2020-10-27 + // `__builtin_unreachable();` would be best here, but it's not + // available with all compilers. So we instead return an arbitrary + // value trusting that this line will, in fact, never be reached. + return false; // Horrible hack + } + + constexpr bool operator!=(const integer_iterator& other) const { + return !(*this == other); + } + + protected: + I value; +}; + +} // namespace detail + +template < + typename I, + bool one_sided = false, + std::enable_if_t, bool> = true> +struct integer_range { + public: + constexpr integer_range(I begin, I end) : begin_(begin), end_(end) {} + using iterator = detail::integer_iterator; + constexpr iterator begin() const { + return begin_; + } + constexpr iterator end() const { + return end_; + } + + private: + iterator begin_; + iterator end_; +}; + +/// Creates an integer range for the half-open interval [begin, end) +/// If end<=begin, then the range is empty. +/// The range has the type of the `end` integer; `begin` integer is +/// cast to this type. +template < + typename Integer1, + typename Integer2, + std::enable_if_t, bool> = true, + std::enable_if_t, bool> = true> +constexpr integer_range irange(Integer1 begin, Integer2 end) { + // If end<=begin then the range is empty; we can achieve this effect by + // choosing the larger of {begin, end} as the loop terminator + return { + static_cast(begin), + std::max(static_cast(begin), end)}; +} + +/// Creates an integer range for the half-open interval [0, end) +/// If end<=begin, then the range is empty +template < + typename Integer, + std::enable_if_t, bool> = true> +constexpr integer_range irange(Integer end) { + return {Integer(), end}; +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/llvmMathExtras.h b/backends/pytorch/include/c10/util/llvmMathExtras.h new file mode 100644 index 000000000..6884e20d1 --- /dev/null +++ b/backends/pytorch/include/c10/util/llvmMathExtras.h @@ -0,0 +1,910 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +//===-- llvm/Support/MathExtras.h - Useful math functions -------*- C++ -*-===// +// +// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. +// See https://llvm.org/LICENSE.txt for license information. +// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception +// +//===----------------------------------------------------------------------===// +// +// This file contains some functions that are useful for math stuff. +// +//===----------------------------------------------------------------------===// + +#pragma once + +#include + +#include +#include +#include +#include +#include +#include +#include +#include + +#ifdef __ANDROID_NDK__ +#include +#endif + +#ifndef __has_builtin +#define __has_builtin(x) 0 +#endif + +#ifndef LLVM_GNUC_PREREQ +#if defined(__GNUC__) && defined(__GNUC_MINOR__) && defined(__GNUC_PATCHLEVEL__) +#define LLVM_GNUC_PREREQ(maj, min, patch) \ + ((__GNUC__ << 20) + (__GNUC_MINOR__ << 10) + __GNUC_PATCHLEVEL__ >= \ + ((maj) << 20) + ((min) << 10) + (patch)) +#elif defined(__GNUC__) && defined(__GNUC_MINOR__) +#define LLVM_GNUC_PREREQ(maj, min, patch) \ + ((__GNUC__ << 20) + (__GNUC_MINOR__ << 10) >= ((maj) << 20) + ((min) << 10)) +#else +#define LLVM_GNUC_PREREQ(maj, min, patch) 0 +#endif +#endif + +#ifdef _MSC_VER +// Declare these intrinsics manually rather including intrin.h. It's very +// expensive, and MathExtras.h is popular. +// #include +extern "C" { +unsigned char _BitScanForward(unsigned long* _Index, unsigned long _Mask); +unsigned char _BitScanForward64(unsigned long* _Index, unsigned __int64 _Mask); +unsigned char _BitScanReverse(unsigned long* _Index, unsigned long _Mask); +unsigned char _BitScanReverse64(unsigned long* _Index, unsigned __int64 _Mask); +} +#endif + +namespace c10::llvm { +/// The behavior an operation has on an input of 0. +enum ZeroBehavior { + /// The returned value is undefined. + ZB_Undefined, + /// The returned value is numeric_limits::max() + ZB_Max, + /// The returned value is numeric_limits::digits + ZB_Width +}; + +namespace detail { +template +struct TrailingZerosCounter { + static std::size_t count(T Val, ZeroBehavior /*unused*/) { + if (!Val) + return std::numeric_limits::digits; + if (Val & 0x1) + return 0; + + // Bisection method. + std::size_t ZeroBits = 0; + T Shift = std::numeric_limits::digits >> 1; + T Mask = std::numeric_limits::max() >> Shift; + while (Shift) { + if ((Val & Mask) == 0) { + Val >>= Shift; + ZeroBits |= Shift; + } + Shift >>= 1; + Mask >>= Shift; + } + return ZeroBits; + } +}; + +#if (defined(__GNUC__) && __GNUC__ >= 4) || defined(_MSC_VER) +template +struct TrailingZerosCounter { + static std::size_t count(T Val, ZeroBehavior ZB) { + if (ZB != ZB_Undefined && Val == 0) + return 32; + +#if __has_builtin(__builtin_ctz) || LLVM_GNUC_PREREQ(4, 0, 0) + return __builtin_ctz(Val); +#elif defined(_MSC_VER) + unsigned long Index; + _BitScanForward(&Index, Val); + return Index; +#endif + } +}; + +#if !defined(_MSC_VER) || defined(_M_X64) +template +struct TrailingZerosCounter { + static std::size_t count(T Val, ZeroBehavior ZB) { + if (ZB != ZB_Undefined && Val == 0) + return 64; + +#if __has_builtin(__builtin_ctzll) || LLVM_GNUC_PREREQ(4, 0, 0) + return __builtin_ctzll(Val); +#elif defined(_MSC_VER) + unsigned long Index; + _BitScanForward64(&Index, Val); + return Index; +#endif + } +}; +#endif +#endif +} // namespace detail + +/// Count number of 0's from the least significant bit to the most +/// stopping at the first 1. +/// +/// Only unsigned integral types are allowed. +/// +/// \param ZB the behavior on an input of 0. Only ZB_Width and ZB_Undefined are +/// valid arguments. +template +std::size_t countTrailingZeros(T Val, ZeroBehavior ZB = ZB_Width) { + static_assert( + std::numeric_limits::is_integer && !std::numeric_limits::is_signed, + "Only unsigned integral types are allowed."); + return llvm::detail::TrailingZerosCounter::count(Val, ZB); +} + +namespace detail { +template +struct LeadingZerosCounter { + static std::size_t count(T Val, ZeroBehavior /*unused*/) { + if (!Val) + return std::numeric_limits::digits; + + // Bisection method. + std::size_t ZeroBits = 0; + for (T Shift = std::numeric_limits::digits >> 1; Shift; Shift >>= 1) { + T Tmp = Val >> Shift; + if (Tmp) + Val = Tmp; + else + ZeroBits |= Shift; + } + return ZeroBits; + } +}; + +#if (defined(__GNUC__) && __GNUC__ >= 4) || defined(_MSC_VER) +template +struct LeadingZerosCounter { + static std::size_t count(T Val, ZeroBehavior ZB) { + if (ZB != ZB_Undefined && Val == 0) + return 32; + +#if __has_builtin(__builtin_clz) || LLVM_GNUC_PREREQ(4, 0, 0) + return __builtin_clz(Val); +#elif defined(_MSC_VER) + unsigned long Index; + _BitScanReverse(&Index, Val); + return Index ^ 31; +#endif + } +}; + +#if !defined(_MSC_VER) || defined(_M_X64) +template +struct LeadingZerosCounter { + static std::size_t count(T Val, ZeroBehavior ZB) { + if (ZB != ZB_Undefined && Val == 0) + return 64; + +#if __has_builtin(__builtin_clzll) || LLVM_GNUC_PREREQ(4, 0, 0) + return __builtin_clzll(Val); +#elif defined(_MSC_VER) + unsigned long Index; + _BitScanReverse64(&Index, Val); + return Index ^ 63; +#endif + } +}; +#endif +#endif +} // namespace detail + +/// Count number of 0's from the most significant bit to the least +/// stopping at the first 1. +/// +/// Only unsigned integral types are allowed. +/// +/// \param ZB the behavior on an input of 0. Only ZB_Width and ZB_Undefined are +/// valid arguments. +template +std::size_t countLeadingZeros(T Val, ZeroBehavior ZB = ZB_Width) { + static_assert( + std::numeric_limits::is_integer && !std::numeric_limits::is_signed, + "Only unsigned integral types are allowed."); + return llvm::detail::LeadingZerosCounter::count(Val, ZB); +} + +/// Get the index of the first set bit starting from the least +/// significant bit. +/// +/// Only unsigned integral types are allowed. +/// +/// \param ZB the behavior on an input of 0. Only ZB_Max and ZB_Undefined are +/// valid arguments. +template +T findFirstSet(T Val, ZeroBehavior ZB = ZB_Max) { + if (ZB == ZB_Max && Val == 0) + return std::numeric_limits::max(); + + return countTrailingZeros(Val, ZB_Undefined); +} + +/// Create a bitmask with the N right-most bits set to 1, and all other +/// bits set to 0. Only unsigned types are allowed. +template +T maskTrailingOnes(unsigned N) { + static_assert(std::is_unsigned_v, "Invalid type!"); + const unsigned Bits = CHAR_BIT * sizeof(T); + assert(N <= Bits && "Invalid bit index"); + return N == 0 ? 0 : (T(-1) >> (Bits - N)); +} + +/// Create a bitmask with the N left-most bits set to 1, and all other +/// bits set to 0. Only unsigned types are allowed. +template +T maskLeadingOnes(unsigned N) { + return ~maskTrailingOnes(CHAR_BIT * sizeof(T) - N); +} + +/// Create a bitmask with the N right-most bits set to 0, and all other +/// bits set to 1. Only unsigned types are allowed. +template +T maskTrailingZeros(unsigned N) { + return maskLeadingOnes(CHAR_BIT * sizeof(T) - N); +} + +/// Create a bitmask with the N left-most bits set to 0, and all other +/// bits set to 1. Only unsigned types are allowed. +template +T maskLeadingZeros(unsigned N) { + return maskTrailingOnes(CHAR_BIT * sizeof(T) - N); +} + +/// Get the index of the last set bit starting from the least +/// significant bit. +/// +/// Only unsigned integral types are allowed. +/// +/// \param ZB the behavior on an input of 0. Only ZB_Max and ZB_Undefined are +/// valid arguments. +template +T findLastSet(T Val, ZeroBehavior ZB = ZB_Max) { + if (ZB == ZB_Max && Val == 0) + return std::numeric_limits::max(); + + // Use ^ instead of - because both gcc and llvm can remove the associated ^ + // in the __builtin_clz intrinsic on x86. + return countLeadingZeros(Val, ZB_Undefined) ^ + (std::numeric_limits::digits - 1); +} + +/// Macro compressed bit reversal table for 256 bits. +/// +/// http://graphics.stanford.edu/~seander/bithacks.html#BitReverseTable +/// NOLINTNEXTLINE(*c-arrays*) +static constexpr unsigned char BitReverseTable256[256] = { +#define R2(n) n, n + 2 * 64, n + 1 * 64, n + 3 * 64 +#define R4(n) R2(n), R2(n + 2 * 16), R2(n + 1 * 16), R2(n + 3 * 16) +#define R6(n) R4(n), R4(n + 2 * 4), R4(n + 1 * 4), R4(n + 3 * 4) + R6(0), + R6(2), + R6(1), + R6(3) +#undef R2 +#undef R4 +#undef R6 +}; + +/// Reverse the bits in \p Val. +template +T reverseBits(T Val) { + // NOLINTNEXTLINE(*c-arrays*) + unsigned char in[sizeof(Val)]; + // NOLINTNEXTLINE(*c-arrays*) + unsigned char out[sizeof(Val)]; + std::memcpy(in, &Val, sizeof(Val)); + for (unsigned i = 0; i < sizeof(Val); ++i) + out[(sizeof(Val) - i) - 1] = BitReverseTable256[in[i]]; + std::memcpy(&Val, out, sizeof(Val)); + return Val; +} + +// NOTE: The following support functions use the _32/_64 extensions instead of +// type overloading so that signed and unsigned integers can be used without +// ambiguity. + +/// Return the high 32 bits of a 64 bit value. +constexpr inline uint32_t Hi_32(uint64_t Value) { + return static_cast(Value >> 32); +} + +/// Return the low 32 bits of a 64 bit value. +constexpr inline uint32_t Lo_32(uint64_t Value) { + return static_cast(Value); +} + +/// Make a 64-bit integer from a high / low pair of 32-bit integers. +constexpr inline uint64_t Make_64(uint32_t High, uint32_t Low) { + return ((uint64_t)High << 32) | (uint64_t)Low; +} + +/// Checks if an integer fits into the given bit width. +template +constexpr inline bool isInt(int64_t x) { + return N >= 64 || + (-(INT64_C(1) << (N - 1)) <= x && x < (INT64_C(1) << (N - 1))); +} +// Template specializations to get better code for common cases. +template <> +constexpr inline bool isInt<8>(int64_t x) { + return static_cast(x) == x; +} +template <> +constexpr inline bool isInt<16>(int64_t x) { + return static_cast(x) == x; +} +template <> +constexpr inline bool isInt<32>(int64_t x) { + return static_cast(x) == x; +} + +/// Checks if a signed integer is an N bit number shifted left by S. +template +constexpr inline bool isShiftedInt(int64_t x) { + static_assert( + N > 0, "isShiftedInt<0> doesn't make sense (refers to a 0-bit number."); + static_assert(N + S <= 64, "isShiftedInt with N + S > 64 is too wide."); + return isInt(x) && (x % (UINT64_C(1) << S) == 0); +} + +/// Checks if an unsigned integer fits into the given bit width. +/// +/// This is written as two functions rather than as simply +/// +/// return N >= 64 || X < (UINT64_C(1) << N); +/// +/// to keep MSVC from (incorrectly) warning on isUInt<64> that we're shifting +/// left too many places. +template +constexpr inline std::enable_if_t<(N < 64), bool> isUInt(uint64_t X) { + static_assert(N > 0, "isUInt<0> doesn't make sense"); + return X < (UINT64_C(1) << N); +} +template +constexpr inline std::enable_if_t= 64, bool> isUInt(uint64_t /*X*/) { + return true; +} + +// Template specializations to get better code for common cases. +template <> +constexpr inline bool isUInt<8>(uint64_t x) { + return static_cast(x) == x; +} +template <> +constexpr inline bool isUInt<16>(uint64_t x) { + return static_cast(x) == x; +} +template <> +constexpr inline bool isUInt<32>(uint64_t x) { + return static_cast(x) == x; +} + +/// Checks if a unsigned integer is an N bit number shifted left by S. +template +constexpr inline bool isShiftedUInt(uint64_t x) { + static_assert( + N > 0, "isShiftedUInt<0> doesn't make sense (refers to a 0-bit number)"); + static_assert( + N + S <= 64, "isShiftedUInt with N + S > 64 is too wide."); + // Per the two static_asserts above, S must be strictly less than 64. So + // 1 << S is not undefined behavior. + return isUInt(x) && (x % (UINT64_C(1) << S) == 0); +} + +/// Gets the maximum value for a N-bit unsigned integer. +inline uint64_t maxUIntN(uint64_t N) { + assert(N > 0 && N <= 64 && "integer width out of range"); + + // uint64_t(1) << 64 is undefined behavior, so we can't do + // (uint64_t(1) << N) - 1 + // without checking first that N != 64. But this works and doesn't have a + // branch. + return UINT64_MAX >> (64 - N); +} + +// Ignore the false warning "Arithmetic overflow" for MSVC +#ifdef _MSC_VER +#pragma warning(push) +#pragma warning(disable : 4146) +#endif + +/// Gets the minimum value for a N-bit signed integer. +inline int64_t minIntN(int64_t N) { + assert(N > 0 && N <= 64 && "integer width out of range"); + // NOLINTNEXTLINE(*-narrowing-conversions) + return -(UINT64_C(1) << (N - 1)); +} + +#ifdef _MSC_VER +#pragma warning(pop) +#endif + +/// Gets the maximum value for a N-bit signed integer. +inline int64_t maxIntN(int64_t N) { + assert(N > 0 && N <= 64 && "integer width out of range"); + + // This relies on two's complement wraparound when N == 64, so we convert to + // int64_t only at the very end to avoid UB. + // NOLINTNEXTLINE(*-narrowing-conversions) + return (UINT64_C(1) << (N - 1)) - 1; +} + +/// Checks if an unsigned integer fits into the given (dynamic) bit width. +inline bool isUIntN(unsigned N, uint64_t x) { + return N >= 64 || x <= maxUIntN(N); +} + +/// Checks if an signed integer fits into the given (dynamic) bit width. +inline bool isIntN(unsigned N, int64_t x) { + return N >= 64 || (minIntN(N) <= x && x <= maxIntN(N)); +} + +/// Return true if the argument is a non-empty sequence of ones starting at the +/// least significant bit with the remainder zero (32 bit version). +/// Ex. isMask_32(0x0000FFFFU) == true. +constexpr inline bool isMask_32(uint32_t Value) { + return Value && ((Value + 1) & Value) == 0; +} + +/// Return true if the argument is a non-empty sequence of ones starting at the +/// least significant bit with the remainder zero (64 bit version). +constexpr inline bool isMask_64(uint64_t Value) { + return Value && ((Value + 1) & Value) == 0; +} + +/// Return true if the argument contains a non-empty sequence of ones with the +/// remainder zero (32 bit version.) Ex. isShiftedMask_32(0x0000FF00U) == true. +constexpr inline bool isShiftedMask_32(uint32_t Value) { + return Value && isMask_32((Value - 1) | Value); +} + +/// Return true if the argument contains a non-empty sequence of ones with the +/// remainder zero (64 bit version.) +constexpr inline bool isShiftedMask_64(uint64_t Value) { + return Value && isMask_64((Value - 1) | Value); +} + +/// Return true if the argument is a power of two > 0. +/// Ex. isPowerOf2_32(0x00100000U) == true (32 bit edition.) +constexpr inline bool isPowerOf2_32(uint32_t Value) { + return Value && !(Value & (Value - 1)); +} + +/// Return true if the argument is a power of two > 0 (64 bit edition.) +constexpr inline bool isPowerOf2_64(uint64_t Value) { + return Value && !(Value & (Value - 1)); +} + +/// Count the number of ones from the most significant bit to the first +/// zero bit. +/// +/// Ex. countLeadingOnes(0xFF0FFF00) == 8. +/// Only unsigned integral types are allowed. +/// +/// \param ZB the behavior on an input of all ones. Only ZB_Width and +/// ZB_Undefined are valid arguments. +template +std::size_t countLeadingOnes(T Value, ZeroBehavior ZB = ZB_Width) { + static_assert( + std::numeric_limits::is_integer && !std::numeric_limits::is_signed, + "Only unsigned integral types are allowed."); + return countLeadingZeros(~Value, ZB); +} + +/// Count the number of ones from the least significant bit to the first +/// zero bit. +/// +/// Ex. countTrailingOnes(0x00FF00FF) == 8. +/// Only unsigned integral types are allowed. +/// +/// \param ZB the behavior on an input of all ones. Only ZB_Width and +/// ZB_Undefined are valid arguments. +template +std::size_t countTrailingOnes(T Value, ZeroBehavior ZB = ZB_Width) { + static_assert( + std::numeric_limits::is_integer && !std::numeric_limits::is_signed, + "Only unsigned integral types are allowed."); + return countTrailingZeros(~Value, ZB); +} + +namespace detail { +template +struct PopulationCounter { + static unsigned count(T Value) { + // Generic version, forward to 32 bits. + static_assert(SizeOfT <= 4, "Not implemented!"); +#if defined(__GNUC__) && __GNUC__ >= 4 + return __builtin_popcount(Value); +#else + uint32_t v = Value; + v = v - ((v >> 1) & 0x55555555); + v = (v & 0x33333333) + ((v >> 2) & 0x33333333); + return ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24; +#endif + } +}; + +template +struct PopulationCounter { + static unsigned count(T Value) { +#if defined(__GNUC__) && __GNUC__ >= 4 + return __builtin_popcountll(Value); +#else + uint64_t v = Value; + v = v - ((v >> 1) & 0x5555555555555555ULL); + v = (v & 0x3333333333333333ULL) + ((v >> 2) & 0x3333333333333333ULL); + v = (v + (v >> 4)) & 0x0F0F0F0F0F0F0F0FULL; + return unsigned((uint64_t)(v * 0x0101010101010101ULL) >> 56); +#endif + } +}; +} // namespace detail + +/// Count the number of set bits in a value. +/// Ex. countPopulation(0xF000F000) = 8 +/// Returns 0 if the word is zero. +template +inline unsigned countPopulation(T Value) { + static_assert( + std::numeric_limits::is_integer && !std::numeric_limits::is_signed, + "Only unsigned integral types are allowed."); + return detail::PopulationCounter::count(Value); +} + +/// Return the log base 2 of the specified value. +inline double Log2(double Value) { +#if defined(__ANDROID_API__) && __ANDROID_API__ < 18 + return __builtin_log(Value) / __builtin_log(2.0); +#else + return log2(Value); +#endif +} + +/// Return the floor log base 2 of the specified value, -1 if the value is zero. +/// (32 bit edition.) +/// Ex. Log2_32(32) == 5, Log2_32(1) == 0, Log2_32(0) == -1, Log2_32(6) == 2 +inline unsigned Log2_32(uint32_t Value) { + return static_cast(31 - countLeadingZeros(Value)); +} + +/// Return the floor log base 2 of the specified value, -1 if the value is zero. +/// (64 bit edition.) +inline unsigned Log2_64(uint64_t Value) { + return static_cast(63 - countLeadingZeros(Value)); +} + +/// Return the ceil log base 2 of the specified value, 32 if the value is zero. +/// (32 bit edition). +/// Ex. Log2_32_Ceil(32) == 5, Log2_32_Ceil(1) == 0, Log2_32_Ceil(6) == 3 +inline unsigned Log2_32_Ceil(uint32_t Value) { + return static_cast(32 - countLeadingZeros(Value - 1)); +} + +/// Return the ceil log base 2 of the specified value, 64 if the value is zero. +/// (64 bit edition.) +inline unsigned Log2_64_Ceil(uint64_t Value) { + return static_cast(64 - countLeadingZeros(Value - 1)); +} + +/// Return the greatest common divisor of the values using Euclid's algorithm. +inline uint64_t GreatestCommonDivisor64(uint64_t A, uint64_t B) { + while (B) { + uint64_t T = B; + B = A % B; + A = T; + } + return A; +} + +/// This function takes a 64-bit integer and returns the bit equivalent double. +inline double BitsToDouble(uint64_t Bits) { + double D = 0; + static_assert(sizeof(uint64_t) == sizeof(double), "Unexpected type sizes"); + memcpy(&D, &Bits, sizeof(Bits)); + return D; +} + +/// This function takes a 32-bit integer and returns the bit equivalent float. +inline float BitsToFloat(uint32_t Bits) { + // TODO: Use std::bit_cast once C++20 becomes available. + return c10::bit_cast(Bits); +} + +/// This function takes a double and returns the bit equivalent 64-bit integer. +/// Note that copying doubles around changes the bits of NaNs on some hosts, +/// notably x86, so this routine cannot be used if these bits are needed. +inline uint64_t DoubleToBits(double Double) { + // NOLINTNEXTLINE(cppcoreguidelines-init-variables) + uint64_t Bits; + static_assert(sizeof(uint64_t) == sizeof(double), "Unexpected type sizes"); + memcpy(&Bits, &Double, sizeof(Double)); + return Bits; +} + +/// This function takes a float and returns the bit equivalent 32-bit integer. +/// Note that copying floats around changes the bits of NaNs on some hosts, +/// notably x86, so this routine cannot be used if these bits are needed. +inline uint32_t FloatToBits(float Float) { + // NOLINTNEXTLINE(cppcoreguidelines-init-variables) + uint32_t Bits; + static_assert(sizeof(uint32_t) == sizeof(float), "Unexpected type sizes"); + memcpy(&Bits, &Float, sizeof(Float)); + return Bits; +} + +/// A and B are either alignments or offsets. Return the minimum alignment that +/// may be assumed after adding the two together. +constexpr inline uint64_t MinAlign(uint64_t A, uint64_t B) { + // The largest power of 2 that divides both A and B. + // + // Replace "-Value" by "1+~Value" in the following commented code to avoid + // MSVC warning C4146 + // return (A | B) & -(A | B); + return (A | B) & (1 + ~(A | B)); +} + +/// Aligns \c Addr to \c Alignment bytes, rounding up. +/// +/// Alignment should be a power of two. This method rounds up, so +/// alignAddr(7, 4) == 8 and alignAddr(8, 4) == 8. +inline uintptr_t alignAddr(const void* Addr, size_t Alignment) { + assert( + Alignment && isPowerOf2_64((uint64_t)Alignment) && + "Alignment is not a power of two!"); + + assert((uintptr_t)Addr + Alignment - 1 >= (uintptr_t)Addr); + + return (((uintptr_t)Addr + Alignment - 1) & ~(uintptr_t)(Alignment - 1)); +} + +/// Returns the necessary adjustment for aligning \c Ptr to \c Alignment +/// bytes, rounding up. +inline size_t alignmentAdjustment(const void* Ptr, size_t Alignment) { + return alignAddr(Ptr, Alignment) - (uintptr_t)Ptr; +} + +/// Returns the next power of two (in 64-bits) that is strictly greater than A. +/// Returns zero on overflow. +inline uint64_t NextPowerOf2(uint64_t A) { + A |= (A >> 1); + A |= (A >> 2); + A |= (A >> 4); + A |= (A >> 8); + A |= (A >> 16); + A |= (A >> 32); + return A + 1; +} + +/// Returns the power of two which is less than or equal to the given value. +/// Essentially, it is a floor operation across the domain of powers of two. +inline uint64_t PowerOf2Floor(uint64_t A) { + if (!A) + return 0; + return 1ull << (63 - countLeadingZeros(A, ZB_Undefined)); +} + +/// Returns the power of two which is greater than or equal to the given value. +/// Essentially, it is a ceil operation across the domain of powers of two. +inline uint64_t PowerOf2Ceil(uint64_t A) { + if (!A) + return 0; + return NextPowerOf2(A - 1); +} + +/// Returns the next integer (mod 2**64) that is greater than or equal to +/// \p Value and is a multiple of \p Align. \p Align must be non-zero. +/// +/// If non-zero \p Skew is specified, the return value will be a minimal +/// integer that is greater than or equal to \p Value and equal to +/// \p Align * N + \p Skew for some integer N. If \p Skew is larger than +/// \p Align, its value is adjusted to '\p Skew mod \p Align'. +/// +/// Examples: +/// \code +/// alignTo(5, 8) = 8 +/// alignTo(17, 8) = 24 +/// alignTo(~0LL, 8) = 0 +/// alignTo(321, 255) = 510 +/// +/// alignTo(5, 8, 7) = 7 +/// alignTo(17, 8, 1) = 17 +/// alignTo(~0LL, 8, 3) = 3 +/// alignTo(321, 255, 42) = 552 +/// \endcode +inline uint64_t alignTo(uint64_t Value, uint64_t Align, uint64_t Skew = 0) { + assert(Align != 0u && "Align can't be 0."); + Skew %= Align; + return (Value + Align - 1 - Skew) / Align * Align + Skew; +} + +/// Returns the next integer (mod 2**64) that is greater than or equal to +/// \p Value and is a multiple of \c Align. \c Align must be non-zero. +template +constexpr inline uint64_t alignTo(uint64_t Value) { + static_assert(Align != 0u, "Align must be non-zero"); + return (Value + Align - 1) / Align * Align; +} + +/// Returns the integer ceil(Numerator / Denominator). +inline uint64_t divideCeil(uint64_t Numerator, uint64_t Denominator) { + return alignTo(Numerator, Denominator) / Denominator; +} + +/// \c alignTo for contexts where a constant expression is required. +/// \sa alignTo +/// +/// \todo FIXME: remove when \c constexpr becomes really \c constexpr +template +struct AlignTo { + static_assert(Align != 0u, "Align must be non-zero"); + template + struct from_value { + static const uint64_t value = (Value + Align - 1) / Align * Align; + }; +}; + +/// Returns the largest uint64_t less than or equal to \p Value and is +/// \p Skew mod \p Align. \p Align must be non-zero +inline uint64_t alignDown(uint64_t Value, uint64_t Align, uint64_t Skew = 0) { + assert(Align != 0u && "Align can't be 0."); + Skew %= Align; + return (Value - Skew) / Align * Align + Skew; +} + +/// Returns the offset to the next integer (mod 2**64) that is greater than +/// or equal to \p Value and is a multiple of \p Align. \p Align must be +/// non-zero. +inline uint64_t OffsetToAlignment(uint64_t Value, uint64_t Align) { + return alignTo(Value, Align) - Value; +} + +/// Sign-extend the number in the bottom B bits of X to a 32-bit integer. +/// Requires 0 < B <= 32. +template +constexpr inline int32_t SignExtend32(uint32_t X) { + static_assert(B > 0, "Bit width can't be 0."); + static_assert(B <= 32, "Bit width out of range."); + return int32_t(X << (32 - B)) >> (32 - B); +} + +/// Sign-extend the number in the bottom B bits of X to a 32-bit integer. +/// Requires 0 < B < 32. +inline int32_t SignExtend32(uint32_t X, unsigned B) { + assert(B > 0 && "Bit width can't be 0."); + assert(B <= 32 && "Bit width out of range."); + return int32_t(X << (32 - B)) >> (32 - B); +} + +/// Sign-extend the number in the bottom B bits of X to a 64-bit integer. +/// Requires 0 < B < 64. +template +constexpr inline int64_t SignExtend64(uint64_t x) { + static_assert(B > 0, "Bit width can't be 0."); + static_assert(B <= 64, "Bit width out of range."); + return int64_t(x << (64 - B)) >> (64 - B); +} + +/// Sign-extend the number in the bottom B bits of X to a 64-bit integer. +/// Requires 0 < B < 64. +inline int64_t SignExtend64(uint64_t X, unsigned B) { + assert(B > 0 && "Bit width can't be 0."); + assert(B <= 64 && "Bit width out of range."); + return int64_t(X << (64 - B)) >> (64 - B); +} + +/// Subtract two unsigned integers, X and Y, of type T and return the absolute +/// value of the result. +template +std::enable_if_t, T> AbsoluteDifference(T X, T Y) { + return std::max(X, Y) - std::min(X, Y); +} + +/// Add two unsigned integers, X and Y, of type T. Clamp the result to the +/// maximum representable value of T on overflow. ResultOverflowed indicates if +/// the result is larger than the maximum representable value of type T. +template +std::enable_if_t, T> SaturatingAdd( + T X, + T Y, + bool* ResultOverflowed = nullptr) { + // NOLINTNEXTLINE(cppcoreguidelines-init-variables) + bool Dummy; + bool& Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy; + // Hacker's Delight, p. 29 + T Z = X + Y; + Overflowed = (Z < X || Z < Y); + if (Overflowed) + return std::numeric_limits::max(); + else + return Z; +} + +/// Multiply two unsigned integers, X and Y, of type T. Clamp the result to the +/// maximum representable value of T on overflow. ResultOverflowed indicates if +/// the result is larger than the maximum representable value of type T. +template +std::enable_if_t, T> SaturatingMultiply( + T X, + T Y, + bool* ResultOverflowed = nullptr) { + // NOLINTNEXTLINE(cppcoreguidelines-init-variables) + bool Dummy; + bool& Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy; + + // Hacker's Delight, p. 30 has a different algorithm, but we don't use that + // because it fails for uint16_t (where multiplication can have undefined + // behavior due to promotion to int), and requires a division in addition + // to the multiplication. + + Overflowed = false; + + // Log2(Z) would be either Log2Z or Log2Z + 1. + // Special case: if X or Y is 0, Log2_64 gives -1, and Log2Z + // will necessarily be less than Log2Max as desired. + int Log2Z = Log2_64(X) + Log2_64(Y); + const T Max = std::numeric_limits::max(); + int Log2Max = Log2_64(Max); + if (Log2Z < Log2Max) { + return X * Y; + } + if (Log2Z > Log2Max) { + Overflowed = true; + return Max; + } + + // We're going to use the top bit, and maybe overflow one + // bit past it. Multiply all but the bottom bit then add + // that on at the end. + T Z = (X >> 1) * Y; + if (Z & ~(Max >> 1)) { + Overflowed = true; + return Max; + } + Z <<= 1; + if (X & 1) + return SaturatingAdd(Z, Y, ResultOverflowed); + + return Z; +} + +/// Multiply two unsigned integers, X and Y, and add the unsigned integer, A to +/// the product. Clamp the result to the maximum representable value of T on +/// overflow. ResultOverflowed indicates if the result is larger than the +/// maximum representable value of type T. +template +std::enable_if_t, T> SaturatingMultiplyAdd( + T X, + T Y, + T A, + bool* ResultOverflowed = nullptr) { + // NOLINTNEXTLINE(cppcoreguidelines-init-variables) + bool Dummy; + bool& Overflowed = ResultOverflowed ? *ResultOverflowed : Dummy; + + T Product = SaturatingMultiply(X, Y, &Overflowed); + if (Overflowed) + return Product; + + return SaturatingAdd(A, Product, &Overflowed); +} + +/// Use this rather than HUGE_VALF; the latter causes warnings on MSVC. +extern const float huge_valf; +} // namespace c10::llvm + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/logging_common.h b/backends/pytorch/include/c10/util/logging_common.h new file mode 100644 index 000000000..8d881f4de --- /dev/null +++ b/backends/pytorch/include/c10/util/logging_common.h @@ -0,0 +1,79 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#ifndef C10_UTIL_LOGGING_COMMON_H_ +#define C10_UTIL_LOGGING_COMMON_H_ + +#include +#include + +namespace c10 { + +// MessageLogger that throws exceptions instead of aborting (glog version) +// or logs and may abort (non-glog version). +class C10_API MessageLogger { + public: + MessageLogger( + const char* file, + int line, + int severity, + bool exit_on_fatal = true); + ~MessageLogger() noexcept(false); + + // Return the stream associated with the logger object. + std::stringstream& stream(); + + private: + // When there is a fatal log, and fatal == true, we abort + // otherwise, we throw. + void DealWithFatal(); + +#if defined(ANDROID) && !defined(C10_USE_GLOG) + const char* tag_{"native"}; +#endif + std::stringstream stream_; + int severity_; + bool exit_on_fatal_; +}; + +// This class is used to explicitly ignore values in the conditional +// logging macros. This avoids compiler warnings like "value computed +// is not used" and "statement has no effect". +class C10_API LoggerVoidify { + public: + LoggerVoidify() = default; + // This has to be an operator with a precedence lower than << but + // higher than ?: + void operator&(const std::ostream& s [[maybe_unused]]) {} +}; + +// Forward declarations for CheckNotNull functions +template +T& CheckNotNullCommon( + const char* file, + int line, + const char* names, + T& t, + bool fatal = true); + +template +T* CheckNotNull( + const char* file, + int line, + const char* names, + T* t, + bool fatal = true); + +template +T& CheckNotNull( + const char* file, + int line, + const char* names, + T& t, + bool fatal = true); + +} // namespace c10 + +#endif // C10_UTIL_LOGGING_COMMON_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/logging_is_not_google_glog.h b/backends/pytorch/include/c10/util/logging_is_not_google_glog.h new file mode 100644 index 000000000..efeffb93a --- /dev/null +++ b/backends/pytorch/include/c10/util/logging_is_not_google_glog.h @@ -0,0 +1,186 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#ifndef C10_UTIL_LOGGING_IS_NOT_GOOGLE_GLOG_H_ +#define C10_UTIL_LOGGING_IS_NOT_GOOGLE_GLOG_H_ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include + +const char CAFFE2_SEVERITY_PREFIX[] = "FEWIV"; + +namespace c10 { + +// Log severity level constants. +const int GLOG_FATAL = 3; +const int GLOG_ERROR = 2; +const int GLOG_WARNING = 1; +const int GLOG_INFO = 0; + +// Helpers for TORCH_CHECK_NOTNULL(). Two are necessary to support both raw +// pointers and smart pointers. +template +T& CheckNotNullCommon( + const char* file, + int line, + const char* names, + T& t, + bool fatal) { + if (t == nullptr) { + MessageLogger(file, line, GLOG_FATAL, fatal).stream() + << "Check failed: '" << names << "' must be non NULL. "; + } + return t; +} + +template +T* CheckNotNull( + const char* file, + int line, + const char* names, + T* t, + bool fatal) { + return CheckNotNullCommon(file, line, names, t, fatal); +} + +template +T& CheckNotNull( + const char* file, + int line, + const char* names, + T& t, + bool fatal) { + return CheckNotNullCommon(file, line, names, t, fatal); +} +} // namespace c10 + +// ---------------------- Logging Macro definitions -------------------------- + +static_assert( + CAFFE2_LOG_THRESHOLD <= ::c10::GLOG_FATAL, + "CAFFE2_LOG_THRESHOLD should at most be GLOG_FATAL."); +// If n is under the compile time caffe log threshold, The _CAFFE_LOG(n) +// should not generate anything in optimized code. +#define LOG(n) \ + if (::c10::GLOG_##n >= CAFFE2_LOG_THRESHOLD) \ + ::c10::MessageLogger(__FILE__, __LINE__, ::c10::GLOG_##n).stream() +#define VLOG(n) \ + if (-n >= CAFFE2_LOG_THRESHOLD) \ + ::c10::MessageLogger(__FILE__, __LINE__, -n).stream() + +#define LOG_IF(n, condition) \ + if (::c10::GLOG_##n >= CAFFE2_LOG_THRESHOLD && (condition)) \ + ::c10::MessageLogger(__FILE__, __LINE__, ::c10::GLOG_##n).stream() +#define VLOG_IF(n, condition) \ + if (-n >= CAFFE2_LOG_THRESHOLD && (condition)) \ + ::c10::MessageLogger(__FILE__, __LINE__, -n).stream() + +#define VLOG_IS_ON(verboselevel) (CAFFE2_LOG_THRESHOLD <= -(verboselevel)) + +// Log with source location information override (to be used in generic +// warning/error handlers implemented as functions, not macros) +#define LOG_AT_FILE_LINE(n, file, line) \ + if (::c10::GLOG_##n >= CAFFE2_LOG_THRESHOLD) \ + ::c10::MessageLogger(file, line, ::c10::GLOG_##n).stream() + +// Log only if condition is met. Otherwise evaluates to void. +#define FATAL_IF(condition) \ + condition ? (void)0 \ + : ::c10::LoggerVoidify() & \ + ::c10::MessageLogger(__FILE__, __LINE__, ::c10::GLOG_FATAL).stream() + +// Check for a given boolean condition. +#define CHECK(condition) FATAL_IF(condition) << "Check failed: " #condition " " + +#ifndef NDEBUG +// Debug only version of CHECK +#define DCHECK(condition) FATAL_IF(condition) << "Check failed: " #condition " " +#define DLOG(severity) LOG(severity) +#else // NDEBUG +// Optimized version - generates no code. +#define DCHECK(condition) \ + while (false) \ + CHECK(condition) + +#define DLOG(n) \ + true ? (void)0 \ + : ::c10::LoggerVoidify() & \ + ::c10::MessageLogger(__FILE__, __LINE__, ::c10::GLOG_##n).stream() +#endif // NDEBUG + +// ---------------------- Support for std objects -------------------------- +// These are adapted from glog to support a limited set of logging capability +// for STL objects. + +namespace std { +// Forward declare these two, and define them after all the container streams +// operators so that we can recurse from pair -> container -> container -> pair +// properly. +template +std::ostream& operator<<(std::ostream& out, const std::pair& p); +} // namespace std + +namespace c10 { +template +void PrintSequence(std::ostream& ss, Iter begin, Iter end); +} // namespace c10 + +namespace std { +#define INSTANTIATE_FOR_CONTAINER(container) \ + template \ + std::ostream& operator<<( \ + std::ostream& out, const container& seq) { \ + c10::PrintSequence(out, seq.begin(), seq.end()); \ + return out; \ + } + +INSTANTIATE_FOR_CONTAINER(std::vector) +INSTANTIATE_FOR_CONTAINER(std::map) +INSTANTIATE_FOR_CONTAINER(std::set) +#undef INSTANTIATE_FOR_CONTAINER + +template +inline std::ostream& operator<<( + std::ostream& out, + const std::pair& p) { + out << '(' << p.first << ", " << p.second << ')'; + return out; +} + +inline std::ostream& operator<<( + std::ostream& out, + const std::nullptr_t& /*unused*/) { + out << "(null)"; + return out; +} +} // namespace std + +namespace c10 { +template +inline void PrintSequence(std::ostream& out, Iter begin, Iter end) { + // Output at most 100 elements -- appropriate if used for logging. + for (int i = 0; begin != end && i < 100; ++i, ++begin) { + if (i > 0) + out << ' '; + out << *begin; + } + if (begin != end) { + out << " ..."; + } +} +} // namespace c10 + +#endif // C10_UTIL_LOGGING_IS_NOT_GOOGLE_GLOG_H_ + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/order_preserving_flat_hash_map.h b/backends/pytorch/include/c10/util/order_preserving_flat_hash_map.h new file mode 100644 index 000000000..e991a567e --- /dev/null +++ b/backends/pytorch/include/c10/util/order_preserving_flat_hash_map.h @@ -0,0 +1,2222 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +// Taken from +// https://github.com/skarupke/flat_hash_map/blob/2c4687431f978f02a3780e24b8b701d22aa32d9c/flat_hash_map.hpp +// with fixes applied: +// - https://github.com/skarupke/flat_hash_map/pull/25 +// - https://github.com/skarupke/flat_hash_map/pull/26 +// - replace size_t with uint64_t to fix it for 32bit +// - add "GCC diagnostic" pragma to ignore -Wshadow +// - make sherwood_v3_table::convertible_to_iterator public because GCC5 seems +// to have issues with it otherwise +// - fix compiler warnings in operator templated_iterator +// - make use of 'if constexpr' and eliminate AssignIfTrue template + +// Copyright Malte Skarupke 2017. +// Distributed under the Boost Software License, Version 1.0. +// (See http://www.boost.org/LICENSE_1_0.txt) + +// Modified to maintain insertion and deletion order through a doubly-linked +// list + +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#ifdef _MSC_VER +#define SKA_NOINLINE(...) __declspec(noinline) __VA_ARGS__ +#else +#define SKA_NOINLINE(...) __VA_ARGS__ __attribute__((noinline)) +#endif + +namespace ska_ordered { + +struct prime_number_hash_policy; +struct power_of_two_hash_policy; +struct fibonacci_hash_policy; + +namespace detailv3 { +template +struct functor_storage : Functor { + functor_storage() = default; + functor_storage(const Functor& functor) : Functor(functor) {} + template + Result operator()(Args&&... args) { + return static_cast(*this)(std::forward(args)...); + } + template + Result operator()(Args&&... args) const { + return static_cast(*this)(std::forward(args)...); + } +}; +template +struct functor_storage { + typedef Result (*function_ptr)(Args...); + function_ptr function; + functor_storage(function_ptr function) : function(function) {} + Result operator()(Args... args) const { + return function(std::forward(args)...); + } + operator function_ptr&() { + return function; + } + operator const function_ptr&() { + return function; + } +}; +template +struct KeyOrValueHasher : functor_storage { + typedef functor_storage hasher_storage; + KeyOrValueHasher() = default; + KeyOrValueHasher(const hasher& hash) : hasher_storage(hash) {} + uint64_t operator()(const key_type& key) { + return static_cast(*this)(key); + } + uint64_t operator()(const key_type& key) const { + return static_cast(*this)(key); + } + uint64_t operator()(const value_type& value) { + return static_cast(*this)(value.first); + } + uint64_t operator()(const value_type& value) const { + return static_cast(*this)(value.first); + } + template + uint64_t operator()(const std::pair& value) { + return static_cast(*this)(value.first); + } + template + uint64_t operator()(const std::pair& value) const { + return static_cast(*this)(value.first); + } +}; +template +struct KeyOrValueEquality : functor_storage { + typedef functor_storage equality_storage; + KeyOrValueEquality() = default; + KeyOrValueEquality(const key_equal& equality) : equality_storage(equality) {} + bool operator()(const key_type& lhs, const key_type& rhs) { + return static_cast(*this)(lhs, rhs); + } + bool operator()(const key_type& lhs, const value_type& rhs) { + return static_cast(*this)(lhs, rhs.first); + } + bool operator()(const value_type& lhs, const key_type& rhs) { + return static_cast(*this)(lhs.first, rhs); + } + bool operator()(const value_type& lhs, const value_type& rhs) { + return static_cast(*this)(lhs.first, rhs.first); + } + template + bool operator()(const key_type& lhs, const std::pair& rhs) { + return static_cast(*this)(lhs, rhs.first); + } + template + bool operator()(const std::pair& lhs, const key_type& rhs) { + return static_cast(*this)(lhs.first, rhs); + } + template + bool operator()(const value_type& lhs, const std::pair& rhs) { + return static_cast(*this)(lhs.first, rhs.first); + } + template + bool operator()(const std::pair& lhs, const value_type& rhs) { + return static_cast(*this)(lhs.first, rhs.first); + } + template + bool operator()(const std::pair& lhs, const std::pair& rhs) { + return static_cast(*this)(lhs.first, rhs.first); + } +}; +static constexpr int8_t min_lookups = 4; +template +// NOLINTNEXTLINE(cppcoreguidelines-special-member-functions) +struct sherwood_v3_entry { + // NOLINTNEXTLINE(modernize-use-equals-default) + sherwood_v3_entry() {} + sherwood_v3_entry(int8_t distance_from_desired) + : distance_from_desired(distance_from_desired) {} + // NOLINTNEXTLINE(modernize-use-equals-default) + ~sherwood_v3_entry() {} + + bool has_value() const { + return distance_from_desired >= 0; + } + bool is_empty() const { + return distance_from_desired < 0; + } + bool is_at_desired_position() const { + return distance_from_desired <= 0; + } + template + void emplace(int8_t distance, Args&&... args) { + new (std::addressof(value)) T(std::forward(args)...); + distance_from_desired = distance; + } + + void destroy_value() { + value.