X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=c73f55660330545d300857b2874287fd54b21563;hp=7d85db8618ca982059b96466d963e29d3b1b5ba7;hb=2ca142a5b4ca200c56cb99495ec51a804983d07d;hpb=c65d67feb53895a34423cc2043edc69741f2ba92 diff --git a/src/thread.cpp b/src/thread.cpp index 7d85db86..c73f5566 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) 2008-2013 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 @@ -23,34 +23,34 @@ #include "movegen.h" #include "search.h" #include "thread.h" -#include "ucioption.h" +#include "uci.h" using namespace Search; ThreadPool Threads; // Global object -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; } } +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 object shall 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 - return th; + 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(); - thread_join(th->handle); // Wait for thread termination + th->join(); // Wait for thread termination delete th; } @@ -61,86 +61,31 @@ namespace { void ThreadBase::notify_one() { - mutex.lock(); + std::unique_lock lk(mutex); sleepCondition.notify_one(); - mutex.unlock(); } -// ThreadBase::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 but does not launch any thread of execution that -// instead will be started only upon 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; - idx = Threads.size(); -} - - -// TimerThread::idle_loop() is where the timer thread waits msec milliseconds -// and then calls check_time(). If msec is 0 thread sleeps until is woken up. -extern void check_time(); - -void TimerThread::idle_loop() { - - while (!exit) - { - mutex.lock(); - - if (!exit) - sleepCondition.wait_for(mutex, msec ? msec : INT_MAX); - - mutex.unlock(); - - if (msec) - check_time(); - } -} - - -// MainThread::idle_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 MainThread::idle_loop() { - - while (true) - { - mutex.lock(); - - thinking = false; - - while (!thinking && !exit) - { - Threads.sleepCondition.notify_one(); // Wake up UI thread if needed - sleepCondition.wait(mutex); - } - - mutex.unlock(); - - if (exit) - return; - - searching = true; - - Search::think(); - - assert(searching); - - searching = false; - } + maxPly = 0; + splitPointsSize = 0; + activeSplitPoint = nullptr; + activePosition = nullptr; + idx = Threads.size(); // Starts from 0 } @@ -157,128 +102,56 @@ bool Thread::cutoff_occurred() const { } -// 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 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). +// 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(const Thread* master) const { +bool Thread::can_join(const SplitPoint* sp) const { 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 size = splitPointsSize; + // 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 & (1ULL << master->idx)); -} - - -// 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 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 ThreadPool::init() { - - sleepWhileIdle = true; - timer = new_thread(); - push_back(new_thread()); - read_uci_options(); + return !size || splitPoints[size - 1].slavesMask.test(sp->master->idx); } -// 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 in advance all possible -// threads, with included pawns and material tables, if only few are used. - -void ThreadPool::read_uci_options() { - - maxThreadsPerSplitPoint = Options["Max Threads per Split Point"]; - minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; - size_t requested = Options["Threads"]; - - assert(requested > 0); - - // Value 0 has a special meaning: We determine the optimal minimum split depth - // automatically. Anyhow the minimumSplitDepth should never be under 4 plies. - if (!minimumSplitDepth) - minimumSplitDepth = (requested < 8 ? 4 : 7) * ONE_PLY; - else - minimumSplitDepth = std::max(4 * ONE_PLY, minimumSplitDepth); - - while (size() < requested) - push_back(new_thread()); - - while (size() > requested) - { - delete_thread(back()); - pop_back(); - } -} - - -// slave_available() 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)->is_available_to(master)) - return *it; - - return NULL; -} - - -// split() does the actual work of distributing the work at a node between +// 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 -// told that they have been assigned work. This will cause them to instantly +// 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. -template -void Thread::split(Position& pos, const Stack* ss, Value alpha, Value beta, Value* bestValue, - Move* bestMove, Depth depth, Move threatMove, int moveCount, +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(*bestValue <= alpha && alpha < beta && beta <= VALUE_INFINITE); - assert(*bestValue > -VALUE_INFINITE); - assert(depth >= Threads.minimumSplitDepth); assert(searching); + assert(-VALUE_INFINITE < *bestValue && *bestValue <= alpha && alpha < beta && beta <= VALUE_INFINITE); + assert(depth >= Threads.minimumSplitDepth); assert(splitPointsSize < MAX_SPLITPOINTS_PER_THREAD); - // Pick the next available split point from the split point stack + // Pick and init the next available split point SplitPoint& sp = splitPoints[splitPointsSize]; - sp.masterThread = this; + sp.spinlock.acquire(); // No contention