Thanks for working on this Kai!

Another note;ndarray.T has a corresponding functionnp.transpose, as part of this PR should we add a corresponding function for.mT e.g.np.matrix_transpose?

I'd say yes we want that (because of the analogue to.T and becausethe array API standard has it). It doesn't have to be in this PR though, if it's already large and easier to do as a follow-up - whatever seems best to you.

This will need a release note.

@rgommers I'm happy to work on it in this PR, but I'm still learning about NumPy internals, meaning that the commit history could become quite messy as I learn (of course I'll try my best to keep it clean). As a summary, I think we should match the functions fortranspose i.e.:

• np.matrix_transpose
• np.ndarray.mT
• np.ndaray.matrix_transpose
• np.ma.matrix_transpose
• np.ma.MaskedArray.mT
• np.ma.MaskedArray.matrix_transpose

The next step for me is to addnp.ndaray.matrix_transpose, which I am having some trouble with. What I have done so far is:

• Moved the implementation into a new function (PyArray_MatrixTranspose) intonumpy/core/src/multiarray/shape.c alongsidePyArray_Transpose
• Added amatrix_transpose entry toarray_methods innumpy/core/src/multiarray/methods.c
• AddedPyArray_Transpose to themultiarray_funcs_api innumpy/core/code_generators/numpy_api.py

I'm not really sure if this is the right direction, or if we even want to addPyArray_MatrixTranspose to the C-API as (I think) I am doing.

I would not touch the NumPy C-API with this PR. It is enough to add the attribute to ndarray and ma. If in the future we wish to export it via the C-API, that can be another PR.

As a summary, I think we should match the functions fortranspose i.e.:

We don't need/want both a.mT property and a.matrix_transpose method. It's duplicate and the plan is to trim down onndarray methods. Just the property and thenp.matrix_transpose function (and their MaskedArray equivalents) is enough.

I rather like the idea of an.mT property :) It would be useful to have a short way to call it in...@...@... constructions.

I've removed the code adding it to the C-API (I think) and stopped working on adding anndarray.matrix_transpose method. I've kept the functionPyArray_MatrixTranspose inshape.c as I want to be able to share it betweenndarray.mT andnp.matrix_transpose - this is causing the test to segfault compared to the earlier version.

I may need some help structuring the code so that I can share it betweenndarray.mT andnp.matrix_transpose.

You have to add the new function to a header file now and include it, probably. If you check the test failures, you see they fail on compiler warnings.

@seberg I'vefinally got the masked array implementation working! I'm not sure if it's done correctly, could you take a look at my last commit?

If it's right, I think I have three things left:

• Take the reference to the function versions out of the release notes
• Add the documentation
• Add type hints for the masked array

@rgommers@seberg I've made the last changes that (I think) are needed, could I please ask you two for another review?

I've decided to leave thenp.matrix_transpose function for now, I'll get to it in another PR

Thanks, lets put it in.

After merging, I noticed that the docs actually render slightly incorrectly. Not a big deal, but maybe you want to follow up?

I guess we need<BLANKLINE> or maybe its fine to just remove the blank line also.

Thanks that's my first C level contribution 👌. Let's hope for many more.

I'll follow up with the docs rendering properly.