search

data-apis / array-api

RFC document, tooling and other content related to the array API standard
https://data-apis.github.io/array-api/latest/
MIT License
211 stars 44 forks source link

Only require axis to be negative in vecdot and cross #740

Closed asmeurer closed 7 months ago

asmeurer commented 8 months ago

Nonnegative axes and negative axes less than the smaller of the two arrays are unspecified.

This is because it is ambiguous in these cases whether the dimension should refer to the axis before or after broadcasting. Previously, the spec stated it should refer to the dimension before broadcasting, but this deviates from NumPy gufunc behavior, and results in ambiguous and confusing situations, where, for instance, the result of a the function is different when the inputs are manually broadcasted together.

Also clean up some of the cross text a little bit since the computed dimension must be exactly size 3.

Fixes #724 Fixes #617

See the discussion in those issues for more details.

I don't think this should be backported, since my vecdot implementation in np.array_api and array-api-compat already followed the spec behavior here, and will need to be updated.

asmeurer commented 8 months ago

We could also allow nonnegative axis when the two input arrays have the same number of dimensions. In that case, there is no ambiguity for something like axix=0 because it would refer to the same dimensions both before and after broadcasting.

leofang commented 7 months ago

@kgryte could you remind me if this will be part of v2023?