int main(int argc, const char * argv[])
{
uint32_t maxThreads = GetNumHardwareThreads();
double* avSpeedUps = new double[ maxThreads ];
for( uint32_t numThreads = 1; numThreads <= maxThreads; ++numThreads )
{
g_TS.Initialize(numThreads);
double avSpeedUp = 0.0;
for( int run = 0; run< REPEATS; ++run )
{
printf("Run %d.....\n", run);
Timer tParallel;
tParallel.Start();
ParallelReductionSumTaskSet m_ParallelReductionSumTaskSet( 10 * 1024 * 1024 );
g_TS.AddTaskSetToPipe( &m_ParallelReductionSumTaskSet );
g_TS.WaitforTaskSet( &m_ParallelReductionSumTaskSet );
tParallel.Stop();
printf("Parallel Example complete in \t%fms,\t sum: %" PRIu64 "\n", tParallel.GetTimeMS(), m_ParallelReductionSumTaskSet.m_FinalSum );
Timer tSerial;
tSerial.Start();
uint64_t sum = 0;
for( uint64_t i = 0; i < (uint64_t)m_ParallelReductionSumTaskSet.m_ParallelSumTaskSet.m_SetSize; ++i )
{
sum += i + 1;
}
tSerial.Stop();
if( run >= WARMUPS )
{
avSpeedUp += tSerial.GetTimeMS() / tParallel.GetTimeMS() / RUNS;
}
printf("Serial Example complete in \t%fms,\t sum: %" PRIu64 "\n", tSerial.GetTimeMS(), sum );
printf("Speed Up Serial / Parallel: %f\n\n", tSerial.GetTimeMS() / tParallel.GetTimeMS() );
}
avSpeedUps[numThreads-1] = avSpeedUp;
printf("\nAverage Speed Up for %d Hardware Threads Serial / Parallel: %f\n", numThreads, avSpeedUp );
}
printf("\nHardware Threads, Av Speed Up/s\n" );
for( uint32_t numThreads = 1; numThreads <= maxThreads; ++numThreads )
{
printf("%d, %f\n", numThreads, avSpeedUps[numThreads-1] );
}
return 0;
}
int main(int argc, const char * argv[])
{
Remotery* rmt;
rmt_CreateGlobalInstance(&rmt);
// Set the callbacks BEFORE initialize or we will get no threadstart nor first waitStart calls
g_TS.GetProfilerCallbacks()->threadStart = threadStartCallback;
g_TS.GetProfilerCallbacks()->waitStart = waitStartCallback;
g_TS.GetProfilerCallbacks()->waitStop = waitStopCallback;
g_TS.Initialize();
rmt_SetCurrentThreadName("Main");
double avSpeedUp = 0.0;
for( int run = 0; run< RUNS; ++run )
{
rmt_ScopedCPUSample(Run);
printf("Run %d.....\n", run);
ParallelReductionSumTaskSet m_ParallelReductionSumTaskSet( SUMS );
{
rmt_ScopedCPUSample(Parallel);
m_ParallelReductionSumTaskSet.Init();
g_TS.AddTaskSetToPipe(&m_ParallelReductionSumTaskSet);
g_TS.WaitforTaskSet(&m_ParallelReductionSumTaskSet);
}
volatile uint64_t sum = 0;
{
rmt_ScopedCPUSample(Serial);
for (uint64_t i = 0; i < (uint64_t)m_ParallelReductionSumTaskSet.m_ParallelSumTaskSet.m_SetSize; ++i)
{
sum += i + 1;
}
}
}
rmt_DestroyGlobalInstance(rmt);
return 0;
}
