Radeon ROCm 3.0 Released With LLVM "AOMP" For Radeon OpenMP, FFT Updates

Announced last month at SuperComputing 19 in Denver was Radeon Open eCosystem 3.0 (ROCm 3.0) but it didn't end up shipping until last night. ROCm 3.0 is a big update to AMD's open-source Linux compute stack for ending out 2019.

Radeon Open eCosystem 3.0 is the first pulling in AMD's "AOMP" LLVM Clang downstream. AOMP is their modified version of LLVM Clang that adds support for OpenMP API offloading for Radeon GPU hardware. Hopefully they'll end up getting this AMD Radeon OpenMP support mainlined into LLVM/Clang but at least for now AOMP is being developed independently. Working OpenMP support for Radeon GPUs should open up more code-bases to now offloading to the AMD graphics processors, but we'll have to give it a spin with some software supporting the latest OpenMP APIs first for seeing how well it works in practice.

Besides having AOMP in ROCm 3.0, there is now CentOS/RHEL 7.7 support as part of this update. Another area with a lot of updates is the Fast Fourier Transform (FFT) handling with performance optimizations and other improvements. There are also performance improvements for the MIGraph graph optimizer and other enhancements.

Sadly not supported with ROCm 3.0 is Navi/GFX10 support... Vega still is the latest hardware appearing to be officially supported by Radeon Open eCosystem with no word yet when Navi will be officially supported.

Radeon Open eCosystem 3.0 can be downloaded from GitHub. I'll be trying out ROCm 3.0 shortly on Vega.