【CUDA学习日记】3.4 动态并行

之前，所有的核函数都是从主机线程中被调用。内核启动是静态的，意味着所有的并行任务必须在程序运行前就确定好。动态并行是指在一个CUDA内核执行时，该内核能够动态地启动其他内核的能力。启用动态并行，你需要在编译CUDA代码时使用-rdc=true选项，并且确保你的GPU支持这一特性。

在动态并行中，内核执行分为两种类型：父母和孩子。父线程、父线程块或父网格启

动一个新的网格，即子网格。子线程、子线程块或子网格被父母启动。子网格必须在父线程、父线程块或父网格完成之前完成。只有在所有的子网格都完成之后，父母才会完成。

3.4.1 GPU 上嵌套Hello world

复制代码

#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <string.h>
#include <windows.h>
#include "../common/common.h"

__global__ void nestedHelloWorld(int const iSize, int iDepth){
    int tid = threadIdx.x;
    printf("Recursion = %d : helloworld from thread %d block %d\n", iDepth, tid, blockIdx.x);

    if (iSize == 1) return;

    int nthreads = iSize >> 1;
    if (tid == 0 && nthreads){
        nestedHelloWorld<<<1, nthreads>>>(nthreads, ++iDepth);
        printf("------> nested execution depth : %d\n" , iDepth);
    }
}

int main(int argc, char **argv)
{
    int size = 8;
    int blocksize = 8;   // initial block size
    int igrid = 1;

    if(argc > 1)
    {
        igrid = atoi(argv[1]);
        size = igrid * blocksize;
    }

    dim3 block (blocksize, 1);
    dim3 grid  ((size + block.x - 1) / block.x, 1);
    printf("%s Execution Configuration: grid %d block %d\n", argv[0], grid.x,
           block.x);

    nestedHelloWorld<<<grid, block>>>(block.x, 0);

    CHECK(cudaDeviceReset());
    return 0;
}

编译： nvcc .\nestedHelloworld.cu -o .\nestedHelloworld -rdc=true

输出：

nestedHelloworld.exe Execution Configuration: grid 1 block 8

Recursion = 0 : helloworld from thread 0 block 0

Recursion = 0 : helloworld from thread 1 block 0

Recursion = 0 : helloworld from thread 2 block 0

Recursion = 0 : helloworld from thread 3 block 0

Recursion = 0 : helloworld from thread 4 block 0

Recursion = 0 : helloworld from thread 5 block 0

Recursion = 0 : helloworld from thread 6 block 0

Recursion = 0 : helloworld from thread 7 block 0

------> nested execution depth : 1

Recursion = 1 : helloworld from thread 0 block 0

Recursion = 1 : helloworld from thread 1 block 0

Recursion = 1 : helloworld from thread 2 block 0

Recursion = 1 : helloworld from thread 3 block 0

------> nested execution depth : 2

Recursion = 2 : helloworld from thread 0 block 0

Recursion = 2 : helloworld from thread 1 block 0

------> nested execution depth : 3

Recursion = 3 : helloworld from thread 0 block 0

由输出可以看出，主机调用的父网格有1个线程块和8个线程。nestedHelloWorld

核函数递归地调用三次，每次调用的线程数是上一次的一半。