Add support of np.swapaxes #4074

[braniii]

This PR is adds np.swapaxes, which is currently unimplemented according to #4074 (or np.swapaxis which is the mis-spelling version). I had help by @stuartarchibald on discourse: (https://numba.discourse.group/t/how-to-use-types-integer-for-indexing-list/647)

This function can be used to simplify the implementation of np.rot90 from #6822, which is also the first step of implementing full further arguments of it.

[esc]

@braniii thank you for submitting this, I have added it to the queue for review.

[stuartarchibald]

This PR is adds np.swapaxes, which is currently unimplemented according to #4074 (or np.swapaxis which is the name in numpy <1.16).

I can't find reference to np.swapaxis anywhere in the git log or change log apart from in places where there's people correcting the spelling to swapaxes, has it ever existed or is it just a common typo/mis-spelling?

[stuartarchibald]

Thanks for working on this feature. I've given this an initial review, there's a few small things to resolve but the patch is generally good. Thanks again!

[stuartarchibald]

I'm reasonably convinced that this will break the "swapaxes returns a view" notion as per the docs, https://numpy.org/doc/stable/reference/generated/numpy.swapaxes.html:

For NumPy >= 1.10.0, if a is an ndarray, then a view of a is returned; otherwise a new array is created.

Asserting that a is expected and a is got in the unittests might highlight this.

[braniii]

a is expected and a is got should always be false (except when exchanging with the same axis). For being true, not only the data but also the view needs to be the same (which gets changed) or one needs to use the base or np.share_memory.

You are right, that the definition breaks the "swapaxes returns a view". But the reason is, that the numba version of np.transpose does not return a view. Here a short code snippet:

import numpy as np
from numba import njit


@njita: False
def transpose(a):
    return np.transpose(a)


# reference arraype command to continue
a = np.arange(8).reshape(2, -1)

b = np.transpose(a)
c = transpose(a)

# this is always false
print(f'b is a: {b is a}')  # false
print(f'c is a: {c is a}')  # false

# using base it works
print(f'b.base is a.base: {b.base is a.base}')  # true
print(f'c.base is a.base: {c.base is a.base}')  # false

[stuartarchibald]

I don't think there's much we can do about this right now? Do you feel that this behaviour is prohibitive to it being useful as-is?

[braniii]

At least from my side there is not much I can do, I have not enough inside in the code base to propose a fix. But I think the function is useful as-is. But in general it could be helpful to document difference from numpy functions, maybe simple at
https://numba.readthedocs.io/en/stable/reference/numpysupported.html

- `np.transpose()`
+ `np.transpose()` (a copy is returned)

[stuartarchibald]

At least from my side there is not much I can do, I have not enough inside in the code base to propose a fix. But I think the function is useful as-is. But in general it could be helpful to document difference from numpy functions, maybe simple at
https://numba.readthedocs.io/en/stable/reference/numpysupported.html

- `np.transpose()`
+ `np.transpose()` (a copy is returned)

Agree, this is probably the best thing to do for now.

[braniii]

This PR is adds np.swapaxes, which is currently unimplemented according to #4074 (or np.swapaxis which is the name in numpy <1.16).

I can't find reference to np.swapaxis anywhere in the git log or change log apart from in places where there's people correcting the spelling to swapaxes, has it ever existed or is it just a common typo/mis-spelling?

When I searched for the function I found multiple results, and I did not consider that numba could be wrong ;) But you are right, I corrected my comment. And thx for the review.

[stuartarchibald]

Thanks for the patch and fixes!