libgomp/testsuite/libgomp.oacc-c-c++-common/loop-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 int main ()
   8 {
   9   int ary[N];
  10   int ix;
  11   int exit = 0;
  12   int ondev = 0;
  13
  14   for (ix = 0; ix < N;ix++)
  15     ary[ix] = -1;
  16
  17 #pragma acc parallel num_gangs(32) num_workers(32) vector_length(32) copy(ary) copy(ondev)
  18   {
  19 #pragma acc loop gang worker vector
  20     for (unsigned ix = 0; ix < N; ix++)
  21       {
  22         if (__builtin_acc_on_device (5))
  23           {
  24             int g = 0, w = 0, v = 0;
  25
  26             __asm__ volatile ("mov.u32 %0,%%ctaid.x;" : "=r" (g));
  27             __asm__ volatile ("mov.u32 %0,%%tid.y;" : "=r" (w));
  28             __asm__ volatile ("mov.u32 %0,%%tid.x;" : "=r" (v));
  29             ary[ix] = (g << 16) | (w << 8) | v;
  30             ondev = 1;
  31           }
  32         else
  33           ary[ix] = ix;
  34       }
  35   }
  36
  37   for (ix = 0; ix < N; ix++)
  38     {
  39       int expected = ix;
  40       if(ondev)
  41         {
  42           int chunk_size = (N + 32*32*32 - 1) / (32*32*32);
  43
  44           int g = ix / (chunk_size * 32 * 32);
  45           int w = ix / 32 % 32;
  46           int v = ix % 32;
  47
  48           expected = (g << 16) | (w << 8) | v;
  49         }
  50
  51       if (ary[ix] != expected)
  52         {
  53           exit = 1;
  54           printf ("ary[%d]=%x expected %x\n", ix, ary[ix], expected);
  55         }
  56     }
  57
  58   return exit;
  59 }