AMD Releases AOMP 15.0-3 Compiler

AMD engineers have released an updated version of AOMP, their LLVM/Clang downstream that carries the company's latest patches around OpenMP offloading to Radeon GPUs.

Not to be confused with AOCC as their LLVM/Clang compiler carrying the latest Zen CPU patches, AOMP is from the graphics side of the house to deliver their latest OpenMP offloading support for Radeon GPUs / Instinct accelerators.

AOMP is part of the ROCm software collection and aimed to deliver the latest Radeon/Instinct OpenMP offloading support until the patches are eventually upstreamed into mainline LLVM.

AOMP 15.0-3 was released on Monday as their newest version tracking LLVM/Clang 15.0 but now that LLVM/Clang 16.0 upstream is the next feature version, AMD will be adapting that for AOMP 16.0-0 moving forward.

In addition to AOMP 15.0-3 re-basing against LLVM Git upstream as of the end of July, there are a few AMD changes included in this update. Most significant is switching to the new OpenMP device run-time library (DeviceRTL) by default. (Update:) This is the new OpenMP GPU Runtime led by researchers from Argonne National Laboratory, Oak Ridge National Laboratory, Lawrence Livermore National Laboratory, Stony Brook University, and the University of Waterloo. This whitepaper outlines this new run-time and optimizations aiming for near-zero overhead execution. Additional details can also be found in this LLVM patch.

AOMP 15.0-3 adds new DeviceRTL APIs for optimized cross-team reduction. Other AOMP 15.0-3 work includes Clang code generation improvements for using the new optimized cross-team reduction APIs, Flang Fortran compiler support to use the new DeviceRTL, and other updates.

More details on AOMP 15.0-3 along with binaries for CentOS / RHEL, SUSE Linux Enterprise, and Ubuntu are available from GitHub.