WIP GPU implementation of FFT-based smoothing in ML

[pierrethibault]

Implemented only in cuda_pycuda for now, and does not support stacks of objects.

[ptim0626]

Hello, I temporarily added a dummy _get_smooth_gradient_fft in the serial version to avoid calling the old GSK while using ML_pycuda.

[jcesardasilva]

Hello, I temporarily added a dummy _get_smooth_gradient_fft in the serial version to avoid calling the old GSK while using ML_pycuda.

Ok, thank you @ptim0626. May I kindly ask for your recommendations on the best course of action to pursue? Additionally, are we looking to explore a more long-term solution to the matter at hand? What are the options?

[ptim0626]

Hi @jcesardasilva, the important thing is making sure ML_pycuda here uses the FFT-based smoothing in every smoothing operation, the fix here ensures this.

As with how to proceed, I would suggest first and foremost this is working in some testing dataset, and get the corresponding kernel in cuda_cupy. What do you think? @pierrethibault @bjoernenders @daurer

[daurer]

The current solution seems to fix the problem that was reported when using the smoothing preconditioner in ML_pycuda. But more work is need before this can be merged.

• should also be implemented in ML_cupy
• should we provide both convolution-base and fft-based version and let user decide or silently replace with FFT version in pycuda/cupy engines?
• do we need a similar change in the object smoothing regulariser in DM etc.?

[pierrethibault]

I would just replace the convolution without adding an option, and use this in DM object smoothing also.

[daurer]

FFT-based smoothing is now implemented as a kernel for both PyCUDA and CuPy.
The fft-based smoothing regulariser is available in both ML_pycuda and ML_cupy by changing the new parameter smooth_gradient_method from "convolution" (default) to "fft".

FFT-based smoothing for other use cases, i.e. object smoothing in DM is still outstanding, see #546