#include <ThreadPool.hh>

Public Types
template<typename KeyT , typename MappedT , typename HashT = KeyT>
using	uomap = std::unordered_map< KeyT, MappedT, std::hash< HashT > >

using	size_type = size_t

using	task_count_type = std::shared_ptr< std::atomic_uintmax_t >

using	atomic_int_type = std::shared_ptr< std::atomic_uintmax_t >

using	pool_state_type = std::shared_ptr< std::atomic_short >

using	atomic_bool_type = std::shared_ptr< std::atomic_bool >

using	task_type = VTask

using	lock_t = std::shared_ptr< Mutex >

using	condition_t = std::shared_ptr< Condition >

using	task_pointer = task_type *

using	task_queue_t = VUserTaskQueue

typedef std::deque< ThreadId >	thread_list_t

typedef std::vector< bool >	bool_list_t

typedef std::map< ThreadId, uintmax_t >	thread_id_map_t

typedef std::map< uintmax_t, ThreadId >	thread_index_map_t

using	thread_vec_t = std::vector< Thread >

typedef std::function< void()>	initialize_func_t

typedef std::function< intmax_t(intmax_t)>	affinity_func_t

Public Member Functions
	ThreadPool (const size_type &pool_size, VUserTaskQueue *task_queue=nullptr, bool _use_affinity=GetEnv< bool >("PTL_CPU_AFFINITY", false), const affinity_func_t &=[](intmax_t) { static std::atomic< intmax_t > assigned;intmax_t _assign=assigned++;return _assign % Thread::hardware_concurrency();})

virtual	~ThreadPool ()

	ThreadPool (const ThreadPool &)=delete

	ThreadPool (ThreadPool &&)=default

ThreadPool &	operator= (const ThreadPool &)=delete

ThreadPool &	operator= (ThreadPool &&)=default

size_type	initialize_threadpool (size_type)

size_type	destroy_threadpool ()

size_type	stop_thread ()

template<typename FuncT >
void	execute_on_all_threads (FuncT &&_func)

size_type	add_task (task_pointer task, int bin=-1)

template<typename ListT >
size_type	add_tasks (ListT &)

Thread *	get_thread (size_type _n) const

Thread *	get_thread (std::thread::id id) const

task_queue_t *	get_queue () const

void	set_initialization (initialize_func_t f)

void	reset_initialization ()

const pool_state_type &	state () const

size_type	size () const

void	resize (size_type _n)

bool	using_affinity () const

bool	is_alive ()

void	notify ()

void	notify_all ()

void	notify (size_type)

bool	is_initialized () const

int	get_active_threads_count () const

void	set_affinity (affinity_func_t f)

void	set_affinity (intmax_t i, Thread &)

void	set_verbose (int n)

int	get_verbose () const

bool	is_master () const

Static Public Member Functions
static bool	using_tbb ()

static void	set_use_tbb (bool val)

static tbb_global_control_t *&	tbb_global_control ()

static const thread_id_map_t &	get_thread_ids ()

static uintmax_t	get_this_thread_id ()

Protected Member Functions
void	execute_thread (VUserTaskQueue *)

int	insert (const task_pointer &, int=-1)

int	run_on_this (task_pointer)

void	record_entry ()

void	record_exit ()

Static Protected Member Functions
static void	start_thread (ThreadPool *, intmax_t=-1)

Detailed Description

Definition at line 63 of file ThreadPool.hh.

Member Typedef Documentation

◆ affinity_func_t

typedef std::function<intmax_t(intmax_t)> PTL::ThreadPool::affinity_func_t

Definition at line 89 of file ThreadPool.hh.

◆ atomic_bool_type

using PTL::ThreadPool::atomic_bool_type = std::shared_ptr<std::atomic_bool>

Definition at line 74 of file ThreadPool.hh.

◆ atomic_int_type

using PTL::ThreadPool::atomic_int_type = std::shared_ptr<std::atomic_uintmax_t>

Definition at line 72 of file ThreadPool.hh.

◆ bool_list_t

typedef std::vector<bool> PTL::ThreadPool::bool_list_t

Definition at line 83 of file ThreadPool.hh.

◆ condition_t

using PTL::ThreadPool::condition_t = std::shared_ptr<Condition>

Definition at line 78 of file ThreadPool.hh.

◆ initialize_func_t

typedef std::function<void()> PTL::ThreadPool::initialize_func_t

Definition at line 88 of file ThreadPool.hh.

◆ lock_t

using PTL::ThreadPool::lock_t = std::shared_ptr<Mutex>

Definition at line 77 of file ThreadPool.hh.

◆ pool_state_type

using PTL::ThreadPool::pool_state_type = std::shared_ptr<std::atomic_short>

Definition at line 73 of file ThreadPool.hh.

◆ size_type

using PTL::ThreadPool::size_type = size_t

Definition at line 70 of file ThreadPool.hh.

◆ task_count_type

using PTL::ThreadPool::task_count_type = std::shared_ptr<std::atomic_uintmax_t>

Definition at line 71 of file ThreadPool.hh.

