libgomp/testsuite/libgomp.oacc-c-c++-common/lib-76.c

   1 /* { dg-do run { target openacc_nvidia_accel_selected } } */
   2 /* { dg-additional-options "-lcuda" } */
   3 /* { dg-require-effective-target openacc_cuda } */
   4
   5 #include <stdio.h>
   6 #include <stdlib.h>
   7 #include <unistd.h>
   8 #include <openacc.h>
   9 #include <cuda.h>
  10 #include "timer.h"
  11
  12 int
  13 main (int argc, char **argv)
  14 {
  15   CUdevice dev;
  16   CUfunction delay;
  17   CUmodule module;
  18   CUresult r;
  19   int N;
  20   int i;
  21   CUstream *streams;
  22   unsigned long *a, *d_a, dticks;
  23   int nbytes;
  24   float atime, dtime, hitime, lotime;
  25   void *kargs[2];
  26   int clkrate;
  27   int devnum, nprocs;
  28
  29   acc_init (acc_device_nvidia);
  30
  31   devnum = acc_get_device_num (acc_device_nvidia);
  32
  33   r = cuDeviceGet (&dev, devnum);
  34   if (r != CUDA_SUCCESS)
  35     {
  36       fprintf (stderr, "cuDeviceGet failed: %d\n", r);
  37       abort ();
  38     }
  39
  40   r =
  41     cuDeviceGetAttribute (&nprocs, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT,
  42                           dev);
  43   if (r != CUDA_SUCCESS)
  44     {
  45       fprintf (stderr, "cuDeviceGetAttribute failed: %d\n", r);
  46       abort ();
  47     }
  48
  49   r = cuDeviceGetAttribute (&clkrate, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, dev);
  50   if (r != CUDA_SUCCESS)
  51     {
  52       fprintf (stderr, "cuDeviceGetAttribute failed: %d\n", r);
  53       abort ();
  54     }
  55
  56   r = cuModuleLoad (&module, "subr.ptx");
  57   if (r != CUDA_SUCCESS)
  58     {
  59       fprintf (stderr, "cuModuleLoad failed: %d\n", r);
  60       abort ();
  61     }
  62
  63   r = cuModuleGetFunction (&delay, module, "delay");
  64   if (r != CUDA_SUCCESS)
  65     {
  66       fprintf (stderr, "cuModuleGetFunction failed: %d\n", r);
  67       abort ();
  68     }
  69
  70   nbytes = nprocs * sizeof (unsigned long);
  71
  72   dtime = 200.0;
  73
  74   dticks = (unsigned long) (dtime * clkrate);
  75
  76   N = nprocs;
  77
  78   a = (unsigned long *) malloc (nbytes);
  79   d_a = (unsigned long *) acc_malloc (nbytes);
  80
  81   acc_map_data (a, d_a, nbytes);
  82
  83   streams = (CUstream *) malloc (N * sizeof (void *));
  84
  85   for (i = 0; i < N; i++)
  86     {
  87       streams[i] = (CUstream) acc_get_cuda_stream (i);
  88       if (streams[i] != NULL)
  89         abort ();
  90
  91       r = cuStreamCreate (&streams[i], CU_STREAM_DEFAULT);
  92       if (r != CUDA_SUCCESS)
  93         {
  94           fprintf (stderr, "cuStreamCreate failed: %d\n", r);
  95           abort ();
  96         }
  97
  98         if (!acc_set_cuda_stream (i, streams[i]))
  99           abort ();
 100     }
 101
 102   init_timers (1);
 103
 104   kargs[0] = (void *) &d_a;
 105   kargs[1] = (void *) &dticks;
 106
 107   start_timer (0);
 108
 109   for (i = 0; i < N; i++)
 110     {
 111       r = cuLaunchKernel (delay, 1, 1, 1, 1, 1, 1, 0, streams[i], kargs, 0);
 112       if (r != CUDA_SUCCESS)
 113         {
 114           fprintf (stderr, "cuLaunchKernel failed: %d\n", r);
 115           abort ();
 116         }
 117
 118       acc_wait (i);
 119     }
 120
 121   atime = stop_timer (0);
 122
 123   hitime = dtime * N;
 124   hitime += hitime * 0.02;
 125
 126   lotime = dtime * N;
 127   lotime -= lotime * 0.02;
 128
 129   if (atime > hitime || atime < lotime)
 130     {
 131       fprintf (stderr, "actual time < delay time\n");
 132       abort ();
 133     }
 134
 135   acc_unmap_data (a);
 136
 137   fini_timers ();
 138
 139   free (streams);
 140   free (a);
 141   acc_free (d_a);
 142
 143   acc_shutdown (acc_device_nvidia);
 144
 145   exit (0);
 146 }
 147
 148 /* { dg-output "" } */