X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=6eb00d63a649b463d6dc1edea86ee3fe51f6b72a;hp=cdb0d5412a501b781e4064815eacf255b358ef66;hb=ef38046e734490499e4b6254dbf4fc434bf37fc9;hpb=3c0fe1d9b252acca8107b7bece636e2bc20d6411 diff --git a/src/thread.cpp b/src/thread.cpp index cdb0d541..6eb00d63 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -2,6 +2,7 @@ Stockfish, a UCI chess playing engine derived from Glaurung 2.1 Copyright (C) 2004-2008 Tord Romstad (Glaurung author) 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 @@ -17,359 +18,202 @@ along with this program. If not, see . */ -#include // For std::count #include +#include // For std::count #include "movegen.h" #include "search.h" #include "thread.h" #include "uci.h" - -using namespace Search; +#include "syzygy/tbprobe.h" +#include "tt.h" ThreadPool Threads; // Global object -extern void check_time(); - -namespace { - - // 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() { - std::thread* th = new T; - *th = std::thread(&T::idle_loop, (T*)th); // Will go to sleep - return (T*)th; - } - - void delete_thread(ThreadBase* th) { - - th->mutex.lock(); - th->exit = true; // Search must be already finished - th->mutex.unlock(); - th->notify_one(); - th->join(); // Wait for thread termination - delete th; - } +/// Thread constructor launches the thread and waits until it goes to sleep +/// in idle_loop(). Note that 'searching' and 'exit' should be already set. -} - - -// ThreadBase::notify_one() wakes up the thread when there is some work to do - -void ThreadBase::notify_one() { +Thread::Thread(size_t n) : idx(n), stdThread(&Thread::idle_loop, this) { - std::unique_lock lk(mutex); - sleepCondition.notify_one(); + wait_for_search_finished(); } -// ThreadBase::wait_for() set the thread to sleep until 'condition' turns true - -void ThreadBase::wait_for(volatile const bool& condition) { - - std::unique_lock lk(mutex); - sleepCondition.wait(lk, [&]{ return condition; }); -} - +/// Thread destructor wakes up the thread in idle_loop() and waits +/// for its termination. Thread should be already waiting. -// Thread c'tor makes some init but does not launch any execution thread that -// will be started only when c'tor returns. +Thread::~Thread() { -Thread::Thread() /* : splitPoints() */ { // Initialization of non POD broken in MSVC + assert(!searching); - searching = false; - maxPly = 0; - splitPointsSize = 0; - activeSplitPoint = nullptr; - activePosition = nullptr; - idx = Threads.size(); // Starts from 0 + 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 = activeSplitPoint; sp; sp = sp->parentSplitPoint) - 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::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). +void Thread::clear() { -bool Thread::can_join(const SplitPoint* sp) const { + counterMoves.fill(MOVE_NONE); + mainHistory.fill(0); + captureHistory.fill(0); - if (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 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 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(sp->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 -// 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.spinlock.acquire(); // No contention here until we don't increment 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; - sp.allSlavesSearching = true; // Must be set under lock protection - - ++splitPointsSize; - activeSplitPoint = &sp; - activePosition = nullptr; - - // Try to allocate available threads - Thread* slave; - - while ( sp.slavesMask.count() < MAX_SLAVES_PER_SPLITPOINT - && (slave = Threads.available_slave(&sp)) != nullptr) - { - slave->spinlock.acquire(); - - if (slave->can_join(activeSplitPoint)) - { - activeSplitPoint->slavesMask.set(slave->idx); - slave->activeSplitPoint = activeSplitPoint; - slave->searching = true; - } - - slave->spinlock.release(); - } - - // 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(); - - 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 decreasing splitPointsSize must be done under lock - // protection to avoid a race with Thread::can_join(). - spinlock.acquire(); +void Thread::start_searching() { + std::lock_guard lk(mutex); searching = true; - --splitPointsSize; - activeSplitPoint = sp.parentSplitPoint; - activePosition = &pos; - - spinlock.release(); - - // Split point data cannot be changed now, so no need to lock protect - pos.set_nodes_searched(pos.nodes_searched() + sp.nodes); - *bestMove = sp.bestMove; - *bestValue = sp.bestValue; + cv.notify_one(); // Wake up the thread in idle_loop() } -// 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. +/// Thread::wait_for_search_finished() blocks on the condition variable +/// until the thread has finished searching. -void