1  
       2  #include <stdio.h>
       3  #include <cuda.h>
       4  
       5  static int _Tnum_timers;
       6  static CUevent *_Tstart_events, *_Tstop_events;
       7  static CUstream _Tstream;
       8  
       9  void
      10  init_timers (int ntimers)
      11  {
      12    int i;
      13    CUresult r;
      14  
      15    _Tnum_timers = ntimers;
      16  
      17    _Tstart_events = (CUevent *) malloc (_Tnum_timers * sizeof (CUevent));
      18    _Tstop_events = (CUevent *) malloc (_Tnum_timers * sizeof (CUevent));
      19  
      20    r = cuStreamCreate (&_Tstream, CU_STREAM_DEFAULT);
      21    if (r != CUDA_SUCCESS)
      22      {
      23        fprintf (stderr, "cuStreamCreate failed: %d\n", r);
      24        abort ();
      25      }
      26  
      27    for (i = 0; i < _Tnum_timers; i++)
      28      {
      29        r = cuEventCreate (&_Tstart_events[i], CU_EVENT_DEFAULT);
      30        if (r != CUDA_SUCCESS)
      31  	{
      32  	  fprintf (stderr, "cuEventCreate failed: %d\n", r);
      33  	  abort ();
      34  	}
      35  
      36        r = cuEventCreate (&_Tstop_events[i], CU_EVENT_DEFAULT);
      37        if (r != CUDA_SUCCESS)
      38  	{
      39  	  fprintf (stderr, "cuEventCreate failed: %d\n", r);
      40  	  abort ();
      41  	}
      42      }
      43  }
      44  
      45  void
      46  fini_timers (void)
      47  {
      48    int i;
      49  
      50    for (i = 0; i < _Tnum_timers; i++)
      51      {
      52        cuEventDestroy (_Tstart_events[i]);
      53        cuEventDestroy (_Tstop_events[i]);
      54      }
      55  
      56    cuStreamDestroy (_Tstream);
      57  
      58    free (_Tstart_events);
      59    free (_Tstop_events);
      60  }
      61  
      62  void
      63  start_timer (int timer)
      64  {
      65    CUresult r;
      66  
      67    r = cuEventRecord (_Tstart_events[timer], _Tstream);
      68    if (r != CUDA_SUCCESS)
      69      {
      70        fprintf (stderr, "cuEventRecord failed: %d\n", r);
      71        abort ();
      72      }
      73  }
      74  
      75  float
      76  stop_timer (int timer)
      77  {
      78    CUresult r;
      79    float etime;
      80  
      81    r = cuEventRecord (_Tstop_events[timer], _Tstream);
      82    if (r != CUDA_SUCCESS)
      83      {
      84        fprintf (stderr, "cuEventRecord failed: %d\n", r);
      85        abort ();
      86      }
      87  
      88    r = cuEventSynchronize (_Tstop_events[timer]);
      89    if (r != CUDA_SUCCESS)
      90      {
      91        fprintf (stderr, "cuEventSynchronize failed: %d\n", r);
      92        abort ();
      93      }
      94  
      95    r = cuEventElapsedTime (&etime, _Tstart_events[timer], _Tstop_events[timer]);
      96    if (r != CUDA_SUCCESS)
      97      {
      98        fprintf (stderr, "cuEventElapsedTime failed: %d\n", r);
      99        abort ();
     100      }
     101  
     102    return etime;
     103  }