X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=1217e3ab020656787c9dd1b8a5886fecdadcaa1a;hp=4fc0953eac6ec7145ba8ade4e16b228db6d70e0f;hb=38112060dc2da351c6dde8f12d0ee5cfaeac5084;hpb=4739037f967ac3c818907e89cc88c7b97021d027 diff --git a/src/thread.cpp b/src/thread.cpp index 4fc0953e..1217e3ab 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) 2008-2014 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 @@ -33,66 +33,63 @@ 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; } } - - // 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 + // 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(); - thread_create(th->handle, start_routine, th); // Will go to sleep + th->nativeThread = std::thread(&ThreadBase::idle_loop, th); // Will go to sleep return th; } void delete_thread(ThreadBase* th) { + + th->mutex.lock(); th->exit = true; // Search must be already finished + th->mutex.unlock(); + th->notify_one(); - thread_join(th->handle); // Wait for thread termination + th->nativeThread.join(); // Wait for thread termination delete th; } } -// notify_one() wakes up the thread when there is some work to do +// ThreadBase::notify_one() wakes up the thread when there is some work to do void ThreadBase::notify_one() { - mutex.lock(); + std::unique_lock(this->mutex); sleepCondition.notify_one(); - mutex.unlock(); } -// wait_for() set the thread to sleep until condition 'b' turns true +// ThreadBase::wait_for() set the thread to sleep until 'condition' turns true -void ThreadBase::wait_for(volatile const bool& b) { +void ThreadBase::wait_for(volatile const bool& condition) { - mutex.lock(); - while (!b) sleepCondition.wait(mutex); - mutex.unlock(); + std::unique_lock lk(mutex); + sleepCondition.wait(lk, [&]{ return condition; }); } -// 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 c'tor makes some init but does not launch any execution thread that +// will be started only when c'tor returns. -Thread::Thread() /* : splitPoints() */ { // Value-initialization bug in MSVC +Thread::Thread() /* : splitPoints() */ { // Initialization of non POD broken in MSVC searching = false; - maxPly = splitPointsSize = 0; - activeSplitPoint = NULL; - activePosition = NULL; + 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 +// 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 { @@ -119,7 +116,7 @@ bool Thread::available_to(const Thread* master) const { // 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; + 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. @@ -127,19 +124,113 @@ bool Thread::available_to(const Thread* master) const { } -// 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. +// 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(); + + sp.allSlavesSearching = true; // Must be set under lock protection + ++splitPointsSize; + activeSplitPoint = &sp; + activePosition = nullptr; + + Thread* slave; + + while ( sp.slavesMask.count() < MAX_SLAVES_PER_SPLITPOINT + && (slave = Threads.available_slave(this)) != nullptr) + { + 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 '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(); + + 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 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(); +} + + +// 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 TimerThread::idle_loop() { while (!exit) { - mutex.lock(); + std::unique_lock lk(mutex); if (!exit) - sleepCondition.wait_for(mutex, run ? Resolution : INT_MAX); + sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX)); - mutex.unlock(); + lk.unlock(); if (run) check_time(); @@ -152,38 +243,38 @@ void TimerThread::idle_loop() { void MainThread::idle_loop() { - while (true) + while (!exit) { - mutex.lock(); + std::unique_lock lk(mutex); thinking = false; while (!thinking && !exit) { Threads.sleepCondition.notify_one(); // Wake up the UI thread if needed - sleepCondition.wait(mutex); + sleepCondition.wait(lk); } - mutex.unlock(); - - if (exit) - return; + lk.unlock(); - searching = true; + if (!exit) + { + searching = true; - Search::think(); + Search::think(); - assert(searching); + assert(searching); - searching = false; + searching = false; + } } } -// 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. +// 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. void ThreadPool::init() { @@ -193,23 +284,23 @@ void ThreadPool::init() { } -// 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. +// 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 - for (iterator it = begin(); it != end(); ++it) - delete_thread(*it); + for (Thread* th : *this) + delete_thread(th); } -// 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. +// 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::read_uci_options() { @@ -233,126 +324,33 @@ void ThreadPool::read_uci_options() { } -// available_slave() tries to find an idle thread which is available as a slave -// for the thread 'master'. +// ThreadPool::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; + for (Thread* th : *this) + if (th->available_to(master)) + return th; - return NULL; + return nullptr; } -// 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. - -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(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 = 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.mutex.lock(); - sp.mutex.lock(); - - sp.allSlavesSearching = true; // Must be set under lock protection - ++splitPointsSize; - activeSplitPoint = &sp; - activePosition = NULL; - - 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 'searching' flag is set. - // The thread will return from the idle loop when all slaves have finished - // their work at this split point. - 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(); - - 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(); -} - -// wait_for_think_finished() waits for main thread to go to sleep then returns +// ThreadPool::wait_for_think_finished() waits for main thread to finish the search void ThreadPool::wait_for_think_finished() { - MainThread* t = main(); - t->mutex.lock(); - while (t->thinking) sleepCondition.wait(t->mutex); - t->mutex.unlock(); + std::unique_lock lk(main()->mutex); + sleepCondition.wait(lk, [&]{ return !main()->thinking; }); } -// 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) { +// 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) { wait_for_think_finished(); SearchTime = Time::now(); // As early as possible @@ -365,14 +363,14 @@ void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, S Limits = limits; if (states.get()) // If we don't set a new position, preserve current state { - SetupStates = states; // Ownership transfer here + SetupStates = std::move(states); // Ownership transfer here assert(!states.get()); } - for (MoveList it(pos); *it; ++it) + for (const auto& m : MoveList(pos)) if ( limits.searchmoves.empty() - || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), *it)) - RootMoves.push_back(RootMove(*it)); + || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m)) + RootMoves.push_back(RootMove(m)); main()->thinking = true; main()->notify_one(); // Starts main thread