~T(); + distance_from_desired = -1; + } + + sherwood_v3_entry* prev = nullptr; + sherwood_v3_entry* next = nullptr; + int8_t distance_from_desired = -1; + static constexpr int8_t special_end_value = 0; + union { + T value; + }; +}; + +inline int8_t log2(uint64_t value) { + static constexpr std::array table = { + 63, 0, 58, 1, 59, 47, 53, 2, 60, 39, 48, 27, 54, 33, 42, 3, + 61, 51, 37, 40, 49, 18, 28, 20, 55, 30, 34, 11, 43, 14, 22, 4, + 62, 57, 46, 52, 38, 26, 32, 41, 50, 36, 17, 19, 29, 10, 13, 21, + 56, 45, 25, 31, 35, 16, 9, 12, 44, 24, 15, 8, 23, 7, 6, 5}; + value |= value >> 1; + value |= value >> 2; + value |= value >> 4; + value |= value >> 8; + value |= value >> 16; + value |= value >> 32; + return table[((value - (value >> 1)) * 0x07EDD5E59A4E28C2) >> 58]; +} + +inline uint64_t next_power_of_two(uint64_t i) { + --i; + i |= i >> 1; + i |= i >> 2; + i |= i >> 4; + i |= i >> 8; + i |= i >> 16; + i |= i >> 32; + ++i; + return i; +} + +// Implementation taken from http://en.cppreference.com/w/cpp/types/void_t +// (it takes CWG1558 into account and also works for older compilers) +template +struct make_void { + typedef void type; +}; +template +using void_t = typename make_void::type; + +template +struct HashPolicySelector { + typedef fibonacci_hash_policy type; +}; +template +struct HashPolicySelector> { + typedef typename T::hash_policy type; +}; + +template < + typename T, + typename FindKey, + typename ArgumentHash, + typename Hasher, + typename ArgumentEqual, + typename Equal, + typename ArgumentAlloc, + typename EntryAlloc> +class sherwood_v3_table : private EntryAlloc, private Hasher, private Equal { + using Entry = detailv3::sherwood_v3_entry; + using AllocatorTraits = std::allocator_traits; + using EntryPointer = typename AllocatorTraits::pointer; + + public: + struct convertible_to_iterator; + + using value_type = T; + using size_type = uint64_t; + using difference_type = std::ptrdiff_t; + using hasher = ArgumentHash; + using key_equal = ArgumentEqual; + using allocator_type = EntryAlloc; + using reference = value_type&; + using const_reference = const value_type&; + using pointer = value_type*; + using const_pointer = const value_type*; + + sherwood_v3_table() = default; + explicit sherwood_v3_table( + size_type bucket_count, + const ArgumentHash& hash = ArgumentHash(), + const ArgumentEqual& equal = ArgumentEqual(), + const ArgumentAlloc& alloc = ArgumentAlloc()) + : EntryAlloc(alloc), Hasher(hash), Equal(equal) { + rehash(bucket_count); + } + sherwood_v3_table(size_type bucket_count, const ArgumentAlloc& alloc) + : sherwood_v3_table( + bucket_count, + ArgumentHash(), + ArgumentEqual(), + alloc) {} + sherwood_v3_table( + size_type bucket_count, + const ArgumentHash& hash, + const ArgumentAlloc& alloc) + : sherwood_v3_table(bucket_count, hash, ArgumentEqual(), alloc) {} + explicit sherwood_v3_table(const ArgumentAlloc& alloc) : EntryAlloc(alloc) {} + template + sherwood_v3_table( + It first, + It last, + size_type bucket_count = 0, + const ArgumentHash& hash = ArgumentHash(), + const ArgumentEqual& equal = ArgumentEqual(), + const ArgumentAlloc& alloc = ArgumentAlloc()) + : sherwood_v3_table(bucket_count, hash, equal, alloc) { + insert(first, last); + } + template + sherwood_v3_table( + It first, + It last, + size_type bucket_count, + const ArgumentAlloc& alloc) + : sherwood_v3_table( + first, + last, + bucket_count, + ArgumentHash(), + ArgumentEqual(), + alloc) {} + template + sherwood_v3_table( + It first, + It last, + size_type bucket_count, + const ArgumentHash& hash, + const ArgumentAlloc& alloc) + : sherwood_v3_table( + first, + last, + bucket_count, + hash, + ArgumentEqual(), + alloc) {} + sherwood_v3_table( + std::initializer_list il, + size_type bucket_count = 0, + const ArgumentHash& hash = ArgumentHash(), + const ArgumentEqual& equal = ArgumentEqual(), + const ArgumentAlloc& alloc = ArgumentAlloc()) + : sherwood_v3_table(bucket_count, hash, equal, alloc) { + if (bucket_count == 0) + rehash(il.size()); + insert(il.begin(), il.end()); + } + sherwood_v3_table( + std::initializer_list il, + size_type bucket_count, + const ArgumentAlloc& alloc) + : sherwood_v3_table( + il, + bucket_count, + ArgumentHash(), + ArgumentEqual(), + alloc) {} + sherwood_v3_table( + std::initializer_list il, + size_type bucket_count, + const ArgumentHash& hash, + const ArgumentAlloc& alloc) + : sherwood_v3_table(il, bucket_count, hash, ArgumentEqual(), alloc) {} + sherwood_v3_table(const sherwood_v3_table& other) + : sherwood_v3_table( + other, + AllocatorTraits::select_on_container_copy_construction( + other.get_allocator())) {} + sherwood_v3_table(const sherwood_v3_table& other, const ArgumentAlloc& alloc) + : EntryAlloc(alloc), + Hasher(other), + Equal(other), + _max_load_factor(other._max_load_factor) { + rehash_for_other_container(other); + try { + insert(other.begin(), other.end()); + } catch (...) { + clear(); + deallocate_data(entries, num_slots_minus_one, max_lookups); + throw; + } + } + sherwood_v3_table(sherwood_v3_table&& other) noexcept + : EntryAlloc(std::move(other)), + Hasher(std::move(other)), + Equal(std::move(other)) { + swap_pointers(other); + } + sherwood_v3_table( + sherwood_v3_table&& other, + const ArgumentAlloc& alloc) noexcept + : EntryAlloc(alloc), Hasher(std::move(other)), Equal(std::move(other)) { + swap_pointers(other); + } + sherwood_v3_table& operator=(const sherwood_v3_table& other) { + if (this == std::addressof(other)) + return *this; + + clear(); + if constexpr (AllocatorTraits::propagate_on_container_copy_assignment:: + value) { + if (static_cast(*this) != + static_cast(other)) { + reset_to_empty_state(); + } + static_cast(*this) = other; + } + _max_load_factor = other._max_load_factor; + static_cast(*this) = other; + static_cast(*this) = other; + rehash_for_other_container(other); + insert(other.begin(), other.end()); + return *this; + } + sherwood_v3_table& operator=(sherwood_v3_table&& other) noexcept { + if (this == std::addressof(other)) + return *this; + else if constexpr (AllocatorTraits::propagate_on_container_move_assignment:: + value) { + clear(); + reset_to_empty_state(); + static_cast(*this) = std::move(other); + swap_pointers(other); + } else if ( + static_cast(*this) == static_cast(other)) { + swap_pointers(other); + } else { + clear(); + _max_load_factor = other._max_load_factor; + rehash_for_other_container(other); + for (T& elem : other) + emplace(std::move(elem)); + other.clear(); + } + static_cast(*this) = std::move(other); + static_cast(*this) = std::move(other); + return *this; + } + ~sherwood_v3_table() { + clear(); + deallocate_data(entries, num_slots_minus_one, max_lookups); + } + + const allocator_type& get_allocator() const { + return static_cast(*this); + } + const ArgumentEqual& key_eq() const { + return static_cast(*this); + } + const ArgumentHash& hash_function() const { + return static_cast(*this); + } + + template + struct templated_iterator { + templated_iterator() = default; + templated_iterator(EntryPointer current) : current(current) {} + EntryPointer current = EntryPointer(); + + using iterator_category = std::forward_iterator_tag; + using value_type = ValueType; + using difference_type = ptrdiff_t; + using pointer = ValueType*; + using reference = ValueType&; + + friend bool operator==( + const templated_iterator& lhs, + const templated_iterator& rhs) { + return lhs.current == rhs.current; + } + friend bool operator!=( + const templated_iterator& lhs, + const templated_iterator& rhs) { + return !(lhs == rhs); + } + + templated_iterator& operator++() { + current = current->next; + return *this; + } + templated_iterator operator++(int) { + templated_iterator copy(*this); + ++*this; + return copy; + } + + ValueType& operator*() const { + return current->value; + } + ValueType* operator->() const { + return std::addressof(current->value); + } + + // the template automatically disables the operator when value_type is + // already const, because that would cause a lot of compiler warnings + // otherwise. + template < + class target_type = const value_type, + class = std::enable_if_t< + std::is_same_v && + !std::is_same_v>> + operator templated_iterator() const { + return {current}; + } + }; + using iterator = templated_iterator; + using const_iterator = templated_iterator; + + iterator begin() { + return sentinel->next; + } + const_iterator begin() const { + return sentinel->next; + } + const_iterator cbegin() const { + return begin(); + } + iterator end() { + return sentinel; + } + const_iterator end() const { + return sentinel; + } + const_iterator cend() const { + return end(); + } + + iterator find(const FindKey& key) { + uint64_t index = + hash_policy.index_for_hash(hash_object(key), num_slots_minus_one); + EntryPointer it = entries + ptrdiff_t(index); + for (int8_t distance = 0; it->distance_from_desired >= distance; + ++distance, ++it) { + if (compares_equal(key, it->value)) + return {it}; + } + return end(); + } + const_iterator find(const FindKey& key) const { + // NOLINTNEXTLINE(cppcoreguidelines-pro-type-const-cast) + return const_cast(this)->find(key); + } + uint64_t count(const FindKey& key) const { + return find(key) == end() ? 0 : 1; + } + std::pair equal_range(const FindKey& key) { + iterator found = find(key); + if (found == end()) + return {found, found}; + else + return {found, std::next(found)}; + } + std::pair equal_range( + const FindKey& key) const { + const_iterator found = find(key); + if (found == end()) + return {found, found}; + else + return {found, std::next(found)}; + } + + template + std::pair emplace(Key&& key, Args&&... args) { + uint64_t index = + hash_policy.index_for_hash(hash_object(key), num_slots_minus_one); + EntryPointer current_entry = entries + ptrdiff_t(index); + int8_t distance_from_desired = 0; + for (; current_entry->distance_from_desired >= distance_from_desired; + ++current_entry, ++distance_from_desired) { + // insertion of an existing key does not change ordering + if (compares_equal(key, current_entry->value)) + return {{current_entry}, false}; + } + return emplace_new_key( + distance_from_desired, + current_entry, + std::forward(key), + std::forward(args)...); + } + + std::pair insert(const value_type& value) { + return emplace(value); + } + std::pair insert(value_type&& value) { + return emplace(std::move(value)); + } + template + iterator emplace_hint(const_iterator /*unused*/, Args&&... args) { + return emplace(std::forward(args)...).first; + } + iterator insert(const_iterator /*unused*/, const value_type& value) { + return emplace(value).first; + } + iterator insert(const_iterator /*unused*/, value_type&& value) { + return emplace(std::move(value)).first; + } + + template + void insert(It begin, It end) { + for (; begin != end; ++begin) { + emplace(*begin); + } + } + void insert(std::initializer_list il) { + insert(il.begin(), il.end()); + } + + void rehash(uint64_t num_buckets) { + num_buckets = std::max( + num_buckets, + static_cast(std::ceil( + static_cast(num_elements) / + static_cast(_max_load_factor)))); + if (num_buckets == 0) { + reset_to_empty_state(); + return; + } + auto new_prime_index = hash_policy.next_size_over(num_buckets); + if (num_buckets == bucket_count()) + return; + int8_t new_max_lookups = compute_max_lookups(num_buckets); + EntryPointer new_buckets( + AllocatorTraits::allocate(*this, num_buckets + new_max_lookups)); + EntryPointer special_end_item = + new_buckets + static_cast(num_buckets + new_max_lookups - 1); + for (EntryPointer it = new_buckets; it != special_end_item; ++it) + it->distance_from_desired = -1; + special_end_item->distance_from_desired = Entry::special_end_value; + std::swap(entries, new_buckets); + std::swap(num_slots_minus_one, num_buckets); + --num_slots_minus_one; + hash_policy.commit(new_prime_index); + int8_t old_max_lookups = max_lookups; + max_lookups = new_max_lookups; + num_elements = 0; + + auto start = sentinel->next; + // point sentinel to itself; + reset_list(); + // reinsert list + for (EntryPointer it = start; it != sentinel;) { + auto next = it->next; + emplace(std::move(it->value)); + it->destroy_value(); + it = next; + } + + deallocate_data(new_buckets, num_buckets, old_max_lookups); + } + + void reserve(uint64_t num_elements_) { + uint64_t required_buckets = num_buckets_for_reserve(num_elements_); + if (required_buckets > bucket_count()) + rehash(required_buckets); + } + + void replace_linked_list_position( + EntryPointer to_be_replaced, + EntryPointer new_node) { + remove_from_list(new_node); + insert_after(new_node, to_be_replaced->prev); + remove_from_list(to_be_replaced); + } + + // the return value is a type that can be converted to an iterator + // the reason for doing this is that it's not free to find the + // iterator pointing at the next element. if you care about the + // next iterator, turn the return value into an iterator + convertible_to_iterator erase(const_iterator to_erase) { + EntryPointer current = to_erase.current; + remove_from_list(current); + current->destroy_value(); + --num_elements; + + for (EntryPointer next = current + ptrdiff_t(1); + !next->is_at_desired_position(); + ++current, ++next) { + // if an entry is being removed, and there are other entries with the + // same hash, the other entries get moved to their desired position by + // reinserting. + current->emplace(next->distance_from_desired - 1, std::move(next->value)); + replace_linked_list_position(next, current); + next->destroy_value(); + } + return {to_erase.current}; + } + + iterator erase(const_iterator begin_it, const_iterator end_it) { + // whenever an entry is removed and there are other entries with the same + // hash, the other entries must get moved to their desired position. + // any reference to a moved entry is invalidated. + // here, we iterate through the range, and make sure that we update + // the pointer to our next entry in the list or the end of the iterator + // when it is invalidated. + + auto curr_iter = begin_it.current; + auto next_iter = curr_iter->next; + auto end_iter = end_it.current; + + while (curr_iter != end_iter) { + remove_from_list(curr_iter); + curr_iter->destroy_value(); + --num_elements; + + for (EntryPointer next_hash_slot = curr_iter + ptrdiff_t(1); + !next_hash_slot->is_at_desired_position(); + ++curr_iter, ++next_hash_slot) { + curr_iter->emplace( + next_hash_slot->distance_from_desired - 1, + std::move(next_hash_slot->value)); + replace_linked_list_position(next_hash_slot, curr_iter); + next_hash_slot->destroy_value(); + + // we are invalidating next_iter or end_iter + if (next_hash_slot == end_iter) { + end_iter = curr_iter; + } else if (next_hash_slot == next_iter) { + next_iter = curr_iter; + } + } + curr_iter = next_iter; + next_iter = curr_iter->next; + } + + return {end_iter}; + } + + uint64_t erase(const FindKey& key) { + auto found = find(key); + if (found == end()) + return 0; + else { + erase(found); + return 1; + } + } + + void clear() { + for (EntryPointer it = entries, + end = it + + static_cast(num_slots_minus_one + max_lookups); + it != end; + ++it) { + if (it->has_value()) + it->destroy_value(); + } + reset_list(); + num_elements = 0; + } + + void shrink_to_fit() { + rehash_for_other_container(*this); + } + + void swap(sherwood_v3_table& other) noexcept { + using std::swap; + swap_pointers(other); + swap(static_cast(*this), static_cast(other)); + swap( + static_cast(*this), static_cast(other)); + if (AllocatorTraits::propagate_on_container_swap::value) + swap(static_cast(*this), static_cast(other)); + } + + uint64_t size() const { + return num_elements; + } + uint64_t max_size() const { + return (AllocatorTraits::max_size(*this)) / sizeof(Entry); + } + uint64_t bucket_count() const { + return num_slots_minus_one ? num_slots_minus_one + 1 : 0; + } + size_type max_bucket_count() const { + return (AllocatorTraits::max_size(*this) - min_lookups) / sizeof(Entry); + } + uint64_t bucket(const FindKey& key) const { + return hash_policy.index_for_hash(hash_object(key), num_slots_minus_one); + } + float load_factor() const { + uint64_t buckets = bucket_count(); + if (buckets) + return static_cast(num_elements) / bucket_count(); + else + return 0; + } + void max_load_factor(float value) { + _max_load_factor = value; + } + float max_load_factor() const { + return _max_load_factor; + } + + bool empty() const { + return num_elements == 0; + } + + private: + EntryPointer entries = empty_default_table(); + uint64_t num_slots_minus_one = 0; + typename HashPolicySelector::type hash_policy; + int8_t max_lookups = detailv3::min_lookups - 1; + float _max_load_factor = 0.5f; + uint64_t num_elements = 0; + std::unique_ptr> sentinel_val; + + // head of doubly linked list + EntryPointer sentinel = initSentinel(); + + EntryPointer initSentinel() { + // needs to be a pointer so that hash map can be used with forward declared + // types + sentinel_val = std::make_unique>(); + sentinel = sentinel_val.get(); + reset_list(); + return sentinel; + } + + EntryPointer empty_default_table() { + EntryPointer result = + AllocatorTraits::allocate(*this, detailv3::min_lookups); + EntryPointer special_end_item = + result + static_cast(detailv3::min_lookups - 1); + for (EntryPointer it = result; it != special_end_item; ++it) + it->distance_from_desired = -1; + special_end_item->distance_from_desired = Entry::special_end_value; + return result; + } + + static int8_t compute_max_lookups(uint64_t num_buckets) { + int8_t desired = detailv3::log2(num_buckets); + return std::max(detailv3::min_lookups, desired); + } + + uint64_t num_buckets_for_reserve(uint64_t num_elements_) const { + return static_cast(std::ceil( + static_cast(num_elements_) / + std::min(0.5, static_cast(_max_load_factor)))); + } + void rehash_for_other_container(const sherwood_v3_table& other) { + rehash( + std::min(num_buckets_for_reserve(other.size()), other.bucket_count())); + } + + void swap_pointers(sherwood_v3_table& other) { + using std::swap; + swap(hash_policy, other.hash_policy); + swap(entries, other.entries); + swap(num_slots_minus_one, other.num_slots_minus_one); + swap(num_elements, other.num_elements); + swap(max_lookups, other.max_lookups); + swap(_max_load_factor, other._max_load_factor); + swap(sentinel, other.sentinel); + swap(sentinel_val, other.sentinel_val); + } + + void reset_list() { + sentinel->next = sentinel; + sentinel->prev = sentinel; + } + + void remove_from_list(EntryPointer elem) { + elem->prev->next = elem->next; + elem->next->prev = elem->prev; + } + + void insert_after(EntryPointer new_elem, EntryPointer prev) { + auto next = prev->next; + + prev->next = new_elem; + new_elem->prev = prev; + + new_elem->next = next; + next->prev = new_elem; + } + + void swap_adjacent_nodes(EntryPointer before, EntryPointer after) { + // sentinel stays constant, so before->prev cannot equal after + auto before_prev = before->prev; + auto after_next = after->next; + + before_prev->next = after; + after->prev = before_prev; + + after_next->prev = before; + before->next = after_next; + + before->prev = after; + after->next = before; + } + + void swap_positions(EntryPointer p1, EntryPointer p2) { + if (p1 == p2) { + return; + } + if (p1->next == p2) { + return swap_adjacent_nodes(p1, p2); + } else if (p2->next == p1) { + return swap_adjacent_nodes(p2, p1); + } + + auto p1_prev = p1->prev; + auto p1_next = p1->next; + + auto p2_prev = p2->prev; + auto p2_next = p2->next; + + p1_prev->next = p2; + p2->prev = p1_prev; + + p1_next->prev = p2; + p2->next = p1_next; + + p2_prev->next = p1; + p1->prev = p2_prev; + + p2_next->prev = p1; + p1->next = p2_next; + } + + void append_to_list(EntryPointer new_tail) { + insert_after(new_tail, sentinel->prev); + } + + template + SKA_NOINLINE(std::pair) + emplace_new_key( + int8_t distance_from_desired, + EntryPointer current_entry, + Key&& key, + Args&&... args) { + using std::swap; + if (num_slots_minus_one == 0 || distance_from_desired == max_lookups || + static_cast(num_elements + 1) > + static_cast(num_slots_minus_one + 1) * + static_cast(_max_load_factor)) { + grow(); + return emplace(std::forward(key), std::forward(args)...); + } else if (current_entry->is_empty()) { + current_entry->emplace( + distance_from_desired, + std::forward(key), + std::forward(args)...); + ++num_elements; + append_to_list(current_entry); + return {{current_entry}, true}; + } + value_type to_insert(std::forward(key), std::forward(args)...); + swap(distance_from_desired, current_entry->distance_from_desired); + // We maintain the invariant that: + // - result.current_entry contains the new value we're inserting + // and is in the LinkedList position of to_insert + // - to_insert contains the value that represents the position of + // result.current_entry + swap(to_insert, current_entry->value); + iterator result = {current_entry}; + for (++distance_from_desired, ++current_entry;; ++current_entry) { + if (current_entry->is_empty()) { + current_entry->emplace(distance_from_desired, std::move(to_insert)); + append_to_list(current_entry); + // now we can swap back the displaced value to its correct position, + // putting the new value we're inserting to the front of the list + swap_positions(current_entry, result.current); + ++num_elements; + return {result, true}; + } else if (current_entry->distance_from_desired < distance_from_desired) { + swap(distance_from_desired, current_entry->distance_from_desired); + swap(to_insert, current_entry->value); + // to maintain our invariants we need to swap positions + // of result.current & current_entry: + swap_positions(result.current, current_entry); + ++distance_from_desired; + } else { + ++distance_from_desired; + if (distance_from_desired == max_lookups) { + // the displaced element gets put back into its correct position + // we grow the hash table, and then try again to reinsert the new + // element + swap(to_insert, result.current->value); + grow(); + return emplace(std::move(to_insert)); + } + } + } + } + + void grow() { + rehash(std::max(uint64_t(4), 2 * bucket_count())); + } + + void deallocate_data( + EntryPointer begin, + uint64_t num_slots_minus_one_, + int8_t max_lookups_) { + AllocatorTraits::deallocate( + *this, begin, num_slots_minus_one_ + max_lookups_ + 1); + } + + void reset_to_empty_state() { + deallocate_data(entries, num_slots_minus_one, max_lookups); + entries = empty_default_table(); + num_slots_minus_one = 0; + hash_policy.reset(); + max_lookups = detailv3::min_lookups - 1; + } + + template + uint64_t hash_object(const U& key) { + return static_cast(*this)(key); + } + template + uint64_t hash_object(const U& key) const { + return static_cast(*this)(key); + } + template + bool compares_equal(const L& lhs, const R& rhs) { + return static_cast(*this)(lhs, rhs); + } + + public: + struct convertible_to_iterator { + EntryPointer it; + + operator iterator() { + if (it->has_value()) + return {it}; + else + return ++iterator{it}; + } + operator const_iterator() { + if (it->has_value()) + return {it}; + else + return ++const_iterator{it}; + } + }; +}; +} // namespace detailv3 + +struct prime_number_hash_policy { + static uint64_t mod0(uint64_t /*unused*/) { + return 0llu; + } + static uint64_t mod2(uint64_t hash) { + return hash % 2llu; + } + static uint64_t mod3(uint64_t hash) { + return hash % 3llu; + } + static uint64_t mod5(uint64_t hash) { + return hash % 5llu; + } + static uint64_t mod7(uint64_t hash) { + return hash % 7llu; + } + static uint64_t mod11(uint64_t hash) { + return hash % 11llu; + } + static uint64_t mod13(uint64_t hash) { + return hash % 13llu; + } + static uint64_t mod17(uint64_t hash) { + return hash % 17llu; + } + static uint64_t mod23(uint64_t hash) { + return hash % 23llu; + } + static uint64_t mod29(uint64_t hash) { + return hash % 29llu; + } + static uint64_t mod37(uint64_t hash) { + return hash % 37llu; + } + static uint64_t mod47(uint64_t hash) { + return hash % 47llu; + } + static uint64_t mod59(uint64_t hash) { + return hash % 59llu; + } + static uint64_t mod73(uint64_t hash) { + return hash % 73llu; + } + static uint64_t mod97(uint64_t hash) { + return hash % 97llu; + } + static uint64_t mod127(uint64_t hash) { + return hash % 127llu; + } + static uint64_t mod151(uint64_t hash) { + return hash % 151llu; + } + static uint64_t mod197(uint64_t hash) { + return hash % 197llu; + } + static uint64_t mod251(uint64_t hash) { + return hash % 251llu; + } + static uint64_t mod313(uint64_t hash) { + return hash % 313llu; + } + static uint64_t mod397(uint64_t hash) { + return hash % 397llu; + } + static uint64_t mod499(uint64_t hash) { + return hash % 499llu; + } + static uint64_t mod631(uint64_t hash) { + return hash % 631llu; + } + static uint64_t mod797(uint64_t hash) { + return hash % 797llu; + } + static uint64_t mod1009(uint64_t hash) { + return hash % 1009llu; + } + static uint64_t mod1259(uint64_t hash) { + return hash % 1259llu; + } + static uint64_t mod1597(uint64_t hash) { + return hash % 1597llu; + } + static uint64_t mod2011(uint64_t hash) { + return hash % 2011llu; + } + static uint64_t mod2539(uint64_t hash) { + return hash % 2539llu; + } + static uint64_t mod3203(uint64_t hash) { + return hash % 3203llu; + } + static uint64_t mod4027(uint64_t hash) { + return hash % 4027llu; + } + static uint64_t mod5087(uint64_t hash) { + return hash % 5087llu; + } + static uint64_t mod6421(uint64_t hash) { + return hash % 6421llu; + } + static uint64_t mod8089(uint64_t hash) { + return hash % 8089llu; + } + static uint64_t mod10193(uint64_t hash) { + return hash % 10193llu; + } + static uint64_t mod12853(uint64_t hash) { + return hash % 12853llu; + } + static uint64_t mod16193(uint64_t hash) { + return hash % 16193llu; + } + static uint64_t mod20399(uint64_t hash) { + return hash % 20399llu; + } + static uint64_t mod25717(uint64_t hash) { + return hash % 25717llu; + } + static uint64_t mod32401(uint64_t hash) { + return hash % 32401llu; + } + static uint64_t mod40823(uint64_t hash) { + return hash % 40823llu; + } + static uint64_t mod51437(uint64_t hash) { + return hash % 51437llu; + } + static uint64_t mod64811(uint64_t hash) { + return hash % 64811llu; + } + static uint64_t mod81649(uint64_t hash) { + return hash % 81649llu; + } + static uint64_t mod102877(uint64_t hash) { + return hash % 102877llu; + } + static uint64_t mod129607(uint64_t hash) { + return hash % 129607llu; + } + static uint64_t mod163307(uint64_t hash) { + return hash % 163307llu; + } + static uint64_t mod205759(uint64_t hash) { + return hash % 205759llu; + } + static uint64_t mod259229(uint64_t hash) { + return hash % 259229llu; + } + static uint64_t mod326617(uint64_t hash) { + return hash % 326617llu; + } + static uint64_t mod411527(uint64_t hash) { + return hash % 411527llu; + } + static uint64_t mod518509(uint64_t hash) { + return hash % 518509llu; + } + static uint64_t mod653267(uint64_t hash) { + return hash % 653267llu; + } + static uint64_t mod823117(uint64_t hash) { + return hash % 823117llu; + } + static uint64_t mod1037059(uint64_t hash) { + return hash % 1037059llu; + } + static uint64_t mod1306601(uint64_t hash) { + return hash % 1306601llu; + } + static uint64_t mod1646237(uint64_t hash) { + return hash % 1646237llu; + } + static uint64_t mod2074129(uint64_t hash) { + return hash % 2074129llu; + } + static uint64_t mod2613229(uint64_t hash) { + return hash % 2613229llu; + } + static uint64_t mod3292489(uint64_t hash) { + return hash % 3292489llu; + } + static uint64_t mod4148279(uint64_t hash) { + return hash % 4148279llu; + } + static uint64_t mod5226491(uint64_t hash) { + return hash % 5226491llu; + } + static uint64_t mod6584983(uint64_t hash) { + return hash % 6584983llu; + } + static uint64_t mod8296553(uint64_t hash) { + return hash % 8296553llu; + } + static uint64_t mod10453007(uint64_t hash) { + return hash % 10453007llu; + } + static uint64_t mod13169977(uint64_t hash) { + return hash % 13169977llu; + } + static uint64_t mod16593127(uint64_t hash) { + return hash % 16593127llu; + } + static uint64_t mod20906033(uint64_t hash) { + return hash % 20906033llu; + } + static uint64_t mod26339969(uint64_t hash) { + return hash % 26339969llu; + } + static uint64_t mod33186281(uint64_t hash) { + return hash % 33186281llu; + } + static uint64_t mod41812097(uint64_t hash) { + return hash % 41812097llu; + } + static uint64_t mod52679969(uint64_t hash) { + return hash % 52679969llu; + } + static uint64_t mod66372617(uint64_t hash) { + return hash % 66372617llu; + } + static uint64_t mod83624237(uint64_t hash) { + return hash % 83624237llu; + } + static uint64_t mod105359939(uint64_t hash) { + return hash % 105359939llu; + } + static uint64_t mod132745199(uint64_t hash) { + return hash % 132745199llu; + } + static uint64_t mod167248483(uint64_t hash) { + return hash % 167248483llu; + } + static uint64_t mod210719881(uint64_t hash) { + return hash % 210719881llu; + } + static uint64_t mod265490441(uint64_t hash) { + return hash % 265490441llu; + } + static uint64_t mod334496971(uint64_t hash) { + return hash % 334496971llu; + } + static uint64_t mod421439783(uint64_t hash) { + return hash % 421439783llu; + } + static uint64_t mod530980861(uint64_t hash) { + return hash % 530980861llu; + } + static uint64_t mod668993977(uint64_t hash) { + return hash % 668993977llu; + } + static uint64_t mod842879579(uint64_t hash) { + return hash % 842879579llu; + } + static uint64_t mod1061961721(uint64_t hash) { + return hash % 1061961721llu; + } + static uint64_t mod1337987929(uint64_t hash) { + return hash % 1337987929llu; + } + static uint64_t mod1685759167(uint64_t hash) { + return hash % 1685759167llu; + } + static uint64_t mod2123923447(uint64_t hash) { + return hash % 2123923447llu; + } + static uint64_t mod2675975881(uint64_t hash) { + return hash % 2675975881llu; + } + static uint64_t mod3371518343(uint64_t hash) { + return hash % 3371518343llu; + } + static uint64_t mod4247846927(uint64_t hash) { + return hash % 4247846927llu; + } + static uint64_t mod5351951779(uint64_t hash) { + return hash % 5351951779llu; + } + static uint64_t mod6743036717(uint64_t hash) { + return hash % 6743036717llu; + } + static uint64_t mod8495693897(uint64_t hash) { + return hash % 8495693897llu; + } + static uint64_t mod10703903591(uint64_t hash) { + return hash % 10703903591llu; + } + static uint64_t mod13486073473(uint64_t hash) { + return hash % 13486073473llu; + } + static uint64_t mod16991387857(uint64_t hash) { + return hash % 16991387857llu; + } + static uint64_t mod21407807219(uint64_t hash) { + return hash % 21407807219llu; + } + static uint64_t mod26972146961(uint64_t hash) { + return hash % 26972146961llu; + } + static uint64_t mod33982775741(uint64_t hash) { + return hash % 33982775741llu; + } + static uint64_t mod42815614441(uint64_t hash) { + return hash % 42815614441llu; + } + static uint64_t mod53944293929(uint64_t hash) { + return hash % 53944293929llu; + } + static uint64_t mod67965551447(uint64_t hash) { + return hash % 67965551447llu; + } + static uint64_t mod85631228929(uint64_t hash) { + return hash % 85631228929llu; + } + static uint64_t mod107888587883(uint64_t hash) { + return hash % 107888587883llu; + } + static uint64_t mod135931102921(uint64_t hash) { + return hash % 135931102921llu; + } + static uint64_t mod171262457903(uint64_t hash) { + return hash % 171262457903llu; + } + static uint64_t mod215777175787(uint64_t hash) { + return hash % 215777175787llu; + } + static uint64_t mod271862205833(uint64_t hash) { + return hash % 271862205833llu; + } + static uint64_t mod342524915839(uint64_t hash) { + return hash % 342524915839llu; + } + static uint64_t mod431554351609(uint64_t hash) { + return hash % 431554351609llu; + } + static uint64_t mod543724411781(uint64_t hash) { + return hash % 543724411781llu; + } + static uint64_t mod685049831731(uint64_t hash) { + return hash % 685049831731llu; + } + static uint64_t mod863108703229(uint64_t hash) { + return hash % 863108703229llu; + } + static uint64_t mod1087448823553(uint64_t hash) { + return hash % 1087448823553llu; + } + static uint64_t mod1370099663459(uint64_t hash) { + return hash % 1370099663459llu; + } + static uint64_t mod1726217406467(uint64_t hash) { + return hash % 1726217406467llu; + } + static uint64_t mod2174897647073(uint64_t hash) { + return hash % 2174897647073llu; + } + static uint64_t mod2740199326961(uint64_t hash) { + return hash % 2740199326961llu; + } + static uint64_t mod3452434812973(uint64_t hash) { + return hash % 3452434812973llu; + } + static uint64_t mod4349795294267(uint64_t hash) { + return hash % 4349795294267llu; + } + static uint64_t mod5480398654009(uint64_t hash) { + return hash % 5480398654009llu; + } + static uint64_t mod6904869625999(uint64_t hash) { + return hash % 6904869625999llu; + } + static uint64_t mod8699590588571(uint64_t hash) { + return hash % 8699590588571llu; + } + static uint64_t mod10960797308051(uint64_t hash) { + return hash % 10960797308051llu; + } + static uint64_t mod13809739252051(uint64_t hash) { + return hash % 13809739252051llu; + } + static uint64_t mod17399181177241(uint64_t hash) { + return hash % 17399181177241llu; + } + static uint64_t mod21921594616111(uint64_t hash) { + return hash % 21921594616111llu; + } + static uint64_t mod27619478504183(uint64_t hash) { + return hash % 27619478504183llu; + } + static uint64_t mod34798362354533(uint64_t hash) { + return hash % 34798362354533llu; + } + static uint64_t mod43843189232363(uint64_t hash) { + return hash % 43843189232363llu; + } + static uint64_t mod55238957008387(uint64_t hash) { + return hash % 55238957008387llu; + } + static uint64_t mod69596724709081(uint64_t hash) { + return hash % 69596724709081llu; + } + static uint64_t mod87686378464759(uint64_t hash) { + return hash % 87686378464759llu; + } + static uint64_t mod110477914016779(uint64_t hash) { + return hash % 110477914016779llu; + } + static uint64_t mod139193449418173(uint64_t hash) { + return hash % 139193449418173llu; + } + static uint64_t mod175372756929481(uint64_t hash) { + return hash % 175372756929481llu; + } + static uint64_t mod220955828033581(uint64_t hash) { + return hash % 220955828033581llu; + } + static uint64_t mod278386898836457(uint64_t hash) { + return hash % 278386898836457llu; + } + static uint64_t mod350745513859007(uint64_t hash) { + return hash % 350745513859007llu; + } + static uint64_t mod441911656067171(uint64_t hash) { + return hash % 441911656067171llu; + } + static uint64_t mod556773797672909(uint64_t hash) { + return hash % 556773797672909llu; + } + static uint64_t mod701491027718027(uint64_t hash) { + return hash % 701491027718027llu; + } + static uint64_t mod883823312134381(uint64_t hash) { + return hash % 883823312134381llu; + } + static uint64_t mod1113547595345903(uint64_t hash) { + return hash % 1113547595345903llu; + } + static uint64_t mod1402982055436147(uint64_t hash) { + return hash % 1402982055436147llu; + } + static uint64_t mod1767646624268779(uint64_t hash) { + return hash % 1767646624268779llu; + } + static uint64_t mod2227095190691797(uint64_t hash) { + return hash % 2227095190691797llu; + } + static uint64_t mod2805964110872297(uint64_t hash) { + return hash % 2805964110872297llu; + } + static uint64_t mod3535293248537579(uint64_t hash) { + return hash % 3535293248537579llu; + } + static uint64_t mod4454190381383713(uint64_t hash) { + return hash % 4454190381383713llu; + } + static uint64_t mod5611928221744609(uint64_t hash) { + return hash % 5611928221744609llu; + } + static uint64_t mod7070586497075177(uint64_t hash) { + return hash % 7070586497075177llu; + } + static uint64_t mod8908380762767489(uint64_t hash) { + return hash % 8908380762767489llu; + } + static uint64_t mod11223856443489329(uint64_t hash) { + return hash % 11223856443489329llu; + } + static uint64_t mod14141172994150357(uint64_t hash) { + return hash % 14141172994150357llu; + } + static uint64_t mod17816761525534927(uint64_t hash) { + return hash % 17816761525534927llu; + } + static uint64_t mod22447712886978529(uint64_t hash) { + return hash % 22447712886978529llu; + } + static uint64_t mod28282345988300791(uint64_t hash) { + return hash % 28282345988300791llu; + } + static uint64_t mod35633523051069991(uint64_t hash) { + return hash % 35633523051069991llu; + } + static uint64_t mod44895425773957261(uint64_t hash) { + return hash % 44895425773957261llu; + } + static uint64_t mod56564691976601587(uint64_t hash) { + return hash % 56564691976601587llu; + } + static uint64_t mod71267046102139967(uint64_t hash) { + return hash % 71267046102139967llu; + } + static uint64_t mod89790851547914507(uint64_t hash) { + return hash % 89790851547914507llu; + } + static uint64_t mod113129383953203213(uint64_t hash) { + return hash % 113129383953203213llu; + } + static uint64_t mod142534092204280003(uint64_t hash) { + return hash % 142534092204280003llu; + } + static uint64_t mod179581703095829107(uint64_t hash) { + return hash % 179581703095829107llu; + } + static uint64_t mod226258767906406483(uint64_t hash) { + return hash % 226258767906406483llu; + } + static uint64_t mod285068184408560057(uint64_t hash) { + return hash % 285068184408560057llu; + } + static uint64_t mod359163406191658253(uint64_t hash) { + return hash % 359163406191658253llu; + } + static uint64_t mod452517535812813007(uint64_t hash) { + return hash % 452517535812813007llu; + } + static uint64_t mod570136368817120201(uint64_t hash) { + return hash % 570136368817120201llu; + } + static uint64_t mod718326812383316683(uint64_t hash) { + return hash % 718326812383316683llu; + } + static uint64_t mod905035071625626043(uint64_t hash) { + return hash % 905035071625626043llu; + } + static uint64_t mod1140272737634240411(uint64_t hash) { + return hash % 1140272737634240411llu; + } + static uint64_t mod1436653624766633509(uint64_t hash) { + return hash % 1436653624766633509llu; + } + static uint64_t mod1810070143251252131(uint64_t hash) { + return hash % 1810070143251252131llu; + } + static uint64_t mod2280545475268481167(uint64_t hash) { + return hash % 2280545475268481167llu; + } + static uint64_t mod2873307249533267101(uint64_t hash) { + return hash % 2873307249533267101llu; + } + static uint64_t mod3620140286502504283(uint64_t hash) { + return hash % 3620140286502504283llu; + } + static uint64_t mod4561090950536962147(uint64_t hash) { + return hash % 4561090950536962147llu; + } + static uint64_t mod5746614499066534157(uint64_t hash) { + return hash % 5746614499066534157llu; + } + static uint64_t mod7240280573005008577(uint64_t hash) { + return hash % 7240280573005008577llu; + } + static uint64_t mod9122181901073924329(uint64_t hash) { + return hash % 9122181901073924329llu; + } + static uint64_t mod11493228998133068689(uint64_t hash) { + return hash % 11493228998133068689llu; + } + static uint64_t mod14480561146010017169(uint64_t hash) { + return hash % 14480561146010017169llu; + } + static uint64_t mod18446744073709551557(uint64_t hash) { + return hash % 18446744073709551557llu; + } + + using mod_function = uint64_t (*)(uint64_t); + + mod_function next_size_over(uint64_t& size) const { + // prime numbers generated by the following method: + // 1. start with a prime p = 2 + // 2. go to wolfram alpha and get p = NextPrime(2 * p) + // 3. repeat 2. until you overflow 64 bits + // you now have large gaps which you would hit if somebody called reserve() + // with an unlucky number. + // 4. to fill the gaps for every prime p go to wolfram alpha and get + // ClosestPrime(p * 2^(1/3)) and ClosestPrime(p * 2^(2/3)) and put those in + // the gaps + // 5. get PrevPrime(2^64) and put it at the end + // NOLINTNEXTLINE(*c-array*) + static constexpr const uint64_t prime_list[] = { + 2llu, + 3llu, + 5llu, + 7llu, + 11llu, + 13llu, + 17llu, + 23llu, + 29llu, + 37llu, + 47llu, + 59llu, + 73llu, + 97llu, + 127llu, + 151llu, + 197llu, + 251llu, + 313llu, + 397llu, + 499llu, + 631llu, + 797llu, + 1009llu, + 1259llu, + 1597llu, + 2011llu, + 2539llu, + 3203llu, + 4027llu, + 5087llu, + 6421llu, + 8089llu, + 10193llu, + 12853llu, + 16193llu, + 20399llu, + 25717llu, + 32401llu, + 40823llu, + 51437llu, + 64811llu, + 81649llu, + 102877llu, + 129607llu, + 163307llu, + 205759llu, + 259229llu, + 326617llu, + 411527llu, + 518509llu, + 653267llu, + 823117llu, + 1037059llu, + 1306601llu, + 1646237llu, + 2074129llu, + 2613229llu, + 3292489llu, + 4148279llu, + 5226491llu, + 6584983llu, + 8296553llu, + 10453007llu, + 13169977llu, + 16593127llu, + 20906033llu, + 26339969llu, + 33186281llu, + 41812097llu, + 52679969llu, + 66372617llu, + 83624237llu, + 105359939llu, + 132745199llu, + 167248483llu, + 210719881llu, + 265490441llu, + 334496971llu, + 421439783llu, + 530980861llu, + 668993977llu, + 842879579llu, + 1061961721llu, + 1337987929llu, + 1685759167llu, + 2123923447llu, + 2675975881llu, + 3371518343llu, + 4247846927llu, + 5351951779llu, + 6743036717llu, + 8495693897llu, + 10703903591llu, + 13486073473llu, + 16991387857llu, + 21407807219llu, + 26972146961llu, + 33982775741llu, + 42815614441llu, + 53944293929llu, + 67965551447llu, + 85631228929llu, + 107888587883llu, + 135931102921llu, + 171262457903llu, + 215777175787llu, + 271862205833llu, + 342524915839llu, + 431554351609llu, + 543724411781llu, + 685049831731llu, + 863108703229llu, + 1087448823553llu, + 1370099663459llu, + 1726217406467llu, + 2174897647073llu, + 2740199326961llu, + 3452434812973llu, + 4349795294267llu, + 5480398654009llu, + 6904869625999llu, + 8699590588571llu, + 10960797308051llu, + 13809739252051llu, + 17399181177241llu, + 21921594616111llu, + 27619478504183llu, + 34798362354533llu, + 43843189232363llu, + 55238957008387llu, + 69596724709081llu, + 87686378464759llu, + 110477914016779llu, + 139193449418173llu, + 175372756929481llu, + 220955828033581llu, + 278386898836457llu, + 350745513859007llu, + 441911656067171llu, + 556773797672909llu, + 701491027718027llu, + 883823312134381llu, + 1113547595345903llu, + 1402982055436147llu, + 1767646624268779llu, + 2227095190691797llu, + 2805964110872297llu, + 3535293248537579llu, + 4454190381383713llu, + 5611928221744609llu, + 7070586497075177llu, + 8908380762767489llu, + 11223856443489329llu, + 14141172994150357llu, + 17816761525534927llu, + 22447712886978529llu, + 28282345988300791llu, + 35633523051069991llu, + 44895425773957261llu, + 56564691976601587llu, + 71267046102139967llu, + 89790851547914507llu, + 113129383953203213llu, + 142534092204280003llu, + 179581703095829107llu, + 226258767906406483llu, + 285068184408560057llu, + 359163406191658253llu, + 452517535812813007llu, + 570136368817120201llu, + 718326812383316683llu, + 905035071625626043llu, + 1140272737634240411llu, + 1436653624766633509llu, + 1810070143251252131llu, + 2280545475268481167llu, + 2873307249533267101llu, + 3620140286502504283llu, + 4561090950536962147llu, + 5746614499066534157llu, + 7240280573005008577llu, + 9122181901073924329llu, + 11493228998133068689llu, + 14480561146010017169llu, + 18446744073709551557llu}; + // NOLINTNEXTLINE(*c-array*) + static constexpr uint64_t (*const mod_functions[])(uint64_t) = { + &mod0, + &mod2, + &mod3, + &mod5, + &mod7, + &mod11, + &mod13, + &mod17, + &mod23, + &mod29, + &mod37, + &mod47, + &mod59, + &mod73, + &mod97, + &mod127, + &mod151, + &mod197, + &mod251, + &mod313, + &mod397, + &mod499, + &mod631, + &mod797, + &mod1009, + &mod1259, + &mod1597, + &mod2011, + &mod2539, + &mod3203, + &mod4027, + &mod5087, + &mod6421, + &mod8089, + &mod10193, + &mod12853, + &mod16193, + &mod20399, + &mod25717, + &mod32401, + &mod40823, + &mod51437, + &mod64811, + &mod81649, + &mod102877, + &mod129607, + &mod163307, + &mod205759, + &mod259229, + &mod326617, + &mod411527, + &mod518509, + &mod653267, + &mod823117, + &mod1037059, + &mod1306601, + &mod1646237, + &mod2074129, + &mod2613229, + &mod3292489, + &mod4148279, + &mod5226491, + &mod6584983, + &mod8296553, + &mod10453007, + &mod13169977, + &mod16593127, + &mod20906033, + &mod26339969, + &mod33186281, + &mod41812097, + &mod52679969, + &mod66372617, + &mod83624237, + &mod105359939, + &mod132745199, + &mod167248483, + &mod210719881, + &mod265490441, + &mod334496971, + &mod421439783, + &mod530980861, + &mod668993977, + &mod842879579, + &mod1061961721, + &mod1337987929, + &mod1685759167, + &mod2123923447, + &mod2675975881, + &mod3371518343, + &mod4247846927, + &mod5351951779, + &mod6743036717, + &mod8495693897, + &mod10703903591, + &mod13486073473, + &mod16991387857, + &mod21407807219, + &mod26972146961, + &mod33982775741, + &mod42815614441, + &mod53944293929, + &mod67965551447, + &mod85631228929, + &mod107888587883, + &mod135931102921, + &mod171262457903, + &mod215777175787, + &mod271862205833, + &mod342524915839, + &mod431554351609, + &mod543724411781, + &mod685049831731, + &mod863108703229, + &mod1087448823553, + &mod1370099663459, + &mod1726217406467, + &mod2174897647073, + &mod2740199326961, + &mod3452434812973, + &mod4349795294267, + &mod5480398654009, + &mod6904869625999, + &mod8699590588571, + &mod10960797308051, + &mod13809739252051, + &mod17399181177241, + &mod21921594616111, + &mod27619478504183, + &mod34798362354533, + &mod43843189232363, + &mod55238957008387, + &mod69596724709081, + &mod87686378464759, + &mod110477914016779, + &mod139193449418173, + &mod175372756929481, + &mod220955828033581, + &mod278386898836457, + &mod350745513859007, + &mod441911656067171, + &mod556773797672909, + &mod701491027718027, + &mod883823312134381, + &mod1113547595345903, + &mod1402982055436147, + &mod1767646624268779, + &mod2227095190691797, + &mod2805964110872297, + &mod3535293248537579, + &mod4454190381383713, + &mod5611928221744609, + &mod7070586497075177, + &mod8908380762767489, + &mod11223856443489329, + &mod14141172994150357, + &mod17816761525534927, + &mod22447712886978529, + &mod28282345988300791, + &mod35633523051069991, + &mod44895425773957261, + &mod56564691976601587, + &mod71267046102139967, + &mod89790851547914507, + &mod113129383953203213, + &mod142534092204280003, + &mod179581703095829107, + &mod226258767906406483, + &mod285068184408560057, + &mod359163406191658253, + &mod452517535812813007, + &mod570136368817120201, + &mod718326812383316683, + &mod905035071625626043, + &mod1140272737634240411, + &mod1436653624766633509, + &mod1810070143251252131, + &mod2280545475268481167, + &mod2873307249533267101, + &mod3620140286502504283, + &mod4561090950536962147, + &mod5746614499066534157, + &mod7240280573005008577, + &mod9122181901073924329, + &mod11493228998133068689, + &mod14480561146010017169, + &mod18446744073709551557}; + const uint64_t* found = std::lower_bound( + std::begin(prime_list), std::end(prime_list) - 1, size); + size = *found; + return mod_functions[1 + found - prime_list]; + } + void commit(mod_function new_mod_function) { + current_mod_function = new_mod_function; + } + void reset() { + current_mod_function = &mod0; + } + + uint64_t index_for_hash(uint64_t hash, uint64_t /*num_slots_minus_one*/) + const { + return current_mod_function(hash); + } + uint64_t keep_in_range(uint64_t index, uint64_t num_slots_minus_one) const { + return index > num_slots_minus_one ? current_mod_function(index) : index; + } + + private: + mod_function current_mod_function = &mod0; +}; + +struct power_of_two_hash_policy { + uint64_t index_for_hash(uint64_t hash, uint64_t num_slots_minus_one) const { + return hash & num_slots_minus_one; + } + uint64_t keep_in_range(uint64_t index, uint64_t num_slots_minus_one) const { + return index_for_hash(index, num_slots_minus_one); + } + int8_t next_size_over(uint64_t& size) const { + size = detailv3::next_power_of_two(size); + return 0; + } + void commit(int8_t /*unused*/) {} + void reset() {} +}; + +struct fibonacci_hash_policy { + uint64_t index_for_hash(uint64_t hash, uint64_t /*num_slots_minus_one*/) + const { + return (11400714819323198485ull * hash) >> shift; + } + uint64_t keep_in_range(uint64_t index, uint64_t num_slots_minus_one) const { + return index & num_slots_minus_one; + } + + int8_t next_size_over(uint64_t& size) const { + size = std::max(uint64_t(2), detailv3::next_power_of_two(size)); + return static_cast(64 - detailv3::log2(size)); + } + void commit(int8_t shift_) { + shift = shift_; + } + void reset() { + shift = 63; + } + + private: + int8_t shift = 63; +}; + +template < + typename K, + typename V, + typename H = std::hash, + typename E = std::equal_to, + typename A = std::allocator>> +class order_preserving_flat_hash_map + : public detailv3::sherwood_v3_table< + std::pair, + K, + H, + detailv3::KeyOrValueHasher, H>, + E, + detailv3::KeyOrValueEquality, E>, + A, + typename std::allocator_traits::template rebind_alloc< + detailv3::sherwood_v3_entry>>> { + using Table = detailv3::sherwood_v3_table< + std::pair, + K, + H, + detailv3::KeyOrValueHasher, H>, + E, + detailv3::KeyOrValueEquality, E>, + A, + typename std::allocator_traits::template rebind_alloc< + detailv3::sherwood_v3_entry>>>; + + public: + using key_type = K; + using mapped_type = V; + + using Table::Table; + order_preserving_flat_hash_map() = default; + + inline V& operator[](const K& key) { + return emplace(key, convertible_to_value()).first->second; + } + inline V& operator[](K&& key) { + return emplace(std::move(key), convertible_to_value()).first->second; + } + V& at(const K& key) { + auto found = this->find(key); + if (found == this->end()) + throw std::out_of_range("Argument passed to at() was not in the map."); + return found->second; + } + const V& at(const K& key) const { + auto found = this->find(key); + if (found == this->end()) + throw std::out_of_range("Argument passed to at() was not in the map."); + return found->second; + } + + using Table::emplace; + std::pair emplace() { + return emplace(key_type(), convertible_to_value()); + } + template + std::pair insert_or_assign( + const key_type& key, + M&& m) { + auto emplace_result = emplace(key, std::forward(m)); + if (!emplace_result.second) + emplace_result.first->second = std::forward(m); + return emplace_result; + } + template + std::pair insert_or_assign( + key_type&& key, + M&& m) { + auto emplace_result = emplace(std::move(key), std::forward(m)); + if (!emplace_result.second) + emplace_result.first->second = std::forward(m); + return emplace_result; + } + template + typename Table::iterator insert_or_assign( + typename Table::const_iterator /*unused*/, + const key_type& key, + M&& m) { + return insert_or_assign(key, std::forward(m)).first; + } + template + typename Table::iterator insert_or_assign( + typename Table::const_iterator /*unused*/, + key_type&& key, + M&& m) { + return insert_or_assign(std::move(key), std::forward(m)).first; + } + + friend bool operator==( + const order_preserving_flat_hash_map& lhs, + const order_preserving_flat_hash_map& rhs) { + if (lhs.size() != rhs.size()) + return false; + for (const typename Table::value_type& value : lhs) { + auto found = rhs.find(value.first); + if (found == rhs.end() || value.second != found->second) + return false; + } + return true; + } + friend bool operator!=( + const order_preserving_flat_hash_map& lhs, + const order_preserving_flat_hash_map& rhs) { + return !(lhs == rhs); + } + + private: + struct convertible_to_value { + operator V() const { + return V(); + } + }; +}; + +template < + typename T, + typename H = std::hash, + typename E = std::equal_to, + typename A = std::allocator> +class flat_hash_set + : public detailv3::sherwood_v3_table< + T, + T, + H, + detailv3::functor_storage, + E, + detailv3::functor_storage, + A, + typename std::allocator_traits::template rebind_alloc< + detailv3::sherwood_v3_entry>> { + using Table = detailv3::sherwood_v3_table< + T, + T, + H, + detailv3::functor_storage, + E, + detailv3::functor_storage, + A, + typename std::allocator_traits::template rebind_alloc< + detailv3::sherwood_v3_entry>>; + + public: + using key_type = T; + + using Table::Table; + flat_hash_set() = default; + + template + std::pair emplace(Args&&... args) { + return Table::emplace(T(std::forward(args)...)); + } + std::pair emplace(const key_type& arg) { + return Table::emplace(arg); + } + std::pair emplace(key_type& arg) { + return Table::emplace(arg); + } + std::pair emplace(const key_type&& arg) { + return Table::emplace(std::move(arg)); + } + std::pair emplace(key_type&& arg) { + return Table::emplace(std::move(arg)); + } + + friend bool operator==(const flat_hash_set& lhs, const flat_hash_set& rhs) { + if (lhs.size() != rhs.size()) + return false; + for (const T& value : lhs) { + if (rhs.find(value) == rhs.end()) + return false; + } + return true; + } + friend bool operator!=(const flat_hash_set& lhs, const flat_hash_set& rhs) { + return !(lhs == rhs); + } +}; + +template +struct power_of_two_std_hash : std::hash { + typedef ska_ordered::power_of_two_hash_policy hash_policy; +}; + +} // namespace ska_ordered + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/overflows.h b/backends/pytorch/include/c10/util/overflows.h new file mode 100644 index 000000000..e414de5aa --- /dev/null +++ b/backends/pytorch/include/c10/util/overflows.h @@ -0,0 +1,105 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include + +#include +#include +#include + +namespace c10 { +// In some versions of MSVC, there will be a compiler error when building. +// C4146: unary minus operator applied to unsigned type, result still unsigned +// C4804: unsafe use of type 'bool' in operation +// It can be addressed by disabling the following warning. +#ifdef _MSC_VER +#pragma warning(push) +#pragma warning(disable : 4146) +#pragma warning(disable : 4804) +#pragma warning(disable : 4018) +#endif + +// The overflow checks may involve float to int conversion which may +// trigger precision loss warning. Re-enable the warning once the code +// is fixed. See T58053069. +C10_CLANG_DIAGNOSTIC_PUSH() +#if C10_CLANG_HAS_WARNING("-Wimplicit-float-conversion") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-float-conversion") +#endif + +// bool can be converted to any type. +// Without specializing on bool, in pytorch_linux_trusty_py2_7_9_build: +// `error: comparison of constant '255' with boolean expression is always false` +// for `f > limit::max()` below +template +std::enable_if_t, bool> overflows( + From /*f*/, + bool strict_unsigned [[maybe_unused]] = false) { + return false; +} + +// skip isnan and isinf check for integral types +template +std::enable_if_t && !std::is_same_v, bool> +overflows(From f, bool strict_unsigned = false) { + using limit = std::numeric_limits::type>; + if constexpr (!limit::is_signed && std::numeric_limits::is_signed) { + // allow for negative numbers to wrap using two's complement arithmetic. + // For example, with uint8, this allows for `a - b` to be treated as + // `a + 255 * b`. + if (!strict_unsigned) { + return greater_than_max(f) || + (c10::is_negative(f) && + -static_cast(f) > static_cast(limit::max())); + } + } + return c10::less_than_lowest(f) || greater_than_max(f); +} + +template +std::enable_if_t, bool> overflows( + From f, + bool strict_unsigned [[maybe_unused]] = false) { + using limit = std::numeric_limits::type>; + if (limit::has_infinity && std::isinf(static_cast(f))) { + return false; + } + if (!limit::has_quiet_NaN && (f != f)) { + return true; + } + return f < limit::lowest() || f > limit::max(); +} + +C10_CLANG_DIAGNOSTIC_POP() + +#ifdef _MSC_VER +#pragma warning(pop) +#endif + +template +std::enable_if_t::value, bool> overflows( + From f, + bool strict_unsigned = false) { + // casts from complex to real are considered to overflow if the + // imaginary component is non-zero + if (!is_complex::value && f.imag() != 0) { + return true; + } + // Check for overflow componentwise + // (Technically, the imag overflow check is guaranteed to be false + // when !is_complex, but any optimizer worth its salt will be + // able to figure it out.) + return overflows< + typename scalar_value_type::type, + typename From::value_type>(f.real(), strict_unsigned) || + overflows< + typename scalar_value_type::type, + typename From::value_type>(f.imag(), strict_unsigned); +} +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/python_stub.h b/backends/pytorch/include/c10/util/python_stub.h new file mode 100644 index 000000000..f457be594 --- /dev/null +++ b/backends/pytorch/include/c10/util/python_stub.h @@ -0,0 +1,9 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +struct _object; +using PyObject = _object; + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/qint32.h b/backends/pytorch/include/c10/util/qint32.h new file mode 100644 index 000000000..2b48a5a89 --- /dev/null +++ b/backends/pytorch/include/c10/util/qint32.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/qint8.h b/backends/pytorch/include/c10/util/qint8.h new file mode 100644 index 000000000..47f7a9e42 --- /dev/null +++ b/backends/pytorch/include/c10/util/qint8.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/quint2x4.h b/backends/pytorch/include/c10/util/quint2x4.h new file mode 100644 index 000000000..b7781bc57 --- /dev/null +++ b/backends/pytorch/include/c10/util/quint2x4.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/quint4x2.h b/backends/pytorch/include/c10/util/quint4x2.h new file mode 100644 index 000000000..b4603a707 --- /dev/null +++ b/backends/pytorch/include/c10/util/quint4x2.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/quint8.h b/backends/pytorch/include/c10/util/quint8.h new file mode 100644 index 000000000..5445be709 --- /dev/null +++ b/backends/pytorch/include/c10/util/quint8.h @@ -0,0 +1,6 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#include + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/safe_numerics.h b/backends/pytorch/include/c10/util/safe_numerics.h new file mode 100644 index 000000000..f376f9dfd --- /dev/null +++ b/backends/pytorch/include/c10/util/safe_numerics.h @@ -0,0 +1,119 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once +#include + +#include +#include +#include + +// GCC has __builtin_mul_overflow from before it supported __has_builtin +#ifdef _MSC_VER +#define C10_HAS_BUILTIN_OVERFLOW() (0) +#include +#include +#else +#define C10_HAS_BUILTIN_OVERFLOW() (1) +#endif + +namespace c10 { + +template , int> = 0> +C10_ALWAYS_INLINE bool add_overflows(T a, T b, T* out) { +#if C10_HAS_BUILTIN_OVERFLOW() + return __builtin_add_overflow(a, b, out); +#else + if constexpr (std::is_signed_v) { + // For signed types, detect overflow by checking sign changes + volatile T tmp = a + b; + *out = tmp; + + // If both operands have the same sign, check if result changed sign + // unexpectedly. + if ((a > 0) == (b > 0)) { + if ((a > 0) && (tmp <= 0)) { + return true; // Positive overflow + } + if ((a < 0) && (tmp >= 0)) { + return true; // Negative overflow + } + } + return false; + } else { + // For unsigned types, overflow causes wrap-around + volatile T tmp = a + b; + *out = tmp; + return (tmp < a || tmp < b); + } +#endif +} + +C10_ALWAYS_INLINE bool add_overflows(uint64_t a, uint64_t b, uint64_t* out) { + return add_overflows(a, b, out); +} + +template , int> = 0> +C10_ALWAYS_INLINE bool mul_overflows(T a, T b, T* out) { +#if C10_HAS_BUILTIN_OVERFLOW() + return __builtin_mul_overflow(a, b, out); +#else + if constexpr (std::is_signed_v) { + // For signed types, use the division-based check + volatile T tmp = a * b; + *out = tmp; + if (a == 0 || b == 0) { + return false; + } + return !(a == tmp / b); + } else { + // For unsigned types, use leading zeros approach + // This test isn't exact, but avoids doing integer division + *out = a * b; + constexpr int bits = sizeof(T) * 8; + return ( + (c10::llvm::countLeadingZeros(a) + c10::llvm::countLeadingZeros(b)) < + bits); + } +#endif +} + +C10_ALWAYS_INLINE bool mul_overflows(uint64_t a, uint64_t b, uint64_t* out) { + return mul_overflows(a, b, out); +} + +template +bool safe_multiplies_u64(It first, It last, uint64_t* out) { +#if C10_HAS_BUILTIN_OVERFLOW() + uint64_t prod = 1; + bool overflow = false; + for (; first != last; ++first) { + overflow |= c10::mul_overflows(prod, *first, &prod); + } + *out = prod; + return overflow; +#else + uint64_t prod = 1; + uint64_t prod_log2 = 0; + bool is_zero = false; + for (; first != last; ++first) { + auto x = static_cast(*first); + prod *= x; + // log2(0) isn't valid, so need to track it specially + is_zero |= (x == 0); + prod_log2 += c10::llvm::Log2_64_Ceil(x); + } + *out = prod; + // This test isn't exact, but avoids doing integer division + return !is_zero && (prod_log2 >= 64); +#endif +} + +template +bool safe_multiplies_u64(const Container& c, uint64_t* out) { + return safe_multiplies_u64(c.begin(), c.end(), out); +} + +} // namespace c10 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/string_utils.h b/backends/pytorch/include/c10/util/string_utils.h new file mode 100644 index 000000000..cbcf0b1f3 --- /dev/null +++ b/backends/pytorch/include/c10/util/string_utils.h @@ -0,0 +1,27 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include + +#if !defined(FBCODE_CAFFE2) && !defined(C10_NO_DEPRECATED) + +namespace c10 { + +// NOLINTNEXTLINE(misc-unused-using-decls) +using std::stod; +// NOLINTNEXTLINE(misc-unused-using-decls) +using std::stoi; +// NOLINTNEXTLINE(misc-unused-using-decls) +using std::stoll; +// NOLINTNEXTLINE(misc-unused-using-decls) +using std::stoull; +// NOLINTNEXTLINE(misc-unused-using-decls) +using std::to_string; + +} // namespace c10 + +#endif + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/string_view.h b/backends/pytorch/include/c10/util/string_view.h new file mode 100644 index 000000000..559cde09f --- /dev/null +++ b/backends/pytorch/include/c10/util/string_view.h @@ -0,0 +1,648 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include + +namespace c10 { + +/** + * Port of std::string_view with methods from C++20. + * Implemented following the interface definition in + * https://en.cppreference.com/w/cpp/string/basic_string_view + * See there for the API documentation. + * + * Difference: We don't have a Traits template parameter because + * std::char_traits isn't constexpr and we'd have to reimplement + * std::char_traits if we wanted to use it with our constexpr basic_string_view. + */ +template +// NOLINTNEXTLINE(cppcoreguidelines-special-member-functions) +class basic_string_view final { + public: + using value_type = CharT; + using pointer = CharT*; + using const_pointer = const CharT*; + using reference = CharT&; + using const_reference = const CharT&; + using const_iterator = const CharT*; + using iterator = const_iterator; + using const_reverse_iterator = std::reverse_iterator; + using reverse_iterator = const_reverse_iterator; + using size_type = std::size_t; + using difference_type = std::ptrdiff_t; + + static constexpr size_type npos = size_type(-1); + + constexpr basic_string_view() noexcept : begin_(nullptr) {} + + explicit constexpr basic_string_view(const_pointer str, size_type count) + : begin_(str), size_(count) {} + + /* implicit */ constexpr basic_string_view(const_pointer str) + : basic_string_view(str, strlen_(str)) {} + + /* implicit */ basic_string_view(const ::std::basic_string& str) + : basic_string_view(str.data(), str.size()) {} + + /* implicit */ constexpr basic_string_view( + const ::std::basic_string_view& str) + : basic_string_view(str.data(), str.size()) {} + + constexpr basic_string_view(const basic_string_view&) noexcept = default; + + constexpr basic_string_view& operator=( + const basic_string_view& rhs) noexcept = default; + + constexpr operator ::std::basic_string_view() const { + return ::std::basic_string_view(data(), size()); + } + + explicit operator ::std::basic_string() const { + return ::std::basic_string(data(), size()); + } + + constexpr const_iterator begin() const noexcept { + return cbegin(); + } + + constexpr const_iterator cbegin() const noexcept { + return begin_; + } + + constexpr const_iterator end() const noexcept { + return cend(); + } + + constexpr const_iterator cend() const noexcept { + return begin_ + size_; + } + + constexpr const_reverse_iterator rbegin() const noexcept { + return crbegin(); + } + + constexpr const_reverse_iterator crbegin() const noexcept { + return const_reverse_iterator(this->end()); + } + + constexpr const_reverse_iterator rend() const noexcept { + return crend(); + } + + constexpr const_reverse_iterator crend() const noexcept { + return const_reverse_iterator(this->begin()); + } + + friend constexpr const_iterator begin(basic_string_view sv) noexcept { + return sv.begin(); + } + + friend constexpr const_iterator end(basic_string_view sv) noexcept { + return sv.end(); + } + + constexpr const_reference operator[](size_type pos) const { + // TODO: split out + return at_(pos); + } + + constexpr const_reference at(size_type pos) const { +#if !defined( \ + __CUDA_ARCH__) // CUDA doesn't like std::out_of_range in device code + return C10_UNLIKELY(pos >= size_) + ? (throw std::out_of_range( + "string_view::operator[] or string_view::at() out of range. Index: " + + std::to_string(pos) + ", size: " + std::to_string(size())), + at_(0)) + : at_(pos); +#else + return at_(pos); +#endif + } + + constexpr const_reference front() const { + return *begin_; + } + + constexpr const_reference back() const { + return *(begin_ + size_ - 1); + } + + constexpr const_pointer data() const noexcept { + return begin_; + } + + constexpr size_type size() const noexcept { + return size_; + } + + constexpr size_type length() const noexcept { + return size(); + } + + constexpr size_type max_size() const noexcept { + return std::numeric_limits::max(); + } + + [[nodiscard]] constexpr bool empty() const noexcept { + return size() == 0; + } + + constexpr void remove_prefix(size_type n) { + if (n > size()) { + throw std::out_of_range( + "basic_string_view::remove_prefix: out of range. PrefixLength: " + + std::to_string(n) + ", size: " + std::to_string(size())); + } + begin_ += n; + size_ -= n; + } + + constexpr void remove_suffix(size_type n) { + if (n > size()) { + throw std::out_of_range( + "basic_string_view::remove_suffix: out of range. SuffixLength: " + + std::to_string(n) + ", size: " + std::to_string(size())); + } + size_ -= n; + } + + constexpr void swap(basic_string_view& sv) noexcept { + auto tmp = *this; + *this = sv; + sv = tmp; + } + + size_type copy(pointer dest, size_type count, size_type pos = 0) const { + if (pos > size_) { + throw std::out_of_range( + "basic_string_view::copy: out of range. Index: " + + std::to_string(pos) + ", size: " + std::to_string(size())); + } + size_type copy_length = std::min(count, size_ - pos); + for (auto iter = begin() + pos, end = iter + copy_length; iter != end;) { + *(dest++) = *(iter++); + } + return copy_length; + } + + constexpr basic_string_view substr(size_type pos = 0, size_type count = npos) + const { +#if !defined( \ + __CUDA_ARCH__) // CUDA doesn't like std::out_of_range in device code + return (pos > size_) + ? (throw std::out_of_range( + "basic_string_view::substr parameter out of bounds. Index: " + + std::to_string(pos) + ", size: " + std::to_string(size())), + substr_()) + : substr_(pos, count); +#else + return substr_(pos, count); +#endif + } + + constexpr int compare(basic_string_view rhs) const noexcept { + // Write it iteratively. This is faster. + for (size_t i = 0, end = std::min(size(), rhs.size()); i < end; ++i) { + if (at_(i) < rhs.at_(i)) { + return -1; + } else if (at_(i) > rhs.at_(i)) { + return 1; + } + } + if (size() < rhs.size()) { + return -1; + } else if (size() > rhs.size()) { + return 1; + } + return 0; + } + + constexpr int compare(size_type pos1, size_type count1, basic_string_view v) + const { + return substr(pos1, count1).compare(v); + } + + constexpr int compare( + size_type pos1, + size_type count1, + basic_string_view v, + size_type pos2, + size_type count2) const { + return substr(pos1, count1).compare(v.substr(pos2, count2)); + } + + constexpr int compare(const_pointer s) const { + return compare(basic_string_view(s)); + } + + constexpr int compare(size_type pos1, size_type count1, const_pointer s) + const { + return substr(pos1, count1).compare(basic_string_view(s)); + } + + constexpr int compare( + size_type pos1, + size_type count1, + const_pointer s, + size_type count2) const { + return substr(pos1, count1).compare(basic_string_view(s, count2)); + } + + friend constexpr bool operator==( + basic_string_view lhs, + basic_string_view rhs) noexcept { + return lhs.equals_(rhs); + } + + friend constexpr bool operator!=( + basic_string_view lhs, + basic_string_view rhs) noexcept { + return !(lhs == rhs); + } + + friend constexpr bool operator<( + basic_string_view lhs, + basic_string_view rhs) noexcept { + return lhs.compare(rhs) < 0; + } + + friend constexpr bool operator>=( + basic_string_view lhs, + basic_string_view rhs) noexcept { + return !(lhs < rhs); + } + + friend constexpr bool operator>( + basic_string_view lhs, + basic_string_view rhs) noexcept { + return rhs < lhs; + } + + friend constexpr bool operator<=( + basic_string_view lhs, + basic_string_view rhs) noexcept { + return !(lhs > rhs); + } + + constexpr bool starts_with(basic_string_view prefix) const noexcept { + return (prefix.size() > size()) ? false + : prefix.equals_(substr_(0, prefix.size())); + } + + constexpr bool starts_with(CharT prefix) const noexcept { + return !empty() && prefix == front(); + } + + constexpr bool starts_with(const_pointer prefix) const { + return starts_with(basic_string_view(prefix)); + } + + constexpr bool ends_with(basic_string_view suffix) const noexcept { + return (suffix.size() > size()) + ? false + : suffix.equals_(substr_(size() - suffix.size(), suffix.size())); + } + + constexpr bool ends_with(CharT suffix) const noexcept { + return !empty() && suffix == back(); + } + + constexpr bool ends_with(const_pointer suffix) const { + return ends_with(basic_string_view(suffix)); + } + + constexpr size_type find(basic_string_view v, size_type pos = 0) + const noexcept { + if (v.empty()) { + return pos <= size() ? pos : npos; + } + + if (pos + v.size() <= size()) { + for (size_type cur = pos, end = size() - v.size(); cur <= end; ++cur) { + if (v.at_(0) == at_(cur) && + v.substr_(1).equals_(substr_(cur + 1, v.size() - 1))) { + return cur; + } + } + } + return npos; + } + + constexpr size_type find(CharT ch, size_type pos = 0) const noexcept { + return find_first_if_(pos, charIsEqual_{ch}); + } + + constexpr size_type find(const_pointer s, size_type pos, size_type count) + const { + return find(basic_string_view(s, count), pos); + } + + constexpr size_type find(const_pointer s, size_type pos = 0) const { + return find(basic_string_view(s), pos); + } + + constexpr size_type rfind(basic_string_view v, size_type pos = npos) + const noexcept { + // Write it iteratively. This is faster. + if (v.empty()) { + return pos <= size() ? pos : size(); + } + + if (v.size() <= size()) { + pos = std::min(size() - v.size(), pos); + do { + if (v.at_(0) == at_(pos) && + v.substr_(1).equals_(substr_(pos + 1, v.size() - 1))) { + return pos; + } + } while (pos-- > 0); + } + return npos; + } + + constexpr size_type rfind(CharT ch, size_type pos = npos) const noexcept { + return find_last_if_(pos, charIsEqual_{ch}); + } + + constexpr size_type rfind(const_pointer s, size_type pos, size_type count) + const { + return rfind(basic_string_view(s, count), pos); + } + + constexpr size_type rfind(const_pointer s, size_type pos = npos) const { + return rfind(basic_string_view(s), pos); + } + + constexpr size_type find_first_of(basic_string_view v, size_type pos = 0) + const noexcept { + return find_first_if_(pos, stringViewContainsChar_{v}); + } + + constexpr size_type find_first_of(CharT ch, size_type pos = 0) + const noexcept { + return find_first_if_(pos, charIsEqual_{ch}); + } + + constexpr size_type find_first_of( + const_pointer s, + size_type pos, + size_type count) const { + return find_first_of(basic_string_view(s, count), pos); + } + + constexpr size_type find_first_of(const_pointer s, size_type pos = 0) const { + return find_first_of(basic_string_view(s), pos); + } + + constexpr size_type find_last_of(basic_string_view v, size_type pos = npos) + const noexcept { + return find_last_if_(pos, stringViewContainsChar_{v}); + } + + constexpr size_type find_last_of(CharT ch, size_type pos = npos) + const noexcept { + return find_last_if_(pos, charIsEqual_{ch}); + } + + constexpr size_type find_last_of( + const_pointer s, + size_type pos, + size_type count) const { + return find_last_of(basic_string_view(s, count), pos); + } + + constexpr size_type find_last_of(const_pointer s, size_type pos = npos) + const { + return find_last_of(basic_string_view(s), pos); + } + + constexpr size_type find_first_not_of(basic_string_view v, size_type pos = 0) + const noexcept { + return find_first_if_(pos, stringViewDoesNotContainChar_{v}); + } + + constexpr size_type find_first_not_of(CharT ch, size_type pos = 0) + const noexcept { + return find_first_if_(pos, charIsNotEqual_{ch}); + } + + constexpr size_type find_first_not_of( + const_pointer s, + size_type pos, + size_type count) const { + return find_first_not_of(basic_string_view(s, count), pos); + } + + constexpr size_type find_first_not_of(const_pointer s, size_type pos = 0) + const { + return find_first_not_of(basic_string_view(s), pos); + } + + constexpr size_type find_last_not_of( + basic_string_view v, + size_type pos = npos) const noexcept { + return find_last_if_(pos, stringViewDoesNotContainChar_{v}); + } + + constexpr size_type find_last_not_of(CharT ch, size_type pos = npos) + const noexcept { + return find_last_if_(pos, charIsNotEqual_{ch}); + } + + constexpr size_type find_last_not_of( + const_pointer s, + size_type pos, + size_type count) const { + return find_last_not_of(basic_string_view(s, count), pos); + } + + constexpr size_type find_last_not_of(const_pointer s, size_type pos = npos) + const { + return find_last_not_of(basic_string_view(s), pos); + } + + private: + static constexpr size_type strlen_(const_pointer str) noexcept { + const_pointer current = str; + while (*current != '\0') { + ++current; + } + return current - str; + } + + constexpr const_reference at_(size_type pos) const noexcept { + return *(begin_ + pos); + } + + constexpr basic_string_view substr_(size_type pos = 0, size_type count = npos) + const { + return basic_string_view{begin_ + pos, std::min(count, size() - pos)}; + } + + template + // NOLINTNEXTLINE(cppcoreguidelines-missing-std-forward) + constexpr size_type find_first_if_(size_type pos, Condition&& condition) + const noexcept { + if (pos + 1 <= size()) { + for (size_type cur = pos; cur < size(); ++cur) { + if (condition(at_(cur))) { + return cur; + } + } + } + return npos; + } + + template + // NOLINTNEXTLINE(cppcoreguidelines-missing-std-forward) + constexpr size_type find_last_if_(size_type pos, Condition&& condition) + const noexcept { + // Write it iteratively. This is faster. + if (!empty()) { + pos = std::min(size() - 1, pos); + do { + if (condition(at_(pos))) { + return pos; + } + } while (pos-- > 0); + } + return npos; + } + + constexpr bool equals_(basic_string_view rhs) const { + // We don't use string_view::compare() here but implement it manually + // because only looking at equality allows for more optimized code. +#if defined(__GNUC__) && !defined(__CUDACC__) + return size() == rhs.size() && + 0 == __builtin_memcmp(data(), rhs.data(), size()); +#else + if (size() != rhs.size()) { + return false; + } + // Yes, memcmp would be laster than this loop, but memcmp isn't constexpr + // and I didn't feel like implementing a constexpr memcmp variant. + // TODO At some point this should probably be done, including tricks + // like comparing one machine word instead of a byte per iteration. + for (typename basic_string_view::size_type pos = 0; pos < size(); + ++pos) { + if (at_(pos) != rhs.at_(pos)) { + return false; + } + } + return true; +#endif + } + + struct charIsEqual_ final { + CharT expected; + constexpr bool operator()(CharT actual) const noexcept { + return expected == actual; + } + }; + + struct charIsNotEqual_ final { + CharT expected; + constexpr bool operator()(CharT actual) const noexcept { + return expected != actual; + } + }; + + struct stringViewContainsChar_ final { + basic_string_view expected; + constexpr bool operator()(CharT ch) const noexcept { + return npos != expected.find(ch); + } + }; + + struct stringViewDoesNotContainChar_ final { + basic_string_view expected; + constexpr bool operator()(CharT ch) const noexcept { + return npos == expected.find(ch); + } + }; + + const_pointer begin_; + size_type size_{}; +}; + +template +inline std::basic_ostream& operator<<( + std::basic_ostream& stream, + basic_string_view sv) { + // The rules for operator<< are quite complex, so lets defer to the + // STL implementation. + using std_string_type = ::std::basic_string_view; + return stream << std_string_type(sv.data(), sv.size()); +} + +template +constexpr inline void swap( + basic_string_view& lhs, + basic_string_view& rhs) noexcept { + lhs.swap(rhs); +} +using string_view = std::string_view; +using c10_string_view = basic_string_view; + +// NOTE: In C++20, this function should be replaced by string_view.starts_with +constexpr bool starts_with( + const std::string_view s, + const std::string_view prefix) noexcept { + return (prefix.size() > s.size()) ? false + : prefix == s.substr(0, prefix.size()); +} + +// NOTE: In C++20, this function should be replaced by string_view.starts_with +constexpr bool starts_with( + const std::string_view s, + const char prefix) noexcept { + return !s.empty() && prefix == s.front(); +} + +// NOTE: In C++20, this function should be replaced by string_view.ends_with +constexpr bool ends_with( + const std::string_view s, + const std::string_view suffix) noexcept { + return (suffix.size() > s.size()) + ? false + : suffix == s.substr(s.size() - suffix.size(), suffix.size()); +} + +// NOTE: In C++20, this function should be replaced by string_view.ends_with +constexpr bool ends_with(const std::string_view s, const char prefix) noexcept { + return !s.empty() && prefix == s.back(); +} + +} // namespace c10 + +namespace std { +template +struct hash<::c10::basic_string_view> { + size_t operator()(::c10::basic_string_view x) const { + // The standard says that std::string_view hashing must do the same as + // std::string hashing but leaves the details of std::string hashing + // up to the implementer. So, to be conformant, we need to reuse and + // existing STL type's hash function. The std::string fallback is probably + // slow but the only way to be conformant. + + using std_string_type = ::std::basic_string_view; + return ::std::hash{}(std_string_type(x.data(), x.size())); + } +}; +} // namespace std + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/c10/util/typeid.h b/backends/pytorch/include/c10/util/typeid.h new file mode 100644 index 000000000..3f7da4264 --- /dev/null +++ b/backends/pytorch/include/c10/util/typeid.h @@ -0,0 +1,720 @@ +#if !