PERF: Micro optimize find_common_type

[eendebakpt]

The find_common_type is a (small) performance bottleneck in one of the scipy algoritms. This PR optimizes the method:

A benchmark (extracted from what is called from scipy):

import numpy as np
from numpy.core.numerictypes import find_common_type

array_types = [np.dtype('float64'), np.dtype('float64')]
scalar_types = []

find_common_type(array_types, scalar_types)

Results of %timeit find_common_type(array_types, scalar_types).

• Main: 8.37 µs ± 500 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
• PR: 1.09 µs ± 57.6 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)
• PR (only first commit): 1.81 µs ± 313 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)

The optmizations are:

• Cache cast of __test_types to dtype
• Make elements of dtypelist unique. This could make other calls a bit slower, but the list(set(dtypelist)) is fast and it makes cases with duplicate dtypes faster
• Fast path in find_common_type for empty arguments.

[seberg]

Sorry, but I am pretty firmly 👎 because this function should probably never be used and I would like to burn this function in a volcano, see also gh-21703. np.promote_types or np.result_typeshould be used instead.

EDIT: But, the code change itself looks simple and fine. So if anyone other core dev likes this independent of the fact that the function should be deprecated, just go ahead.

[mattip]

Can we suggest a fix to scipy to avoid this function that is also more performant?

[mattip]

Ahh, you already did :). Since the alternatives are more performant, I think we should close this and deprecate find_common_type

[eendebakpt]

Ahh, you already did :). Since the alternatives are more performant, I think we should close this and deprecate find_common_type

Since there is a PR at scipy I am fine with closing this. The only reason to merge would be if (even when deprecating) the find_common_type would stick around a long time to support external pacakges.

[eendebakpt]

Closing this PR. The issue has been solved in scipy.