AMD software engineers today released AOMP 21.0-0 as the newest snapshot of their LLVM/Clang compiler downstream focused on providing the best OpenMP/OpenACC GPU offloading support to AMD GPUs and Instinct accelerators via the ROCm software stack.
AOMP 21.0-0 is their first release now based against the upstream LLVM/Clang 21 Git state in its early development form following the release LLVM 20 release. Plus AOMP 21.0-0 is layered in all of the latest AMD patches for working to help the GPU-accelerated offloading experience.
AMD changes in AOMP 21.0-0 include replacing the existing “Flang” Fortran compiler (Flang Classic) with using the new Flang “flang-new” compiler. The new/modern Flang compiler in LLVM is superior to the prior Flang implementation with better performance, greater Fortran features, and more.
AOMP 21.0-0 also brings performance improvements on min and max reductions via fmin/fmax functions as well as new infrastructure for executing host APIs in target regions with Emissary APIs. These AMD Emissary APIs use the offload RPC mechanism to transparently execution functions called from the target region on the host.
AOMP 21.0-0 also switches all C/C++/Fortran OpenMP toolchains to using the “clang-linker-wrapper” by default, updates from the ROCm 6.3 software sources, changes to the AOMP build infrastructure, and various other changes.
Those interested in this AMD LLVM downstream for GPU/device offloading can find more details on today’s 21.0-0 release via GitHub.