here until we don't increment splitPointsSize + + sp.master = this; sp.parentSplitPoint = activeSplitPoint; - sp.slavesMask = 1ULL << idx; + sp.slavesMask = 0, sp.slavesMask.set(idx); sp.depth = depth; sp.bestValue = *bestValue; sp.bestMove = *bestMove; - sp.threatMove = threatMove; sp.alpha = alpha; sp.beta = beta; sp.nodeType = nodeType; @@ -289,88 +162,199 @@ void Thread::split(Position& pos, const Stack* ss, Value alpha, Value beta, Valu sp.nodes = 0; sp.cutoff = false; sp.ss = ss; + sp.allSlavesSearching = true; // Must be set under lock protection - // 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++; + ++splitPointsSize; activeSplitPoint = &sp; - activePosition = NULL; + activePosition = nullptr; - size_t slavesCnt = 1; // This thread is always included + // Try to allocate available threads Thread* slave; - while ( (slave = Threads.available_slave(this)) != NULL - && ++slavesCnt <= Threads.maxThreadsPerSplitPoint && !Fake) + while ( sp.slavesMask.count() < MAX_SLAVES_PER_SPLITPOINT + && (slave = Threads.available_slave(&sp)) != nullptr) { - sp.slavesMask |= 1ULL << slave->idx; - slave->activeSplitPoint = &sp; - slave->searching = true; // Slave leaves idle_loop() - slave->notify_one(); // Could be sleeping + 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. - if (slavesCnt > 1 || Fake) - { - sp.mutex.unlock(); - Threads.mutex.unlock(); - - Thread::idle_loop(); // Force a call to base class idle_loop() + sp.spinlock.release(); - // 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(!searching); - assert(!activePosition); + Thread::idle_loop(); // Force a call to base class idle_loop() - // 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::is_available_to(). - Threads.mutex.lock(); - sp.mutex.lock(); - } + // 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); searching = true; - splitPointsSize--; + + // 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(). + sp.spinlock.acquire(); + + --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(); + sp.spinlock.release(); } -// Explicit template instantiations -template void Thread::split(Position&, const Stack*, Value, Value, Value*, Move*, Depth, Move, int, MovePicker*, int, bool); -template void Thread::split< true>(Position&, const Stack*, Value, Value, Value*, Move*, Depth, Move, int, MovePicker*, int, bool); +// 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. -// wait_for_think_finished() waits for main thread to go to sleep then returns +void TimerThread::idle_loop() { -void ThreadPool::wait_for_think_finished() { + while (!exit) + { + std::unique_lock lk(mutex); - MainThread* t = main(); - t->mutex.lock(); - while (t->thinking) sleepCondition.wait(t->mutex); - t->mutex.unlock(); + if (!exit) + sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX)); + + lk.unlock(); + + if (run) + check_time(); + } } -// start_thinking() wakes up the main thread sleeping in MainThread::idle_loop() -// so to start a new search, then returns immediately. +// 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 ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, - const std::vector& searchMoves, StateStackPtr& states) { - wait_for_think_finished(); +void MainThread::idle_loop() { + + while (!exit) + { + std::unique_lock lk(mutex); + + thinking = false; + + while (!thinking && !exit) + { + sleepCondition.notify_one(); // Wake up the UI thread if needed + sleepCondition.wait(lk); + } + + lk.unlock(); + + if (!exit) + { + searching = true; + + Search::think(); + + assert(searching); + + searching = false; + } + } +} + + +// MainThread::join() waits for main thread to finish the search + +void MainThread::join() { + + std::unique_lock lk(mutex); + sleepCondition.wait(lk, [&]{ return !thinking; }); +} + + +// 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. - SearchTime = Time::now(); // As early as possible +void ThreadPool::init() { + + timer = new_thread(); + push_back(new_thread()); + read_uci_options(); +} + + +// 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; + + for (Thread* th : *this) + delete_thread(th); + + clear(); // Get rid of stale pointers +} + + +// 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() { + + 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 = 5 * ONE_PLY ; + + while (size() < requested) + push_back(new_thread()); + + while (size() > requested) + { + delete_thread(back()); + pop_back(); + } +} + + +// ThreadPool::available_slave() tries to find an idle thread which is available +// to join SplitPoint 'sp'. + +Thread* ThreadPool::available_slave(const SplitPoint* sp) const { + + 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; @@ -380,15 +364,15 @@ void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, 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) - if ( searchMoves.empty() - || std::count(searchMoves.begin(), searchMoves.end(), *it)) - RootMoves.push_back(RootMove(*it)); + for (const auto& m : MoveList(pos)) + if ( limits.searchmoves.empty() + || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m)) + RootMoves.push_back(RootMove(m)); main()->thinking = true; - main()->notify_one(); // Starts main thread + main()->notify_one(); // Wake up main thread: 'thinking' must be already set }