RADV ACO SMEM Patches Land - Taking ACO To Feature Parity With AMDGPU LLVM

As of today in Mesa 20.2-devel Git, the Radeon Vulkan driver (RADV) with the ACO back-end is now effectively at feature-parity to the default AMDGPU LLVM shader compiler back-end.

Earlier this month we reported on the matter that RADV was looking at switching to ACO by default as its shader compiler back-end in place of the common AMDGPU LLVM back-end. But for that to happen they first needed to hit feature parity.

Marked as blockers were FP16 features that since merged and also handling some SMEM bits for 8/16-bit accesses where it's safe. That big merge request was pulled into Git master today for Mesa 20.2 with the ACO and relevant SPIR-V/NIR changes.

In other words, the previously expressed blockers for switching to RADV ACO by default are now merged.

So it's quite possible now in the very near future we'll see RADV using ACO by default for offering quicker shader compile times and in turn quicker game load times while for many games offering faster Vulkan performance than the AMDGPU LLVM code path. There is talk of Valve working on ACO for RadeonSI but at this point the change is just in the context of RADV.

There is about one month to the Mesa 20.2 feature freeze so it's quite likely for this next release due out at the end of August we'll be seeing the RADV ACO default. Stay tuned for the likely change-over and subsequent benchmarks.