X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=541610a1ca4b2bd5a63d16323c4d2aed991aae21;hp=3b87a72415ddd0fd3fc57ba345a55c3af78fbc97;hb=e53774bc49dd0aaa1c129ee98c09e1a56ef974fb;hpb=b1768c115cf2bbe7ed6f89dc53a8db85b4442353 diff --git a/src/thread.cpp b/src/thread.cpp index 3b87a724..ff52576b 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 @@ -17,131 +17,84 @@ 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; -ThreadsManager Threads; // Global object +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 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(). - -#if defined(_WIN32) || defined(_WIN64) - DWORD WINAPI start_routine(LPVOID thread) { -#else - void* start_routine(void* thread) { -#endif - - Thread* th = (Thread*)thread; - - if (th->threadID == 0) - th->main_loop(); - - else if (th->threadID == MAX_THREADS) - th->timer_loop(); - - else - th->idle_loop(NULL); - - return 0; - } - -} } - - -// 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 check_time(); -void Thread::timer_loop() { +namespace { - while (!do_exit) - { - lock_grab(sleepLock); - timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX); - lock_release(sleepLock); - check_time(); - } -} + // Helpers to launch a thread after creation and joining before delete. Must be + // outside Thread c'tor and d'tor because the object must be fully initialized + // when start_routine (and hence virtual idle_loop) is called and when joining. + template T* new_thread() { + T* th = new T(); + th->nativeThread = std::thread(&ThreadBase::idle_loop, th); // Will go to sleep + return 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 delete_thread(ThreadBase* th) { -void Thread::main_loop() { - - while (true) - { - lock_grab(sleepLock); + th->mutex.lock(); + th->exit = true; // Search must be already finished + th->mutex.unlock(); - do_sleep = true; // Always return to sleep after a search - is_searching = false; + th->notify_one(); + th->nativeThread.join(); // Wait for thread termination + delete th; + } - while (do_sleep && !do_exit) - { - cond_signal(Threads.sleepCond); // Wake up UI thread if needed - cond_wait(sleepCond, sleepLock); - } +} - lock_release(sleepLock); - if (do_exit) - return; +// ThreadBase::notify_one() wakes up the thread when there is some work to do - is_searching = true; +void ThreadBase::notify_one() { - Search::think(); - } + std::unique_lock(this->mutex); + sleepCondition.notify_one(); } -// Thread::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::wait_for() set the thread to sleep until 'condition' turns true -void Thread::wake_up() { +void ThreadBase::wait_for(volatile const bool& condition) { - lock_grab(sleepLock); - cond_signal(sleepCond); - lock_release(sleepLock); + std::unique_lock lk(mutex); + sleepCondition.wait(lk, [&]{ return condition; }); } -// 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. - -void Thread::wait_for_stop_or_ponderhit() { +// Thread c'tor makes some init but does not launch any execution thread that +// will be started only when c'tor returns. - Signals.stopOnPonderhit = true; +Thread::Thread() /* : splitPoints() */ { // Initialization of non POD broken in MSVC - lock_grab(sleepLock); - - while (!Signals.stop) - cond_wait(sleepCond, sleepLock); - - lock_release(sleepLock); + searching = false; + maxPly = 0; + splitPointsSize = 0; + activeSplitPoint = nullptr; + activePosition = nullptr; + idx = Threads.size(); // Starts from 0 } -// cutoff_occurred() checks whether a beta cutoff has occurred in the current -// active split point, or in some ancestor of the split point. +// 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 Thread::cutoff_occurred() const { - for (SplitPoint* sp = splitPoint; sp; sp = sp->parent) + for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint) if (sp->cutoff) return true; @@ -149,323 +102,275 @@ bool Thread::cutoff_occurred() const { } -// 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::can_join() checks whether the thread is available to join the split +// point 'sp'. 