ENH: faster chebyshev distance with ndarray.max

[lozanodorian]

Is your feature request related to a problem? Please describe.

Context

The Chebyshev distance $\mathrm{d}(u, v) \overset{\text{def}}{=} \underset{1\leq i \leq n}{\max}\left|u_i-v_i\right|$ can be computed in scipy.spatial.distance by using chebyshev(u, v).
The last line of this chebyshev function is:

return max(abs(u - v))

where u and v are two NDArrays.

Benchmark

A basic benchmark shows that for len(a) > 13, the use of max(a) is less efficient than the use of np.max(a) or the use of a.max().

In this benchmark:

• a = np.random.random(n),
• each point quantifies the time of 20 computations of the considered max function.

Describe the solution you'd like.

Replace:

return max(abs(u - v))

By:

return abs(u - v).max()

At least if len(u) > something.

Describe alternatives you've considered.

No response

Additional context (e.g. screenshots, GIFs)

Code of the benchmark

The code of the benchmark is short, we include it in this additional context:

import perfplot
import numpy as np

perfplot.show(
    setup = lambda n: np.random.random(n),
    kernels=[
            lambda a : [max(a) for _ in range(20)],
            lambda a : [np.max(a) for _ in range(20)],
            lambda a : [a.max() for _ in range(20)],
            ],
    labels=['max', 'np.max', 'ndarray.max'],
    n_range=[int(1.2**k) for k in range(10, 92, 2)],
    xlabel='len(a)',
    title='Comparison of time for max, np.max, and ndarray.max of a NumPy array'
    )

Additional remark

Note that the differences of time for the abs functions/method are negligible (ndarray.abs is here ndarray.__abs__).

[tupui]

Thank you @lozanodorian. Want to make a PR?

np.max(a) should be preferred to make it future proof to work with the Array API standard as oppose to a.max().

[lozanodorian]

Thank you @tupui, I'm going to make a PR.