PerformanceCounter.cc

#include "Common/Config.h"

#include "Common/OpenCL_Support.h"
#include "Common/PerformanceCounter.h"

#include <iomanip>
#include <iostream>


unsigned PerformanceCounter::nrCounters;


PerformanceCounter::PerformanceCounter(const std::string &name, bool profiling)
:
#if defined MEASURE_POWER
  totalJoules(0),
#endif
  totalNrOperations(0),
  totalNrBytesRead(0),
  totalNrBytesWritten(0),
  totalTime(0),
  totalEvents(0),
  name(name),
  profiling(profiling)
{
//#pragma omp atomic
  counterNumber = nrCounters ++;
}


PerformanceCounter::~PerformanceCounter() noexcept(false)
{
  if (totalTime > 0)
#pragma omp critical (cout)
  {
    std::cout << std::setw(12) << name
	      << std::setprecision(3)
	      << ": " << totalTime / totalEvents * 1e3 << " ms";

    if (totalNrOperations != 0)
      std::cout << ", " << totalNrOperations / totalTime * 1e-12 << " TFLOPS";

    if (totalNrBytesRead != 0 || totalNrBytesWritten != 0)
      std::cout << ", " << totalNrBytesRead / totalTime * 1e-9 << '+'
		<< totalNrBytesWritten / totalTime * 1e-9 << '='
		<< (totalNrBytesRead + totalNrBytesWritten) / totalTime * 1e-9 << " GB/s (R+W)";

#if defined MEASURE_POWER
    if (totalNrOperations > 0)
      std::cout << ", " << totalJoules / totalTime << " W"
		", " << totalNrOperations / totalJoules * 1e-9 << " GFLOPS/W";
#endif

    std::cout << std::endl;
  }
}


void PerformanceCounter::eventCompleteCallBack(cl_event ev, cl_int /*status*/, void *counter)
{
#if 0
  try {
    clRetainEvent(ev); // otherwise event.~Event() will destroy object
    cl::Event event(ev);

    cl_ulong queued = event.getProfilingInfo<CL_PROFILING_COMMAND_QUEUED>();
    cl_ulong submitted = event.getProfilingInfo<CL_PROFILING_COMMAND_SUBMIT>();
    cl_ulong start = event.getProfilingInfo<CL_PROFILING_COMMAND_START>();
    cl_ulong stop = event.getProfilingInfo<CL_PROFILING_COMMAND_END>();
    double seconds = (stop - start) / 1e9;