#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <cuda_runtime.h>

extern  "C" void vecadd(int,float*,float*,float*);
extern  "C" void set_value (int, float*, float*);

int main()
{
  float *a, *b, *c;
  float *result;
  int i;
  int n;
  n = 1024;

  result = (float*)malloc(n*sizeof(float));

  cudaMalloc((void**)&a,(size_t)n*sizeof(float));
  cudaMalloc((void**)&b,(size_t)n*sizeof(float));
  cudaMalloc((void**)&c,(size_t)n*sizeof(float));

    set_value(n, a, b);

  vecadd(n, a, b, c);

  cudaMemcpy (result, c, sizeof(float)*n, cudaMemcpyDeviceToHost);
  
  for( i = 0; i < 10; i++)
    fprintf(stdout, "c[%d] = %f\n", i, result[i]);
  return 0;
}