intel / intel-xpu-backend-for-triton

OpenAI Triton backend for Intel® GPUs
MIT License
131 stars 39 forks source link

[Productize GEMM Performance] Features #1450

Closed vlad-penkin closed 1 month ago

vlad-penkin commented 3 months ago

This is the GEMM Performance features productization umbrella ticket. Before converting this ticket umbrella ticket, please:

cc: @chengjunlu, @quintinwang5, @Dewei-Wang-sh, @etiotto, @whitneywhtsang, @tdeng5, @LiyangLingIntel

vlad-penkin commented 3 months ago

This ticket is related to:

Dewei-Wang-sh commented 3 months ago

feature needed to productize gemm with block pointer

whitneywhtsang commented 2 months ago

Still need to enable split barrier.

vlad-penkin commented 1 month ago

@whitneywhtsang what is the ticket # for enabling split barrier?

whitneywhtsang commented 1 month ago

@whitneywhtsang what is the ticket # for enabling split barrier?

There was a functional failure when enabling split barrier, it is fixed in https://github.com/intel/intel-xpu-backend-for-triton/pull/1708. As commented in https://github.com/intel/intel-xpu-backend-for-triton/pull/1708#issuecomment-2253097323, there is no performance gain in GEMM geomean, so it is not enabled.