ENH: Add np.asanycontiguousarray #7217
Conversation
Numpy has np.ascontiguousarray() and np.asfortranarray(), but doesn't have a function to return an array of either contiguity depending on the input, making a copy only when necessary.
|
Probably should post it to the NumPy mailing list if you haven't already. If you have, maybe a link to the mailing list entry would be nice and vice versa. |
| if is_f_contiguous: | ||
| return array(a, dtype, copy=False, order='F', ndmin=1) | ||
| else: | ||
| return array(a, dtype, copy=False, order='C', ndmin=1) |
There was a problem hiding this comment.
Not sure, I like the idea too much, but that is not a deal breaker (also the any has a bit of a name clash with asanyarray). But I think you can simplify this to a single line function:
return array(a, dtype, copy=False, order="A", ndmin=1)
(not that I like the ndmin=1 either, what is so bad about 0-D arrays? But I guess the ascontiguousarray functions do that too). The whole f_contiguous stuff should be handled by "A" is my guess.
|
Needs release note. |
|
If @seberg is right that this is just |
|
I'm also -0.5 for this. We need to limit the growth of small functions at some point. |
|
Unfortunatly, I am not right. "A" typically works like that, so I |
|
Can we fix A? |
|
If it is not really used much, I guess it is an option. It is |
|
Well, "A" typically means "any layout", so I'm not sure why changing it to "any contiguous layout" in a single API would make things easier for the user. |
|
(I certainly wouldn't guess that calling |
|
Another option would be to add another order value, like |
|
None of the options sound very convincing to me on first sight :(. "A" stands for any, that is true, but it seems (almost?) only used in the "fortran if not C (and actually fortran)" sense, though sometimes in terms of iteration order and not memory layout (we have "K" as a weaker "A"). "ForC" or variations of this, may make some sense since we use it in |
|
This is looking like a reasonable addition. Does anyone object? @pitrou Needs mention in 1.12.0 release notes. |
|
I'm still -0 I guess since it isn't clear to me how this could have broad utility for multiple users (what is the use case, BTW?), and because it isn't clear that adding a new function is better than adding a new option to |
|
The use case we have in mind is writing wrappers for third-party libraries such as BLAS, which often accept either C- or F-ordered arrays, but not non-contiguous arrays. |
|
☔ The latest upstream changes (presumably #7539) made this pull request unmergeable. Please resolve the merge conflicts. |
|
I think the name conflict with |
|
A better way to add the desired functionality would be via adding a kwarg to |
|
Good point, giving that an order keyword argument should be pretty uncontroversial. @pitrou maybe that would be a solution to continue here? |
Numpy has np.ascontiguousarray() and np.asfortranarray(), but doesn't have
a function to return an array of either contiguity depending on the input,
making a copy only when necessary.