[llvm] AMDGPU: Correct cycle counts for f64 mfma on gfx940 (PR #83782)

Matt Arsenault via llvm-commits llvm-commits at lists.llvm.org
Tue Mar 5 09:23:06 PST 2024


arsenm wrote:

>  The code related to 8 pass DGEMM now needs to be moved from 'case 16' to 'case 8' in the 
> `GCNHazardRecognizer::checkMAIVALUHazards`.

This one avoids the test changes

> `GCNHazardRecognizer::checkMAIHazards90A` 

This one seems to be unnecessary since it's special casing the MFMA_F64 opcodes before that switch 


https://github.com/llvm/llvm-project/pull/83782


More information about the llvm-commits mailing list