主页 / coding / 问题 / 79543662
Accepted
Bbbbbunn2023
Asked: 2025-03-30 01:47:03 +0800 CST

CUDA 在全局内存中分配并初始化一个数组，但一直出现分段错误

  • 772
void CudaRenderer::render() {

    dim3 blockDim(16, 16, 1);
    dim3 gridDim((image->width + blockDim.x - 1) / blockDim.x,
                  (image->height + blockDim.y - 1) / blockDim.y);

    int *filteredCircles,
        *lastIndices,
        sz = gridDim.y * gridDim.x ;
    cudaMalloc((void **)&filteredCircles, sizeof(int) * sz * 2000);
    cudaMalloc((void **)&lastIndices, sizeof(int) * sz);
    cudaMemset(lastIndices, 0, sizeof(int) * sz);
    filterCircles<<<gridDim, blockDim>>>(filteredCircles, lastIndices);
        for (int i = 0; i < 10; ++i)
                printf("lastIndices[%d] = %d\n", i, lastIndices[i]);
    kernelRenderCircles<<<gridDim, blockDim>>>(filteredCircles, lastIndices);
    cudaFree(filteredCircles);
    cudaDeviceSynchronize();

在添加打印之前，代码编译正常，但产生了错误的结果。在添加打印以检查 lastIndices 之后，它一直给我分段错误。我在这里做错了什么？

cuda

1 个回答

  1. Best Answer

    lastIndices我认为您想要做的是在 filterCircles 运行后打印值。

    为此，你需要明白

    A：GPU__global__功能与 CPU 异步运行。B
    ：您无法在主机（CPU）上访问设备（GPU）指针。

    诀窍是将代码更改如下：

    老的：

        cudaMalloc((void **)&lastIndices, sizeof(int) * sz);
    cudaMemset(lastIndices, 0, sizeof(int) * sz);
    filterCircles<<<gridDim, blockDim>>>(filteredCircles, lastIndices);
    //Problem 1: the CPU for loop will run concurrently with the GPU code.
    //meaning lastIndices may not have the filterCircles data yet.
    for (int i = 0; i < 10; ++i) {
        //problem 2: lastIndices is in GPU memory that the CPU cannot see.
        printf("lastIndices[%d] = %d\n", i, lastIndices[i]);
    }

    新的：

        cudaMalloc((void **)&lastIndices, sizeof(int) * sz);
    cudaMemset(lastIndices, 0, sizeof(int) * sz);
    filterCircles<<<gridDim, blockDim>>>(filteredCircles, lastIndices);
    cudaDeviceSynchronize(); //wait for filterCircles to finish
    //allocate space on the host to hold the cpu data
    int* hostLastIndices;
    hostLastIndices = (int*)malloc(sizeof(int) * sz);
    //copy data from device to host
    cudaMemcpy(hostLastIndices, lastIndices, sizeof(int) * sz, cudaMemcopyDeviceToHost);
    for (int i = 0; i < sz; ++i) {
        printf("lastIndices[%d] = %d\n", i, hostLastIndices[i]);
    }
    • 0

相关问题