Accuracy: Round to nearest, ties to even for generic_float - #5193
Open
CharlieL7 wants to merge 6 commits into
Open
Accuracy: Round to nearest, ties to even for generic_float#5193CharlieL7 wants to merge 6 commits into
generic_float#5193CharlieL7 wants to merge 6 commits into
Conversation
Contributor
There was a problem hiding this comment.
Pull request overview
Updates MIGraphX’s generic_float (notably migraphx::half and migraphx::bf16) float32 host conversion to use round-to-nearest, ties-to-even (RNE), aligning with typical hardware conversion behavior and improving numeric agreement with allclose()-style comparisons.
Changes:
- Implemented RNE rounding in
generic_floatfloat32→target conversion, including correct carry/overflow behavior (to infinity when appropriate). - Updated unit tests and ONNX verify gold data to reflect RNE results and overflow-to-inf behavior.
- Adjusted
rewrite_resize_testcomparisons to use an elementwise tolerance for narrow types to avoid 1-ULP mismatches.
Reviewed changes
Copilot reviewed 8 out of 8 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
src/include/migraphx/generic_float.hpp |
Implements RNE rounding logic for float32 conversions used by half/bf16. |
test/half.cpp |
Adds/updates tests to pin RNE rounding behavior and overflow/subnormal corner cases for half. |
test/bf16.cpp |
Adds/updates tests to pin RNE rounding behavior and overflow/subnormal corner cases for bf16. |
test/rewrite_resize_test.cpp |
Uses elementwise allclose tolerance for narrow output types where interpolation precision differs. |
test/onnx/verify/gelu_default_test.cpp |
Updates expected outputs to match ORT CPU EP rounding behavior. |
test/onnx/verify/gelu_default_bf16_test.cpp |
Refreshes bf16 expectations and comments for the verify test. |
test/onnx/verify/group_query_attention_test.cpp |
Updates expected values/comments to reflect more precise ORT-sourced gold data. |
CHANGELOG.md |
Documents the behavior change for float32→half/bf16 host conversion. |
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
| * Allow for 1 arg slicing over a dynamic dimension. (#5015) | ||
| * Route convolutions and dot operations through rocMLIR when MIOpen or GEMM libraries are disabled at build time (#5059). | ||
| * Rejected symbolic input shapes in the multi-input `slice` calls, and pointed both symbolic `slice` errors at `dyn_slice` (#5088). | ||
| * Changed host conversion of float32 to `migraphx::half` and `migraphx::bf16` to round to nearest with ties to even instead of truncating, matching the hardware conversions these types emulate. Values above the target type's overflow threshold, including `std::numeric_limits<float>::max()`, now convert to infinity rather than to the largest finite value, so fp16 and bf16 model output can shift by up to one ULP (#5193). |
pfultz2
reviewed
Aug 26, 2026
| constexpr const int diff = float32_parts::exponent_bias() - exponent_bias(); | ||
| constexpr const int drop = int(float32_parts::mantissa_width() - MantissaSize); | ||
|
|
||
| if(f.exponent != 0 and f.exponent != float32_parts::max_exponent() and |
Collaborator
There was a problem hiding this comment.
So fp32 and bf16 uses round_normal. Is this an optimization? So are we doing the same RNE algorithm and just taking advantage that it has the same exponent as float to do this automatically?
Regressions detected 🔴 * No develop baseline was found for this PR's branch point; compared against the latest available develop run instead. |
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
generic_floatwhen trying to compare withallclose().Technical Details
generic_floatso that the reference values are correctly rounded when compared to the target results.Changelog Category
Add a
CHANGELOG.mdentry for any option other thanNot ApplicableFollow the LLVM AI Tool Use Policy for contributions using AI.