◆ task_pointer

using PTL::ThreadPool::task_pointer = task_type*

Definition at line 79 of file ThreadPool.hh.

◆ task_queue_t

using PTL::ThreadPool::task_queue_t = VUserTaskQueue

Definition at line 80 of file ThreadPool.hh.

◆ task_type

using PTL::ThreadPool::task_type = VTask

Definition at line 76 of file ThreadPool.hh.

◆ thread_id_map_t

typedef std::map<ThreadId, uintmax_t> PTL::ThreadPool::thread_id_map_t

Definition at line 84 of file ThreadPool.hh.

◆ thread_index_map_t

typedef std::map<uintmax_t, ThreadId> PTL::ThreadPool::thread_index_map_t

Definition at line 85 of file ThreadPool.hh.

◆ thread_list_t

typedef std::deque<ThreadId> PTL::ThreadPool::thread_list_t

Definition at line 82 of file ThreadPool.hh.

◆ thread_vec_t

using PTL::ThreadPool::thread_vec_t = std::vector<Thread>

Definition at line 86 of file ThreadPool.hh.

◆ uomap

template<typename KeyT , typename MappedT , typename HashT = KeyT>

using PTL::ThreadPool::uomap = std::unordered_map<KeyT, MappedT, std::hash<HashT> >

Definition at line 67 of file ThreadPool.hh.

Constructor & Destructor Documentation

◆ ThreadPool() [1/3]

ThreadPool::ThreadPool	(	const size_type &	pool_size,
		VUserTaskQueue *	task_queue = `nullptr`,
		bool	_use_affinity = `GetEnv<bool>("PTL_CPU_AFFINITY", false)`,
		const affinity_func_t &	_affinity_func = `[](intmax_t) { static std::atomic<intmax_t> assigned; intmax_t _assign = assigned++; return _assign % Thread::hardware_concurrency(); }`
	)

Definition at line 138 of file ThreadPool.cc.

: m_use_affinity(_use_affinity)
, m_tbb_tp(false)
, m_verbose(0)
, m_pool_size(0)
, m_master_tid(ThisThread::get_id())
, m_alive_flag(std::make_shared<std::atomic_bool>(false))
, m_pool_state(std::make_shared<std::atomic_short>(thread_pool::state::NONINIT))
, m_thread_awake(std::make_shared<std::atomic_uintmax_t>(0))
, m_task_lock(std::make_shared<Mutex>())
, m_task_cond(std::make_shared<Condition>())
, m_task_queue(task_queue)
, m_tbb_task_group(nullptr)
, m_init_func([]() { return; })
, m_affinity_func(_affinity_func)
{
    m_verbose = GetEnv<int>("PTL_VERBOSE", m_verbose);
 
    auto master_id = get_this_thread_id();
    if(master_id != 0 && m_verbose > 1)
        std::cerr << "ThreadPool created on non-master slave" << std::endl;
 
    thread_data() = new ThreadData(this);
 
    // initialize after get_this_thread_id so master is zero
    this->initialize_threadpool(pool_size);
 
    if(!m_task_queue)
        m_task_queue = new UserTaskQueue(m_pool_size);
}

◆ ~ThreadPool()

ThreadPool::~ThreadPool ( )

virtual

Definition at line 172 of file ThreadPool.cc.

{
    if(m_alive_flag->load())
    {
        std::cerr << "Warning! ThreadPool was not properly destroyed! Call "
                     "destroy_threadpool() before deleting the ThreadPool object to "
                     "eliminate this message."
                  << std::endl;
        m_pool_state->store(thread_pool::state::STOPPED);
        m_task_lock->lock();
        CONDITIONBROADCAST(m_task_cond.get());
        m_task_lock->unlock();
        for(auto& itr : m_threads)
            itr.join();
        m_threads.clear();
    }
}

◆ ThreadPool() [2/3]

PTL::ThreadPool::ThreadPool ( const ThreadPool & )

delete

◆ ThreadPool() [3/3]

PTL::ThreadPool::ThreadPool ( ThreadPool && )

default

Member Function Documentation

◆ add_task()

ThreadPool::size_type PTL::ThreadPool::add_task	(	task_pointer	task,
		int	bin = `-1`
	)

inline

Definition at line 328 of file ThreadPool.hh.

{
    // if not native (i.e. TBB) then return
    if(!task->is_native_task())
        return 0;
 
    // if we haven't built thread-pool, just execute
    if(!m_alive_flag->load())
        return static_cast<size_type>(run_on_this(task));
 
    return static_cast<size_type>(insert(task, bin));
}

Referenced by PTL::TaskManager::async(), PTL::TaskGroup< Tp, Arg >::exec(), PTL::TaskManager::exec(), and PTL::TaskGroup< Tp, Arg >::run().

◆ add_tasks()

template<typename ListT >

ThreadPool::size_type PTL::ThreadPool::add_tasks ( ListT & c )

inline

Definition at line 343 of file ThreadPool.hh.

