search

intel / intel-xpu-backend-for-triton

OpenAI Triton backend for Intel® GPUs
MIT License
142 stars 43 forks source link

spirv.UGreaterThan does not support 'i1' #235

Closed CuiYifeng closed 9 months ago

CuiYifeng commented 10 months ago

Issue Description

Intel-xpu-backend-for-triton > operator does not support i1 but triton for CUDA can.

Code Example

tmp83 = tmp77 != tmp77
tmp84 = tmp80 != tmp80
tmp85 = tmp83 > tmp84

Please find example in attachment. maml_cuda.txt is triton code generated with Pytorch CUDA. The above triton code segment also appears in it. maml_xpu_to_cuda.txt is triton code generated with XPU backend. It works with Pytorch CUDA after replacing the header.

System Info

aregm commented 9 months ago

This is related to SPIR-V lowering path. Can you please recheck after migration to the llvm-target.

tdeng5 commented 9 months ago

Will double check if this issue happens in LLVM path when E2E could run.

vlad-penkin commented 9 months ago

The llvm branch has reached parity with the spirv branch in terms of the UT pass rate. Closing the issue.