cython如何實現更好的循環速度性能？

Question

我已經在python中啟動了一個項目，該項目主要由循環組成。 幾天前，我讀到了有關cython的信息，它可以幫助您通過靜態鍵入來獲得更快的代碼。 我開發了這兩個函數來檢查性能（一個在python中，另一個在cython中）：

import numpy as np
from time import clock

size = 11
board = np.random.randint(2, size=(size, size))

def py_playout(board, N):
    black_rave = []
    white_rave = []
    for i in range(N):
        for x in range(board.shape[0]):
            for y in range(board.shape[1]):
                if board[(x,y)] == 0:
                    black_rave.append((x,y))
                else:
                    white_rave.append((x,y))
    return black_rave, white_rave

cdef cy_playout(board, int N):
    cdef list white_rave = [], black_rave = []
    cdef int M = board.shape[0], L = board.shape[1]
    cdef int i=0, x=0, y=0
    for i in range(N):
        for x in range(M):
            for y in range(L):
                if board[(x,y)] == 0:
                    black_rave.append((x,y))
                else:
                    white_rave.append((x,y))
    return black_rave, white_rave

我畢竟使用下面的代碼來測試性能：

t1 = clock()
a = playout(board, 1000)
t2 = clock()
b = playout1(board, 1000)
t3 = clock()

py = t2 - t1
cy = t3 - t2
print('cy is %a times better than py'% str(py / cy))

但是我沒有發現任何明顯的改進。 我還沒有使用Typed-Memoryviews。 有人可以提出有用的解決方案來提高速度，還是可以幫助我使用typed-memoryview重寫代碼？

Answer 1

沒錯，沒有在cython函數中的board參數中添加類型，加速不是很多：

%timeit py_playout(board, 1000)
# 321 ms ± 19.3 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
%timeit cy_playout(board, 1000)
# 186 ms ± 541 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

但這仍然快了兩倍。 通過添加類型，例如

cdef cy_playout(int[:, :] board, int N):
    # ...

# or if you want explicit types:
# cimport numpy as np
# cdef cy_playout(np.int64_t[:, :] board, int N):  # or np.int32_t

它快得多（快了將近十倍）：

%timeit cy_playout(board, 1000)
# 38.7 ms ± 1.84 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

我還使用了timeit （可以使用IPython的魔術%timeit ）來獲得更准確的計時。

---

請注意，您也可以使用numba來實現極大的加速，而無需任何其他靜態類型：

import numba as nb

nb_playout = nb.njit(py_playout)  # Just decorated your python function

%timeit nb_playout(board, 1000)
# 37.5 ms ± 154 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

Answer 2

我實現了一個運行速度更快的功能。 我只是將black_rave和white_rave聲明為memoryviews並將它們放入返回值中：

cdef tuple cy_playout1(int[:, :] board, int N):
    cell_size = int((size ** 2) / 2) + 10
    cdef int[:, :] black_rave = np.empty([cell_size, 2], dtype=np.int32)
    cdef int[:, :] white_rave = np.empty([cell_size, 2], dtype=np.int32)

    cdef int i, j, x, y, h
    i, j = 0, 0
    cdef int M,L
    M = board.shape[0]
    L = board.shape[1]
    for h in range(N):
        for x in range(M):
            for y in range(L):
                if board[x,y] == 0:
                    black_rave[i][0], black_rave[i][1] = x, y
                    i += 1
                elif board[x,y] == 1:
                    white_rave[j][0], white_rave[j][1] = x, y
                    j += 1
        i = 0
        j = 0

    return black_rave[:i], white_rave[:j]

這是速度測試結果：

%timeit py_playout(board, 1000)
%timeit cy_playout(board, 1000)
%timeit cy_playout1(board, 1000)
# 1 loop, best of 3: 200 ms per loop
# 100 loops, best of 3: 9.26 ms per loop
# 100 loops, best of 3: 4.88 ms per loop