Numba nopython 模式不能接受二维布尔索引

Question

我正在尝试使用numba（目前我正在使用numba 0.45.1）加速代码，但遇到了布尔索引问题。代码如下：

from numba import njit
import numpy as np

n_max = 1000

n_arr = np.hstack((np.arange(1,3),
                   np.arange(3,n_max, 3)
                   ))

@njit
def func(arr):
    idx =  np.arange(arr[-1]).reshape((-1,1)) < arr -2
    result = np.zeros(idx.shape)
    result[idx] = 10.1
    return result

new_arr = func(n_arr)

运行代码后，我会立即收到以下消息

TypingError: Invalid use of Function(<built-in function setitem>) with argument(s) of type(s): (array(float64, 2d, C), array(bool, 2d, C), float64)
 * parameterized
In definition 0:
    All templates rejected with literals.
In definition 1:
    All templates rejected without literals.
In definition 2:
    All templates rejected with literals.
In definition 3:
    All templates rejected without literals.
In definition 4:
    All templates rejected with literals.
In definition 5:
    All templates rejected without literals.
In definition 6:
    All templates rejected with literals.
In definition 7:
    All templates rejected without literals.
In definition 8:
    TypeError: unsupported array index type array(bool, 2d, C) in [array(bool, 2d, C)]
    raised from C:\Users\User\Anaconda3\lib\site-packages\numba\typing\arraydecl.py:71
In definition 9:
    TypeError: unsupported array index type array(bool, 2d, C) in [array(bool, 2d, C)]
    raised from C:\Users\User\Anaconda3\lib\site-packages\numba\typing\arraydecl.py:71
This error is usually caused by passing an argument of a type that is unsupported by the named function.
[1] During: typing of setitem at C:/Users/User/Desktop/all python file/5.5.5/numba index broadcasting2.py (29)

请注意，最后一行的(29) 对应于第 29 行，即result[idx] = 10.1，这是我尝试为索引为idx（一个二维布尔索引）的结果赋值的行。

---

我想解释一下，在 @njit 中包含该声明 result[idx] = 10.1 是必须的。尽管我想在@njit 中排除这条语句，但我不能，因为这条线正好位于我正在处理的代码的中间。

如果我坚持将赋值语句result[idx] = 10.1 包含在@njit 中，究竟需要更改什么才能使其工作？如果可能的话，我希望看到一些代码示例，其中涉及可以运行的@njit 内的二维布尔索引。

谢谢

Answer 1

Numba 目前不支持二维数组的精美索引。见：

https://numba.pydata.org/numba-doc/dev/reference/numpysupported.html#array-access

但是，您可以通过使用 for 循环显式重写函数而不是依赖广播来获得等效的行为：

from numba import njit
import numpy as np

n_max = 1000

n_arr = np.hstack((np.arange(1,3),
                   np.arange(3,n_max, 3)
                   ))

def func(arr):
    idx =  np.arange(arr[-1]).reshape((-1,1)) < arr -2
    result = np.zeros(idx.shape)
    result[idx] = 10.1
    return result

@njit
def func2(arr):
    M = arr[-1]
    N = arr.shape[0]
    result = np.zeros((M, N))
    for i in range(M):
        for j in range(N):
            if i < arr[j] - 2:
                result[i, j] = 10.1

    return result

new_arr = func(n_arr)
new_arr2 = func2(n_arr)
print(np.allclose(new_arr, new_arr2))  # True

在我的机器上，使用您提供的示例输入，func2 比 func 快​​大约 3.5 倍。