[FEA] Multi-GPU support for cupyimg about cucim

[jakirkham]

cc @grlee77 (who may know more)

[jakirkham]

Dask-image has an affine_transform, which uses CuPy under-the-hood. So maybe that can be used as-is.

There isn't currently a phase_cross_correlation implementation for multiple GPUs.

If the problem can be solved in smaller chunks, using map_blocks or map_overlap may make sense. This can then be followed by some computation on the aggregate result. For example one might identify shifts needed for smaller chunks of the image and then average these over the whole image or similar. This could give a rough answer that may be good enough.

Alternatively Dask has its own FFTs and Dask-image has fourier_shift. So it might be possible to roll your own registration algorithm using these techniques.

[b2jia]

@gigony (and perhaps @caryr35 ) I'm following up maybe too late. Happy to see phase_cross_correlation with GPU support, and I gave cucim another try. The results of cucim.skimage.registration.phase_cross_correlation often do not match that computed on CPU (skimage.registration.phase_cross_correlation. Is there a reason why?

CUCIM version: 23.02.00
skimage version: 0.18.3

from cucim.skimage.registration import phase_cross_correlation as pcc_gpu
from skimage.registration import phase_cross_correlation as pcc_cpu

drift_gpu = pcc_gpu(cp.asarray(img_stack_v1[:, :512, :512]), 
                    cp.asarray(img_stack_v2[:, :512, :512]),
                    upsample_factor=100)
drift_cpu = pcc_cpu(img_stack_v1[:, :512, :512], 
                    img_stack_v2[:, :512, :512],
                    upsample_factor=100)
print("Drift gpu", drift_gpu[0])
print("Drift cpu", drift_cpu[0])
>>>Drift gpu (0.85, 0.5, -0.11)
>>>Drift cpu [ 0.03  0.21 -0.04]