Refactor kernel writing in TensileCreateLibrary - Githubissues

ROCm / Tensile

Stretching GPU performance for GEMMs and tensor contractions.

MIT License

208 stars 144 forks source link

Refactor kernel writing in TensileCreateLibrary #1960

Closed bstefanuk closed 1 month ago

bstefanuk commented 1 month ago

Objectives

The objective of this PR is to refactor the writeKernels function within TensileCreateLibrary into multiple testable and profileable functions. This refactor enhances performance assessment capabilities and ensures future changes maintain or improve current functionality without regressions.

Outcomes

Code blocks in writeKernels are refactored into 6 new functions.
Unit tests are implemented for each new function.
Inline documentation is added for writeKernels and each new function.
Type-hinting is added for the parameter and return types of each function.

Testing

Functional outputs of TensileCreateLibrary were evaluated to ensure they remain unchanged when building rocBLAS, e.g., TensileManifest.txt.
All new unit tests are integrated into CI pipelines to ensure they pass consistently.

bstefanuk commented 1 month ago

PR closed because it was broken out into smaller components.