The method PyUFuncOverride_GetNonDefaultArrayUfunc is expensive on numpy scalars because these objects do not have a __array_ufunc__ set and for a missing attribute lookup cpython generates an exception that is later cleared by numpy. This is a performance bottleneck, see #21455.

An issue has been submitted to cpython (python/cpython#92216). But even if this is addressed in cpython, it will take untill python 3.12+ before this will be useable by numpy.

As an alternative solution, this PR adds a fast path to PyUFuncOverride_GetNonDefaultArrayUfunc to determine whether an object is a numpy scalar. Some remarks:

• Currently the check is only on PyDoubleArrType_Type and PyIntArrType_Type. Perhaps we can check all types easily? Or select a different subset of all numpy scalar types.
• Subclasses of the numpy scalars are not handled
• An alternative would be to add __array_ufunc__ to the numpy scalars, but this is not backwards compatible (PERF: Improve performance of special attribute lookups #21423 (comment))

Microbenchmark on np.sqrt(np.float64(1.1)):

main: 0.54
PR: 0.40

import numpy as np
import math
import time
print(np.__version__)

w=np.float64(1.1)
#w=1.1

niter=600_000
for kk in range(3):
    t0=time.perf_counter()
    for ii in range(niter):
        _=np.sqrt(w)
    dt=time.perf_counter()-t0
    print(f'loop {kk}: {dt}')