search

intel / intel-xpu-backend-for-triton

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

Fix coalescing pass #2760

Closed etiotto closed 2 days ago

etiotto commented 3 days ago

Fix Intel coalescing pass for cases where the result of a SCF loop (containing a coalescable block ptr load) is used by an operation with operands that do not have block ptr type (e.g. tt.reduce)

etiotto commented 3 days ago

Can you provide a PR description?

Added a description.

etiotto commented 2 days ago

This PR can resolve this particular issue, but there can still be issues e.g. the operation uses part of the result that should not be propagated and it has blocked pointer type, then the pass would incorrectly propagated the type. Please create an issue to handle that.

Tracked using https://github.com/intel/intel-xpu-backend-for-triton/issues/2762