Note

This repo is only for learning and practice purpose.

My Kernels

A self-own kernel library with CUDA and PyTorch integration for learning purpose.

Included

• CPU implementation
• CUDA implementation
• C++ API
• PyTorch registration
• Python Wrapper
• Tests
  • Python correctness/dispatch/perf smoke
  • GoogleTest-based C++ tests discovered by CTest
• Benchmark against SOTA kernels
  • Python helpers
  • C++ executable benchmarks

Repo structure

my-kernels/
├── README.md
├── pyproject.toml
├── CMakeLists.txt
├── include/
│   └── my_kernels/
├── benchmarks/
│   ├── cpp/
│   │   ├── add/
│   │   └── support/
│   └── python/
│       └── utils/
├── src/
│   ├── pytorch/
│   │   └── register_ops.cpp
│   └── ops/
├── tests/
│   ├── cpp/
│   │   ├── add/
│   │   └── support/
│   └── python/
│       ├── correctness/
│       ├── dispatch/
│       ├── perf_smoke/
│       └── reference/
├── python/
│   └── my_kernels/
│       ├── __init__.py
│       ├── _load_lib.py
│       └── ops.py
└── scripts/

Build and run

just build python_bin=.venv/bin/python
just test
just test-fast
just test-python
just test-cpp
./build/benchmarks/cpp/bench_add --device cpu --size 1048576 --iters 100

Use the Python interpreter that already has torch installed when running just build. Use just build-cpp when you only want the native library, tests, and benchmarks without the Python extension.

Best-practice layout

• Keep implementation under src/ops/<op>/ and public declarations under include/my_kernels/.
• Keep Python validation under tests/python/<category>/ and import benchmark helpers from benchmarks/python/.
• Keep native validation under tests/cpp/<op>/ and register each GoogleTest executable in tests/cpp/CMakeLists.txt.
• Keep native benchmarking under benchmarks/cpp/<op>/ and report timings instead of asserting hard performance thresholds.
• Expose one public C++ entry point per op, then test that API in C++ and the registered op in Python.

Test orchestration

• Use ctest as the top-level runner for both C++ and Python suites.
• Filter by label:
  • ctest --test-dir build -L cpp
  • ctest --test-dir build -L python
  • ctest --test-dir build -LE perf
• Python tests run against the built extension staged under build/python/, with JIT fallback disabled so packaging regressions surface in CI.

Adding a new kernel

For a new op such as mul, mirror the same slices:

1. Add include/my_kernels/mul.h plus src/ops/mul/.
2. Add the C++ dispatch entry point that routes to CPU/CUDA implementations.
3. Add Python correctness and dispatch tests under tests/python/.
4. Add native C++ tests under tests/cpp/mul/.
5. Add a native benchmark under benchmarks/cpp/mul/.