Faster input validation

[dcherian]

I found that ravel_multi_index was taking a lot of time with reductions I tend to run; nD array, 1D group_idx, axis=-1 .

This is an alternate algorithm from https://stackoverflow.com/questions/46256279/bin-elements-per-row-vectorized-2d-bincount-for-numpy

I timed it with this script:

import timeit

import numpy as np
import numpy_groupies as npg


def time_call(method):
    import numpy_groupies as npg

    group_idx = np.repeat([1, 2, 3, 4], repeats=3)
    times = []
    for exp in np.arange(6):
        a = np.ones(
            (
                10 ** exp,
                100,
                12,
            ),
            dtype=np.int32,
        )
        time = timeit.timeit(
            f"npg.utils_numpy.input_validation(group_idx, a, axis=-1, func='sum', method={method!r})",
            number=10,
            globals=locals(),
        )

        times.append(time)

    np.testing.assert_array_equal(
        npg.utils_numpy.input_validation(group_idx, a, axis=-1, func="sum", method="ravel")[0],
        npg.utils_numpy.input_validation(group_idx, a, axis=-1, func="sum", method="offset")[0],
    )
    return times


ravel = time_call("ravel")
offset = time_call("offset")

import matplotlib.pyplot as plt

numel = 12 * 100 * 10**np.arange(len(ravel))
plt.plot(numel, ravel)
plt.plot(numel, offset)
plt.legend(["current npg", "proposed"])
plt.yscale("log")
plt.xscale("log")
plt.grid(True)

It's an ≈ 2x speedup for decent sized arrays