LLVM Lands Support For OpenMP Offloading To AMD Radeon GPUs

[phoronix]

Phoronix: LLVM Lands Support For OpenMP Offloading To AMD Radeon GPUs

While the AMDGPU back-end has been part of mainline LLVM for years, not until now has it supported an OpenMP toolchain for offloading to Radeon GPUs...

http://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-LLVM-OpenMP-Merged

[Setif]

Phoronix Test Suite still doesn't have OpenMP offloading tests.

[ms178]

As I am not a programmer, that might be a naive question: But is the compiler capable of automatically offloading suitable code to this OpenMP target or are pragmas in the code still needed? I suppose it is not that easy to let the compiler handle all the optimization and offloading work without touching a single line of code.

[oleid]

As I am not a programmer, that might be a naive question: But is the compiler capable of automatically offloading suitable code to this OpenMP target or are pragmas in the code still needed? I suppose it is not that easy to let the compiler handle all the optimization and offloading work without touching a single line of code.

The #pragma are still needed. But instead of running multi-threaded on your CPU, your GPU can now be used as well (via LLVM). This is already possible via GCC for some time.

[Djip007]

https://github.com/RadeonOpenCompute/ROCm#AOMP-Deprecation

AOMP is already deprecated in favore of more LLVM upstream... I think...

[MadeUpName]

Hard to see this as any thing other than good news. Sill got my fingers crossed that Fedora gets their AMD openCL act together for 34.