AOMP 20.0-1 was released on Tuesday as the newest version of this LLVM/Clang downstream focused on shipping the latest AMD patches around Radeon/Instinct OpenMP accelerator offload support.

The [1]AOMP compiler leverages AMD's latest OpenMP GPU/accelerator offloading patches prior to being upstreamed into LLVM proper for providing the very best AMD GPU compiler offload experience. AOMP also bundles in the related ROCm components it uses as part of its GPU compute stack to serve as a standalone build for having all the user-space bits needed for jiving with the AMDGPU/AMDKFD kernel drivers and not needing a separate ROCm install.

With the AOMP 20.0-1 release besides updating against a newer state of LLVM 20 Git, it also updates against the [2]ROCm 6.3 source code. There are also various libomp target changes, adding official support for Ubuntu 24.04 LTS and SLES 15 SP5 and other newer Linux distributions, and adding the SPIRV-LLVM-Translator to the mix with initial SPIR-V JIT offloading capabilities. This new SPIRV-LLVM-Translator component is named "amd-llvm-spirv" and can be used for SPIR-V to LLVM IR translation. This SPIR-V JIT offloading is considered experimental and still in active development.

AOMP 20.0-1 also brings various other component changes, renames the "flang-legacy" binary to "flang-classic", and other updates.

See the AOMP 20.0-1 release notes on [3]GitHub for more details on the changes in this latest AOMP compiler release. As a reminder, this isn't to be confused with [4]AOCC that is AMD's LLVM compiler downstream focused on the latest CPU (Zen) support.



[1] https://www.phoronix.com/search/AOMP

[2] https://www.phoronix.com/news/AMD-ROCm-6.3-GitHub

[3] https://github.com/ROCm/aomp/releases/tag/rel_20.0-1

[4] https://www.phoronix.com/search/AOCC