Include ptxas in cudatoolkit package

[jayfurmanek]

It was discovered that XLA does not work with the latest TensorFlow GPU builds:
AnacondaRecipes/tensorflow_recipes#24 (comment)

This is because XLA requires the ptxas compiler and libdevice.10.bc.

This PR has two changes to the cudatoolkit package to fix XLA:

• the libdevice.10.bc package is in the wrong location for XLA. It's shipped in $CONDA_HOME/lib when it should be in $CONDA_HOME/nvvm/libdevice/ so this PR copies it to this location (and keeps it in /lib too in case some things need it there)

• the ptxas binary, which is used by the XLA compiler, is not included in the package at all currently so this PR adds it into $CONDA_HOME/bin

Unfortunately, this feedstock doesn't have any branches for older cuda versions, so this PR is only against the latest (11.0).

[jayfurmanek]

@katietz FYI

[jayfurmanek]

@jjhelmus

[jayfurmanek]

Note I didn't test this on Windows since I don't have a Windows machine.

[jakirkham]

@jayfurmanek would suggest submitting this to the conda-forge feedstock ( https://github.com/conda-forge/cudatoolkit-feedstock )

Also Jonathan left Anaconda FYI

[leofang]

@jayfurmanek Could you also fire up a PR to https://github.com/conda-forge/cudatoolkit-feedstock if you have time? 🙂 It is maintained by some NVIDIA folks so it's in a better shape. I am less certain about whether or not ptxas is redistributable as per NVIDIA's EULA, but the change for libdevice.10.bc looks like no harm at least for Linux.

btw I think Jonathon is no longer a maintainer here. You might have to ping someone else from Anaconda (try https://gitter.im/conda/conda).

cc: @jakirkham @kkraus14

[jayfurmanek]

@jakirkham What do you think about ptxas as being redistributable? It is part of the Cuda Toolkit and presumably under the same terms?

I can look at putting this up in forge as well.

[kkraus14]

@jakirkham What do you think about ptxas as being redistributable? It is part of the Cuda Toolkit and presumably under the same terms?

I can look at putting this up in forge as well.

ptxas is not redistributable per the EULA: https://docs.nvidia.com/cuda/eula/index.html#attachment-a

[kkraus14]

Will ping some folks internally to see what can be done here

[leofang]

Aha John and I replied at the same time!

[katietz]

The part about libdevice.10.bc is for sure something I would merge into Anconda's recipe. The ptxas part needs to be delayed. If there might be a solution found for it, ok. So I would suggest to make at least for the libdevice.10.bc part a new PR.