X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=9bd7498979761942ba0b0e01f315e2c444385c26;hp=536f22c1194cf6912eb578f550cf1d298d21d959;hb=c465f4c4df1a8ad3d5c1e3759c6aa27b777b8a77;hpb=6809b57cfc47321826f01253241afef8b4380612 diff --git a/src/thread.cpp b/src/thread.cpp index 536f22c1..9bd74989 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) 2008-2010 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2008-2012 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 @@ -17,473 +17,369 @@ along with this program. If not, see . */ +#include #include +#include "movegen.h" #include "search.h" #include "thread.h" #include "ucioption.h" -ThreadsManager Threads; // Global object definition +using namespace Search; + +ThreadPool Threads; // Global object namespace { extern "C" { // 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 - // last two threads are dedicated to read input from GUI and to mimic a - // timer, so they run in listener_loop() and timer_loop() respectively. - -#if defined(_MSC_VER) - DWORD WINAPI start_routine(LPVOID thread) { -#else - void* start_routine(void* thread) { -#endif - - if (((Thread*)thread)->threadID == 0) - ((Thread*)thread)->main_loop(); - - else if (((Thread*)thread)->threadID == MAX_THREADS) - ((Thread*)thread)->timer_loop(); - else - ((Thread*)thread)->idle_loop(NULL); + // is launched. It is a wrapper to the virtual function idle_loop(). - return 0; - } + long start_routine(Thread* th) { th->idle_loop(); 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. - -void Thread::wake_up() { - - lock_grab(&sleepLock); - cond_signal(&sleepCond); - lock_release(&sleepLock); -} - +// Thread c'tor starts a newly-created thread of execution that will call +// the the virtual function idle_loop(), going immediately to sleep. -// cutoff_occurred() checks whether a beta cutoff has occurred in the current -// active split point, or in some ancestor of the split point. +Thread::Thread() : splitPoints() { -bool Thread::cutoff_occurred() const { + is_searching = do_exit = false; + maxPly = splitPointsCnt = 0; + curSplitPoint = NULL; + idx = Threads.size(); - for (SplitPoint* sp = splitPoint; sp; sp = sp->parent) - if (sp->is_betaCutoff) - return true; - return false; + if (!thread_create(handle, start_routine, this)) + { + std::cerr << "Failed to create thread number " << idx << std::endl; + ::exit(EXIT_FAILURE); + } } -// 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). +// Thread d'tor waits for thread termination before to return -bool Thread::is_available_to(int master) const { +Thread::~Thread() { - if (is_searching) - return false; + do_exit = true; // Search must be already finished + notify_one(); + thread_join(handle); // Wait for thread termination +} - // 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; - // 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; +// TimerThread::idle_loop() is where the timer thread waits msec milliseconds +// and then calls check_time(). If msec is 0 thread sleeps until is woken up. +extern void check_time(); - return false; -} +void TimerThread::idle_loop() { + while (!do_exit) + { + mutex.lock(); -// read_uci_options() updates number of active threads and other internal -// parameters according to the UCI options values. It is called before -// to start a new search. - -void ThreadsManager::read_uci_options() { + if (!do_exit) + sleepCondition.wait_for(mutex, msec ? msec : INT_MAX); - maxThreadsPerSplitPoint = Options["Maximum Number of Threads per Split Point"].value(); - minimumSplitDepth = Options["Minimum Split Depth"].value() * ONE_PLY; - useSleepingThreads = Options["Use Sleeping Threads"].value(); + mutex.unlock(); - set_size(Options["Threads"].value()); + if (msec) + check_time(); + } } -// set_size() changes the number of active threads and raises do_sleep flag for -// all the unused threads that will go immediately to sleep. +// MainThread::idle_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 ThreadsManager::set_size(int cnt) { +void MainThread::idle_loop() { - assert(cnt > 0 && cnt <= MAX_THREADS); + while (true) + { + mutex.lock(); - activeThreads = cnt; + is_finished = true; // Always return to sleep after a search + is_searching = false; - for (int i = 1; i < MAX_THREADS; i++) // Ignore main thread - if (i < activeThreads) + while (is_finished && !do_exit) { - // 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 as, for instance, on mobile - // devices where memory is scarce and allocating for MAX_THREADS could - // even result in a crash. - threads[i].pawnTable.init(); - threads[i].materialTable.init(); - - threads[i].do_sleep = false; + Threads.sleepCondition.notify_one(); // Wake up UI thread if needed + sleepCondition.wait(mutex); } - else - threads[i].do_sleep = true; -} - - -// 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 used to block waiting for end of searching - cond_init(&sleepCond); + mutex.unlock(); - // Initialize threads lock, used when allocating slaves during splitting - lock_init(&threadsLock); - - // Initialize sleep and split point locks - for (int i = 0; i <= MAX_THREADS; i++) - { - lock_init(&threads[i].sleepLock); - cond_init(&threads[i].sleepCond); + if (do_exit) + return; - for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_init(&(threads[i].splitPoints[j].lock)); - } + is_searching = true; - // Initialize main thread's associated data - threads[0].pawnTable.init(); - threads[0].materialTable.init(); + Search::think(); - // Create and launch all the threads, threads will go immediately to sleep - for (int i = 0; i <= MAX_THREADS; i++) - { - threads[i].is_searching = false; - threads[i].do_sleep = true; - threads[i].threadID = i; - -#if defined(_MSC_VER) - threads[i].handle = CreateThread(NULL, 0, start_routine, (LPVOID)&threads[i], 0, NULL); - bool ok = (threads[i].handle != NULL); -#else - bool ok = (pthread_create(&threads[i].handle, NULL, start_routine, (void*)&threads[i]) == 0); -#endif - - if (!ok) - { - std::cerr << "Failed to create thread number " << i << std::endl; - ::exit(EXIT_FAILURE); - } + assert(is_searching); } } -// exit() is called to cleanly terminate the threads when the program finishes - -void ThreadsManager::exit() { +// Thread::notify_one() wakes up the thread, normally at split time - for (int i = 0; i <= MAX_THREADS; i++) - { - threads[i].do_terminate = true; - threads[i].wake_up(); - - // Wait for slave termination -#if defined(_MSC_VER) - WaitForSingleObject(threads[i].handle, 0); - CloseHandle(threads[i].handle); -#else - pthread_join(threads[i].handle, NULL); -#endif - - // Now we can safely destroy 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)); - } +void Thread::notify_one() { - lock_destroy(&threadsLock); - cond_destroy(&sleepCond); + mutex.lock(); + sleepCondition.notify_one(); + mutex.unlock(); } -// 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 { +// Thread::wait_for() set the thread to sleep until condition 'b' turns true - assert(master >= 0 && master < activeThreads); - - for (int i = 0; i < activeThreads; i++) - if (i != master && threads[i].is_available_to(master)) - return true; +void Thread::wait_for(volatile const bool& b) { - return false; + mutex.lock(); + while (!b) sleepCondition.wait(mutex); + mutex.unlock(); } -// split_point_finished() checks if all the slave threads of a given split -// point have finished searching. +// Thread::cutoff_occurred() checks whether a beta cutoff has occurred in the +// current active split point, or in some ancestor of the split point. -bool ThreadsManager::split_point_finished(SplitPoint* sp) const { +bool Thread::cutoff_occurred() const { - for (int i = 0; i < activeThreads; i++) - if (sp->is_slave[i]) - return false; + for (SplitPoint* sp = curSplitPoint; sp; sp = sp->parent) + if (sp->cutoff) + return true; - 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 we tell our helper threads 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, SearchStack* 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; +// 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). - // Pick the next available split point object from the split point stack - SplitPoint* sp = masterThread.splitPoints + masterThread.activeSplitPoints; - - // Initialize the split point object - 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 +bool Thread::is_available_to(Thread* master) const { - // Try to allocate available threads and ask them to start searching setting - // the state to Thread::WORKISWAITING, this must be done under lock protection - // to avoid concurrent allocation of the same slave by another master. - lock_grab(&threadsLock); + if (is_searching) + return false; - for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++) - if (i != master && threads[i].is_available_to(master)) - { - workersCnt++; - sp->is_slave[i] = true; - threads[i].splitPoint = sp; + // 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; - // This makes the slave to exit from idle_loop() - threads[i].is_searching = true; + // 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)); +} - if (useSleepingThreads) - threads[i].wake_up(); - } - lock_release(&threadsLock); +// 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. - // We failed to allocate even one slave, return - if (!Fake && workersCnt == 1) - return bestValue; +void ThreadPool::init() { - masterThread.splitPoint = sp; - masterThread.activeSplitPoints++; + sleepWhileIdle = true; + timer = new TimerThread(); + threads.push_back(new MainThread()); + read_uci_options(); +} - // 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, and - // because here is all finished is not possible master is booked. - assert(!masterThread.is_searching); +// exit() cleanly terminates the threads before the program exits. - // 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); +void ThreadPool::exit() { - masterThread.is_searching = true; - masterThread.activeSplitPoints--; + delete timer; // As first becuase check_time() accesses threads data - lock_release(&threadsLock); + for (size_t i = 0; i < threads.size(); i++) + delete threads[i]; +} - masterThread.splitPoint = sp->parent; - pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); - return sp->bestValue; -} +// 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. -// Explicit template instantiations -template Value ThreadsManager::split(Position&, SearchStack*, Value, Value, Value, Depth, Move, int, MovePicker*, int); -template Value ThreadsManager::split(Position&, SearchStack*, Value, Value, Value, Depth, Move, int, MovePicker*, int); +void ThreadPool::read_uci_options() { + maxThreadsPerSplitPoint = Options["Max Threads per Split Point"]; + minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; + size_t requested = Options["Threads"]; -// Thread::timer_loop() is where the timer thread waits maxPly milliseconds -// and then calls do_timer_event(). + assert(requested > 0); -void Thread::timer_loop() { + while (threads.size() < requested) + threads.push_back(new Thread()); - while (!do_terminate) + while (threads.size() > requested) { - lock_grab(&sleepLock); - timed_wait(&sleepCond, &sleepLock, maxPly ? maxPly : INT_MAX); - lock_release(&sleepLock); - do_timer_event(); + delete threads.back(); + threads.pop_back(); } } -// ThreadsManager::set_timer() is used to set the timer to trigger after msec -// milliseconds. If msec is 0 then timer is stopped. +// available_slave_exists() tries to find an idle thread which is available as +// a slave for the thread 'master'. -void ThreadsManager::set_timer(int msec) { +bool ThreadPool::available_slave_exists(Thread* master) const { - Thread& timer = threads[MAX_THREADS]; + for (size_t i = 0; i < threads.size(); i++) + if (threads[i]->is_available_to(master)) + return true; - lock_grab(&timer.sleepLock); - timer.maxPly = msec; - cond_signal(&timer.sleepCond); // Wake up and restart the timer - lock_release(&timer.sleepLock); + return false; } -// 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. +// 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. -void Thread::main_loop() { +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) { - while (true) - { - lock_grab(&sleepLock); + assert(pos.pos_is_ok()); + assert(bestValue > -VALUE_INFINITE); + assert(bestValue <= alpha); + assert(alpha < beta); + assert(beta <= VALUE_INFINITE); + assert(depth > DEPTH_ZERO); - do_sleep = true; // Always return to sleep after a search + Thread* master = pos.this_thread(); - is_searching = false; + if (master->splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD) + return bestValue; - while (do_sleep && !do_terminate) - { - cond_signal(&Threads.sleepCond); // Wake up UI thread if needed - cond_wait(&sleepCond, &sleepLock); - } + // 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 = ∓ + sp.moveCount = moveCount; + sp.pos = &pos; + sp.nodes = 0; + sp.ss = ss; + + assert(master->is_searching); + + master->curSplitPoint = &sp; + int slavesCnt = 0; - is_searching = true; + // 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. + mutex.lock(); + sp.mutex.lock(); - lock_release(&sleepLock); + for (size_t i = 0; i < threads.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() + threads[i]->notify_one(); // Could be sleeping - if (do_terminate) - return; + if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included + break; + } - Search::think(); - } -} + master->splitPointsCnt++; + sp.mutex.unlock(); + mutex.unlock(); -// ThreadsManager::wait_end_of_search() blocks UI thread until main thread has -// returned to sleep in main_loop(). It is needed becuase xboard sends immediately -// new position to search after a "stop" due to ponder miss. + // 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. + // The thread will return from the idle loop when all slaves have finished + // their work at this split point. + if (slavesCnt || Fake) + { + master->Thread::idle_loop(); // Force a call to base class idle_loop() -void ThreadsManager::wait_end_of_search() { + // 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); + } - Thread& main = threads[0]; + // 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(). + mutex.lock(); + sp.mutex.lock(); - lock_grab(&main.sleepLock); + master->is_searching = true; + master->splitPointsCnt--; + master->curSplitPoint = sp.parent; + pos.set_nodes_searched(pos.nodes_searched() + sp.nodes); + *bestMove = sp.bestMove; - while (!main.do_sleep) - cond_wait(&sleepCond, &main.sleepLock); + sp.mutex.unlock(); + mutex.unlock(); - lock_release(&main.sleepLock); + 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); -// 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(bool asyncMode) { - - Thread& main = threads[0]; - - // Wait main thread has finished before to launch a new search - wait_end_of_search(); - - lock_grab(&main.sleepLock); - - // Reset signals before to start the search - memset((void*)&Search::Signals, 0, sizeof(Search::Signals)); - main.do_sleep = false; - cond_signal(&main.sleepCond); // Wake up main thread +// wait_for_search_finished() waits for main thread to go to sleep, this means +// search is finished. Then returns. - if (!asyncMode) - cond_wait(&sleepCond, &main.sleepLock); +void ThreadPool::wait_for_search_finished() { - lock_release(&main.sleepLock); + MainThread* t = main_thread(); + t->mutex.lock(); + while (!t->is_finished) sleepCondition.wait(t->mutex); + t->mutex.unlock(); } -// 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 return, after which the bestmove and pondermove will be printed. +// start_searching() wakes up the main thread sleeping in main_loop() so to start +// a new search, then returns immediately. -void ThreadsManager::wait_for_stop_or_ponderhit() { +void ThreadPool::start_searching(const Position& pos, const LimitsType& limits, + const std::vector& searchMoves, StateStackPtr& states) { + wait_for_search_finished(); - Search::Signals.stopOnPonderhit = true; + SearchTime = Time::now(); // As early as possible - Thread& main = threads[0]; + Signals.stopOnPonderhit = Signals.firstRootMove = false; + Signals.stop = Signals.failedLowAtRoot = false; - lock_grab(&main.sleepLock); + RootPos = pos; + Limits = limits; + SetupStates = states; // Ownership transfer here + RootMoves.clear(); - while (!Search::Signals.stop) - cond_wait(&main.sleepCond, &main.sleepLock); + for (MoveList ml(pos); !ml.end(); ++ml) + if (searchMoves.empty() || count(searchMoves.begin(), searchMoves.end(), ml.move())) + RootMoves.push_back(RootMove(ml.move())); - lock_release(&main.sleepLock); + main_thread()->is_finished = false; + main_thread()->notify_one(); // Starts main thread }