Question

我制作了一个简短的示例程序来测试与线程模块结合的numpy函数。因此，我用不同大小的随机数组调用了np.dot函数。

总体调查结果基本符合预期。方法 test_serial 比方法 test_threading 快一点（因为开销更少？）。

尽管如此，我不明白为什么 threaded 点操作为什么一个接一个地执行，而不是同时执行。我确保内存使用不会有问题。

我期望的是

thread_0启动（所有用于np.dot（arg_1，arg_2）的资源
线程_1启动（资源在np.dot（arg_1，arg_2）和np.dot（arg_3，arg_4）之间分配
依此类推
线程3和4完成（因为我为它们使用了较小的数组）
线程0、1、2完成

似乎numpy一次只能执行一个点运算。有人可以确认吗？进一步的测试表明，np.multiply操作不是这种情况。

输出 np.dot ：

--- start threaded test ---
--- start threads ---
start: 0
start: 1
start: 2
start: 3
start: 4
--- start threads done ---
-8.263092041015625
--- end ---  0
--- join thread ---
-16.34356927871704
--- end ---  1
--- join thread ---
-24.071669578552246
--- end ---  2
--- join thread ---
-23.95636224746704
--- end ---  3
--- join thread ---
-23.962880611419678
--- end ---  4
--- join thread ---
-24.32594323158264
--- end threaded test ---

输出乘以 np ：

--- start threaded test ---
--- start threads ---
start: 0
start: 1
start: 2
start: 3
start: 4
--- start threads done ---
-0.007001638412475586
--- end ---  3
-0.007020235061645508
--- end ---  4
-0.3390059471130371
--- end ---  1
-0.3525271415710449
--- end ---  2
-0.3660862445831299
--- end ---  0
--- join thread ---
--- join thread ---
--- join thread ---
--- join thread ---
--- join thread ---
-0.36977648735046387
--- end threaded test ---

from threading import Thread, Lock
print_access = Lock()
import gc

def test_f1(arg_1, arg_2,out, num):
    with print_access:
        print("start:", num)
    xx = time.time()
    np.dot(arg_1, arg_2,out)
    with print_access:
        print(xx-time.time())
        print("--- end --- ", num)

def test_f2(arg_1, arg_2,out,  num):
    with print_access:
        print("start:", num)
    xx = time.time()
    np.multiply(arg_1, arg_2,out)
    with print_access:
        print(xx-time.time())
        print("--- end --- ", num)

def test_f3(arg_1, arg_2, out, num):
    with print_access:
        print("start:", num)
    xx = time.time()
    np.matmul(arg_1, arg_2,out)
    with print_access:
        print(xx-time.time())
        print("--- end --- ", num)


def test_serial(arg_1, arg_2, arg_3, arg_4,arg_5,arg_6,arg_7,arg_8, arg_9,arg_10,out_1, out_2, out_3, out_4, out_5):
    test_f2(arg_1,arg_2,out_1, 0)
    test_f2(arg_3,arg_4,out_2, 1)
    test_f2(arg_5,arg_6,out_3, 2)
    test_f2(arg_7,arg_8,out_4, 3)
    test_f2(arg_9,arg_10,out_5, 4)

def test_thread(arg_1, arg_2, arg_3, arg_4,arg_5,arg_6,arg_7,arg_8, arg_9,arg_10,out_1, out_2, out_3, out_4, out_5):
    t_1 = Thread(target = test_f2, args = (arg_1,arg_2,out_1,0))
    t_2 = Thread(target = test_f2, args = (arg_3,arg_4,out_2,1))
    t_3 = Thread(target = test_f2, args = (arg_5,arg_6,out_3,2))
    t_4 = Thread(target = test_f2, args = (arg_7,arg_8,out_4,3))
    t_5 = Thread(target = test_f2, args = (arg_9,arg_10,out_5,4))
    print("--- start threads ---")
    t_1.start()
    t_2.start()
    t_3.start()
    t_4.start()
    t_5.start()
    print("--- start threads done ---")
    t_1.join()
    print("--- join thread ---")
    t_2.join()
    print("--- join thread ---")
    t_3.join()
    print("--- join thread ---")
    t_4.join()
    print("--- join thread ---")
    t_5.join()
    print("--- join thread ---")

if __name__ == "__main__":
    x = 10000
    y = 1000
    array_1 = np.random.rand(x,x)
    array_2 = np.random.rand(x,x)

    array_3 = np.random.rand(x,x)
    array_4 = np.random.rand(x,x)
    array_5 = np.random.rand(x,x)
    array_6 = np.random.rand(x,x)
    array_7 = np.random.rand(y,y)
    array_8 = np.random.rand(y,y)
    array_9 = np.random.rand(y,y)
    array_10 = np.random.rand(y,y)

    out_1 = np.random.rand(x,x)
    out_2 = np.random.rand(x,x)
    out_3 = np.random.rand(x,x)
    out_4 = np.random.rand(y,y)
    out_5 = np.random.rand(y,y)


    print("--- start ---")
    xx = time.time()
    test_serial(array_1, array_2,array_3, array_4,array_5, array_6,array_7, array_8,array_9, array_10, \
                out_1,out_2,out_3, out_4, out_5)
    print(xx-time.time())
    print("--- end ---")

    time.sleep(0.5)
    gc.collect()
    time.sleep(0.5)

    print("--- start ---")
    xx = time.time()
    test_thread(array_1, array_2,array_3, array_4,array_5, array_6,array_7, array_8,array_9, array_10, \
                out_1,out_2,out_3, out_4, out_5)
    print(xx-time.time())
    print("--- end ---")
```

numpy.dot函数不能同时执行

0 个答案: