[PATCH] D64935: [AMDGPU] Add llvm.amdgcn.softwqm intrinsic

[Carl Ritson via Phabricator via llvm-commits]

critson added a comment.

In D64935#1593333 <https://reviews.llvm.org/D64935#1593333>, @nhaehnle wrote:

> Have you checked that this actually fixes the reported CTS failure?


Yes, with the associated (minimal) frontend changes this fixes the CTS failure.

In D64935#1593333 <https://reviews.llvm.org/D64935#1593333>, @nhaehnle wrote:

> The derivative calculation enables WQM, but then we may leave WQM again for the subgroup operations which is unexpected (since helper lanes are expected to participate). So softwqm needs to seed WQM requirements, but only if there is at least one hard wqm requirement in the shader.


While my understanding of "seed requirements" means "for the whole shader", this code does what you expect.
If there are any hard WQM requirements for the shader, then all softwqm instructions (and their dependencies) are marked WQM.


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D64935/new/

https://reviews.llvm.org/D64935