Allow multiple GPU architectures at compilation

[gpenazzi]

An enhancement suggestion.

It would be convenient to compile ELPA with Nvidia GPU support for multiple architecture, such that the library will contain compiled cubin for multiple architectures. This allows better portability, especially in heterogeneous environments, and facilitates cross-compilation. It is a common option in other GPU-enabled libraries which would be nice to have.

I attach an example of a patch for configure.ac. Would you consider to take this feature upstream?

The patch reflects the pattern we use, although a common one, which is to generate PTX for the highest architecture (ensures forward compatibility) and cubin for multiple architectures (best performances and forward driver compatibility within the same major version). Currently we only use ELPA1 on GPU, therefore it is not tested (yet) with the options related to the ELPA2 kernel.

2025.01.002-allow_for_multi_gpu_arch.txt

[terminationshock]

Dear @gpenazzi ,
thank you very much for your suggestion and your patch. We are currently discussing the details internally, but are in principle open for this change. I have adapted your patch in the way that the user has to set, for example, sm_80 instead of 80 to preserve backwards compatibility. Also the default value for the config setting will be unchanged.
We will keep you updated...

[gpenazzi]

Thank you, it sounds good!

[terminationshock]

Brief update: We are working on it, especially to make sure that our handwritten kernels are working correctly on heterogeneous GPU clusters.