explicitly cast the dtype of where's condition parameter to bool

[keewis]

(this is necessary because array-api-strict raises an error otherwise)

• towards ⚠️ Nightly upstream-dev CI failed ⚠️ #10084

[keewis]

looks like it is not that simple... at some point we're trying to call

duck_array_ops.where(True, np.arange(5), np.float64(np.nan))

where numpy implements astype by forwarding to the ndarray method (and our own wrapper does the same).

To resolve this, we probably need to make cond a 0d array:

Suggested change

	return xp.where(xp.astype(condition, xp.bool), *as_shared_dtype([x, y], xp=xp))
	try:
	cond_ = astype(condition, xp.bool)
	except AttributeError:
	cond_ = xp.asarray(condition, dtype=xp.bool)
	return xp.where(cond_, *as_shared_dtype([x, y], xp=xp))

however, I'm not sure I like the way this looks. Another option would be to make astype work for this case

[eni-awowale]

Hi @keewis ! I got stuck on the same thing. What I did:

# Check array type
    if isinstance(condition, np.ndarray):
        condition = condition.astype(dtype=np.bool)

    elif isinstance(condition, xp_strict._array_object.Array):
        condition = xp_strict.asarray(condition, dtype=xp_strict.bool)
    else:
        raise TypeError("Unsupported array type")

    xp = get_array_namespace(condition, x, y)
    return xp.where(condition, *as_shared_dtype([x, y], xp=xp))

This passes test_array_api but test_computation fails with a bunch of TypeErrors

[keewis]

thanks, that gave me some ideas on how to possibly fix this in a cleaner way (and I also found a potential bug in the groupby code)

[dcherian]

thanks!