@grlee77, if you have time, could you confirm that this PR resolves all of the bugs you found? Thanks!

Note: as a bug fix that involves minimal code changes (at the Cython level), this PR meets the original intention for supporting C-contiguous arrays. However, with proper reduction axes (or axis=None) F-contiguous arrays can also use CUB. I will address this in the next PR.

I can confirm the c-contiguity fix works. I haven't tested the others yet

I can confirm that all of the previously failing cases from #2634 pass with these changes. Thanks for the quick fix @leofang!

Not related to the test failures, but I think there is still a "bug" in terms of not using cub for some compatible axis combinations in _routines_statistics.pyx. This one is just a performance loss for cases where the user has specified the axis

specifically the two places with:

if axis is None and cub.can_use_device_reduce(

should just be

if cub.can_use_device_reduce(

can_use_device_reduce will check the axis via _cub_device_reduce_axis_compatible.

Probably worth just fixing that here as well

Good catch, @grlee77! That was a residual from the segmented reduce experiment. I just stole your commit and corrected it 🤣

Jenkins, test this please!

Jenkins, test this please!

Jenkins CI test (for commit 291b2a3, target branch master) succeeded!

Thank you @grlee77 for finding those bugs and testing this fix and @take-cheeze for review!

@take-cheeze I suppose this PR can be tagged with v7.0.0?

@leofang Thank you for noticing. Should be added now