{
    if(!m_alive_flag)  // if we haven't built thread-pool, just execute
    {
        for(auto& itr : c)
            run(itr);
        c.clear();
        return 0;
    }
 
    // TODO: put a limit on how many tasks can be added at most
    auto c_size = c.size();
    for(auto& itr : c)
    {
        if(!itr->is_native_task())
            --c_size;
        else
        {
            //++(m_task_queue);
            m_task_queue->InsertTask(itr);
        }
    }
    c.clear();
 
    // notify sleeping threads
    notify(c_size);
 
    return c_size;
}

◆ destroy_threadpool()

ThreadPool::size_type ThreadPool::destroy_threadpool ( )

Definition at line 366 of file ThreadPool.cc.

{
    // Note: this is not for synchronization, its for thread communication!
    // destroy_threadpool() will only be called from the main thread, yet
    // the modified m_pool_state may not show up to other threads until its
    // modified in a lock!
    //------------------------------------------------------------------------//
    m_pool_state->store(thread_pool::state::STOPPED);
 
    //--------------------------------------------------------------------//
    // handle tbb task scheduler
#ifdef PTL_USE_TBB
    if(m_tbb_task_group)
    {
        m_tbb_task_group->wait();
        delete m_tbb_task_group;
        m_tbb_task_group = nullptr;
    }
    if(m_tbb_tp && tbb_global_control())
    {
        tbb_global_control_t*& _global_control = tbb_global_control();
        delete _global_control;
        _global_control = nullptr;
        m_tbb_tp        = false;
        std::cout << "ThreadPool [TBB] destroyed" << std::endl;
    }
#endif
 
    if(!m_alive_flag->load())
        return 0;
 
    //------------------------------------------------------------------------//
    // notify all threads we are shutting down
    m_task_lock->lock();
    CONDITIONBROADCAST(m_task_cond.get());
    m_task_lock->unlock();
    //------------------------------------------------------------------------//
 
    if(m_is_joined.size() != m_main_threads.size())
    {
        std::stringstream ss;
        ss << "   ThreadPool::destroy_thread_pool - boolean is_joined vector "
           << "is a different size than threads vector: " << m_is_joined.size() << " vs. "
           << m_main_threads.size() << " (tid: " << std::this_thread::get_id() << ")";
 
        throw std::runtime_error(ss.str());
    }
 
    for(size_type i = 0; i < m_is_joined.size(); i++)
    {
        //--------------------------------------------------------------------//
        //
        if(i < m_threads.size())
            m_threads.at(i).join();
 
        //--------------------------------------------------------------------//
        // if its joined already, nothing else needs to be done
        if(m_is_joined.at(i))
            continue;
 
        //--------------------------------------------------------------------//
        // join
        if(std::this_thread::get_id() == m_main_threads[i])
            continue;
 
        //--------------------------------------------------------------------//
        // thread id and index
        auto _tid = m_main_threads[i];
 
        //--------------------------------------------------------------------//
        // erase thread from thread ID list
        if(f_thread_ids.find(_tid) != f_thread_ids.end())
            f_thread_ids.erase(f_thread_ids.find(_tid));
 
        //--------------------------------------------------------------------//
        // it's joined
        m_is_joined.at(i) = true;
 
        //--------------------------------------------------------------------//
        // try waking up a bunch of threads that are still waiting
        CONDITIONBROADCAST(m_task_cond.get());
        //--------------------------------------------------------------------//
    }
 
    m_threads.clear();
    m_main_threads.clear();
    m_is_joined.clear();
    m_alive_flag->store(false);
 
    auto start   = std::chrono::steady_clock::now();
    auto elapsed = std::chrono::duration<double>{};
    // wait maximum of 30 seconds for threads to exit
    while(m_thread_active->load() > 0 && elapsed.count() < 30)
    {
        std::this_thread::sleep_for(std::chrono::milliseconds(50));
        elapsed = std::chrono::steady_clock::now() - start;
    }
 
    auto _active = m_thread_active->load();
 
    if(_active == 0)
        std::cout << "ThreadPool destroyed" << std::endl;
    else
        std::cout << "ThreadPool destroyed but " << _active
                  << " threads might still be active (and cause a termination error)"
                  << std::endl;
 
    return 0;
}

Referenced by PTL::TaskManager::finalize(), PTL::TaskRunManager::Terminate(), and G4TaskRunManager::~G4TaskRunManager().

◆ execute_on_all_threads()

template<typename FuncT >

void PTL::ThreadPool::execute_on_all_threads ( FuncT && _func )

inline

Definition at line 375 of file ThreadPool.hh.

