1  #include <stdlib.h>
       2  
       3  #define N (1024 * 512)
       4  #define COUNTERTYPE unsigned int
       5  
       6  int
       7  main (void)
       8  {
       9    unsigned int *__restrict a;
      10    unsigned int *__restrict b;
      11    unsigned int *__restrict c;
      12  
      13    a = (unsigned int *__restrict)malloc (N * sizeof (unsigned int));
      14    b = (unsigned int *__restrict)malloc (N * sizeof (unsigned int));
      15    c = (unsigned int *__restrict)malloc (N * sizeof (unsigned int));
      16  
      17  #pragma acc data copyout (a[0:N], b[0:N], c[0:N])
      18    {
      19  #pragma acc kernels present (a[0:N])
      20      {
      21        for (COUNTERTYPE i = 0; i < N; i++)
      22  	a[i] = i * 2;
      23      }
      24  
      25  #pragma acc kernels present (b[0:N])
      26      {
      27        for (COUNTERTYPE i = 0; i < N; i++)
      28  	b[i] = i * 4;
      29      }
      30  
      31  #pragma acc kernels present (a[0:N], b[0:N], c[0:N])
      32      {
      33        for (COUNTERTYPE ii = 0; ii < N; ii++)
      34  	c[ii] = a[ii] + b[ii];
      35      }
      36    }
      37  
      38    for (COUNTERTYPE i = 0; i < N; i++)
      39      if (c[i] != a[i] + b[i])
      40        abort ();
      41  
      42    free (a);
      43    free (b);
      44    free (c);
      45  
      46    return 0;
      47  }