Question

My problem is this: I have an array in 3D and I cannot use the malloc3D, and I need to convert and manipulate 1D array on GPU. But I don’t know how do it. In this moment I am using

#define nx  8
#define ny  6
#define nz  4

to define the matriz array.. 4 matrices of 6 row with 8 columns with index i,j,k.

u[i][j][k]

and I declaration of:

cudaMalloc( (void**)&dev_u, ny * nx * nz * sizeof(float) ) ;
cudaMemcpy( dev_u, u, ny * nx * nz * sizeof(float), cudaMemcpyHostToDevice );
dim3 dimBlock(nx,ny,nz);
dim3 dimGrid(1,1);
FTCS3D<<<dimGrid, dimBlock>>>( dev_u );
cudaMemcpy( u, dev_u, ny * nx * nz * sizeof(float), cudaMemcpyDeviceToHost );

Inside the GPU:

__global__ void FTCS3D( float *u )
{
    int i = threadIdx.y+blockDim.y*blockIdx.y;
    int j = threadIdx.x+blockDim.x*blockIdx.x;
    int k = threadIdx.z+blockDim.z*blockIdx.z;
    int offset = i * nx + j + ny * nx * z;
    int totid=nx*ny*nz;

    if (offset < totid)
    {
        if ( offset ==1 )
           u[offset]=5.0;
   }
}

The number 5 appears in other matriz not in u[0][0][1], I do not have any idea about how to index all variables inside the offset remember I HAVE TO DO IT in this way of 1D vector.

Answer 1

If you have a array3D [HEIGHT][WIDTH] [DEPTH] then you could turn it into array1D [HEIGHT * WIDTH * DEPTH].

Out side your kernel you convert the 3D to 1D array

for (int x = 0, k=0; x < HEIGHT; x++)
  for (int y = 0; y < WIDTH; y++)
     for (int z = 0; z < DEPTH; z++)
        a1D[k++] = a3D[x][y][z]

Why not only one dimension in you cuda?

__global__ void FTCS3D( float *u,int HEIGHT, int WIDTH, int DEPTH)
{   
    int x = threadIdx.x+blockDim.x*blockIdx.x;
    int totid = HEIGHT * WIDTH * DEPTH;

    if (x < totid)
    {
       if (x==1 )
          u[x]=5.0;
     }

}