How to use Cuda Memory 3D using cudaMalloc3D

Question

I am learning 3D array manipulation in cuda. I implemented the following code, however I cannot get the expected result. That is I am taking the array and changing the elements from 0 to 1. I tried to find the mistake but I cannot locate it. Can someone point where my mistake is in the code.

 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";
        }
    }
}

OUTPUT: I get all 0 instead of 1

Answer 1

The only error I could find in your code is that you are passing an int array to the kernel, but inside the kernel you are manipulating it as if it were a float array:

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

When I fix that issue, the test code I created from yours runs without error, producing all 1 values at the final test:

$ 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
$