CuPy support

[dcherian]

1. We should test with numpy-groupies. CuPy provides bincount, Dispatch ufunc.at methods cupy/cupy#7561
2. We'd have to avoid factorizing with Pandas unfortunately and use np.searchsorted or np.digitize; or use CuDF?

[dcherian]

Turns out scatter_* are a repacement for ufunc.at not ufunc.reduceat. The former does not take an axis kwarg so we can't use it for now.

cupy 12 might have more ufunc support, so let's wait and see

[kmaehashi]

Hi, CuPy maintainer here! We've created a dedicated tracker for ufunc features: cupy/cupy#7082
Do you need reduceat for specific ufuncs or all ufuncs?

[dcherian]

Thanks for checking in @kmaehashi

Technically we could use

• add.reduceat, multiply.reduceat, maximum.reduceat, and minimum.reduceat (with engine="flox") and
• multiply.at, maximum.at, and minimum.at (for engine="numpy" which uses numpy_groupies)

The second one is the default now, supporting .at would be most helpful

For reduceat, add.reduceat would be the most useful since that lets us do sum, mean, var and std

[kmaehashi]

Thanks! Sounds it aligns with our v12 target 😄
cc/ @asi1024