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 enter data create (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 parallel present (b[0:N])
      26    {
      27  #pragma acc loop
      28      for (COUNTERTYPE i = 0; i < N; i++)
      29        b[i] = i * 4;
      30    }
      31  
      32  #pragma acc kernels present (a[0:N], b[0:N], c[0:N])
      33    {
      34      for (COUNTERTYPE ii = 0; ii < N; ii++)
      35        c[ii] = a[ii] + b[ii];
      36    }
      37  
      38  #pragma acc exit data copyout (a[0:N], b[0:N], c[0:N])
      39  
      40    for (COUNTERTYPE i = 0; i < N; i++)
      41      if (c[i] != a[i] + b[i])
      42        abort ();
      43  
      44    free (a);
      45    free (b);
      46    free (c);
      47  
      48    return 0;
      49  }