BUG: Fix the implementation of numpy.array_api.vecdot

[asmeurer]

See https://data-apis.org/array-api/latest/API_specification/generated/signatures.linear_algebra_functions.vecdot.html

[seberg]

Seems OK, I think I slightly prefer the np.moveaxis() solution:

x1 = np.moveaxis(x1, axis, -1)
x2 = np.moveaxis(x2, axis, -1)

res = x1[..., None, :] @ x2[..., None]
return res[..., 0, 0]

Plus the wrapping/unwrapping stuff. Or use einsum("...i,...i->...", x1, x2).

[asmeurer]

OK, made it use moveaxis + matmul instead of einsum.

[seberg]

This will go away with the change.

[asmeurer]

Maybe this can be simplified, but without this, the shape checking is not correct (because size 1 dimensions broadcast).

Also, the broadcasting is necessary because, assuming I am understanding the idea in the spec correctly, nonnegative axis should refer to the broadcasted dimensions.

[seberg]

You can move the checking to later and use -1 as axis. That probably gives you better errors for bad axis values anyway.

[asmeurer]

I'm sorry, I'm not following your suggestion here.

[seberg]

I am saying that you can do the check after the broadcast+moveaxis at the point where the axis is properly normalized to -1.

[seberg]

I took the liberty to do that change, unless you disagree will merge next time around (or earlier with feedback).

[asmeurer]

You missed some dots here.

[asmeurer]

I guess the spec's a little ambiguous here, but my understanding is that vecdot should not broadcast along the specified axis. With my code, it does not (e.g., vecdot(empty((1, 3)), empty((2, 3)), axis=0) gives an error), but with your code it does. Not allowing broadcasting along the contracted dimension is inline with other functions like matmul.

We should probably figure out which behavior we want here and update the spec to be clearer.

[charris]

Just to note that we ignore the lint errors in the array_api.

[seberg]

Yeah, I missed the boardcasting part, lets go with the previous version for now.

[seberg]

Thanks @asmeurer, sorry I had lost track of this, lets put it in.