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include + +#include + +/* + * TypeIdentifier is a small type containing an id. + * Types must be registered using CAFFE_DECLARE_KNOWN_TYPE() (in their header) + * and CAFFE_DEFINE_KNOWN_TYPE() (in their .cpp file) for them to have a type + * id. If a type is registered, you can also create an object containing meta + * data like constructor, destructor, stringified name, ... about the type by + * calling TypeMeta::Make. This returns a TypeMeta() object, which is + * basically just a pointer to the type information, so it's cheap to pass + * around. + */ + +// TODO: This file is still in the caffe2 namespace, despite living +// in the ATen directory. This is because the macro +// CAFFE_KNOWN_TYPE (and CAFFE_DECLARE_KNOWN_TYPE) defines a template +// specialization, which relies +// on the namespace of TypeMeta matching the namespace where the macro is +// called. This requires us to fix all of the call-sites, which I want to do +// later. So the namespace is not fixed at the moment. + +// Make at::Half a fundamental type. + +namespace c10::guts { +template <> +struct is_fundamental : std::true_type {}; +} // namespace c10::guts + +namespace caffe2 { + +/** + * A type id is a unique id for a given C++ type. + * You need to register your types using CAFFE_KNOWN_TYPE(MyType) to be able to + * use TypeIdentifier with custom types. This is for example used to store the + * dtype of tensors. + */ +class C10_API TypeIdentifier final + : public at::IdWrapper { + public: + friend std::ostream& operator<<(std::ostream& stream, TypeIdentifier typeId); + friend constexpr bool operator<(TypeIdentifier lhs, TypeIdentifier rhs); + + /** + * Returns the unique id for the given type T. The id is unique for the type T + * in the sense that for any two different types, their ids are different; for + * the same type T, the id remains the same over different calls of the + * function. However, this is not guaranteed over different runs, as the id + * is generated during run-time. Do NOT serialize the id for storage. + */ + template + static constexpr TypeIdentifier Get() noexcept { + return TypeIdentifier(c10::util::get_type_index()); + } + + static constexpr TypeIdentifier uninitialized() { + return TypeIdentifier(c10::util::type_index{0}); + } + + private: + constexpr explicit TypeIdentifier(c10::util::type_index id) : IdWrapper(id) {} +}; + +// Allow usage in std::map / std::set +// TODO Disallow this and rather use std::unordered_map/set everywhere +inline constexpr bool operator<(TypeIdentifier lhs, TypeIdentifier rhs) { + return lhs.underlyingId() < rhs.underlyingId(); +} + +inline std::ostream& operator<<( + std::ostream& stream, + caffe2::TypeIdentifier typeId) { + return stream << typeId.underlyingId(); +} + +} // namespace caffe2 + +namespace at { +using DataType = caffe2::TypeIdentifier; +} + +C10_DEFINE_HASH_FOR_IDWRAPPER(caffe2::TypeIdentifier) + +namespace caffe2 { + +namespace detail { + +// This struct holds the actual type information. There will be +// one allocated per type. TypeMeta objects will then point to the struct +// instance for the type they're configured for. +struct TypeMetaData final { + using New = void*(); + using PlacementNew = void(void*, size_t); + using Copy = void(const void*, void*, size_t); + using PlacementDelete = void(void*, size_t); + using Delete = void(void*); + + constexpr TypeMetaData() noexcept + : itemsize_(0), + new_(nullptr), + placementNew_(nullptr), + copy_(nullptr), + placementDelete_(nullptr), + delete_(nullptr), + id_(TypeIdentifier::uninitialized()), + name_("nullptr (uninitialized)") {} + + constexpr TypeMetaData( + size_t itemsize, + New* newFn, + PlacementNew* placementNew, + Copy* copy, + PlacementDelete* placementDelete, + Delete* deleteFn, + TypeIdentifier id, + std::string_view name) noexcept + : itemsize_(itemsize), + new_(newFn), + placementNew_(placementNew), + copy_(copy), + placementDelete_(placementDelete), + delete_(deleteFn), + id_(id), + name_(name) {} + + size_t itemsize_; + New* new_; + PlacementNew* placementNew_; + Copy* copy_; + PlacementDelete* placementDelete_; + Delete* delete_; + TypeIdentifier id_; + std::string_view name_; +}; + +// Mechanism for throwing errors which can't be prevented at compile time +// due to type erasure. E.g. somebody calling TypeMeta::copy() for +// non-copyable type. Right now just throws exception but is implemented +// in .cpp to manage dependencies +[[noreturn]] C10_API void _ThrowRuntimeTypeLogicError(const std::string& msg); + +/** + * Placement new function for the type. + */ +template +inline void _PlacementNew(void* ptr, size_t n) { + T* typed_ptr = static_cast(ptr); + for (const auto i : c10::irange(n)) { + new (typed_ptr + i) T; + } +} + +template +inline void _PlacementNewNotDefault(void* /*ptr*/, size_t /*n*/) { + _ThrowRuntimeTypeLogicError( + "Type " + std::string(c10::util::get_fully_qualified_type_name()) + + " is not default-constructible."); +} + +template < + typename T, + std::enable_if_t>* = nullptr> +inline constexpr TypeMetaData::PlacementNew* _PickPlacementNew() { + return (c10::guts::is_fundamental::value || std::is_pointer_v) + ? nullptr + : &_PlacementNew; +} + +template < + typename T, + std::enable_if_t>* = nullptr> +inline constexpr TypeMetaData::PlacementNew* _PickPlacementNew() { + static_assert( + !c10::guts::is_fundamental::value && !std::is_pointer_v, + "this should have picked the other SFINAE case"); + return &_PlacementNewNotDefault; +} + +template +inline void* _New() { + return new T; +} + +template +inline void* _NewNotDefault() { + _ThrowRuntimeTypeLogicError( + "Type " + std::string(c10::util::get_fully_qualified_type_name()) + + " is not default-constructible."); +} + +template < + typename T, + std::enable_if_t>* = nullptr> +inline constexpr TypeMetaData::New* _PickNew() { + return &_New; +} + +template < + typename T, + std::enable_if_t>* = nullptr> +inline constexpr TypeMetaData::New* _PickNew() { + return &_NewNotDefault; +} + +/** + * Typed copy function for classes. + */ +template +inline void _Copy(const void* src, void* dst, size_t n) { + const T* typed_src = static_cast(src); + T* typed_dst = static_cast(dst); + for (const auto i : c10::irange(n)) { + typed_dst[i] = typed_src[i]; + } +} + +/** + * A placeholder function for types that do not allow assignment. + */ +template +inline void _CopyNotAllowed(const void* /*src*/, void* /*dst*/, size_t /*n*/) { + _ThrowRuntimeTypeLogicError( + "Type " + std::string(c10::util::get_fully_qualified_type_name()) + + " does not allow assignment."); +} + +template >* = nullptr> +inline constexpr TypeMetaData::Copy* _PickCopy() { + return (c10::guts::is_fundamental::value || std::is_pointer_v) + ? nullptr + : &_Copy; +} + +template < + typename T, + std::enable_if_t>* = nullptr> +inline constexpr TypeMetaData::Copy* _PickCopy() { + static_assert( + !c10::guts::is_fundamental::value && !std::is_pointer_v, + "this should have picked the other SFINAE case"); + return &_CopyNotAllowed; +} + +/** + * Destructor for non-fundamental types. + */ +template +inline void _PlacementDelete(void* ptr, size_t n) { + T* typed_ptr = static_cast(ptr); + for (const auto i : c10::irange(n)) { + typed_ptr[i].~T(); + } +} + +template +inline constexpr TypeMetaData::PlacementDelete* _PickPlacementDelete() { + return (c10::guts::is_fundamental::value || std::is_pointer_v) + ? nullptr + : &_PlacementDelete; +} + +template +inline void _Delete(void* ptr) { + T* typed_ptr = static_cast(ptr); + delete typed_ptr; +} + +template +inline constexpr TypeMetaData::Delete* _PickDelete() noexcept { + return &_Delete; +} + +class _Uninitialized final {}; + +} // namespace detail + +// +// note: this is outside TypeMeta bc gcc seems to have trouble +// with scalarTypeItemSizes as a constexpr static member used by +// a public inline instance method +// + +// item sizes for TypeMeta::itemsize() fast path +static constexpr std::array scalarTypeItemSizes = { +#define SCALAR_TYPE_SIZE(T, name) sizeof(T), + AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(SCALAR_TYPE_SIZE) +#undef SCALAR_TYPE_SIZE + 0, // Undefined +}; + +/** + * TypeMeta is a thin class that allows us to store the type of a container such + * as a blob, or the data type of a tensor, with a unique run-time id. It also + * stores some additional data such as the item size and the name of the type + * for run-time inspection. + */ +class C10_API TypeMeta final { + public: + using New = detail::TypeMetaData::New; + using PlacementNew = detail::TypeMetaData::PlacementNew; + using Copy = detail::TypeMetaData::Copy; + using PlacementDelete = detail::TypeMetaData::PlacementDelete; + using Delete = detail::TypeMetaData::Delete; + + /** Create a dummy TypeMeta object. To create a TypeMeta object for a specific + * type, use TypeMeta::Make(). + */ + TypeMeta() noexcept; + ~TypeMeta() = default; + + /** + * Copy constructor. + */ + TypeMeta(const TypeMeta& src) noexcept = default; + + /** + * Assignment operators. + */ + TypeMeta& operator=(const TypeMeta& src) noexcept = default; + + TypeMeta& operator=(TypeMeta&& src) noexcept = default; + TypeMeta(TypeMeta&& rhs) noexcept = default; + + inline TypeMeta& operator=(ScalarType scalar_type) noexcept { + index_ = static_cast(scalar_type); + return *this; + } + + private: + // TypeMeta can only be created by Make, making sure that we do not + // create incorrectly mixed up TypeMeta objects. + explicit TypeMeta(const uint16_t index) noexcept : index_(index) {} + + public: + /** + * Returns the type id. + */ + TypeIdentifier id() const noexcept { + return data().id_; + } + /** + * true if we represent some ScalarType type + */ + inline bool isScalarType() const noexcept { + return index_ < NumScalarTypes; + } + /** + * true if we represent ScalarType scalar_type + */ + inline bool isScalarType(ScalarType scalar_type) const noexcept { + return index_ == static_cast(scalar_type); + } + /** + * Returns the size of the item. + */ + inline size_t itemsize() const noexcept { + if (C10_LIKELY(isScalarType())) { + return scalarTypeItemSizes[index_]; + } + return data().itemsize_; + } + /** + * Returns the new function pointer for individual items. + */ + New* newFn() const noexcept { + return data().new_; + } + /** + * Returns the placement new function pointer for individual items. + */ + PlacementNew* placementNew() const noexcept { + return data().placementNew_; + } + /** + * Returns the typed copy function pointer for individual items. + */ + Copy* copy() const noexcept { + return data().copy_; + } + /** + * Returns the destructor function pointer for individual items. + */ + PlacementDelete* placementDelete() const noexcept { + return data().placementDelete_; + } + Delete* deleteFn() const noexcept { + return data().delete_; + } + /** + * Returns a printable name for the type. + */ + std::string_view name() const noexcept { + return data().name_; + } + + friend bool operator==(const TypeMeta& lhs, const TypeMeta& rhs) noexcept; + + template + bool Match() const noexcept { + return (*this == Make()); + } + + // Below are static functions that can be called by passing a specific type. + + template + static constexpr TypeIdentifier Id() noexcept { + return TypeIdentifier::Get(); + } + + template + static std::string_view TypeName() noexcept { + return c10::util::get_fully_qualified_type_name(); + } + + template + static constexpr size_t ItemSize() noexcept { + return sizeof(T); + } + + /** + * Returns a TypeMeta object that corresponds to the typename T. + */ + template + static TypeMeta Make() { + // The instance pointed to is declared here, but defined in a .cpp file. + // We need to silence the compiler warning about using an undefined + // variable template. '-Wpragmas' and '-Wunknown-warning-option' has to be + // disabled for compilers that don't know '-Wundefined-var-template' and + // would error at our attempt to disable it. +#ifndef _MSC_VER +#pragma GCC diagnostic push +#pragma GCC diagnostic ignored "-Wpragmas" +#pragma GCC diagnostic ignored "-Wunknown-warning-option" +#pragma GCC diagnostic ignored "-Wundefined-var-template" +#endif + return TypeMeta(_typeMetaData()); +#ifndef _MSC_VER +#pragma GCC diagnostic pop +#endif + } + + /** + * convert ScalarType enum values to TypeMeta handles + */ + static inline caffe2::TypeMeta fromScalarType(ScalarType scalar_type) { + const auto index = static_cast(scalar_type); + TORCH_INTERNAL_ASSERT_DEBUG_ONLY( + index < NumScalarTypes, + "Unrecognized Scalartype ", + scalar_type, + " (please report this error)"); + return TypeMeta(index); + } + + /** + * convert TypeMeta handles to ScalarType enum values + */ + inline ScalarType toScalarType() const { + if (C10_LIKELY(isScalarType())) { + return static_cast(index_); + } + error_unsupported_typemeta(*this); + } + + private: + [[noreturn]] static void error_unsupported_typemeta(caffe2::TypeMeta dtype); + + // hard limit number of registered types + // note: constexpr provokes Windows compilation error "member may not be + // initialized" static constexpr size_t MaxTypeIndex = 32; + // +#if defined C10_MOBILE +// The reason for this not to be UINT8_MAX is that the array +// initialization takes space which is proportional to the size of the array. +// The compiler seems to add code (or data padding) to initialize the array with +// empty elements. Please see +// https://github.com/pytorch/pytorch/pull/51881 for details. +// +#define MaxTypeIndex \ + (NumScalarTypes + 15 /* number of CAFFE_DEFINE_KNOWN_TYPE in typeid.cpp */ + \ + 1 /* 1 more for caffe2 tensor */) +#else +#define MaxTypeIndex UINT8_MAX +#endif + + // Protects type metadata allocation. + // NOLINTNEXTLINE(facebook-hte-NonPodStaticDeclaration) + static std::mutex& getTypeMetaDatasLock(); + static uint16_t nextTypeIndex; + + static detail::TypeMetaData* typeMetaDatas(); + + static uint16_t existingMetaDataIndexForType(TypeIdentifier identifier); + + public: +#ifdef __CUDACC__ + // NOTE [ TypeIdentifier::Get nvcc/clang discrepancy] + // nvcc and clang do not produce identical results for + // TypeIdentifier::Get, because TypeIdentifier::Get relies on + // __PRETTY_FUNCTION__ and they don't agree on the canonical names + // of types (e.g., nvcc normalizes to `short unsigned int`, but clang + // calls it `unsigned short`). Hide the implementation of this function + // from nvcc so that we always use clang (or whatever host C++ compiler) + // for TypeIdentifier::Get. + template + C10_EXPORT static uint16_t addTypeMetaData(); +#else + template + C10_EXPORT static uint16_t addTypeMetaData() { + const auto identifier = TypeIdentifier::Get(); + // Need to hold this for the rest of the function, protecting: + // 1) existingMetaDataIndexForType() + // 2) nextTypeIndex++ + // 3) the write into typeMetaDatas() + std::lock_guard lock(getTypeMetaDatasLock()); + // It may exist already if added in a different dynamic shared library. + const uint16_t existing_index = existingMetaDataIndexForType(identifier); + if (existing_index != MaxTypeIndex) { + return existing_index; + } + const uint16_t index = nextTypeIndex++; + TORCH_CHECK( + index <= MaxTypeIndex, + "Maximum number of CAFFE_KNOWN_TYPE declarations has been exceeded. ", + "Please report this issue."); + typeMetaDatas()[index] = detail::TypeMetaData{ + sizeof(T), + detail::_PickNew(), + detail::_PickPlacementNew(), + detail::_PickCopy(), + detail::_PickPlacementDelete(), + detail::_PickDelete(), + identifier, + c10::util::get_fully_qualified_type_name()}; + return index; + } +#endif + + private: + // specializations return indexes into typeMetaDataInstances() + template + C10_API static uint16_t _typeMetaData() noexcept; + + // + // TypeMeta just wraps this index + // + + uint16_t index_; + + inline const detail::TypeMetaData& data() const { + return typeMetaDatas()[index_]; + } +}; + +// specializations of TypeMeta::_typeMetaData for ScalarType types + +#define DEFINE_SCALAR_METADATA_INSTANCE(T, name) \ + template <> \ + constexpr uint16_t TypeMeta::_typeMetaData() noexcept { \ + return static_cast(ScalarType::name); \ + } +AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(DEFINE_SCALAR_METADATA_INSTANCE) +#undef DEFINE_SCALAR_METADATA_INSTANCE + +template <> +C10_EXPORT constexpr uint16_t TypeMeta::_typeMetaData< + detail::_Uninitialized>() noexcept { + return static_cast(ScalarType::Undefined); +} + +inline TypeMeta::TypeMeta() noexcept + : index_(_typeMetaData()) {} + +inline bool operator==(const TypeMeta& lhs, const TypeMeta& rhs) noexcept { + return (lhs.index_ == rhs.index_); +} +inline bool operator!=(const TypeMeta& lhs, const TypeMeta& rhs) noexcept { + return !operator==(lhs, rhs); +} + +inline std::ostream& operator<<( + std::ostream& stream, + caffe2::TypeMeta typeMeta) { + return stream << typeMeta.name(); +} + +/** + * Register unique id for a type so it can be used in TypeMeta context, e.g. be + * used as a type for Blob or for Tensor elements. + * + * CAFFE_KNOWN_TYPE is deprecated; prefer CAFFE_DECLARE_KNOWN_TYPE and + * CAFFE_DEFINE_KNOWN_TYPE. + * + * CAFFE_KNOWN_TYPE does explicit instantiation of TypeIdentifier::Get + * template function and thus needs to be put in a single translation unit (.cpp + * file) for a given type T. Other translation units that use type T as a type + * of the caffe2::Blob or element type of caffe2::Tensor need to depend on the + * translation unit that contains CAFFE_KNOWN_TYPE declaration via regular + * linkage dependencies. + * + * NOTE: the macro needs to be invoked in ::caffe2 namespace + */ +// Implementation note: in MSVC, we will need to prepend the C10_API +// keyword in order to get things compiled properly. in Linux, gcc seems to +// create attribute ignored error for explicit template instantiations, see +// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2017/p0537r0.html +// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=51930 +// and as a result, we define these two macros slightly differently. +#if defined(_MSC_VER) || defined(__clang__) +#define EXPORT_IF_NOT_GCC C10_EXPORT +#else +#define EXPORT_IF_NOT_GCC +#endif + +// CAFFE_KNOWN_TYPE is deprecated! Use CAFFE_DECLARE_KNOWN_TYPE and +// CAFFE_DEFINE_KNOWN_TYPE instead. +#define CAFFE_KNOWN_TYPE(T) \ + template uint16_t TypeMeta::addTypeMetaData(); \ + template <> \ + EXPORT_IF_NOT_GCC uint16_t TypeMeta::_typeMetaData() noexcept { \ + static const uint16_t index = addTypeMetaData(); \ + return index; \ + } + +#define CAFFE_DEFINE_KNOWN_TYPE(T, ident) \ + template uint16_t TypeMeta::addTypeMetaData(); \ + namespace detail { \ + EXPORT_IF_NOT_GCC const uint16_t ident##_metadata_index = \ + TypeMeta::addTypeMetaData(); \ + } // namespace detail + +// Unlike CAFFE_KNOWN_TYPE, CAFFE_DECLARE_KNOWN_TYPE avoids a function +// call to access _typeMetaData in the common case. +#define CAFFE_DECLARE_KNOWN_TYPE(T, ident) \ + extern template uint16_t TypeMeta::addTypeMetaData(); \ + namespace detail { \ + extern C10_API const uint16_t ident##_metadata_index; \ + } /* namespace detail */ \ + template <> \ + EXPORT_IF_NOT_GCC C10_ALWAYS_INLINE uint16_t \ + TypeMeta::_typeMetaData() noexcept { \ + return detail::ident##_metadata_index; \ + } + +#define CAFFE_KNOWN_TYPE_NOEXPORT(T) \ + template <> \ + uint16_t TypeMeta::_typeMetaData() noexcept { \ + static const uint16_t index = addTypeMetaData(); \ + return index; \ + } + +CAFFE_DECLARE_KNOWN_TYPE(std::string, std_string) +CAFFE_DECLARE_KNOWN_TYPE(char, char) +CAFFE_DECLARE_KNOWN_TYPE(std::unique_ptr, std_unique_ptr_std_mutex) +CAFFE_DECLARE_KNOWN_TYPE( + std::unique_ptr>, + std_unique_ptr_std_atomic_bool) +CAFFE_DECLARE_KNOWN_TYPE(std::vector, std_vector_int32_t) +CAFFE_DECLARE_KNOWN_TYPE(std::vector, std_vector_int64_t) +CAFFE_DECLARE_KNOWN_TYPE(std::vector, std_vector_unsigned_long) +CAFFE_DECLARE_KNOWN_TYPE(bool*, bool_ptr) +CAFFE_DECLARE_KNOWN_TYPE(char*, char_ptr) +CAFFE_DECLARE_KNOWN_TYPE(int*, int_ptr) + +// For some of the compilers, long is defined separately from int32_t and +// int64_t. As a result we will need to actually define them separately. +// It is recommended that one does NOT use long - use int32_t and int64_t +// explicitly. Explicit long type annotation may go away in the future. +// details: This hack works by defining a _guard_long_unique type, which is +// long iff the compiler has a separate long type and is a dummy type otherwise. +// we then allocate a type id to that _guard_long_unique. If the compiler has a +// separate long type, this allocates a type id for long. Otherwise, it +// allocates a type id for the dummy type, which doesn't matter. +namespace detail { +template +class _guard_long_unique_dummy final {}; +template +using _guard_long_unique = std::conditional_t< + std::is_same_v || std::is_same_v, + _guard_long_unique_dummy, + T>; +} // namespace detail + +CAFFE_DECLARE_KNOWN_TYPE( + detail::_guard_long_unique, + detail_guard_long_unique_long) +CAFFE_DECLARE_KNOWN_TYPE( + detail::_guard_long_unique>, + detail_guard_long_unique_std_vector_long) + +CAFFE_DECLARE_KNOWN_TYPE(float*, float_ptr) +CAFFE_DECLARE_KNOWN_TYPE(at::Half*, at_Half) + +} // namespace caffe2 + +#else +#error "This file should not be included when either TORCH_STABLE_ONLY or TORCH_TARGET_VERSION is defined." +#endif // !defined(TORCH_STABLE_ONLY) && !defined(TORCH_TARGET_VERSION) diff --git a/backends/pytorch/include/torch/headeronly/core/DeviceType.h b/backends/pytorch/include/torch/headeronly/core/DeviceType.h new file mode 100644 index 000000000..9db3ef256 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/core/DeviceType.h @@ -0,0 +1,128 @@ +#pragma once + +// This is directly synchronized with caffe2/proto/caffe2.proto, but +// doesn't require me to figure out how to get Protobuf headers into +// ATen/core (which would require a lot more build system hacking.) +// If you modify me, keep me synchronized with that file. + +#include +#include + +#include +#include +#include + +namespace c10 { + +// These contains all device types that also have a BackendComponent +// and therefore participate in per-backend functionality dispatch keys. +// This is most backends except PrivateUse2 and PrivateUse3 +#define C10_FORALL_BACKEND_DEVICE_TYPES(_, extra) \ + _(CPU, extra) \ + _(CUDA, extra) \ + _(HIP, extra) \ + _(XLA, extra) \ + _(MPS, extra) \ + _(IPU, extra) \ + _(XPU, extra) \ + _(HPU, extra) \ + _(VE, extra) \ + _(Lazy, extra) \ + _(Meta, extra) \ + _(MTIA, extra) \ + _(PrivateUse1, extra) + +enum class DeviceType : int8_t { + CPU = 0, + CUDA = 1, // CUDA. + MKLDNN = 2, // Reserved for explicit MKLDNN + OPENGL = 3, // OpenGL + OPENCL = 4, // OpenCL + IDEEP = 5, // IDEEP. + HIP = 6, // AMD HIP + FPGA = 7, // FPGA + MAIA = 8, // ONNX Runtime / Microsoft + XLA = 9, // XLA / TPU + Vulkan = 10, // Vulkan + Metal = 11, // Metal + XPU = 12, // XPU + MPS = 13, // MPS + Meta = 14, // Meta (tensors with no data) + HPU = 15, // HPU / HABANA + VE = 16, // SX-Aurora / NEC + Lazy = 17, // Lazy Tensors + IPU = 18, // Graphcore IPU + MTIA = 19, // Meta training and inference devices + PrivateUse1 = 20, // PrivateUse1 device + // NB: If you add more devices: + // - Change the implementations of DeviceTypeName and isValidDeviceType + // in c10/core/DeviceType.cpp + // - Change the number below + COMPILE_TIME_MAX_DEVICE_TYPES = 21, +}; + +constexpr DeviceType kCPU = DeviceType::CPU; +constexpr DeviceType kCUDA = DeviceType::CUDA; +constexpr DeviceType kHIP = DeviceType::HIP; +constexpr DeviceType kFPGA = DeviceType::FPGA; +constexpr DeviceType kMAIA = DeviceType::MAIA; +constexpr DeviceType kXLA = DeviceType::XLA; +constexpr DeviceType kMPS = DeviceType::MPS; +constexpr DeviceType kMeta = DeviceType::Meta; +constexpr DeviceType kVulkan = DeviceType::Vulkan; +constexpr DeviceType kMetal = DeviceType::Metal; +constexpr DeviceType kXPU = DeviceType::XPU; +constexpr DeviceType kHPU = DeviceType::HPU; +constexpr DeviceType kVE = DeviceType::VE; +constexpr DeviceType kLazy = DeviceType::Lazy; +constexpr DeviceType kIPU = DeviceType::IPU; +constexpr DeviceType kMTIA = DeviceType::MTIA; +constexpr DeviceType kPrivateUse1 = DeviceType::PrivateUse1; + +// define explicit int constant +constexpr int COMPILE_TIME_MAX_DEVICE_TYPES = + static_cast(DeviceType::COMPILE_TIME_MAX_DEVICE_TYPES); + +static_assert( + COMPILE_TIME_MAX_DEVICE_TYPES <= 21, + "Hey! You seem to be adding a lot of new DeviceTypes. The intent was " + "for this constant to reflect the actual number of DeviceTypes we support " + "in PyTorch; it's important that this number is not too large as we " + "use this to allocate stack arrays in some places in our code. If you " + "are indeed just adding the 20th device type, feel free to change " + "the check to 32; but if you are adding some sort of extensible device " + "types registration, please be aware that you are affecting code that " + "this number is small. Try auditing uses of this constant."); + +} // namespace c10 + +namespace std { +template <> +struct hash { + std::size_t operator()(c10::DeviceType k) const { + return std::hash()(static_cast(k)); + } +}; +} // namespace std + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) +using c10::COMPILE_TIME_MAX_DEVICE_TYPES; +using c10::DeviceType; +using c10::kCPU; +using c10::kCUDA; +using c10::kFPGA; +using c10::kHIP; +using c10::kHPU; +using c10::kIPU; +using c10::kLazy; +using c10::kMAIA; +using c10::kMeta; +using c10::kMetal; +using c10::kMPS; +using c10::kMTIA; +using c10::kPrivateUse1; +using c10::kVE; +using c10::kVulkan; +using c10::kXLA; +using c10::kXPU; +HIDDEN_NAMESPACE_END(torch, headeronly) diff --git a/backends/pytorch/include/torch/headeronly/core/Layout.h b/backends/pytorch/include/torch/headeronly/core/Layout.h new file mode 100644 index 000000000..62e34ff67 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/core/Layout.h @@ -0,0 +1,44 @@ +#pragma once + +#include +#include + +#include +#include + +namespace c10 { + +enum class Layout : int8_t { + Strided, + Sparse, + SparseCsr, + Mkldnn, + SparseCsc, + SparseBsr, + SparseBsc, + Jagged, + NumOptions +}; + +constexpr auto kStrided = Layout::Strided; +constexpr auto kSparse = Layout::Sparse; +constexpr auto kSparseCsr = Layout::SparseCsr; +constexpr auto kMkldnn = Layout::Mkldnn; +constexpr auto kSparseCsc = Layout::SparseCsc; +constexpr auto kSparseBsr = Layout::SparseBsr; +constexpr auto kSparseBsc = Layout::SparseBsc; +constexpr auto kJagged = Layout::Jagged; + +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) +using c10::kJagged; +using c10::kMkldnn; +using c10::kSparse; +using c10::kSparseBsc; +using c10::kSparseBsr; +using c10::kSparseCsc; +using c10::kSparseCsr; +using c10::kStrided; +using c10::Layout; +HIDDEN_NAMESPACE_END(torch, headeronly) diff --git a/backends/pytorch/include/torch/headeronly/core/MemoryFormat.h b/backends/pytorch/include/torch/headeronly/core/MemoryFormat.h new file mode 100644 index 000000000..ad02a901e --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/core/MemoryFormat.h @@ -0,0 +1,46 @@ +#pragma once + +#include +#include + +#include +#include + +// Memory format is not the property of a Tensor. It is the way to tell an +// operator how the result should be organized in memory and nothing more. That +// means memory format should never be used as return value for any tensor state +// interrogation functions (internally and externally). +// +// Possible options are: +// Preserve: +// If any of the input tensors is in channels_last format, operator output +// should be in channels_last format +// +// Contiguous: +// Regardless of input tensors format, the output should be contiguous +// Tensor. +// +// ChannelsLast: +// Regardless of input tensors format, the output should be in channels_last +// format. + +namespace c10 { + +enum class MemoryFormat : int8_t { + Contiguous, + Preserve, + ChannelsLast, + ChannelsLast3d, + NumOptions +}; + +inline MemoryFormat get_contiguous_memory_format() { + return MemoryFormat::Contiguous; +} + +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) +using c10::get_contiguous_memory_format; +using c10::MemoryFormat; +HIDDEN_NAMESPACE_END(torch, headeronly) diff --git a/backends/pytorch/include/torch/headeronly/core/ScalarType.h b/backends/pytorch/include/torch/headeronly/core/ScalarType.h new file mode 100644 index 000000000..ce43ce686 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/core/ScalarType.h @@ -0,0 +1,381 @@ +#pragma once + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include + +C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wswitch-enum") + +namespace c10 { + +// dummy struct for uint1 to uint7, actual functionality +// of these dtypes will be implemented in python with Tensor subclass +template +struct dummy_uint1_7_t {}; + +// dummy struct for int1 to int7, actual functionality +// of these dtypes will be implemented in python with Tensor subclass +template +struct dummy_int1_7_t {}; + +// [dtype Macros note] For the macros below: +// +// For users: If you want to macro some code for all non-QInt scalar types +// (i.e. types with complete information, you probably want one of the +// AT_FORALL_SCALAR_TYPES / AT_FORALL_SCALAR_TYPES_AND macros below, which are +// designed to behave similarly to the Dispatch macros with the same name. +// +// For adding a new dtype: In the beginning, we had an idea that there was a +// list of all scalar types, and you could use AT_FORALL_SCALAR_TYPES to +// iterate over them. But over the years we added weird types which couldn't +// be handled uniformly everywhere and so in the end we ended up with some +// mish-mosh of some helper macros, but mostly use sites making a call about +// what dtypes they can or can't support. So if you want to add a new dtype, +// the preferred resolution is to find a dtype similar to what you want, +// grep for it and edit all the sites you find this way. If you need to add +// a completely new kind of dtype, you're going to have to laboriously audit +// all of the sites everywhere to figure out how it should work. Consulting +// some old PRs where we added new dtypes (check history of this file) can +// help give you an idea where to start. + +// If you want to support ComplexHalf for real, add ComplexHalf +// into this macro (and change the name). But beware: convert() +// doesn't work for all the conversions you need... +// +// TODO: To add unsigned int types here, we must define accumulate type. +// But uint8 currently accumulates into int64, so we would have to make +// an inconsistent choice for the larger types. Difficult. +#define AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_EXCEPT_COMPLEX_HALF_F8NZ(_) \ + _(uint8_t, Byte) \ + _(int8_t, Char) \ + _(int16_t, Short) \ + _(int, Int) \ + _(int64_t, Long) \ + _(c10::Half, Half) \ + _(float, Float) \ + _(double, Double) \ + _(c10::complex, ComplexFloat) \ + _(c10::complex, ComplexDouble) \ + _(bool, Bool) \ + _(c10::BFloat16, BFloat16) \ + _(c10::Float8_e5m2, Float8_e5m2) \ + _(c10::Float8_e4m3fn, Float8_e4m3fn) + +// This macro controls many of our C++ APIs, including constructors +// for Scalar as well as the data() and item() accessors on Tensor +#define AT_FORALL_SCALAR_TYPES_WITH_COMPLEX(_) \ + _(uint8_t, Byte) \ + _(int8_t, Char) \ + _(int16_t, Short) \ + _(int, Int) \ + _(int64_t, Long) \ + _(c10::Half, Half) \ + _(float, Float) \ + _(double, Double) \ + _(c10::complex, ComplexHalf) \ + _(c10::complex, ComplexFloat) \ + _(c10::complex, ComplexDouble) \ + _(bool, Bool) \ + _(c10::BFloat16, BFloat16) \ + _(c10::Float8_e5m2, Float8_e5m2) \ + _(c10::Float8_e4m3fn, Float8_e4m3fn) \ + _(c10::Float8_e5m2fnuz, Float8_e5m2fnuz) \ + _(c10::Float8_e4m3fnuz, Float8_e4m3fnuz) \ + _(c10::Float8_e8m0fnu, Float8_e8m0fnu) + +// NB: Order matters for this macro; it is relied upon in +// _promoteTypesLookup and the serialization format. +#define AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(_) \ + _(uint8_t, Byte) /* 0 */ \ + _(int8_t, Char) /* 1 */ \ + _(int16_t, Short) /* 2 */ \ + _(int, Int) /* 3 */ \ + _(int64_t, Long) /* 4 */ \ + _(c10::Half, Half) /* 5 */ \ + _(float, Float) /* 6 */ \ + _(double, Double) /* 7 */ \ + _(c10::complex, ComplexHalf) /* 8 */ \ + _(c10::complex, ComplexFloat) /* 9 */ \ + _(c10::complex, ComplexDouble) /* 10 */ \ + _(bool, Bool) /* 11 */ \ + _(c10::qint8, QInt8) /* 12 */ \ + _(c10::quint8, QUInt8) /* 13 */ \ + _(c10::qint32, QInt32) /* 14 */ \ + _(c10::BFloat16, BFloat16) /* 15 */ \ + _(c10::quint4x2, QUInt4x2) /* 16 */ \ + _(c10::quint2x4, QUInt2x4) /* 17 */ \ + _(c10::bits1x8, Bits1x8) /* 18 */ \ + _(c10::bits2x4, Bits2x4) /* 19 */ \ + _(c10::bits4x2, Bits4x2) /* 20 */ \ + _(c10::bits8, Bits8) /* 21 */ \ + _(c10::bits16, Bits16) /* 22 */ \ + _(c10::Float8_e5m2, Float8_e5m2) /* 23 */ \ + _(c10::Float8_e4m3fn, Float8_e4m3fn) /* 24 */ \ + _(c10::Float8_e5m2fnuz, Float8_e5m2fnuz) /* 25 */ \ + _(c10::Float8_e4m3fnuz, Float8_e4m3fnuz) /* 26 */ \ + _(uint16_t, UInt16) /* 27 */ \ + _(uint32_t, UInt32) /* 28 */ \ + _(uint64_t, UInt64) /* 29 */ \ + _(c10::dummy_uint1_7_t<1>, UInt1) /* 30 */ \ + _(c10::dummy_uint1_7_t<2>, UInt2) /* 31 */ \ + _(c10::dummy_uint1_7_t<3>, UInt3) /* 32 */ \ + _(c10::dummy_uint1_7_t<4>, UInt4) /* 33 */ \ + _(c10::dummy_uint1_7_t<5>, UInt5) /* 34 */ \ + _(c10::dummy_uint1_7_t<6>, UInt6) /* 35 */ \ + _(c10::dummy_uint1_7_t<7>, UInt7) /* 36 */ \ + _(c10::dummy_int1_7_t<1>, Int1) /* 37 */ \ + _(c10::dummy_int1_7_t<2>, Int2) /* 38 */ \ + _(c10::dummy_int1_7_t<3>, Int3) /* 39 */ \ + _(c10::dummy_int1_7_t<4>, Int4) /* 40 */ \ + _(c10::dummy_int1_7_t<5>, Int5) /* 41 */ \ + _(c10::dummy_int1_7_t<6>, Int6) /* 42 */ \ + _(c10::dummy_int1_7_t<7>, Int7) /* 43 */ \ + _(c10::Float8_e8m0fnu, Float8_e8m0fnu) /* 44 */ \ + _(c10::Float4_e2m1fn_x2, Float4_e2m1fn_x2) /* 45 */ + +// NB: despite its generic sounding name, the macros that don't take _AND +// are mostly only used by tensorexpr +#define AT_FORALL_INT_TYPES(_) \ + _(uint8_t, Byte) \ + _(int8_t, Char) \ + _(int16_t, Short) \ + _(int, Int) \ + _(int64_t, Long) + +#define AT_FORALL_SCALAR_TYPES(_) \ + _(uint8_t, Byte) \ + _(int8_t, Char) \ + _(int16_t, Short) \ + _(int, Int) \ + _(int64_t, Long) \ + _(float, Float) \ + _(double, Double) + +// These macros are often controlling how many template instantiations we +// create for kernels. It is typically inappropriate to add new dtypes here, +// instead, new types should be added to use sites on a case-by-case basis. +// We generally are not accepting new dtypes due to binary size concerns. + +#define AT_FORALL_SCALAR_TYPES_AND(SCALARTYPE, _) \ + _(uint8_t, Byte) \ + _(int8_t, Char) \ + _(int16_t, Short) \ + _(int, Int) \ + _(int64_t, Long) \ + _(float, Float) \ + _(double, Double) \ + _(c10::impl::ScalarTypeToCPPTypeT, SCALARTYPE) + +#define AT_FORALL_SCALAR_TYPES_AND2(SCALARTYPE1, SCALARTYPE2, _) \ + _(uint8_t, Byte) \ + _(int8_t, Char) \ + _(int16_t, Short) \ + _(int, Int) \ + _(int64_t, Long) \ + _(float, Float) \ + _(double, Double) \ + _(c10::impl::ScalarTypeToCPPTypeT, \ + SCALARTYPE1) \ + _(c10::impl::ScalarTypeToCPPTypeT, SCALARTYPE2) + +#define AT_FORALL_SCALAR_TYPES_AND3(SCALARTYPE1, SCALARTYPE2, SCALARTYPE3, _) \ + _(uint8_t, Byte) \ + _(int8_t, Char) \ + _(int16_t, Short) \ + _(int, Int) \ + _(int64_t, Long) \ + _(float, Float) \ + _(double, Double) \ + _(c10::impl::ScalarTypeToCPPTypeT, \ + SCALARTYPE1) \ + _(c10::impl::ScalarTypeToCPPTypeT, \ + SCALARTYPE2) \ + _(c10::impl::ScalarTypeToCPPTypeT, SCALARTYPE3) + +#define AT_FORALL_SCALAR_TYPES_AND7( \ + SCALARTYPE1, \ + SCALARTYPE2, \ + SCALARTYPE3, \ + SCALARTYPE4, \ + SCALARTYPE5, \ + SCALARTYPE6, \ + SCALARTYPE7, \ + _) \ + _(uint8_t, Byte) \ + _(int8_t, Char) \ + _(int16_t, Short) \ + _(int, Int) \ + _(int64_t, Long) \ + _(float, Float) \ + _(double, Double) \ + _(c10::impl::ScalarTypeToCPPTypeT, \ + SCALARTYPE1) \ + _(c10::impl::ScalarTypeToCPPTypeT, \ + SCALARTYPE2) \ + _(c10::impl::ScalarTypeToCPPTypeT, \ + SCALARTYPE3) \ + _(c10::impl::ScalarTypeToCPPTypeT, \ + SCALARTYPE4) \ + _(c10::impl::ScalarTypeToCPPTypeT, \ + SCALARTYPE5) \ + _(c10::impl::ScalarTypeToCPPTypeT, \ + SCALARTYPE6) \ + _(c10::impl::ScalarTypeToCPPTypeT, SCALARTYPE7) + +#define AT_FORALL_QINT_TYPES(_) \ + _(c10::qint8, QInt8) \ + _(c10::quint8, QUInt8) \ + _(c10::qint32, QInt32) \ + _(c10::quint4x2, QUInt4x2) \ + _(c10::quint2x4, QUInt2x4) + +#define AT_FORALL_FLOAT8_TYPES(_) \ + _(c10::Float8_e5m2, Float8_e5m2) \ + _(c10::Float8_e4m3fn, Float8_e4m3fn) \ + _(c10::Float8_e5m2fnuz, Float8_e5m2fnuz) \ + _(c10::Float8_e4m3fnuz, Float8_e4m3fnuz) \ + _(c10::Float8_e8m0fnu, Float8_e8m0fnu) + +#define AT_FORALL_COMPLEX_TYPES(_) \ + _(c10::complex, ComplexFloat) \ + _(c10::complex, ComplexDouble) + +enum class ScalarType : int8_t { +#define DEFINE_ST_ENUM_VAL_(_1, n) n, + AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(DEFINE_ST_ENUM_VAL_) +#undef DEFINE_ENUM_ST_ENUM_VAL_ + Undefined, + NumOptions +}; + +constexpr uint16_t NumScalarTypes = + static_cast(ScalarType::NumOptions); + +// Map from C++ type to ScalarType enum +template +struct CppTypeToScalarType; + +#define SPECIALIZE_CppTypeToScalarType(cpp_type, scalar_type) \ + template <> \ + struct CppTypeToScalarType \ + : std:: \ + integral_constant { \ + }; + +AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(SPECIALIZE_CppTypeToScalarType) + +#undef SPECIALIZE_CppTypeToScalarType + +namespace impl { + +// These are used to map ScalarTypes to C++ types. + +template +struct ScalarTypeToCPPType; + +#define SPECIALIZE_ScalarTypeToCPPType(cpp_type, scalar_type) \ + template <> \ + struct ScalarTypeToCPPType { \ + using type = cpp_type; \ + \ + /* This is a workaround for the CUDA bug which prevents */ \ + /* ::detail::ScalarTypeToCType::type being used directly due to */ \ + /* ambiguous reference which can't to be resolved. For some reason it */ \ + /* can't pick between at::detail and at::cuda::detail. */ \ + /* For repro example, please see: */ \ + /* https://gist.github.com/izdeby/952ae7cf256ddb740a73776d39a7e7ba */ \ + /* UPDATE: while the CUDA bug is fixed, we cannot remove the */ \ + /* workaround as it is BC breaking. However, it is recommended to */ \ + /* update any code that contains */ \ + /* decltype(ScalarTypeToCPPType::t) */ \ + /* with */ \ + /* ScalarTypeToCPPTypeT */ \ + static type t; \ + }; + +AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(SPECIALIZE_ScalarTypeToCPPType) + +#undef SPECIALIZE_ScalarTypeToCPPType + +template +using ScalarTypeToCPPTypeT = typename ScalarTypeToCPPType::type; + +} // namespace impl + +inline const char* toString(ScalarType t) { +#define DEFINE_CASE(_, name) \ + case ScalarType::name: \ + return #name; + + switch (t) { + AT_FORALL_SCALAR_TYPES_WITH_COMPLEX_AND_QINTS(DEFINE_CASE) + default: + return "UNKNOWN_SCALAR"; + } +#undef DEFINE_CASE +} + +inline std::ostream& operator<<( + std::ostream& stream, + c10::ScalarType scalar_type) { + return stream << toString(scalar_type); +} + +inline bool isQIntType(ScalarType t) { + // Don't forget to extend this when adding new QInt types + return t == ScalarType::QInt8 || t == ScalarType::QUInt8 || + t == ScalarType::QInt32 || t == ScalarType::QUInt4x2 || + t == ScalarType::QUInt2x4; +} + +inline ScalarType toUnderlying(ScalarType t) { + switch (t) { + case ScalarType::QUInt8: + case ScalarType::QUInt4x2: + [[fallthrough]]; + case ScalarType::QUInt2x4: + return ScalarType::Byte; + case ScalarType::QInt8: + return ScalarType::Char; + case ScalarType::QInt32: + return ScalarType::Int; + default: + return t; + } +} + +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) +using c10::CppTypeToScalarType; +using c10::dummy_int1_7_t; +using c10::dummy_uint1_7_t; +using c10::NumScalarTypes; +using c10::ScalarType; +using c10::toString; +using c10::operator<<; +using c10::isQIntType; +using c10::toUnderlying; + +namespace impl { +using c10::impl::ScalarTypeToCPPTypeT; +} // namespace impl + +HIDDEN_NAMESPACE_END(torch, headeronly) + +C10_DIAGNOSTIC_POP() diff --git a/backends/pytorch/include/torch/headeronly/core/TensorAccessor.h b/backends/pytorch/include/torch/headeronly/core/TensorAccessor.h new file mode 100644 index 000000000..9019c7ac3 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/core/TensorAccessor.h @@ -0,0 +1,462 @@ +#pragma once + +#include +#include +#include + +#include +#include +#include +#include + +namespace torch::headeronly { + +// The PtrTraits argument to the TensorAccessor/GenericPackedTensorAccessor +// is used to enable the __restrict__ keyword/modifier for the data +// passed to cuda. +template +struct DefaultPtrTraits { + typedef T* PtrType; +}; + +#if defined(__CUDACC__) || defined(__HIPCC__) +template +struct RestrictPtrTraits { + typedef T* __restrict__ PtrType; +}; +#endif + +namespace detail { +// Template classes in torch::headeronly::detail namespace are used +// to construct accessor template classes with custom ArrayRef and +// index bound check implementations. For instance, +// at::TensorAccessor and torch::headeronly::TensorAccessor template +// classes use c10::IntArrayRef and +// torch::headeronly::IntHeaderOnlyArrayRef classes, respectively, +// as return value types of sizes() and strides() methods. + +// TensorAccessorBase and TensorAccessor are used for both CPU and CUDA tensors. +// For CUDA tensors it is used in device code (only). This means that we +// restrict ourselves to functions and types available there (e.g. IntArrayRef +// isn't). + +// The PtrTraits argument is only relevant to cuda to support `__restrict__` +// pointers. +template < + class ArrayRefCls, + typename T, + size_t N, + template class PtrTraits = DefaultPtrTraits, + typename index_t = int64_t> +class TensorAccessorBase { + public: + typedef typename PtrTraits::PtrType PtrType; + + C10_HOST_DEVICE TensorAccessorBase( + PtrType data_, + const index_t* sizes_, + const index_t* strides_) + : data_(data_), sizes_(sizes_), strides_(strides_) {} + C10_HOST ArrayRefCls sizes() const { + return ArrayRefCls(sizes_, N); + } + C10_HOST ArrayRefCls strides() const { + return