如何使用cudaMalloc3D使用Cuda Memory 3D

Question

我正在學習cuda中的3D數組操縱。 我實現了以下代碼，但是無法獲得預期的結果。 那就是我正在使用數組並將元素從0更改為1。我試圖找到錯誤，但是找不到它。 有人可以指出我的錯誤在代碼中的位置嗎？

 int iDivUp(int a, int b) {
  return ((a % b) != 0) ? (a / b + 1) : (a / b);
}

__global__
void kernel(cudaPitchedPtr d_pitched_ptr, int COLS, int ROWS, int D) {
    int t_idx = threadIdx.x + blockIdx.x * blockDim.x;
    int t_idy = threadIdx.y + blockIdx.y * blockDim.y;

    char* d_ptr = static_cast<char*>(d_pitched_ptr.ptr);
    size_t pitch = d_pitched_ptr.pitch;

    float* element  = (float*)(d_ptr + t_idy * pitch) + t_idx;
    element[0] = 1;
    element[1] = 1;
    element[2] = 1;
}

void iFilter() {
    const int ROWS = 100
    const int COLS = 120
    const int DEPTH = 3;
    int pixels[COLS][ROWS][DEPTH];

    for (int j = 0; j < ROWS; j++) {
        for (int i = 0; i < COLS; i++) {
            for (int k = 0; k < DEPTH; k++) {
                pixels[i][j][k] = 0;
            }
        }
    }
    cudaExtent extent = make_cudaExtent(COLS * sizeof(int), ROWS, DEPTH);
    cudaPitchedPtr d_pitched_ptr;
    cudaMalloc3D(&d_pitched_ptr, extent);

    cudaMemcpy3DParms d_parms = {0};
    d_parms.srcPtr.ptr = pixels;
    d_parms.srcPtr.pitch = COLS * sizeof(int);
    d_parms.srcPtr.xsize = COLS;
    d_parms.srcPtr.ysize = ROWS;

    d_parms.dstPtr.ptr = d_pitched_ptr.ptr;
    d_parms.dstPtr.pitch = d_pitched_ptr.pitch;
    d_parms.dstPtr.xsize = COLS;
    d_parms.dstPtr.ysize = ROWS;

    d_parms.extent.width = COLS * sizeof(int);
    d_parms.extent.height = ROWS;
    d_parms.extent.depth = DEPTH;
    d_parms.kind = cudaMemcpyHostToDevice;

    cudaMemcpy3D(&d_parms);

    dim3 block_size(blocksize, blocksize);
    dim3 grid_size(iDivUp(COLS, blocksize), iDivUp(ROWS, blocksize));

    kernel<<<grid_size, block_size>>>(
        d_pitched_ptr, COLS, ROWS, DEPTH);

    int download_pixels[COLS][ROWS][DEPTH];
    d_parms.srcPtr.ptr = d_pitched_ptr.ptr;
    d_parms.srcPtr.pitch = d_pitched_ptr.pitch;
    d_parms.dstPtr.ptr = download_pixels;
    d_parms.dstPtr.pitch = COLS * sizeof(int);
    d_parms.kind = cudaMemcpyDeviceToHost;

    cudaMemcpy3D(&d_parms);

    for (int j = 0; j < ROWS; j++) {
        for (int i = 0; i < COLS; i++) {
            for (int k = 0; k < DEPTH; k++) {
                image.at<cv::Vec3b>(j, i)[k] = download_pixels[i][j][k];
                 std::cout << download_pixels[i][j][k]  << " ";
            }
             std::cout << "\n";
        }
    }
}

輸出：我得到的全部是0而不是1

Answer 1

我在您的代碼中發現的唯一錯誤是，您正在將一個int數組傳遞給內核，但是在內核內部您正在對其進行操作，就好像它是一個float組一樣：

float* element  = (float*)(d_ptr + t_idy * pitch) + t_idx;

解決此問題后，我從您創建的測試代碼將運行無錯誤，並在最終測試中生成所有1個值：

$ cat t1114.cu
#include <iostream>
const int blocksize = 16;
 int iDivUp(int a, int b) {
  return ((a % b) != 0) ? (a / b + 1) : (a / b);
}

__global__
void kernel(cudaPitchedPtr d_pitched_ptr, int COLS, int ROWS, int D) {
    int t_idx = threadIdx.x + blockIdx.x * blockDim.x;
    int t_idy = threadIdx.y + blockIdx.y * blockDim.y;

    char* d_ptr = static_cast<char*>(d_pitched_ptr.ptr);
    size_t pitch = d_pitched_ptr.pitch;

    int * element  = (int *)(d_ptr + t_idy * pitch) + t_idx;
    element[0] = 1;
    element[1] = 1;
    element[2] = 1;
}

int main() {
    const int ROWS = 100;
    const int COLS = 120;
    const int DEPTH = 3;
    int pixels[COLS][ROWS][DEPTH];

    for (int j = 0; j < ROWS; j++) {
        for (int i = 0; i < COLS; i++) {
            for (int k = 0; k < DEPTH; k++) {
                pixels[i][j][k] = 0;
            }
        }
    }
    cudaExtent extent = make_cudaExtent(COLS * sizeof(int), ROWS, DEPTH);
    cudaPitchedPtr d_pitched_ptr;
    cudaMalloc3D(&d_pitched_ptr, extent);

    cudaMemcpy3DParms d_parms = {0};
    d_parms.srcPtr.ptr = pixels;
    d_parms.srcPtr.pitch = COLS * sizeof(int);
    d_parms.srcPtr.xsize = COLS;
    d_parms.srcPtr.ysize = ROWS;

    d_parms.dstPtr.ptr = d_pitched_ptr.ptr;
    d_parms.dstPtr.pitch = d_pitched_ptr.pitch;
    d_parms.dstPtr.xsize = COLS;
    d_parms.dstPtr.ysize = ROWS;

    d_parms.extent.width = COLS * sizeof(int);
    d_parms.extent.height = ROWS;
    d_parms.extent.depth = DEPTH;
    d_parms.kind = cudaMemcpyHostToDevice;

    cudaMemcpy3D(&d_parms);

    dim3 block_size(blocksize, blocksize);
    dim3 grid_size(iDivUp(COLS, blocksize), iDivUp(ROWS, blocksize));

    kernel<<<grid_size, block_size>>>(
        d_pitched_ptr, COLS, ROWS, DEPTH);

    int download_pixels[COLS][ROWS][DEPTH];
    d_parms.srcPtr.ptr = d_pitched_ptr.ptr;
    d_parms.srcPtr.pitch = d_pitched_ptr.pitch;
    d_parms.dstPtr.ptr = download_pixels;
    d_parms.dstPtr.pitch = COLS * sizeof(int);
    d_parms.kind = cudaMemcpyDeviceToHost;

    cudaMemcpy3D(&d_parms);
    for (int j = 0; j < ROWS; j++) {
        for (int i = 0; i < COLS; i++) {
            for (int k = 0; k < DEPTH; k++) {
                if (download_pixels[i][j][k] != 1) std::cout << i << "," << j << "," << k << ": " << download_pixels[i][j][k]  << " error! " << std::endl; return 1;
            }
        }
    }
    return 0;
}
$ nvcc -o t1114 t1114.cu
$ cuda-memcheck ./t1114
========= CUDA-MEMCHECK
========= ERROR SUMMARY: 0 errors
$