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

   1 /* This code uses nvptx inline assembly guarded with acc_on_device, which is
   2    not optimized away at -O0, and then confuses the target assembler.
   3    { dg-skip-if "" { *-*-* } { "-O0" } { "" } } */
   4
   5 #include <stdio.h>
   6
   7 #define N (32*32*32+17)
   8 int main ()
   9 {
  10   int ary[N];
  11   int ix;
  12   int exit = 0;
  13   int ondev = 0;
  14
  15   for (ix = 0; ix < N;ix++)
  16     ary[ix] = -1;
  17
  18 #pragma acc parallel num_gangs(32) num_workers(32) vector_length(32) copy(ary) copy(ondev)
  19   {
  20 #pragma acc loop gang worker vector
  21     for (unsigned ix = 0; ix < N; ix++)
  22       {
  23         if (__builtin_acc_on_device (5))
  24           {
  25             int g = 0, w = 0, v = 0;
  26
  27             __asm__ volatile ("mov.u32 %0,%%ctaid.x;" : "=r" (g));
  28             __asm__ volatile ("mov.u32 %0,%%tid.y;" : "=r" (w));
  29             __asm__ volatile ("mov.u32 %0,%%tid.x;" : "=r" (v));
  30             ary[ix] = (g << 16) | (w << 8) | v;
  31             ondev = 1;
  32           }
  33         else
  34           ary[ix] = ix;
  35       }
  36   }
  37
  38   for (ix = 0; ix < N; ix++)
  39     {
  40       int expected = ix;
  41       if(ondev)
  42         {
  43           int chunk_size = (N + 32*32*32 - 1) / (32*32*32);
  44
  45           int g = ix / (chunk_size * 32 * 32);
  46           int w = ix / 32 % 32;
  47           int v = ix % 32;
  48
  49           expected = (g << 16) | (w << 8) | v;
  50         }
  51
  52       if (ary[ix] != expected)
  53         {
  54           exit = 1;
  55           printf ("ary[%d]=%x expected %x\n", ix, ary[ix], expected);
  56         }
  57     }
  58
  59   return exit;
  60 }