[PATCH] D112373: [AMDGPU] Use max waves for scheduler's initial occupancy target

Austin Kerbow via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Sat Oct 23 17:19:37 PDT 2021


kerbowa added a comment.

I wanted to clarify some things that I saw with the test schedule-regpressure-limit3.ll. Before this change, the scheduler is actually targeting max occupancy in this test when only one wave was requested. This is the main motivation for this patch. However what we actually see is higher register usage, despite the scheduler trying to limit VGPR RP at basically every step. I found that the reason for this was that disabling amdgpu-aa so that the flat stores may alias the LDS loads was greatly restricting the ability of the scheduler to reduce RP. So I've removed -enable-amdgpu-aa from this test.
This test is also a good example of the machine schedulers' seeming shortcomings when trying to maximize ILP. Even when running -misched=ilpmax we are getting worse results than some naive ILP heuristics.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D112373/new/

https://reviews.llvm.org/D112373



More information about the llvm-commits mailing list