I have tested this with #14327 and this is able to compile and run several examples from the HIP repository

So a bit of a challenge with this build. To get clang to correctly link with HIP I had to export ROCM_PATH=$EBROOTHIP. This enables hipcc to properly function and is then able to compile HIP sources for AMD GPUs. However, one thing that I did not notice right away is that this same export makes clang unable to pick up it's own device lib.

In other words, if one just loads the Clang-AOMP module and tries to compile an OpenMP offloading code that works. Loading HIP is then able to build HIP GPU codes, but now OpenMP offloading will not work. It works for HIP since hipcc sets up the path to device lib manually.

This has effects on downstream packages. I wanted the ROCm package to be the one to depend on when you need ROCm, but when trying to build hipSYCL this problem (the path exporting) ended up making it impossible to build.

So a bit of a challenge with this build. To get clang to correctly link with HIP I had to export ROCM_PATH=$EBROOTHIP. This enables hipcc to properly function and is then able to compile HIP sources for AMD GPUs. However, one thing that I did not notice right away is that this same export makes clang unable to pick up it's own device lib.

In other words, if one just loads the Clang-AOMP module and tries to compile an OpenMP offloading code that works. Loading HIP is then able to build HIP GPU codes, but now OpenMP offloading will not work. It works for HIP since hipcc sets up the path to device lib manually.

This has effects on downstream packages. I wanted the ROCm package to be the one to depend on when you need ROCm, but when trying to build hipSYCL this problem (the path exporting) ended up making it impossible to build.

Managed to solve this by exporting HIP_DEVICE_LIB_PATH in Clang-AOMP since that is what Clang is using. With that export loading this module or the full ROCm module works without issue. I also managed to solve the problem for the hipSYCL downstream package (which was a bit unrelated).

@nordmoen Required patch file is missing in the PR?

Must have spaced out on that, added now, in addition to your suggestion about sources.

Test report by @boegel
SUCCESS
Build succeeded for 1 out of 1 (1 easyconfigs in total)
pg-lab02 - Linux centos linux 8.3.2011, x86_64, AMD EPYC 7542 32-Core Processor (zen2), Python 3.6.8
See https://gist.github.com/af93175791988d81f07e8bbdda626447 for a full test report.

Test report by @boegel
SUCCESS
Build succeeded for 1 out of 1 (1 easyconfigs in total)
pg-lab02 - Linux centos linux 8.3.2011, x86_64, AMD EPYC 7542 32-Core Processor (zen2), Python 3.6.8
See https://gist.github.com/d441cdd7ccf8f33aa60d383ef63c81ae for a full test report.

Going in, thanks @nordmoen!