AMD Releases AOMP 16.0-0 As Their Radeon GPU OpenMP Offloading Compiler
AOMP is AMD's staging area or playground for their GPU offloading patches until they are upstreamed within LLVM itself. AOMP isn't to be confused with AOCC as AMD's Optimizing C/C++ Compiler focused on their Zen-enriched downstream of LLVM/Clang for the CPU side of the house. With AOMP 16.0-0, they have re-based against the latest LLVM upstream code as of last week while also adding a few new Radeon offloading patches.
AOMP 16.0-0 captures the upstream LLVM 16 Git state as of 14 September. AOMP also pulls in the relevant ROCm components it deals with too sans the AMDGPU/AMDKFD kernel code. There are only two new AMD-led changes with AOMP 16.0-0:
- Adds new flag -fopenmp-target-fast to group enable a set of OpenMP target optimizations.
- Enhancements and bug fixes for No-Loop and cross-team reduction support.
The new -fopenmp-target-fast option sounds interesting in the name of performance. AOMP 16.0-0 sources as well as binaries for Ubuntu / SUSE / CentOS (RHEL) Linux systems can be found via GitHub.