[llvm-branch-commits] [llvm] [AMDGPU] Hint the MFMA destination into the accumulator register (PR #228533)

[Dmitry Sidorov via llvm-branch-commits]

MrSidims wrote:

Tested on gluon tutorial sample v9 on gfx950 against the current main including a branch, where I've assembled few other fixes in opened PRs. The results:
 | | main  | main + this patch   | assembled branch base | assembled branch + this patch |
  |-------------------|-----------|-----------|---------------------------------|---|
  | v9   | 243 TFlops  | 423 TFlops  | 561 TFlops                        |   798 TFlops | 

https://github.com/llvm/llvm-project/pull/228533