Radeon GCC Back-End Updated For Running Single-Threaded C & Fortran On AMD GPUs

Back in September Code Sourcery / Mentor Graphics posted the Radeon GCC back-end they have been developing with the cooperation of AMD. This is for allowing the GCC compiler to eventually offload nicely to Radeon GPUs with its different programming languages and supported parallel programming models, particularly with OpenMP and OpenACC in mind. But for now this patch series just works with single-threaded C and Fortran programs. The second version of this port was posted for review.

Hitting the GCC mailing list on Friday was the updated version of this AMD GCN port targeting Tonga/Fiji through Vega graphics hardware. Code Sourcery will post the OpenACC/OpenMP support bits at a later date while for now the code works with single-threaded C/Fortran programs with C++ not yet supported, among other initial shortcomings. For now the AMDGPU LLVM back-end is far more mature in comparison, which is what's currently used by the open-source AMD Linux driver compute and graphics stacks.

The latest version of these patches can be found on the mailing list.

While GCC 9 is embarking on its third stage of development where it's mostly about fixing and documentation, new ports can still land during this stage as long as they don't touch existing code. So we'll see if/what of this Radeon back-end makes it into the GCC 9 compiler release due out in 2019.

This GCC port is also used by Code Sourcery / Mentor Graphics' CodeBench Lite software for GPU offloading.