{
    if(m_tbb_tp && m_tbb_task_group)
    {
#if defined(PTL_USE_TBB)
        // TBB lazily activates threads to process tasks and the master thread
        // participates in processing the tasks so getting a specific
        // function to execute only on the worker threads requires some trickery
        //
        auto                      master_tid = ThisThread::get_id();
        std::set<std::thread::id> _first;
        Mutex                     _mutex;
        // init function which executes function and returns 1 only once
        auto _init = [&]() {
            static thread_local int _once = 0;
            _mutex.lock();
            if(_first.find(std::this_thread::get_id()) == _first.end())
            {
                // we need to reset this thread-local static for multiple invocations
                // of the same template instantiation
                _once = 0;
                _first.insert(std::this_thread::get_id());
            }
            _mutex.unlock();
            if(_once++ == 0)
            {
                _func();
                return 1;
            }
            return 0;
        };
        // consumes approximately N milliseconds of cpu time
        auto _consume = [](long n) {
            using stl_mutex_t   = std::mutex;
            using unique_lock_t = std::unique_lock<stl_mutex_t>;
            // a mutex held by one lock
            stl_mutex_t mutex;
            // acquire lock
            unique_lock_t hold_lk(mutex);
            // associate but defer
            unique_lock_t try_lk(mutex, std::defer_lock);
            // get current time
            auto now = std::chrono::steady_clock::now();
            // try until time point
            while(std::chrono::steady_clock::now() < (now + std::chrono::milliseconds(n)))
                try_lk.try_lock();
        };
        // this will collect the number of threads which have
        // executed the _init function above
        std::atomic<size_t> _total_init{ 0 };
        // this is the task passed to the task-group
        auto _init_task = [&]() {
            int _ret = 0;
            // don't let the master thread execute the function
            if(ThisThread::get_id() != master_tid)
            {
                // execute the function
                _ret = _init();
                // add the result
                _total_init += _ret;
            }
            // if the function did not return anything, put it to sleep
            // so TBB will wake other threads to execute the remaining tasks
            if(_ret == 0)
                _consume(100);
        };
 
        // TBB won't oversubscribe so we need to limit by ncores - 1
        size_t nitr  = 0;
        size_t _maxp = tbb_global_control()->active_value(
            tbb::global_control::max_allowed_parallelism);
        size_t _sz         = size();
        size_t _ncore      = Threading::GetNumberOfCores() - 1;
        size_t _num        = std::min(_maxp, std::min(_sz, _ncore));
        auto   _fname      = __FUNCTION__;
        auto   _write_info = [&]() {
            std::cerr << "[" << _fname << "]> Total initalized: " << _total_init
                      << ", expected: " << _num << ", max-parallel: " << _maxp
                      << ", size: " << _sz << ", ncore: " << _ncore << std::endl;
        };
        while(_total_init < _num)
        {
            auto _n = _num;
            while(--_n > 0)
                m_tbb_task_group->run(_init_task);
            m_tbb_task_group->wait();
            // don't loop infinitely but use a strict condition
            if(nitr++ > 2 * (_num + 1) && (_total_init - 1) == _num)
            {
                _write_info();
                break;
            }
            // at this point we need to exit
            if(nitr > 4 * (_ncore + 1))
            {
                _write_info();
                break;
            }
        }
        if(get_verbose() > 3)
            _write_info();
#endif
    }
    else if(get_queue())
    {
        get_queue()->ExecuteOnAllThreads(this, std::forward<FuncT>(_func));
    }
}

Referenced by G4TaskRunManager::CreateAndStartWorkers(), G4TaskRunManager::RequestWorkersProcessCommandsStack(), G4TaskRunManager::TerminateWorkers(), and G4TaskRunManager::WaitForEndEventLoopWorkers().

◆ execute_thread()

void ThreadPool::execute_thread ( VUserTaskQueue * _task_queue )

protected

Definition at line 520 of file ThreadPool.cc.

{
    // how long the thread waits on condition variable
    // static int wait_time = GetEnv<int>("PTL_POOL_WAIT_TIME", 5);
 
    ++(*m_thread_awake);
 
    // initialization function
    m_init_func();
 
    ThreadId    tid  = ThisThread::get_id();
    ThreadData* data = thread_data();
    // auto        thread_bin = _task_queue->GetThreadBin();
    // auto        workers    = _task_queue->workers();
 
    auto start   = std::chrono::steady_clock::now();
    auto elapsed = std::chrono::duration<double>{};
    // check for updates for 60 seconds max
    while(!_task_queue && elapsed.count() < 60)
    {
        elapsed = std::chrono::steady_clock::now() - start;
        data->update();
        _task_queue = data->current_queue;
    }
 
    if(!_task_queue)
    {
        --(*m_thread_awake);
        throw std::runtime_error("No task queue was found after 60 seconds!");
    }
 
    assert(data->current_queue != nullptr);
    assert(_task_queue == data->current_queue);
 
    // essentially a dummy run
    if(_task_queue)
    {
        data->within_task = true;
        auto _task        = _task_queue->GetTask();
        if(_task)
        {
            (*_task)();
            if(!_task->group())
                delete _task;
        }
        data->within_task = false;
    }
 
    // threads stay in this loop forever until thread-pool destroyed
    while(true)
    {
        static thread_local auto p_task_lock = m_task_lock;
 
