Enable CUDA builds using conda-forge packages

[jaimergp]

Checklist

• Used a personal fork of the feedstock to propose changes
• Bumped the build number (if the version is unchanged)
• Reset the build number to 0 (if the version changed)
• Re-rendered with the latest conda-smithy (Use the phrase @conda-forge-admin, please rerender in a comment in this PR for automated rerendering)
• Ensured the license file is being packaged.

Pending tasks

Linux 64

None so far.

Windows

• Provide drivers for CI -- see Install CUDA on Windows if cuda_compiler_version is present in %CONFIG% conda-forge-ci-setup-feedstock#129
• nvcc does not exist yet -- see Add Windows nvcc-feedstock#49
• Update matrix for cuda_compiler and cuda_version in conda-forge-pinning -- see Add migrator for CUDA on Windows conda-forge-pinning-feedstock#914
  • Added manual migrator with commit fc40e07

Note: We might need to drop CUDA 9.2 on Windows because it requires vs2015, now deprecated in conda-forge.

MacOS

None so far. Cuda not available.

Linux PPC

• nvcc does not exist yet
• Update matrix for cuda_compiler and cuda_version in conda-forge-pinning

ARM

None so far. Cuda not available.

[conda-forge-linter]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe) and found it was in an excellent condition.

[jaimergp]

I am going to merge later today so we can have something to test locally. I will change the label so it's test to mimic cudatoolkit. This might need an update in some of our downstream CI, @jchodera, but I will handle that too.

This is being labeled as 7.5.0, btw, but shouldn't cause too much trouble given that we are in a different subchannel anyway.

PowerPC will be addressed in a separate PR since it's not as urgent.

[jakirkham]

Thanks so much for working on this Jaime! 😄

On ppc64le we are still missing some cudatoolkit packages ( conda-forge/cudatoolkit-feedstock#12 ). So we would probably want to resolve that first (if we are interested in those versions).

[jaimergp]

As far as OpenMM is concerned we only need one CUDA build that works nicely at Summit. We are hoping 10.2 is that version 😬

If it's not, I'll put some work on 10.1 and 10.0 (and I guess 9.2 for free?). Before that, I will submit PRs for nvcc and conda-forge-pinning though (even if it's only for 10.2).

Btw, while I have your attention, can you take a look at conda-forge/conda-forge-ci-setup-feedstock#131 ? Ideally there's a Nvidia-approved way to get nvcuda.dll installed on GPU-less machines :)

[jchodera]

@jaimergp Did we ever get the ppc64le issues sorted out? Is it just the missing cudatoolkit package we're waiting on?

[jaimergp]

Docker images and cudatoolkit packages are in place. We only need the ppc64le nvcc wrapper (and I guess the relevant updates in the pinnings). However, I wanted to wait a bit to do this until we could run the full CI suite with the conda compilers (GCC9), which we are doing now in openmm/openmm#2964. There are some non-trivial errors that'd need to be fixed before releasing the ppc64le builds.

[jchodera]

@jaimergp Would you be able to comment on what non-trivial errors remain to be addressed in OpenMM for ppc64le?

[jaimergp]

CI for master passes for a configuration compatible with conda-forge, so it should work for the upcoming 7.5.1, discussed here openmm/openmm#2993