AMD Releases AOMP 18.0-1 Compiler Based On ROCm 6.0, Defaults To -O2 Optimizations

As the first new tagged version of AMD's AOMP LLVM-based OpenMP-focused compiler for offloading to their Instinct / Radeon GPUs, AOMP 18.0-1 was released today with many changes.

AOMP is AMD's LLVM downstream focused on providing the latest OpenMP device offloading support to their hardware ahead of the relevant patches being upstreamed into LLVM itself. AOMP 18.0-1 is the latest version based in turn on LLVM 18 Git, with LLVM 18 stable expected in March.

The AOMP 18.0-1 compiler has re-based its sources against targeting the big ROCm 6.0 release that debuted in December. ROCm 6.0 provides Instinct MI300 series support, new AI capabilities, and more for this open-source GPU compute stack.

AOMP 18.0-1 also notably now defaults to the -O2 optimization level if no other optimization level is specified. Up to now AOMP would default to no compiler optimizations (-O0) if not specified so at least now it's using the sensible -O2 as the default behavior for better out-of-the-box performance.

AOMP 18.0-1 also contains build system enhancements to make it easier to compile this AMD OpenMP compiler from scratch, test infrastructure improvements, and various fixes.

The AOMP 18.0-1 sources as well as CentOS/RHEL, SUSE, and Ubuntu binaries are available via GitHub for those wanting to enjoy the latest AMD OpenMP device offloading support.