search

intel / intel-xpu-backend-for-triton

OpenAI Triton backend for Intel® GPUs
MIT License
143 stars 44 forks source link

Disable LLVM post processing #2626

Closed whitneywhtsang closed 2 weeks ago

whitneywhtsang commented 2 weeks ago

IGCVectorizer of driver agama 1032 has improved. This PR disables the LLVM post processing Triton performed by default, which includes SLPVectorizer.

By disabling LLVM post processing, the performance impact to the 3 key workloads (FA, GEMM, Softmax) are all positive. For example, GEMM out of box has improved by 16%.  Screenshot 2024-11-04 184048

CI: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/11664443045, https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/11674882693

chengjunlu commented 2 weeks ago

It is good news that we can remove the SLPVectorizer. It is observed that the SLPVectorizer at Triton side may change the order of the operation which make the IR to be more complicate to IGC for optimize.

Changes are LGTM.