X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=0a8bacf7d09acebf34d40145697338b8fb970a10;hp=832648015cfd362cf9488cca93c663b005fc1af4;hb=0b36ba74fc0a80388cac43a35962ffc73c01b071;hpb=d282cf6964d493706b6face7109e1859eab1742f diff --git a/src/thread.cpp b/src/thread.cpp index 83264801..ba5d13d5 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 @@ -17,486 +17,361 @@ along with this program. If not, see . */ -#include +#include // For std::count +#include +#include "movegen.h" #include "search.h" #include "thread.h" -#include "ucioption.h" +#include "uci.h" using namespace Search; -ThreadsManager Threads; // Global object +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 simply calls idle_loop() of the supplied thread. The first - // and last thread are special. First one is the main search thread while the - // last one mimics a timer, they run in main_loop() and timer_loop(). +namespace { -#if defined(_MSC_VER) - DWORD WINAPI start_routine(LPVOID thread) { -#else - void* start_routine(void* thread) { -#endif + // 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. - Thread* th = (Thread*)thread; + template T* new_thread() { + T* th = new T(); + th->nativeThread = std::thread(&ThreadBase::idle_loop, th); // Will go to sleep + return th; + } - if (th->threadID == 0) - th->main_loop(); + void delete_thread(ThreadBase* th) { - else if (th->threadID == MAX_THREADS) - th->timer_loop(); + th->mutex.lock(); + th->exit = true; // Search must be already finished + th->mutex.unlock(); - else - th->idle_loop(NULL); + th->notify_one(); + th->nativeThread.join(); // Wait for thread termination + delete th; + } - return 0; - } - -} } +} -// wake_up() wakes up the thread, normally at the beginning of the search or, -// if "sleeping threads" is used, when there is some work to do. +// ThreadBase::notify_one() wakes up the thread when there is some work to do -void Thread::wake_up() { +void ThreadBase::notify_one() { - lock_grab(&sleepLock); - cond_signal(&sleepCond); - lock_release(&sleepLock); + std::unique_lock(this->mutex); + sleepCondition.notify_one(); } -// cutoff_occurred() checks whether a beta cutoff has occurred in the current -// active split point, or in some ancestor of the split point. +// ThreadBase::wait_for() set the thread to sleep until 'condition' turns true -bool Thread::cutoff_occurred() const { - - for (SplitPoint* sp = splitPoint; sp; sp = sp->parent) - if (sp->is_betaCutoff) - return true; +void ThreadBase::wait_for(volatile const bool& condition) { - return false; + std::unique_lock lk(mutex); + sleepCondition.wait(lk, [&]{ return condition; }); } -// is_available_to() checks whether the thread is available to help the thread with -// threadID "master" at a split point. An obvious requirement is that thread must be -// idle. With more than two threads, this is not by itself sufficient: If the thread -// is the master of some active split point, it is only available as a slave to the -// threads which are busy searching the split point at the top of "slave"'s split -// point stack (the "helpful master concept" in YBWC terminology). - -bool Thread::is_available_to(int master) const { - - if (is_searching) - return false; - - // 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 localActiveSplitPoints = activeSplitPoints; +// Thread c'tor makes some init but does not launch any execution thread that +// will be started only when c'tor returns. - // 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. - if ( !localActiveSplitPoints - || splitPoints[localActiveSplitPoints - 1].is_slave[master]) - return true; +Thread::Thread() /* : splitPoints() */ { // Initialization of non POD broken in MSVC - return false; + searching = false; + maxPly = 0; + splitPointsSize = 0; + activeSplitPoint = nullptr; + activePosition = nullptr; + idx = Threads.size(); // Starts from 0 } -// read_uci_options() updates number of active threads and other parameters -// according to the UCI options values. It is called before to start a new 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. -void ThreadsManager::read_uci_options() { +bool