        //--------------------------------------------------------------------//
        // Try to pick a task
        AutoLock _task_lock(*p_task_lock, std::defer_lock);
        //--------------------------------------------------------------------//
 
        auto leave_pool = [&]() {
            auto _state      = [&]() { return static_cast<int>(m_pool_state->load()); };
            auto _pool_state = _state();
            if(_pool_state > 0)
            {
                // stop whole pool
                if(_pool_state == thread_pool::state::STOPPED)
                {
                    if(_task_lock.owns_lock())
                        _task_lock.unlock();
                    return true;
                }
                // single thread stoppage
                else if(_pool_state == thread_pool::state::PARTIAL)  // NOLINT
                {
                    if(!_task_lock.owns_lock())
                        _task_lock.lock();
                    if(!m_is_stopped.empty() && m_is_stopped.back())
                    {
                        m_stop_threads.push_back(tid);
                        m_is_stopped.pop_back();
                        if(_task_lock.owns_lock())
                            _task_lock.unlock();
                        // exit entire function
                        return true;
                    }
                    if(_task_lock.owns_lock())
                        _task_lock.unlock();
                }
            }
            return false;
        };
 
        // We need to put condition.wait() in a loop for two reasons:
        // 1. There can be spurious wake-ups (due to signal/ENITR)
        // 2. When mutex is released for waiting, another thread can be woken up
        //    from a signal/broadcast and that thread can mess up the condition.
        //    So when the current thread wakes up the condition may no longer be
        //    actually true!
        while(_task_queue->empty())
        {
            auto _state = [&]() { return static_cast<int>(m_pool_state->load()); };
            auto _size  = [&]() { return _task_queue->true_size(); };
            auto _empty = [&]() { return _task_queue->empty(); };
            auto _wake  = [&]() { return (!_empty() || _size() > 0 || _state() > 0); };
 
            if(leave_pool())
                return;
 
            if(_task_queue->true_size() == 0)
            {
                if(m_thread_awake && m_thread_awake->load() > 0)
                    --(*m_thread_awake);
 
                // lock before sleeping on condition
                if(!_task_lock.owns_lock())
                    _task_lock.lock();
 
                // Wait until there is a task in the queue
                // Unlocks mutex while waiting, then locks it back when signaled
                // use lambda to control waking
                m_task_cond->wait(_task_lock, _wake);
 
                if(_state() == thread_pool::state::STOPPED)
                    return;
 
                // unlock if owned
                if(_task_lock.owns_lock())
                    _task_lock.unlock();
 
                // notify that is awake
                if(m_thread_awake && m_thread_awake->load() < m_pool_size)
                    ++(*m_thread_awake);
            }
            else
                break;
        }
 
        // release the lock
        if(_task_lock.owns_lock())
            _task_lock.unlock();
 
        //----------------------------------------------------------------//
 
        // leave pool if conditions dictate it
        if(leave_pool())
            return;
 
        // activate guard against recursive deadlock
        data->within_task = true;
        //----------------------------------------------------------------//
 
        // execute the task(s)
        while(!_task_queue->empty())
        {
            auto _task = _task_queue->GetTask();
            if(_task)
            {
                (*_task)();
                if(!_task->group())
                    delete _task;
            }
        }
        //----------------------------------------------------------------//
 
        // disable guard against recursive deadlock
        data->within_task = false;
        //----------------------------------------------------------------//
    }
}

◆ get_active_threads_count()

int PTL::ThreadPool::get_active_threads_count ( ) const

inline

Definition at line 161 of file ThreadPool.hh.

    {
        return (m_thread_awake) ? m_thread_awake->load() : 0;
    }

◆ get_queue()

task_queue_t * PTL::ThreadPool::get_queue ( ) const

inline

Definition at line 135 of file ThreadPool.hh.

135{ return m_task_queue; }

Referenced by execute_on_all_threads(), PTL::ThreadData::update(), and PTL::VTaskGroup::wait().

◆ get_this_thread_id()

uintmax_t ThreadPool::get_this_thread_id ( )

static

Definition at line 120 of file ThreadPool.cc.

{
    auto _tid = ThisThread::get_id();
    {
        AutoLock lock(TypeMutex<ThreadPool>(), std::defer_lock);
        if(!lock.owns_lock())
            lock.lock();
        if(f_thread_ids.find(_tid) == f_thread_ids.end())
        {
            auto _idx          = f_thread_ids.size();
            f_thread_ids[_tid] = _idx;
        }
    }
    return f_thread_ids[_tid];
}

Referenced by PTL::UserTaskQueue::GetThreadBin(), initialize_threadpool(), G4TaskRunManagerKernel::InitializeWorker(), PTL::VTask::operator--(), and PTL::UserTaskQueue::UserTaskQueue().

◆ get_thread() [1/2]

Thread * PTL::ThreadPool::get_thread ( size_type _n ) const

◆ get_thread() [2/2]

Thread * PTL::ThreadPool::get_thread ( std::thread::id id ) const

◆ get_thread_ids()

const ThreadPool::thread_id_map_t & ThreadPool::get_thread_ids ( )

static

Definition at line 112 of file ThreadPool.cc.

{
  return f_thread_ids;
}

◆ get_verbose()

int PTL::ThreadPool::get_verbose ( ) const

inline

Definition at line 170 of file ThreadPool.hh.

170{ return m_verbose; }

Referenced by execute_on_all_threads().

◆ initialize_threadpool()

ThreadPool::size_type ThreadPool::initialize_threadpool ( size_type proposed_size )

Definition at line 223 of file ThreadPool.cc.

