libgomp/testsuite/libgomp.oacc-c-c++-common/routine-wv-2.c

/* This code uses nvptx inline assembly guarded with acc_on_device, which is
   not optimized away at -O0, and then confuses the target assembler.
   { dg-skip-if "" { *-*-* } { "-O0" } { "" } } */

#include <stdio.h>
#include <openacc.h>

#define NUM_WORKERS 16
#define NUM_VECTORS 32
#define WIDTH 64
#define HEIGHT 32

#define WORK_ID(I,N)						\
  (acc_on_device (acc_device_nvidia)				\
   ? ({unsigned __r;						\
       __asm__ volatile ("mov.u32 %0,%%tid.y;" : "=r" (__r));	\
       __r; }) : (I % N))
#define VEC_ID(I,N)						\
  (acc_on_device (acc_device_nvidia)				\
   ? ({unsigned __r;						\
       __asm__ volatile ("mov.u32 %0,%%tid.x;" : "=r" (__r));	\
       __r; }) : (I % N))

#pragma acc routine worker
void __attribute__ ((noinline))
  WorkVec (int *ptr, int w, int h, int nw, int nv)
{
#pragma acc loop worker
  for (int i = 0; i < h; i++)
#pragma acc loop vector
    for (int j = 0; j < w; j++)
      ptr[i*w + j] = (WORK_ID (i, nw) << 8) | VEC_ID(j, nv);
}

int DoWorkVec (int nw)
{
  int ary[HEIGHT][WIDTH];
  int err = 0;

  for (int ix = 0; ix != HEIGHT; ix++)
    for (int jx = 0; jx != WIDTH; jx++)
      ary[ix][jx] = 0xdeadbeef;

  printf ("spawning %d ...", nw); fflush (stdout);
  
#pragma acc parallel num_workers(nw) vector_length (NUM_VECTORS) copy (ary)
  {
    WorkVec ((int *)ary, WIDTH, HEIGHT, nw, NUM_VECTORS);
  }

  for (int ix = 0; ix != HEIGHT; ix++)
    for (int jx = 0; jx != WIDTH; jx++)
      {
	int exp = ((ix % nw) << 8) | (jx % NUM_VECTORS);
	
	if (ary[ix][jx] != exp)
	  {
	    printf ("\nary[%d][%d] = %#x expected %#x", ix, jx,
		    ary[ix][jx], exp);
	    err = 1;
	  }
      }
  printf (err ? " failed\n" : " ok\n");
  
  return err;
}

int main ()
{
  int err = 0;

  for (int W = 1; W <= NUM_WORKERS; W <<= 1)
    err |= DoWorkVec (W);

  return err;
}