c++ - cudaMemcpyDeviceToHost() 失败

Question

这个问题已经有答案了:

Interpretation of "too many resources for launch" (1 个回答)

已关闭 6 年前。

我有以下代码(假设所有内容都已正确定义):

#include "OurIncludes.h"
#include <ctime>

__global__ void kernel_testing(int *d_intersects, Circle *part1, Circle *part2)
{
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    int j = blockIdx.y * blockDim.y + threadIdx.y;
    if (i < 10 && j < 10) {
        int index = i + j * 10;
        d_intersects[index] = part1[i].intersect(part2[j]);
    }
}


int main(void)
{
    dim3 GRID(1, 1);
    dim3 BLOCK(10, 10);

    short randomNum;
    RandObj randGenerator;
    Circle* obj = new Circle[10];
    Circle* obj2 = new Circle[10];
    Circle *d_obj;
    Circle *d_obj2;
    int intersects[100];
    int *d_intersects;

    if (cudaSuccess != cudaMalloc((void **)&d_obj, sizeof(Circle) * 10)) {
        fprintf(stderr, "Failed to allocate memory for d_result\n");
    }
    if (cudaSuccess != cudaMalloc((void **)&d_obj2, sizeof(Circle) * 10)) {
        fprintf(stderr, "Failed to allocate memory for d_result\n");
    }
    if (cudaSuccess != cudaMalloc((void **)&d_intersects, sizeof(int) * 100)) {
        fprintf(stderr, "Failed to allocate memory for d_result\n");
    }
    for (int i = 0; i < 10; i++) {
        obj[i] = (*randGenerator.makeRandomCircle());
    }

    for (int i = 0; i < 10; i++) {
        obj2[i] = (*randGenerator.makeRandomCircle());
    }
    size_t size = sizeof(Circle);
    if (cudaSuccess != cudaMemcpy(d_obj, obj, size * 10, cudaMemcpyHostToDevice)) {
        fprintf(stderr, "Failed to copy data to d_obj\n");
    }
    if (cudaSuccess != cudaMemcpy(d_obj2, obj2, size * 10, cudaMemcpyHostToDevice)) {
        fprintf(stderr, "Failed to copy data to d_obj2\n");
    }

    kernel_testing << < GRID, BLOCK >> >(d_intersects, d_obj, d_obj2);


    cudaError_t s = cudaMemcpy(intersects, d_intersects, sizeof(int) * 100, cudaMemcpyDeviceToHost);
    fprintf(stderr, "Error is: %s", cudaGetErrorString(s));
    cudaFree(d_intersects);
    cudaFree(d_obj);
    cudaFree(d_obj2);
    return 0;   
}

由于某种原因，代码总是在cudaMemcpyDeviceToHost处失败。 ，而且我看不出为什么应该这样做的原因。我尝试过使用不同的对象(三角形、球体等)启动，但当我需要将数据从设备复制回主机时，它总是失败。感谢任何帮助和/或建议，我对使用 CUDA 编程非常陌生。谢谢。

编辑:错误代码表示遇到了非法内存访问，但我不明白为什么会发生这种情况。

编辑2:所以我删除了所有双指针并“展平”我的数组，但我仍然遇到同样的问题。我现在完全没有想法了。

Answer 1

我想我曾经遇到过这样的问题，我的解决方案是:

cudaError_t status = cudaMemcpy(devPtr, srcPtr, size * sizeof(int), cudaMemcpyHostToDevice);
if (status == cudaSuccess) { ... }

尝试不要直接将函数与 cudaSuccess 进行比较，而是通过变量进行比较。