在这个简单的例子中,为什么 Matlab 看起来比 Python 慢得多

Why does Matlab seem so much slower than Python in this simple case

我最近问a question on the optimization of a mask function in Matlab。我得到了两个对我有很大帮助的答案,但根据我的时间安排,似乎所有 Matlab 解决方案都比其中一个 Numpy 解决方案慢得多。不同功能的代码可以在我之前的问题中找到,但为了让您了解我在做什么,我给出了 Numpy "loop" 解决方案,这当然不是最快的,但可能是最简单的阅读:

def dealiasing2d(where, data):
    nk, n0, n1 = data.shape
    for i0 in xrange(n0):
        for i1 in xrange(n1):
            if where[i0, i1]:
                data[:, i0, i1] = 0.

我获得(使用 Matlab R2014b 和 "basic" Numpy 1.9.1 与 Blas 和 Lapack 链接)(n0 = n1 = N):

N = 500 ; nk = 500:
Method          | time (s) | normalized      
----------------|----------|------------
Numpy           |    0.05  |     1.0
Numpy loop      |    0.05  |     1.0
Matlab find     |    0.74  |    14.8
Matlab bsxfun2  |    0.76  |    15.2
Matlab bsxfun   |    0.78  |    15.6
Matlab loop     |    0.78  |    15.6
Matlab repmat   |    0.89  |    17.8

N = 500 ; nk = 100:
Method          | time (s) | normalized      
----------------|----------|------------
Numpy           |    0.01  |     1.0
Numpy loop      |    0.03  |     3.0
Matlab find     |    0.15  |    13.6
Matlab bsxfun2  |    0.15  |    13.6
Matlab bsxfun   |    0.16  |    14.5
Matlab loop     |    0.16  |    14.5
Matlab repmat   |    0.18  |    16.4

N = 2000 ; nk = 10:
Method          | time (s) | normalized      
----------------|----------|------------
Numpy           |    0.02  |     1.0
Matlab find     |    0.23  |    13.8
Matlab bsxfun2  |    0.23  |    13.8
Matlab bsxfun   |    0.26  |    15.6
Matlab repmat   |    0.28  |    16.8
Matlab loop     |    0.34  |    20.4
Numpy loop      |    0.42  |    25.1

在我看来,这些结果很奇怪。对我来说,Numpy 和 Matlab 在科学计算方面非常相似,所以性能应该相似,而这里有超过 10 倍!所以我的第一个猜测是我比较这两种语言的方式有问题。另一种可能是我的 Matlab 设置有问题,但我不明白为什么。或者 Matlab 和 Numpy 之间真正的深层区别?

任何人都可以对这些函数计时来验证这些结果吗?你知道为什么在这个简单的例子中 Matlab 看起来比 Python 慢得多吗?

为了给 Matlab 函数计时,我使用了一个文件:

N = 500;
n0 = N;
n1 = N;
nk = 500;
disp(['N = ', num2str(N), ' ; nk = ', num2str(nk)])

where = false([n1, n0]);
where(1:100, 1:100) = 1;
data = (5.+1i)*ones([n1, n0, nk]);

disp('time dealiasing2d_loops:')
time = timeit(@() dealiasing2d_loops(where, data));
disp(['     ', num2str(time), ' s'])

disp('time dealiasing2d_find:')
time = timeit(@() dealiasing2d_find(where, data));
disp(['     ', num2str(time), ' s'])

disp('time dealiasing2d_bsxfun:')
time = timeit(@() dealiasing2d_bsxfun(where, data));
disp(['     ', num2str(time), ' s'])

disp('time dealiasing2d_bsxfun2:')
time = timeit(@() dealiasing2d_bsxfun2(where, data));
disp(['     ', num2str(time), ' s'])

disp('time dealiasing2d_repmat:')
time = timeit(@() dealiasing2d_repmat(where, data));
disp(['     ', num2str(time), ' s'])

我用

测量 Python 函数的性能
from __future__ import print_function

import numpy as np
from timeit import timeit, repeat

import_lines = {
    'numpy_bad': 'from dealiasing_numpy_bad import dealiasing2d as dealiasing',
    'numpy': 'from dealiasing_numpy import dealiasing'}
tools = import_lines.keys()

time_approx_one_bench = 5.

setup = """
import numpy as np

N = 500
n0, n1 = N, N
nk = 500
where = np.zeros((n0, n1), dtype=np.uint8)
where[0:100, 0:100] = 1
data = (5.+1j)*np.ones((nk, n0, n1), dtype=np.complex128)
"""
exec(setup)

print('n0 = n1 = {}, nk = {}'.format(N, nk))
print(13*' ' + 'min          mean')
durations = np.empty([len(tools)])

for it, tool in enumerate(tools):
    setup_tool = import_lines[tool] + setup
    duration = timeit(setup_tool + 'dealiasing(where, data)', number=1)
    nb_repeat = int(round((time_approx_one_bench - duration)/duration))
    nb_repeat = max(1, nb_repeat)
    ds = np.array(repeat('dealiasing(where, data)',
                         setup=setup_tool, number=1, repeat=nb_repeat))
    duration = ds.min()
    print('{:11s} {:8.2e} s ; {:8.2e} s'.format(
        tool.capitalize() + ':', duration, ds.mean()))
    durations[it] = duration

fastest = tools[durations.argmin()].capitalize()

durations = durations / durations.min()
print('Durations normalized by the fastest method (' + fastest + '):')
for it, tool in enumerate(tools):
    print('{:11s} {:8.3f}'.format(tool.capitalize() + ':', durations[it]))

我认为这主要与复制 data 变量有关。如果您运行整理东西以便 MATLAB 的写时复制行为对您有利,那么您可以获得相当不错的时间。我用线性索引写了一个简单的版本

function data = dealiasing2d2(where_dealiased, data)
[n1, n2, nk] = size(data);
where_li = find(where_dealiased);
for idx = 1:nk
  offset = n1 * n2 * (idx-1);
  data(where_li + offset) = 0;
end

我 运行 喜欢这个(请注意,重要的是 timed 是一个 函数 而不是允许 data 执行的脚本被重复使用)

function timed
N = 2000;
nk = 10;

where = false([N, N]);
where(1:100, 1:100) = 1;
data = (5.+1j)*ones([N, N, nk]);
tic, data = dealiasing2d2(where,data); toc

这在我的 GLNXA64 机器上运行 0.00437 秒 运行 R2014b。