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 for the split points below his active +// one (the "helpful master" concept in YBWC terminology). -bool Thread::is_available_to(int master) const { +bool Thread::can_join(const SplitPoint* sp) const { - if (is_searching) + if (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 sp_count = activeSplitPoints; + // 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; - // No active split points means that the thread is available as a slave for any + // 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 !sp_count || (splitPoints[sp_count - 1].slavesMask & (1ULL << master)); + return !size || splitPoints[size - 1].slavesMask.test(sp->master->idx); } -// read_uci_options() updates internal threads parameters from the corresponding -// UCI options. It is called before to start a new search. - -void ThreadsManager::read_uci_options() { - - maxThreadsPerSplitPoint = Options["Max Threads per Split Point"]; - minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; - useSleepingThreads = Options["Use Sleeping Threads"]; -} +// 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; + // 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(); -// set_size() changes the number of active threads and raises do_sleep flag for -// all the unused threads that will go immediately to sleep. + sp.allSlavesSearching = true; // Must be set under lock protection + ++splitPointsSize; + activeSplitPoint = &sp; + activePosition = nullptr; -void ThreadsManager::set_size(int cnt) { + Thread* slave; - assert(cnt > 0 && cnt < MAX_THREADS); + while ( sp.slavesMask.count() < MAX_SLAVES_PER_SPLITPOINT + && (slave = Threads.available_slave(activeSplitPoint)) != nullptr) + { + sp.slavesMask.set(slave->idx); + slave->activeSplitPoint = activeSplitPoint; + slave->searching = true; // Slave leaves idle_loop() + slave->notify_one(); // Could be sleeping + } - activeThreads = cnt; + // 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(); - for (int i = 0; i < MAX_THREADS; i++) - 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(); - threads[i].maxPly = 0; + Thread::idle_loop(); // Force a call to base class idle_loop() - threads[i].do_sleep = false; + // 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); - if (!useSleepingThreads) - threads[i].wake_up(); - } - else - threads[i].do_sleep = true; + // 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(); } -// init() is called during startup. Initializes locks and condition variables -// and launches all threads sending them immediately to sleep. +// 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 ThreadsManager::init() { +void TimerThread::idle_loop() { - cond_init(sleepCond); - lock_init(splitLock); - - for (int i = 0; i <= MAX_THREADS; i++) + while (!exit) { - lock_init(threads[i].sleepLock); - cond_init(threads[i].sleepCond); + std::unique_lock lk(mutex); - for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_init(threads[i].splitPoints[j].lock); - } + if (!exit) + sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX)); - // Allocate main thread tables to call evaluate() also when not searching - threads[0].pawnTable.init(); - threads[0].materialTable.init(); + lk.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 = (i != 0); // Avoid a race with start_thinking() - threads[i].threadID = i; - - if (!thread_create(threads[i].handle, start_routine, threads[i])) - { - 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 (!exit) { - assert(threads[i].do_sleep); + std::unique_lock lk(mutex); - threads[i].do_exit = true; // Search must be already finished - threads[i].wake_up(); + thinking = false; - thread_join(threads[i].handle); // Wait for thread termination + while (!thinking && !exit) + { + Threads.sleepCondition.notify_one(); // Wake up the UI thread if needed + sleepCondition.wait(lk); + } - lock_destroy(threads[i].sleepLock); - cond_destroy(threads[i].sleepCond); + lk.unlock(); - for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_destroy(threads[i].splitPoints[j].lock); - } + if (!exit) + { + searching = true; - lock_destroy(splitLock); - cond_destroy(sleepCond); -} + Search::think(); + assert(searching); -// available_slave_exists() tries to find an idle thread which is available as -// a slave for the thread with threadID 'master'. + searching = false; + } + } +} -bool ThreadsManager::available_slave_exists(int master) const { - assert(master >= 0 && master < activeThreads); +// 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. - for (int i = 0; i < activeThreads; i++) - if (threads[i].is_available_to(master)) - return true; +void ThreadPool::init() { - return false; + timer = new_thread(); + push_back(new_thread()); + read_uci_options(); } -// 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 master = pos.thread(); - Thread& masterThread = threads[master]; - - 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]; - - sp->parent = masterThread.splitPoint; - sp->master = master; - sp->cutoff = false; - sp->slavesMask = 1ULL << master; - 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; - - assert(masterThread.is_searching); - - int slavesCnt = 0; +// ThreadPool::exit() terminates the threads before the program exits. Cannot be +// done in d'tor because threads must be terminated before freeing us. - // Try to