Radeon AOMP 14.0-2 Compiler Released For OpenMP GPU Offloading

Along with this week bringing ROCm 5.0, on Friday AMD engineers released Radeon AOMP 14.0-2 as their downstream of LLVM/Clang focused on carrying the latest patches for OpenMP offloading to Radeon GPUs.

AOMP 14.0-2 is their newest release of this LLVM/Clang downstream that carries AMD's latest GPU offloading patches until they are ready for mainline/upstream integration. Back in November they introduced AOMP 14.0 as they began tracking the LLVM/Clang 14.0 Git development, updated with a new AOMP snapshot in January, and yesterday released AOMP 14.0-2. Their LLVM/Clang snapshot is that of upstream as of earlier this month prior to LLVM 14.0's branching with the latest Git code upstream now working on LLVM 15.0.

AOMP 14.0-2 brings device runtime performance improvements thanks to overlapping multiple copies between the host and the device. There is also now a static build of hwloc in its OpenMP library to support the use of "PLACES" for CPU affinity. Another feature with AOMP 14.0-2 is support for memory managed allocations in OpenMP. Other AOMP 14.0-2 changes include numerous fixes.

AOMP 14.0-2 downloads and more details on the changes via the AMD/ROCm GitHub.