Python cuda内核for循环中的Break语句出现问题
我最近在玩cuda/numba代码。我有一个MxN矩阵,比如说,cumul_a,其中每一行都是累积概率分布。我想通过从均匀随机分布映射样本,从这些累积分布中抽取样本。简单来说,假设从均匀随机分布中抽取的样本为0.3。cuda内核应该选择一行“cumul_a”,并将该行的每个元素从该行的第一个元素开始与0.3进行比较。一旦得到大于0.3的值,内核应该将元素的索引存储在输出参数中,并中断for循环。我无法让这个看似简单的内核工作。break语句是否在内核中引起任何问题? 下面提供了最低限度的工作示例Python cuda内核for循环中的Break语句出现问题,python,cuda,numba,Python,Cuda,Numba,我最近在玩cuda/numba代码。我有一个MxN矩阵,比如说,cumul_a,其中每一行都是累积概率分布。我想通过从均匀随机分布映射样本,从这些累积分布中抽取样本。简单来说,假设从均匀随机分布中抽取的样本为0.3。cuda内核应该选择一行“cumul_a”,并将该行的每个元素从该行的第一个元素开始与0.3进行比较。一旦得到大于0.3的值,内核应该将元素的索引存储在输出参数中,并中断for循环。我无法让这个看似简单的内核工作。break语句是否在内核中引起任何问题? 下面提供了最低限度的工作示例
from __future__ import division
from __future__ import print_function
import numpy as np
from numba import vectorize, cuda, jit
np.set_printoptions(precision=4, suppress=True)
# Number of rows
M = 10
# Number of columns
N = 20
# ======= 1-D GRIDS =======
# Set the number of threads in a block
threadsperblock_1d = 4
# Calculate the number of thread blocks in the grid
blockspergrid_1d = np.int(np.ceil(M / threadsperblock_1d))
# ======= 1-D GRIDS =======
@cuda.jit('void(float32[:, :], float32[:], int32[:])')
def get_randomchoice(cumul_a, random_nos, output):
x = cuda.grid(1)
if x < cumul_a.shape[0]:
for y in range(cumul_a.shape[1]):
if random_nos[x] > cumul_a[x, y]:
output[x] = y
break # return
if __name__ == '__main__':
# Prepare the matrix whise each row is a cumulative probability distribution
A = np.random.rand(M, N).astype(np.float32)
A = np.divide(A,np.sum(A,axis=1,keepdims=True))
cumul_A = np.cumsum(A, axis=1)
# Put an assertion that cumul_A is indeed cumulative
assert np.allclose(cumul_A[:,-1],np.ones(M))
# Draw values from uniform distribution
RandValues = np.random.rand(M).astype(np.float32)
# Output array in numpy
Y = np.zeros(M, dtype=np.int32)
for iStep in range(M):
Y[iStep] = np.argwhere(RandValues[iStep] <= cumul_A[iStep])[0]
print('From numpy:\n{}'.format(Y))
# Transfer to GPU
cumul_A_gpu = cuda.to_device(cumul_A)
RandValues_gpu = cuda.to_device(RandValues)
# Return array from GPU
random_idx_gpu = cuda.device_array(M, dtype=np.int32)
get_randomchoice[blockspergrid_1d, threadsperblock_1d](cumul_A_gpu, RandValues_gpu, random_idx_gpu)
random_idx = random_idx_gpu.copy_to_host()
print('From cuda:\n{}'.format(random_idx))
任何帮助都将不胜感激。这是虚惊一场!代码中有个小故障。行if random_nos[x]>cumul_a[x,y]:将是if random_nos[x]