[Python] What's the fastest way to do a cross-correlation/convolution on GPU in python?

I'm trying to speed up a python analysis that uses libraries including numpy, scipy, and opencv. I have been mainly doing this by using cupy to replace numpy and scipy. There is one step that uses the opencv filter2D function that I want to speed up. I tried switching to scipy.ndimage convolve2d and then using the cupyx equivalent, but even on GPU it is an order of magnitude slower (~2s vs. 0.1s in my test). Does anyone know why it is so much slower and what other options I have?

Continue reading...