search

intel / intel-xpu-backend-for-triton

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

[Upstream] Try to upstream the supporting of the nested layout slice->dot->mma without LinearLayout #1843

Closed chengjunlu closed 3 weeks ago

chengjunlu commented 3 months ago

The changes to the common TritonGPU dialect in the PR https://github.com/intel/intel-xpu-backend-for-triton/pull/1782

chengjunlu commented 2 months ago

Since we have supported the conversion to LinearLayout of the DPAS layout, we are not going to upstream the changes in legacy layout interface which is going to be rejected most likely.

The proper way is to support the conversion from dot layout to LinearLayout in upstream which is not supported yet.

@etiotto , @whitneywhtsang , Should we upstream the code to Triton or should we wait upstream to finish that?

whitneywhtsang commented 2 months ago

Since we already have the change, we can attend to upstream it (assuming the cost is low) with a comment in the description that the change is not needed when dot layout to linear layout conversion is supported, and ask when do they plan to support it. WDYT @etiotto @chengjunlu

chengjunlu commented 2 months ago

Since we already have the change, we can attend to upstream it (assuming the cost is low) with a comment in the description that the change is not needed when dot layout to linear layout conversion is supported, and ask when do they plan to support it. WDYT @etiotto @chengjunlu

Make sense. I will create the PR to upstream the changes and request for comments.

chengjunlu commented 2 months ago

Created an RFC to the public Triton: https://github.com/triton-lang/triton/pull/4598

chengjunlu commented 2 months ago

The PR has been rejected.

This is blocked. Need to wait the upstream to support the DotOp layout to LL conversion.

chengjunlu commented 3 weeks ago

The DotOp layout to LL has been supported in upstream Triton. And the code has been merged to Intel XPU downstream repo. We can close this JIRA because there is no need to support the nested layout in legacy path any more.

whitneywhtsang commented 3 weeks ago

Can we remove the changes we made in PR https://github.com/intel/intel-xpu-backend-for-triton/pull/1782 then?

chengjunlu commented 3 weeks ago

Can we remove the changes we made in PR #1782 then?

We can track the clean up task in this issue https://github.com/intel/intel-xpu-backend-for-triton/issues/2030.

We can clean up our local logic after replacing the intel::emitIndices when we fully reuses this file include/triton/Conversion/TritonGPUToLLVM/Utility.h