Thread::cutoff_occurred() const { - maxThreadsPerSplitPoint = Options["Max Threads per Split Point"]; - minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; - useSleepingThreads = Options["Use Sleeping Threads"]; + for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint) + if (sp->cutoff) + return true; - set_size(Options["Threads"]); + return false; } -// set_size() changes the number of active threads and raises do_sleep flag for -// all the unused threads that will go immediately to sleep. - -void ThreadsManager::set_size(int cnt) { +// Thread::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 split point, it is only available as a slave to the slaves +// which are busy searching the split point at the top of slave's split point +// stack (the "helpful master concept" in YBWC terminology). - assert(cnt > 0 && cnt <= MAX_THREADS); +bool Thread::available_to(const Thread* master) const { - activeThreads = cnt; + if (searching) + return false; - for (int i = 1; i < MAX_THREADS; i++) // Ignore main thread - if (i < activeThreads) - { - // Dynamically allocate pawn and material hash tables according to the - // number of active threads. This avoids preallocating memory for all - // possible threads if only few are used. - threads[i].pawnTable.init(); - threads[i].materialTable.init(); + // Make a local copy to be sure it doesn't become zero under our feet while + // testing next condition and so leading to an out of bounds access. + const size_t size = splitPointsSize; - threads[i].do_sleep = false; - } - else - threads[i].do_sleep = true; + // No split points means that the thread is available as a slave for any + // other thread otherwise apply the "helpful master" concept if possible. + return !size || splitPoints[size - 1].slavesMask.test(master->idx); } -// init() is called during startup. Initializes locks and condition variables -// and launches all threads sending them immediately to sleep. - -void ThreadsManager::init() { - - // Initialize sleep condition and lock used by thread manager - cond_init(&sleepCond); - lock_init(&threadsLock); +// Thread::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), 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 +// informed 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. + +void Thread::split(Position& pos, Stack* ss, Value alpha, Value beta, Value* bestValue, + Move* bestMove, Depth depth, int moveCount, + MovePicker* movePicker, int nodeType, bool cutNode) { + + assert(searching); + assert(-VALUE_INFINITE < *bestValue && *bestValue <= alpha && alpha < beta && beta <= VALUE_INFINITE); + assert(depth >= Threads.minimumSplitDepth); + assert(splitPointsSize < MAX_SPLITPOINTS_PER_THREAD); + + // Pick and init the next available split point + SplitPoint& sp = splitPoints[splitPointsSize]; + + sp.master = this; + sp.parentSplitPoint = activeSplitPoint; + sp.slavesMask = 0, sp.slavesMask.set(idx); + sp.depth = depth; + sp.bestValue = *bestValue; + sp.bestMove = *bestMove; + sp.alpha = alpha; + sp.beta = beta; + sp.nodeType = nodeType; + sp.cutNode = cutNode; + sp.movePicker = movePicker; + sp.moveCount = moveCount; + sp.pos = &pos; + sp.nodes = 0; + sp.cutoff = false; + sp.ss = ss; - // Initialize thread's sleep conditions and split point locks - for (int i = 0; i <= MAX_THREADS; i++) - { - lock_init(&threads[i].sleepLock); - cond_init(&threads[i].sleepCond); + // Try to allocate available threads and ask them to start searching setting + // 'searching' flag. This must be done under lock protection to avoid concurrent + // allocation of the same slave by another master. + Threads.spinlock.acquire(); + sp.spinlock.acquire(); - for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_init(&(threads[i].splitPoints[j].lock)); - } + sp.allSlavesSearching = true; // Must be set under lock protection + ++splitPointsSize; + activeSplitPoint = &sp; + activePosition = nullptr; - // Allocate main thread tables to call evaluate() also when not searching - threads[0].pawnTable.init(); - threads[0].materialTable.init(); + Thread* slave; - // Create and launch all