| --- |
| license: apache-2.0 |
| tags: |
| - cuda |
| - gpu-optimization |
| - kernels |
| - gemm |
| - hpc |
| --- |
| |
| # PyC CUDA kernel lab |
|
|
| This repository documents 19 CUDA kernel-lab entries from PyC. It is a source |
| and evidence release, not a compiled binary distribution and not a claim that |
| all entries are wired into PyC runtime dispatch. |
|
|
| ## Contents |
|
|
| - `kernels/prototypes/`: standalone CUDA prototype sources. |
| - `manifests/lab_kernels.json`: the 19-entry lab catalog, including build/run commands. |
| - `manifests/registry_kernels.json`: the catalog mirrored into the registry release. |
| - `PERFORMANCE_SUMMARY.md`: selected H100 campaign measurements and the optimization progression. |
|
|
| ## Optimization themes |
|
|
| The progression covers shared-memory tiling, WMMA Tensor Core execution, BF16 |
| versus FP16, `cp.async` double buffering, CTA shape, K-stage depth, warp work |
| assignment, and cuBLASLt as a hardware-library ceiling/control. |
|
|
| Performance numbers are campaign-specific measurements. They should be read |
| with the GPU, CUDA toolchain, matrix shape, correctness mode, and timing method |
| from the accompanying evidence; they are not universal benchmarks. |
|
|
| ## Reproduce |
|
|
| The commands in `manifests/kernels.json` use `{nvcc}`, `{source}`, and |
| `{build_dir}` placeholders. Replace them with a CUDA 12.x toolchain, a suitable |
| Hopper or Ada GPU, and a local build directory before running. |
|
|
