Build system limitations for older CUDA versions

[rykerfish]

I was attempting to set up libMobility on Brennan's desktop. We ran into a few issues that I thought might have wider applicability to any potential user base.

The main problem is an inability to support older CUDA versions. I think we're limited to CUDA 12+ due to needing span to manage GPU/CPU memory intelligently (is that right?). Because of this, it seems unlikely that we could easily support CUDA 11. Even though the most recent CUDA version 11 were released almost three years ago, do you think it's worth attempting to support them? I think that, in general, HPCs should be able to accommodate CUDA 12, but I'm wondering if this would be a restriction for other people in the research community.

The secondary issue is that the current conda environment.yml file isn't very flexible. Because we install specific CUDA packages, they can install conflicting dependencies if you set a version restriction on one package. Using a meta package with more coverage like cudatoolkit-dev would install dependencies of the correct versions, although it does make the conda environment much bulkier. I don't think this issue particularly matters since we can't support older CUDA versions at the moment anyways, but it may be relevant in the future.

[RaulPPelaez]

Ok so, the main reason why I stay in CUDA >=12 is because the support for it in conda-forge is much better. The conda-forge repo swtiched to a new, standardized, treatment of CUDA starting from CUDA 12.0.
It makes dealing with CUDA super smooth, but adopting it means that it is awkward to support CUDA 11 and prior.

IMHO it is best to stick with the current environment for future proofing.

We can write a small guide in the docs explaining installation for CUDA<12. It amounts to changing the names of the packages in the environment to cudatoolkit-dev as you noticed, or adding the nvidia conda channel, for which they allow CUDA 11: https://anaconda.org/nvidia/cuda

Other than that, the main reason that liMobility is restricted to CUDA 12 is that I set the C++/CUDA standard to C++20, which is only supported by CUDA12. The changes to remove this restriction are doable but damn, C++20 has been around for 5 years already, I do not think we have the bandwidth to support older software stacks...

[rykerfish]

Maybe I didn't do everything right, but I'm not sure if libMobility will even work with CUDA<12. When I tried on Brennan's machine, it didn't seem able to since it seems like span was only introduced in C++20, and since CUDA 11 only has partial support for C++20 it failed.

I agree with you that we can't support CUDA<12, especially since all of the memory handling relies heavily on span. Maybe we just add something to the README about it and move on?

[RaulPPelaez]

Yes I would do that and thats it.
BTW which GPU does Brennan have? Is it like super old?