{
    //--------------------------------------------------------------------//
    // return before initializing
    if(proposed_size < 1)
        return 0;
 
    //--------------------------------------------------------------------//
    // store that has been started
    if(!m_alive_flag->load())
        m_pool_state->store(thread_pool::state::STARTED);
 
        //--------------------------------------------------------------------//
        // handle tbb task scheduler
#ifdef PTL_USE_TBB
    if(f_use_tbb)
    {
        m_tbb_tp                               = true;
        m_pool_size                            = proposed_size;
        tbb_global_control_t*& _global_control = tbb_global_control();
        // delete if wrong size
        if(m_pool_size != proposed_size)
        {
            delete _global_control;
            _global_control = nullptr;
        }
 
        if(!_global_control)
        {
            _global_control = new tbb_global_control_t(
                tbb::global_control::max_allowed_parallelism, proposed_size + 1);
            if(m_verbose > 0)
            {
                std::cout << "ThreadPool [TBB] initialized with " << m_pool_size
                          << " threads." << std::endl;
            }
        }
 
        // create task group (used for async)
        if(!m_tbb_task_group)
            m_tbb_task_group = new tbb_task_group_t();
        return m_pool_size;
    }
#endif
 
    m_alive_flag->store(true);
 
    //--------------------------------------------------------------------//
    // if started, stop some thread if smaller or return if equal
    if(m_pool_state->load() == thread_pool::state::STARTED)
    {
        if(m_pool_size > proposed_size)
        {
            while(stop_thread() > proposed_size)
                ;
            if(m_verbose > 0)
            {
                std::cout << "ThreadPool initialized with " << m_pool_size << " threads."
                          << std::endl;
            }
            if(!m_task_queue)
                m_task_queue = new UserTaskQueue(m_pool_size);
            return m_pool_size;
        }
        else if(m_pool_size == proposed_size)  // NOLINT
        {
            if(m_verbose > 0)
            {
                std::cout << "ThreadPool initialized with " << m_pool_size << " threads."
                          << std::endl;
            }
            if(!m_task_queue)
                m_task_queue = new UserTaskQueue(m_pool_size);
            return m_pool_size;
        }
    }
 
    //--------------------------------------------------------------------//
    // reserve enough space to prevent realloc later
    {
        AutoLock _task_lock(*m_task_lock);
        m_is_joined.reserve(proposed_size);
    }
 
    auto this_tid = get_this_thread_id();
    for(size_type i = m_pool_size; i < proposed_size; ++i)
    {
        // add the threads
        try
        {
            Thread thr(ThreadPool::start_thread, this, this_tid + i + 1);
            // only reaches here if successful creation of thread
            ++m_pool_size;
            // store thread
            m_main_threads.push_back(thr.get_id());
            // list of joined thread booleans
            m_is_joined.push_back(false);
            // set the affinity
            if(m_use_affinity)
                set_affinity(i, thr);
            // store
            m_threads.emplace_back(std::move(thr));
        } catch(std::runtime_error& e)
        {
            std::cerr << e.what() << std::endl;  // issue creating thread
            continue;
        } catch(std::bad_alloc& e)
        {
            std::cerr << e.what() << std::endl;
            continue;
        }
    }
    //------------------------------------------------------------------------//
 
    AutoLock _task_lock(*m_task_lock);
 
    // thread pool size doesn't match with join vector
    // this will screw up joining later
    if(m_is_joined.size() != m_main_threads.size())
    {
        std::stringstream ss;
        ss << "ThreadPool::initialize_threadpool - boolean is_joined vector "
           << "is a different size than threads vector: " << m_is_joined.size() << " vs. "
           << m_main_threads.size() << " (tid: " << std::this_thread::get_id() << ")";
 
        throw std::runtime_error(ss.str());
    }
 
    if(m_verbose > 0)
    {
        std::cout << "ThreadPool initialized with " << m_pool_size << " threads."
                  << std::endl;
    }
 
    if(!m_task_queue)
        m_task_queue = new UserTaskQueue(m_main_threads.size());
 
    return m_main_threads.size();
}

Referenced by resize().

◆ insert()

int PTL::ThreadPool::insert	(	const task_pointer &	task,
		int	bin = `-1`
	)

inlineprotected

Definition at line 317 of file ThreadPool.hh.

{
    static thread_local ThreadData* _data = ThreadData::GetInstance();
 
    // pass the task to the queue
    auto ibin = m_task_queue->InsertTask(task, _data, bin);
    notify();
    return ibin;
}

Referenced by add_task().

◆ is_alive()

bool PTL::ThreadPool::is_alive ( )

inline

Definition at line 156 of file ThreadPool.hh.

156{ return m_alive_flag->load(); }

Referenced by PTL::VTaskGroup::wait().

◆ is_initialized()

bool ThreadPool::is_initialized ( ) const

Definition at line 193 of file ThreadPool.cc.

{
    return !(m_pool_state->load() == thread_pool::state::NONINIT);
}

◆ is_master()

bool PTL::ThreadPool::is_master ( ) const

inline

Definition at line 171 of file ThreadPool.hh.

