AMD Releases AOMP 20.0-0 For Radeon/Instinct Compiler Offloading

Following last week's release of the LLVM/Clang-downstream AOCC 5.0 for optimized compiler support extended to Zen 5 CPUs, the GPU side of the house at AMD this week released AOMP 20.0-0 as their LLVM/Clang downstream focused on GPU device offloading.

AOMP remains the AMD-maintained LLVM/Clang downstream compiler focused on OpenMP API device offloading to Radeon GPUs and Instinct accelerators. AOMP pulls in all of the latest AMDGPU compiler back-end bits and associated device offload patches for what isn't currently within the upstream LLVM repository. But unlike the AOCC compiler, AOMP is open-source. The AOMP developers also do a good job ensuring their patches do work their way upstream where relevant in a timely fashion.

AOMP continues to track LLVM Git and hence the version 20.0-0 to mark their first release since LLVM Git was bumped to version 20 development. AOMP 20.0-0 pulls in the latest ROCm 6.2.2 source code, AOMP now creates the relevant rocm/clang/clang++ configuration files, the gpurun utility now has multiple device support using the "-md" argument, updates to AOMP example code, and other updates.

AOMP 20.0-0 also adds in ROCm SMI and AMD SMI as AOMP components. Lastly the gfx90c, gfx1103, gfx1150, gfx1151, and gfx1152 AMD GPU targets are now enabled. GFX115* are for the RDNA 3.5 refresh parts and gfx1103 is for the Radeon 780M and gfx90c is for the AMD APUs with Vega graphics.

Downloads and more details on the AOMP 20.0-0 release via ROCm on GitHub.