ArrayRefCls(strides_, N); + } + C10_HOST_DEVICE index_t stride(index_t i) const { + return strides_[i]; + } + C10_HOST_DEVICE index_t size(index_t i) const { + return sizes_[i]; + } + C10_HOST_DEVICE PtrType data() { + return data_; + } + C10_HOST_DEVICE const PtrType data() const { + return data_; + } + + protected: + PtrType data_; + const index_t* sizes_; + const index_t* strides_; +}; + +// The `TensorAccessor` is typically instantiated for CPU `Tensor`s using +// `Tensor.accessor()`. +// For CUDA `Tensor`s, `GenericPackedTensorAccessor` is used on the host and +// only indexing on the device uses `TensorAccessor`s. +template < + class ArrayRefCls, + typename T, + size_t N, + template class PtrTraits = DefaultPtrTraits, + typename index_t = int64_t> +class TensorAccessor + : public TensorAccessorBase { + public: + typedef typename PtrTraits::PtrType PtrType; + + C10_HOST_DEVICE TensorAccessor( + PtrType data_, + const index_t* sizes_, + const index_t* strides_) + : TensorAccessorBase( + data_, + sizes_, + strides_) {} + + C10_HOST_DEVICE TensorAccessor + operator[](index_t i) { + return TensorAccessor( + this->data_ + this->strides_[0] * i, + this->sizes_ + 1, + this->strides_ + 1); + } + + C10_HOST_DEVICE const TensorAccessor< + ArrayRefCls, + T, + N - 1, + PtrTraits, + index_t> + operator[](index_t i) const { + return TensorAccessor( + this->data_ + this->strides_[0] * i, + this->sizes_ + 1, + this->strides_ + 1); + } +}; + +template < + class ArrayRefCls, + typename T, + template class PtrTraits, + typename index_t> +class TensorAccessor + : public TensorAccessorBase { + public: + typedef typename PtrTraits::PtrType PtrType; + + C10_HOST_DEVICE TensorAccessor( + PtrType data_, + const index_t* sizes_, + const index_t* strides_) + : TensorAccessorBase( + data_, + sizes_, + strides_) {} + C10_HOST_DEVICE T& operator[](index_t i) { + // NOLINTNEXTLINE(clang-analyzer-core.NullDereference) + return this->data_[this->strides_[0] * i]; + } + C10_HOST_DEVICE const T& operator[](index_t i) const { + return this->data_[this->strides_[0] * i]; + } +}; + +// GenericPackedTensorAccessorBase and GenericPackedTensorAccessor are used on +// for CUDA `Tensor`s on the host and as in contrast to `TensorAccessor`s, they +// copy the strides and sizes on instantiation (on the host) in order to +// transfer them on the device when calling kernels. On the device, indexing of +// multidimensional tensors gives to `TensorAccessor`s. Use RestrictPtrTraits as +// PtrTraits if you want the tensor's data pointer to be marked as __restrict__. +// Instantiation from data, sizes, strides is only needed on the host and +// std::copy isn't available on the device, so those functions are host only. +template < + typename IndexBoundsCheck, + typename T, + size_t N, + template class PtrTraits = DefaultPtrTraits, + typename index_t = int64_t> +class GenericPackedTensorAccessorBase { + public: + typedef typename PtrTraits::PtrType PtrType; + C10_HOST GenericPackedTensorAccessorBase( + PtrType data_, + const index_t* sizes_, + const index_t* strides_) + : data_(data_) { + std::copy(sizes_, sizes_ + N, std::begin(this->sizes_)); + std::copy(strides_, strides_ + N, std::begin(this->strides_)); + } + + // if index_t is not int64_t, we want to have an int64_t constructor + template < + typename source_index_t, + class = std::enable_if_t>> + C10_HOST GenericPackedTensorAccessorBase( + PtrType data_, + const source_index_t* sizes_, + const source_index_t* strides_) + : data_(data_) { + for (size_t i = 0; i < N; ++i) { + this->sizes_[i] = sizes_[i]; + this->strides_[i] = strides_[i]; + } + } + + C10_HOST_DEVICE index_t stride(index_t i) const { + return strides_[i]; + } + C10_HOST_DEVICE index_t size(index_t i) const { + return sizes_[i]; + } + C10_HOST_DEVICE PtrType data() { + return data_; + } + C10_HOST_DEVICE const PtrType data() const { + return data_; + } + + protected: + PtrType data_; + // NOLINTNEXTLINE(*c-arrays*) + index_t sizes_[N]; + // NOLINTNEXTLINE(*c-arrays*) + index_t strides_[N]; + C10_HOST void bounds_check_(index_t i) const { + IndexBoundsCheck _(i); + } +}; + +template < + typename ItemAccessor, + typename IndexBoundsCheck, + typename T, + size_t N, + template class PtrTraits = DefaultPtrTraits, + typename index_t = int64_t> +class GenericPackedTensorAccessor : public GenericPackedTensorAccessorBase< + IndexBoundsCheck, + T, + N, + PtrTraits, + index_t> { + public: + typedef typename PtrTraits::PtrType PtrType; + + C10_HOST GenericPackedTensorAccessor( + PtrType data_, + const index_t* sizes_, + const index_t* strides_) + : GenericPackedTensorAccessorBase< + IndexBoundsCheck, + T, + N, + PtrTraits, + index_t>(data_, sizes_, strides_) {} + + // if index_t is not int64_t, we want to have an int64_t constructor + template < + typename source_index_t, + class = std::enable_if_t>> + C10_HOST GenericPackedTensorAccessor( + PtrType data_, + const source_index_t* sizes_, + const source_index_t* strides_) + : GenericPackedTensorAccessorBase< + IndexBoundsCheck, + T, + N, + PtrTraits, + index_t>(data_, sizes_, strides_) {} + + C10_DEVICE ItemAccessor operator[](index_t i) { + index_t* new_sizes = this->sizes_ + 1; + index_t* new_strides = this->strides_ + 1; + return ItemAccessor( + this->data_ + this->strides_[0] * i, new_sizes, new_strides); + } + + C10_DEVICE const ItemAccessor operator[](index_t i) const { + const index_t* new_sizes = this->sizes_ + 1; + const index_t* new_strides = this->strides_ + 1; + return ItemAccessor( + this->data_ + this->strides_[0] * i, new_sizes, new_strides); + } + + /// Returns a PackedTensorAccessor of the same dimension after transposing the + /// two dimensions given. Does not actually move elements; transposition is + /// made by permuting the size/stride arrays. If the dimensions are not valid, + /// asserts. + C10_HOST GenericPackedTensorAccessor< + ItemAccessor, + IndexBoundsCheck, + T, + N, + PtrTraits, + index_t> + transpose(index_t dim1, index_t dim2) const { + this->bounds_check_(dim1); + this->bounds_check_(dim2); + GenericPackedTensorAccessor< + ItemAccessor, + IndexBoundsCheck, + T, + N, + PtrTraits, + index_t> + result(this->data_, this->sizes_, this->strides_); + std::swap(result.strides_[dim1], result.strides_[dim2]); + std::swap(result.sizes_[dim1], result.sizes_[dim2]); + return result; + } +}; + +template < + typename ItemAccessor, + typename IndexBoundsCheck, + typename T, + template class PtrTraits, + typename index_t> +class GenericPackedTensorAccessor< + ItemAccessor, + IndexBoundsCheck, + T, + 1, + PtrTraits, + index_t> + : public GenericPackedTensorAccessorBase< + IndexBoundsCheck, + T, + 1, + PtrTraits, + index_t> { + public: + typedef typename PtrTraits::PtrType PtrType; + C10_HOST GenericPackedTensorAccessor( + PtrType data_, + const index_t* sizes_, + const index_t* strides_) + : GenericPackedTensorAccessorBase< + IndexBoundsCheck, + T, + 1, + PtrTraits, + index_t>(data_, sizes_, strides_) {} + + // if index_t is not int64_t, we want to have an int64_t constructor + template < + typename source_index_t, + class = std::enable_if_t>> + C10_HOST GenericPackedTensorAccessor( + PtrType data_, + const source_index_t* sizes_, + const source_index_t* strides_) + : GenericPackedTensorAccessorBase< + IndexBoundsCheck, + T, + 1, + PtrTraits, + index_t>(data_, sizes_, strides_) {} + + C10_DEVICE T& operator[](index_t i) { + return this->data_[this->strides_[0] * i]; + } + C10_DEVICE const T& operator[](index_t i) const { + return this->data_[this->strides_[0] * i]; + } + + // Same as in the general N-dimensional case, but note that in the + // 1-dimensional case the returned PackedTensorAccessor will always be an + // identical copy of the original + C10_HOST GenericPackedTensorAccessor< + ItemAccessor, + IndexBoundsCheck, + T, + 1, + PtrTraits, + index_t> + transpose(index_t dim1, index_t dim2) const { + this->bounds_check_(dim1); + this->bounds_check_(dim2); + return GenericPackedTensorAccessor< + ItemAccessor, + IndexBoundsCheck, + T, + 1, + PtrTraits, + index_t>(this->data_, this->sizes_, this->strides_); + } +}; + +template +struct HeaderOnlyIndexBoundsCheck { + HeaderOnlyIndexBoundsCheck(index_t i) { + STD_TORCH_CHECK( + 0 <= i && i < index_t{N}, + "Index ", + i, + " is not within bounds of a tensor of dimension ", + N); + } +}; + +} // namespace detail + +// HeaderOnlyTensorAccessorBase is same as at::TensorAccessorBase +// except sizes() and strides() return IntHeaderOnlyArrayRef instead +// of IntArrayRef. +template < + typename T, + size_t N, + template class PtrTraits = DefaultPtrTraits, + typename index_t = int64_t> +using HeaderOnlyTensorAccessorBase = detail::TensorAccessorBase< + torch::headeronly::IntHeaderOnlyArrayRef, + T, + N, + PtrTraits, + index_t>; + +// HeaderOnlyTensorAccessor is same as at::TensorAccessor except +// sizes() and strides() return IntHeaderOnlyArrayRef instead of +// IntArrayRef. +template < + typename T, + size_t N, + template class PtrTraits = DefaultPtrTraits, + typename index_t = int64_t> +using HeaderOnlyTensorAccessor = detail::TensorAccessor< + torch::headeronly::IntHeaderOnlyArrayRef, + T, + N, + PtrTraits, + index_t>; + +// HeaderOnlyGenericPackedTensorAccessorBase is same as +// at::GenericPackedTensorAccessorBase except sizes() and strides() +// return IntHeaderOnlyArrayRef instead of IntArrayRef. +template < + typename T, + size_t N, + template class PtrTraits = DefaultPtrTraits, + typename index_t = int64_t> +using HeaderOnlyGenericPackedTensorAccessorBase = + detail::GenericPackedTensorAccessorBase< + detail::HeaderOnlyIndexBoundsCheck, + T, + N, + PtrTraits, + index_t>; + +// HeaderOnlyGenericPackedTensorAccessor is same as +// at::GenericPackedTensorAccessor except sizes() and strides() return +// IntHeaderOnlyArrayRef instead of IntArrayRef, and bounds check uses +// STD_TORCH_CHECK instead of TORCH_CHECK_INDEX. +template < + typename T, + size_t N, + template class PtrTraits = DefaultPtrTraits, + typename index_t = int64_t> +using HeaderOnlyGenericPackedTensorAccessor = + detail::GenericPackedTensorAccessor< + HeaderOnlyTensorAccessor, + detail::HeaderOnlyIndexBoundsCheck, + T, + N, + PtrTraits, + index_t>; + +} // namespace torch::headeronly diff --git a/backends/pytorch/include/torch/headeronly/macros/Export.h b/backends/pytorch/include/torch/headeronly/macros/Export.h new file mode 100644 index 000000000..8dd25419e --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/macros/Export.h @@ -0,0 +1,153 @@ +#pragma once + +#ifndef C10_MACROS_EXPORT_H_ +#define C10_MACROS_EXPORT_H_ + +#ifndef C10_USING_CUSTOM_GENERATED_MACROS +#include +#endif // C10_USING_CUSTOM_GENERATED_MACROS + +/* Header file to define the common scaffolding for exported symbols. + * + * Export is by itself a quite tricky situation to deal with, and if you are + * hitting this file, make sure you start with the background here: + * - Linux: https://gcc.gnu.org/wiki/Visibility + * - Windows: + * https://docs.microsoft.com/en-us/cpp/cpp/dllexport-dllimport?view=vs-2017 + * + * Do NOT include this file directly. Instead, use c10/macros/Macros.h + */ + +// You do not need to edit this part of file unless you are changing the core +// pytorch export abstractions. +// +// This part defines the C10 core export and import macros. This is controlled +// by whether we are building shared libraries or not, which is determined +// during build time and codified in c10/core/cmake_macros.h. +// When the library is built as a shared lib, EXPORT and IMPORT will contain +// visibility attributes. If it is being built as a static lib, then EXPORT +// and IMPORT basically have no effect. + +// As a rule of thumb, you should almost NEVER mix static and shared builds for +// libraries that depend on c10. AKA, if c10 is built as a static library, we +// recommend everything dependent on c10 to be built statically. If c10 is built +// as a shared library, everything dependent on it should be built as shared. In +// the PyTorch project, all native libraries shall use the macro +// C10_BUILD_SHARED_LIB to check whether pytorch is building shared or static +// libraries. + +// For build systems that do not directly depend on CMake and directly build +// from the source directory (such as Buck), one may not have a cmake_macros.h +// file at all. In this case, the build system is responsible for providing +// correct macro definitions corresponding to the cmake_macros.h.in file. +// +// In such scenarios, one should define the macro +// C10_USING_CUSTOM_GENERATED_MACROS +// to inform this header that it does not need to include the cmake_macros.h +// file. + +#ifdef _WIN32 +#define C10_HIDDEN +#if defined(C10_BUILD_SHARED_LIBS) +#define C10_EXPORT __declspec(dllexport) +#define C10_IMPORT __declspec(dllimport) +#else +#define C10_EXPORT +#define C10_IMPORT +#endif +#else // _WIN32 +#if defined(__GNUC__) +#define C10_EXPORT __attribute__((__visibility__("default"))) +#define C10_HIDDEN __attribute__((__visibility__("hidden"))) +#else // defined(__GNUC__) +#define C10_EXPORT +#define C10_HIDDEN +#endif // defined(__GNUC__) +#define C10_IMPORT C10_EXPORT +#endif // _WIN32 + +#ifdef NO_EXPORT +#undef C10_EXPORT +#define C10_EXPORT +#endif + +// Definition of an adaptive XX_API macro, that depends on whether you are +// building the library itself or not, routes to XX_EXPORT and XX_IMPORT. +// Basically, you will need to do this for each shared library that you are +// building, and the instruction is as follows: assuming that you are building +// a library called libawesome.so. You should: +// (1) for your cmake target (usually done by "add_library(awesome, ...)"), +// define a macro called AWESOME_BUILD_MAIN_LIB using +// target_compile_options. +// (2) define the AWESOME_API macro similar to the one below. +// And in the source file of your awesome library, use AWESOME_API to +// annotate public symbols. + +// Here, for the C10 library, we will define the macro C10_API for both import +// and export. + +// This one is being used by libc10.so +#ifdef C10_BUILD_MAIN_LIB +#define C10_API C10_EXPORT +#else +#define C10_API C10_IMPORT +#endif + +// This one is being used by libtorch.so +#ifdef CAFFE2_BUILD_MAIN_LIB +#define TORCH_API C10_EXPORT +#else +#define TORCH_API C10_IMPORT +#endif + +// You may be wondering why we have TORCH_CUDA_CPP_API and TORCH_CUDA_CU_API +// belonging to the same library instead of just one TORCH_CUDA_API. Well, it +// can indeed just be one TORCH_CUDA_API (and used to be)! TORCH_CUDA_CPP_API +// and TORCH_CUDA_CU_API are artifacts of when we needed a split build to +// avoid relocation marker linking errors. The context is as follows: +// +// Once upon a time, there _was_ only TORCH_CUDA_API. All was happy until we +// tried to compile PyTorch for CUDA 11.1, which ran into relocation marker +// issues when linking big binaries. +// (https://github.com/pytorch/pytorch/issues/39968) We had two choices: +// (1) Stop supporting so many GPU architectures +// (2) Do something else +// We chose #2 and decided to split the behemoth that was torch_cuda into two +// smaller libraries, one with most of the core kernel functions (torch_cuda_cu) +// and the other that had..well..everything else (torch_cuda_cpp). The idea was +// this: instead of linking our static libraries (like the hefty +// libcudnn_static.a) with another huge library, torch_cuda, and run into pesky +// relocation marker issues, we could link our static libraries to a smaller +// part of torch_cuda (torch_cuda_cpp) and avoid the issues. + +// libtorch_cuda.so (where torch_cuda_cu and torch_cuda_cpp are a part of the +// same api) +#ifdef TORCH_CUDA_BUILD_MAIN_LIB +#define TORCH_CUDA_CPP_API C10_EXPORT +#define TORCH_CUDA_CU_API C10_EXPORT +#else +#define TORCH_CUDA_CPP_API C10_IMPORT +#define TORCH_CUDA_CU_API C10_IMPORT +#endif + +#if defined(TORCH_HIP_BUILD_MAIN_LIB) +#define TORCH_HIP_CPP_API C10_EXPORT +#define TORCH_HIP_API C10_EXPORT +#else +#define TORCH_HIP_CPP_API C10_IMPORT +#define TORCH_HIP_API C10_IMPORT +#endif + +#if defined(TORCH_XPU_BUILD_MAIN_LIB) +#define TORCH_XPU_API C10_EXPORT +#else +#define TORCH_XPU_API C10_IMPORT +#endif + +// Enums only need to be exported on windows for non-CUDA files +#if defined(_WIN32) && defined(__CUDACC__) +#define C10_API_ENUM C10_API +#else +#define C10_API_ENUM +#endif +#endif // C10_MACROS_EXPORT_H_ diff --git a/backends/pytorch/include/torch/headeronly/macros/Macros.h b/backends/pytorch/include/torch/headeronly/macros/Macros.h new file mode 100644 index 000000000..63aa0d20d --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/macros/Macros.h @@ -0,0 +1,694 @@ +#ifndef C10_MACROS_MACROS_H_ +#define C10_MACROS_MACROS_H_ + +#ifdef __cplusplus +#include +#else +#include +#endif + +/* Main entry for torch/headeronly/macros (used to be c10/macros). + * + * In your code, include torch/headeronly/macros/Macros.h directly, instead of + * individual files in this folder. + */ + +// For build systems that do not directly depend on CMake and directly build +// from the source directory (such as Buck), one may not have a cmake_macros.h +// file at all. In this case, the build system is responsible for providing +// correct macro definitions corresponding to the cmake_macros.h.in file. +// +// In such scenarios, one should define the macro +// C10_USING_CUSTOM_GENERATED_MACROS +// to inform this header that it does not need to include the cmake_macros.h +// file. + +#ifndef C10_USING_CUSTOM_GENERATED_MACROS +#include +#endif // C10_USING_CUSTOM_GENERATED_MACROS + +#include + +#if defined(__clang__) +#define __ubsan_ignore_float_divide_by_zero__ \ + __attribute__((no_sanitize("float-divide-by-zero"))) +#define __ubsan_ignore_undefined__ __attribute__((no_sanitize("undefined"))) +#define __ubsan_ignore_signed_int_overflow__ \ + __attribute__((no_sanitize("signed-integer-overflow"))) +#define __ubsan_ignore_pointer_overflow__ \ + __attribute__((no_sanitize("pointer-overflow"))) +#define __ubsan_ignore_function__ __attribute__((no_sanitize("function"))) +#define __ubsan_ignore_float_cast_overflow__ \ + __attribute__((no_sanitize("float-cast-overflow"))) +#else +#define __ubsan_ignore_float_divide_by_zero__ +#define __ubsan_ignore_undefined__ +#define __ubsan_ignore_signed_int_overflow__ +#define __ubsan_ignore_pointer_overflow__ +#define __ubsan_ignore_function__ +#define __ubsan_ignore_float_cast_overflow__ +#endif + +// Detect address sanitizer as some stuff doesn't work with it +#undef C10_ASAN_ENABLED + +// for clang +#if defined(__has_feature) +#if ((__has_feature(address_sanitizer))) +#define C10_ASAN_ENABLED 1 +#endif +#endif + +// for gcc +#if defined(__SANITIZE_ADDRESS__) +#if __SANITIZE_ADDRESS__ +#if !defined(C10_ASAN_ENABLED) +#define C10_ASAN_ENABLED 1 +#endif +#endif +#endif + +#if !defined(C10_ASAN_ENABLED) +#define C10_ASAN_ENABLED 0 +#endif + +// Detect undefined-behavior sanitizer (UBSAN) +#undef C10_UBSAN_ENABLED + +// for clang or gcc >= 14 +// NB: gcc 14 adds support for Clang's __has_feature +// https://gcc.gnu.org/gcc-14/changes.html +// gcc < 14 doesn't have a macro for UBSAN +// (e.g. __SANITIZE_UNDEFINED__ does not exist in gcc) +// https://github.com/google/sanitizers/issues/765 +#if defined(__has_feature) +#if ((__has_feature(undefined_behavior_sanitizer))) +#define C10_UBSAN_ENABLED 1 +#endif +#endif + +#if !defined(C10_UBSAN_ENABLED) +#define C10_UBSAN_ENABLED 0 +#endif + +// Disable the copy and assignment operator for a class. Note that this will +// disable the usage of the class in std containers. +#define C10_DISABLE_COPY_AND_ASSIGN(classname) \ + classname(const classname&) = delete; \ + classname& operator=(const classname&) = delete + +#define C10_CONCATENATE_IMPL(s1, s2) s1##s2 +#define C10_CONCATENATE(s1, s2) C10_CONCATENATE_IMPL(s1, s2) + +#define C10_MACRO_EXPAND(args) args + +#define C10_STRINGIZE_IMPL(x) #x +#define C10_STRINGIZE(x) C10_STRINGIZE_IMPL(x) + +/** + * C10_ANONYMOUS_VARIABLE(str) introduces a new identifier which starts with + * str and ends with a unique number. + */ +#ifdef __COUNTER__ +#define C10_UID __COUNTER__ +#define C10_ANONYMOUS_VARIABLE(str) C10_CONCATENATE(str, __COUNTER__) +#else +#define C10_UID __LINE__ +#define C10_ANONYMOUS_VARIABLE(str) C10_CONCATENATE(str, __LINE__) +#endif + +#ifdef __has_cpp_attribute +#define C10_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x) +#else +#define C10_HAS_CPP_ATTRIBUTE(x) (0) +#endif + +#ifndef FBCODE_CAFFE2 +/// DEPRECATED: Warn if a type or return value is discarded. +#define C10_NODISCARD [[nodiscard]] + +/// DEPRECATED: Suppress an unused variable. +#define C10_UNUSED [[maybe_unused]] +#endif + +#if !defined(__has_attribute) +#define __has_attribute(x) 0 +#endif + +// Direct port of LLVM_ATTRIBUTE_USED. +#if __has_attribute(used) +#define C10_USED __attribute__((__used__)) +#else +#define C10_USED +#endif + +#define C10_RESTRICT __restrict + +#ifdef __cplusplus + +// Simply define the namespace, in case a dependent library want to refer to +// the c10 namespace but not any nontrivial files. +namespace c10 {} +namespace c10::cuda {} +namespace c10::hip {} +namespace c10::xpu {} + +// Since C10 is the core library for caffe2 (and aten), we will simply reroute +// all abstractions defined in c10 to be available in caffe2 as well. +// This is only for backwards compatibility. Please use the symbols from the +// c10 namespace where possible. +namespace caffe2 { +using namespace c10; +} +namespace at { +using namespace c10; +} +namespace at::cuda { +using namespace c10::cuda; +} // namespace at::cuda + +// WARNING!!! THIS IS A GIANT HACK!!! +// This line means you cannot simultaneously include c10/hip +// and c10/cuda and then use them from the at::cuda namespace. +// This is true in practice, because HIPIFY works inplace on +// files in ATen/cuda, so it assumes that c10::hip is available +// from at::cuda. This namespace makes that happen. When +// HIPIFY is no longer out-of-place, we can switch the cuda +// here to hip and everyone is happy. +namespace at::cuda { +using namespace c10::hip; +} // namespace at::cuda + +namespace at::xpu { +using namespace c10::xpu; +} // namespace at::xpu + +#endif // __cplusplus + +// C10_LIKELY/C10_UNLIKELY +// +// These macros provide parentheses, so you can use these macros as: +// +// if C10_LIKELY(some_expr) { +// ... +// } +// +// NB: static_cast to boolean is mandatory in C++, because __builtin_expect +// takes a long argument, which means you may trigger the wrong conversion +// without it. +// +#if defined(__GNUC__) || defined(__ICL) || defined(__clang__) +#define C10_LIKELY(expr) (__builtin_expect(static_cast(expr), 1)) +#define C10_UNLIKELY(expr) (__builtin_expect(static_cast(expr), 0)) +#else +#define C10_LIKELY(expr) (expr) +#define C10_UNLIKELY(expr) (expr) +#endif + +/// C10_NOINLINE - Functions whose declaration is annotated with this will not +/// be inlined. +#ifdef __GNUC__ +#define C10_NOINLINE __attribute__((noinline)) +#elif _MSC_VER +#define C10_NOINLINE __declspec(noinline) +#else +#define C10_NOINLINE +#endif + +#if defined(_MSC_VER) +#define C10_ALWAYS_INLINE __forceinline +#elif __has_attribute(always_inline) || defined(__GNUC__) +#define C10_ALWAYS_INLINE __attribute__((__always_inline__)) inline +#else +#define C10_ALWAYS_INLINE inline +#endif + +// Unlike C10_ALWAYS_INLINE, C10_ALWAYS_INLINE_ATTRIBUTE can be used +// on a lambda. +#if defined(_MSC_VER) +// MSVC 14.39 is reasonably recent and doesn't like +// [[msvc::forceinline]] on a lambda, so don't try to use it. +#define C10_ALWAYS_INLINE_ATTRIBUTE +#elif __has_attribute(always_inline) || defined(__GNUC__) +#define C10_ALWAYS_INLINE_ATTRIBUTE __attribute__((__always_inline__)) +#else +#define C10_ALWAYS_INLINE_ATTRIBUTE +#endif + +#if defined(_MSC_VER) +#define C10_ATTR_VISIBILITY_HIDDEN +#elif defined(__GNUC__) +#define C10_ATTR_VISIBILITY_HIDDEN __attribute__((__visibility__("hidden"))) +#else +#define C10_ATTR_VISIBILITY_HIDDEN +#endif + +#define C10_ERASE C10_ALWAYS_INLINE C10_ATTR_VISIBILITY_HIDDEN + +#ifdef __cplusplus +#include +#else +#include +#endif + +#ifdef __HIPCC__ +// Unlike CUDA, HIP requires a HIP header to be included for __host__ to work. +// We do this #include here so that C10_HOST_DEVICE and friends will Just Work. +// See https://github.com/ROCm/hip/issues/441 +#include +#endif + +#if defined(__CUDACC__) || defined(__HIPCC__) +// Designates functions callable from the host (CPU) and the device (GPU) +#define C10_HOST_DEVICE __host__ __device__ +#define C10_DEVICE __device__ +#define C10_HOST __host__ +// constants from +// (https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#features-and-technical-specifications) +// The maximum number of threads per multiprocessor is 1024 for Turing +// architecture (7.5), 1536 for Geforce Ampere (8.6)/Jetson Orin (8.7), and +// 2048 for all other architectures. You'll get warnings if you exceed these +// constants. Hence, the following macros adjust the input values from the user +// to resolve potential warnings. +#if __CUDA_ARCH__ == 750 +constexpr uint32_t CUDA_MAX_THREADS_PER_SM = 1024; +#elif __CUDA_ARCH__ == 860 || __CUDA_ARCH__ == 870 || __CUDA_ARCH__ == 890 || \ + __CUDA_ARCH__ == 1200 +constexpr uint32_t CUDA_MAX_THREADS_PER_SM = 1536; +#else +constexpr uint32_t CUDA_MAX_THREADS_PER_SM = 2048; +#endif +// CUDA_MAX_THREADS_PER_BLOCK is same for all architectures currently +constexpr uint32_t CUDA_MAX_THREADS_PER_BLOCK = 1024; +// CUDA_THREADS_PER_BLOCK_FALLBACK is the "canonical fallback" choice of block +// size. 256 is a good number for this fallback and should give good occupancy +// and versatility across all architectures. +constexpr uint32_t CUDA_THREADS_PER_BLOCK_FALLBACK = 256; +// NOTE: if you are thinking of constexpr-ify the inputs to launch bounds, it +// turns out that although __launch_bounds__ can take constexpr, it +// can't take a constexpr that has anything to do with templates. +// Currently we use launch_bounds that depend on template arguments in +// Loops.cuh, Reduce.cuh and LossCTC.cuh. Hence, C10_MAX_THREADS_PER_BLOCK +// and C10_MIN_BLOCKS_PER_SM are kept as macros. +// Suppose you were planning to write __launch_bounds__(a, b), based on your +// performance tuning on a modern GPU. Instead, you should write +// __launch_bounds__(C10_MAX_THREADS_PER_BLOCK(a), C10_MIN_BLOCKS_PER_SM(a, b)), +// which will also properly respect limits on old architectures. +#define C10_MAX_THREADS_PER_BLOCK(val) \ + (((val) <= CUDA_MAX_THREADS_PER_BLOCK) ? (val) \ + : CUDA_THREADS_PER_BLOCK_FALLBACK) +#define C10_MIN_BLOCKS_PER_SM(threads_per_block, blocks_per_sm) \ + ((((threads_per_block) * (blocks_per_sm) <= CUDA_MAX_THREADS_PER_SM) \ + ? (blocks_per_sm) \ + : ((CUDA_MAX_THREADS_PER_SM + (threads_per_block) - 1) / \ + (threads_per_block)))) +// C10_LAUNCH_BOUNDS is analogous to __launch_bounds__ +#define C10_LAUNCH_BOUNDS_0 \ + __launch_bounds__( \ + 256, 4) // default launch bounds that should give good occupancy and + // versatility across all architectures. +#define C10_LAUNCH_BOUNDS_1(max_threads_per_block) \ + __launch_bounds__((C10_MAX_THREADS_PER_BLOCK((max_threads_per_block)))) +#define C10_LAUNCH_BOUNDS_2(max_threads_per_block, min_blocks_per_sm) \ + __launch_bounds__( \ + (C10_MAX_THREADS_PER_BLOCK((max_threads_per_block))), \ + (C10_MIN_BLOCKS_PER_SM((max_threads_per_block), (min_blocks_per_sm)))) +#else +#define C10_HOST_DEVICE +#define C10_HOST +#define C10_DEVICE +#endif + +#if defined(USE_ROCM) +#define C10_HIP_HOST_DEVICE __host__ __device__ +#else +#define C10_HIP_HOST_DEVICE +#endif + +#if defined(USE_ROCM) +// C10_WARP_SIZE is only allowed for device code. +// Host code _must_ use at::cuda::warp_size() +// HIP header used to define warpSize as a constexpr that was either 32 or 64 +// depending on the target device, and then always set it to 64 for host code. +// Host pass of HIP compiler needs C10_WARP_SIZE defined to _something_ so we +// set it to something unreasonable to trigger obvious host code errors. + +namespace at::cuda { +TORCH_CUDA_CPP_API int warp_size(); +} +#ifdef __HIPCC__ +static inline int __host__ C10_WARP_SIZE_INTERNAL() { + return at::cuda::warp_size(); +} + +static inline constexpr int __device__ C10_WARP_SIZE_INTERNAL() { +#if defined(__GFX9__) + return 64; +#else // __GFX9__ + return 32; +#endif // __GFX9__ +} +#else // __HIPCC__ +static inline int C10_WARP_SIZE_INTERNAL() { + return at::cuda::warp_size(); +} +#endif // __HIPCC__ + +#define C10_WARP_SIZE (C10_WARP_SIZE_INTERNAL()) +#define C10_WARP_SIZE_STATIC 64 + +#else // defined(USE_ROCM) +#define C10_WARP_SIZE 32 +#endif + +#if defined(_MSC_VER) && _MSC_VER <= 1900 +#define __func__ __FUNCTION__ +#endif + +// CUDA_KERNEL_ASSERT checks the assertion +// even when NDEBUG is defined. This is useful for important assertions in CUDA +// code that would otherwise be suppressed when building Release. +#if defined(__ANDROID__) || defined(__APPLE__) || defined(__FreeBSD__) +// Those platforms do not support assert() +#define CUDA_KERNEL_ASSERT(cond) +#define CUDA_KERNEL_ASSERT_MSG(cond, msg) +#define CUDA_KERNEL_ASSERT_PRINTF(cond, msg, ...) +#define SYCL_KERNEL_ASSERT(cond) +#elif defined(_MSC_VER) +#if defined(NDEBUG) +extern "C" { +C10_IMPORT +#if defined(__SYCL_DEVICE_ONLY__) +extern SYCL_EXTERNAL void _wassert( + const wchar_t* wexpr, + const wchar_t* wfile, + unsigned line); +#else +#if defined(__CUDA_ARCH__) +__host__ __device__ +#endif // __CUDA_ARCH__ + void + _wassert(wchar_t const* _Message, wchar_t const* _File, unsigned _Line); +#endif // __SYCL_DEVICE_ONLY__ +} +#endif // NDEBUG +#define CUDA_KERNEL_ASSERT(cond) \ + if (C10_UNLIKELY(!(cond))) { \ + (void)(_wassert( \ + _CRT_WIDE(#cond), \ + _CRT_WIDE(__FILE__), \ + static_cast(__LINE__)), \ + 0); \ + } +// TODO: This doesn't assert the message because I (chilli) couldn't figure out +// a nice way to convert a char* to a wchar_t* +#define CUDA_KERNEL_ASSERT_MSG(cond, msg) \ + if (C10_UNLIKELY(!(cond))) { \ + (void)(_wassert( \ + _CRT_WIDE(#cond), \ + _CRT_WIDE(__FILE__), \ + static_cast(__LINE__)), \ + 0); \ + } +#define CUDA_KERNEL_ASSERT_PRINTF(cond, msg, ...) \ + if (C10_UNLIKELY(!(cond))) { \ + (void)(printf( \ + "[CUDA_KERNEL_ASSERT] " __FILE__ ":" C10_STRINGIZE( \ + __LINE__) ": %s: block: [%d,%d,%d], thread: [%d,%d,%d]: " \ + "Assertion failed: `" #cond "`: " msg "\n", \ + __func__, \ + blockIdx.x, \ + blockIdx.y, \ + blockIdx.z, \ + threadIdx.x, \ + threadIdx.y, \ + threadIdx.z, \ + ##__VA_ARGS__)); \ + (void)(_wassert( \ + _CRT_WIDE(#cond), \ + _CRT_WIDE(__FILE__), \ + static_cast(__LINE__)), \ + 0); \ + } +#define SYCL_KERNEL_ASSERT(cond) \ + if (C10_UNLIKELY(!(cond))) { \ + (void)(_wassert( \ + _CRT_WIDE(#cond), \ + _CRT_WIDE(__FILE__), \ + static_cast(__LINE__)), \ + 0); \ + } +#else // __APPLE__, _MSC_VER +#if defined(NDEBUG) +extern "C" { +#if defined(__SYCL_DEVICE_ONLY__) +extern SYCL_EXTERNAL void __assert_fail( + const char* expr, + const char* file, + unsigned int line, + const char* func); +#elif (defined(__EMSCRIPTEN__)) +// As defined in assert.h in the Emscripten stdlib +_Noreturn void __assert_fail( + const char* expr, + const char* file, + int line, + const char* func); +#else // __SYCL_DEVICE_ONLY__ +#if (defined(__CUDA_ARCH__) && !(defined(__clang__) && defined(__CUDA__))) +// CUDA supports __assert_fail function which are common for both device +// and host side code. +__host__ __device__ +#endif + + // This forward declaration matching the declaration of __assert_fail + // exactly how it is in glibc in case parts of the program are compiled with + // different NDEBUG settings. Otherwise we might get 'ambiguous declaration' + // error. Note: On ROCm - this declaration serves for host side compilation. + void + __assert_fail( + const char* assertion, + const char* file, + unsigned int line, + const char* function) noexcept __attribute__((__noreturn__)); + +#endif // __SYCL_DEVICE_ONLY__ +} +#endif // NDEBUG +// ROCm disables kernel assert by default for performance considerations. +// Though ROCm supports __assert_fail, it uses kernel printf which has +// a non-negligible performance impact even if the assert condition is +// never triggered. We choose to use abort() instead which will still +// terminate the application but without a more useful error message. +#if !defined(C10_USE_ROCM_KERNEL_ASSERT) && defined(USE_ROCM) +#define CUDA_KERNEL_ASSERT(cond) \ + if C10_UNLIKELY (!(cond)) { \ + abort(); \ + } +#define CUDA_KERNEL_ASSERT_MSG(cond, msg) \ + if C10_UNLIKELY (!(cond)) { \ + abort(); \ + } +#define CUDA_KERNEL_ASSERT_PRINTF(cond, msg, ...) \ + if C10_UNLIKELY (!(cond)) { \ + abort(); \ + } +#define SYCL_KERNEL_ASSERT(cond) \ + if C10_UNLIKELY (!(cond)) { \ + abort(); \ + } +#else +#define CUDA_KERNEL_ASSERT(cond) \ + if (C10_UNLIKELY(!(cond))) { \ + __assert_fail( \ + #cond, __FILE__, static_cast(__LINE__), __func__); \ + } +#define CUDA_KERNEL_ASSERT_MSG(cond, msg) \ + if (C10_UNLIKELY(!(cond))) { \ + __assert_fail( \ + msg, __FILE__, static_cast(__LINE__), __func__); \ + } +#define CUDA_KERNEL_ASSERT_PRINTF(cond, msg, ...) \ + if (C10_UNLIKELY(!(cond))) { \ + printf( \ + "[CUDA_KERNEL_ASSERT] " __FILE__ ":" C10_STRINGIZE( \ + __LINE__) ": %s: block: [%d,%d,%d], thread: [%d,%d,%d]: " \ + "Assertion failed: `" #cond "`: " msg "\n", \ + __func__, \ + blockIdx.x, \ + blockIdx.y, \ + blockIdx.z, \ + threadIdx.x, \ + threadIdx.y, \ + threadIdx.z, \ + ##__VA_ARGS__); \ + __assert_fail( \ + #cond, __FILE__, static_cast(__LINE__), __func__); \ + } +#define SYCL_KERNEL_ASSERT(cond) \ + if (C10_UNLIKELY(!(cond))) { \ + __assert_fail( \ + #cond, __FILE__, static_cast(__LINE__), __func__); \ + } +#endif // C10_USE_ROCM_KERNEL_ASSERT && USE_ROCM +#endif // __APPLE__ + +// Compile-time switch to control how assertions are logged inside CUDA kernels. +// If C10_CUDA_VERBOSE_ASSERT is defined, CUDA_KERNEL_ASSERT_VERBOSE will +// take addition information passed to the macro and forward them to +// CUDA_KERNEL_ASSERT_PRINTF If C10_CUDA_VERBOSE_ASSERT is not defined, +// CUDA_KERNEL_ASSERT_VERBOSE will behave the same as CUDA_KERNEL_ASSERT. +#ifdef C10_ENABLE_VERBOSE_ASSERT +#define CUDA_KERNEL_ASSERT_VERBOSE(cond, ...) \ + CUDA_KERNEL_ASSERT_PRINTF(cond, __VA_ARGS__) +#else +#define CUDA_KERNEL_ASSERT_VERBOSE(cond, ...) CUDA_KERNEL_ASSERT(cond) +#endif + +#ifdef __APPLE__ +#include +#endif + +#if defined(__ANDROID__) +#define C10_ANDROID 1 +#define C10_MOBILE 1 +#elif ( \ + defined(__APPLE__) && \ + (TARGET_IPHONE_SIMULATOR || TARGET_OS_SIMULATOR || TARGET_OS_IPHONE)) +#define C10_IOS 1 +#define C10_MOBILE 1 +#endif // ANDROID / IOS + +#if defined(C10_MOBILE) && C10_MOBILE +#define C10_ALWAYS_INLINE_UNLESS_MOBILE inline +#else +#define C10_ALWAYS_INLINE_UNLESS_MOBILE C10_ALWAYS_INLINE +#endif + +#if !defined(FBCODE_CAFFE2) && !defined(C10_NODEPRECATED) +#define CONSTEXPR_EXCEPT_WIN_CUDA constexpr +#define C10_HOST_CONSTEXPR_EXCEPT_WIN_CUDA constexpr + +#define STATIC_CONSTEXPR_STR_INL_EXCEPT_WIN_CUDA(field, val) \ + static constexpr const char field[] = val; +#define STATIC_CONST_STR_OUT_OF_LINE_FOR_WIN_CUDA(cls, field, val) +#endif // !defined(FBCODE_CAFFE2) && !defined(C10_NODEPRECATED) + +#ifndef HAS_DEMANGLE +#if defined(__ANDROID__) || defined(_WIN32) || defined(__EMSCRIPTEN__) +#define HAS_DEMANGLE 0 +#elif defined(__APPLE__) && \ + (TARGET_IPHONE_SIMULATOR || TARGET_OS_SIMULATOR || TARGET_OS_IPHONE) +#define HAS_DEMANGLE 0 +#else +#define HAS_DEMANGLE 1 +#endif +#endif // HAS_DEMANGLE + +#define _C10_PRAGMA__(string) _Pragma(#string) +#define _C10_PRAGMA_(string) _C10_PRAGMA__(string) + +#ifdef __clang__ +#define C10_CLANG_DIAGNOSTIC_PUSH() _Pragma("clang diagnostic push") +#define C10_CLANG_DIAGNOSTIC_POP() _Pragma("clang diagnostic pop") +#define C10_CLANG_DIAGNOSTIC_IGNORE(flag) \ + _C10_PRAGMA_(clang diagnostic ignored flag) +#define C10_CLANG_HAS_WARNING(flag) __has_warning(flag) +#else +#define C10_CLANG_DIAGNOSTIC_PUSH() +#define C10_CLANG_DIAGNOSTIC_POP() +#define C10_CLANG_DIAGNOSTIC_IGNORE(flag) +#define C10_CLANG_HAS_WARNING(flag) 0 +#endif + +#ifdef __clang__ + +#define C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED(warning) \ + _C10_PRAGMA_(clang diagnostic push) \ + _C10_PRAGMA_(clang diagnostic ignored "-Wunknown-warning-option") \ + _C10_PRAGMA_(clang diagnostic ignored warning) + +#define C10_DIAGNOSTIC_POP() _C10_PRAGMA_(clang diagnostic pop) + +#elif __GNUC__ + +#define C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED(warning) \ + _C10_PRAGMA_(GCC diagnostic push) \ + _C10_PRAGMA_(GCC diagnostic ignored "-Wpragmas") \ + _C10_PRAGMA_(GCC diagnostic ignored warning) + +#define C10_DIAGNOSTIC_POP() _C10_PRAGMA_(GCC diagnostic pop) + +#else + +#define C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED(warning) +#define C10_DIAGNOSTIC_POP() + +#endif + +// This macro is used to find older C++ compilers +// that don't support move optimization for return values. + +#if (defined(__GNUC__) && __GNUC__ < 13) || \ + (defined(__clang_major__) && __clang_major__ < 13) +#define C10_RETURN_MOVE_IF_OLD_COMPILER 1 +#else +#define C10_RETURN_MOVE_IF_OLD_COMPILER 0 +#endif + +// The HIDDEN_NAMESPACE_BEGIN and HIDDEN_NAMESPACE_END below +// are needed for maintaining robustness in our header APIs in +// torch/headeronly and torch/csrc/stable under the namespaces +// torch::headeronly and torch::stable respectively. We enforce +// hidden visibility for these APIs because we want to enable +// loading custom extensions compiled against different libtorch +// versions where these APIs may have changed. + +// Helper macros to handle 1-3 hidden namespace levels when not windows +#define _HIDDEN_NS_GET_MACRO(_1, _2, _3, NAME, ...) NAME +#define _HIDDEN_NS_1(n1) namespace n1 __attribute__((visibility("hidden"))) { +#define _HIDDEN_NS_2(n1, n2) \ + namespace n1 { \ + namespace n2 __attribute__((visibility("hidden"))) { +#define _HIDDEN_NS_3(n1, n2, n3) \ + namespace n1::n2 { \ + namespace n3 __attribute__((visibility("hidden"))) { + +// Helper macros to close namespaces when not windows +#define _HIDDEN_NS_END_1(n1) } +#define _HIDDEN_NS_END_N(n1, ...) \ + } \ + } + +// Helper macros to join strs with :: (for win, where symbols are hidden by +// default) +#define _EXPAND(...) __VA_ARGS__ +#define _JOIN_GET_MACRO(_1, _2, _3, NAME, ...) NAME +#define _JOIN_NS1(a) a +#define _JOIN_NS2(a, b) a::b +#define _JOIN_NS3(a, b, c) a::b::c + +#if !defined(HIDDEN_NAMESPACE_BEGIN) +#if defined(__GNUG__) && !defined(_WIN32) +#define HIDDEN_NAMESPACE_BEGIN(...) \ + _HIDDEN_NS_GET_MACRO( \ + __VA_ARGS__, _HIDDEN_NS_3, _HIDDEN_NS_2, _HIDDEN_NS_1)(__VA_ARGS__) +#else +#define HIDDEN_NAMESPACE_BEGIN(...) \ + namespace _EXPAND(_JOIN_GET_MACRO( \ + __VA_ARGS__, _JOIN_NS3, _JOIN_NS2, _JOIN_NS1)(__VA_ARGS__)) { +#endif +#endif + +#if !defined(HIDDEN_NAMESPACE_END) +#if defined(__GNUG__) && !defined(_WIN32) +#define HIDDEN_NAMESPACE_END(...) \ + _HIDDEN_NS_GET_MACRO( \ + __VA_ARGS__, _HIDDEN_NS_END_N, _HIDDEN_NS_END_N, _HIDDEN_NS_END_1)( \ + __VA_ARGS__) +#else +#define HIDDEN_NAMESPACE_END(...) } +#endif +#endif + +#endif // C10_MACROS_MACROS_H_ diff --git a/backends/pytorch/include/torch/headeronly/macros/cmake_macros.h b/backends/pytorch/include/torch/headeronly/macros/cmake_macros.h new file mode 100644 index 000000000..8b8894ca9 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/macros/cmake_macros.h @@ -0,0 +1,14 @@ +#ifndef C10_MACROS_CMAKE_MACROS_H_ +#define C10_MACROS_CMAKE_MACROS_H_ + +// Automatically generated header file for the C10 library. +// Do not include this file directly. Instead, include torch/headeronly/macros/Macros.h. + +#define C10_BUILD_SHARED_LIBS +/* #undef C10_USE_GLOG */ +/* #undef C10_USE_GFLAGS */ +/* #undef C10_USE_NUMA */ +/* #undef C10_USE_MSVC_STATIC_RUNTIME */ +/* #undef C10_USE_ROCM_KERNEL_ASSERT */ + +#endif // C10_MACROS_CMAKE_MACROS_H_ diff --git a/backends/pytorch/include/torch/headeronly/util/BFloat16.h b/backends/pytorch/include/torch/headeronly/util/BFloat16.h new file mode 100644 index 000000000..64479ba36 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/BFloat16.h @@ -0,0 +1,480 @@ +#pragma once + +// Defines the bloat16 type (brain floating-point). This representation uses +// 1 bit for the sign, 8 bits for the exponent and 7 bits for the mantissa. + +#include +#include + +#include +#include +#include +#include +#include + +#if defined(__CUDACC__) && !defined(USE_ROCM) +#include +#endif + +#if defined(CL_SYCL_LANGUAGE_VERSION) +#include // for SYCL 1.2.1 +#elif defined(SYCL_LANGUAGE_VERSION) +#include // for SYCL 2020 +#endif + +namespace c10 { + +struct alignas(2) BFloat16 { + uint16_t x; + + // HIP wants __host__ __device__ tag, CUDA does not +#if defined(USE_ROCM) && defined(__HIPCC__) + C10_HOST_DEVICE BFloat16() = default; +#else + BFloat16() = default; +#endif + + struct from_bits_t {}; + static constexpr C10_HOST_DEVICE from_bits_t from_bits() { + return