AMD Rebases Their OpenMP For Radeon GPUs Against LLVM 11

[jrch2k8]

Honestly the best path for AMD compute is Clover Over NIR(not the original one), So far CL1.1 is solid and is already way faster then in luxmark than BLOB CL on Polaris, it works OOB, comes with mesa and once all the compute bits land in NIR should be extremely trivial(outside few niche things here and there) to actually have an CUDA state tracker that reuses Clover works.

If AMD want compute their focus should be on iron out HMM in the AMDGPU kernel module ASAP and maybe put some people to help Karol dealing with NIR missing bits

[boxie]

Or make OneAPI ROCm compliant

either way - or option 3 - you both work together to get the most out of accelerators using all the open shinies