[WMMA][Dot] Support WMMA layout in TritonAMDGPUAccelerateMatmulPass

ROCm / triton

Development repository for the Triton language and compiler

MIT License

83 stars 27 forks source link

[WMMA][Dot] Support WMMA layout in TritonAMDGPUAccelerateMatmulPass #423

Closed joviliast closed 8 months ago

joviliast commented 8 months ago

-Introduce WmmaEncodingAttr for WMMA output (disabled for now) -Introduce BlockedToWMMA rewrite pattern in TritonAMDGPUAccelerateMatmulPass

Signed-off-by: joviliast iveselov.nn@gmail.com

joviliast commented 8 months ago

Thank you @joviliast for refactor the arch version stuff. Now it looks more organized and easy to extend in the future.

Approved. But I'd suggest remove enableWmmaTransform since it is not used.

Sure! I put it in a separate patch. I'll just revert it after enabling