from_bits_t(); + } + + constexpr C10_HOST_DEVICE BFloat16( + unsigned short bits, + from_bits_t /*unused*/) + : x(bits) {} + /* implicit */ inline C10_HOST_DEVICE BFloat16(float value); + inline C10_HOST_DEVICE operator float() const; + +#if defined(__CUDACC__) && !defined(USE_ROCM) + inline C10_HOST_DEVICE BFloat16(const __nv_bfloat16& value); + explicit inline C10_HOST_DEVICE operator __nv_bfloat16() const; +#endif + +#if defined(SYCL_EXT_ONEAPI_BFLOAT16_MATH_FUNCTIONS) + inline C10_HOST_DEVICE BFloat16(const sycl::ext::oneapi::bfloat16& value); + explicit inline C10_HOST_DEVICE operator sycl::ext::oneapi::bfloat16() const; +#endif +}; + +inline std::ostream& operator<<(std::ostream& out, const BFloat16& value) { + out << (float)value; + return out; +} + +namespace detail { +inline C10_HOST_DEVICE float f32_from_bits(uint16_t src) { + float res = 0; + uint32_t tmp = src; + tmp <<= 16; + +#if defined(USE_ROCM) && defined(__HIPCC__) + float* tempRes; + + // We should be using memcpy in order to respect the strict aliasing rule + // but it fails in the HIP environment. + tempRes = reinterpret_cast(&tmp); + res = *tempRes; +#else + std::memcpy(&res, &tmp, sizeof(tmp)); +#endif + + return res; +} + +inline C10_HOST_DEVICE uint16_t bits_from_f32(float src) { + uint32_t res = 0; + +#if defined(USE_ROCM) && defined(__HIPCC__) + // We should be using memcpy in order to respect the strict aliasing rule + // but it fails in the HIP environment. + uint32_t* tempRes = reinterpret_cast(&src); + res = *tempRes; +#else + std::memcpy(&res, &src, sizeof(res)); +#endif + + return res >> 16; +} + +inline C10_HOST_DEVICE uint16_t round_to_nearest_even(float src) { +#if defined(USE_ROCM) && defined(__HIPCC__) + if (src != src) { +#elif defined(_MSC_VER) + if (isnan(src)) { +#else + if (std::isnan(src)) { +#endif + return UINT16_C(0x7FC0); + } else { + const uint32_t U32 = c10::bit_cast(src); + uint32_t rounding_bias = ((U32 >> 16) & 1) + UINT32_C(0x7FFF); + return static_cast((U32 + rounding_bias) >> 16); + } +} + +} // namespace detail + +//-------- the following is copied from c10/util/BFloat16-inl.h ---------// +C10_CLANG_DIAGNOSTIC_PUSH() +#if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion") +#endif + +/// Constructors +inline C10_HOST_DEVICE BFloat16::BFloat16(float value) + : +#if defined(__CUDACC__) && !defined(USE_ROCM) && defined(__CUDA_ARCH__) && \ + __CUDA_ARCH__ >= 800 + x(__bfloat16_as_ushort(__float2bfloat16(value))) +#elif defined(__SYCL_DEVICE_ONLY__) && \ + defined(SYCL_EXT_ONEAPI_BFLOAT16_MATH_FUNCTIONS) + x(c10::bit_cast(sycl::ext::oneapi::bfloat16(value))) +#else + // RNE by default + x(detail::round_to_nearest_even(value)) +#endif +{ +} + +/// Implicit conversions +inline C10_HOST_DEVICE BFloat16::operator float() const { +#if defined(__CUDACC__) && !defined(USE_ROCM) + return __bfloat162float(*reinterpret_cast(&x)); +#elif defined(__SYCL_DEVICE_ONLY__) && \ + defined(SYCL_EXT_ONEAPI_BFLOAT16_MATH_FUNCTIONS) + return float(*reinterpret_cast(&x)); +#else + return detail::f32_from_bits(x); +#endif +} + +#if defined(__CUDACC__) && !defined(USE_ROCM) +inline C10_HOST_DEVICE BFloat16::BFloat16(const __nv_bfloat16& value) { + x = *reinterpret_cast(&value); +} +inline C10_HOST_DEVICE BFloat16::operator __nv_bfloat16() const { + return *reinterpret_cast(&x); +} +#endif + +#if defined(SYCL_EXT_ONEAPI_BFLOAT16_MATH_FUNCTIONS) +inline C10_HOST_DEVICE BFloat16::BFloat16( + const sycl::ext::oneapi::bfloat16& value) { + x = *reinterpret_cast(&value); +} +inline C10_HOST_DEVICE BFloat16::operator sycl::ext::oneapi::bfloat16() const { + return *reinterpret_cast(&x); +} +#endif + +// CUDA intrinsics + +#if defined(__CUDACC__) || defined(__HIPCC__) +inline C10_DEVICE BFloat16 __ldg(const BFloat16* ptr) { +#if !defined(USE_ROCM) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800 + return __ldg(reinterpret_cast(ptr)); +#else + return *ptr; +#endif +} +#endif + +/// Arithmetic + +inline C10_HOST_DEVICE BFloat16 +operator+(const BFloat16& a, const BFloat16& b) { + return static_cast(a) + static_cast(b); +} + +inline C10_HOST_DEVICE BFloat16 +operator-(const BFloat16& a, const BFloat16& b) { + return static_cast(a) - static_cast(b); +} + +inline C10_HOST_DEVICE BFloat16 +operator*(const BFloat16& a, const BFloat16& b) { + return static_cast(a) * static_cast(b); +} + +inline C10_HOST_DEVICE BFloat16 operator/(const BFloat16& a, const BFloat16& b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / static_cast(b); +} + +inline C10_HOST_DEVICE BFloat16 operator-(const BFloat16& a) { + return -static_cast(a); +} + +inline C10_HOST_DEVICE BFloat16& operator+=(BFloat16& a, const BFloat16& b) { + a = a + b; + return a; +} + +inline C10_HOST_DEVICE BFloat16& operator-=(BFloat16& a, const BFloat16& b) { + a = a - b; + return a; +} + +inline C10_HOST_DEVICE BFloat16& operator*=(BFloat16& a, const BFloat16& b) { + a = a * b; + return a; +} + +inline C10_HOST_DEVICE BFloat16& operator/=(BFloat16& a, const BFloat16& b) { + a = a / b; + return a; +} + +inline C10_HOST_DEVICE BFloat16& operator|(BFloat16& a, const BFloat16& b) { + a.x = a.x | b.x; + return a; +} + +inline C10_HOST_DEVICE BFloat16& operator^(BFloat16& a, const BFloat16& b) { + a.x = a.x ^ b.x; + return a; +} + +inline C10_HOST_DEVICE BFloat16& operator&(BFloat16& a, const BFloat16& b) { + a.x = a.x & b.x; + return a; +} + +/// Arithmetic with floats + +inline C10_HOST_DEVICE float operator+(BFloat16 a, float b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE float operator-(BFloat16 a, float b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE float operator*(BFloat16 a, float b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE float operator/(BFloat16 a, float b) { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE float operator+(float a, BFloat16 b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE float operator-(float a, BFloat16 b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE float operator*(float a, BFloat16 b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE float operator/(float a, BFloat16 b) { + return a / static_cast(b); +} + +inline C10_HOST_DEVICE float& operator+=(float& a, const BFloat16& b) { + return a += static_cast(b); +} +inline C10_HOST_DEVICE float& operator-=(float& a, const BFloat16& b) { + return a -= static_cast(b); +} +inline C10_HOST_DEVICE float& operator*=(float& a, const BFloat16& b) { + return a *= static_cast(b); +} +inline C10_HOST_DEVICE float& operator/=(float& a, const BFloat16& b) { + return a /= static_cast(b); +} + +/// Arithmetic with doubles + +inline C10_HOST_DEVICE double operator+(BFloat16 a, double b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE double operator-(BFloat16 a, double b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE double operator*(BFloat16 a, double b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE double operator/(BFloat16 a, double b) { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE double operator+(double a, BFloat16 b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE double operator-(double a, BFloat16 b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE double operator*(double a, BFloat16 b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE double operator/(double a, BFloat16 b) { + return a / static_cast(b); +} + +/// Arithmetic with ints + +inline C10_HOST_DEVICE BFloat16 operator+(BFloat16 a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE BFloat16 operator-(BFloat16 a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE BFloat16 operator*(BFloat16 a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE BFloat16 operator/(BFloat16 a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE BFloat16 operator+(int a, BFloat16 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE BFloat16 operator-(int a, BFloat16 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE BFloat16 operator*(int a, BFloat16 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE BFloat16 operator/(int a, BFloat16 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) / b; +} + +//// Arithmetic with int64_t + +inline C10_HOST_DEVICE BFloat16 operator+(BFloat16 a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE BFloat16 operator-(BFloat16 a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE BFloat16 operator*(BFloat16 a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE BFloat16 operator/(BFloat16 a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE BFloat16 operator+(int64_t a, BFloat16 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE BFloat16 operator-(int64_t a, BFloat16 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE BFloat16 operator*(int64_t a, BFloat16 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE BFloat16 operator/(int64_t a, BFloat16 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) / b; +} + +// Overloading < and > operators, because std::max and std::min use them. + +inline C10_HOST_DEVICE bool operator>(BFloat16& lhs, BFloat16& rhs) { + return float(lhs) > float(rhs); +} + +inline C10_HOST_DEVICE bool operator<(BFloat16& lhs, BFloat16& rhs) { + return float(lhs) < float(rhs); +} + +C10_CLANG_DIAGNOSTIC_POP() +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) + +namespace detail { +using c10::detail::bits_from_f32; +using c10::detail::f32_from_bits; +using c10::detail::round_to_nearest_even; +} // namespace detail + +using c10::BFloat16; +using c10::operator+; +using c10::operator-; +using c10::operator*; +using c10::operator/; +using c10::operator+=; +using c10::operator-=; +using c10::operator*=; +using c10::operator/=; +using c10::operator<; +using c10::operator>; +using c10::operator<<; +HIDDEN_NAMESPACE_END(torch, headeronly) + +namespace std { + +template <> +class numeric_limits { + public: + static constexpr bool is_signed = true; + static constexpr bool is_specialized = true; + static constexpr bool is_integer = false; + static constexpr bool is_exact = false; + static constexpr bool has_infinity = true; + static constexpr bool has_quiet_NaN = true; + static constexpr bool has_signaling_NaN = true; + static constexpr auto has_denorm = numeric_limits::has_denorm; + static constexpr auto has_denorm_loss = + numeric_limits::has_denorm_loss; + static constexpr auto round_style = numeric_limits::round_style; + static constexpr bool is_iec559 = false; + static constexpr bool is_bounded = true; + static constexpr bool is_modulo = false; + static constexpr int digits = 8; + static constexpr int digits10 = 2; + static constexpr int max_digits10 = 4; + static constexpr int radix = 2; + static constexpr int min_exponent = -125; + static constexpr int min_exponent10 = -37; + static constexpr int max_exponent = 128; + static constexpr int max_exponent10 = 38; + static constexpr auto traps = numeric_limits::traps; + static constexpr auto tinyness_before = + numeric_limits::tinyness_before; + + static constexpr c10::BFloat16 min() { + return c10::BFloat16(0x0080, c10::BFloat16::from_bits()); + } + static constexpr c10::BFloat16 lowest() { + return c10::BFloat16(0xFF7F, c10::BFloat16::from_bits()); + } + static constexpr c10::BFloat16 max() { + return c10::BFloat16(0x7F7F, c10::BFloat16::from_bits()); + } + static constexpr c10::BFloat16 epsilon() { + return c10::BFloat16(0x3C00, c10::BFloat16::from_bits()); + } + static constexpr c10::BFloat16 round_error() { + return c10::BFloat16(0x3F00, c10::BFloat16::from_bits()); + } + static constexpr c10::BFloat16 infinity() { + return c10::BFloat16(0x7F80, c10::BFloat16::from_bits()); + } + static constexpr c10::BFloat16 quiet_NaN() { + return c10::BFloat16(0x7FC0, c10::BFloat16::from_bits()); + } + static constexpr c10::BFloat16 signaling_NaN() { + return c10::BFloat16(0x7F80, c10::BFloat16::from_bits()); + } + static constexpr c10::BFloat16 denorm_min() { + return c10::BFloat16(0x0001, c10::BFloat16::from_bits()); + } +}; + +} // namespace std diff --git a/backends/pytorch/include/torch/headeronly/util/Deprecated.h b/backends/pytorch/include/torch/headeronly/util/Deprecated.h new file mode 100644 index 000000000..88440a024 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Deprecated.h @@ -0,0 +1,102 @@ +#pragma once + +/** + * This file provides portable macros for marking declarations + * as deprecated. You should generally use C10_DEPRECATED, + * except when marking 'using' declarations as deprecated, + * in which case you should use C10_DEFINE_DEPRECATED_USING + * (due to portability concerns). + */ + +// Sample usage: +// +// C10_DEPRECATED void bad_func(); +// struct C10_DEPRECATED BadStruct { +// ... +// }; + +// NB: __cplusplus doesn't work for MSVC, so for now MSVC always uses +// the "__declspec(deprecated)" implementation and not the C++14 +// "[[deprecated]]" attribute. We tried enabling "[[deprecated]]" for C++14 on +// MSVC, but ran into issues with some older MSVC versions. +#if (defined(__cplusplus) && __cplusplus >= 201402L) +#define C10_DEPRECATED [[deprecated]] +#define C10_DEPRECATED_MESSAGE(message) [[deprecated(message)]] +#elif defined(__GNUC__) +#define C10_DEPRECATED __attribute__((deprecated)) +// TODO Is there some way to implement this? +#define C10_DEPRECATED_MESSAGE(message) __attribute__((deprecated)) + +#elif defined(_MSC_VER) +#define C10_DEPRECATED __declspec(deprecated) +#define C10_DEPRECATED_MESSAGE(message) __declspec(deprecated(message)) +#else +#warning "You need to implement C10_DEPRECATED for this compiler" +#define C10_DEPRECATED +#endif + +// Sample usage: +// +// C10_DEFINE_DEPRECATED_USING(BadType, int) +// +// which is the portable version of +// +// using BadType [[deprecated]] = int; + +// technically [[deprecated]] syntax is from c++14 standard, but it works in +// many compilers. +#if defined(__has_cpp_attribute) +#if __has_cpp_attribute(deprecated) && !defined(__CUDACC__) +#define C10_DEFINE_DEPRECATED_USING(TypeName, TypeThingy) \ + using TypeName [[deprecated]] = TypeThingy; +#endif +#endif + +#if defined(_MSC_VER) +#if defined(__CUDACC__) +// neither [[deprecated]] nor __declspec(deprecated) work on nvcc on Windows; +// you get the error: +// +// error: attribute does not apply to any entity +// +// So we just turn the macro off in this case. +#if defined(C10_DEFINE_DEPRECATED_USING) +#undef C10_DEFINE_DEPRECATED_USING +#endif +#define C10_DEFINE_DEPRECATED_USING(TypeName, TypeThingy) \ + using TypeName = TypeThingy; +#else +// [[deprecated]] does work in windows without nvcc, though msc doesn't support +// `__has_cpp_attribute` when c++14 is supported, otherwise +// __declspec(deprecated) is used as the alternative. +#ifndef C10_DEFINE_DEPRECATED_USING +#if defined(_MSVC_LANG) && _MSVC_LANG >= 201402L +#define C10_DEFINE_DEPRECATED_USING(TypeName, TypeThingy) \ + using TypeName [[deprecated]] = TypeThingy; +#else +#define C10_DEFINE_DEPRECATED_USING(TypeName, TypeThingy) \ + using TypeName = __declspec(deprecated) TypeThingy; +#endif +#endif +#endif +#endif + +#if !defined(C10_DEFINE_DEPRECATED_USING) && defined(__GNUC__) +// nvcc has a bug where it doesn't understand __attribute__((deprecated)) +// declarations even when the host compiler supports it. We'll only use this gcc +// attribute when not cuda, and when using a GCC compiler that doesn't support +// the c++14 syntax we checked for above (available in __GNUC__ >= 5) +#if !defined(__CUDACC__) +#define C10_DEFINE_DEPRECATED_USING(TypeName, TypeThingy) \ + using TypeName __attribute__((deprecated)) = TypeThingy; +#else +// using cuda + gcc < 5, neither deprecated syntax is available so turning off. +#define C10_DEFINE_DEPRECATED_USING(TypeName, TypeThingy) \ + using TypeName = TypeThingy; +#endif +#endif + +#if !defined(C10_DEFINE_DEPRECATED_USING) +#warning "You need to implement C10_DEFINE_DEPRECATED_USING for this compiler" +#define C10_DEFINE_DEPRECATED_USING +#endif diff --git a/backends/pytorch/include/torch/headeronly/util/Exception.h b/backends/pytorch/include/torch/headeronly/util/Exception.h new file mode 100644 index 000000000..0e067244a --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Exception.h @@ -0,0 +1,83 @@ +#pragma once + +#include +#include + +#include +#include + +namespace c10 { +// On nvcc, C10_UNLIKELY thwarts missing return statement analysis. In cases +// where the unlikely expression may be a constant, use this macro to ensure +// return statement analysis keeps working (at the cost of not getting the +// likely/unlikely annotation on nvcc). +// https://github.com/pytorch/pytorch/issues/21418 +// +// Currently, this is only used in the error reporting macros below. If you +// want to use it more generally, move me to Macros.h +// +// TODO: Brian Vaughan observed that we might be able to get this to work on +// nvcc by writing some sort of C++ overload that distinguishes constexpr inputs +// from non-constexpr. Since there isn't any evidence that losing C10_UNLIKELY +// in nvcc is causing us perf problems, this is not yet implemented, but this +// might be an interesting piece of C++ code for an intrepid bootcamper to +// write. +#if defined(__CUDACC__) +#define C10_UNLIKELY_OR_CONST(e) e +#else +#define C10_UNLIKELY_OR_CONST(e) C10_UNLIKELY(e) +#endif + +} // namespace c10 + +// STD_TORCH_CHECK throws std::runtime_error instead of c10::Error which is +// useful when certain headers are used in a libtorch-independent way, +// e.g. when Vectorized is used in AOTInductor generated code, or +// for custom ops to have an ABI stable dependency on libtorch. +#ifdef STRIP_ERROR_MESSAGES +#define STD_TORCH_CHECK_MSG(cond, type, ...) \ + (#cond #type " CHECK FAILED at " C10_STRINGIZE(__FILE__)) +#else // so STRIP_ERROR_MESSAGES is not defined +HIDDEN_NAMESPACE_BEGIN(torch, headeronly, detail) +template +std::string stdTorchCheckMsgImpl(const char* /*msg*/, const Args&... args) { + // This is similar to the one in c10/util/Exception.h, but does + // not depend on the more complex c10::str() function. ostringstream + // supports fewer data types than c10::str(), but should be sufficient + // in the headeronly world. + std::ostringstream oss; + ((oss << args), ...); + return oss.str(); +} + +inline const char* stdTorchCheckMsgImpl(const char* msg) { + return msg; +} +// If there is just 1 user-provided C-string argument, use it. +inline const char* stdTorchCheckMsgImpl(const char* /*msg*/, const char* args) { + return args; +} +HIDDEN_NAMESPACE_END(torch, headeronly, detail) + +#define STD_TORCH_CHECK_MSG(cond, type, ...) \ + (torch::headeronly::detail::stdTorchCheckMsgImpl( \ + "Expected " #cond \ + " to be true, but got false. " \ + "(Could this error message be improved? If so, " \ + "please report an enhancement request to PyTorch.)", \ + ##__VA_ARGS__)) +#endif // STRIP_ERROR_MESSAGES + +#define STD_TORCH_CHECK(cond, ...) \ + if (C10_UNLIKELY_OR_CONST(!(cond))) { \ + throw std::runtime_error(STD_TORCH_CHECK_MSG( \ + cond, \ + "", \ + __func__, \ + ", ", \ + __FILE__, \ + ":", \ + __LINE__, \ + ", ", \ + ##__VA_ARGS__)); \ + } diff --git a/backends/pytorch/include/torch/headeronly/util/Float4_e2m1fn_x2.h b/backends/pytorch/include/torch/headeronly/util/Float4_e2m1fn_x2.h new file mode 100644 index 000000000..00075914c --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Float4_e2m1fn_x2.h @@ -0,0 +1,47 @@ +#pragma once +#include + +#include + +/// Defines the Float4_e2m1fn_x2 type (4-bit floating-point, two elements packed +/// into one byte). This is the FP4 dtype from the OCP MX format spec +/// (https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf, +/// Section 5.3.3) +/// +/// Given two high precision values val0 and val1, here is the +/// binary configuration of their packed representation, from MSB to LSB: +/// +/// original value | val1 : val0 +/// ======================================== +/// bit index (MSB==7, LSB==0) | 7654 : 3210 +/// sign/exponent/mantissa | seem : seem +/// + +namespace c10 { + +struct alignas(1) Float4_e2m1fn_x2 { + uint8_t val_; + Float4_e2m1fn_x2() = default; + C10_HOST_DEVICE explicit Float4_e2m1fn_x2(uint8_t val) : val_(val) {} +}; + +/// Comparison operators +inline C10_HOST_DEVICE bool operator==( + const Float4_e2m1fn_x2& a, + const Float4_e2m1fn_x2& b) { + return a.val_ == b.val_; +} + +inline C10_HOST_DEVICE bool operator!=( + const Float4_e2m1fn_x2& a, + const Float4_e2m1fn_x2& b) { + return a.val_ != b.val_; +} + +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) +using c10::Float4_e2m1fn_x2; +using c10::operator==; +using c10::operator!=; +HIDDEN_NAMESPACE_END(torch, headeronly) diff --git a/backends/pytorch/include/torch/headeronly/util/Float8_e4m3fn.h b/backends/pytorch/include/torch/headeronly/util/Float8_e4m3fn.h new file mode 100644 index 000000000..b35fb1a7a --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Float8_e4m3fn.h @@ -0,0 +1,531 @@ +#pragma once + +/// Defines the Float8_e4m3fn type (8-bit floating-point) including conversions +/// to standard C types and basic arithmetic operations. Note that arithmetic +/// operations are implemented by converting to floating point and +/// performing the operation in float32. +/// Binary configuration: +/// s eeee mmm +/// 1 sign bit +/// 4 exponent bits +/// 3 mantissa bits +/// bias = 7 +/// +/// Implementation based on the paper https://arxiv.org/pdf/2209.05433.pdf +/// and inspired by Half implementation from pytorch/c10/util/Half.h + +#include +#include + +#if defined(__cplusplus) +#include +#include +#elif !defined(__OPENCL_VERSION__) +#include +#include +#endif + +#ifdef _MSC_VER +#include +#endif + +#include +#include + +namespace c10 { + +struct alignas(1) Float8_e4m3fn { + uint8_t x; + + struct from_bits_t {}; + C10_HOST_DEVICE static constexpr from_bits_t from_bits() { + return from_bits_t(); + } + + Float8_e4m3fn() = default; + + constexpr C10_HOST_DEVICE Float8_e4m3fn(uint8_t bits, from_bits_t /*unused*/) + : x(bits) {} + inline C10_HOST_DEVICE Float8_e4m3fn(float value); + inline C10_HOST_DEVICE operator float() const; + inline C10_HOST_DEVICE bool isnan() const; +}; + +inline std::ostream& operator<<(std::ostream& out, const Float8_e4m3fn& value) { + out << (float)value; + return out; +} + +namespace detail { + +/* + * Convert a 8-bit floating-point number in fp8 E4M3FN format, in bit + * representation, to a 32-bit floating-point number in IEEE single-precision + * format, in bit representation. + * + * @note The implementation doesn't use any floating-point operations. + */ +inline C10_HOST_DEVICE float fp8e4m3fn_to_fp32_value(uint8_t input) { + /* + * Extend the fp8 E4M3FN number to 32 bits and shift to the + * upper part of the 32-bit word: + * +---+----+---+-----------------------------+ + * | S |EEEE|MMM|0000 0000 0000 0000 0000 0000| + * +---+----+---+-----------------------------+ + * Bits 31 27-30 24-26 0-23 + * + * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 + * - zero bits. + */ + const uint32_t w = (uint32_t)input << 24; + /* + * Extract the sign of the input number into the high bit of the 32-bit word: + * + * +---+----------------------------------+ + * | S |0000000 00000000 00000000 00000000| + * +---+----------------------------------+ + * Bits 31 0-31 + */ + const uint32_t sign = w & UINT32_C(0x80000000); + /* + * Extract mantissa and biased exponent of the input number into the bits 0-30 + * of the 32-bit word: + * + * +---+----+---+-----------------------------+ + * | S |EEEE|MMM|0000 0000 0000 0000 0000 0000| + * +---+----+---+-----------------------------+ + * Bits 31 27-30 24-26 0-23 + */ + const uint32_t nonsign = w & UINT32_C(0x7FFFFFFF); + /* + * Renorm shift is the number of bits to shift mantissa left to make the + * half-precision number normalized. If the initial number is normalized, some + * of its high 5 bits (sign == 0 and 4-bit exponent) equals one. In this case + * renorm_shift == 0. If the number is denormalize, renorm_shift > 0. Note + * that if we shift denormalized nonsign by renorm_shift, the unit bit of + * mantissa will shift into exponent, turning the biased exponent into 1, and + * making mantissa normalized (i.e. without leading 1). + */ +#if defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__) + uint32_t renorm_shift = __clz(nonsign); +#elif defined(__SYCL_DEVICE_ONLY__) + // Note: zero is not a supported input into `__builtin_clz` + uint32_t renorm_shift = + nonsign != 0 ? __builtin_clz(nonsign) : sizeof(uint32_t) * CHAR_BIT; +#elif defined(_MSC_VER) && !defined(__clang__) + unsigned long nonsign_bsr; + _BitScanReverse(&nonsign_bsr, (unsigned long)nonsign); + uint32_t renorm_shift = (uint32_t)nonsign_bsr ^ 31; +#else + // Note: zero is not a supported input into `__builtin_clz` + uint32_t renorm_shift = + nonsign != 0 ? __builtin_clz(nonsign) : sizeof(uint32_t) * CHAR_BIT; +#endif + renorm_shift = renorm_shift > 4 ? renorm_shift - 4 : 0; + /* + * Iff fp8e4m3fn number has all exponent and mantissa bits set to 1, + * the addition overflows it into bit 31, and the subsequent shift turns the + * high 9 bits into 1. Thus inf_nan_mask == 0x7F800000 if the fp8e4m3fn number + * is Nan, 0x00000000 otherwise + */ + const int32_t inf_nan_mask = + ((int32_t)(nonsign + 0x01000000) >> 8) & INT32_C(0x7F800000); + /* + * Iff nonsign is 0, it overflows into 0xFFFFFFFF, turning bit 31 + * into 1. Otherwise, bit 31 remains 0. The signed shift right by 31 + * broadcasts bit 31 into all bits of the zero_mask. Thus zero_mask == + * 0xFFFFFFFF if the half-precision number was zero (+0.0h or -0.0h) + * 0x00000000 otherwise + */ + const int32_t zero_mask = (int32_t)(nonsign - 1) >> 31; + /* + * 1. Shift nonsign left by renorm_shift to normalize it (if the input + * was denormal) + * 2. Shift nonsign right by 4 so the exponent (4 bits originally) + * becomes an 8-bit field and 3-bit mantissa shifts into the 3 high + * bits of the 23-bit mantissa of IEEE single-precision number. + * 3. Add 0x78 to the exponent (starting at bit 23) to compensate the + * different in exponent bias (0x7F for single-precision number less 0x07 + * for fp8e4m3fn number). + * 4. Subtract renorm_shift from the exponent (starting at bit 23) to + * account for renormalization. As renorm_shift is less than 0x78, this + * can be combined with step 3. + * 5. Binary OR with inf_nan_mask to turn the exponent into 0xFF if the + * input was NaN or infinity. + * 6. Binary ANDNOT with zero_mask to turn the mantissa and exponent + * into zero if the input was zero. + * 7. Combine with the sign of the input number. + */ + uint32_t result = sign | + ((((nonsign << renorm_shift >> 4) + ((0x78 - renorm_shift) << 23)) | + inf_nan_mask) & + ~zero_mask); + return fp32_from_bits(result); +} + +/* + * Convert a 32-bit floating-point number in IEEE single-precision format to a + * 8-bit floating-point number in fp8 E4M3FN format, in bit representation. + */ +inline C10_HOST_DEVICE uint8_t fp8e4m3fn_from_fp32_value(float f) { + /* + * Binary representation of 480.0f, which is the first value + * not representable in fp8e4m3fn range: + * 0 1111 111 - fp8e4m3fn + * 0 10000111 11100000000000000000000 - fp32 + */ + constexpr uint32_t fp8_max = UINT32_C(1087) << 20; + + /* + * A mask for converting fp32 numbers lower than fp8e4m3fn normal range + * into denorm representation + * magic number: ((127 - 7) + (23 - 3) + 1) + */ + constexpr uint32_t denorm_mask = UINT32_C(141) << 23; + + uint32_t f_bits = fp32_to_bits(f); + + uint8_t result = 0u; + + /* + * Extract the sign of the input number into the high bit of the 32-bit word: + * + * +---+----------------------------------+ + * | S |0000000 00000000 00000000 00000000| + * +---+----------------------------------+ + * Bits 31 0-31 + */ + const uint32_t sign = f_bits & UINT32_C(0x80000000); + + /* + * Set sign bit to 0 + */ + f_bits ^= sign; + + if (f_bits >= fp8_max) { + // NaN - all exponent and mantissa bits set to 1 + result = 0x7f; + } else { + if (f_bits < (UINT32_C(121) << 23)) { + // Input number is smaller than 2^(-6), which is the smallest + // fp8e4m3fn normal number + f_bits = + fp32_to_bits(fp32_from_bits(f_bits) + fp32_from_bits(denorm_mask)); + result = static_cast(f_bits - denorm_mask); + } else { + // resulting mantissa is odd + uint8_t mant_odd = (f_bits >> 20) & 1; + + // update exponent, rounding bias part 1 + f_bits += ((uint32_t)(7 - 127) << 23) + 0x7FFFF; + + // rounding bias part 2 + f_bits += mant_odd; + + // take the bits! + result = static_cast(f_bits >> 20); + } + } + + result |= static_cast(sign >> 24); + return result; +} + +} // namespace detail + +// -------- below is copied from c10/util/Float8_e4m3fn-inl.h --------// +C10_CLANG_DIAGNOSTIC_PUSH() +#if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion") +#endif + +/// Constructors + +inline C10_HOST_DEVICE Float8_e4m3fn::Float8_e4m3fn(float value) + : x(detail::fp8e4m3fn_from_fp32_value(value)) {} + +/// Implicit conversions + +inline C10_HOST_DEVICE Float8_e4m3fn::operator float() const { + return detail::fp8e4m3fn_to_fp32_value(x); +} + +/// Special values helper + +inline C10_HOST_DEVICE bool Float8_e4m3fn::isnan() const { + return (x & 0b01111111) == 0b01111111; +} + +/// Arithmetic + +inline C10_HOST_DEVICE Float8_e4m3fn +operator+(const Float8_e4m3fn& a, const Float8_e4m3fn& b) { + return static_cast(a) + static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fn +operator-(const Float8_e4m3fn& a, const Float8_e4m3fn& b) { + return static_cast(a) - static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fn +operator*(const Float8_e4m3fn& a, const Float8_e4m3fn& b) { + return static_cast(a) * static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fn operator/( + const Float8_e4m3fn& a, + const Float8_e4m3fn& b) __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fn operator-(const Float8_e4m3fn& a) { + return -static_cast(a); +} + +inline C10_HOST_DEVICE Float8_e4m3fn& operator+=( + Float8_e4m3fn& a, + const Float8_e4m3fn& b) { + a = a + b; + return a; +} + +inline C10_HOST_DEVICE Float8_e4m3fn& operator-=( + Float8_e4m3fn& a, + const Float8_e4m3fn& b) { + a = a - b; + return a; +} + +inline C10_HOST_DEVICE Float8_e4m3fn& operator*=( + Float8_e4m3fn& a, + const Float8_e4m3fn& b) { + a = a * b; + return a; +} + +inline C10_HOST_DEVICE Float8_e4m3fn& operator/=( + Float8_e4m3fn& a, + const Float8_e4m3fn& b) { + a = a / b; + return a; +} + +/// Arithmetic with floats + +inline C10_HOST_DEVICE float operator+(Float8_e4m3fn a, float b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE float operator-(Float8_e4m3fn a, float b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE float operator*(Float8_e4m3fn a, float b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE float operator/(Float8_e4m3fn a, float b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE float operator+(float a, Float8_e4m3fn b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE float operator-(float a, Float8_e4m3fn b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE float operator*(float a, Float8_e4m3fn b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE float operator/(float a, Float8_e4m3fn b) + __ubsan_ignore_float_divide_by_zero__ { + return a / static_cast(b); +} + +inline C10_HOST_DEVICE float& operator+=(float& a, const Float8_e4m3fn& b) { + return a += static_cast(b); +} +inline C10_HOST_DEVICE float& operator-=(float& a, const Float8_e4m3fn& b) { + return a -= static_cast(b); +} +inline C10_HOST_DEVICE float& operator*=(float& a, const Float8_e4m3fn& b) { + return a *= static_cast(b); +} +inline C10_HOST_DEVICE float& operator/=(float& a, const Float8_e4m3fn& b) { + return a /= static_cast(b); +} + +/// Arithmetic with doubles + +inline C10_HOST_DEVICE double operator+(Float8_e4m3fn a, double b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE double operator-(Float8_e4m3fn a, double b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE double operator*(Float8_e4m3fn a, double b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE double operator/(Float8_e4m3fn a, double b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE double operator+(double a, Float8_e4m3fn b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE double operator-(double a, Float8_e4m3fn b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE double operator*(double a, Float8_e4m3fn b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE double operator/(double a, Float8_e4m3fn b) + __ubsan_ignore_float_divide_by_zero__ { + return a / static_cast(b); +} + +/// Arithmetic with ints + +inline C10_HOST_DEVICE Float8_e4m3fn operator+(Float8_e4m3fn a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fn operator-(Float8_e4m3fn a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fn operator*(Float8_e4m3fn a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fn operator/(Float8_e4m3fn a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fn operator+(int a, Float8_e4m3fn b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE Float8_e4m3fn operator-(int a, Float8_e4m3fn b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE Float8_e4m3fn operator*(int a, Float8_e4m3fn b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE Float8_e4m3fn operator/(int a, Float8_e4m3fn b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) / b; +} + +//// Arithmetic with int64_t + +inline C10_HOST_DEVICE Float8_e4m3fn operator+(Float8_e4m3fn a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fn operator-(Float8_e4m3fn a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fn operator*(Float8_e4m3fn a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fn operator/(Float8_e4m3fn a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fn operator+(int64_t a, Float8_e4m3fn b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE Float8_e4m3fn operator-(int64_t a, Float8_e4m3fn b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE Float8_e4m3fn operator*(int64_t a, Float8_e4m3fn b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE Float8_e4m3fn operator/(int64_t a, Float8_e4m3fn b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) / b; +} + +/// NOTE: we do not define comparisons directly and instead rely on the implicit +/// conversion from c10::Float8_e4m3fn to float. + +C10_CLANG_DIAGNOSTIC_POP() + +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) +using c10::Float8_e4m3fn; +using c10::operator<<; +using c10::operator+; +using c10::operator-; +using c10::operator*; +using c10::operator/; +using c10::operator+=; +using c10::operator-=; +using c10::operator*=; +using c10::operator/=; +HIDDEN_NAMESPACE_END(torch, headeronly) + +namespace std { + +template <> +class numeric_limits { + public: + static constexpr bool is_specialized = true; + static constexpr bool is_signed = true; + static constexpr bool is_integer = false; + static constexpr bool is_exact = false; + static constexpr bool has_infinity = false; + static constexpr bool has_quiet_NaN = true; + static constexpr bool has_signaling_NaN = false; + static constexpr auto has_denorm = true; + static constexpr auto has_denorm_loss = true; + static constexpr auto round_style = numeric_limits::round_style; + static constexpr bool is_iec559 = false; + static constexpr bool is_bounded = true; + static constexpr bool is_modulo = false; + static constexpr int digits = 4; + static constexpr int digits10 = 0; + static constexpr int max_digits10 = 3; + static constexpr int radix = 2; + static constexpr int min_exponent = -5; + static constexpr int min_exponent10 = -1; + static constexpr int max_exponent = 8; + static constexpr int max_exponent10 = 2; + static constexpr auto traps = numeric_limits::traps; + static constexpr auto tinyness_before = false; + + static constexpr c10::Float8_e4m3fn min() { + return c10::Float8_e4m3fn(0x08, c10::Float8_e4m3fn::from_bits()); + } + static constexpr c10::Float8_e4m3fn lowest() { + return c10::Float8_e4m3fn(0xFE, c10::Float8_e4m3fn::from_bits()); + } + static constexpr c10::Float8_e4m3fn max() { + return c10::Float8_e4m3fn(0x7E, c10::Float8_e4m3fn::from_bits()); + } + static constexpr c10::Float8_e4m3fn epsilon() { + return c10::Float8_e4m3fn(0x20, c10::Float8_e4m3fn::from_bits()); + } + static constexpr c10::Float8_e4m3fn round_error() { + return c10::Float8_e4m3fn(0x30, c10::Float8_e4m3fn::from_bits()); + } + static constexpr c10::Float8_e4m3fn quiet_NaN() { + return c10::Float8_e4m3fn(0x7F, c10::Float8_e4m3fn::from_bits()); + } + static constexpr c10::Float8_e4m3fn denorm_min() { + return c10::Float8_e4m3fn(0x01, c10::Float8_e4m3fn::from_bits()); + } +}; + +} // namespace std diff --git a/backends/pytorch/include/torch/headeronly/util/Float8_e4m3fnuz.h b/backends/pytorch/include/torch/headeronly/util/Float8_e4m3fnuz.h new file mode 100644 index 000000000..e361a2f92 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Float8_e4m3fnuz.h @@ -0,0 +1,444 @@ +#pragma once + +/// Defines the Float8_e4m3fnuz type (8-bit floating-point) including +/// conversions to standard C types and basic arithmetic operations. Note that +/// arithmetic operations are implemented by converting to floating point and +/// performing the operation in float32. +/// Binary configuration remains the same as Float8_e4m3fn: +/// s eeee mmm +/// 1 sign bit +/// 4 exponent bits +/// 3 mantissa bits +/// The key differences versus Float8_e4m3fn are: +/// bias = 8 +/// no infinities or negative zero +/// NaN only when sign bit is 1, rest all 0s +/// +/// Implementation based on the paper https://arxiv.org/pdf/2206.02915.pdf and +/// the existing Float8_e4m3fn implementation. + +#include +#include +#include + +#include + +#if defined(__cplusplus) +#include +#elif !defined(__OPENCL_VERSION__) +#include +#include +#endif + +#include +#include + +namespace c10 { + +struct alignas(1) Float8_e4m3fnuz { + uint8_t x; + + struct from_bits_t {}; + C10_HOST_DEVICE static constexpr from_bits_t from_bits() { + return from_bits_t(); + } + + Float8_e4m3fnuz() = default; + + constexpr C10_HOST_DEVICE Float8_e4m3fnuz( + uint8_t bits, + from_bits_t /*unused*/) + : x(bits) {} + inline C10_HOST_DEVICE Float8_e4m3fnuz(float value); + inline C10_HOST_DEVICE operator float() const; + inline C10_HOST_DEVICE bool isnan() const; +}; + +inline std::ostream& operator<<( + std::ostream& out, + const Float8_e4m3fnuz& value) { + out << (float)value; + return out; +} + +namespace detail { + +/* + * Convert a 32-bit floating-point number in IEEE single-precision format to a + * 8-bit floating-point number in fp8 E4M3FNUZ format, in bit representation. + */ +inline C10_HOST_DEVICE uint8_t fp8e4m3fnuz_from_fp32_value(float f) { + /* + * Binary representation of 256.0f, which is the first value not representable + * (i.e. the first value which would overflow in to the sign bit, resulting in + * a NaN) in fp8e4m3fnuz range: + * 1 0000 000 - fp8e4m3fnuz + * 0 10000111 00000000000000000000000 - fp32 + */ + constexpr uint32_t fnuz_max = UINT32_C(0x87) << 23; + + /* + * A mask for converting fp32 numbers lower than fp8e4m3fnuz normal range + * into denorm representation + * magic number: ((127 - 8) + (23 - 3) + 1) + */ + constexpr uint32_t denorm_mask = UINT32_C(0x8C) << 23; + + uint32_t f_bits = fp32_to_bits(f); + + uint32_t result = 0u; + + /* + * Extract the sign of the input number into the high bit of the 32-bit word: + * + * +---+----------------------------------+ + * | S |0000000 00000000 00000000 00000000| + * +---+----------------------------------+ + * Bits 31 0-31 + */ + const uint32_t sign = f_bits & UINT32_C(0x80000000); + + /* + * Set sign bit to 0 + */ + f_bits ^= sign; + + if (f_bits >= fnuz_max) { + // NaN -- sign bit set to 1, rest 0s. + return 0x80; + } + + if (f_bits < (UINT32_C(0x78) << 23) /* 2^-7 in float32 */) { + // Input exponent is less than -7, the smallest e4m3fnuz exponent, so the + // number will become subnormal. + f_bits = fp32_to_bits(fp32_from_bits(f_bits) + fp32_from_bits(denorm_mask)); + result = static_cast(f_bits - denorm_mask); + if (result == 0) { + // fnuz types don't have negative zero. + return 0; + } + } else { + // resulting mantissa is odd + uint8_t mant_odd = (f_bits >> 20) & 1; + + // update exponent, rounding bias part 1 + f_bits += ((uint32_t)(8 - 127) << 23) + 0x7FFFF; + + // rounding bias part 2 + f_bits += mant_odd; + + // take the bits! + result = static_cast(f_bits >> 20); + } + + result |= sign >> 24; + return result; +} + +} // namespace detail + +//------ below is copied from c10/util/Float8_e4m3fnuz-inl.h ------// +C10_CLANG_DIAGNOSTIC_PUSH() +#if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion") +#endif + +/// Constructors + +inline C10_HOST_DEVICE Float8_e4m3fnuz::Float8_e4m3fnuz(float value) + : x(detail::fp8e4m3fnuz_from_fp32_value(value)) {} + +/// Implicit conversions + +inline C10_HOST_DEVICE Float8_e4m3fnuz::operator float() const { + return torch::headeronly::detail::fp8_fnuz_to_fp32_value<4, 3>(x); +} + +/// Special values helper + +inline C10_HOST_DEVICE bool Float8_e4m3fnuz::isnan() const { + return x == 0b10000000; +} + +/// Arithmetic + +inline C10_HOST_DEVICE Float8_e4m3fnuz +operator+(const Float8_e4m3fnuz& a, const Float8_e4m3fnuz& b) { + return static_cast(a) + static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fnuz +operator-(const Float8_e4m3fnuz& a, const Float8_e4m3fnuz& b) { + return static_cast(a) - static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fnuz +operator*(const Float8_e4m3fnuz& a, const Float8_e4m3fnuz& b) { + return static_cast(a) * static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fnuz operator/( + const Float8_e4m3fnuz& a, + const Float8_e4m3fnuz& b) __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fnuz operator-(const Float8_e4m3fnuz& a) { + return -static_cast(a); +} + +inline C10_HOST_DEVICE Float8_e4m3fnuz& operator+=( + Float8_e4m3fnuz& a, + const Float8_e4m3fnuz& b) { + a = a + b; + return a; +} + +inline C10_HOST_DEVICE Float8_e4m3fnuz& operator-=( + Float8_e4m3fnuz& a, + const Float8_e4m3fnuz& b) { + a = a - b; + return a; +} + +inline C10_HOST_DEVICE Float8_e4m3fnuz& operator*=( + Float8_e4m3fnuz& a, + const Float8_e4m3fnuz& b) { + a = a * b; + return a; +} + +inline C10_HOST_DEVICE Float8_e4m3fnuz& operator/=( + Float8_e4m3fnuz& a, + const Float8_e4m3fnuz& b) { + a = a / b; + return a; +} + +/// Arithmetic with floats + +inline C10_HOST_DEVICE float operator+(Float8_e4m3fnuz a, float b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE float operator-(Float8_e4m3fnuz a, float b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE float operator*(Float8_e4m3fnuz a, float b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE float operator/(Float8_e4m3fnuz a, float b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE float operator+(float a, Float8_e4m3fnuz b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE float operator-(float a, Float8_e4m3fnuz b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE float operator*(float a, Float8_e4m3fnuz b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE float operator/(float a, Float8_e4m3fnuz b) + __ubsan_ignore_float_divide_by_zero__ { + return a / static_cast(b); +} + +inline C10_HOST_DEVICE float& operator+=(float& a, const Float8_e4m3fnuz& b) { + return a += static_cast(b); +} +inline C10_HOST_DEVICE float& operator-=(float& a, const Float8_e4m3fnuz& b) { + return a -= static_cast(b); +} +inline C10_HOST_DEVICE float& operator*=(float& a, const Float8_e4m3fnuz& b) { + return a *= static_cast(b); +} +inline C10_HOST_DEVICE float& operator/=(float& a, const Float8_e4m3fnuz& b) { + return a /= static_cast(b); +} + +/// Arithmetic with doubles + +inline C10_HOST_DEVICE double operator+(Float8_e4m3fnuz a, double b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE double operator-(Float8_e4m3fnuz a, double b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE double operator*(Float8_e4m3fnuz a, double b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE double operator/(Float8_e4m3fnuz a, double b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE double operator+(double a, Float8_e4m3fnuz b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE double operator-(double a, Float8_e4m3fnuz b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE double operator*(double a, Float8_e4m3fnuz b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE double operator/(double a, Float8_e4m3fnuz b) + __ubsan_ignore_float_divide_by_zero__ { + return a / static_cast(b); +} + +/// Arithmetic with ints + +inline C10_HOST_DEVICE Float8_e4m3fnuz operator+(Float8_e4m3fnuz a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator-(Float8_e4m3fnuz a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator*(Float8_e4m3fnuz a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator/(Float8_e4m3fnuz a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fnuz operator+(int a, Float8_e4m3fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator-(int a, Float8_e4m3fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator*(int a, Float8_e4m3fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator/(int a, Float8_e4m3fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) / b; +} + +//// Arithmetic with int64_t + +inline C10_HOST_DEVICE Float8_e4m3fnuz operator+(Float8_e4m3fnuz a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator-(Float8_e4m3fnuz a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator*(Float8_e4m3fnuz a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator/(Float8_e4m3fnuz a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e4m3fnuz operator+(int64_t a, Float8_e4m3fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator-(int64_t a, Float8_e4m3fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator*(int64_t a, Float8_e4m3fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE Float8_e4m3fnuz operator/(int64_t a, Float8_e4m3fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) / b; +} + +/// NOTE: we do not define comparisons directly and instead rely on the implicit +/// conversion from c10::Float8_e4m3fnuz to float. + +C10_CLANG_DIAGNOSTIC_POP() + +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) +using c10::Float8_e4m3fnuz; +using c10::operator+; +using c10::operator-; +using c10::operator*; +using c10::operator/; +using c10::operator+=; +using c10::operator-=; +using c10::operator*=; +using c10::operator/=; +using c10::operator<<; + +namespace detail { +using c10::detail::fp8e4m3fnuz_from_fp32_value; +} // namespace detail + +HIDDEN_NAMESPACE_END(torch, headeronly) + +namespace std { + +template <> +class numeric_limits { + public: + static constexpr bool is_specialized = true; + static constexpr bool is_signed = true; + static constexpr bool is_integer = false; + static constexpr bool is_exact = false; + static constexpr bool has_infinity = false; + static constexpr bool has_quiet_NaN = true; + static constexpr bool has_signaling_NaN = false; + static constexpr auto has_denorm = true; + static constexpr auto has_denorm_loss = true; + static constexpr auto round_style = numeric_limits::round_style; + static constexpr bool is_iec559 = false; + static constexpr bool is_bounded = true; + static constexpr bool is_modulo = false; + static constexpr int digits = 4; + static constexpr int digits10 = 0; + static constexpr int max_digits10 = 3; + static constexpr int radix = 2; + static constexpr int min_exponent = -6; + static constexpr int min_exponent10 = -1; + static constexpr int max_exponent = 8; + static constexpr int max_exponent10 = 2; + static constexpr auto traps = numeric_limits::traps; + static constexpr auto tinyness_before = false; + + static constexpr c10::Float8_e4m3fnuz min() { + return c10::Float8_e4m3fnuz(0x08, c10::Float8_e4m3fnuz::from_bits()); + } + static constexpr c10::Float8_e4m3fnuz lowest() { + return c10::Float8_e4m3fnuz(0xFF, c10::Float8_e4m3fnuz::from_bits()); + } + static constexpr c10::Float8_e4m3fnuz max() { + return c10::Float8_e4m3fnuz(0x7F, c10::Float8_e4m3fnuz::from_bits()); + } + static constexpr c10::Float8_e4m3fnuz epsilon() { + return c10::Float8_e4m3fnuz(0x28, c10::Float8_e4m3fnuz::from_bits()); + } + static constexpr c10::Float8_e4m3fnuz round_error() { + return c10::Float8_e4m3fnuz(0x38, c10::Float8_e4m3fnuz::from_bits()); + } + static constexpr c10::Float8_e4m3fnuz infinity() { + // NaN (no infinities) + return c10::Float8_e4m3fnuz(0x80, c10::Float8_e4m3fnuz::from_bits()); + } + static constexpr c10::Float8_e4m3fnuz quiet_NaN() { + return c10::Float8_e4m3fnuz(0x80, c10::Float8_e4m3fnuz::from_bits()); + } + static constexpr c10::Float8_e4m3fnuz denorm_min() { + return c10::Float8_e4m3fnuz(0x01, c10::Float8_e4m3fnuz::from_bits()); + } +}; + +} // namespace std diff --git a/backends/pytorch/include/torch/headeronly/util/Float8_e5m2.h b/backends/pytorch/include/torch/headeronly/util/Float8_e5m2.h new file mode 100644 index 000000000..0aa856d0d --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Float8_e5m2.h @@ -0,0 +1,458 @@ +#pragma once + +/// Defines the Float8_e5m2 type (8-bit floating-point) including conversions +/// to standard C types and basic arithmetic operations. Note that arithmetic +/// operations are implemented by converting to floating point and +/// performing the operation in float32. +/// Binary configuration: +/// s eeeee mm +/// 1 sign bit +/// 5 exponent bits +/// 2 mantissa bits +/// bias = 15 +/// +/// Implementation based on the paper https://arxiv.org/pdf/2209.05433.pdf +/// and inspired by Half implementation from pytorch/c10/util/Half.h + +#include +#include + +#include + +namespace c10 { + +struct alignas(1) Float8_e5m2 { + uint8_t x; + + struct from_bits_t {}; + C10_HOST_DEVICE static constexpr from_bits_t from_bits() { + return from_bits_t(); + } + + Float8_e5m2() = default; + + constexpr C10_HOST_DEVICE Float8_e5m2(uint8_t bits, from_bits_t /*unused*/) + : x(bits) {} + inline C10_HOST_DEVICE Float8_e5m2(float value); + inline C10_HOST_DEVICE operator float() const; + inline C10_HOST_DEVICE bool isnan() const; + inline C10_HOST_DEVICE bool isinf() const; +}; + +inline std::ostream& operator<<(std::ostream& out, const Float8_e5m2& value) { + out << (float)value; + return out; +} + +namespace detail { + +/* + * Convert a 8-bit floating-point number in fp8 E5M2 format, in bit + * representation, to a 32-bit floating-point number in IEEE single-precision + * format, in bit representation. + * + * @note The implementation doesn't use any floating-point operations. + */ +inline C10_HOST_DEVICE float fp8e5m2_to_fp32_value(uint8_t input) { + /* + * Extend the fp8 E5M2 number to 32 bits and shift to the + * upper part of the 32-bit word: + * +---+----+---+-----------------------------+ + * | S |EEEEE|MM|0000 0000 0000 0000 0000 0000| + * +---+----+---+-----------------------------+ + * Bits 31 26-30 24-25 0-23 + * + * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 + * - zero bits. + */ + uint16_t half_representation = input; + half_representation <<= 8; + return fp16_ieee_to_fp32_value(half_representation); +} + +/* + * Convert a 32-bit floating-point number in IEEE single-precision format to a + * 8-bit floating-point number in fp8 E5M2 format, in bit representation. + */ +inline C10_HOST_DEVICE uint8_t fp8e5m2_from_fp32_value(float f) { + /* + * Binary representation of fp32 infinity + * 0 11111111 00000000000000000000000 + */ + constexpr uint32_t fp32_inf = UINT32_C(255) << 23; + + /* + * Binary representation of 65536.0f, which is the first value + * not representable in fp8e5m2 range: + * 0 11111 00 - fp8e5m2 + * 0 10001111 00000000000000000000000 - fp32 + */ + constexpr uint32_t fp8_max = UINT32_C(143) << 23; + + /* + * A mask for converting fp32 numbers lower than fp8e5m2 normal range + * into denorm representation + * magic number: ((127 - 15) + (23 - 2) + 1) + */ + constexpr uint32_t denorm_mask = UINT32_C(134) << 23; + + uint32_t f_bits = fp32_to_bits(f); + uint8_t result = 0u; + + /* + * Extract the sign of the input number into the high bit of the 32-bit word: + * + * +---+----------------------------------+ + * | S |0000000 00000000 00000000 00000000| + * +---+----------------------------------+ + * Bits 31 0-31 + */ + const uint32_t sign = f_bits & UINT32_C(0x80000000); + + /* + * Set sign bit to 0 + */ + f_bits ^= sign; + + if (f_bits >= fp8_max) { + // NaN - all exponent and mantissa bits set to 1 + result = f_bits > fp32_inf ? UINT8_C(0x7F) : UINT8_C(0x7C); + } else { + if (f_bits < (UINT32_C(113) << 23)) { + // Input number is smaller than 2^(-14), which is the smallest + // fp8e5m2 normal number + f_bits = + fp32_to_bits(fp32_from_bits(f_bits) + fp32_from_bits(denorm_mask)); + result = static_cast(f_bits - denorm_mask); + } else { + // resulting mantissa is odd + uint32_t mant_odd = (f_bits >> 21) & 1; + + // update exponent, rounding bias part 1 + f_bits += ((uint32_t)(15 - 127) << 23) + 0xFFFFF; + + // rounding bias part 2 + f_bits += mant_odd; + + // take the bits! + result = static_cast(f_bits >> 21); + } + } + + result |= static_cast(sign >> 24); + return result; +} + +} // namespace detail + +// -------- below is copied from c10/util/Float8_e5m2-inl.h --------// +C10_CLANG_DIAGNOSTIC_PUSH() +#if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion") +#endif + +#define EXP_WIDTH_FP8 5 +#define MAN_WIDTH_FP8 2 +#define EXP_BIAS_FP8 15 + +/// Constructors + +inline C10_HOST_DEVICE Float8_e5m2::Float8_e5m2(float value) + : x(detail::fp8e5m2_from_fp32_value(value)) {} + +/// Implicit conversions + +inline C10_HOST_DEVICE Float8_e5m2::operator float() const { + return detail::fp8e5m2_to_fp32_value(x); +} + +/// Special values helpers + +inline C10_HOST_DEVICE bool Float8_e5m2::isnan() const { + return (x & 0b01111111) > 0b01111100; +} + +inline C10_HOST_DEVICE bool Float8_e5m2::isinf() const { + return (x & 0b01111111) == 0b01111100; +} + +/// Arithmetic + +inline C10_HOST_DEVICE Float8_e5m2 +operator+(const Float8_e5m2& a, const Float8_e5m2& b) { + return static_cast(a) + static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2 +operator-(const Float8_e5m2& a, const Float8_e5m2& b) { + return static_cast(a) - static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2 +operator*(const Float8_e5m2& a, const Float8_e5m2& b) { + return static_cast(a) * static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2 operator/( + const Float8_e5m2& a, + const Float8_e5m2& b) __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2 operator-(const Float8_e5m2& a) { + return -static_cast(a); +} + +inline C10_HOST_DEVICE Float8_e5m2& operator+=( + Float8_e5m2& a, + const Float8_e5m2& b) { + a = a + b; + return a; +} + +inline C10_HOST_DEVICE Float8_e5m2& operator-=( + Float8_e5m2& a, + const Float8_e5m2& b) { + a = a - b; + return a; +} + +inline C10_HOST_DEVICE Float8_e5m2& operator*=( + Float8_e5m2& a, + const Float8_e5m2& b) { + a = a * b; + return a; +} + +inline C10_HOST_DEVICE Float8_e5m2& operator/=( + Float8_e5m2& a, + const Float8_e5m2& b) { + a = a / b; + return a; +} + +/// Arithmetic with floats + +inline C10_HOST_DEVICE float operator+(Float8_e5m2 a, float b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE float operator-(Float8_e5m2 a, float b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE float operator*(Float8_e5m2 a, float b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE float operator/(Float8_e5m2 a, float b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE float operator+(float a, Float8_e5m2 b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE float operator-(float a, Float8_e5m2 b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE float operator*(float a, Float8_e5m2 b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE float operator/(float a, Float8_e5m2 b) + __ubsan_ignore_float_divide_by_zero__ { + return a / static_cast(b); +} + +inline C10_HOST_DEVICE float& operator+=(float& a, const Float8_e5m2& b) { + return a += static_cast(b); +} +inline C10_HOST_DEVICE float& operator-=(float& a, const Float8_e5m2& b) { + return a -= static_cast(b); +} +inline C10_HOST_DEVICE float& operator*=(float& a, const Float8_e5m2& b) { + return a *= static_cast(b); +} +inline C10_HOST_DEVICE float& operator/=(float& a, const Float8_e5m2& b) { + return a /= static_cast(b); +} + +/// Arithmetic with doubles + +inline C10_HOST_DEVICE double operator+(Float8_e5m2 a, double b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE double operator-(Float8_e5m2 a, double b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE double operator*(Float8_e5m2 a, double b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE double operator/(Float8_e5m2 a, double b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE double operator+(double a, Float8_e5m2 b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE double operator-(double a, Float8_e5m2 b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE double operator*(double a, Float8_e5m2 b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE double operator/(double a, Float8_e5m2 b) + __ubsan_ignore_float_divide_by_zero__ { + return a / static_cast(b); +} + +/// Arithmetic with ints + +inline C10_HOST_DEVICE Float8_e5m2 operator+(Float8_e5m2 a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2 operator-(Float8_e5m2 a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2 operator*(Float8_e5m2 a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2 operator/(Float8_e5m2 a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2 operator+(int a, Float8_e5m2 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE Float8_e5m2 operator-(int a, Float8_e5m2 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE Float8_e5m2 operator*(int a, Float8_e5m2 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE Float8_e5m2 operator/(int a, Float8_e5m2 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) / b; +} + +//// Arithmetic with int64_t + +inline C10_HOST_DEVICE Float8_e5m2 operator+(Float8_e5m2 a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2 operator-(Float8_e5m2 a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2 operator*(Float8_e5m2 a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2 operator/(Float8_e5m2 a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2 operator+(int64_t a, Float8_e5m2 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE Float8_e5m2 operator-(int64_t a, Float8_e5m2 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE Float8_e5m2 operator*(int64_t a, Float8_e5m2 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE Float8_e5m2 operator/(int64_t a, Float8_e5m2 b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) / b; +} + +/// NOTE: we do not define comparisons directly and instead rely on the implicit +/// conversion from c10::Float8_e5m2 to float. +C10_CLANG_DIAGNOSTIC_POP() +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) +using c10::Float8_e5m2; +using c10::operator<<; +using c10::operator+; +using c10::operator-; +using c10::operator*; +using c10::operator/; +using c10::operator+=; +using c10::operator-=; +using c10::operator*=; +using c10::operator/=; + +namespace detail { +using c10::detail::fp8e5m2_from_fp32_value; +using c10::detail::fp8e5m2_to_fp32_value; +} // namespace detail +HIDDEN_NAMESPACE_END(torch, headeronly) + +namespace std { + +template <> +class numeric_limits { + public: + static constexpr bool is_signed = true; + static constexpr bool is_integer = false; + static constexpr bool is_specialized = true; + static constexpr bool is_exact = false; + static constexpr bool has_infinity = true; + static constexpr bool has_quiet_NaN = true; + static constexpr bool has_signaling_NaN = false; + static constexpr auto has_denorm = true; + static constexpr auto has_denorm_loss = true; + static constexpr auto round_style = numeric_limits::round_style; + static constexpr bool is_iec559 = false; + static constexpr bool is_bounded = true; + static constexpr bool is_modulo = false; + static constexpr int digits = 3; + static constexpr int digits10 = 0; + static constexpr int max_digits10 = 2; + static constexpr int radix = 2; + static constexpr int min_exponent = -13; + static constexpr int min_exponent10 = -4; + static constexpr int max_exponent = 16; + static constexpr int max_exponent10 = 4; + static constexpr auto traps = numeric_limits::traps; + static constexpr auto tinyness_before = + numeric_limits::tinyness_before; + + static constexpr c10::Float8_e5m2 min() { + return c10::Float8_e5m2(0x4, c10::Float8_e5m2::from_bits()); + } + static constexpr c10::Float8_e5m2 max() { + return c10::Float8_e5m2(0x7B, c10::Float8_e5m2::from_bits()); + } + static constexpr c10::Float8_e5m2 lowest() { + return c10::Float8_e5m2(0xFB, c10::Float8_e5m2::from_bits()); + } + static constexpr c10::Float8_e5m2 epsilon() { + return c10::Float8_e5m2(0x34, c10::Float8_e5m2::from_bits()); + } + static constexpr c10::Float8_e5m2 round_error() { + return c10::Float8_e5m2(0x38, c10::Float8_e5m2::from_bits()); + } + static constexpr c10::Float8_e5m2 infinity() { + return c10::Float8_e5m2(0x7C, c10::Float8_e5m2::from_bits()); + } + static constexpr c10::Float8_e5m2 quiet_NaN() { + return c10::Float8_e5m2(0x7F, c10::Float8_e5m2::from_bits()); + } + static constexpr c10::Float8_e5m2 denorm_min() { + return c10::Float8_e5m2(0x01, c10::Float8_e5m2::from_bits()); + } +}; + +} // namespace std diff --git a/backends/pytorch/include/torch/headeronly/util/Float8_e5m2fnuz.h b/backends/pytorch/include/torch/headeronly/util/Float8_e5m2fnuz.h new file mode 100644 index 000000000..6b0f79b75 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Float8_e5m2fnuz.h @@ -0,0 +1,448 @@ +#pragma once + +/// Defines the Float8_e5m2fnuz type (8-bit floating-point) including +/// conversions to standard C types and basic arithmetic operations. Note that +/// arithmetic operations are implemented by converting to floating point and +/// performing the operation in float32. +/// Binary configuration remains the same as e5m2: +/// s eeeee mm +/// 1 sign bit +/// 5 exponent bits +/// 2 mantissa bits +/// The key differences that e5m2fnuz brings are: +/// bias = 16 +/// no infinities or negative zero +/// NaN only when sign bit is 1, rest all 0s +/// +/// Implementation based on the paper https://arxiv.org/pdf/2206.02915.pdf and +/// the existing Float8_e4m3fn implementation. + +#include +#include +#include +#include + +#if defined(__cplusplus) +#include +#elif !defined(__OPENCL_VERSION__) +#include +#include +#endif + +#include +#include + +namespace c10 { + +struct alignas(1) Float8_e5m2fnuz { + uint8_t x; + + struct from_bits_t {}; + C10_HOST_DEVICE static constexpr from_bits_t from_bits() { + return from_bits_t(); + } + + Float8_e5m2fnuz() = default; + + constexpr C10_HOST_DEVICE Float8_e5m2fnuz( + uint8_t bits, + from_bits_t /*unused*/) + : x(bits) {} + inline C10_HOST_DEVICE Float8_e5m2fnuz(float value); + inline C10_HOST_DEVICE operator float() const; + inline C10_HOST_DEVICE bool isnan() const; + inline C10_HOST_DEVICE bool isinf() const; +}; + +inline std::ostream& operator<<( + std::ostream& out, + const Float8_e5m2fnuz& value) { + out << (float)value; + return out; +} + +namespace detail { + +/* + * Convert a 32-bit floating-point number in IEEE single-precision format to a + * 8-bit floating-point number in fp8 E5M2 format, in bit representation. + */ +inline C10_HOST_DEVICE uint8_t fp8e5m2fnuz_from_fp32_value(float f) { + /* + * Binary representation of 65536.0f, which is the first value not + * representable (i.e. the first value which would overflow in to the sign + * bit, resulting in a NaN) in fp8e4m3fnuz range: + * 1 00000 00 - fp8e5m2fnuz + * 0 10001111 00000000000000000000000 - fp32 + */ + constexpr uint32_t fnuz_max = UINT32_C(0x8F) << 23; + + /* + * A mask for converting fp32 numbers lower than fp8e5m2fnuz normal range + * into denormalized representation. + * magic number: ((127 - 16) + (23 - 2) + 1) + */ + constexpr uint32_t denorm_mask = UINT32_C(0x85) << 23; + + uint32_t f_bits = fp32_to_bits(f); + uint32_t result = 0u; + + /* + * Extract the sign of the input number into the high bit of the 32-bit word: + * + * +---+----------------------------------+ + * | S |0000000 00000000 00000000 00000000| + * +---+----------------------------------+ + * Bits 31 0-31 + */ + const uint32_t sign = f_bits & UINT32_C(0x80000000); + + /* + * Set sign bit to 0 + */ + f_bits ^= sign; + + if (f_bits >= fnuz_max) { + // NaN -- sign bit set to 1, rest 0s + return 0x80; + } + + if (f_bits < (UINT32_C(0x70) << 23) /* 2^-15 in float32 */) { + // Input exponent is less than -15, the smallest e5m2fnuz exponent, so the + // number will become subnormal. + f_bits = fp32_to_bits(fp32_from_bits(f_bits) + fp32_from_bits(denorm_mask)); + result = static_cast(f_bits - denorm_mask); + if (result == 0) { + // fnuz types don't have negative zero. + return 0; + } + } else { + // resulting mantissa is odd + uint8_t mant_odd = (f_bits >> 21) & 1; + + // update exponent, rounding bias part 1 + f_bits += ((uint32_t)(16 - 127) << 23) + 0xFFFFF; + + // rounding bias part 2 + f_bits += mant_odd; + + // take the bits! + result = static_cast(f_bits >> 21); + } + + result |= sign >> 24; + return result; +} + +} // namespace detail + +//------ below is copied from c10/util/Float8_e5m2fnuz-inl.h ------// +C10_CLANG_DIAGNOSTIC_PUSH() +#if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion") +#endif + +/// Constructors + +inline C10_HOST_DEVICE Float8_e5m2fnuz::Float8_e5m2fnuz(float value) + : x(detail::fp8e5m2fnuz_from_fp32_value(value)) {} + +/// Implicit conversions + +inline C10_HOST_DEVICE Float8_e5m2fnuz::operator float() const { + return torch::headeronly::detail::fp8_fnuz_to_fp32_value<5, 2>(x); +} + +/// Special values helpers + +inline C10_HOST_DEVICE bool Float8_e5m2fnuz::isnan() const { + return x == 0b10000000; +} + +inline C10_HOST_DEVICE bool Float8_e5m2fnuz::isinf() const { + return false; +} + +/// Arithmetic + +inline C10_HOST_DEVICE Float8_e5m2fnuz +operator+(const Float8_e5m2fnuz& a, const Float8_e5m2fnuz& b) { + return static_cast(a) + static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2fnuz +operator-(const Float8_e5m2fnuz& a, const Float8_e5m2fnuz& b) { + return static_cast(a) - static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2fnuz +operator*(const Float8_e5m2fnuz& a, const Float8_e5m2fnuz& b) { + return static_cast(a) * static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2fnuz operator/( + const Float8_e5m2fnuz& a, + const Float8_e5m2fnuz& b) __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2fnuz operator-(const Float8_e5m2fnuz& a) { + return -static_cast(a); +} + +inline C10_HOST_DEVICE Float8_e5m2fnuz& operator+=( + Float8_e5m2fnuz& a, + const Float8_e5m2fnuz& b) { + a = a + b; + return a; +} + +inline C10_HOST_DEVICE Float8_e5m2fnuz& operator-=( + Float8_e5m2fnuz& a, + const Float8_e5m2fnuz& b) { + a = a - b; + return a; +} + +inline C10_HOST_DEVICE Float8_e5m2fnuz& operator*=( + Float8_e5m2fnuz& a, + const Float8_e5m2fnuz& b) { + a = a * b; + return a; +} + +inline C10_HOST_DEVICE Float8_e5m2fnuz& operator/=( + Float8_e5m2fnuz& a, + const Float8_e5m2fnuz& b) { + a = a / b; + return a; +} + +/// Arithmetic with floats + +inline C10_HOST_DEVICE float operator+(Float8_e5m2fnuz a, float b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE float operator-(Float8_e5m2fnuz a, float b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE float operator*(Float8_e5m2fnuz a, float b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE float operator/(Float8_e5m2fnuz a, float b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE float operator+(float a, Float8_e5m2fnuz b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE float operator-(float a, Float8_e5m2fnuz b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE float operator*(float a, Float8_e5m2fnuz b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE float operator/(float a, Float8_e5m2fnuz b) + __ubsan_ignore_float_divide_by_zero__ { + return a / static_cast(b); +} + +inline C10_HOST_DEVICE float& operator+=(float& a, const Float8_e5m2fnuz& b) { + return a += static_cast(b); +} +inline C10_HOST_DEVICE float& operator-=(float& a, const Float8_e5m2fnuz& b) { + return a -= static_cast(b); +} +inline C10_HOST_DEVICE float& operator*=(float& a, const Float8_e5m2fnuz& b) { + return a *= static_cast(b); +} +inline C10_HOST_DEVICE float& operator/=(float& a, const Float8_e5m2fnuz& b) { + return a /= static_cast(b); +} + +/// Arithmetic with doubles + +inline C10_HOST_DEVICE double operator+(Float8_e5m2fnuz a, double b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE double operator-(Float8_e5m2fnuz a, double b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE double operator*(Float8_e5m2fnuz a, double b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE double operator/(Float8_e5m2fnuz a, double b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE double operator+(double a, Float8_e5m2fnuz b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE double operator-(double a, Float8_e5m2fnuz b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE double operator*(double a, Float8_e5m2fnuz b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE double operator/(double a, Float8_e5m2fnuz b) + __ubsan_ignore_float_divide_by_zero__ { + return a / static_cast(b); +} + +/// Arithmetic with ints + +inline C10_HOST_DEVICE Float8_e5m2fnuz operator+(Float8_e5m2fnuz a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator-(Float8_e5m2fnuz a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator*(Float8_e5m2fnuz a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator/(Float8_e5m2fnuz a, int b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2fnuz operator+(int a, Float8_e5m2fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator-(int a, Float8_e5m2fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator*(int a, Float8_e5m2fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator/(int a, Float8_e5m2fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) / b; +} + +//// Arithmetic with int64_t + +inline C10_HOST_DEVICE Float8_e5m2fnuz operator+(Float8_e5m2fnuz a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator-(Float8_e5m2fnuz a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator*(Float8_e5m2fnuz a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator/(Float8_e5m2fnuz a, int64_t b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE Float8_e5m2fnuz operator+(int64_t a, Float8_e5m2fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator-(int64_t a, Float8_e5m2fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator*(int64_t a, Float8_e5m2fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE Float8_e5m2fnuz operator/(int64_t a, Float8_e5m2fnuz b) { + // NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions) + return static_cast(a) / b; +} + +/// NOTE: we do not define comparisons directly and instead rely on the implicit +/// conversion from c10::Float8_e5m2fnuz to float. + +C10_CLANG_DIAGNOSTIC_POP() + +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) +using c10::Float8_e5m2fnuz; +using c10::operator<<; +using c10::operator+; +using c10::operator-; +using c10::operator*; +using c10::operator/; +using c10::operator+=; +using c10::operator-=; +using c10::operator*=; +using c10::operator/=; + +namespace detail { +using c10::detail::fp8e5m2fnuz_from_fp32_value; +} +HIDDEN_NAMESPACE_END(torch, headeronly) + +namespace std { + +template <> +class numeric_limits { + public: + static constexpr bool is_signed = true; + static constexpr bool is_integer = false; + static constexpr bool is_specialized = true; + static constexpr bool is_exact = false; + static constexpr bool has_infinity = false; + static constexpr bool has_quiet_NaN = true; + static constexpr bool has_signaling_NaN = false; + static constexpr auto has_denorm = true; + static constexpr auto has_denorm_loss = true; + static constexpr auto round_style = numeric_limits::round_style; + static constexpr bool is_iec559 = false; + static constexpr bool is_bounded = true; + static constexpr bool is_modulo = false; + static constexpr int digits = 3; + static constexpr int digits10 = 0; + static constexpr int max_digits10 = 2; + static constexpr int radix = 2; + static constexpr int min_exponent = -14; + static constexpr int min_exponent10 = -4; + static constexpr int max_exponent = 16; + static constexpr int max_exponent10 = 4; + static constexpr auto traps = numeric_limits::traps; + static constexpr auto tinyness_before = + numeric_limits::tinyness_before; + + static constexpr c10::Float8_e5m2fnuz min() { + return c10::Float8_e5m2fnuz(0x04, c10::Float8_e5m2fnuz::from_bits()); + } + static constexpr c10::Float8_e5m2fnuz max() { + return c10::Float8_e5m2fnuz(0x7F, c10::Float8_e5m2fnuz::from_bits()); + } + static constexpr c10::Float8_e5m2fnuz lowest() { + return c10::Float8_e5m2fnuz(0xFF, c10::Float8_e5m2fnuz::from_bits()); + } + static constexpr c10::Float8_e5m2fnuz epsilon() { + return c10::Float8_e5m2fnuz(0x34, c10::Float8_e5m2fnuz::from_bits()); + } + static constexpr c10::Float8_e5m2fnuz round_error() { + return c10::Float8_e5m2fnuz(0x38, c10::Float8_e5m2fnuz::from_bits()); + } + static constexpr c10::Float8_e5m2fnuz infinity() { + return c10::Float8_e5m2fnuz(0x80, c10::Float8_e5m2fnuz::from_bits()); + } + // TODO(future): we are mapping neg_zero to both inf and NaN, this is + // surprising and we should figure out what to do about it. + static constexpr c10::Float8_e5m2fnuz quiet_NaN() { + return c10::Float8_e5m2fnuz(0x80, c10::Float8_e5m2fnuz::from_bits()); + } + static constexpr c10::Float8_e5m2fnuz denorm_min() { + return c10::Float8_e5m2fnuz(0x01, c10::Float8_e5m2fnuz::from_bits()); + } +}; + +} // namespace std diff --git a/backends/pytorch/include/torch/headeronly/util/Float8_e8m0fnu.h b/backends/pytorch/include/torch/headeronly/util/Float8_e8m0fnu.h new file mode 100644 index 000000000..80d89439c --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Float8_e8m0fnu.h @@ -0,0 +1,226 @@ +#pragma once + +/// Defines the Float8_e8m0fnu type (8-bit floating-point) including +/// conversions to standard C types +/// Binary configuration : +/// eeeeeeee +/// no sign bits +/// 8 exponent bits +/// no mantissa bits +/// +/// This is the E8M0 dtype from the OCP MX format spec +/// (https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf, +/// Section 5.4.1) + +#include +#include + +// TODO(#146647): do we need to special case OPENCL? +#if defined(__cplusplus) +#include +#elif !defined(__OPENCL_VERSION__) +#include +#include +#endif + +#include +#include +#include + +namespace c10 { + +struct alignas(1) Float8_e8m0fnu { + uint8_t x; + + struct from_bits_t {}; + C10_HOST_DEVICE static constexpr from_bits_t from_bits() { + return from_bits_t(); + } + + Float8_e8m0fnu() = default; + + constexpr C10_HOST_DEVICE Float8_e8m0fnu(uint8_t bits, from_bits_t /*unused*/) + : x(bits) {} + inline C10_HOST_DEVICE Float8_e8m0fnu(float value); + inline C10_HOST_DEVICE operator float() const; + inline C10_HOST_DEVICE bool isnan() const; +}; + +inline std::ostream& operator<<( + std::ostream& out, + const Float8_e8m0fnu& value) { + out << (float)value; + return out; +} + +namespace detail { +/* + * Convert a 32-bit floating-point number in IEEE single-precision format to a + * 8-bit floating-point number in fp8 e8m0fnu format, in bit representation. + */ +inline C10_HOST_DEVICE uint8_t fp8e8m0fnu_from_fp32_value(float f) { + // TODO(#146647): maybe rewrite without control flow + + uint32_t f_bits = c10::detail::fp32_to_bits(f); + + // extract the exponent + uint32_t exponent = (f_bits >> 23) & 0b11111111; + + // special case float32 NaN and +-inf to map to e8m0 nan + if (exponent == 0b11111111) { + return exponent; + } + + // next, we use guard, round, sticky bits and the LSB to implement round to + // nearest, with ties to even + + // guard bit - bit 23, or 22 zero-indexed + uint8_t g = (f_bits & 0x400000) > 0; + // round bit - bit 22, or 21 zero-indexed + uint8_t r = (f_bits & 0x200000) > 0; + // sticky bit - bits 21 to 1, or 20 to 0 zero-indexed + uint8_t s = (f_bits & 0x1FFFFF) > 0; + // in casting to e8m0, LSB is the implied mantissa bit. It equals to 0 if the + // original float32 is denormal, and to 1 if the original float32 is normal. + uint8_t lsb = exponent > 0; + + // implement the RNE logic + bool round_up = false; + + // if g == 0, round down (no-op) + if (g == 1) { + if ((r == 1) || (s == 1)) { + // round up + round_up = true; + } else { + if (lsb == 1) { + // round up + round_up = true; + } + // if lsb == 0, round down (no-op) + } + } + + if (round_up) { + // adjust exponent + // note that if exponent was 255 we would have already returned earlier, so + // we know we can add one safely without running out of bounds + exponent++; + } + + return exponent; +} + +} // namespace detail + +//------- the below is from c10/util/Float8_e8m0fnu-inl.h ------// +// TODO(#146647): Can we remove the below warning? +C10_CLANG_DIAGNOSTIC_PUSH() +#if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion") +#endif + +/// Constructors +inline C10_HOST_DEVICE Float8_e8m0fnu::Float8_e8m0fnu(float value) + : x(detail::fp8e8m0fnu_from_fp32_value(value)) {} + +/// Implicit conversions + +inline C10_HOST_DEVICE Float8_e8m0fnu::operator float() const { + // TODO(#146647): maybe rewrite without control flow + + // if exponent is zero, need to special case to return 2^-127 instead of zero + if (x == 0) { + return c10::detail::fp32_from_bits(0x00400000); + } + + // if exponent is NaN, need to special case to return properly encoded NaN + if (isnan()) { + return c10::detail::fp32_from_bits(0x7f800001); + } + + // leave sign at 0, set the exponent bits, leave stored mantissa at 0 + uint32_t res = x << 23; + + return c10::detail::fp32_from_bits(res); +} + +/// Special values helper + +inline C10_HOST_DEVICE bool Float8_e8m0fnu::isnan() const { + return x == 0b11111111; +} + +/// NOTE: we do not define comparisons directly and instead rely on the implicit +/// conversion from c10::Float8_e8m0fnu to float. +C10_CLANG_DIAGNOSTIC_POP() + +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) +using c10::Float8_e8m0fnu; +using c10::operator<<; + +namespace detail { +using c10::detail::fp8e8m0fnu_from_fp32_value; +} // namespace detail +HIDDEN_NAMESPACE_END(torch, headeronly) + +namespace std { + +template <> +class numeric_limits { + public: + static constexpr bool is_specialized = true; + static constexpr bool is_signed = false; + static constexpr bool is_integer = false; + static constexpr bool is_exact = false; + static constexpr bool has_infinity = false; + static constexpr bool has_quiet_NaN = true; + static constexpr bool has_signaling_NaN = false; + static constexpr auto has_denorm = false; + static constexpr auto has_denorm_loss = false; + static constexpr auto round_style = numeric_limits::round_style; + static constexpr bool is_iec559 = false; + static constexpr bool is_bounded = true; + static constexpr bool is_modulo = false; + static constexpr int digits = 1; + static constexpr int digits10 = 0; + static constexpr int max_digits10 = 1; // just a 2! + static constexpr int radix = 2; + static constexpr int min_exponent = -126; + static constexpr int min_exponent10 = -38; + static constexpr int max_exponent = 128; + static constexpr int max_exponent10 = 38; + static constexpr auto traps = numeric_limits::traps; + static constexpr auto tinyness_before = false; + + static constexpr c10::Float8_e8m0fnu min() { + // 2^-127 + return c10::Float8_e8m0fnu(0b00000000, c10::Float8_e8m0fnu::from_bits()); + } + static constexpr c10::Float8_e8m0fnu lowest() { + // 2^-127 + return c10::Float8_e8m0fnu(0b00000000, c10::Float8_e8m0fnu::from_bits()); + } + static constexpr c10::Float8_e8m0fnu max() { + // 254 biased, which is 127 unbiased, so 2^127 + return c10::Float8_e8m0fnu(0b11111110, c10::Float8_e8m0fnu::from_bits()); + } + static constexpr c10::Float8_e8m0fnu epsilon() { + // according to https://en.cppreference.com/w/cpp/types/numeric_limits, this + // is "the difference between 1.0 and the next representable value of the + // given floating-point type". The next representable value is 2.0, so the + // difference is 1.0 which is 2^0. 0 unbiased is 127 biased. + return c10::Float8_e8m0fnu(0b01111111, c10::Float8_e8m0fnu::from_bits()); + } + static constexpr c10::Float8_e8m0fnu round_error() { + // 0.5 in float, which is 2^-1, and -1 + 127 = 126 + return c10::Float8_e8m0fnu(0b01111110, c10::Float8_e8m0fnu::from_bits()); + } + static constexpr c10::Float8_e8m0fnu quiet_NaN() { + return c10::Float8_e8m0fnu(0b11111111, c10::Float8_e8m0fnu::from_bits()); + } +}; + +} // namespace std diff --git a/backends/pytorch/include/torch/headeronly/util/Float8_fnuz_cvt.h b/backends/pytorch/include/torch/headeronly/util/Float8_fnuz_cvt.h new file mode 100644 index 000000000..59fab8cc2 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Float8_fnuz_cvt.h @@ -0,0 +1,69 @@ +#pragma once + +#include +#include + +#include + +#if defined(SYCL_LANGUAGE_VERSION) +#include +#endif + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly, detail) + +/* + * Convert a 8-bit floating-point number in either f8 E4M3FNUZ or bf8 E5M2FNUZ + * format, in bit representation, to a 32-bit floating-point number. + */ +template +inline C10_HOST_DEVICE float fp8_fnuz_to_fp32_value(uint8_t x) { + static_assert((we == 4 && wm == 3) || (we == 5 && wm == 2)); + constexpr uint32_t weo = 8; + constexpr uint32_t wmo = 23; + + if (x == 0) { + return 0; + } + + if (x == 0x80) { + constexpr uint32_t ifNaN = 0x7F800001; + return fp32_from_bits(ifNaN); + } + + uint32_t mantissa = x & ((1 << wm) - 1); + uint32_t exponent = (x & 0x7F) >> wm; + + // subnormal input + if (exponent == 0) { + // guaranteed mantissa!