[PATCH] D75475: AMDGPU: Fix computation for getOccupancyWithLocalMemSize

Matt Arsenault via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Mon Mar 2 12:29:29 PST 2020


arsenm marked an inline comment as done.
arsenm added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp:361
+  // FIXME: Needs to be a multiple of the group size?
+  //MaxWaves = MaxGroupNumWaves * (MaxWaves / MaxGroupNumWaves);
+
----------------
rampitec wrote:
> Yes, we cannot split a group. I think it is better to uncomment this.
The problem is this assumes the very worst case. This would hit 0 with the default / maximum group size, and need clamping to 1. This breaks every test. I was thinking I would try to apply this after changing this to report a range of occupancies


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D75475/new/

https://reviews.llvm.org/D75475





More information about the llvm-commits mailing list