Support cuTENSOR

[leofang]

According to cuTENSOR license the .so and .h files are redistributable:

2. Distribution. The following portions of the SDK are distributable under the Agreement: the runtimes files ending with .so and .h as part of your application.

I suppose the package should be added upstream. Then, we just add cutensor to our dependencies, and set CUTENSOR_PATH to $PREFIX prior to building.

I was able to build cuTENSOR support with our experimental self-hosted agent, but I don't know how to include its files in the package. (UPDATE) the module is not importable, see my comment below.

[leofang]

@jakirkham I think the glibc version used in CF's docker images (condaforge/linux-anvil-cuda:10.0 etc) is too old (2.12); cuTENSOR requires at least 2.14: https://dev.azure.com/nsls2forge/nsls2forge/_build/results?buildId=2526&view=logs&j=d0d954b5-f111-5dc4-4d76-03b6c9d0cf7e&t=6d4b912b-175d-51da-0fd9-4d30fe1eb4e7&l=2413

[leofang]

cc: @jjhelmus

[jjhelmus]

The CUDA Toolkit 10.0 release notes indicated that support for RHEL 6.x was deprecated and may be dropped in a future release. I would not be surprised if cuTENSOR did not support RHEL 6.x.

RHEL 6 / CentOS 6 use glibc 2.12, the later is the base for the conda-forge/linux-anvil-cuda:10.0 image.

[leofang]

Thanks, @jjhelmus! Does it mean that to enable cuTENSOR support here we should

1. Update conda-forge's CUDA dockers to CentOS 7
2. Package cuTENSOR and upload to Anaconda

[leofang]

1. Update conda-forge's CUDA dockers to CentOS 7

This is brought to upstream: https://github.com/conda-forge/docker-images/issues/137.

[leofang]

cuTENSOR can help cupy.sum(): cupy/cupy#2939

[jakirkham]

We are working on getting CentOS 7 up and running in conda-forge, but there are a few moving parts involved. So won't be done overnight, but we are working on it 😉

[leofang]

Closed by #94 (master) and #95 (rc).

[leofang]

(What a long journey we've been through!)