X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=src%2Fthread.cpp;h=88b4592143950c8ca2077b080d1bf1248c044396;hb=55758344d3ccf49353bcd8f3a06a4553ff1b753a;hp=0a8bacf7d09acebf34d40145697338b8fb970a10;hpb=4b19430103ac75b574a6b269db447d359814b603;p=stockfish diff --git a/src/thread.cpp b/src/thread.cpp index 0a8bacf7..88b45921 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -1,7 +1,7 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2012 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -17,434 +17,243 @@ along with this program. If not, see . */ +#include // For std::count #include -#include #include "movegen.h" #include "search.h" #include "thread.h" -#include "ucioption.h" +#include "uci.h" using namespace Search; ThreadPool Threads; // Global object -namespace { extern "C" { +extern void check_time(); - // start_routine() is the C function which is called when a new thread - // is launched. It is a wrapper to member function pointed by start_fn. +namespace { - long start_routine(Thread* th) { (th->*(th->start_fn))(); return 0; } + // Helpers to launch a thread after creation and joining before delete. Must be + // outside Thread c'tor and d'tor because the object must be fully initialized + // when start_routine (and hence virtual idle_loop) is called and when joining. -} } + template T* new_thread() { + std::thread* th = new T; + *th = std::thread(&T::idle_loop, (T*)th); // Will go to sleep + return (T*)th; + } + void delete_thread(ThreadBase* th) { -// Thread c'tor starts a newly-created thread of execution that will call -// the idle loop function pointed by start_fn going immediately to sleep. + th->mutex.lock(); + th->exit = true; // Search must be already finished + th->mutex.unlock(); -Thread::Thread(Fn fn) { + th->notify_one(); + th->join(); // Wait for thread termination + delete th; + } - is_searching = do_exit = false; - maxPly = splitPointsCnt = 0; - curSplitPoint = NULL; - start_fn = fn; - idx = Threads.size(); +} - do_sleep = (fn != &Thread::main_loop); // Avoid a race with start_searching() - lock_init(sleepLock); - cond_init(sleepCond); +// ThreadBase::notify_one() wakes up the thread when there is some work to do - for (size_t j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++) - lock_init(splitPoints[j].lock); +void ThreadBase::notify_one() { - if (!thread_create(handle, start_routine, this)) - { - std::cerr << "Failed to create thread number " << idx << std::endl; - ::exit(EXIT_FAILURE); - } + std::unique_lock lk(mutex); + sleepCondition.notify_one(); } -// Thread d'tor waits for thread termination before to return. +// ThreadBase::wait() set the thread to sleep until 'condition' turns true -Thread::~Thread() { +void ThreadBase::wait(volatile const bool& condition) { - assert(do_sleep); + std::unique_lock lk(mutex); + sleepCondition.wait(lk, [&]{ return condition; }); +} - do_exit = true; // Search must be already finished - wake_up(); - thread_join(handle); // Wait for thread termination +// ThreadBase::wait_while() set the thread to sleep until 'condition' turns false - lock_destroy(sleepLock); - cond_destroy(sleepCond); +void ThreadBase::wait_while(volatile const bool& condition) { - for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++) - lock_destroy(splitPoints[j].lock); + std::unique_lock lk(mutex); + sleepCondition.wait(lk, [&]{ return !condition; }); } -// Thread::timer_loop() is where the timer thread waits maxPly milliseconds and -// then calls check_time(). If maxPly is 0 thread sleeps until is woken up. -extern void check_time(); +// Thread c'tor makes some init but does not launch any execution thread that +// will be started only when c'tor returns. -void Thread::timer_loop() { +Thread::Thread() /* : splitPoints() */ { // Initialization of non POD broken in MSVC - while (!do_exit) - { - lock_grab(sleepLock); - timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX); - lock_release(sleepLock); - check_time(); - } + searching = false; + maxPly = 0; + idx = Threads.size(); // Starts from 0 } -// Thread::main_loop() is where the main thread is parked waiting to be started -// when there is a new search. Main thread will launch all the slave threads. +// TimerThread::idle_loop() is where the timer thread waits Resolution milliseconds +// and then calls check_time(). When not searching, thread sleeps until it's woken up. -void Thread::main_loop() { +void TimerThread::idle_loop() { - while (true) + while (!exit) { - lock_grab(sleepLock); - - do_sleep = true; // Always return