#include <thread>
#include <atomic>

#define THREADPOOL_THREADMAX (16)

class ThreadPool
{
public:
    typedef uint64_t Progress;

    template<typename T>
    void Start(T Function, const size_t NewThreadNum)
    {
        if (ThreadNum > 0) return;//läuft schon
        ThreadNum = NewThreadNum;

        for (int i = 0; i < ThreadNum; ++i)
            Thread[i] = std::thread(Function);
    }
    void Join()
    {
        for (int i = 0; i < ThreadNum; ++i)
            Thread[i].join();
        ThreadNum = 0;
    }

    void Sync(Progress& LocalProgress)
    {
        ++LocalProgress;
        if ((FinishCounter.fetch_add(1) + 1) == ThreadNum)
        {
            FinishCounter = 0;
            ++GlobalProgress;
        }

        while (LocalProgress > GlobalProgress)
            std::this_thread::sleep_for(std::chrono::milliseconds(0));
    }

private:
    size_t ThreadNum = 0;

    std::thread Thread[THREADPOOL_THREADMAX];
    std::atomic<size_t> FinishCounter = 0;
    Progress GlobalProgress = 0;
};