X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=6516192d3700cc4dea05ba26f396cc09a6e37594;hp=60c0ce71c683b0f917a8b2e28535eb8ff2548b2f;hb=4c57cf0ead29536504ca452b876d350a8e2edbdc;hpb=a1a7bc84da4a6f180179e437494dd485cba1711d diff --git a/src/thread.cpp b/src/thread.cpp index 60c0ce71..6516192d 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) 2008-2014 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2015-2018 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 @@ -23,369 +24,176 @@ #include "movegen.h" #include "search.h" #include "thread.h" -#include "ucioption.h" - -using namespace Search; +#include "uci.h" +#include "syzygy/tbprobe.h" ThreadPool Threads; // Global object -extern void check_time(); - -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 constructor launches the thread and waits until it goes to sleep +/// in idle_loop(). Note that 'searching' and 'exit' should be alredy set. - // 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. - - template T* new_thread() { - T* th = new T(); - thread_create(th->handle, start_routine, th); // Will go to sleep - return th; - } - - 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; - } +Thread::Thread(size_t n) : idx(n), stdThread(&Thread::idle_loop, this) { + wait_for_search_finished(); } -// notify_one() wakes up the thread when there is some work to do - -void ThreadBase::notify_one() { - - mutex.lock(); - sleepCondition.notify_one(); - mutex.unlock(); -} - +/// Thread destructor wakes up the thread in idle_loop() and waits +/// for its termination. Thread should be already waiting. -// wait_for() set the thread to sleep until condition 'b' turns true +Thread::~Thread() { -void ThreadBase::wait_for(volatile const bool& b) { + assert(!searching); - mutex.lock(); - while (!b) sleepCondition.wait(mutex); - mutex.unlock(); + exit = true; + start_searching(); + stdThread.join(); } -// Thread c'tor just inits data and does not launch any execution thread. -// Such a thread will only be started when c'tor returns. - -Thread::Thread() /* : splitPoints() */ { // Value-initialization bug in MSVC - - searching = false; - maxPly = splitPointsSize = 0; - activeSplitPoint = NULL; - activePosition = NULL; - idx = Threads.size(); // Starts from 0 -} - +/// Thread::clear() reset histories, usually before a new game -// cutoff_occurred() checks whether a beta cutoff has occurred in the -// current active split point, or in some ancestor of the split point. +void Thread::clear() { -bool Thread::cutoff_occurred() const { + counterMoves.fill(MOVE_NONE); + mainHistory.fill(0); + captureHistory.fill(0); - for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint) - if (sp->cutoff) - return true; + for (auto& to : contHistory) + for (auto& h : to) + h.fill(0); - return false; + contHistory[NO_PIECE][0].fill(Search::CounterMovePruneThreshold - 1); } +/// Thread::start_searching() wakes up the thread that will start the search -// 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 { - - if (searching) - return false; +void Thread::start_searching() { - // 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. - int 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 & (1ULL << master->idx)); + std::lock_guard lk(mutex); + searching = true; + cv.notify_one(); // Wake up the thread in idle_loop() } -// 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 TimerThread::idle_loop() { - - while (!exit) - { - mutex.lock(); - - if (!exit) - sleepCondition.wait_for(mutex, run ? Resolution : INT_MAX); +/// Thread::wait_for_search_finished() blocks on the condition variable +/// until the thread has finished searching. - mutex.unlock(); +void Thread::wait_for_search_finished() { - 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 Thread::idle_loop() { -void MainThread::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 (true) { - mutex.lock(); - - thinking = false; - - while (!thinking && !exit) - { - Threads.sleepCondition.notify_one(); // Wake up the UI thread if needed - sleepCondition.wait(mutex); - } - - mutex.unlock(); + std::unique_lock lk(mutex); + searching = false; + cv.notify_one(); // Wake up anyone waiting for search finished + cv.wait(lk, [&]{ return searching; }); if (exit) return; - searching = true; + lk.unlock(); - Search::think(); - - assert(searching); - - searching = false; + search(); } } +/// ThreadPool::set() creates/destroys threads to match the requested number. +/// Created and launced threads wil go immediately to sleep in idle_loop. +/// Upon resizing, threads are recreated to allow for binding if necessary. -// init() is called at startup to create and launch requested threads, that will -// go immediately to sleep due to 'sleepWhileIdle' set to true. 