libgomp/testsuite/libgomp.oacc-c-c++-common/routine-gwv-1.c

   1 /* { dg-do run } */
   2 /* { dg-additional-options "-O2" */
   3
   4 #include <stdio.h>
   5
   6 #define N (32*32*32+17)
   7
   8 #pragma acc routine gang
   9 void __attribute__ ((noinline)) gang (int ary[N])
  10 {
  11 #pragma acc loop gang worker vector
  12   for (unsigned ix = 0; ix < N; ix++)
  13     {
  14       if (__builtin_acc_on_device (5))
  15         {
  16           int g = 0, w = 0, v = 0;
  17
  18           __asm__ volatile ("mov.u32 %0,%%ctaid.x;" : "=r" (g));
  19           __asm__ volatile ("mov.u32 %0,%%tid.y;" : "=r" (w));
  20           __asm__ volatile ("mov.u32 %0,%%tid.x;" : "=r" (v));
  21           ary[ix] = (g << 16) | (w << 8) | v;
  22         }
  23       else
  24         ary[ix] = ix;
  25     }
  26 }
  27
  28 int main ()
  29 {
  30   int ary[N];
  31   int ix;
  32   int exit = 0;
  33   int ondev = 0;
  34
  35   for (ix = 0; ix < N;ix++)
  36     ary[ix] = -1;
  37
  38 #pragma acc parallel num_gangs(32) num_workers(32) vector_length(32) copy(ary) copy(ondev)
  39   {
  40     ondev = __builtin_acc_on_device (5);
  41     gang (ary);
  42   }
  43
  44   for (ix = 0; ix < N; ix++)
  45     {
  46       int expected = ix;
  47       if(ondev)
  48         {
  49           int chunk_size = (N + 32*32*32 - 1) / (32*32*32);
  50
  51           int g = ix / (chunk_size * 32 * 32);
  52           int w = ix / 32 % 32;
  53           int v = ix % 32;
  54
  55           expected = (g << 16) | (w << 8) | v;
  56         }
  57
  58       if (ary[ix] != expected)
  59         {
  60           exit = 1;
  61           printf ("ary[%d]=%x expected %x\n", ix, ary[ix], expected);
  62         }
  63     }
  64
  65   return exit;
  66 }