cuda内核for循环太长了?

时间:2017-06-09 11:26:52

标签: c++ memory cuda

经过一段长时间的手动调整和半自动调试后,我终于找到了内核崩溃的原因。是不是你不能在内核中制作一个for循环太长时间? 这是最小完整和可验证的代码: 错误是:未指定的启动失败

#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <iostream>
#include <stdio.h>
using namespace std;
typedef unsigned char uchar;

struct NodePointer {
    int id;
    uchar dist;
    int idintree;
    NodePointer() :id(0), dist(0), idintree(-1){}
};
struct TreeNode {
    NodePointer father;
    NodePointer children[4];
    int id;
    int childrenNum;
    int idintree;
    TreeNode() :id(0), childrenNum(0), idintree(-1){}
};

__global__ void
kernel4(int NumTN, TreeNode* tempthistree, int size)
{
    int index = blockIdx.x*blockDim.x + threadIdx.x;
    if (index < NumTN)
    {
        for (int i = 0; i < size; i++)
        {
            TreeNode node1 = tempthistree[i];
            printf("              node %d in tree %d, its id in tree is %d, its child num is %d\n", i, index, node1.idintree, node1.childrenNum);

        }
    }
}


int main()
{
    int n1 = 33417;
    TreeNode * testtree;
    cudaMallocManaged(&testtree, n1*sizeof(TreeNode));
    for (int i = 0; i < n1; i++)
    {
        TreeNode c;
        c.idintree = i;
        c.id = i;
        c.father.id = i - 1;
        c.father.dist = 1;
        c.childrenNum = i % 4;
        int aaa = i % 4;
        for (int j = 0; j < 4; j++)
        {
            c.children[j].dist = j;
            c.children[j].id = 1;
            c.children[j].idintree = 10;
        }
        testtree[i] = c;
    }


    kernel4 << <1, 1 >> >(4000, testtree, n1);
    cudaDeviceSynchronize();
    cudaError_t err = cudaGetLastError();
    if (err != cudaSuccess)
    {
        printf("Kernel3 error :%s\n", cudaGetErrorString(err));
        system("pause");
    }
}

2 个答案:

答案 0 :(得分:0)

假设您使用的是Windows,可能会遇到超时检测和恢复问题,这意味着您的内核执行时间过长。

看到这个答案: C++: Simple CUDA volume reconstruction code crashing

答案 1 :(得分:0)

首先,内核调用:kernel4 << <1, 1 >> >(4000, testtree, n1)

您正在使用1个块,只有1个线程。所以你的索引只在这里变为0。您正在为此线程提供以下任务:CPU,请执行for循环。所以那里没有并行性。

你可以在内核中使用for循环,但我建议你把它放在外面。至少在这里有可能:

__global__ void
kernel4(int NumTN, TreeNode* tempthistree, int size) 
{
    int index = blockIdx.x*blockDim.x + threadIdx.x;
    for (int i=0;i<NumTN;i++)
    {
        if (index < size)
        {
            TreeNode node1 = tempthistree[index];
            printf("node %d in tree %d, its id in tree is %d, its child num is %d\n", index, i, node1.idintree, node1.childrenNum);

        }
    }
}

现在你需要更多的线程,n1最小。在一个块中有1024个线程(现在)的限制,因此您需要更多的块才能达到n1。例如ceil(n1 / 1024)块。

另外,我在C中不确定,但在结构中使用数组时要小心。尝试使用内存传输的简短示例,看看你是否得到了正确的答案(我在使用fortran时遇到了一些问题)。

PS:还要检查你的计算能力:printf仅适用于计算能力> = 2.0的GPU。