AMD's ROCm AOMP Compiler 11.7-1 Brings OMPD Support, ROCgdb

The AMD ROCm developer tool engineers have released a new build of AOMP, their LLVM Clang compiler downstream that adds OpenMP support for Radeon GPU offloading until that support ultimately makes it back upstream into LLVM/Clang.

The ROCm engineers working on AOMP have been doing a great job on keeping their code re-based against the newest upstream LLVM code, which with this release is from just two weeks ago prior to the LLVM 11.0 branching. The AMD developers have been working on upstreaming more of their LLVM/Clang changes albeit that is a lengthy process especially with new Radeon OpenMP code continuing to be written and fine tuned.

Besides re-basing the LLVM code-base, AOMP 11.7-1 brings OMPD support, a build of their ROCgdb GNU Debugger, a host runtime optimization, and fixes.

OMPD support is the OpenMP standard providing an interface to third-party tools like debuggers. OMPD allows for third-party tools to inspect the OpenMP state in real-time of programs. OMPD is all contingent upon the OpenMP implementation being used and now AMD is shipping a supported library to assist in the debugging of workloads being offloaded to Radeon GPUs.

The host runtime optimization for AOMP in this release is GPU image information is now "mostly" read from the host CPU instead of from the GPU memory.

More details on AOMP 11.7-1 along with source downloads as well as x86_64 binary builds for RHEL 7, SLES 15 SP1, and Ubuntu 16.04/18.04 can be found from GitHub.