=0 since cases 0x0 and 0x80 are handled above +#if defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__) + uint32_t renorm_shift = __clz(mantissa); +#elif defined(__SYCL_DEVICE_ONLY__) + uint32_t renorm_shift = sycl::clz(mantissa); +#elif defined(_MSC_VER) + unsigned long nonsign_bsr; + _BitScanReverse(&nonsign_bsr, (unsigned long)mantissa); + uint32_t renorm_shift = (uint32_t)nonsign_bsr ^ 31; +#else + uint32_t renorm_shift = __builtin_clz(mantissa); +#endif + uint32_t sh = 1 + renorm_shift - (32 - wm); + mantissa <<= sh; + exponent += 1 - sh; + mantissa &= ((1 << wm) - 1); + } + + const uint32_t exp_low_cutoff = (1 << (weo - 1)) - (1 << (we - 1)); + exponent += exp_low_cutoff - 1; + mantissa <<= wmo - wm; + + uint32_t sign = x >> 7; + uint32_t retval = (sign << 31) | (exponent << 23) | mantissa; + return fp32_from_bits(retval); +} + +HIDDEN_NAMESPACE_END(torch, headeronly, detail) + +namespace c10::detail { +using torch::headeronly::detail::fp8_fnuz_to_fp32_value; +} diff --git a/backends/pytorch/include/torch/headeronly/util/Half.h b/backends/pytorch/include/torch/headeronly/util/Half.h new file mode 100644 index 000000000..a9c0b166b --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Half.h @@ -0,0 +1,788 @@ +#pragma once + +/// Defines the Half type (half-precision floating-point) including conversions +/// to standard C types and basic arithmetic operations. Note that arithmetic +/// operations are implemented by converting to floating point and +/// performing the operation in float32, instead of using CUDA half intrinsics. +/// Most uses of this type within ATen are memory bound, including the +/// element-wise kernels, and the half intrinsics aren't efficient on all GPUs. +/// If you are writing a compute bound kernel, you can use the CUDA half +/// intrinsics directly on the Half type from device code. + +#include +#include +#include + +#if defined(__cplusplus) +#include +#elif !defined(__OPENCL_VERSION__) +#include +#endif + +#ifdef _MSC_VER +#include +#endif + +#include +#include +#include + +#ifdef __CUDACC__ +#include +#endif + +#ifdef __HIPCC__ +#include +#endif + +#if defined(CL_SYCL_LANGUAGE_VERSION) +#include // for SYCL 1.2.1 +#elif defined(SYCL_LANGUAGE_VERSION) +#include // for SYCL 2020 +#endif + +#if (defined(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_AVX512)) && \ + !defined(__APPLE__) +#include +#endif + +#if defined(__aarch64__) && !defined(__CUDACC__) +#include +#endif + +#if defined(__GNUC__) || defined(__clang__) +#if defined(__x86_64__) || defined(_M_X64) || defined(__i386) || \ + defined(_M_IX86) +#if defined(__F16C__) && \ + !(defined(__CUDA_ARCH__) || defined(__CUDACC__) || \ + defined(__HIP_DEVICE_COMPILE__)) +#define C10_X86_F16 1 +#include // import conversion ops from f16cintrin.h +#endif // defined(__F16C__) && !(defined(__CUDA_ARCH__) || defined(__CUDACC__) + // || defined(__HIP_DEVICE_COMPILE__)) +#endif // __x86_64__ || _M_X64 || __i386 || _M_IX86 +#endif // __GNUC__ || __clang__ + +namespace c10 { + +struct alignas(2) Half { + unsigned short x; + + struct from_bits_t {}; + C10_HOST_DEVICE static constexpr from_bits_t from_bits() { + return from_bits_t(); + } + + // HIP wants __host__ __device__ tag, CUDA does not +#if defined(USE_ROCM) + C10_HOST_DEVICE Half() = default; +#else + Half() = default; +#endif + + constexpr C10_HOST_DEVICE Half(unsigned short bits, from_bits_t /*unused*/) + : x(bits) {} +#if defined(__aarch64__) && !defined(__CUDACC__) + inline Half(float16_t value); + inline operator float16_t() const; +#else + inline C10_HOST_DEVICE Half(float value); + inline C10_HOST_DEVICE operator float() const; +#endif + +#if defined(__CUDACC__) || defined(__HIPCC__) + inline C10_HOST_DEVICE Half(const __half& value); + inline C10_HOST_DEVICE operator __half() const; +#endif +#ifdef SYCL_LANGUAGE_VERSION + inline C10_HOST_DEVICE Half(const sycl::half& value); + inline C10_HOST_DEVICE operator sycl::half() const; +#endif +}; + +inline std::ostream& operator<<(std::ostream& out, const Half& value) { + out << (float)value; + return out; +} + +namespace detail { +/* + * Convert a 16-bit floating-point number in IEEE half-precision format, in bit + * representation, to a 32-bit floating-point number in IEEE single-precision + * format. + * + * @note The implementation relies on IEEE-like (no assumption about rounding + * mode and no operations on denormals) floating-point operations and bitcasts + * between integer and floating-point variables. + */ +C10_HOST_DEVICE inline float fp16_ieee_to_fp32_value(uint16_t h) { +#ifdef C10_X86_F16 + return _cvtsh_ss(h); +#else + /* + * Extend the half-precision floating-point number to 32 bits and shift to the + * upper part of the 32-bit word: + * +---+-----+------------+-------------------+ + * | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000| + * +---+-----+------------+-------------------+ + * Bits 31 26-30 16-25 0-15 + * + * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 + * - zero bits. + */ + const uint32_t w = (uint32_t)h << 16; + /* + * Extract the sign of the input number into the high bit of the 32-bit word: + * + * +---+----------------------------------+ + * | S |0000000 00000000 00000000 00000000| + * +---+----------------------------------+ + * Bits 31 0-31 + */ + const uint32_t sign = w & UINT32_C(0x80000000); + /* + * Extract mantissa and biased exponent of the input number into the high bits + * of the 32-bit word: + * + * +-----+------------+---------------------+ + * |EEEEE|MM MMMM MMMM|0 0000 0000 0000 0000| + * +-----+------------+---------------------+ + * Bits 27-31 17-26 0-16 + */ + const uint32_t two_w = w + w; + + /* + * Shift mantissa and exponent into bits 23-28 and bits 13-22 so they become + * mantissa and exponent of a single-precision floating-point number: + * + * S|Exponent | Mantissa + * +-+---+-----+------------+----------------+ + * |0|000|EEEEE|MM MMMM MMMM|0 0000 0000 0000| + * +-+---+-----+------------+----------------+ + * Bits | 23-31 | 0-22 + * + * Next, there are some adjustments to the exponent: + * - The exponent needs to be corrected by the difference in exponent bias + * between single-precision and half-precision formats (0x7F - 0xF = 0x70) + * - Inf and NaN values in the inputs should become Inf and NaN values after + * conversion to the single-precision number. Therefore, if the biased + * exponent of the half-precision input was 0x1F (max possible value), the + * biased exponent of the single-precision output must be 0xFF (max possible + * value). We do this correction in two steps: + * - First, we adjust the exponent by (0xFF - 0x1F) = 0xE0 (see exp_offset + * below) rather than by 0x70 suggested by the difference in the exponent bias + * (see above). + * - Then we multiply the single-precision result of exponent adjustment by + * 2**(-112) to reverse the effect of exponent adjustment by 0xE0 less the + * necessary exponent adjustment by 0x70 due to difference in exponent bias. + * The floating-point multiplication hardware would ensure than Inf and + * NaN would retain their value on at least partially IEEE754-compliant + * implementations. + * + * Note that the above operations do not handle denormal inputs (where biased + * exponent == 0). However, they also do not operate on denormal inputs, and + * do not produce denormal results. + */ + constexpr uint32_t exp_offset = UINT32_C(0xE0) << 23; + // const float exp_scale = 0x1.0p-112f; + constexpr uint32_t scale_bits = (uint32_t)15 << 23; + float exp_scale_val = 0; +#if defined(_MSC_VER) && defined(__clang__) + __builtin_memcpy(&exp_scale_val, &scale_bits, sizeof(exp_scale_val)); +#else + std::memcpy(&exp_scale_val, &scale_bits, sizeof(exp_scale_val)); +#endif + + const float exp_scale = exp_scale_val; + const float normalized_value = + fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale; + + /* + * Convert denormalized half-precision inputs into single-precision results + * (always normalized). Zero inputs are also handled here. + * + * In a denormalized number the biased exponent is zero, and mantissa has + * on-zero bits. First, we shift mantissa into bits 0-9 of the 32-bit word. + * + * zeros | mantissa + * +---------------------------+------------+ + * |0000 0000 0000 0000 0000 00|MM MMMM MMMM| + * +---------------------------+------------+ + * Bits 10-31 0-9 + * + * Now, remember that denormalized half-precision numbers are represented as: + * FP16 = mantissa * 2**(-24). + * The trick is to construct a normalized single-precision number with the + * same mantissa and thehalf-precision input and with an exponent which would + * scale the corresponding mantissa bits to 2**(-24). A normalized + * single-precision floating-point number is represented as: FP32 = (1 + + * mantissa * 2**(-23)) * 2**(exponent - 127) Therefore, when the biased + * exponent is 126, a unit change in the mantissa of the input denormalized + * half-precision number causes a change of the constructed single-precision + * number by 2**(-24), i.e. the same amount. + * + * The last step is to adjust the bias of the constructed single-precision + * number. When the input half-precision number is zero, the constructed + * single-precision number has the value of FP32 = 1 * 2**(126 - 127) = + * 2**(-1) = 0.5 Therefore, we need to subtract 0.5 from the constructed + * single-precision number to get the numerical equivalent of the input + * half-precision number. + */ + constexpr uint32_t magic_mask = UINT32_C(126) << 23; + constexpr float magic_bias = 0.5f; + const float denormalized_value = + fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias; + + /* + * - Choose either results of conversion of input as a normalized number, or + * as a denormalized number, depending on the input exponent. The variable + * two_w contains input exponent in bits 27-31, therefore if its smaller than + * 2**27, the input is either a denormal number, or zero. + * - Combine the result of conversion of exponent and mantissa with the sign + * of the input number. + */ + constexpr uint32_t denormalized_cutoff = UINT32_C(1) << 27; + const uint32_t result = sign | + (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) + : fp32_to_bits(normalized_value)); + return fp32_from_bits(result); +#endif // C10_X86_F16 +} + +/* + * Convert a 32-bit floating-point number in IEEE single-precision format to a + * 16-bit floating-point number in IEEE half-precision format, in bit + * representation. + * + * @note The implementation relies on IEEE-like (no assumption about rounding + * mode and no operations on denormals) floating-point operations and bitcasts + * between integer and floating-point variables. + */ +inline uint16_t fp16_ieee_from_fp32_value(float f) { +#ifdef C10_X86_F16 + return _cvtss_sh(f, _MM_FROUND_TO_NEAREST_INT); +#else + // const float scale_to_inf = 0x1.0p+112f; + // const float scale_to_zero = 0x1.0p-110f; + constexpr uint32_t scale_to_inf_bits = (uint32_t)239 << 23; + constexpr uint32_t scale_to_zero_bits = (uint32_t)17 << 23; + float scale_to_inf_val = 0, scale_to_zero_val = 0; + std::memcpy(&scale_to_inf_val, &scale_to_inf_bits, sizeof(scale_to_inf_val)); + std::memcpy( + &scale_to_zero_val, &scale_to_zero_bits, sizeof(scale_to_zero_val)); + const float scale_to_inf = scale_to_inf_val; + const float scale_to_zero = scale_to_zero_val; + +#if defined(_MSC_VER) && _MSC_VER == 1916 + float base = ((signbit(f) != 0 ? -f : f) * scale_to_inf) * scale_to_zero; +#else + float base = (fabsf(f) * scale_to_inf) * scale_to_zero; +#endif + + const uint32_t w = fp32_to_bits(f); + const uint32_t shl1_w = w + w; + const uint32_t sign = w & UINT32_C(0x80000000); + uint32_t bias = shl1_w & UINT32_C(0xFF000000); + if (bias < UINT32_C(0x71000000)) { + bias = UINT32_C(0x71000000); + } + + base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base; + const uint32_t bits = fp32_to_bits(base); + const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00); + const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF); + const uint32_t nonsign = exp_bits + mantissa_bits; + return static_cast( + (sign >> 16) | + (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign)); +#endif // C10_X86_F16 +} + +/* + * Convert a 16-bit floating-point number in IEEE half-precision format, in bit + * representation, to a 32-bit floating-point number in IEEE single-precision + * format, in bit representation. + * + * @note The implementation doesn't use any floating-point operations. + */ +inline uint32_t fp16_ieee_to_fp32_bits(uint16_t h) { + /* + * Extend the half-precision floating-point number to 32 bits and shift to the + * upper part of the 32-bit word: + * +---+-----+------------+-------------------+ + * | S |EEEEE|MM MMMM MMMM|0000 0000 0000 0000| + * +---+-----+------------+-------------------+ + * Bits 31 26-30 16-25 0-15 + * + * S - sign bit, E - bits of the biased exponent, M - bits of the mantissa, 0 + * - zero bits. + */ + const uint32_t w = (uint32_t)h << 16; + /* + * Extract the sign of the input number into the high bit of the 32-bit word: + * + * +---+----------------------------------+ + * | S |0000000 00000000 00000000 00000000| + * +---+----------------------------------+ + * Bits 31 0-31 + */ + const uint32_t sign = w & UINT32_C(0x80000000); + /* + * Extract mantissa and biased exponent of the input number into the bits 0-30 + * of the 32-bit word: + * + * +---+-----+------------+-------------------+ + * | 0 |EEEEE|MM MMMM MMMM|0000 0000 0000 0000| + * +---+-----+------------+-------------------+ + * Bits 30 27-31 17-26 0-16 + */ + const uint32_t nonsign = w & UINT32_C(0x7FFFFFFF); + /* + * Renorm shift is the number of bits to shift mantissa left to make the + * half-precision number normalized. If the initial number is normalized, some + * of its high 6 bits (sign == 0 and 5-bit exponent) equals one. In this case + * renorm_shift == 0. If the number is denormalize, renorm_shift > 0. Note + * that if we shift denormalized nonsign by renorm_shift, the unit bit of + * mantissa will shift into exponent, turning the biased exponent into 1, and + * making mantissa normalized (i.e. without leading 1). + */ +#ifdef _MSC_VER + unsigned long nonsign_bsr; + _BitScanReverse(&nonsign_bsr, (unsigned long)nonsign); + uint32_t renorm_shift = (uint32_t)nonsign_bsr ^ 31; +#else + uint32_t renorm_shift = __builtin_clz(nonsign); +#endif + renorm_shift = renorm_shift > 5 ? renorm_shift - 5 : 0; + /* + * Iff half-precision number has exponent of 15, the addition overflows + * it into bit 31, and the subsequent shift turns the high 9 bits + * into 1. Thus inf_nan_mask == 0x7F800000 if the half-precision number + * had exponent of 15 (i.e. was NaN or infinity) 0x00000000 otherwise + */ + const int32_t inf_nan_mask = + ((int32_t)(nonsign + 0x04000000) >> 8) & INT32_C(0x7F800000); + /* + * Iff nonsign is 0, it overflows into 0xFFFFFFFF, turning bit 31 + * into 1. Otherwise, bit 31 remains 0. The signed shift right by 31 + * broadcasts bit 31 into all bits of the zero_mask. Thus zero_mask == + * 0xFFFFFFFF if the half-precision number was zero (+0.0h or -0.0h) + * 0x00000000 otherwise + */ + const int32_t zero_mask = (int32_t)(nonsign - 1) >> 31; + /* + * 1. Shift nonsign left by renorm_shift to normalize it (if the input + * was denormal) + * 2. Shift nonsign right by 3 so the exponent (5 bits originally) + * becomes an 8-bit field and 10-bit mantissa shifts into the 10 high + * bits of the 23-bit mantissa of IEEE single-precision number. + * 3. Add 0x70 to the exponent (starting at bit 23) to compensate the + * different in exponent bias (0x7F for single-precision number less 0xF + * for half-precision number). + * 4. Subtract renorm_shift from the exponent (starting at bit 23) to + * account for renormalization. As renorm_shift is less than 0x70, this + * can be combined with step 3. + * 5. Binary OR with inf_nan_mask to turn the exponent into 0xFF if the + * input was NaN or infinity. + * 6. Binary ANDNOT with zero_mask to turn the mantissa and exponent + * into zero if the input was zero. + * 7. Combine with the sign of the input number. + */ + return sign | + ((((nonsign << renorm_shift >> 3) + ((0x70 - renorm_shift) << 23)) | + inf_nan_mask) & + ~zero_mask); +} + +#ifdef C10_X86_F16 +#undef C10_X86_F16 +#endif // C10_X86_F16 + +#if defined(__aarch64__) && !defined(__CUDACC__) +inline float16_t fp16_from_bits(uint16_t h) { + return c10::bit_cast(h); +} + +inline uint16_t fp16_to_bits(float16_t f) { + return c10::bit_cast(f); +} + +// According to https://godbolt.org/z/frExdbsWG it would translate to single +// fcvt s0, h0 +inline float native_fp16_to_fp32_value(uint16_t h) { + return static_cast(fp16_from_bits(h)); +} + +inline uint16_t native_fp16_from_fp32_value(float f) { + return fp16_to_bits(static_cast(f)); +} +#endif + +} // namespace detail + +//---------- below is copied from c10/util/Half-inl.h ----------------// +C10_CLANG_DIAGNOSTIC_PUSH() +#if C10_CLANG_HAS_WARNING("-Wimplicit-int-float-conversion") +C10_CLANG_DIAGNOSTIC_IGNORE("-Wimplicit-int-float-conversion") +#endif + +#if defined(__aarch64__) && !defined(__CUDACC__) +/// Constructors +inline Half::Half(float16_t value) : x(detail::fp16_to_bits(value)) {} +inline Half::operator float16_t() const { + return detail::fp16_from_bits(x); +} +#else + +inline C10_HOST_DEVICE Half::Half(float value) + : +#if defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__) + x(__half_as_short(__float2half(value))) +#elif defined(__SYCL_DEVICE_ONLY__) + x(c10::bit_cast(sycl::half(value))) +#elif (defined(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_AVX512)) && \ + !defined(__APPLE__) + x(at::vec::float2half_scalar(value)) +#else + x(detail::fp16_ieee_from_fp32_value(value)) +#endif +{ +} + +/// Implicit conversions + +inline C10_HOST_DEVICE Half::operator float() const { +#if defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__) + return __half2float(*reinterpret_cast(&x)); +#elif defined(__SYCL_DEVICE_ONLY__) + return float(c10::bit_cast(x)); +#elif (defined(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_AVX512)) && \ + !defined(__APPLE__) + return at::vec::half2float_scalar(x); +#elif defined(__aarch64__) && !defined(__CUDACC__) + return detail::native_fp16_to_fp32_value(x); +#else + return detail::fp16_ieee_to_fp32_value(x); +#endif +} + +#endif /* !defined(__aarch64__) || defined(__CUDACC__) \ + */ + +#if defined(__CUDACC__) || defined(__HIPCC__) +inline C10_HOST_DEVICE Half::Half(const __half& value) { + x = *reinterpret_cast(&value); +} +inline C10_HOST_DEVICE Half::operator __half() const { + return *reinterpret_cast(&x); +} +#endif + +#ifdef SYCL_LANGUAGE_VERSION +inline C10_HOST_DEVICE Half::Half(const sycl::half& value) { + x = *reinterpret_cast(&value); +} +inline C10_HOST_DEVICE Half::operator sycl::half() const { + return *reinterpret_cast(&x); +} +#endif + +// CUDA intrinsics + +#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 350)) || \ + (defined(__clang__) && defined(__CUDA__)) +inline __device__ Half __ldg(const Half* ptr) { + return __ldg(reinterpret_cast(ptr)); +} +#endif + +/// Arithmetic + +inline C10_HOST_DEVICE Half operator+(const Half& a, const Half& b) { + return static_cast(a) + static_cast(b); +} + +inline C10_HOST_DEVICE Half operator-(const Half& a, const Half& b) { + return static_cast(a) - static_cast(b); +} + +inline C10_HOST_DEVICE Half operator*(const Half& a, const Half& b) { + return static_cast(a) * static_cast(b); +} + +inline C10_HOST_DEVICE Half operator/(const Half& a, const Half& b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / static_cast(b); +} + +inline C10_HOST_DEVICE Half operator-(const Half& a) { +#if (defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530) || \ + defined(__HIP_DEVICE_COMPILE__) + return __hneg(a); +#elif defined(__SYCL_DEVICE_ONLY__) + return -c10::bit_cast(a); +#else + return -static_cast(a); +#endif +} + +inline C10_HOST_DEVICE Half& operator+=(Half& a, const Half& b) { + a = a + b; + return a; +} + +inline C10_HOST_DEVICE Half& operator-=(Half& a, const Half& b) { + a = a - b; + return a; +} + +inline C10_HOST_DEVICE Half& operator*=(Half& a, const Half& b) { + a = a * b; + return a; +} + +inline C10_HOST_DEVICE Half& operator/=(Half& a, const Half& b) { + a = a / b; + return a; +} + +/// Arithmetic with floats + +inline C10_HOST_DEVICE float operator+(Half a, float b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE float operator-(Half a, float b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE float operator*(Half a, float b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE float operator/(Half a, float b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE float operator+(float a, Half b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE float operator-(float a, Half b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE float operator*(float a, Half b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE float operator/(float a, Half b) + __ubsan_ignore_float_divide_by_zero__ { + return a / static_cast(b); +} + +inline C10_HOST_DEVICE float& operator+=(float& a, const Half& b) { + return a += static_cast(b); +} +inline C10_HOST_DEVICE float& operator-=(float& a, const Half& b) { + return a -= static_cast(b); +} +inline C10_HOST_DEVICE float& operator*=(float& a, const Half& b) { + return a *= static_cast(b); +} +inline C10_HOST_DEVICE float& operator/=(float& a, const Half& b) { + return a /= static_cast(b); +} + +/// Arithmetic with doubles + +inline C10_HOST_DEVICE double operator+(Half a, double b) { + return static_cast(a) + b; +} +inline C10_HOST_DEVICE double operator-(Half a, double b) { + return static_cast(a) - b; +} +inline C10_HOST_DEVICE double operator*(Half a, double b) { + return static_cast(a) * b; +} +inline C10_HOST_DEVICE double operator/(Half a, double b) + __ubsan_ignore_float_divide_by_zero__ { + return static_cast(a) / b; +} + +inline C10_HOST_DEVICE double operator+(double a, Half b) { + return a + static_cast(b); +} +inline C10_HOST_DEVICE double operator-(double a, Half b) { + return a - static_cast(b); +} +inline C10_HOST_DEVICE double operator*(double a, Half b) { + return a * static_cast(b); +} +inline C10_HOST_DEVICE double operator/(double a, Half b) + __ubsan_ignore_float_divide_by_zero__ { + return a / static_cast(b); +} + +/// Arithmetic with ints + +inline C10_HOST_DEVICE Half operator+(Half a, int b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE Half operator-(Half a, int b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE Half operator*(Half a, int b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE Half operator/(Half a, int b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE Half operator+(int a, Half b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE Half operator-(int a, Half b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE Half operator*(int a, Half b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE Half operator/(int a, Half b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return static_cast(a) / b; +} + +//// Arithmetic with int64_t + +inline C10_HOST_DEVICE Half operator+(Half a, int64_t b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return a + static_cast(b); +} +inline C10_HOST_DEVICE Half operator-(Half a, int64_t b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return a - static_cast(b); +} +inline C10_HOST_DEVICE Half operator*(Half a, int64_t b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return a * static_cast(b); +} +inline C10_HOST_DEVICE Half operator/(Half a, int64_t b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return a / static_cast(b); +} + +inline C10_HOST_DEVICE Half operator+(int64_t a, Half b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return static_cast(a) + b; +} +inline C10_HOST_DEVICE Half operator-(int64_t a, Half b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return static_cast(a) - b; +} +inline C10_HOST_DEVICE Half operator*(int64_t a, Half b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return static_cast(a) * b; +} +inline C10_HOST_DEVICE Half operator/(int64_t a, Half b) { + // NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions) + return static_cast(a) / b; +} + +/// NOTE: we do not define comparisons directly and instead rely on the implicit +/// conversion from c10::Half to float. + +C10_CLANG_DIAGNOSTIC_POP() + +} // namespace c10 + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly) + +using c10::Half; +using c10::operator+; +using c10::operator-; +using c10::operator*; +using c10::operator/; +using c10::operator+=; +using c10::operator-=; +using c10::operator*=; +using c10::operator/=; +using c10::operator<<; + +namespace detail { +#if defined(__aarch64__) && !defined(__CUDACC__) +using c10::detail::fp16_from_bits; +using c10::detail::fp16_to_bits; +using c10::detail::native_fp16_from_fp32_value; +using c10::detail::native_fp16_to_fp32_value; +#endif + +using c10::detail::fp16_ieee_from_fp32_value; +using c10::detail::fp16_ieee_to_fp32_bits; +using c10::detail::fp16_ieee_to_fp32_value; +} // namespace detail + +HIDDEN_NAMESPACE_END(torch, headeronly) + +namespace std { + +template <> +class numeric_limits { + public: + static constexpr bool is_specialized = true; + static constexpr bool is_signed = true; + static constexpr bool is_integer = false; + static constexpr bool is_exact = false; + static constexpr bool has_infinity = true; + static constexpr bool has_quiet_NaN = true; + static constexpr bool has_signaling_NaN = true; + static constexpr auto has_denorm = numeric_limits::has_denorm; + static constexpr auto has_denorm_loss = + numeric_limits::has_denorm_loss; + static constexpr auto round_style = numeric_limits::round_style; + static constexpr bool is_iec559 = true; + static constexpr bool is_bounded = true; + static constexpr bool is_modulo = false; + static constexpr int digits = 11; + static constexpr int digits10 = 3; + static constexpr int max_digits10 = 5; + static constexpr int radix = 2; + static constexpr int min_exponent = -13; + static constexpr int min_exponent10 = -4; + static constexpr int max_exponent = 16; + static constexpr int max_exponent10 = 4; + static constexpr auto traps = numeric_limits::traps; + static constexpr auto tinyness_before = + numeric_limits::tinyness_before; + static constexpr c10::Half min() { + return c10::Half(0x0400, c10::Half::from_bits()); + } + static constexpr c10::Half lowest() { + return c10::Half(0xFBFF, c10::Half::from_bits()); + } + static constexpr c10::Half max() { + return c10::Half(0x7BFF, c10::Half::from_bits()); + } + static constexpr c10::Half epsilon() { + return c10::Half(0x1400, c10::Half::from_bits()); + } + static constexpr c10::Half round_error() { + return c10::Half(0x3800, c10::Half::from_bits()); + } + static constexpr c10::Half infinity() { + return c10::Half(0x7C00, c10::Half::from_bits()); + } + static constexpr c10::Half quiet_NaN() { + return c10::Half(0x7E00, c10::Half::from_bits()); + } + static constexpr c10::Half signaling_NaN() { + return c10::Half(0x7D00, c10::Half::from_bits()); + } + static constexpr c10::Half denorm_min() { + return c10::Half(0x0001, c10::Half::from_bits()); + } +}; + +} // namespace std diff --git a/backends/pytorch/include/torch/headeronly/util/HeaderOnlyArrayRef.h b/backends/pytorch/include/torch/headeronly/util/HeaderOnlyArrayRef.h new file mode 100644 index 000000000..751ffef32 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/HeaderOnlyArrayRef.h @@ -0,0 +1,248 @@ +#pragma once + +#include +#include + +#include +#include +#include +#include +#include +#include +#include +#include + +namespace c10 { + +/// HeaderOnlyArrayRef - A subset of ArrayRef that is implemented only +/// in headers. This will be a base class from which ArrayRef inherits, so that +/// we can keep much of the implementation shared. +/// +/// [HeaderOnlyArrayRef vs ArrayRef note] +/// As HeaderOnlyArrayRef is a subset of ArrayRef, it has slightly less +/// functionality than ArrayRef. We document the minor differences below: +/// 1. ArrayRef has an extra convenience constructor for SmallVector. +/// 2. ArrayRef uses TORCH_CHECK. HeaderOnlyArrayRef uses header-only +/// STD_TORCH_CHECK, which will output a std::runtime_error vs a +/// c10::Error. Consequently, you should use ArrayRef when possible +/// and HeaderOnlyArrayRef only when necessary to support headeronly code. +/// In all other aspects, HeaderOnlyArrayRef is identical to ArrayRef, with the +/// positive benefit of being header-only and thus independent of libtorch.so. +template +class HeaderOnlyArrayRef { + public: + using iterator = const T*; + using const_iterator = const T*; + using size_type = size_t; + using value_type = T; + + using reverse_iterator = std::reverse_iterator; + + protected: + /// The start of the array, in an external buffer. + const T* Data; + + /// The number of elements. + size_type Length; + + public: + /// @name Constructors + /// @{ + + /// Construct an empty HeaderOnlyArrayRef. + /* implicit */ constexpr HeaderOnlyArrayRef() : Data(nullptr), Length(0) {} + + /// Construct a HeaderOnlyArrayRef from a single element. + // TODO Make this explicit + constexpr HeaderOnlyArrayRef(const T& OneElt) : Data(&OneElt), Length(1) {} + + /// Construct a HeaderOnlyArrayRef from a pointer and length. + constexpr HeaderOnlyArrayRef(const T* data, size_t length) + : Data(data), Length(length) {} + + /// Construct a HeaderOnlyArrayRef from a range. + constexpr HeaderOnlyArrayRef(const T* begin, const T* end) + : Data(begin), Length(end - begin) {} + + template < + typename Container, + typename U = decltype(std::declval().data()), + typename = std::enable_if_t< + (std::is_same_v || std::is_same_v)>> + /* implicit */ HeaderOnlyArrayRef(const Container& container) + : Data(container.data()), Length(container.size()) {} + + /// Construct a HeaderOnlyArrayRef from a std::vector. + // The enable_if stuff here makes sure that this isn't used for + // std::vector, because ArrayRef can't work on a std::vector + // bitfield. + template + /* implicit */ HeaderOnlyArrayRef(const std::vector& Vec) + : Data(Vec.data()), Length(Vec.size()) { + static_assert( + !std::is_same_v, + "HeaderOnlyArrayRef cannot be constructed from a std::vector bitfield."); + } + + /// Construct a HeaderOnlyArrayRef from a std::array + template + /* implicit */ constexpr HeaderOnlyArrayRef(const std::array& Arr) + : Data(Arr.data()), Length(N) {} + + /// Construct a HeaderOnlyArrayRef from a C array. + template + // NOLINTNEXTLINE(*c-arrays*) + /* implicit */ constexpr HeaderOnlyArrayRef(const T (&Arr)[N]) + : Data(Arr), Length(N) {} + + /// Construct a HeaderOnlyArrayRef from a std::initializer_list. + /* implicit */ constexpr HeaderOnlyArrayRef( + const std::initializer_list& Vec) + : Data( + std::begin(Vec) == std::end(Vec) ? static_cast(nullptr) + : std::begin(Vec)), + Length(Vec.size()) {} + + /// @} + /// @name Simple Operations + /// @{ + + constexpr iterator begin() const { + return this->Data; + } + constexpr iterator end() const { + return this->Data + this->Length; + } + + // These are actually the same as iterator, since ArrayRef only + // gives you const iterators. + constexpr const_iterator cbegin() const { + return this->Data; + } + constexpr const_iterator cend() const { + return this->Data + this->Length; + } + + constexpr reverse_iterator rbegin() const { + return reverse_iterator(end()); + } + constexpr reverse_iterator rend() const { + return reverse_iterator(begin()); + } + + /// Check if all elements in the array satisfy the given expression + constexpr bool allMatch(const std::function& pred) const { + return std::all_of(cbegin(), cend(), pred); + } + + /// empty - Check if the array is empty. + constexpr bool empty() const { + return this->Length == 0; + } + + constexpr const T* data() const { + return this->Data; + } + + /// size - Get the array size. + constexpr size_t size() const { + return this->Length; + } + + /// front - Get the first element. + constexpr const T& front() const { + STD_TORCH_CHECK( + !this->empty(), + "HeaderOnlyArrayRef: attempted to access front() of empty list"); + return this->Data[0]; + } + + /// back - Get the last element. + constexpr const T& back() const { + STD_TORCH_CHECK( + !this->empty(), + "HeaderOnlyArrayRef: attempted to access back() of empty list"); + return this->Data[this->Length - 1]; + } + + /// equals - Check for element-wise equality. + constexpr bool equals(HeaderOnlyArrayRef RHS) const { + return this->Length == RHS.Length && + std::equal(begin(), end(), RHS.begin()); + } + + /// slice(n, m) - Take M elements of the array starting at element N + constexpr HeaderOnlyArrayRef slice(size_t N, size_t M) const { + STD_TORCH_CHECK( + N + M <= this->size(), + "HeaderOnlyArrayRef: invalid slice, N = ", + N, + "; M = ", + M, + "; size = ", + this->size()); + return HeaderOnlyArrayRef(this->data() + N, M); + } + + /// slice(n) - Chop off the first N elements of the array. + constexpr HeaderOnlyArrayRef slice(size_t N) const { + STD_TORCH_CHECK( + N <= this->size(), + "HeaderOnlyArrayRef: invalid slice, N = ", + N, + "; size = ", + this->size()); + return slice(N, this->size() - N); + } + + /// @} + /// @name Operator Overloads + /// @{ + constexpr const T& operator[](size_t Index) const { + return this->Data[Index]; + } + + /// Vector compatibility + constexpr const T& at(size_t Index) const { + STD_TORCH_CHECK( + Index < this->Length, + "HeaderOnlyArrayRef: invalid index Index = ", + Index, + "; Length = ", + this->Length); + return this->Data[Index]; + } + + /// Disallow accidental assignment from a temporary. + /// + /// The declaration here is extra complicated so that "arrayRef = {}" + /// continues to select the move assignment operator. + template + std::enable_if_t, HeaderOnlyArrayRef>& operator=( + // NOLINTNEXTLINE(cppcoreguidelines-missing-std-forward) + U&& Temporary) = delete; + + /// Disallow accidental assignment from a temporary. + /// + /// The declaration here is extra complicated so that "arrayRef = {}" + /// continues to select the move assignment operator. + template + std::enable_if_t, HeaderOnlyArrayRef>& operator=( + std::initializer_list) = delete; + + /// @} + /// @name Expensive Operations + /// @{ + std::vector vec() const { + return std::vector(this->Data, this->Data + this->Length); + } + + /// @} +}; + +} // namespace c10 + +namespace torch::headeronly { +using c10::HeaderOnlyArrayRef; +using IntHeaderOnlyArrayRef = HeaderOnlyArrayRef; +} // namespace torch::headeronly diff --git a/backends/pytorch/include/torch/headeronly/util/Metaprogramming.h b/backends/pytorch/include/torch/headeronly/util/Metaprogramming.h new file mode 100644 index 000000000..2589f338d --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/Metaprogramming.h @@ -0,0 +1,237 @@ +#pragma once + +#include +#include +#include + +namespace c10::guts { + +/** + * Access information about result type or arguments from a function type. + * Example: + * using A = function_traits::return_type // A == int + * using A = function_traits::parameter_types::tuple_type + * // A == tuple + */ +template +struct function_traits { + static_assert( + !std::is_same_v, + "In function_traits, Func must be a plain function type."); +}; +template +struct function_traits { + using func_type = Result(Args...); + using return_type = Result; + using parameter_types = typelist::typelist; + static constexpr auto number_of_parameters = sizeof...(Args); +}; + +/** + * infer_function_traits: creates a `function_traits` type for a simple + * function (pointer) or functor (lambda/struct). Currently does not support + * class methods. + */ + +template +struct infer_function_traits { + using type = function_traits< + c10::guts::detail::strip_class_t>; +}; + +template +struct infer_function_traits { + using type = function_traits; +}; + +template +struct infer_function_traits { + using type = function_traits; +}; + +template +using infer_function_traits_t = typename infer_function_traits::type; + +/** + * make_function_traits: creates a `function_traits` type given a Return type + * and a typelist of Argument types + * + * Example: + * bool f(int, int); + * + * infer_function_traits_t == make_function_traits_t> + */ +template +struct make_function_traits { + static_assert( + false_t::value, + "In guts::make_function_traits, the ArgList argument must be typelist<...>."); +}; + +template +struct make_function_traits> { + using type = function_traits; +}; + +template +using make_function_traits_t = + typename make_function_traits::type; + +/** + * make_offset_index_sequence + * Like make_index_sequence, but starting from Start instead of 0. + * + * Example: + * make_offset_index_sequence<10, 3> == std::index_sequence<10, 11, 12> + */ +template +struct make_offset_index_sequence_impl + : make_offset_index_sequence_impl { + static_assert( + static_cast(Start) >= 0, + "make_offset_index_sequence: Start < 0"); + static_assert(static_cast(N) >= 0, "make_offset_index_sequence: N < 0"); +}; + +template +struct make_offset_index_sequence_impl { + typedef std::index_sequence type; +}; + +template +using make_offset_index_sequence = + typename make_offset_index_sequence_impl::type; + +/** + * Use tuple_elements to extract a position-indexed subset of elements + * from the argument tuple into a result tuple. + * + * Example: + * std::tuple t = std::make_tuple(0, "HEY", 2.0); + * std::tuple result = tuple_elements(t, std::index_sequence<0, + * 2>()); + */ +template +constexpr auto tuple_elements(Tuple t, std::index_sequence /*unused*/) { + return std::tuple...>(std::get(t)...); +} + +/** + * Use tuple_take to extract the first or last n elements from the argument + * tuple into a result tuple. + * + * Example: + * std::tuple t = std::make_tuple(0, "HEY", 2.0); + * std::tuple first_two = tuple_take(t); + * std::tuple last_two = tuple_take(t); + */ +template +struct TupleTake {}; + +template +struct TupleTake= 0, void>> { + static auto call(Tuple t) { + constexpr size_t size = std::tuple_size(); + static_assert(N <= size, "tuple_take: N > size"); + return tuple_elements(t, std::make_index_sequence{}); + } +}; + +template + struct TupleTake < Tuple, + N, std::enable_if_t> { + static auto call(Tuple t) { + constexpr size_t size = std::tuple_size(); + static_assert(-N <= size, "tuple_take: -N > size"); + return tuple_elements(t, make_offset_index_sequence{}); + } +}; + +template +auto tuple_take(Tuple t) { + return TupleTake::call(t); +} + +/** + * Use tuple_slice to extract a contiguous subtuple from the argument. + * + * Example: + * std::tuple t = std::make_tuple(0, + * "HEY", 2.0, false); std::tuple middle_two = + * tuple_slice(t); + */ +template +constexpr auto tuple_slice(Tuple t) { + constexpr size_t size = std::tuple_size(); + static_assert(Start + N <= size, "tuple_slice: Start + N > size"); + return tuple_elements(t, make_offset_index_sequence{}); +} + +/** + * Use tuple_map to run a mapping function over a tuple to get a new tuple. + * + * Example 1: + * auto result = tuple_map(std::tuple(3, 4, 5), [] + * (int32_t a) -> int16_t {return a+1;}); + * // result == std::tuple(4, 5, 6) + * + * Example 2: + * struct Mapper { + * std::string operator()(int32_t a) const { + * return std::to_string(a); + * } + * int64_t operator()(const std::string& a) const { + * return atoi(a.c_str()); + * } + * }; + * auto result = tuple_map(std::tuple(3, "4"), + * Mapper()); + * // result == std::tuple("3", 4) + * + * Example 3: + * struct A final { + * int32_t func() { + * return 5; + * } + * }; + * struct B final { + * std::string func() { + * return "5"; + * } + * }; + * auto result = tuple_map(std::make_tuple(A(), B()), [] (auto a) { return + * a.func(); }); + * // result == std::tuple(5, "5"); + */ +namespace detail { +template +auto tuple_map( + // NOLINTNEXTLINE(cppcoreguidelines-rvalue-reference-param-not-moved) + std::tuple&& tuple, + const Mapper& mapper, + std::index_sequence /*unused*/) { + return std::tuple(std::get( + tuple))))...>(mapper(std::forward(std::get(tuple)))...); +} +} // namespace detail + +template +auto tuple_map(std::tuple&& tuple, const Mapper& mapper) { + return detail::tuple_map( + std::move(tuple), mapper, std::index_sequence_for()); +} + +} // namespace c10::guts + +HIDDEN_NAMESPACE_BEGIN(torch, headeronly, guts) + +using c10::guts::function_traits; +using c10::guts::infer_function_traits_t; +using c10::guts::make_function_traits_t; +using c10::guts::tuple_elements; +using c10::guts::tuple_map; +using c10::guts::tuple_slice; +using c10::guts::tuple_take; + +HIDDEN_NAMESPACE_END(torch, headeronly, guts) diff --git a/backends/pytorch/include/torch/headeronly/util/TypeList.h b/backends/pytorch/include/torch/headeronly/util/TypeList.h new file mode 100644 index 000000000..cd81f0cc1 --- /dev/null +++ b/backends/pytorch/include/torch/headeronly/util/TypeList.h @@ -0,0 +1,548 @@ +#pragma once + +#include +#include +#include +#include +#include +#include +#include + +namespace c10::guts { + +template +struct false_t : std::false_type {}; +template