X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=680cd3adb8c3dc52bf87c9225f2a08db33e65cd7;hp=b8686f4c48e5f69d0bed038ab9cab3bcfdf8cc18;hb=cff9a8672c1da7d36bc54d168d10ea2b1ce5c728;hpb=699f700162f410519e5510c667aebc9940d4e91e diff --git a/src/thread.cpp b/src/thread.cpp index b8686f4c..680cd3ad 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -1,7 +1,8 @@ /* 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 + Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, 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 @@ -18,436 +19,201 @@ */ #include -#include +#include // For std::count #include "movegen.h" #include "search.h" #include "thread.h" -#include "ucioption.h" +#include "uci.h" +#include "syzygy/tbprobe.h" +#include "tt.h" -using namespace Search; +ThreadPool Threads; // Global object -ThreadsManager Threads; // Global object -namespace { extern "C" { +/// Thread constructor launches the thread and waits until it goes to sleep +/// in idle_loop(). Note that 'searching' and 'exit' should be already set. - // 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. +Thread::Thread(size_t n) : idx(n), stdThread(&Thread::idle_loop, this) { - long start_routine(Thread* th) { - - Threads.set_this_thread(th); // Save pointer into thread local storage - (th->*(th->start_fn))(); - 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) { - - is_searching = do_exit = false; - maxPly = splitPointsCnt = 0; - curSplitPoint = NULL; - start_fn = fn; - idx = Threads.size(); - - do_sleep = (fn != &Thread::main_loop); // Avoid a race with start_searching() - - lock_init(sleepLock); - cond_init(sleepCond); - - for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++) - lock_init(splitPoints[j].lock); - - if (!thread_create(handle, start_routine, this)) - { - std::cerr << "Failed to create thread number " << idx << std::endl; - ::exit(EXIT_FAILURE); - } + wait_for_search_finished(); } -// Thread d'tor waits for thread termination before to return. +/// Thread destructor wakes up the thread in idle_loop() and waits +/// for its termination. Thread should be already waiting. Thread::~Thread() { - assert(do_sleep); - - do_exit = true; // Search must be already finished - wake_up(); - - thread_join(handle); // Wait for thread termination - - lock_destroy(sleepLock); - cond_destroy(sleepCond); - - for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++) - lock_destroy(splitPoints[j].lock); -} - - -// 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(); - -void Thread::timer_loop() { - - while (!do_exit) - { - lock_grab(sleepLock); - timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX); - lock_release(sleepLock); - check_time(); - } -} - - -// 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_exit) - { - cond_signal(Threads.sleepCond); // Wake up UI thread if needed - cond_wait(sleepCond, sleepLock); - } - - lock_release(sleepLock); - - if (do_exit) - return; - - is_searching = true; - - Search::think(); - } -} - - -// Thread::wake_up() wakes up the thread, normally at the beginning of the search -// or, if "sleeping threads" is used at split time. - -void Thread::wake_up() { - - lock_grab(sleepLock); - cond_signal(sleepCond); - lock_release(sleepLock); -} - - -// 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() { + assert(!searching); - Signals.stopOnPonderhit = true; - - lock_grab(sleepLock); - while (!Signals.stop) cond_wait(sleepCond, sleepLock); - lock_release(sleepLock); + exit = true; + start_searching(); + stdThread.join(); } +/// Thread::bestMoveCount(Move move) return best move counter for the given root move -// 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 { +int Thread::best_move_count(Move move) { - for (SplitPoint* sp = curSplitPoint; sp; sp = sp->parent) - if (sp->cutoff) - return true; + auto rm = std::find(rootMoves.begin() + pvIdx, + rootMoves.begin() + pvLast, move); - return false; + return rm != rootMoves.begin() + pvLast ? rm->bestMoveCount : 0; } +/// Thread::clear() reset histories, usually before a new game -// 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). +void Thread::clear() { -bool Thread::is_available_to(const Thread& master) const { + counterMoves.fill(MOVE_NONE); + mainHistory.fill(0); + captureHistory.fill(0); - if (is_searching) - return false; + for (bool inCheck : { false, true }) + for (StatsType c : { NoCaptures, Captures }) + for (auto& to : continuationHistory[inCheck][c]) + for (auto& h : to) + h->fill(0); - // 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; - - // 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)); + for (bool inCheck : { false, true }) + for (StatsType c : { NoCaptures, Captures }) + continuationHistory[inCheck][c][NO_PIECE][0]->fill(Search::CounterMovePruneThreshold - 1); } +/// Thread::start_searching() wakes up the thread that will start the search -// 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. - -void ThreadsManager::init() { +void Thread::start_searching() { - tls_init(tlsKey); - cond_init(sleepCond); - lock_init(splitLock); - timer = new Thread(&Thread::timer_loop); - threads.push_back(new Thread(&Thread::main_loop)); - read_uci_options(); + std::lock_guard lk(mutex); + searching = true; + cv.notify_one(); // Wake up the thread in idle_loop() } -// d'tor cleanly terminates the threads when the program exits. - -ThreadsManager::~ThreadsManager() { +/// Thread::wait_for_search_finished() blocks on the condition variable +/// until the thread has finished searching. - for (int i = 0; i < size(); i++) - delete threads[i]; +void Thread::wait_for_search_finished() { - delete timer; - lock_destroy(splitLock); - cond_destroy(sleepCond); - tls_destroy(tlsKey); + std::unique_lock lk(mutex); + cv.wait(lk, [&]{ return !searching; }); } -// 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. - -void ThreadsManager::read_uci_options() { - - maxThreadsPerSplitPoint = Options["Max Threads per Split Point"]; - minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; - useSleepingThreads = Options["Use Sleeping Threads"]; - int requested = Options["Threads"]; +/// Thread::idle_loop() is where the thread is parked, blocked on the +/// condition variable, when it has no work to do. - assert(requested > 0); +void Thread::idle_loop() { - while (size() < requested) - threads.push_back(new Thread(&Thread::idle_loop)); + // If OS already scheduled us on a different group than 0 then don't overwrite + // the choice, eventually we are one of many one-threaded processes running on + // some Windows NUMA hardware, for instance in fishtest. To make it simple, + // just check if running threads are below a threshold, in this case all this + // NUMA machinery is not needed. + if (Options["Threads"] > 8) + WinProcGroup::bindThisThread(idx); - while (size() > requested) + while (true) { - delete threads.back(); - threads.pop_back(); - } -} - - -// 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. + std::unique_lock lk(mutex); + searching = false; + cv.notify_one(); // Wake up anyone waiting for search finished + cv.wait(lk, [&]{ return searching; }); -void ThreadsManager::wake_up() const { + if (exit) + return; - for (int i = 0; i < size(); i++) - { - threads[i]->maxPly = 0; - threads[i]->do_sleep = false; + lk.unlock(); - if (!useSleepingThreads) - threads[i]->wake_up(); + search(); } } +/// ThreadPool::set() creates/destroys threads to match the requested number. +/// Created and launched threads will immediately go to sleep in idle_loop. +/// Upon resizing, threads are recreated to allow for binding if necessary. -// sleep() is called after the search finishes to ask all the threads but the -// main one to go waiting on a sleep condition. - -void ThreadsManager::sleep() const { - - for (int i = 1; i < size(); i++) // Main thread will go to sleep by itself - threads[i]->do_sleep = true; // to avoid a race with start_searching() -} - - -// available_slave_exists() tries to find an idle thread which is available as -// a slave for the thread 'master'. - -bool ThreadsManager::available_slave_exists(const Thread& master) const { +void ThreadPool::set(size_t requested) { - for (int i = 0; i < size(); i++) - if (threads[i]->is_available_to(master)) - return true; - - return false; -} + if (size() > 0) { // destroy any existing thread(s) + main()->wait_for_search_finished(); - -// 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, 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); - - Thread& master = pos.this_thread(); - - if (master.splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD) - return bestValue; - - // 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 = mp; - sp->moveCount = moveCount; - sp->pos = &pos; - sp->nodes = 0; - 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 - // allocation of the same slave by another master. - lock_grab(sp->lock); - lock_grab(splitLock); - - for (int i = 0; i < 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(); - - if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included - break; - } - - lock_release(splitLock); - lock_release(sp->lock); - - // 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) - { - master.idle_loop(sp); - - // In helpful master concept a master can help