X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=ca87e1ddaec9426ee95c181f62397ac3d7b0f1a2;hp=536f22c1194cf6912eb578f550cf1d298d21d959;hb=f9571e8d57381275f08ffbfb960358319d4c34dd;hpb=6809b57cfc47321826f01253241afef8b4380612 diff --git a/src/thread.cpp b/src/thread.cpp index 536f22c1..ca87e1dd 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) 2008-2014 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 +17,363 @@ 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" -ThreadsManager Threads; // Global object definition +using namespace Search; -namespace { extern "C" { +ThreadPool Threads; // Global object - // 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 +extern void check_time(); - if (((Thread*)thread)->threadID == 0) - ((Thread*)thread)->main_loop(); +namespace { - else if (((Thread*)thread)->threadID == MAX_THREADS) - ((Thread*)thread)->timer_loop(); - else - ((Thread*)thread)->idle_loop(NULL); + // start_routine() is the C function which is called when a new thread + // is launched. It is a wrapper to the virtual function idle_loop(). - return 0; - } + extern "C" { long start_routine(ThreadBase* th) { th->idle_loop(); 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 will be fully initialized + // when start_routine (and hence virtual idle_loop) is called and when joining. -// 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. + template T* new_thread() { + T* th = new T(); + thread_create(th->handle, start_routine, th); // Will go to sleep + return th; + } -void Thread::wake_up() { + void delete_thread(ThreadBase* th) { + th->exit = true; // Search must be already finished + th->notify_one(); + thread_join(th->handle); // Wait for thread termination + delete th; + } - lock_grab(&sleepLock); - cond_signal(&sleepCond); - lock_release(&sleepLock); } -// cutoff_occurred() checks whether a beta cutoff has occurred in the current -// active split point, or in some ancestor of the split point. +// notify_one() wakes up the thread when there is some work to do -bool Thread::cutoff_occurred() const { +void ThreadBase::notify_one() { - for (SplitPoint* sp = splitPoint; sp; sp = sp->parent) - if (sp->is_betaCutoff) - return true; - return false; + mutex.lock(); + sleepCondition.notify_one(); + mutex.unlock(); } -// 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). +// wait_for() set the thread to sleep until 'condition' turns true -bool Thread::is_available_to(int master) const { +void ThreadBase::wait_for(volatile const bool& condition) { - if (is_searching) - return false; + mutex.lock(); + while (!condition) sleepCondition.wait(mutex); + mutex.unlock(); +} - // 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; +// Thread c'tor just inits data and does not launch any execution thread. +// Such a thread will only be started when c'tor returns. - return false; +Thread::Thread() /* : splitPoints() */ { // Value-initialization bug in MSVC + + searching = false; + maxPly = splitPointsSize = 0; + activeSplitPoint = NULL; + activePosition = NULL; + idx = Threads.size(); // Starts from 0 } -// 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. +// 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["Maximum Number of Threads per Split Point"].value(); - minimumSplitDepth = Options["Minimum Split Depth"].value() * ONE_PLY; - useSleepingThreads = Options["Use Sleeping Threads"].value(); + for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint) + if (sp->cutoff) + return true; - set_size(Options["Threads"].value()); + 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. +// 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). -void ThreadsManager::set_size(int cnt) { +bool Thread::available_to(const Thread* master) const { - assert(cnt > 0 && cnt <= MAX_THREADS); + if (searching) + return false; - activeThreads = cnt; + // 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 int size = splitPointsSize; - 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 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; - } - 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() { +// TimerThread::idle_loop() is where the timer thread waits msec milliseconds +// and then calls check_time(). If msec is 0 thread sleeps until it's woken up. - // Initialize sleep condition used to block waiting for end of searching - cond_init(&sleepCond); +void TimerThread::idle_loop() { - // 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++) + while (!exit) { - lock_init(&threads[i].sleepLock); - cond_init(&threads[i].sleepCond); + mutex.lock(); - for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_init(&(threads[i].splitPoints[j].lock)); - } + if (!exit) + sleepCondition.wait_for(mutex, run ? Resolution : INT_MAX); - // Initialize main thread's associated data - threads[0].pawnTable.init(); - threads[0].materialTable.init(); + mutex.unlock(); - // 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); - } + if (run) + check_time(); } } -// exit() is called to cleanly terminate the threads when the program finishes +// 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. -void ThreadsManager::exit() { +void MainThread::idle_loop() { - for (int i = 0; i <= MAX_THREADS; i++) + while (true) { - 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)); - } + mutex.lock(); - lock_destroy(&threadsLock); - cond_destroy(&sleepCond); -} + thinking = false; + while (!thinking && !exit) + { + Threads.sleepCondition.notify_one(); // Wake up the UI thread if needed + sleepCondition.wait(mutex); + } -// 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 (i != master && threads[i].is_available_to(master)) - return true; - - return false; -} + mutex.unlock(); + if (exit) + return; -// split_point_finished() checks if all the slave threads of a given split -// point have finished searching. + searching = true; -bool ThreadsManager::split_point_finished(SplitPoint* sp) const { + Search::think(); - for (int i = 0; i < activeThreads; i++) - if (sp->is_slave[i]) - return false; + assert(searching); - return true; + searching = 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; - - // 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 +// 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. - // 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); +void ThreadPool::init() { - 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; - - // This makes the slave to exit from idle_loop() - threads[i].is_searching = true; - - if (useSleepingThreads) - threads[i].wake_up(); - } + timer = new_thread(); + push_back(new_thread()); + read_uci_options(); +} - lock_release(&threadsLock); - // We failed to allocate even one slave, return - if (!Fake && workersCnt == 1) - return bestValue; +// exit() cleanly terminates the threads before the program exits. Cannot be done in +// d'tor because we have to terminate the threads before to free ThreadPool object. - masterThread.splitPoint = sp; - masterThread.activeSplitPoints++; +void ThreadPool::exit() { - // 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); + delete_thread(timer); // As first because check_time() accesses threads data - // 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); - - // 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--; + for (iterator it = begin(); it != end(); ++it) + delete_thread(*it); +} - lock_release(&threadsLock); - masterThread.splitPoint = sp->parent; - pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); +// 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. - return sp->bestValue; -} +void ThreadPool::read_uci_options() { -// 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); + minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; + size_t requested = Options["Threads"]; + assert(requested > 0); -// Thread::timer_loop() is where the timer thread waits maxPly milliseconds -// and then calls do_timer_event(). + // If zero (default) then set best minimum split depth automatically + if (!minimumSplitDepth) + minimumSplitDepth = requested < 8 ? 4 * ONE_PLY : 7 * ONE_PLY; -void Thread::timer_loop() { + while (size() < requested) + push_back(new_thread()); - while (!do_terminate) + while (size() > requested) { - lock_grab(&sleepLock); - timed_wait(&sleepCond, &sleepLock, maxPly ? maxPly : INT_MAX); - lock_release(&sleepLock); - do_timer_event(); + delete_thread(back()); + 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() tries to find an idle thread which is available as a slave +// for the thread 'master'. -void ThreadsManager::set_timer(int msec) { +Thread* ThreadPool::available_slave(const Thread* master) const { - Thread& timer = threads[MAX_THREADS]; + for (const_iterator it = begin(); it != end(); ++it) + if ((*it)->available_to(master)) + return *it; - lock_grab(&timer.sleepLock); - timer.maxPly = msec; - cond_signal(&timer.sleepCond); // Wake up and restart the timer - lock_release(&timer.sleepLock); + return NULL; } -// 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); - - do_sleep = true; // Always return to sleep after a search - - is_searching = false; - - while (do_sleep && !do_terminate) - { - cond_signal(&Threads.sleepCond); // Wake up UI thread if needed - cond_wait(&sleepCond, &sleepLock); - } +// 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 +// 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::split(Position& pos, Stack* ss, Value alpha, Value beta, Value* bestValue, + Move* bestMove, Depth depth, int moveCount, + MovePicker* movePicker, int nodeType, bool cutNode) { - is_searching = true; + assert(pos.pos_is_ok()); + assert(-VALUE_INFINITE < *bestValue && *bestValue <= alpha && alpha < beta && beta <= VALUE_INFINITE); + assert(depth >= Threads.minimumSplitDepth); + assert(searching); + assert(splitPointsSize < MAX_SPLITPOINTS_PER_THREAD); + + // Pick the next available split point from the split point stack + SplitPoint& sp = splitPoints[splitPointsSize]; + + sp.masterThread = 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; - lock_release(&sleepLock); + // 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.mutex.lock(); + sp.mutex.lock(); - if (do_terminate) - return; + sp.allSlavesSearching = true; // Must be set under lock protection + ++splitPointsSize; + activeSplitPoint = &sp; + activePosition = NULL; - Search::think(); + for (Thread* slave; (slave = Threads.available_slave(this)) != NULL; ) + { + sp.slavesMask.set(slave->idx); + slave->activeSplitPoint = &sp; + slave->searching = true; // Slave leaves idle_loop() + slave->notify_one(); // Could be sleeping } -} - -// 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. - -void ThreadsManager::wait_end_of_search() { - - Thread& main = threads[0]; + // 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.mutex.unlock(); + Threads.mutex.unlock(); - lock_grab(&main.sleepLock); + Thread::idle_loop(); // Force a call to base class idle_loop() - while (!main.do_sleep) - cond_wait(&sleepCond, &main.sleepLock); + // 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_release(&main.sleepLock); + // We have returned from the idle loop, which means that all threads are + // finished. Note that setting 'searching' and decreasing splitPointsSize is + // done under lock protection to avoid a race with Thread::available_to(). + Threads.mutex.lock(); + sp.mutex.lock(); + + searching = true; + --splitPointsSize; + activeSplitPoint = sp.parentSplitPoint; + activePosition = &pos; + pos.set_nodes_searched(pos.nodes_searched() + sp.nodes); + *bestMove = sp.bestMove; + *bestValue = sp.bestValue; + + sp.mutex.unlock(); + Threads.mutex.unlock(); } +// wait_for_think_finished() waits for main thread to go to sleep then returns -// 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 ThreadPool::wait_for_think_finished() { -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 - - if (!asyncMode) - cond_wait(&sleepCond, &main.sleepLock); - - lock_release(&main.sleepLock); + MainThread* th = main(); + th->mutex.lock(); + while (th->thinking) sleepCondition.wait(th->mutex); + th->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_thinking() wakes up the main thread sleeping in MainThread::idle_loop() +// so to start a new search, then returns immediately. + +void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, StateStackPtr& states) { -void ThreadsManager::wait_for_stop_or_ponderhit() { + wait_for_think_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); + RootMoves.clear(); + RootPos = pos; + Limits = limits; + if (states.get()) // If we don't set a new position, preserve current state + { + SetupStates = states; // Ownership transfer here + assert(!states.get()); + } - while (!Search::Signals.stop) - cond_wait(&main.sleepCond, &main.sleepLock); + for (MoveList it(pos); *it; ++it) + if ( limits.searchmoves.empty() + || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), *it)) + RootMoves.push_back(RootMove(*it)); - lock_release(&main.sleepLock); + main()->thinking = true; + main()->notify_one(); // Starts main thread }