the threads, threads will go immediately to sleep - for (int i = 0; i <= MAX_THREADS; i++) + while ( sp.slavesMask.count() < MAX_SLAVES_PER_SPLITPOINT + && (slave = Threads.available_slave(this)) != nullptr) { - threads[i].is_searching = false; - threads[i].do_sleep = (i != 0); // Avoid a race with start_thinking() - threads[i].threadID = i; - -#if defined(_MSC_VER) - threads[i].handle = CreateThread(NULL, 0, start_routine, &threads[i], 0, NULL); - bool ok = (threads[i].handle != NULL); -#else - bool ok = !pthread_create(&threads[i].handle, NULL, start_routine, &threads[i]); -#endif - - if (!ok) - { - std::cerr << "Failed to create thread number " << i << std::endl; - ::exit(EXIT_FAILURE); - } + sp.slavesMask.set(slave->idx); + slave->activeSplitPoint = &sp; + slave->searching = true; // Slave leaves idle_loop() + slave->notify_one(); // Could be sleeping } -} + // Everything is set up. The master thread enters the idle loop, from which + // it will instantly launch a search, because its 'searching' flag is set. + // The thread will return from the idle loop when all slaves have finished + // their work at this split point. + sp.spinlock.release(); + Threads.spinlock.release(); -// exit() is called to cleanly terminate the threads when the program finishes - -void ThreadsManager::exit() { + Thread::idle_loop(); // Force a call to base class idle_loop() - for (int i = 0; i <= MAX_THREADS; i++) - { - threads[i].do_terminate = true; // Search must be already finished - threads[i].wake_up(); - - // Wait for thread termination -#if defined(_MSC_VER) - WaitForSingleObject(threads[i].handle, INFINITE); - CloseHandle(threads[i].handle); -#else - pthread_join(threads[i].handle, NULL); -#endif - - // Now we can safely destroy associated locks and wait conditions - lock_destroy(&threads[i].sleepLock); - cond_destroy(&threads[i].sleepCond); - - for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_destroy(&(threads[i].splitPoints[j].lock)); - } + // In the helpful master concept, a master can help only a sub-tree of its + // split point and because everything is finished here, it's not possible + // for the master to be booked. + assert(!searching); + assert(!activePosition); - lock_destroy(&threadsLock); - cond_destroy(&sleepCond); + // We have returned from the idle loop, which means that all threads are + // finished. Note that setting 'searching' and decreasing splitPointsSize must + // be done under lock protection to avoid a race with Thread::available_to(). + Threads.spinlock.acquire(); + sp.spinlock.acquire(); + + searching = true; + --splitPointsSize; + activeSplitPoint = sp.parentSplitPoint; + activePosition = &pos; + pos.set_nodes_searched(pos.nodes_searched() + sp.nodes); + *bestMove = sp.bestMove; + *bestValue = sp.bestValue; + + sp.spinlock.release(); + Threads.spinlock.release(); } -// available_slave_exists() tries to find an idle thread which is available as -// a slave for the thread with threadID 'master'. - -bool ThreadsManager::available_slave_exists(int master) const { - - assert(master >= 0 && master < activeThreads); - - for (int i = 0; i < activeThreads; i++) - if (threads[i].is_available_to(master)) - return true; - - return false; -} +// 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 TimerThread::idle_loop() { -// split_point_finished() checks if all the slave threads of a given split -// point have finished searching. + while (!exit) + { + std::unique_lock lk(mutex); -bool ThreadsManager::split_point_finished(SplitPoint* sp) const { + if (!exit) + sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX)); - for (int i = 0; i < activeThreads; i++) - if (sp->is_slave[i]) - return false; + lk.unlock(); - return true; + if (run) + check_time(); + } } -// 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 ThreadsManager::split(Position& pos, Stack* ss, Value alpha, Value beta, - Value bestValue, 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); - assert(pos.thread() >= 0 && pos.thread() < activeThreads); - assert(activeThreads > 1); - - int i, master = pos.thread(); - Thread& masterThread = threads[master]; - - // If we already have too many active split points, don't split - if (masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS) - return bestValue; - - // Pick the next available split point from the split point stack - SplitPoint* sp = &masterThread.splitPoints[masterThread.activeSplitPoints]; - - // Initialize the split point - sp->parent = masterThread.splitPoint; - sp->master = master; - sp->is_betaCutoff = false; - sp->depth = depth; - 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; - - for (i = 0; i < activeThreads; i++) - sp->is_slave[i] = false; - - // If we are here it means we are not available - assert(masterThread.is_searching); - - int workersCnt = 1; // At least the master is included +// 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. - // 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(&threadsLock); +void MainThread::idle_loop() { - for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++) - if (threads[i].is_available_to(master)) - { - workersCnt++; - sp->is_slave[i] = true; - threads[i].splitPoint = sp; + while (!exit) + { + std::unique_lock lk(mutex); - // This makes the slave to exit from idle_loop() - threads[i].is_searching = true; + thinking = false; - if (useSleepingThreads) - threads[i].wake_up(); + while (!thinking && !exit) + { + Threads.sleepCondition.notify_one(); // Wake up the UI thread if needed + sleepCondition.wait(lk); } - lock_release(&threadsLock); - - // We failed to allocate even one slave, return - if (!Fake && workersCnt == 1) - return bestValue; - - masterThread.splitPoint = sp; - masterThread.activeSplitPoints++; + lk.unlock(); - // 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. - masterThread.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(!masterThread.is_searching); - - // We have returned from the idle loop, which means that all threads are - // finished. Note that changing state and decreasing activeSplitPoints is done - // under lock protection to avoid a race with Thread::is_available_to(). - lock_grab(&threadsLock); - - masterThread.is_searching = true; - masterThread.activeSplitPoints--; + if (!exit) + { + searching = true; - lock_release(&threadsLock); + Search::think(); - masterThread.splitPoint = sp->parent; - pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); + assert(searching); - return sp->bestValue; + searching = false; + } + } } -// Explicit template instantiations -template Value ThreadsManager::split(Position&, Stack*, Value, Value, Value, Depth, Move, int, MovePicker*, int); -template Value ThreadsManager::split(Position&, Stack*, Value, Value, Value, Depth, Move, int, MovePicker*, int); +// 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. -// Thread::timer_loop() is where the timer thread waits maxPly milliseconds and -// then calls do_timer_event(). If maxPly is 0 thread sleeps until is woken up. -extern void do_timer_event(); +void ThreadPool::init() { -void Thread::timer_loop() { - - while (!do_terminate) - { - lock_grab(&sleepLock); - timed_wait(&sleepCond, &sleepLock, maxPly ? maxPly : INT_MAX); - lock_release(&sleepLock); - do_timer_event(); - } + timer = new_thread(); + push_back(new_thread()); + read_uci_options(); } -// ThreadsManager::set_timer() is used to set the timer to trigger after msec -// milliseconds. If msec is 0 then timer is stopped. +// ThreadPool::exit() terminates the threads before the program exits. Cannot be +// done in d'tor because threads must be terminated before freeing us. -void ThreadsManager::set_timer(int msec) { +void ThreadPool::exit() { - Thread& timer = threads[MAX_THREADS]; + delete_thread(timer); // As first because check_time() accesses threads data - lock_grab(&timer.sleepLock); - timer.maxPly = msec; - cond_signal(&timer.sleepCond); // Wake up and restart the timer - lock_release(&timer.sleepLock); + for (Thread* th : *this) + delete_thread(th); } -// 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. - -void Thread::main_loop() { - - while (true) - { - lock_grab(&sleepLock); +// 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. - do_sleep = true; // Always return to sleep after a search - is_searching = false; +void ThreadPool::read_uci_options() { - while (do_sleep && !do_terminate) - { - cond_signal(&Threads.sleepCond); // Wake up UI thread if needed - cond_wait(&sleepCond, &sleepLock); - } + minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; + size_t requested = Options["Threads"]; - is_searching = true; + assert(requested > 0); - lock_release(&sleepLock); + // If zero (default) then set best minimum split depth automatically + if (!minimumSplitDepth) + minimumSplitDepth = requested < 8 ? 