Build CUDA under host_injections and make EESSI aware of host CUDA drivers

[ocaisa]

No description provided.

[eessi-bot]

Instance eessi-bot-mc-aws is configured to build:

• arch x86_64/generic for repo eessi-hpc.org-2023.06-compat
• arch x86_64/generic for repo eessi-hpc.org-2023.06-software
• arch x86_64/intel/haswell for repo eessi-hpc.org-2023.06-compat
• arch x86_64/intel/haswell for repo eessi-hpc.org-2023.06-software
• arch x86_64/intel/skylake_avx512 for repo eessi-hpc.org-2023.06-compat
• arch x86_64/intel/skylake_avx512 for repo eessi-hpc.org-2023.06-software
• arch x86_64/amd/zen2 for repo eessi-hpc.org-2023.06-compat
• arch x86_64/amd/zen2 for repo eessi-hpc.org-2023.06-software
• arch x86_64/amd/zen3 for repo eessi-hpc.org-2023.06-compat
• arch x86_64/amd/zen3 for repo eessi-hpc.org-2023.06-software
• arch aarch64/generic for repo eessi-hpc.org-2023.06-compat
• arch aarch64/generic for repo eessi-hpc.org-2023.06-software
• arch aarch64/neoverse_n1 for repo eessi-hpc.org-2023.06-compat
• arch aarch64/neoverse_n1 for repo eessi-hpc.org-2023.06-software
• arch aarch64/neoverse_v1 for repo eessi-hpc.org-2023.06-compat
• arch aarch64/neoverse_v1 for repo eessi-hpc.org-2023.06-software

[ocaisa]

This can be tested by checking out the branch, starting the build container from the PR and running, e.g.,

source /cvmfs/pilot.eessi-hpc.org/versions/2023.06/init/bash 
./gpu_support/nvidia/install_cuda_host_injections.sh 12.1.1

[ocaisa]

But I need to transition the PR to (also?) use eessi_container.sh

[ocaisa]

This PR can be tested without GPUs, use

./eessi_container.sh -m shell --nvidia install -v

to start the build container with the necessary settings, then you can install CUDA (under host_injections) with

source /cvmfs/pilot.eessi-hpc.org/versions/2023.06/init/bash 
./gpu_support/nvidia/install_cuda_host_injections.sh 12.1.1

@trz42 There are some sizable changes to the build container in the PR so you may wish to take it for a test drive.

[ocaisa]

@trz42 I need some advice here. I want to probably always trigger the ability to install CUDA, but I only want to trigger the host_injections build of CUDA when it is actually required (or we do it always but reuse the same space so it only actually gets done once)

[ocaisa]

@casparvl I've addressed your review as much as I can and I've added better command line options, including a critical acceptance of the CUDA EULA. This should be read for showtime now

[ocaisa]

I think I'll fold the script to the linking to the host drivers in here as well...working on it now

[casparvl]

Lgtm! Awaiting rerun of a final CI test that crapped out because of rate limiting.