CODEWITHSUNDEEP
arrayspointerscudastructurememcpy
Roshan
Roshan

Reputation: 568

Copy structure pointer array member to another array

Say, I have a structure like

struct vertex
{
  int x;
  int y;
  int z;
}

and I make a pointer array like

vertex *points = new vertex[100];

so, at a point I require just the x members of that pointer array. So how can i do that?

eg a new pointer : int *xPoints=new int[100];

and in this i want to store all the x members of the vertex. Is there any command to do that? i dont want to do that using loop, i want to do that using special memcopy or something.

Upvotes: 0

Views: 106

Answers (1)

Robert Crovella
Robert Crovella

Reputation: 151879

cudaMemcpy2D can be used to copy between host and device when either the source or destination (or both) data should be strided.

Here's a worked example:

$ cat t553.cu
#include <stdio.h>
#define DSIZE 4

struct vertex {
  int x,y,z;
};

__global__ void mykernel(int *data, unsigned length){
  for (int i = 0; i < length; i ++) printf("kernel data[%d] = %d\n",i,data[i]);
}

int main(){

  vertex *points = new vertex[DSIZE];
  for (int i = 0; i < DSIZE; i++){
    points[i].x = 1;
    points[i].y = 2;
    points[i].z = 3;}
  int *d_ypoints;
  cudaMalloc(&d_ypoints, DSIZE*sizeof(int));
  cudaMemcpy2D(d_ypoints, sizeof(int), ((int *)points)+1, 3*sizeof(int), sizeof(int), DSIZE, cudaMemcpyHostToDevice);
  mykernel<<<1,1>>>(d_ypoints, DSIZE);
  cudaDeviceSynchronize();
  return 0;
}


$ nvcc -arch=sm_20 -o t553 t553.cu
$ cuda-memcheck ./t553
========= CUDA-MEMCHECK
kernel data[0] = 2
kernel data[1] = 2
kernel data[2] = 2
kernel data[3] = 2
========= ERROR SUMMARY: 0 errors
$

Parsing out the cudaMemcpy2D operation:

  cudaMemcpy2D(d_ypoints,  // starting pointer on the device (destination)
             sizeof(int),  // stride on device (i.e. no stride)
       ((int *)points)+1,  // starting pointer on host (.y element of first struct)
           3*sizeof(int),  // stride on host (distance between consecutive .y elements)
             sizeof(int),  // number of bytes to transfer per "row"
                   DSIZE,  // number of "rows" to transfer
 cudaMemcpyHostToDevice);  // direction of transfer

Upvotes: 1

Related Questions