OpenMP AMD GPU Offloading

[Cordell Bloor]

Hi folks,

I discovered recently that OpenMP GPU offloading has been packaged by 
the LLVM Team. On Debian Testing or Unstable you will need the following 
packages installed:

     hipcc libomp-21-dev liboffload-21-dev

If you have a sample program, like veccopy.c [1], it can be built with 
the command:

     clang-21 -O2 -fopenmp --offload-arch=native veccopy.c -o veccopy

You can test that it's correctly offloading to the GPU by running the 
test with an environment variable to disable CPU fallback:

     OMP_TARGET_OFFLOAD=MANDATORY ./veccopy

I wonder if we should have a metapackage that installs the necessary 
requirements for OpenMP GPU offloading? That might be something to 
consider for src:rocm.

I've found that it's possible to use OpenMP GPU offloading on Trixie 
with ROCm 5.7 and LLVM 17, but the reorganization of the packaging for 
liboffload in later LLVM versions fixed some library path problems, so 
the newer versions are easier to use. On Trixie, you'll need to use a 
custom LD_LIBRARY_PATH to get it working. I should probably put 
instructions for all this in the ROCm section of the Debian wiki.

I haven't found a lot of "Getting Started" documentation for these 
features, but I did come across a great presentation by Jan-Patrick Lehr 
at SC23 on "OpenMP Target Offloading for AMD GPUs" [2].

The LLVM Team seems to have gotten this working without any great 
fanfare. A big thanks to them!

Sincerely,
Cory Bloor

[1]: 
https://github.com/ROCm/aomp/blob/rocm-7.2.0/test/smoke/veccopy/veccopy.c
[2]: https://www.youtube.com/watch?v=m0MqPYmlGbM