1  #include <stdio.h>
       2  #include <openacc.h>
       3  #include <gomp-constants.h>
       4  
       5  #define N (32*32*32+17)
       6  
       7  #pragma acc routine worker
       8  void __attribute__ ((noinline)) worker (int ary[N])
       9  {
      10  #pragma acc loop worker vector
      11    for (unsigned ix = 0; ix < N; ix++)
      12      {
      13        if (acc_on_device (acc_device_not_host))
      14  	{
      15  	  int g, w, v;
      16  
      17  	  g = __builtin_goacc_parlevel_id (GOMP_DIM_GANG);
      18  	  w = __builtin_goacc_parlevel_id (GOMP_DIM_WORKER);
      19  	  v = __builtin_goacc_parlevel_id (GOMP_DIM_VECTOR);
      20  	  ary[ix] = (g << 16) | (w << 8) | v;
      21  	}
      22        else
      23  	ary[ix] = ix;
      24      }
      25  }
      26  
      27  int main ()
      28  {
      29    int ary[N];
      30    int ix;
      31    int exit = 0;
      32    int ondev = 0;
      33    int workersize, vectorsize;
      34  
      35    for (ix = 0; ix < N;ix++)
      36      ary[ix] = -1;
      37    
      38  #define NW 32
      39  #define VL 32
      40  #pragma acc parallel num_workers(NW) vector_length(VL) \
      41  	    copy(ary) copy(ondev)
      42    {
      43      ondev = acc_on_device (acc_device_not_host);
      44      worker (ary);
      45    }
      46    workersize = NW;
      47    vectorsize = VL;
      48  #ifdef ACC_DEVICE_TYPE_radeon
      49    /* AMD GCN has an upper limit of 'num_workers(16)'.  */
      50    if (workersize > 16)
      51      workersize = 16;
      52    /* AMD GCN uses the autovectorizer for the vector dimension: the use
      53       of a function call in vector-partitioned code in this test is not
      54       currently supported.  */
      55    vectorsize = 1;
      56  #endif
      57  
      58    for (ix = 0; ix < N; ix++)
      59      {
      60        int expected = ix;
      61        if(ondev)
      62  	{
      63  	  int g = 0;
      64  	  int w = (ix / vectorsize) % workersize;
      65  	  int v = ix % vectorsize;
      66  
      67  	  expected = (g << 16) | (w << 8) | v;
      68  	}
      69        
      70        if (ary[ix] != expected)
      71  	{
      72  	  exit = 1;
      73  	  printf ("ary[%d]=%x expected %x\n", ix, ary[ix], expected);
      74  	}
      75      }
      76    
      77    return exit;
      78  }