More performance for (cu)finufft

[paquiteau]

Following

flatironinstitute/finufft#564 (comment)

(Let's continue here to avoid flooding the PR review)

We installed cufinufft with pip. Manual installation with optimized flags may bring even better results.

@DiamonDinoia

[DiamonDinoia]

Hi @paquiteau,

(unistalling finufft, cufinufft or using --reinstall might be required)
you could do as showed in the most recent docs: https://finufft.readthedocs.io/en/latest/python.html

pip install --no-binary finufft finufft

in finufft for 2D and 3D ducc is faster than fftw so one can try:

pip install --no-binary finufft finufft --config-settings=cmake.define.FINUFFT_USE_DUCC0=ON finufft

and

pip install --no-binary cufinufft cufinufft

I assume you have a c++17 compiler (gcc-13 recommended) and nvcc (12 is best) installed.
I am not sure you need cmake or others as this might be brought in by pip.

Thanks,
Marco

[paquiteau]

Hello @DiamonDinoia , our engineer @Lenoush work on a revised edition of the benchmark. The results are in #5 . It seems that the --no-binary installation has worse performances than the classic one for cufinufft.

[DiamonDinoia]

Hi could you provide the images here? From the PR is not clear to me what I should be looking at.

cufinufft being slower but using less vram is possible. May I ask the exact version of cuda used?
finufft using GPU memory I cannot explain it. There is no cudaMalloc in the code. @janden could you have a look as if there is any python cuda allocation in finufft maybe we missed something.

[paquiteau]

The cuda allocation in finufft is probably a bug on our side, the process had probably some remaining GPU memory before running finufft.

The results for the updated benchmark are here to browse: https://github.com/mind-inria/mri-nufft-benchmark/tree/52a38f328124070a955b62bf51ec48adfdd27af2/results

[DiamonDinoia]

Can I ask the CUDA version?

[paquiteau]

Yes, we are using CUDA 12.2

[DiamonDinoia]

With 12.2 the use of alloca is enabled. I think that differently from my tests using alloca makes things a bit slower but by using the stack as much as possible it reduces VRAM consumption. I will test this further once I have time but the use of alloca in cufinufft might need to be made optional.

[DiamonDinoia]

It is not clear to me the binary and non binary installation of finufft? Did you also test with ducc0 fft ?