Benchmark script for numpy#5509

bench_inf_norm.py

from __future__ import print_function
import numpy as np
from numpy.testing import measure

def bench_inf_norm_1d(dtype='d', strided=False, with_nan=False):
    vec_size = [10, 100, 1000, 10000, 100000, 1000000 ]
    repeat = [100000, 100000, 100000, 40000, 7500, 1000]

    call_both = lambda x: abs(x).max()
    call_norm = lambda x: np.linalg.norm(x, np.inf)
    call_max = np.maximum.reduce
    max_abs = np.max_abs.reduce
    print(np.dtype(dtype))
    for size, rep in zip(vec_size, repeat):
        if strided: 
            v = np.random.randn(3*size).astype(dtype)[::3]
        else:
            v = np.random.randn(size).astype(dtype)

        if with_nan:
            v[2] = np.nan
        t_both = measure('call_both(v)', rep)
        t_norm = measure('call_norm(v)', rep)
        t_max = measure('call_max(v)', rep)
        t_ma = measure('max_abs(v)', rep)

        print('{0:7} {1:6.2f} {2:6.2f} {3:6.2f} {4:6.2f} {5:6} {6:6.2f}'.format(size, t_norm, t_both, t_ma, t_max, rep, t_both/t_ma))

def bench_inf_norm_2d(dtype):
    size_rep = np.array([[3, 10, 50000],
                         [3, 100, 50000],
                         [3, 1000, 25000],
                         [3, 10000, 5000],
                         [100, 10, 10000],
                         [100, 100, 20000],
                         [100, 1000, 2500],
                         [100, 10000, 250],
                         [1000, 10, 10000],
                         [1000, 100, 2500],
                         [1000, 1000, 250],
                         [1000, 10000, 30],
                         [10000, 10, 1000],
                         [10000, 100, 250],
                         [10000, 1000, 25],
                         [10000, 10000, 3]])
    
    vec_size = [3, 16, 100, 500, 1000, 10000]
    n_vec = [1, 5, 10, 100, 1000, 10000]

    call_both = lambda x, a: np.absolute(x).max(axis=a)
    call_norm = lambda x, a: np.linalg.norm(x, np.inf, axis=a)
    call_max = lambda x, a: np.maximum.reduce(x, axis=a)
    max_abs = lambda x, a: np.max_abs.reduce(x, axis=a)
    print(np.dtype(dtype))

    for size0, size1, rep in size_rep:
        v = np.random.randn(size0, size1).astype(dtype)
        t_both0 = measure('call_both(v, 0)', rep)
        t_norm0 = measure('call_norm(v, 0)', rep)
        t_max0 = measure('call_max(v, 0)', rep)
        t_ma0 = measure('max_abs(v, 0)', rep)
        print('{0:6} ({1:6}, {2:6}) {3:6.2f} {4:6.2f} {5:6.2f} {6:6.2f} {7:6} {8:6.2f}'.format(0, size0, size1, t_norm0, t_both0, t_ma0, t_max0, rep, t_both0/t_ma0))
        t_both1 = measure('call_both(v, 1)', rep)
        t_norm1 = measure('call_norm(v, 1)', rep)
        t_max1 = measure('call_max(v, 1)', rep)
        t_ma1 = measure('max_abs(v, 1)', rep)

        print('{0:6} ({1:6}, {2:6}) {3:6.2f} {4:6.2f} {5:6.2f} {6:6.2f} {7:6} {8:6.2f}'.format(1, size0, size1, t_norm1, t_both1, t_ma1, t_max1, rep, t_both1/t_ma1))