3x3-matvec vs manual matvec

matvec-vs-manual.py

import perfplot
import numpy as np
import npx


A = np.array([[0.0, 1.0, 0.0], [125 / 29, -125 / 29, 0.0], [0.0, 50 / 29, -50 / 29]])


def manual(data):
    fx, fy, fz = data
    return np.array([fy, 125 / 29 * (fx - fy), 50 / 29 * (fy - fz)])


def at(data):
    return A @ data


def np_dot(data):
    return np.dot(A, data)


def npx_dot(data):
    return npx.dot(A, data)


b = perfplot.bench(
    setup=lambda n: np.random.rand(3, n),
    kernels=[manual, at, npx_dot, np_dot],
    n_range=[2 ** k for k in range(25)],
)
b.save("out.png")
b.show()

The proper matvec is almost always faster: