libgomp/testsuite/libgomp.oacc-c-c++-common/lib-75.c

   1 /* { dg-do run { target openacc_nvidia_accel_selected } } */
   2 /* { dg-additional-options "-lcuda" } */
   3
   4 #include <stdio.h>
   5 #include <unistd.h>
   6 #include <stdlib.h>
   7 #include <openacc.h>
   8 #include <cuda.h>
   9 #include "timer.h"
  10
  11 int
  12 main (int argc, char **argv)
  13 {
  14   CUdevice dev;
  15   CUfunction delay;
  16   CUmodule module;
  17   CUresult r;
  18   int N;
  19   int i;
  20   CUstream stream;
  21   unsigned long *a, *d_a, dticks;
  22   int nbytes;
  23   float atime, dtime, hitime, lotime;
  24   void *kargs[2];
  25   int clkrate;
  26   int devnum, nprocs;
  27
  28   acc_init (acc_device_nvidia);
  29
  30   devnum = acc_get_device_num (acc_device_nvidia);
  31
  32   r = cuDeviceGet (&dev, devnum);
  33   if (r != CUDA_SUCCESS)
  34     {
  35       fprintf (stderr, "cuDeviceGet failed: %d\n", r);
  36       abort ();
  37     }
  38
  39   r =
  40     cuDeviceGetAttribute (&nprocs, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
  41                           dev);
  42   if (r != CUDA_SUCCESS)
  43     {
  44       fprintf (stderr, "cuDeviceGetAttribute failed: %d\n", r);
  45       abort ();
  46     }
  47
  48   r = cuDeviceGetAttribute (&clkrate, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, dev);
  49   if (r != CUDA_SUCCESS)
  50     {
  51       fprintf (stderr, "cuDeviceGetAttribute failed: %d\n", r);
  52       abort ();
  53     }
  54
  55   r = cuModuleLoad (&module, "subr.ptx");
  56   if (r != CUDA_SUCCESS)
  57     {
  58       fprintf (stderr, "cuModuleLoad failed: %d\n", r);
  59       abort ();
  60     }
  61
  62   r = cuModuleGetFunction (&delay, module, "delay");
  63   if (r != CUDA_SUCCESS)
  64     {
  65       fprintf (stderr, "cuModuleGetFunction failed: %d\n", r);
  66       abort ();
  67     }
  68
  69   nbytes = nprocs * sizeof (unsigned long);
  70
  71   dtime = 200.0;
  72
  73   dticks = (unsigned long) (dtime * clkrate);
  74
  75   N = nprocs;
  76
  77   a = (unsigned long *) malloc (nbytes);
  78   d_a = (unsigned long *) acc_malloc (nbytes);
  79
  80   acc_map_data (a, d_a, nbytes);
  81
  82   stream = (CUstream) acc_get_cuda_stream (0);
  83   if (stream != NULL)
  84     abort ();
  85
  86   r = cuStreamCreate (&stream, CU_STREAM_DEFAULT);
  87   if (r != CUDA_SUCCESS)
  88     {
  89       fprintf (stderr, "cuStreamCreate failed: %d\n", r);
  90       abort ();
  91     }
  92
  93   if (!acc_set_cuda_stream (0, stream))
  94     abort ();
  95
  96   init_timers (1);
  97
  98   kargs[0] = (void *) &d_a;
  99   kargs[1] = (void *) &dticks;
 100
 101   start_timer (0);
 102
 103   for (i = 0; i < N; i++)
 104     {
 105       r = cuLaunchKernel (delay, 1, 1, 1, 1, 1, 1, 0, stream, kargs, 0);
 106       if (r != CUDA_SUCCESS)
 107         {
 108           fprintf (stderr, "cuLaunchKernel failed: %d\n", r);
 109           abort ();
 110         }
 111
 112       acc_wait (0);
 113     }
 114
 115   atime = stop_timer (0);
 116
 117   hitime = dtime * N;
 118   hitime += hitime * 0.02;
 119
 120   lotime = dtime * N;
 121   lotime -= lotime * 0.02;
 122
 123   if (atime > hitime || atime < lotime)
 124     {
 125       fprintf (stderr, "actual time < delay time\n");
 126       abort ();
 127     }
 128
 129   acc_unmap_data (a);
 130
 131   fini_timers ();
 132
 133   free (a);
 134   acc_free (d_a);
 135
 136   acc_shutdown (acc_device_nvidia);
 137
 138   exit (0);
 139 }
 140
 141 /* { dg-output "" } */