Compiling with Cuda 11.0

[gaspardbb]

Hi, [Sorry, just realized it might be more relevant in the PR section; can close this and edit this as a PR if you wish]
I had to use Falkon, and it worked great. Thanks for all the work you put in!
Still, I ran into a couple of errors before managing to run it on GPU. Here how I fixed them, if you find it useful:

Not compiling with Cuda

If Falkon is not compiled with Cuda (WITH_CUDA in your cmake file), the extensions are not built and running Falkon will fail with an ModuleNotFoundError at from falkon.ooc_ops.cuda import parallel_potrf in falkon/ooc_ops/ooc_potrf.py. It's a bit hard to track down that this is an issue coming from the first compilation. Maybe catching the exception and providing an error message would be useful, e.g:

try: 
    from falkon.ooc_ops.cuda import parallel_potrf
except ModuleNotFoundError as e:
    print(f"Got exception {e} when importing `cuda`. Did you compile with Cuda support?")

Patched version of PyKeops

I had trouble compiling you patched version of PyKeops. I had only two Cuda compilers available: versions 10.1 and 11.0.

1. With Cuda 10.1, had the issue described here: https://forums.developer.nvidia.com/t/cuda-10-1-nvidia-youre-now-fixing-gcc-bugs-that-gcc-doesnt-even-have/71063. It is apparently a cuda specific issue, I did not try solving it.
2. So I used Cuda 11.0. Except I fell in this problem: Update cmake minimum version to 3.18 getkeops/keops#122
   which is basically Cmake not forwarding c++ 17 flags to Cuda. Solution is to use:

• Cmake >= 3.18
• AND apply this commit: getkeops/keops@e19e2a1 which is specific to Cuda 11.0 (and not 11.1).

Then things worked fine; all in all, I would suggest:

• Providing information when compilation is done without Cuda (e.g. flags which makes the installation fails if WITH_CUDA is false in the setup.py.
• Fixed the case with Cuda 11.0 by checking CMake version and applying the commit

Anyway, thanks for the package!

[Giodiro]

Hi Gaspar!

Thanks a lot for the detailed info, I will make sure to address these issues soon, especially the better error handling when compiled without CUDA would make the install experience a lot better!

As a general comment I've been slowly trying to setup a way to build wheels for the package so that the user doesn't have to spend a lot of time compiling to install.. It's not been particularly fun, but it kept a few fixes back in a dev branch (such as for example the keops issue you mentioned).

Hopefully I manage to merge it soon!

Thanks again for your input, glad you found the package useful!