only a sub-tree of its split - // point, and because here is all finished is not possible master is booked. - assert(!master.is_searching); + while (size() > 0) + delete back(), pop_back(); } - // 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(). - lock_grab(sp->lock); // To protect sp->nodes - lock_grab(splitLock); + if (requested > 0) { // create new thread(s) + push_back(new MainThread(0)); - master.is_searching = true; - master.splitPointsCnt--; - master.curSplitPoint = sp->parent; - pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); - *bestMove = sp->bestMove; + while (size() < requested) + push_back(new Thread(size())); + clear(); - lock_release(splitLock); - lock_release(sp->lock); + // Reallocate the hash with the new threadpool size + TT.resize(Options["Hash"]); - return sp->bestValue; + // Init thread number dependent search params. + Search::init(); + } } -// Explicit template instantiations -template Value ThreadsManager::split(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker*, int); -template Value ThreadsManager::split(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker*, int); - +/// ThreadPool::clear() sets threadPool data to initial values. -// ThreadsManager::set_timer() is used to set the timer to trigger after msec -// milliseconds. If msec is 0 then timer is stopped. +void ThreadPool::clear() { -void ThreadsManager::set_timer(int msec) { + for (Thread* th : *this) + th->clear(); - lock_grab(timer->sleepLock); - timer->maxPly = msec; - cond_signal(timer->sleepCond); // Wake up and restart the timer - lock_release(timer->sleepLock); + main()->callsCnt = 0; + main()->previousScore = VALUE_INFINITE; + main()->previousTimeReduction = 1.0; } +/// ThreadPool::start_thinking() wakes up main thread waiting in idle_loop() and +/// returns immediately. Main thread will wake up other threads and start the search. -// ThreadsManager::wait_for_search_finished() waits for main thread to go to -// sleep, this means search is finished. Then returns. +void ThreadPool::start_thinking(Position& pos, StateListPtr& states, + const Search::LimitsType& limits, bool ponderMode) { -void ThreadsManager::wait_for_search_finished() { + main()->wait_for_search_finished(); - Thread* t = main_thread(); - lock_grab(t->sleepLock); - cond_signal(t->sleepCond); // In case is waiting for stop or ponderhit - while (!t->do_sleep) cond_wait(sleepCond, t->sleepLock); - lock_release(t->sleepLock); -} + main()->stopOnPonderhit = stop = false; + main()->ponder = ponderMode; + Search::Limits = limits; + Search::RootMoves rootMoves; + for (const auto& m : MoveList(pos)) + if ( limits.searchmoves.empty() + || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m)) + rootMoves.emplace_back(m); -// ThreadsManager::start_searching() wakes up the main thread sleeping in -// main_loop() so to start a new search, then returns immediately. + if (!rootMoves.empty()) + Tablebases::rank_root_moves(pos, rootMoves); -void ThreadsManager::start_searching(const Position& pos, const LimitsType& limits, - const std::vector& searchMoves) { - wait_for_search_finished(); + // After ownership transfer 'states' becomes empty, so if we stop the search + // and call 'go' again without setting a new position states.get() == NULL. + assert(states.get() || setupStates.get()); - SearchTime.restart(); // As early as possible + if (states.get()) + setupStates = std::move(states); // Ownership transfer, states is now empty - Signals.stopOnPonderhit = Signals.firstRootMove = false; - Signals.stop = Signals.failedLowAtRoot = false; + // We use Position::set() to set root position across threads. But there are + // some StateInfo fields (previous, pliesFromNull, capturedPiece) that cannot + // be deduced from a fen string, so set() clears them and to not lose the info + // we need to backup and later restore setupStates->back(). Note that setupStates + // is shared by threads but is accessed in read-only mode. + StateInfo tmp = setupStates->back(); - RootPosition.copy(pos, main_thread()); - Limits = limits; - RootMoves.clear(); + for (Thread* th : *this) + { + th->shuffleExts = th->nodes = th->tbHits = th->nmpMinPly = 0; + th->rootDepth = th->completedDepth = 0; + th->rootMoves = rootMoves; + th->rootPos.set(pos.fen(), pos.is_chess960(), &setupStates->back(), th); + } - for (MoveList ml(pos); !ml.end(); ++ml) - if (searchMoves.empty() || count(searchMoves.begin(), searchMoves.end(), ml.move())) - RootMoves.push_back(RootMove(ml.move())); + setupStates->back() = tmp; - main_thread()->do_sleep = false; - main_thread()->wake_up(); + main()->start_searching(); }