How to transpose?

[grothesque]

In current mdarray, is it possible to perform a (generalized) transpose operation, i.e. change the order of dimensions without copying the data?

Since a strided layout is supported this should be technically possible.

[fre-hu]

See the reorder, reorder_mut and into_reordered methods. Are they what you mean?

[grothesque]

Ah, I didn't find this one since I was searching for "Strided" on the documentation page for Slice.

Yes, this solves this provides transpose, but actually for my own project I need general reordering of dimensions. Technically this is possible with mdarray, right? Is it just a matter of writing a function, or would that need some reorganization of types as well?

[fre-hu]

There are functions for general permutations also (permute/permut_mut/into_permuted). Like the view functions they are implemented for fixed rank, so you might have issues with dynamic rank. Maybe the interface for all of these should be updated in some way.

Yes, you can also write a standalone function that uses View::new_unchecked with the pointer and a new mapping from StridedMapping::new. For dynamic rank you don't need any reorganization of types.

[grothesque]

Thanks!

IMHO the view/permute interface should be either (eventually) updated, or the dynamic rank removed. It feels incomplete in the current state. See #7 for my thoughts. No idea if/how C++ mdarray handles the dynamic rank case.

(BTW I hate to use the term rank for number of dimensions, because for me the rank of a matrix is something different. But I see that C++ mdarray uses the same term, and also it's compact.)

[fre-hu]

I've made one more commit where the permutation can be dynamic, and also axis arguments in other functions. This simplifies the typical case with dynamic dimensions and is more aligned with ndarray. There are also axis index functions for simple views, which work with dynamic rank. Hopefully this should improve useability.

C++ mdarray does not support dynamic rank.

[grothesque]

With the updated permute this one is solved, isn't it? (I tried to test permutations of DynRank-arrays with many dimensions, but got stuck.) Edit: It works, only that I don't know how to use static arrays as shapes.

[fre-hu]

Yes I think so, I will close it. I'm not sure what you mean with static arrays, but the permutation must be an IntoShape (i.e Dim tuple, [usize; N], &[usize] or similar).

[grothesque]

I'm not sure what you mean with static arrays

I meant what you answered in #6 (comment). Thanks.

[grothesque]

Having used reorder for a while, I cannot get used to the name. To me, “reorder” is a synonym of “permute”, while the method reorder performs a very specific reordering/permutation.

How about renaming reorder to transpose? This naming might be not perfectly adequate when there are more than 2 dimensions, but since there is no well-known multi-dimensional generalization of the transpose operation, I think that would be OK. In terms of precedent, at least NumPy works like this.

If “transpose” is not acceptable, how about “reverse_axes”, or “reverse_dims”, or something like that? Or just “reverse”, but IMHO that could be confused with reversing the order of entries inside the array. (I do think that “transpose” is clearer, though.)

If it has to remain reorder, I suggest to at least clarify the documentation string which currently says: “Returns a reordered array view of the array slice.”

[fre-hu]

I agree, reorder is not so good name.

I was unsure how to define the transposition operation, but it is probably best to follow the convention in Numpy and rename reorder to transpose.

[fre-hu]

Fixed, see latest commit.

[grothesque]

Thanks!