Python GPU提高效率 python 详解怎样使用GPU大幅提高效率

cupy我觉得可以理解为cuda for numpy，安装方式pip install cupy，假设

import numpy as np
import cupy as cp

那么对于np.XXX一般可以直接替代为cp.XXX。

其实numpy已经够快了，毕竟是C写的，每次运行的时候都会尽其所能地调用系统资源。为了验证这一点，我们可以用矩阵乘法来测试一下：在形式上通过多线程并发、多进程并行以及单线程的方式，来比较一下numpy的速度和对资源的调度情况，代码为

# th_pr_array.py
from threading import Thread
from multiprocessing import Process
from time import time as Now
import numpy as np
import sys

N = 3000

def MatrixTest(n,name,t):
    x = np.random.rand(n,n)
    x = x@x
    print(f"{name} @ {t} : {Now()-t}")

def thTest():
    t = Now()
    for i in range(5):
        Thread(target=MatrixTest,args=[N,f'th{i}',t]).start()

def prTest():
    t = Now()
    for i in range(5):
        Process(target=MatrixTest,args=[N,f'pr{i}',t]).start()

if __name__=="__main__":
    if sys.argv[1]=="th":
        thTest()
    elif sys.argv[1]=="pr":
        prTest()
    else:
        t = Now()
        for i in range(5):
            MatrixTest(N,"single",t)

运行结果为

(base) E:\Documents\00\1108>python th_pr_numpy.py th
th0 @ 1636357422.3703225 : 15.23965334892273
th1 @ 1636357422.3703225 : 17.726242780685425
th2 @ 1636357422.3703225 : 19.001763582229614
th3 @ 1636357422.3703225 : 19.06676197052002
th4 @ 1636357422.3703225 : 19.086761951446533

(base) E:\Documents\00\1108>python th_pr_numpy.py pr
pr3 @ 1636357462.4170427 : 4.031360864639282
pr0 @ 1636357462.4170427 : 4.55387806892395
pr1 @ 1636357462.4170427 : 4.590881824493408
pr4 @ 1636357462.4170427 : 4.674877643585205
pr2 @ 1636357462.4170427 : 4.702877759933472

(base) E:\Documents\00\1108>python th_pr_numpy.py single
single @ 1636357567.8899782 : 0.36359524726867676
single @ 1636357567.8899782 : 0.8137514591217041
single @ 1636357567.8899782 : 1.237830400466919
single @ 1636357567.8899782 : 1.683635950088501
single @ 1636357567.8899782 : 2.098794937133789

所以说在numpy中就别用python内置的并行和并发了，反而会称为累赘。而且这么一比更会印证numpy的强大性能。

但在cupy面前，这个速度会显得十分苍白，下面连续5次创建5000x5000的随机矩阵并进行矩阵乘法，

#np_cp.py
import numpy as np
import cupy as cp
import sys
from time import time as Now

N = 5000

def testNp(t):
    for i in range(5):
        x = np.random.rand(N,N)
        x = x@x
    print(f"np:{Now()-t}")

def testCp(t):
    for i in range(5):
        x = cp.random.rand(N,N)
        x = x@x
    print(f"cp:{Now()-t}")


if __name__ == "__main__":
    t = Now()
    if sys.argv[1] == 'np':
        testNp(t)
    elif sys.argv[1]=='cp':
        testCp(t)

最后的结果是

(base) E:\Documents\00\1108>python np_cp.py np
np:8.914457082748413

(base) E:\Documents\00\1108>python np_cp.py cp
cp:0.545649528503418

而且非常霸道的是，当矩阵维度从5000x5000升到15000x15000后，cupy的计算时间并没有什么变化，充其量是线性增长，毕竟只要缓存吃得下，无论多么大的矩阵，乘法数也无非是按行或者按列增加而已。