Benchmark numpy and scipy decompositions

benchmark_decompositions.py

import time

import scipy as sp
import numpy as np


def benchmark(decomp, shapes, *, n_experiments=10):
    def get_fn_name(fn):
        return f"{fn.__module__}.{fn.__name__}"

    print(
        f"{'method':^14} {'shape':>16} " f"{'avg':^8} {'std':^8} {'min':^8} {'max':^8}"
    )
    for i, shape in enumerate(shapes):
        mat = np.random.random(size=shape)

        timings = []
        for _ in range(n_experiments):
            start_time = time.perf_counter()
            _ = decomp(mat)
            end_time = time.perf_counter()
            timings.append(end_time - start_time)

        avg, std = np.mean(timings), np.std(timings)
        min_, max_ = np.min(timings), np.max(timings)

        fn_name = f"{decomp.__module__}.{decomp.__name__}"
        print(
            f"{fn_name if i == 0 else ' ':^14} {str(shape):>16} "
            f"{avg:^8.2f} {std:^8.2f} {min_:^8.2f} {max_:^8.2f}"
        )


shapes = [
    (10, 10),
    (100, 100),
    (500, 500),
    (1000, 1000),
    (5000, 1000),
]

benchmark(np.linalg.svd, shapes)
benchmark(sp.linalg.svd, shapes)

Results on AMD Ryzen 5 5625U with 6 cores (12 threads):

Timings are reported in seconds. Each experiment was run 10 times.

method	shape	avg	std	min	max
numpy.linalg.svd	(10, 10)	0.00	0.00	0.00	0.00
	(100, 100)	0.00	0.00	0.00	0.01
	(500, 500)	0.05	0.01	0.04	0.07
	(1000, 1000)	0.23	0.02	0.21	0.27
	(5000, 1000)	4.07	0.15	3.90	4.39
scipy.linalg.svd	(10, 10)	0.00	0.00	0.00	0.00
	(100, 100)	0.00	0.00	0.00	0.01
	(500, 500)	0.04	0.00	0.04	0.05
	(1000, 1000)	0.20	0.01	0.19	0.23
	(5000, 1000)	3.92	0.14	3.80	4.25

scipy v1.11.4 and numpy v1.26.3