171{ return ThisThread::get_id() == m_master_tid; }

◆ notify() [1/2]

void PTL::ThreadPool::notify ( )

inline

Definition at line 240 of file ThreadPool.hh.

{
    // wake up one thread that is waiting for a task to be available
    if(m_thread_awake && m_thread_awake->load() < m_pool_size)
    {
        AutoLock l(*m_task_lock);
        m_task_cond->notify_one();
    }
}

Referenced by add_tasks(), and insert().

◆ notify() [2/2]

void PTL::ThreadPool::notify ( size_type ntasks )

inline

Definition at line 259 of file ThreadPool.hh.

{
    if(ntasks == 0)
        return;
 
    // wake up as many threads that tasks just added
    if(m_thread_awake && m_thread_awake->load() < m_pool_size)
    {
        AutoLock l(*m_task_lock);
        if(ntasks < this->size())
        {
            for(size_type i = 0; i < ntasks; ++i)
                m_task_cond->notify_one();
        }
        else
            m_task_cond->notify_all();
    }
}

◆ notify_all()

void PTL::ThreadPool::notify_all ( )

inline

Definition at line 251 of file ThreadPool.hh.

{
    // wake all threads
    AutoLock l(*m_task_lock);
    m_task_cond->notify_all();
}

◆ operator=() [1/2]

ThreadPool & PTL::ThreadPool::operator= ( const ThreadPool & )

delete

◆ operator=() [2/2]

ThreadPool & PTL::ThreadPool::operator= ( ThreadPool && )

default

◆ record_entry()

void PTL::ThreadPool::record_entry ( )

inlineprotected

Definition at line 187 of file ThreadPool.hh.

    {
        if(m_thread_active)
            ++(*m_thread_active);
    }

◆ record_exit()

void PTL::ThreadPool::record_exit ( )

inlineprotected

Definition at line 193 of file ThreadPool.hh.

    {
        if(m_thread_active)
            --(*m_thread_active);
    }

◆ reset_initialization()

void PTL::ThreadPool::reset_initialization ( )

inline

Definition at line 141 of file ThreadPool.hh.

    {
        auto f      = []() {};
        m_init_func = f;
    }

◆ resize()

void PTL::ThreadPool::resize ( size_type _n )

inline

Definition at line 287 of file ThreadPool.hh.

{
    if(_n == m_pool_size)
        return;
    initialize_threadpool(_n);
    m_task_queue->resize(static_cast<intmax_t>(_n));
}

Referenced by PTL::TaskRunManager::Initialize(), and G4TaskRunManager::SetNumberOfThreads().

◆ run_on_this()

int PTL::ThreadPool::run_on_this ( task_pointer task )

inlineprotected

Definition at line 296 of file ThreadPool.hh.

{
    auto _func = [=]() {
        (*task)();
        if(!task->group())
            delete task;
    };
 
    if(m_tbb_tp && m_tbb_task_group)
    {
        m_tbb_task_group->run(_func);
    }
    else
    {
        _func();
    }
    // return the number of tasks added to task-list
    return 0;
}

Referenced by add_task().

◆ set_affinity() [1/2]

void PTL::ThreadPool::set_affinity ( affinity_func_t f )

inline

Definition at line 166 of file ThreadPool.hh.

166{ m_affinity_func = f; }

Referenced by initialize_threadpool().

◆ set_affinity() [2/2]

void ThreadPool::set_affinity	(	intmax_t	i,
		Thread &	_thread
	)

Definition at line 201 of file ThreadPool.cc.

{
    try
    {
        NativeThread native_thread = _thread.native_handle();
        intmax_t     _pin          = m_affinity_func(i);
        if(m_verbose > 0)
        {
            std::cout << "Setting pin affinity for thread " << _thread.get_id() << " to "
                      << _pin << std::endl;
        }
        Threading::SetPinAffinity(_pin, native_thread);
    } catch(std::runtime_error& e)
    {
        std::cout << "Error setting pin affinity" << std::endl;
        std::cerr << e.what() << std::endl;  // issue assigning affinity
    }
}

◆ set_initialization()

void PTL::ThreadPool::set_initialization ( initialize_func_t f )

inline

Definition at line 140 of file ThreadPool.hh.

140{ m_init_func = f; }

◆ set_use_tbb()

void ThreadPool::set_use_tbb ( bool val )

static

Definition at line 100 of file ThreadPool.cc.

{
#if defined(PTL_USE_TBB)
    f_use_tbb = enable;
#else
    ConsumeParameters<bool>(enable);
#endif
}

Referenced by G4TaskRunManager::G4TaskRunManager(), and PTL::TaskRunManager::TaskRunManager().

◆ set_verbose()

void PTL::ThreadPool::set_verbose ( int n )

inline

Definition at line 169 of file ThreadPool.hh.

169{ m_verbose = n; }

◆ size()

size_type PTL::ThreadPool::size ( ) const

inline

Definition at line 151 of file ThreadPool.hh.

