CUDA中的累积和 [英] Cumulative summation in CUDA

问题描述

有人可以指出我在正确的方向上如何做这种类型的计算并行，或告诉我这个方法的一般名称是什么？我不认为这些会返回相同的结果。

Can someone please point me in the right direction on how to do this type of calculation in parallel, or tell me what the general name of this method is? I don't think these will return the same result.

C ++

for (int i = 1; i < width; i++)
        x[i] = x[i] + x[i-1];

CUDA

int i = blockIdx.x * blockDim.x + threadIdx.x

if ((i > 0) && (i < (width)))
    X[i] = X[i] + X[i-1];

推荐答案

其中 x [i] 的最终值是所有值的总和 x [0] ... x [i]

This looks like a cumulative sum operation, in which the final value of x[i] is the sum of all values x[0]...x[i] in the original array.

在CUDA中，这被称为扫描或前缀和操作，并且可以有效地并行化。参见例如本讲座的示例。

In CUDA, this is called a scan or prefix-sum operation, and it can be efficiently parallelized. See e.g. this lecture for examples.

这篇关于CUDA中的累积和的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！