ENH Add SBSA aarch64 CTK pkgs for CUDA 11.X

[mike-wendt]

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.

Add aarch64 support with the SBSA runfile for CUDA 11.X versions

NOTE: This does not support the Jetson/AGX platform. Those images use L4T which is not compatible with the SBSA installs. Currently Jetson/AGX only support CUDA 10.2 and SBSA only supports CUDA 11.X so there is no overlap at the moment. We're actively working with internal teams on how to best address this when AGX adopts CUDA 11.X support but that is still TBD.

For Jetson/AGX builds that use conda it's recommended to use a selector and omit the CTK from recipes for now. This means users have to ensure they have the correct version of CUDA installed on their system and it doesn't allow for versioning but the currently JetPack for AGX/Jetson only supports CUDA 10.2

[conda-forge-linter]

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

I wanted to let you know that I linted all conda-recipes in your PR (recipe) and found some lint.

Here's what I've got...

For recipe:

• The recipe license should not include the word "License".

For recipe:

• License is not an SPDX identifier (or a custom LicenseRef) nor an SPDX license expression.

Documentation on acceptable licenses can be found here.

[mike-wendt]

@jakirkham I did not make this change but I wanted input on this section. There are sysroot versions for aarch64 and ppc64le should I add them to this section or not?

cudatoolkit-feedstock/recipe/meta.yaml

Lines 294 to 296 in ab72362

	{% if "sysroot_version" in cudavars[major_minor] %}
	- sysroot_linux-64 {{ cudavars[major_minor]["sysroot_version"] }} # [linux64]
	{% endif %}

[mike-wendt]

FYI I've been seeing a lot of failures with retrieving the CUDA runfiles and the site itself takes a while to load. May have to wait and re-run the CI later when the CDN is in a better state.

[jakirkham]

So that comes from PR ( #16 ). AIUI this was only needed for x86_64 as we support CentOS 6 & 7 there and needed a way to force CentOS 7 be used with CUDA 11.0+. With aarch64 and ppc64le, we only support CentOS 7. So don't think that should be needed

[isuruf]

So don't think that should be needed

AFAIK, sbsa aarch64 doesn't support centos 7.

[mike-wendt]

So don't think that should be needed

AFAIK, sbsa aarch64 doesn't support centos 7.

Right that's why I didn't include it, just didn't know if there was a higher glibc version to match the Centos 8 requirement or not. If the answer is to leave as-is that works for me

[isuruf]

You can merge it here without it, but we'll not be able to use it anywhere.

[mike-wendt]

So the aarch64 jobs are failing due to curand and the glibc version it expects. I'm guessing this is because we are building on CentOS 7 images for a product that supports CentOS 8, Ubuntu 18.04/20.04

Finding curand from Conda environment
	located at $PREFIX/lib/libcurand.so.10.2.1.245
	trying to open library...	ERROR: failed to open curand:
/lib64/libm.so.6: version `GLIBC_2.27' not found (required by $PREFIX/lib/libcurand.so.10.2.1.245)

Are there any workarounds for this? I'm guessing this is no without a heavy lift to support cos8 directly in conda-forge.

[jakirkham]

So don't think that should be needed

AFAIK, sbsa aarch64 doesn't support centos 7.

Yep, you're right.

Also seeing ppc64le requires CentOS 8 for CUDA 11.2 as well. Does anyone know when that changed?

https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#system-requirements

---

In terms of support in conda-forge, here are the things we will need:

• Docker images for CentOS 8
• Rebuild of compilers for CentOS 8
• Updating cudatoolkit (this PR)
• Updating nvcc package
• Migrator (or some strategy to handle migration) to add this to relevant feedstock
• Anything else?

cc @beckermr (as well who went through this before with CentOS 7 in case he spots anything)

[mike-wendt]

Are there any workarounds for this?

Also seeing ppc64le requires CentOS 8 for CUDA 11.2 as well. Does anyone know when that changed?

https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#system-requirements

@jakirkham From scrubbing through the docs it looks this looks like the breakdown, starting with CUDA 11.0 CentOS 8 was required

CUDA	Released	x86-64 OS support	aarch64 OS support	ppc64le OS support
10.2	Nov 2019	RHEL 6.10/7.7/8.1	N/A	RHEL 7.6 & Ubuntu 18.04.1
11.0_GA	May 2020	RHEL 7.x/8.y (x <= 8, y <= 2)	RHEL 8.y (y <= 2) & Ubuntu 18.04.z (z <= 4)	RHEL 8.y (y <= 2) & Ubuntu 18.04.z (z <= 4)
11.1.0	Sep 2020	RHEL 7.x/8.y (x <= 8, y <= 2)	RHEL 8.y (y <= 2) & SLES 15.y (y <= 2) & Ubuntu 18.04.z (z <= 5)	RHEL 8.y (y <= 2) & Ubuntu 18.04 LTS
11.2.0	Dec 2020	RHEL 7.x/8.y (x <= 9, y <= 3)	RHEL 8.y (y <= 3) & SLES 15.y (y <= 2) & Ubuntu 18.04.z (z <= 5)	RHEL 8.y (y <= 3) & Ubuntu 18.04 LTS
11.2.1	Feb 2021	RHEL 7.x/8.y (x <= 9, y <= 3)	RHEL 8.y (y <= 3) & SLES 15.y (y <= 2) & Ubuntu 18.04.z (z <= 5)	RHEL 8.y (y <= 3)
11.2.2	Mar 2021	RHEL 7.x/8.y (x <= 9, y <= 3)	RHEL 8.y (y <= 3) & SLES 15.y (y <= 2) & Ubuntu 18.04.z (z <= 5)	RHEL 8.y (y <= 3)

Notes

• Even though docs for CUDA 11.2.0 show ppc64le support for Ubuntu 18.04 there are no downloads

[jakirkham]

^ @jaimergp

[jaimergp]

Hm, that will require some rebuilds, I guess? Or at least some repodata patching. I'm knee deep in other stuff at work but we will be releasing a new OpenMM version soon that brings full PPC support, so I guess I will put some work into the non-x64 builds too!

[jakirkham]

Yep just letting you know 🙂

[jakirkham]

Also Anthony had started working on a Docker image for ARM in PR ( conda-forge/docker-images#158 )

[jakirkham]

@conda-forge-admin, please re-render