CUDA中结构的简单操作:分段故障

Question

这是我第一次在CUDA中实现结构.在下面的程序中,我将结构复制到GPU并对数据执行基本操作,并将结果复制回主机.

#include<stdio.h>

inline cudaError_t checkCuda(cudaError_t result)
{
    #if defined(DEBUG) || defined(_DEBUG)
        if (result != cudaSuccess) {
            fprintf(stderr, "CUDA Runtime Error: %sn", cudaGetErrorString(result));
        assert(result == cudaSuccess);
        }
    #endif
    return result;
}

typedef struct myStruct {
    int* a;
    int b;
}MyStruct;

__global__ void structOperation(MyStruct *d_data){
    int idx = threadIdx.x;

    d_data->a[idx] += 10;
}

int main(){
    MyStruct *h_data, *d_data, *out_data;

    size_t structSize = sizeof(MyStruct);
    size_t intSize = sizeof(int);


    h_data = (MyStruct *) malloc(structSize * 1);
    h_data->b = 32;
    h_data->a = (int *)malloc(intSize * h_data->b);

    out_data = (MyStruct *) malloc(structSize * 1);
    out_data->b = 32;
    out_data->a = (int *)malloc(intSize * out_data->b);

    for(int i = 0; i<32; i++){
        h_data->a[i] = i;   
    }

    //Memory allocation for the Struct
    checkCuda(cudaMalloc(&d_data, sizeof(MyStruct) * 1));
    checkCuda(cudaMalloc(&(d_data->a), sizeof(int) * 32));


    checkCuda(cudaMemcpy(&d_data, &h_data, sizeof(MyStruct) * 1, cudaMemcpyHostToDevice));
    checkCuda(cudaMemcpy(&(d_data->a), &(h_data->a), sizeof(int) * 32, cudaMemcpyHostToDevice)); 


    structOperation<<<1,32>>>(d_data);


    checkCuda(cudaMemcpy(&out_data, &d_data, sizeof(myStruct) * 1, cudaMemcpyDeviceToHost));
  //cudaMemcpy(&(out_data->a), &(d_data->a), sizeof(int) * d_data->b, cudaMemcpyDeviceToHost); 

    printf("\nDataElements : ");
    for(int i = 0; i<32; i++){
        printf("    %d",out_data->a[i]);
    }
    printf("\n");
}

作为执行结果,我得到'Segmentation Fault'.我想我正在错误地操作结构.实施的正确方法是什么？

Answer 1

提供的代码中有几个无效的内存访问.

1. cudaMalloc从主机访问设备内存(使用分配)d_data->a会导致未定义的行为(分段错误等).
2. cudaMemcpy将指针作为参数,而不是指针的地址.所以 cudaMemcpy(&d_data, &h_data...应该替换为cudaMemcpy(d_data, h_data....

使用设备指针作为成员分配设备对象有点棘手.它可以实现如下:

1. 分配临时主机对象(MyStruct temp).
2. 将设备内存分配给我们在设备上所需的成员(cudaMalloc(&temp.a, bytes)).
3. 分配设备对象(cudaMalloc(&d_data, sizeof(MyStruct)).
4. 将临时主机对象复制到设备对象(cudaMemcpy(d_data, &temp, sizeof(MyStruct), cudaMemcpyHostToDevice)).

请记住,当您修改d_data->a设备上的内容时,temp.a也会进行修改,因为它们实际上指向设备上的相同内存位置.

您的最终主要功能将如下所示:

int main(){
    MyStruct *h_data, *d_data, *out_data;

    size_t structSize = sizeof(MyStruct);
    size_t intSize = sizeof(int);


    h_data = (MyStruct *) malloc(structSize * 1);
    h_data->b = 32;
    h_data->a = (int *)malloc(intSize * h_data->b);

    out_data = (MyStruct *) malloc(structSize * 1);
    out_data->b = 32;
    out_data->a = (int *)malloc(intSize * out_data->b);

    for(int i = 0; i<32; i++){
        h_data->a[i] = i;   
    }

    //Create temporary MyStruct object on host and allocate memory to its member "a" on device
    MyStruct temp;
    temp.b = h_data->b;
    checkCuda(cudaMalloc(&temp.a, 32 * sizeof(int)));

    //Copy host data to temp.a
    checkCuda(cudaMemcpy(temp.a, h_data->a, 32 * sizeof(int), cudaMemcpyHostToDevice));

    //Memory allocation for the device MyStruct
    checkCuda(cudaMalloc(&d_data, sizeof(MyStruct) * 1));
    //Copy actual object to device
    checkCuda(cudaMemcpy(d_data, &temp, sizeof(MyStruct) * 1, cudaMemcpyHostToDevice));


    structOperation<<<1,32>>>(d_data);

    //temp.a will be updated after kernel launch
    checkCuda(cudaMemcpy(out_data->a, temp.a, 32 * sizeof(int), cudaMemcpyDeviceToHost)); 

    printf("\nDataElements : ");
    for(int i = 0; i<32; i++)
    {
        printf("    %d",out_data->a[i]);
    }
    printf("\n");

    checkCuda(cudaFree(temp.a));
    checkCuda(cudaFree(d_data));

    free(h_data->a);
    free(out_data->a);
    free(h_data); 
    free(out_data);
}