CUDA on ppc64le / aarch64

[jaimergp]

Now that cudatoolkit==10.2 is available on defaults and linux-anvil-ppc64le-cuda is on Docker, is it possible to add this configuration to the cuda_compiler_version list and corresponding zipped keys?

-- edit see summary on current status here --

[jakirkham]

Sounds fine to me. Do you want to add a PR?

Probably involves doing something like this with nvcc first ( conda-forge/nvcc-feedstock#38 ) and then reverting it later ( conda-forge/nvcc-feedstock#39 ) once conda-forge-pinning is updated as before ( #377 ).

I guess the other question is, do we want the older CUDA versions there as well ( AnacondaRecipes/cudatoolkit-feedstock#8 )?

[leofang]

@jaimergp Is a PR to this feedstock the final step to get it up and running?

[jaimergp]

@leofang - I think the current support for PPC64LE is only in "manual mode", aka maintaining your own conda_build_config.yaml, as hinted in conda-forge/nvcc-feedstock#66 (comment)

However, if the team thinks that a migration is worth for PPC-10.2, I'd be ok with it.

[jakirkham]

It might be worth migrating, but it might be better to do this once we have ARM as well

[leofang]

@jakirkham Is there anyone working on ARM now?

[jaimergp]

I have summarized the status in conda-forge/cudatoolkit-feedstock#64, where I think it makes more sense to have this conversation. I'd rather close this one here because it took me ages to find it :D

cc @leofang @jakirkham -- note that right now there's no way to add functioning ARM support due to the CentOS8. CUDA 10.2 for PPC is technically there and can be migrated to if we want.

[jakirkham]

Closing now that PR ( #2527 ) is merged