TimerThread::idle_loop() { - - while (!exit) - { - std::unique_lock lk(mutex); +void Thread::wait_for_search_finished() { - if (!exit) - sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX)); - - lk.unlock(); - - if (run) - check_time(); - } + std::unique_lock lk(mutex); + cv.wait(lk, [&]{ return !searching; }); } -// 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. +/// Thread::idle_loop() is where the thread is parked, blocked on the +/// condition variable, when it has no work to do. -void MainThread::idle_loop() { +void Thread::idle_loop() { - while (!exit) - { - std::unique_lock lk(mutex); + // 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); - thinking = false; + while (true) + { + std::unique_lock lk(mutex); + searching = false; + cv.notify_one(); // Wake up anyone waiting for search finished + cv.wait(lk, [&]{ return searching; }); - while (!thinking && !exit) - { - sleepCondition.notify_one(); // Wake up the UI thread if needed - sleepCondition.wait(lk); - } + if (exit) + return; lk.unlock(); - if (!exit) - { - searching = true; - - Search::think(); - - assert(searching); - - searching = false; - } + 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. -// MainThread::join() waits for main thread to finish the search +void ThreadPool::set(size_t requested) { -void MainThread::join() { + if (size() > 0) { // destroy any existing thread(s) + main()->wait_for_search_finished(); - std::unique_lock lk(mutex); - sleepCondition.wait(lk, [&]{ return !thinking; }); -} + while (size() > 0) + delete back(), pop_back(); + } + if (requested > 0) { // create new thread(s) + push_back(new MainThread(0)); -// 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. + while (size() < requested) + push_back(new Thread(size())); + clear(); -void ThreadPool::init() { + // Reallocate the hash with the new threadpool size + TT.resize(Options["Hash"]); - timer = new_thread(); - push_back(new_thread()); - read_uci_options(); + // Init thread number dependent search params. + Search::init(); + } } +/// ThreadPool::clear() sets threadPool data to initial values. -// ThreadPool::exit() terminates the threads before the program exits. Cannot be -// done in d'tor because threads must be terminated before freeing us. - -void ThreadPool::exit() { - - delete_thread(timer); // As first because check_time() accesses threads data - timer = nullptr; +void ThreadPool::clear() { for (Thread* th : *this) - delete_thread(th); + th->clear(); - clear(); // Get rid of stale pointers + 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. -// 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. +void ThreadPool::start_thinking(Position& pos, StateListPtr& states, + const Search::LimitsType& limits, bool ponderMode) { -void ThreadPool::read_uci_options() { + main()->wait_for_search_finished(); - minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; - size_t requested = Options["Threads"]; + main()->stopOnPonderhit = stop = false; + main()->ponder = ponderMode; + Search::Limits = limits; + Search::RootMoves rootMoves; - assert(requested > 0); - - while (size() < requested) - push_back(new_thread()); + for (const auto& m : MoveList(pos)) + if ( limits.searchmoves.empty() + || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m)) + rootMoves.emplace_back(m); - while (size() > requested) - { - delete_thread(back()); - pop_back(); - } -} + if (!rootMoves.empty()) + Tablebases::rank_root_moves(pos, rootMoves); + // 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()); -// ThreadPool::available_slave() tries to find an idle thread which is available -// to join SplitPoint 'sp'. + if (states.get()) + setupStates = std::move(states); // Ownership transfer, states is now empty -Thread* ThreadPool::available_slave(const SplitPoint* sp) const { + // 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(); for (Thread* th : *this) - if (th->can_join(sp)) - return th; - - return nullptr; -} - - -// ThreadPool::start_thinking() wakes up the main thread sleeping in -// MainThread::idle_loop() and starts a new search, then returns immediately. - -void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, - StateStackPtr& states) { - main()->join(); - - Signals.stopOnPonderhit = Signals.firstRootMove = false; - Signals.stop = Signals.failedLowAtRoot = false; - - 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()); + 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 (const auto& m : MoveList(pos)) - if ( limits.searchmoves.empty() - || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m)) - RootMoves.push_back(RootMove(m)); + setupStates->back() = tmp; - main()->thinking = true; - main()->notify_one(); // Wake up main thread: 'thinking' must be already set + main()->start_searching(); }