151{ return m_pool_size; }

Referenced by G4TaskRunManager::ComputeNumberOfTasks(), execute_on_all_threads(), PTL::TaskRunManager::GetNumberActiveThreads(), PTL::TaskRunManager::GetNumberOfThreads(), PTL::TaskRunManager::Initialize(), G4TaskRunManagerKernel::InitializeWorker(), notify(), PTL::TaskManager::size(), and PTL::VTaskGroup::wait().

◆ start_thread()

void ThreadPool::start_thread	(	ThreadPool *	tp,
		intmax_t	_idx = `-1`
	)

staticprotected

Definition at line 72 of file ThreadPool.cc.

{
    {
        AutoLock lock(TypeMutex<ThreadPool>(), std::defer_lock);
        if(!lock.owns_lock())
            lock.lock();
        if(_idx < 0)
            _idx = f_thread_ids.size();
        f_thread_ids[std::this_thread::get_id()] = _idx;
    }
    static thread_local std::unique_ptr<ThreadData> _unique_data(new ThreadData(tp));
    thread_data() = _unique_data.get();
    tp->record_entry();
    tp->execute_thread(thread_data()->current_queue);
    tp->record_exit();
}

Referenced by initialize_threadpool().

◆ state()

const pool_state_type & PTL::ThreadPool::state ( ) const

inline

Definition at line 149 of file ThreadPool.hh.

149{ return m_pool_state; }

Referenced by PTL::VTaskGroup::wait().

◆ stop_thread()

ThreadPool::size_type ThreadPool::stop_thread ( )

Definition at line 479 of file ThreadPool.cc.

{
    if(!m_alive_flag->load() || m_pool_size == 0)
        return 0;
 
    //------------------------------------------------------------------------//
    // notify all threads we are shutting down
    m_task_lock->lock();
    m_is_stopped.push_back(true);
    CONDITIONNOTIFY(m_task_cond.get());
    m_task_lock->unlock();
    //------------------------------------------------------------------------//
 
    // lock up the task queue
    AutoLock _task_lock(*m_task_lock);
 
    while(!m_stop_threads.empty())
    {
        auto tid = m_stop_threads.front();
        // remove from stopped
        m_stop_threads.pop_front();
        // remove from main
        for(auto itr = m_main_threads.begin(); itr != m_main_threads.end(); ++itr)
        {
            if(*itr == tid)
            {
                m_main_threads.erase(itr);
                break;
            }
        }
        // remove from join list
        m_is_joined.pop_back();
    }
 
    m_pool_size = m_main_threads.size();
    return m_main_threads.size();
}

Referenced by initialize_threadpool().

◆ tbb_global_control()

tbb_global_control_t *& PTL::ThreadPool::tbb_global_control ( )

inlinestatic

Definition at line 280 of file ThreadPool.hh.

{
    static thread_local tbb_global_control_t* _instance = nullptr;
    return _instance;
}

Referenced by destroy_threadpool(), execute_on_all_threads(), and initialize_threadpool().

◆ using_affinity()

bool PTL::ThreadPool::using_affinity ( ) const

inline

Definition at line 155 of file ThreadPool.hh.

155{ return m_use_affinity; }

◆ using_tbb()

bool ThreadPool::using_tbb ( )

static

Definition at line 92 of file ThreadPool.cc.

{
  return f_use_tbb;
}

Referenced by PTL::TaskRunManager::Initialize(), and G4TaskRunManager::InitializeThreadPool().

The documentation for this class was generated from the following files:

Public Types

Public Member Functions

Static Public Member Functions

Protected Member Functions

Static Protected Member Functions

Detailed Description

Member Typedef Documentation

◆ affinity_func_t

◆ atomic_bool_type

◆ atomic_int_type

◆ bool_list_t

◆ condition_t

◆ initialize_func_t

◆ lock_t

◆ pool_state_type

◆ size_type

◆ task_count_type

◆ task_pointer

◆ task_queue_t

◆ task_type

◆ thread_id_map_t

◆ thread_index_map_t

◆ thread_list_t

◆ thread_vec_t

◆ uomap

Constructor & Destructor Documentation

◆ ThreadPool() [1/3]

◆ ~ThreadPool()

◆ ThreadPool() [2/3]

◆ ThreadPool() [3/3]

Member Function Documentation

◆ add_task()

◆ add_tasks()

◆ destroy_threadpool()

◆ execute_on_all_threads()

◆ execute_thread()

◆ get_active_threads_count()

◆ get_queue()

◆ get_this_thread_id()

◆ get_thread() [1/2]

◆ get_thread() [2/2]

◆ get_thread_ids()

◆ get_verbose()

◆ initialize_threadpool()

◆ insert()

◆ is_alive()

◆ is_initialized()

◆ is_master()

◆ notify() [1/2]

◆ notify() [2/2]

◆ notify_all()

◆ operator=() [1/2]

◆ operator=() [2/2]

◆ record_entry()

◆ record_exit()

◆ reset_initialization()

◆ resize()

◆ run_on_this()

◆ set_affinity() [1/2]

◆ set_affinity() [2/2]

◆ set_initialization()

◆ set_use_tbb()

◆ set_verbose()

◆ size()

◆ start_thread()

◆ state()

◆ stop_thread()

◆ tbb_global_control()

◆ using_affinity()

◆ using_tbb()