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::set(size_t requested) { -void ThreadPool::init() { + if (size() > 0) { // destroy any existing thread(s) + main()->wait_for_search_finished(); - sleepWhileIdle = true; - timer = new_thread(); - push_back(new_thread()); - read_uci_options(); -} - - -// exit() cleanly terminates the threads before the program exits - -void ThreadPool::exit() { - - delete_thread(timer); // As first because check_time() accesses threads data - - 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 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() { - - minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; - size_t requested = Options["Threads"]; - - assert(requested > 0); - - // If zero (default) then set best minimum split depth automatically - if (!minimumSplitDepth) - minimumSplitDepth = requested < 8 ? 4 * ONE_PLY : 7 * ONE_PLY; - - while (size() < requested) - push_back(new_thread()); - - while (size() > requested) - { - delete_thread(back()); - pop_back(); + while (size() > 0) + delete back(), pop_back(); } -} - - -// available_slave() tries to find an idle thread which is available as a slave -// for the thread 'master'. - -Thread* ThreadPool::available_slave(const Thread* master) const { - for (const_iterator it = begin(); it != end(); ++it) - if ((*it)->available_to(master)) - return *it; + if (requested > 0) { // create new thread(s) + push_back(new MainThread(0)); - return NULL; + while (size() < requested) + push_back(new Thread(size())); + clear(); + } } +/// ThreadPool::clear() sets threadPool data to initial values. -// 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. - -template -void Thread::split(Position& pos, const Stack* ss, Value alpha, Value beta, Value* bestValue, - Move* bestMove, Depth depth, int moveCount, - MovePicker* movePicker, int nodeType, bool cutNode) { - - 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 = 1ULL << 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.mutex.lock(); - sp.mutex.lock(); - - ++splitPointsSize; - activeSplitPoint = &sp; - activePosition = NULL; - - int slavesCnt = 1; // This thread is always included - Thread* slave; - - while (!Fake && (slave = Threads.available_slave(this)) != NULL) - { - ++slavesCnt; - sp.slavesMask |= 1ULL << slave->idx; - slave->activeSplitPoint = &sp; - slave->searching = true; // Slave leaves idle_loop() - slave->notify_one(); // Could be sleeping - } +void ThreadPool::clear() { - // 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. - if (slavesCnt > 1 || Fake) - { - sp.mutex.unlock(); - Threads.mutex.unlock(); - - 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 'searching' and decreasing splitPointsSize is - // done under lock protection to avoid a race with Thread::available_to(). - Threads.mutex.lock(); - sp.mutex.lock(); - } + for (Thread* th : *this) + th->clear(); - 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(); + main()->callsCnt = 0; + main()->previousScore = VALUE_INFINITE; + main()->previousTimeReduction = 1.0; } -// Explicit template instantiations -template void Thread::split(Position&, const Stack*, Value, Value, Value*, Move*, Depth, int, MovePicker*, int, bool); -template void Thread::split< true>(Position&, const Stack*, Value, Value, Value*, Move*, Depth, int, MovePicker*, int, bool); +/// 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. +void ThreadPool::start_thinking(Position& pos, StateListPtr& states, + const Search::LimitsType& limits, bool ponderMode) { -// wait_for_think_finished() waits for main thread to go to sleep then returns + main()->wait_for_search_finished(); -void ThreadPool::wait_for_think_finished() { - - MainThread* t = main(); - t->mutex.lock(); - while (t->thinking) sleepCondition.wait(t->mutex); - t->mutex.unlock(); -} + stopOnPonderhit = stop = false; + 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); -// 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) { + if (!rootMoves.empty()) + Tablebases::filter_root_moves(pos, rootMoves); - wait_for_think_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 = Time::now(); // 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(); - RootMoves.clear(); - RootPos = pos; - Limits = limits; - if (states.get()) // If we don't set a new position, preserve current state + for (Thread* th : *this) { - SetupStates = states; // Ownership transfer here - assert(!states.get()); + th->nodes = th->tbHits = th->nmp_ply = th->nmp_odd = 0; + th->rootDepth = th->completedDepth = DEPTH_ZERO; + th->rootMoves = rootMoves; + th->rootPos.set(pos.fen(), pos.is_chess960(), &setupStates->back(), th); } - for (MoveList it(pos); *it; ++it) - if ( limits.searchmoves.empty() - || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), *it)) - RootMoves.push_back(RootMove(*it)); + setupStates->back() = tmp; - main()->thinking = true; - main()->notify_one(); // Starts main thread + main()->start_searching(); }