X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=4fc0953eac6ec7145ba8ade4e16b228db6d70e0f;hp=271890c6f008b85eedee8d2302f0ebe4f2353e57;hb=4739037f967ac3c818907e89cc88c7b97021d027;hpb=ed0fb0b05fa72ccc6333bf5331eb9abeb7c86457 diff --git a/src/thread.cpp b/src/thread.cpp index 271890c6..4fc0953e 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) 2008-2014 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 @@ -17,420 +17,363 @@ 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" { - - // 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. +extern void check_time(); - long start_routine(Thread* th) { (th->*(th->start_fn))(); return 0; } +namespace { -} } + // 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(). + extern "C" { long start_routine(ThreadBase* th) { th->idle_loop(); return 0; } } -// 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. -Thread::Thread(Fn fn) : splitPoints() { + // 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. - is_searching = do_exit = false; - maxPly = splitPointsCnt = 0; - curSplitPoint = NULL; - start_fn = fn; - idx = Threads.size(); + template T* new_thread() { + T* th = new T(); + thread_create(th->handle, start_routine, th); // Will go to sleep + return th; + } - do_sleep = (fn != &Thread::main_loop); // Avoid a race with start_searching() + 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; + } - if (!thread_create(handle, start_routine, this)) - { - std::cerr << "Failed to create thread number " << idx << std::endl; - ::exit(EXIT_FAILURE); - } } -// Thread d'tor waits for thread termination before to return. +// notify_one() wakes up the thread when there is some work to do -Thread::~Thread() { +void ThreadBase::notify_one() { - assert(do_sleep); - - do_exit = true; // Search must be already finished - wake_up(); - thread_join(handle); // Wait for thread termination + mutex.lock(); + sleepCondition.notify_one(); + mutex.unlock(); } -// 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(); +// wait_for() set the thread to sleep until condition 'b' turns true -void Thread::timer_loop() { +void ThreadBase::wait_for(volatile const bool& b) { - while (!do_exit) - { - mutex.lock(); - sleepCondition.wait_for(mutex, maxPly ? maxPly : INT_MAX); - mutex.unlock(); - check_time(); - } + mutex.lock(); + while (!b) sleepCondition.wait(mutex); + mutex.unlock(); } -// 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. +// Thread c'tor just inits data and does not launch any execution thread. +// Such a thread will only be started when c'tor returns. -void Thread::main_loop() { +Thread::Thread() /* : splitPoints() */ { // Value-initialization bug in MSVC - while (true) - { - mutex.lock(); + searching = false; + maxPly = splitPointsSize = 0; + activeSplitPoint = NULL; + activePosition = NULL; + idx = Threads.size(); // Starts from 0 +} - do_sleep = true; // Always return to sleep after a search - is_searching = false; - while (do_sleep && !do_exit) - { - Threads.sleepCondition.notify_one(); // Wake up UI thread if needed - sleepCondition.wait(mutex); - } +// cutoff_occurred() checks whether a beta cutoff has occurred in the +// current active split point, or in some ancestor of the split point. - mutex.unlock(); +bool Thread::cutoff_occurred() const { - if (do_exit) - return; + for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint) + if (sp->cutoff) + return true; - is_searching = true; + return false; +} - Search::think(); - assert(is_searching); - } -} +// 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). +bool Thread::available_to(const Thread* master) const { -// Thread::wake_up() wakes up the thread, normally at the beginning of the search -// or, if "sleeping threads" is used at split time. + if (searching) + return false; -void Thread::wake_up() { + // 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; - mutex.lock(); - sleepCondition.notify_one(); - mutex.unlock(); + // 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); } -// 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. +// 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. -void Thread::wait_for_stop_or_ponderhit() { +void TimerThread::idle_loop() { - Signals.stopOnPonderhit = true; + while (!exit) + { + mutex.lock(); - mutex.lock(); - while (!Signals.stop) sleepCondition.wait(mutex);; - mutex.unlock(); + if (!exit) + sleepCondition.wait_for(mutex, run ? Resolution : INT_MAX); + + mutex.unlock(); + + if (run) + check_time(); + } } -// 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 (true) + { + mutex.lock(); - return false; -} + thinking = false; + while (!thinking && !exit) + { + Threads.sleepCondition.notify_one(); // Wake up the UI thread if needed + sleepCondition.wait(mutex); + } -// 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). + mutex.unlock(); -bool Thread::is_available_to(Thread* master) const { + if (exit) + return; - if (is_searching) - return false; + searching = true; - // 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; + Search::think(); - // 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)); + assert(searching); + + searching = false; + } } -// 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. +// 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() { - timer = new Thread(&Thread::timer_loop); - threads.push_back(new Thread(&Thread::main_loop)); + timer = new_thread(); + push_back(new_thread()); read_uci_options(); } -// exit() cleanly terminates the threads before the program exits. +// 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. void ThreadPool::exit() { - delete timer; // As first becuase check_time() accesses threads data + delete_thread(timer); // As first because check_time() accesses threads data - for (size_t i = 0; i < threads.size(); i++) - delete threads[i]; + for (iterator it = begin(); it != end(); ++it) + delete_thread(*it); } // 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. +// 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"]; + minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; + size_t requested = Options["Threads"]; assert(requested > 0); - while (threads.size() < requested) - threads.push_back(new Thread(&Thread::idle_loop)); - - while (threads.size() > requested) - { - delete threads.back(); - threads.pop_back(); - } -} - + // If zero (default) then set best minimum split depth automatically + if (!minimumSplitDepth) + minimumSplitDepth = requested < 8 ? 