BUG: spatial.transform.Rotation.from_rotvec slow on arm64

[mhavu]

Describe your issue.

I am running macOS 12.4 on Apple M1 Pro. I have a conda environment that runs osx-arm64 code natively and another one that runs osx-64 (x86-64) code via Rosetta 2 translation. Both environments use the following package versions from conda-forge:

• python 3.9.13
• numpy 1.23.2
• scipy 1.9.0
• libopenblas 0.3.21
• liblapack 3.9.0.

The arm64 native version of scipy.spatial.transform.Rotation.from_rotvec is up to four times slower than the translated version.

Reproducing Code Example

import numpy as np
from scipy.spatial.transform import Rotation

rotvec = np.random.rand(10000, 3)
%timeit rot = Rotation.from_rotvec(rotvec)
# osx-arm64: 407 µs ± 106 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
# osx-64:    145 µs ± 88.2 ns per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

Error message

There are no error messages.

SciPy/NumPy/Python version information

1.9.0 1.23.2 sys.version_info(major=3, minor=9, micro=13, releaselevel='final', serial=0)