Copying a dynamically allocated 2D array from host to device in CUDA

Question

I want to copy a dynamically allocated 2D array from host to device to get its Discrete Fourier Transform.

I'm using below code to copy the array to the device

cudaMalloc((void**)&array_d, sizeof(cufftComplex)*NX*(NY/2+1));
cudaMemcpy(array_d, array_h, sizeof(float)*NX*NY, cudaMemcpyHostToDevice);

This works fine with static arrays, i get the intended output from my fft. But it doesn't work with dynamic arrays. After little bit searching I learnt I can not copy dynamic arrays like this from host to device. So I found this solution.

cudaMalloc((void**)&array_d, sizeof(cufftComplex)*NX*(NY/2+1));
for(int i=0; i<NX; ++i){
    cudaMemcpy(array_d+ i*NY, array_h[i], sizeof(float)*NY, cudaMemcpyHostToDevice);
}

But it's also not doing the task properly since I get wrong values from my fft.

Given below is my fft code.

cufftPlanMany(&plan, NRANK, n,NULL, 1, 0,NULL, 1, 0,CUFFT_R2C,BATCH);
cufftSetCompatibilityMode(plan, CUFFT_COMPATIBILITY_NATIVE);
cufftExecR2C(plan, (cufftReal*)data, data);
cudaThreadSynchronize();
cudaMemcpy(c, data, sizeof(float)*NX*NY, cudaMemcpyDeviceToHost);

How can I overcome this problem ?

EDIT

given below is the code

#define NX 4
#define NY 5
#define NRANK 2
#define BATCH 10

#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <cufft.h>
#include <stdio.h> 
#include <iostream>

int check();

int main()
    { 


    // static array
    float b[NX][NY] ={ 
        {0.7943 ,   0.6020 ,   0.7482  ,  0.9133  ,  0.9961},
        {0.3112 ,   0.2630 ,   0.4505  ,  0.1524  ,  0.0782},
         {0.5285 ,   0.6541 ,   0.0838  ,  0.8258  ,  0.4427},
         {0.1656 ,   0.6892 ,   0.2290  ,  0.5383  ,  0.1067}
        };

    // dynamic array
    float **a = new float*[NX];     
    for (int r = 0; r < NX; ++r)  
        {
        a[r] = new float[NY];
        for (int c = 0; c < NY; ++c)
            {            
                a[r][c] = b[r][c];           
            }
        }

    // arrray to store the results - host side   
    float c[NX][NY] = { 0 };

    cufftHandle plan;
    cufftComplex *data;
    int n[NRANK] = {NX, NY};

    cudaMalloc((void**)&data, sizeof(cufftComplex)*NX*(NY/2+1));
    cudaMemcpy(data, b, sizeof(float)*NX*NY, cudaMemcpyHostToDevice);

    /* Create a 2D FFT plan. */
    cufftPlanMany(&plan, NRANK, n,NULL, 1, 0,NULL, 1, 0,CUFFT_R2C,BATCH);
    cufftSetCompatibilityMode(plan, CUFFT_COMPATIBILITY_NATIVE);
    cufftExecR2C(plan, (cufftReal*)data, data);
    cudaThreadSynchronize();
    cudaMemcpy(c, data, sizeof(float)*NX*NY, cudaMemcpyDeviceToHost);

    cufftDestroy(plan);
    cudaFree(data);

    return 0;
    }

Answer 1

data is of type cufftComplex which is series of typedefs eventually resulting in a float2. That means data + n will advance data by n objects of type float2, or by 2 * n object of type float. This makes your "dynamic array" copying incorrect; you have to halve the increment of data.

EDIT

Looking at the parameter types of cufftExecR2C(), I think this should work:

for(int i=0; i<NX; ++i){
  cudaMemcpy(reinterpret_cast<float*>(data) + i*NY, a[i], sizeof(float)*NY, cudaMemcpyHostToDevice);
}

---

Side note: you don't actually have a dynamic 2D array (that would be new float[NX * NY]). What you have is a dynamic array of pointers to dynamic arrays of floats. I believe it would make more sense for you to use a true 2D array instead, which would allow you to keep the static-case copy code as well.

And since you've tagged this C++, you should seriously consider using std::vector instead of managing your dynamic memory manually. That is, change a like this:

std::vector<float> a(NX * NY);

And while you're at it, I'd suggest turning NX, NY etc. from macros to constants:

const size_t NX = 4;
const size_t NY = 5;

etc.