HSA Accelerator Support Should Be Usable For GCC 6

One of the big focuses as of late for GCC compiler development is device offloading support (e.g. GPGPU, MICs, etc) and as part of that support for the Heterogeneous System Architecture (HSA).

With GCC 5 released earlier this year there is offloading support in the form of OpenMP to Intel MIC targets (Xeon Phi) and OpenACC to NVIDIA PTX targets. That existing offloading support has been covered in previous Phoronix articles and is also documented at the GCC Offloading Wiki page. For GCC 6, HSA accelerator support should be in place plus other OpenMP 4 improvements.

Earlier this month at the GNU Tools Cauldron in Prague, there were talks about the offloading/accelerator support for the GNU Compiler Collection.

Martin Jambor, one of the SUSE developers working with AMD on HSA support, talked about compiling for HSA accelerators with GCC using OpenMP 4. You can see his PDF slides for all of the technical information if interested. Right now that HSA support is in a GCC SVN branch, but for the GCC 6 release in 2016, the plan is to have it mainlined, Jambor said, "So even given the current problems we plan to merge the branch to gcc 6." The problems relate to libgomp and handling of different constraints around OpenMP 4 for HSA accelerators.

On a related note, Intel also had developers at the cauldron where they talked about OpenMP 4 offloading features for GCC. Of course, their interest in device offloading by GCC is for Xeon Phi targets. Those slides can be found here.