BUG: ravel_multi_index does not work on 0D arrays

[eric-wieser]

Reproduction:

>>> def test(n):
...     inds = np.zeros((n, 5), dtype=np.intp)
...     return np.ravel_multi_index(inds, (1,)*n)
>>> test(2)
array([0, 0, 0, 0, 0], dtype=int64)
>>> test(1)
array([0, 0, 0, 0, 0], dtype=int64)
>>> test(0)
ValueError: At least one iterator operand must be non-NULL

The expected result is array([0, 0, 0, 0, 0], dtype=int64) for all cases.

Relates to gh-13934.

[eric-wieser]

Hmm, I think I've seen this before (edit: gh-580). The issue is that the signature of ravel_multi_index is really:

ravel_multi_index(multi_index: Sequence[np.ndarray], shape: tuple)

and not

ravel_multi_index(multi_index: np.ndarray, shape: tuple)

As a result, ravel_multi_index has no shape information available.

This is the reverse of the arr_0d.nonzero() problem.

[madphysicist]

FWIW, this has the same problem:

def test(n):
    inds = np.zeros(5, dtype=np.intp)
    return np.ravel_multi_index((inds,) * n, (1,)*n)

ValueError: At least one iterator operand must be non-NULL

[madphysicist]

I had to put in a workaround for this recently, so let me know if you start working on a solution.