AMD Releases AOMP 13.0-3 For Their Latest Radeon OpenMP Offload Compiler

The latest work happening on AMD's Radeon Open eCosystem (ROCm) front is a new release of AOMP, their LLVM Clang downstream focused on carrying the latest patches around Radeon OpenMP offloading support.

AOMP 13.0-3 is based on an early snapshot of the LLVM/Clang 13.0 Git state as of April while adding in a lot of AMD's own patches that haven't yet had the time to go through the processes to be upstreamed.

AOMP 13.0-3 has "significant" enhancements around multi-image support and target-id handling. AOMP also now supports the "--offload-arch" option for specifying the offload configuration / OpenMP target more easily. Along with the offload-arch option for easing the OpenMP offloading arguments needing to be passed to the compiler, AOMP also adds a new utility called "offload-arch". In the simplest form, offload-arch simply prints out the name of the GPU/family to the console for the detected GPU on the system. Offload-arch is intended just to be a helper to provide the GPU name for the system for knowing what to pass to --offload-arch= and related options.

AOMP 13.0-3 also adds new runtime checking, handling changes around archive files, and a range of different fixes.

More details on this updated Radeon OpenMP GPU offloading compiler via ROCm's GitHub.