allocate available threads and ask them to start searching setting - // is_searching flag. This must be done under lock protection to avoid concurrent - // allocation of the same slave by another master. - lock_grab(sp->lock); - lock_grab(splitLock); +void ThreadPool::exit() { - for (int i = 0; i < activeThreads && !Fake; i++) - if (threads[i].is_available_to(master)) - { - sp->slavesMask |= 1ULL << i; - threads[i].splitPoint = sp; - threads[i].is_searching = true; // Slave leaves idle_loop() + delete_thread(timer); // As first because check_time() accesses threads data - if (useSleepingThreads) - threads[i].wake_up(); + for (Thread* th : *this) + delete_thread(th); +} - if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included - break; - } - masterThread.splitPoint = sp; - masterThread.activeSplitPoints++; +// 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. - lock_release(splitLock); - lock_release(sp->lock); +void ThreadPool::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. - if (slavesCnt || Fake) - masterThread.idle_loop(sp); + minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; + size_t requested = Options["Threads"]; - // We have returned from the idle loop, which means that all threads are - // finished. Note that setting is_searching and decreasing activeSplitPoints is - // done under lock protection to avoid a race with Thread::is_available_to(). - lock_grab(sp->lock); // To protect sp->nodes - lock_grab(splitLock); + assert(requested > 0); - masterThread.is_searching = true; - masterThread.activeSplitPoints--; - masterThread.splitPoint = sp->parent; - pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); + // If zero (default) then set best minimum split depth automatically + if (!minimumSplitDepth) + minimumSplitDepth = requested < 8 ? 4 * ONE_PLY : 7 * ONE_PLY; - lock_release(splitLock); - lock_release(sp->lock); + while (size() < requested) + push_back(new_thread()); - return sp->bestValue; + while (size() > requested) + { + delete_thread(back()); + pop_back(); + } } -// 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); - -// ThreadsManager::set_timer() is used to set the timer to trigger after msec -// milliseconds. If msec is 0 then timer is stopped. +// ThreadPool::available_slave() tries to find an idle thread which is available +// to join SplitPoint 'sp'. -void ThreadsManager::set_timer(int msec) { +Thread* ThreadPool::available_slave(const SplitPoint* sp) const { - Thread& timer = threads[MAX_THREADS]; + for (Thread* th : *this) + if (th->can_join(sp)) + return th; - lock_grab(timer.sleepLock); - timer.maxPly = msec; - cond_signal(timer.sleepCond); // Wake up and restart the timer - lock_release(timer.sleepLock); + return nullptr; } -// 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::set& searchMoves, bool async) { - Thread& main = threads[0]; - - lock_grab(main.sleepLock); - - // 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; - RootMoves.clear(); - - // Populate RootMoves with all the legal moves (default) or, if a searchMoves - // set is given, with the subset of legal moves to search. - for (MoveList ml(pos); !ml.end(); ++ml) - if (searchMoves.empty() || searchMoves.count(ml.move())) - RootMoves.push_back(RootMove(ml.move())); - - // Reset signals before to start the new search - Signals.stopOnPonderhit = Signals.firstRootMove = false; - Signals.stop = Signals.failedLowAtRoot = false; - - main.do_sleep = false; - cond_signal(main.sleepCond); // Wake up main thread and start searching +// ThreadPool::wait_for_think_finished() waits for main thread to finish the search - if (!async) - while (!main.do_sleep) - cond_wait(sleepCond, main.sleepLock); +void ThreadPool::wait_for_think_finished() { - lock_release(main.sleepLock); + std::unique_lock lk(main()->mutex); + sleepCondition.wait(lk, [&]{ return !main()->thinking; }); } -// 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. +// ThreadPool::start_thinking() wakes up the main thread sleeping in +// MainThread::idle_loop() and starts a new search, then returns immediately. -void ThreadsManager::stop_thinking() { +void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, + StateStackPtr& states) { + wait_for_think_finished(); - Thread& main = threads[0]; + SearchTime = now(); // As early as possible - Search::Signals.stop = true; - - lock_grab(main.sleepLock); + Signals.stopOnPonderhit = Signals.firstRootMove = false; + Signals.stop = Signals.failedLowAtRoot = false; - cond_signal(main.sleepCond); // In case is waiting for stop or ponderhit + 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 (!main.do_sleep) - cond_wait(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 }