4 * ONE_PLY : 7 * ONE_PLY; - if (do_terminate) - return; + while (size() < requested) + push_back(new_thread()); - think(); // This is the search entry point + while (size() > requested) + { + delete_thread(back()); + pop_back(); } } -// ThreadsManager::start_thinking() is used by UI thread to wake up the main -// thread parked in main_loop() and starting a new search. If asyncMode is true -// then function returns immediately, otherwise caller is blocked waiting for -// the search to finish. - -void ThreadsManager::start_thinking(const Position& pos, const LimitsType& limits, - const std::vector& searchMoves, bool asyncMode) { - Thread& main = threads[0]; - - lock_grab(&main.sleepLock); +// ThreadPool::available_slave() tries to find an idle thread which is available +// as a slave for the thread 'master'. - // Wait main thread has finished before to launch a new search - while (!main.do_sleep) - cond_wait(&sleepCond, &main.sleepLock); - - // Copy input arguments to initialize the search - RootPosition.copy(pos, 0); - Limits = limits; - SearchMoves = searchMoves; +Thread* ThreadPool::available_slave(const Thread* master) const { - // Reset signals before to start the new search - memset((void*)&Signals, 0, sizeof(Signals)); + for (Thread* th : *this) + if (th->available_to(master)) + return th; - main.do_sleep = false; - cond_signal(&main.sleepCond); // Wake up main thread and start searching - - if (!asyncMode) - while (!main.do_sleep) - cond_wait(&sleepCond, &main.sleepLock); - - lock_release(&main.sleepLock); + return nullptr; } -// ThreadsManager::stop_thinking() is used by UI thread to raise a stop request -// and to wait for the main thread finishing the search. Needed to wait exiting -// and terminate the threads after a 'quit' command. - -void ThreadsManager::stop_thinking() { - - Thread& main = threads[0]; - - Search::Signals.stop = true; +// ThreadPool::wait_for_think_finished() waits for main thread to finish the search - lock_grab(&main.sleepLock); +void ThreadPool::wait_for_think_finished() { - cond_signal(&main.sleepCond); // In case is waiting for stop or ponderhit - - while (!main.do_sleep) - cond_wait(&sleepCond, &main.sleepLock); - - lock_release(&main.sleepLock); + std::unique_lock lk(main()->mutex); + sleepCondition.wait(lk, [&]{ return !main()->thinking; }); } -// ThreadsManager::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. +// ThreadPool::start_thinking() wakes up the main thread sleeping in +// MainThread::idle_loop() and starts a new search, then returns immediately. -void ThreadsManager::wait_for_stop_or_ponderhit() { +void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, + StateStackPtr& states) { + wait_for_think_finished(); - Signals.stopOnPonderhit = true; + SearchTime = now(); // As early as possible - Thread& main = threads[0]; + Signals.stopOnPonderhit = Signals.firstRootMove = false; + Signals.stop = Signals.failedLowAtRoot = false; - lock_grab(&main.sleepLock); + RootMoves.clear(); + RootPos = pos; + Limits = limits; + if (states.get()) // If we don't set a new position, preserve current state + { + SetupStates = std::move(states); // Ownership transfer here + assert(!states.get()); + } - while (!Signals.stop) - cond_wait(&main.sleepCond, &main.sleepLock); + for (const auto& m : MoveList(pos)) + if ( limits.searchmoves.empty() + || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m)) + RootMoves.push_back(RootMove(m)); - lock_release(&main.sleepLock); + main()->thinking = true; + main()->notify_one(); // Starts main thread }