NUMBA - 如何使用“cuda”目标在@guvectorize 中生成随机数？

Question

在这个（愚蠢的）示例中，我试图通过计算 (0, 1) x (0, 1) 中落入单位圆的随机选择点的数量来计算 pi。

@guvectorize(['void(float64[:], int32, float64[:])'], '(n),()->(n)', target='cuda')
def guvec_compute_pi(arr, iters, res):
    n = arr.shape[0]
    for t in range(n):
        inside = 0
        for i in range(iters):
            x = np.random.random()
            y = np.random.random()
            if x ** 2 + y ** 2 <= 1.0:
               inside += 1
        res[t] = 4.0 * inside / iters

编译时弹出这个异常：

numba.errors.UntypedAttributeError: Failed at nopython (nopython frontend)
Unknown attribute 'random' of type Module(<module 'numpy.random' from '...'>)
File "scratch.py", line 34
[1] During: typing of get attribute at /.../scratch.py (34)

我天真地认为使用here 描述的 RNG 可以解决问题。我修改后的代码如下：

@guvectorize(['void(float64[:], int32, float64[:])'], '(n),()->(n)', target='cuda')
def guvec_compute_pi(arr, iters, res):
    n = arr.shape[0]
    rng = create_xoroshiro128p_states(n, seed=1)
    for t in range(n):
        inside = 0
        for i in range(iters):
            x = xoroshiro128p_uniform_float64(rng, t)
            y = xoroshiro128p_uniform_float64(rng, t)
            if x ** 2 + y ** 2 <= 1.0:
                inside += 1
        res[t] = 4.0 * inside / iters

但是会弹出类似的错误：

numba.errors.TypingError: Failed at nopython (nopython frontend)
Untyped global name 'create_xoroshiro128p_states': cannot determine Numba type of <class 'function'>
File "scratch.py", line 28

当我尝试更改为target='parallel' 时，使用numpy.random.random 的原始代码无论nopython=True 与否都可以正常工作。是什么导致了target='cuda' 的问题，有没有办法在@guvectorize-d 块中获取随机数？

Answer 1

函数 create_xoroshiro128p_states 旨在在 CPU 上运行，如 Numba 文档中的此示例所示，重复如下：

from __future__ import print_function, absolute_import

from numba import cuda
from numba.cuda.random import create_xoroshiro128p_states, 
xoroshiro128p_uniform_float32
import numpy as np

@cuda.jit
def compute_pi(rng_states, iterations, out):
    """Find the maximum value in values and store in result[0]"""
    thread_id = cuda.grid(1)

    # Compute pi by drawing random (x, y) points and finding what
    # fraction lie inside a unit circle
    inside = 0
    for i in range(iterations):
        x = xoroshiro128p_uniform_float32(rng_states, thread_id)
        y = xoroshiro128p_uniform_float32(rng_states, thread_id)
        if x**2 + y**2 <= 1.0:
            inside += 1

    out[thread_id] = 4.0 * inside / iterations

threads_per_block = 64
blocks = 24
rng_states = create_xoroshiro128p_states(threads_per_block * blocks, seed=1)
out = np.zeros(threads_per_block * blocks, dtype=np.float32)

compute_pi[blocks, threads_per_block](rng_states, 10000, out)
print('pi:', out.mean())

它生成一个随机初始化数据数组，使 GPU 上的随机数生成独立于线程。该数据最终在设备端，这有点令人困惑。但它允许您将随机状态数据传递给您的 GPU 内核。