Singularity: Update Singularity recipe for eddy_cuda9.1 support

[araikes]

Cuda support in the official Singularity container is currently not functional #2383 (comment)). FSL's Eddy requires CUDA 9.1 or 8. However, Ubuntu 20.02 does not inherently support CUDA < 10.

Modifications to the Singularity recipe installs CUDA 9.1 toolkit only as well as relevant patches. I build this locally (CentOS 7, CUDA 11.4 installed) and eddy_cuda9.1 runs for dwifslpreproc.

Ubiquitous testing requires multiple OS options, which I do not have access to.

[Lestropie]

@araikes Do you have any recollection of whether it was necessary to no longer wipe the contents of these directories, or whether this is just residual from testing?

[Lestropie]

Note: This was originally addressed in comment in #2383.

[Lestropie]

Sorry for taking forever to take a look at this @araikes.

I'd wanted to merge this to master and slip it into the pending 3.0.4, given I've been doing a bunch of other fixes around the new containers as well (#2461, #2370, #2360). Couple of concerns though:

• This approximately triples the size of the container. I'm guessing that only a fraction of the libraries brought in are actually necessary to make use of eddy_cuda9.1. It would be preferable to make use of neurodocker-minify to produce a reduced download file. But not only would this require getting CUDA working in the container used for detection of dependencies (https://github.com/MRtrix3/containers), there would also potentially be a risk of that operation omitting dependencies that change depending on the host system and/or target hardware. I can hopefully look into this at some point, but it won't happen for 3.0.4.

• The Linux system I'd been intending to use for testing the proposed change in retrospect is not a good choice, as it's an Arch laptop with non-functional dual-GPU hardware. I think I only have easy access to one other system with adequate GPU hardware to run eddy_cuda. I can test on a couple of HPCs prior to merging, but those are slow to test on and each will likely have their own requisite wrapper configurations.

I'll try to test what I can soon and post back here.

[Lestropie]

I haven't had any success getting this or alternative solutions to work on either of my testing systems. So I'll transition from 3.0.4 to 3.1.0 milestone. But I do want to get this working.

[araikes]

Hi @Lestropie,

I'll try and circle back around to this next week. There may be some opportunity for container size optimization, watching what's happening with FastSurfer development.