einsum is faster if the tail survives

compare.py

import perfplot
import numpy


def setup(n):
    a = numpy.random.rand(n, 3)
    b = numpy.ascontiguousarray(a.T)
    return a, b


def einsum_i(data):
    return numpy.einsum("ij,ij->i", data[0], data[0])


def einsum_j(data):
    return numpy.einsum("ij,ij->j", data[1], data[1])


perfplot.show(
    setup=setup,
    kernels=[einsum_i, einsum_j],
    n_range=[2 ** k for k in range(26)],
    relative_to=1,
)