richardson_lucy example crashes with large filter size

[eddy16112]

The example is https://github.com/nv-legate/cunumeric/blob/branch-22.05/examples/richardson_lucy.py
If running with richardson_lucy.py -t -fx=512 -fy=256 -fz=151 -x=2048 -y=1024 -z=604, then I am running to the following errors:

Internal cuFFT failure with error code 4 in file cunumeric/convolution/convolve.cu at line 1133
Internal cuFFT failure with error code 1 in file cunumeric/cudalibs.cu at line 88

Even if I set x=fx, y=fy, z=fz, it still crashes.

[magnatelee]

this should be fixed by #279. @eddy16112 can you pull my branch and see if that fixes the issue?

[magnatelee]

leaving a note to myself and others: the issue was caused by the convolve implementation trying to call cuFFT with pointers that are not 8B aligned. This is actually a documented behavior: https://docs.nvidia.com/cuda/cufft/index.html#function-cufftexecr2c-cufftexecd2z Though this fix addresses the reported issue, I'll revert #204, which originally introduced this bug; while fixing the bug, it turns out that cuFFT callbacks make the convolution slower more than 3X for some unfortunate shapes, which is too great to justify the saving on memory space for temporary allocations.

[magnatelee]

@eddy16112 Please pull and try again.

[eddy16112]

It works for me. Thank you!