4 * ONE_PLY : 7 * ONE_PLY; -// 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. + while (size() < requested) + push_back(new_thread()); -void ThreadPool::wake_up() const { - - for (size_t i = 0; i < threads.size(); i++) + while (size() > requested) { - 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. - -void ThreadPool::sleep() const { - - // Main thread will go to sleep by itself to avoid a race with start_searching() - for (size_t i = 1; i < threads.size(); i++) - threads[i]->do_sleep = true; -} - - -// available_slave_exists() tries to find an idle thread which is available as -// a slave for the thread 'master'. +// available_slave() tries to find an idle thread which is available as a slave +// for the thread 'master'. -bool ThreadPool::available_slave_exists(Thread* master) const { +Thread* ThreadPool::available_slave(const Thread* master) const { - for (size_t i = 0; i < threads.size(); i++) - if (threads[i]->is_available_to(master)) - return true; + for (const_iterator it = begin(); it != end(); ++it) + if ((*it)->available_to(master)) + return *it; - return false; + return NULL; } // 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); +// (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. - Thread* master = pos.this_thread(); +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) { - if (master->splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD) - return bestValue; + 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 = master->splitPoints[master->splitPointsCnt]; + SplitPoint& sp = splitPoints[splitPointsSize]; - sp.parent = master->curSplitPoint; - sp.master = master; - sp.cutoff = false; - sp.slavesMask = 1ULL << master->idx; + sp.masterThread = this; + sp.parentSplitPoint = activeSplitPoint; + sp.slavesMask = 0, sp.slavesMask.set(idx); sp.depth = depth; + sp.bestValue = *bestValue; sp.bestMove = *bestMove; - sp.threatMove = threatMove; sp.alpha = alpha; sp.beta = beta; sp.nodeType = nodeType; - sp.bestValue = bestValue; - sp.mp = mp; + sp.cutNode = cutNode; + sp.movePicker = movePicker; sp.moveCount = moveCount; sp.pos = &pos; sp.nodes = 0; + sp.cutoff = false; 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 + // 'searching' flag. This must be done under lock protection to avoid concurrent // allocation of the same slave by another master. - mutex.lock(); + Threads.mutex.lock(); sp.mutex.lock(); - 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() - - if (useSleepingThreads) - threads[i]->wake_up(); + sp.allSlavesSearching = true; // Must be set under lock protection + ++splitPointsSize; + activeSplitPoint = &sp; + activePosition = NULL; - if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included - break; - } - - master->splitPointsCnt++; - - sp.mutex.unlock(); - mutex.unlock(); + 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 + } // 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. + // 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. - if (slavesCnt || Fake) - { - master->idle_loop(); + sp.mutex.unlock(); + Threads.mutex.unlock(); - // 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::idle_loop(); // Force a call to base class idle_loop() + + // 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); // 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(); + // 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(); - master->is_searching = true; - master->splitPointsCnt--; - master->curSplitPoint = sp.parent; + 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(); - mutex.unlock(); - - return sp.bestValue; + Threads.mutex.unlock(); } -// 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) { - - timer->mutex.lock(); - timer->maxPly = msec; - timer->sleepCondition.notify_one(); // Wake up and restart the timer - timer->mutex.unlock(); -} +// wait_for_think_finished() waits for main thread to go to sleep then returns +void ThreadPool::wait_for_think_finished() { -// 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(); + MainThread* t = main(); t->mutex.lock(); - t->sleepCondition.notify_one(); // In case is waiting for stop or ponderhit - while (!t->do_sleep) sleepCondition.wait(t->mutex); + while (t->thinking) sleepCondition.wait(t->mutex); t->mutex.unlock(); } -// start_searching() wakes up the main thread sleeping in main_loop() so to start -// a new search, then returns immediately. +// start_thinking() wakes up the main thread sleeping in MainThread::idle_loop() +// so to start a new search, then returns immediately. -void ThreadPool::start_searching(const Position& pos, const LimitsType& limits, - const std::vector& searchMoves, StateStackPtr& states) { - wait_for_search_finished(); +void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, StateStackPtr& states) { + + wait_for_think_finished(); SearchTime = Time::now(); // As early as possible Signals.stopOnPonderhit = Signals.firstRootMove = false; Signals.stop = Signals.failedLowAtRoot = false; - RootPosition = pos; - Limits = limits; - SetupStates = states; // Ownership transfer here 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()); + } - for (MoveList ml(pos); !ml.end(); ++ml) - if (searchMoves.empty() || count(searchMoves.begin(), searchMoves.end(), ml.move())) - RootMoves.push_back(RootMove(ml.move())); + for (MoveList it(pos); *it; ++it) + if ( limits.searchmoves.empty() + || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), *it)) + RootMoves.push_back(RootMove(*it)); - main_thread()->do_sleep = false; - main_thread()->wake_up(); + main()->thinking = true; + main()->notify_one(); // Starts main thread }