#include <windows.h>
#include <process.h>
#include <cmath>
#include <iostream>
using namespace std;

static const int NumThreads = 20;

LARGE_INTEGER g_freq, g_bef;
volatile long g_go = 0, g_count = 0;

void lot_o_math()   // just waste a reliable amount of time
{
  float x = 1;
  for(size_t i=0; i<1000000; ++i)
    x = sin(x);
}

unsigned int __stdcall do_math(void*)
{
  while(!g_go);
  lot_o_math();
  InterlockedIncrement(&g_count);

  if (g_count == NumThreads) // if last thread, display timing results
  {
    LARGE_INTEGER aft;
    QueryPerformanceCounter(&aft);
    cout << "time: " << double(aft.QuadPart-g_bef.QuadPart)/g_freq.QuadPart << endl;
    ++g_count;
  }
  return 0;
}

int main()
{
  LARGE_INTEGER bef, aft;
  QueryPerformanceFrequency(&g_freq);

  // single threaded
  QueryPerformanceCounter(&bef);
  for(size_t i=0; i<NumThreads; ++i)
    lot_o_math(); 
  QueryPerformanceCounter(&aft);
  cout << "time: " << double(aft.QuadPart-bef.QuadPart)/g_freq.QuadPart << endl;

  // create a bunch of threads (busy waiting on g_go)
  for(size_t i=0; i<20; ++i)
    CloseHandle((HANDLE)_beginthreadex(0, 0, do_math, 0, 0, 0));

  // record start time and release the threads
  QueryPerformanceCounter(&g_bef);
  InterlockedIncrement(&g_go);
  
  // wait until all threads are finished
  while(g_count < NumThreads+1)
    Sleep(1000);

  cout << "Enter a key" << endl;
  char c;
  cin >> c;
}

Quote:Original post by Iftah
switching between threads isnt much more than pushing the IP (instruction pointer) and stack pointer and popping two others - almost same performance hit of calling a function.