to sleep after a search - is_searching = false; + std::unique_lock lk(mutex); - while (do_sleep && !do_exit) - { - cond_signal(Threads.sleepCond); // Wake up UI thread if needed - cond_wait(sleepCond, sleepLock); - } - - lock_release(sleepLock); + if (!exit) + sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX)); - if (do_exit) - return; + lk.unlock(); - is_searching = true; - - Search::think(); + if (!exit && run) + check_time(); } } -// Thread::wake_up() wakes up the thread, normally at the beginning of the search -// or, if "sleeping threads" is used at split time. +// Thread::idle_loop() is where the thread is parked when it has no work to do -void Thread::wake_up() { +void Thread::idle_loop() { - lock_grab(sleepLock); - cond_signal(sleepCond); - lock_release(sleepLock); -} - - -// Thread::wait_for_stop_or_ponderhit() is called when the maximum depth is -// reached while the program is pondering. The point is to work around a wrinkle -// in the UCI protocol: When pondering, the engine is not allowed to give a -// "bestmove" before the GUI sends it a "stop" or "ponderhit" command. We simply -// wait here until one of these commands (that raise StopRequest) is sent and -// then return, after which the bestmove and pondermove will be printed. + while (!exit) + { + std::unique_lock lk(mutex); -void Thread::wait_for_stop_or_ponderhit() { + while (!searching && !exit) + sleepCondition.wait(lk); - Signals.stopOnPonderhit = true; + lk.unlock(); - lock_grab(sleepLock); - while (!Signals.stop) cond_wait(sleepCond, sleepLock); - lock_release(sleepLock); + if (!exit && searching) + search(); + } } -// Thread::cutoff_occurred() checks whether a beta cutoff has occurred in the -// current active split point, or in some ancestor of the split point. +// MainThread::idle_loop() is where the main thread is parked waiting to be started +// when there is a new search. The main thread will launch all the slave threads. -bool Thread::cutoff_occurred() const { +void MainThread::idle_loop() { - for (SplitPoint* sp = curSplitPoint; sp; sp = sp->parent) - if (sp->cutoff) - return true; + while (!exit) + { + std::unique_lock lk(mutex); - return false; -} + thinking = false; + while (!thinking && !exit) + { + sleepCondition.notify_one(); // Wake up the UI thread if needed + sleepCondition.wait(lk); + } -// Thread::is_available_to() checks whether the thread is available to help the -// thread 'master' at a split point. An obvious requirement is that thread must -// be idle. With more than two threads, this is not sufficient: If the thread is -// the master of some active split point, it is only available as a slave to the -// slaves which are busy searching the split point at the top of slaves split -// point stack (the "helpful master concept" in YBWC terminology). + lk.unlock(); + + if (!exit) + think(); + } +} -bool Thread::is_available_to(Thread* master) const { - if (is_searching) - return false; +// MainThread::join() waits for main thread to finish thinking - // Make a local copy to be sure doesn't become zero under our feet while - // testing next condition and so leading to an out of bound access. - int spCnt = splitPointsCnt; +void MainThread::join() { - // No active split points means that the thread is available as a slave for any - // other thread otherwise apply the "helpful master" concept if possible. - return !spCnt || (splitPoints[spCnt - 1].slavesMask & (1ULL << master->idx)); + std::unique_lock lk(mutex); + sleepCondition.wait(lk, [&]{ return !thinking; }); } -// init() is called at startup. Initializes lock and condition variable and -// launches requested threads sending them immediately to sleep. We cannot use -// a c'tor becuase Threads is a static object and we need a fully initialized -// engine at this point due to allocation of endgames in Thread c'tor. +// ThreadPool::init() is called at startup to create and launch requested threads, +// that will go immediately to sleep. We cannot use a c'tor because Threads is a +// static object and we need a fully initialized engine at this point due to +// allocation of Endgames in Thread c'tor. void ThreadPool::init() { - cond_init(sleepCond); - lock_init(splitLock); - timer = new Thread(&Thread::timer_loop); - threads.push_back(new Thread(&Thread::main_loop)); + timer = new_thread(); + push_back(new_thread()); read_uci_options(); } -// d'tor cleanly terminates the threads when the program exits. +// ThreadPool::exit() terminates the threads before the program exits. Cannot be +// done in d'tor because threads must be terminated before freeing us. -ThreadPool::~ThreadPool() { +void ThreadPool::exit() { - for (size_t i = 0; i < size(); i++) - delete threads[i]; + delete_thread(timer); // As first because check_time() accesses threads data + timer = nullptr; - delete timer; - lock_destroy(splitLock); - cond_destroy(sleepCond); + for (Thread* th : *this) + delete_thread(th); + + clear(); // Get rid of stale pointers } -// read_uci_options() updates internal threads parameters from the corresponding -// UCI options and creates/destroys threads to match the requested number. Thread -// objects are dynamically allocated to avoid creating in advance all possible -// threads, with included pawns and material tables, if only few are used. +// ThreadPool::read_uci_options() updates internal threads parameters from the +// corresponding UCI options and creates/destroys threads to match the requested +// number. Thread objects are dynamically allocated to avoid creating all possible +// threads in advance (which include pawns and material tables), even if only a +// few are to be used. void ThreadPool::read_uci_options() { - maxThreadsPerSplitPoint = Options["Max Threads per Split Point"]; - minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; - useSleepingThreads = Options["Use Sleeping Threads"]; - size_t requested = Options["Threads"]; + size_t requested = Options["Threads"]; assert(requested > 0); while (size() < requested) - threads.push_back(new Thread(&Thread::idle_loop)); + push_back(new_thread()); while (size() > requested) { - delete threads.back(); - threads.pop_back(); - } -} - - -// wake_up() is called before a new search to start the threads that are waiting -// on the sleep condition and to reset maxPly. When useSleepingThreads is set -// threads will be woken up at split time. - -void ThreadPool::wake_up() const { - - for (size_t i = 0; i < size(); i++) - { - threads[i]->maxPly = 0; - threads[i]->do_sleep = false; - - if (!useSleepingThreads) - threads[i]->wake_up(); + delete_thread(back()); + pop_back(); } } -// sleep() is called after the search finishes to ask all the threads but the -// main one to go waiting on a sleep condition. +// ThreadPool::nodes_searched() returns the number of nodes searched -void ThreadPool::sleep() const { +int64_t ThreadPool::nodes_searched() { - for (size_t i = 1; i < size(); i++) // Main thread will go to sleep by itself - threads[i]->do_sleep = true; // to avoid a race with start_searching() + int64_t nodes = 0; + for (Thread *th : *this) + nodes += th->rootPos.nodes_searched(); + return nodes; } -// available_slave_exists() tries to find an idle thread which is available as -// a slave for the thread 'master'. +// ThreadPool::start_thinking() wakes up the main thread sleeping in +// MainThread::idle_loop() and starts a new search, then returns immediately. -bool ThreadPool::available_slave_exists(Thread* master) const { - - for (size_t i = 0; i < size(); i++) - if (threads[i]->is_available_to(master)) - return true; - - return false; -} - - -// split() does the actual work of distributing the work at a node between -// several available threads. If it does not succeed in splitting the node -// (because no idle threads are available, or because we have no unused split -// point objects), the function immediately returns. If splitting is possible, a -// SplitPoint object is initialized with all the data that must be copied to the -// helper threads and then helper threads are told that they have been assigned -// work. This will cause them to instantly leave their idle loops and call -// search(). When all threads have returned from search() then split() returns. - -template -Value ThreadPool::split(Position& pos, Stack* ss, Value alpha, Value beta, - Value bestValue, Move* bestMove, Depth depth, - Move threatMove, int moveCount, MovePicker* mp, int nodeType) { - - assert(pos.pos_is_ok()); - assert(bestValue > -VALUE_INFINITE); - assert(bestValue <= alpha); - assert(alpha < beta); - assert(beta <= VALUE_INFINITE); - assert(depth > DEPTH_ZERO); - - Thread* master = pos.this_thread(); - - if (master->splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD) - return bestValue; - - // Pick the next available split point from the split point stack - SplitPoint* sp = &master->splitPoints[master->splitPointsCnt]; - - sp->parent = master->curSplitPoint; - sp->master = master; - sp->cutoff = false; - sp->slavesMask = 1ULL << master->idx; - sp->depth = depth; - sp->bestMove = *bestMove; - sp->threatMove = threatMove; - sp->alpha = alpha; - sp->beta = beta; - sp->nodeType = nodeType; - sp->bestValue = bestValue; - sp->mp = mp; - sp->moveCount = moveCount; - sp->pos = &pos; - sp->nodes = 0; - sp->ss = ss; - - assert(master->is_searching); - - master->curSplitPoint = sp; - int slavesCnt = 0; - - // Try to allocate available threads and ask them to start searching setting - // is_searching flag. This must be done under lock protection to avoid concurrent - // allocation of the same slave by another master. - lock_grab(sp->lock); - lock_grab(splitLock); - - for (size_t i = 0; i < size() && !Fake; ++i) - if (threads[i]->is_available_to(master)) - { - sp->slavesMask |= 1ULL << i; - threads[i]->curSplitPoint = sp; - threads[i]->is_searching = true; // Slave leaves idle_loop() - - if (useSleepingThreads) - threads[i]->wake_up(); - - if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included - break; - } - - master->splitPointsCnt++; - - lock_release(splitLock); - lock_release(sp->lock); - - // Everything is set up. The master thread enters the idle loop, from which - // it will instantly launch a search, because its is_searching flag is set. - // We pass the split point as a parameter to the idle loop, which means that - // the thread will return from the idle loop when all slaves have finished - // their work at this split point. - if (slavesCnt || Fake) - { - master->idle_loop(sp); - - // In helpful master concept a master can help only a sub-tree of its split - // point, and because here is all finished is not possible master is booked. - assert(!master->is_searching); - } - - // We have returned from the idle loop, which means that all threads are - // finished. Note that setting is_searching and decreasing splitPointsCnt is - // done under lock protection to avoid a race with Thread::is_available_to(). - lock_grab(sp->lock); // To protect sp->nodes - lock_grab(splitLock); - - master->is_searching = true; - master->splitPointsCnt--; - master->curSplitPoint = sp->parent; - pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); - *bestMove = sp->bestMove; - - lock_release(splitLock); - lock_release(sp->lock); - - return sp->bestValue; -} - -// Explicit template instantiations -template Value ThreadPool::split(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker*, int); -template Value ThreadPool::split(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker*, int); - - -// set_timer() is used to set the timer to trigger after msec milliseconds. -// If msec is 0 then timer is stopped. - -void ThreadPool::set_timer(int msec) { - - lock_grab(timer->sleepLock); - timer->maxPly = msec; - cond_signal(timer->sleepCond); // Wake up and restart the timer - lock_release(timer->sleepLock); -} - - -// wait_for_search_finished() waits for main thread to go to sleep, this means -// search is finished. Then returns. - -void ThreadPool::wait_for_search_finished() { - - Thread* t = main_thread(); - lock_grab(t->sleepLock); - cond_signal(t->sleepCond); // In case is waiting for stop or ponderhit - while (!t->do_sleep) cond_wait(sleepCond, t->sleepLock); - lock_release(t->sleepLock); -} - - -// start_searching() wakes up the main thread sleeping in main_loop() so to start -// a new search, then returns immediately. - -void ThreadPool::start_searching(const Position& pos, const LimitsType& limits, - const std::vector& searchMoves) { - wait_for_search_finished(); - - SearchTime.restart(); // As early as possible +void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, + StateStackPtr& states) { + main()->join(); Signals.stopOnPonderhit = Signals.firstRootMove = false; Signals.stop = Signals.failedLowAtRoot = false; - RootPosition = pos; + main()->rootMoves.clear(); + main()->rootPos = pos; Limits = limits; - RootMoves.clear(); + if (states.get()) // If we don't set a new position, preserve current state + { + SetupStates = std::move(states); // Ownership transfer here + assert(!states.get()); + } - for (MoveList ml(pos); !ml.end(); ++ml) - if (searchMoves.empty() || count(searchMoves.begin(), searchMoves.end(), ml.move())) - RootMoves.push_back(RootMove(ml.move())); + for (const auto& m : MoveList(pos)) + if ( limits.searchmoves.empty() + || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m)) + main()->rootMoves.push_back(RootMove(m)); - main_thread()->do_sleep = false; - main_thread()->wake_up(); + main()->thinking = true; + main()->notify_one(); // Wake up main thread: 'thinking' must be already set }