BUG: CPU detection fails in windows VM running on ARM Mac

[ngoldbaum]

Describe the issue:

When I try to build numpy in the default configuration on an ARM Mac running an ARM Windows 11 VM via qemu, I get the build error below.

Building with spin build --clean -- -Ddisable-optimization=True lets me build though so this isn't a showstopper for my debugging session.

Reproduce the code example:

N/A

Error message:

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "C:\Users\Nathan Goldbaum\numpy\build-install\usr\Lib\site-packages\numpy\__init__.py", line 114, in <module>
    from numpy.__config__ import show as show_config
  File "C:\Users\Nathan Goldbaum\numpy\build-install\usr\Lib\site-packages\numpy\__config__.py", line 4, in <module>
    from numpy._core._multiarray_umath import (
  File "C:\Users\Nathan Goldbaum\numpy\build-install\usr\Lib\site-packages\numpy\_core\__init__.py", line 23, in <module>
    from . import multiarray
  File "C:\Users\Nathan Goldbaum\numpy\build-install\usr\Lib\site-packages\numpy\_core\multiarray.py", line 10, in <module>
    from . import overrides
  File "C:\Users\Nathan Goldbaum\numpy\build-install\usr\Lib\site-packages\numpy\_core\overrides.py", line 8, in <module>
    from numpy._core._multiarray_umath import (
RuntimeError: NumPy was built with baseline optimizations:
(NEON NEON_FP16 NEON_VFPV4 ASIMD) but your machine doesn't support:
(NEON).

Python and NumPy Versions:

python 3.12.1 installed via python.org installer, numpy main.

Runtime Environment:

numpy doesn't install so I can't get this far.

Context for the issue:

No response

[mattip]

Weird, I thought qemu on M1 supported NEON. Maybe this is a windows-arm64 problem? What happens if you run an aarch64 VM in qemu, do you get NEON?

[ngoldbaum]

It compiles on an aarch64 Ubuntu VM, here's the show_runtime() output, not sure if NEON not being in 'found' means we didn't get NEON.

>>> import numpy; print(numpy.show_runtime())
[{'numpy_version': '2.0.0.dev0+git20240117.448e4da',
'python': '3.11.7 (main, Dec 12 2023, 21:30:01) [GCC 11.4.0]',
'uname': uname_result(system='Linux', node='ubuntuvm', release='5.15.0-91-generic', version='#101-Ubuntu SMP Tue Nov 14 13:29:11 UTC 2023', machine='aarch64')},
{'simd_extensions': {'baseline': ['NEON', 'NEON_FP16', 'NEON_VFPV4', 'ASIMD'],
'found': ['ASIMDHP', 'ASIMDFHM'],
'not_found': ['SVE']}}]
None

[mattip]

Successfully importing NumPy with this line means these CPU features are required and found:

simd_extensions': {'baseline': ['NEON', 'NEON_FP16', 'NEON_VFPV4', 'ASIMD'],

I wonder if npy__cpu_init_features() has a branch for windows + arm64. It seems the compiler defines macros differently to linux, so this check might be failing.

It is hard to get python + numpy running under the visual studio debugger to step through the code. Could you add some fprintf() debug statements and recompile to see which code is active there?