LLVM 11 Merges AMD Radeon GCN Offloading For OpenMP

While AMD has been working on AOMP for Radeon OpenMP offloading as their downstream of the LLVM/Clang compiler suited for GPU compute offloading to their hardware, at least some of that work is beginning to appear back in upstream LLVM.

Merged today into LLVM 11 Git is support for OpenMP offloading for the AMD GCN architecture, including setting of LLVM's CUDA mode. This part 1 patch is just a few dozen lines of code thanks to the AMDGPU back-end long being in upstream LLVM as well as the necessary infrastructure already being there for OpenMP device offload. Great to see this happening and hopefully more of these Radeon GPU compute changes will be hitting the LLVM tree shortly.

This patch does come via AMD's Radeon Open eCosystem team. Hopefully we'll see more of the AOMP changes working their way back upstream in time for the LLVM 11.0 release due out around September.

The recently released GCC 10 compiler also has initial Radeon OpenMP/OpenACC offloading albeit there the functionality is quite restricted and supports just a small subset of GPUs.