A new version of AMD’s Radeon OpenMP Offloading Compiler, AOMP 16.0-2, has been released

The most recent snapshot of their compiler is version AOMP 16.0-2, and it includes enhancements to GPU offloading for graphics cards that use the Radeon and Instinct brands. It is constructed on top of the code base that is currently being watched over by the upstream LLVM 16 project.

The AOMP 16.0-2 release has been rebased against the LLVM 16 Git state that existed around the middle of October. It does away with support for Ubuntu 18.04, introduces a lot of improvements, improves xteam reductions, adds support for the target team loops directive, and forces synchronous execution of regions controlled with OMPX FORCE SYNC REGIONS. In addition to that, it includes the introduction of a new environment variable known as GPU MAX HW QUEUES. Users now have the ability to pick the number of HSA queues that are formed thanks to a new environment option called GPU MAX HW QUEUES. The default setting for this option is four queues.

On GitHub, users

of RHEL/CentOS, Ubuntu, and SUSE Linux Enterprise can obtain downloadable versions of the source code as well as binaries for usage with their distributions. The vast majority of these essential bug patches for the AMD Radeon and Instinct compilers should, if given enough time, make it into the upstream version of LLVM. At the very least, in contrast to the AOCC compiler that runs on the CPU side, the AOMP compiler makes the fixes publicly available for anyone who are interested in the changes that have been made to the upstream version of LLVM.