Check that nvcc can compile to compute cap in build.rs

[coreylowman]

Throw helpful error message if not

[quietlychris]

I was just looking into this, actually (sorry that I've been offline with respect to the question you asked in #624, I haven't had a lot of spare capacity for open source recently).

Unfortunately, it looks like I'm unable to update nvcc in order to get that matching; I've double-checked that I'm using the latest drivers that are available through System76 packaging (see this and this). While I might be able to do a manual install, I'd like to avoid this if possible, and I think that it would be beneficial from a community perspective to support this type of system configuration; Pop!_OS is commonly used in the single-machine deep learning community due to System76's built-in CUDA management, and it's relatively common due to its popularity within the Linux gaming community, all of whom are potential users of dfdx.

I was thinking that maybe we could try a nested solution? Throw an error if the current check in build.rs doesn't work, which can then be caught and try to use the code I submitted in #619, before otherwise erroring out? I know it's not the most elegant solution, but I think it could cover quite a few potential set-ups.

[coreylowman]

Thanks for the update. I think nested would work well. Sounds like the logic would be:

1. Grab compute_cap from nvidia-smi
2. Grab available gpu codes from nvcc (using your code)
3. If nvidia-smi compute_cap is higher than the highest gpu code from nvcc (your case), then choose the highest gpu code in nvcc
4. Otherwise, nvidia-smi compute_cap should be present in nvcc gpu codes, and we can use that

The case that I had was nvidia-smi compute_cap was less than highest nvcc gpu code, so in the new logic we'd just choose nvidia-smi since it's present in nvcc gpu codes.

[coreylowman]

Thanks for the PR @quietlychris, works great on my end!