libgomp/testsuite/libgomp.oacc-c-c++-common/routine-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
   9 #pragma acc routine gang
  10 void __attribute__ ((noinline)) gang (int ary[N])
  11 {
  12 #pragma acc loop gang worker vector
  13   for (unsigned ix = 0; ix < N; ix++)
  14     {
  15       if (__builtin_acc_on_device (5))
  16         {
  17           int g = 0, w = 0, v = 0;
  18
  19           __asm__ volatile ("mov.u32 %0,%%ctaid.x;" : "=r" (g));
  20           __asm__ volatile ("mov.u32 %0,%%tid.y;" : "=r" (w));
  21           __asm__ volatile ("mov.u32 %0,%%tid.x;" : "=r" (v));
  22           ary[ix] = (g << 16) | (w << 8) | v;
  23         }
  24       else
  25         ary[ix] = ix;
  26     }
  27 }
  28
  29 int main ()
  30 {
  31   int ary[N];
  32   int ix;
  33   int exit = 0;
  34   int ondev = 0;
  35
  36   for (ix = 0; ix < N;ix++)
  37     ary[ix] = -1;
  38
  39 #pragma acc parallel num_gangs(32) num_workers(32) vector_length(32) copy(ary) copy(ondev)
  40   {
  41     ondev = __builtin_acc_on_device (5);
  42     gang (ary);
  43   }
  44
  45   for (ix = 0; ix < N; ix++)
  46     {
  47       int expected = ix;
  48       if(ondev)
  49         {
  50           int chunk_size = (N + 32*32*32 - 1) / (32*32*32);
  51
  52           int g = ix / (chunk_size * 32 * 32);
  53           int w = ix / 32 % 32;
  54           int v = ix % 32;
  55
  56           expected = (g << 16) | (w << 8) | v;
  57         }
  58
  59       if (ary[ix] != expected)
  60         {
  61           exit = 1;
  62           printf ("ary[%d]=%x expected %x\n", ix, ary[ix], expected);
  63         }
  64     }
  65
  66   return exit;
  67 }