Numpy array initialization

This benchmark compares the execution time of several numpy.ndarray initialization routines.

import numpy as np
import pyquickbench

dtypes_dict = {
    "float32" : np.float32,
    "float64" : np.float64,
}

def zeros(n, real_dtype):
    np.zeros((n), dtype=dtypes_dict[real_dtype])

def ones(n, real_dtype):
    np.ones((n), dtype=dtypes_dict[real_dtype])

def empty(n, real_dtype):
    np.empty((n), dtype=dtypes_dict[real_dtype])

def full(n, real_dtype):
    np.full((n), 0., dtype=dtypes_dict[real_dtype])

all_funs = [
    zeros   ,
    ones    ,
    empty   ,
    full    ,
]

all_args = {
    "n" : np.array([2**n for n in range(0, 30)]),
    "real_dtype" : ["float32", "float64"],
}

def setup(n, real_dtype):
    return {'n': n, 'real_dtype':real_dtype}

all_timings = pyquickbench.run_benchmark(
    all_args                ,
    all_funs                ,
    setup = setup           ,
    filename = filename     ,
    StopOnExcept = True     ,
    ShowProgress = True     ,

)

plot_intent = {
    "n" : 'points'                              ,
    "real_dtype" : 'curve_linestyle'            ,
    pyquickbench.fun_ax_name :  'curve_color'   ,
}

pyquickbench.plot_benchmark(
    all_timings                             ,
    all_args                                ,
    all_funs                                ,
    plot_intent = plot_intent               ,
    show = True                             ,
)

While these measurement seem surprizing, they are explained in the numpy documentation :

Be mindful that large arrays created with np.empty or np.zeros might not be allocated in physical memory until the memory is accessed. […] One can force pagefaults to occur in the setup phase either by calling np.ones or arr.fill(value) after creating the array.