Address more comments

[lithomas1]

Reference issue

What does this implement/fix?

Additional information

[lithomas1]

cleared about half of the comments, more to follow.

[crusaderky]

Not true. All arrays are non-writeabla for jax. All arrays are writeable for dask.

[lithomas1]

r.e. writeability, I am not referring to the mutability of the original dask array, but the numpy array created from the dask array with np.asarray.

MRE

import dask.array as da
a = da.zeros(10)
a_np = np.asarray(a)
a_np[1] = 10
Traceback (most recent call last):
  File "<python-input-4>", line 1, in <module>
    a_np[1] = 10
    ~~~~^^^
ValueError: assignment destination is read-only

I'll remove the jax comment (and I'll raise an issue on scipy to disallow the output keyword for jax since it looks like there's no way to make that work).

[crusaderky]

Yes, that's because the chunks of a are broadcasted, and so is a_np unless it's a concatenation of chunks.

>>> da.zeros(10).__array__().strides
(0,)
>>> da.zeros(10, chunk=5).__array__().strides
()

Broadcasted arrays are read-only as the whole thing is only 8 bytes in size.

[crusaderky]

The whole idea of having an output= parameter makes no sense if you are going to convert it to numpy.

This makes sense:

def f(x: Array, output: Array) -> None:
    output[:] = x * 2

This does not:

def f(x: Array, output: Array) -> None:
    x = np.asarray(x)
    output = np.asarray(output)
    output[:] = x * 2

In the second example, even without the broadcasting issue you're facing, the output item that the user is holding will remain completely blank, unless np.asarray returns a view of the original memory. which, for dask, can never be.

[crusaderky]

Note that np.asarray(output, copy=False) will not crash, while it should. This is a bug in Dask.

[lithomas1]

Yeah, makes sense.

I'll put in a separate PR to the main scipy repo disabling the output kwarg for jax (to keep the diff down here).

[crusaderky]

FYI dask/dask#11697

[crusaderky]

Also this makes no sense to me.

[crusaderky]

FYI data-apis/array-api-compat#239

[crusaderky]

A fully Array API compliant function would be able to suppor the output= kwarg with dask.
The problem is that you're calling np.asarray(x) will always return a buffer that is not shared with the input parameter.

Please change reason to "converts dask output array to numpy"

[lithomas1]

I changed the comment.

[crusaderky]

only a small nit

[lithomas1]

Thanks for the review
@lucascolley this is ready for merge