BUG: PyArray_MultiIter_SIZE is not defined in 1.26.4 headers despite documentation stating it is new in 1.26.0

[oleksandr-pavlyk]

Describe the issue:

Doc page https://numpy.org/doc/stable/reference/c-api/array.html for 2.0 stable states that PyArray_MultiIter_SIZE is new in 1.26.0.

It is defined with NumPy 2.0.

Lack of this function makes it difficult to update mkl_random to NumPy 2.0

Reproduce the code example:

(b_mkl_packages) opavlyk@opavlyk-mobl:~/repos/mkl_random$ conda list numpy
# packages in environment at /home/opavlyk/mamba3/envs/b_mkl_packages:
#
# Name                    Version                   Build  Channel
numpy                     1.26.4                   pypi_0    pypi
(b_mkl_packages) opavlyk@opavlyk-mobl:~/repos/mkl_random$ grep -r PyArray_MultiIter_SIZE $CONDA_PREFIX/lib/python3.12/site-packages/numpy
(b_mkl_packages) opavlyk@opavlyk-mobl:~/repos/mkl_random$

### Error message:

```shell
Building native extension in `mkl_random` fails with the like of


mkl_random/mklrand.cpp:10517:26: error: ‘PyArray_MultiIter_SIZE’ was not declared in this scope; did you mean ‘PyArray_MultiIter_GOTO’?
10517 |             __pyx_t_10 = PyArray_MultiIter_SIZE(__pyx_v_multi);
      |                          ^~~~~~~~~~~~~~~~~~~~~~
      |                          PyArray_MultiIter_GOTO
mkl_random/mklrand.cpp:10732:18: error: ‘PyArray_MultiIter_SIZE’ was not declared in this scope; did you mean ‘PyArray_MultiIter_GOTO’?
10732 |     __pyx_t_1 = (PyArray_MultiIter_SIZE(__pyx_v_multi) != __pyx_v_res_size);
      |                  ^~~~~~~~~~~~~~~~~~~~~~
      |                  PyArray_MultiIter_GOTO
mkl_random/mklrand.cpp:10970:24: error: ‘PyArray_MultiIter_NDIM’ was not declared in this scope; did you mean ‘PyArray_MultiIter_NEXT’?
10970 |     __pyx_v_multi_nd = PyArray_MultiIter_NDIM(__pyx_v_multi);
      |                        ^~~~~~~~~~~~~~~~~~~~~~
      |                        PyArray_MultiIter_NEXT
mkl_random/mklrand.cpp:10979:26: error: ‘PyArray_MultiIter_DIMS’ was not declared in this scope; did you mean ‘PyArray_MultiIter_DATA’?
10979 |     __pyx_v_multi_dims = PyArray_MultiIter_DIMS(__pyx_v_multi);
      |                          ^~~~~~~~~~~~~~~~~~~~~~
      |                          PyArray_MultiIter_DATA

### Python and NumPy Versions:

Python 3.12, NumPy 1.26.4 

### Runtime Environment:

_No response_

### Context for the issue:

_No response_

[oleksandr-pavlyk]

[ngoldbaum]

It seems @charris found it too hard to backport and it never shipped: #24658 (comment)

I don't think we're doing updates to the 1.26 branch anymore so I think this might just be a known issue in a past release of NumPy?

[ngoldbaum]

Ping @seberg since you might have ideas to ease the C API transition for mkl_random.

[seberg]

Either force compiling with 2.0 or copy a custom definintion of the struct over (e.g. from the Cython 2 .pyd.

[oleksandr-pavlyk]

This is what I ended up doing https://github.com/IntelPython/mkl_random/blob/792b3f7181ef2970a71da26a1b1a0164926fd7cc/mkl_random/src/numpy_multiiter_workaround.h#L46-L78

[seberg]

Going to close. Might have been nice to backport, but it was largely a Cython 2->3 thing, IIRC, and didn't seem used a lot even then.

The workarounds are not that bad (otherwise I am sure we would have backported it).

[seberg]

@oleksandr-pavlyk one thing I just realized, I think your error above is a fluke. If you are using these macro's the Cython .pyx was generated with NumPy 2.0.

That is a problem I have seen before: Cython doesn't notice that regeneration is required due to a dependency .pyd change (and those files can be hidden deep in the project and/or build folders).

The problem that does exist is that if you compile with Cython 3 (which you should), you cannot access the fields of them.

TL;DR: Unless you are using the broadcast attributes in your code, this problem is a build issue and can be fixed by thoroughly cleaning up the generated .c/.cpp files.