AMD Releases AOMP 16.0-2 For Its Radeon OpenMP Offloading Compiler

In addition to AMD this week having released the AMD Optimizing C/C++ Compiler "AOCC" 4.0 as their LLVM/Clang downstream now with various optimizations for Zen 4, the company also released AOMP 16.0-2 as the newest version of their other LLVM/Clang downstream... AOMP is their downstream LLVM/Clang compiler focused on providing the latest Radeon OpenMP GPU offloading support.

AOMP 16.0-2 is the newest snapshot of their compiler with Radeon / Instinct GPU offloading patches atop what is currently tracking the LLVM 16 upstream code-base.

AMD

AOMP 16.0-2 is re-based against an LLVM 16 Git state from mid-October, drops support for Ubuntu 18.04, brings various fixes, improves xteam reductions, adds support for the target team loops directive, forces synchronous execution of regions controlled via OMPX_FORCE_SYNC_REGIONS, and adds a new environment variable GPU_MAX_HW_QUEUES. The new GPU_MAX_HW_QUEUES environment variable can be used for controlling the number of HSA queues created while the default is four queues.

Source downloads as well as binaries for RHEL/CentOS, Ubuntu, and SUSE Linux Enterprise are available from GitHub. With time most of these relevant AMD Radeon/Instinct compiler patches should work their way into upstream LLVM. Unlike the AOCC compiler on the CPU side, at least AOMP has the patches publicly available for those interested in the changes over upstream LLVM.