X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=b04b8adbb1719e586e818e366b2077fdc01053ab;hp=31c5961d2fc104f7af2d12afec86f7ee73f56ea2;hb=bf706c4a4ff06e91edf8c24685d24130146f88d4;hpb=c47a74ec62c8c4fdacf6d16eef068b62634122f1 diff --git a/src/thread.cpp b/src/thread.cpp index 31c5961d..b04b8adb 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -17,6 +17,7 @@ along with this program. If not, see . */ +#include // For std::count #include #include @@ -27,121 +28,126 @@ using namespace Search; -ThreadsManager Threads; // Global object +ThreadPool Threads; // Global object namespace { extern "C" { // start_routine() is the C function which is called when a new thread - // is launched. It simply calls idle_loop() of the supplied thread. The first - // and last thread are special. First one is the main search thread while the - // last one mimics a timer, they run in main_loop() and timer_loop(). + // is launched. It is a wrapper to the virtual function idle_loop(). -#if defined(_WIN32) || defined(_WIN64) - DWORD WINAPI start_routine(LPVOID thread) { -#else - void* start_routine(void* thread) { -#endif + long start_routine(Thread* th) { th->idle_loop(); return 0; } + +} } - Thread* th = (Thread*)thread; - if (th->threadID == 0) - th->main_loop(); +// Thread c'tor starts a newly-created thread of execution that will call +// the the virtual function idle_loop(), going immediately to sleep. - else if (th->threadID == MAX_THREADS) - th->timer_loop(); +Thread::Thread() : splitPoints() { - else - th->idle_loop(NULL); + searching = exit = false; + maxPly = splitPointsSize = 0; + activeSplitPoint = NULL; + idx = Threads.size(); - return 0; + if (!thread_create(handle, start_routine, this)) + { + std::cerr << "Failed to create thread number " << idx << std::endl; + ::exit(EXIT_FAILURE); } +} -} } +// Thread d'tor waits for thread termination before to return -// Thread::timer_loop() is where the timer thread waits maxPly milliseconds and -// then calls do_timer_event(). If maxPly is 0 thread sleeps until is woken up. +Thread::~Thread() { + + exit = true; // Search must be already finished + notify_one(); + thread_join(handle); // Wait for thread termination +} + + +// 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 Thread::timer_loop() { +void TimerThread::idle_loop() { - while (!do_exit) + while (!exit) { - lock_grab(sleepLock); - timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX); - lock_release(sleepLock); - check_time(); + mutex.lock(); + + if (!exit) + sleepCondition.wait_for(mutex, msec ? msec : INT_MAX); + + mutex.unlock(); + + if (msec) + check_time(); } } -// Thread::main_loop() is where the main thread is parked waiting to be started +// 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 Thread::main_loop() { +void MainThread::idle_loop() { while (true) { - lock_grab(sleepLock); + mutex.lock(); - do_sleep = true; // Always return to sleep after a search - is_searching = false; + thinking = false; - while (do_sleep && !do_exit) + while (!thinking && !exit) { - cond_signal(Threads.sleepCond); // Wake up UI thread if needed - cond_wait(sleepCond, sleepLock); + Threads.sleepCondition.notify_one(); // Wake up UI thread if needed + sleepCondition.wait(mutex); } - lock_release(sleepLock); + mutex.unlock(); - if (do_exit) + if (exit) return; - is_searching = true; + searching = true; Search::think(); + + assert(searching); + + searching = false; } } -// Thread::wake_up() wakes up the thread, normally at the beginning of the search -// or, if "sleeping threads" is used, when there is some work to do. +// Thread::notify_one() wakes up the thread when there is some search to do -void Thread::wake_up() { +void Thread::notify_one() { - lock_grab(sleepLock); - cond_signal(sleepCond); - lock_release(sleepLock); + mutex.lock(); + sleepCondition.notify_one(); + mutex.unlock(); } -// 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() { +// Thread::wait_for() set the thread to sleep until condition 'b' turns true - Signals.stopOnPonderhit = true; +void Thread::wait_for(volatile const bool& b) { - lock_grab(sleepLock); - - while (!Signals.stop) - cond_wait(sleepCond, sleepLock); - - lock_release(sleepLock); + mutex.lock(); + while (!b) sleepCondition.wait(mutex); + mutex.unlock(); } -// cutoff_occurred() checks whether a beta cutoff has occurred in the current -// active split point, or in some ancestor of the split point. +// 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 { - for (SplitPoint* sp = curSplitPoint; sp; sp = sp->parent) + for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint) if (sp->cutoff) return true; @@ -149,138 +155,84 @@ bool Thread::cutoff_occurred() const { } -// is_available_to() checks whether the thread is available to help the thread with -// threadID "master" at a split point. An obvious requirement is that thread must be -// idle. With more than two threads, this is not by itself sufficient: If the thread -// is the master of some active split point, it is only available as a slave to the -// threads which are busy searching the split point at the top of "slave"'s split -// point stack (the "helpful master concept" in YBWC terminology). +// 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). -bool Thread::is_available_to(int master) const { +bool Thread::is_available_to(Thread* master) const { - if (is_searching) + 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 spCnt = splitPointsCnt; + int size = splitPointsSize; - // No active split points means that the thread is available as a slave for any + // 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 !spCnt || (splitPoints[spCnt - 1].slavesMask & (1ULL << master)); + return !size || (splitPoints[size - 1].slavesMask & (1ULL << master->idx)); } -// read_uci_options() updates internal threads parameters from the corresponding -// UCI options. It is called before to start a new search. +// 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 ThreadsManager::read_uci_options() { +void ThreadPool::init() { - maxThreadsPerSplitPoint = Options["Max Threads per Split Point"]; - minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; - useSleepingThreads = Options["Use Sleeping Threads"]; + sleepWhileIdle = true; + timer = new TimerThread(); + push_back(new MainThread()); + read_uci_options(); } -// set_size() changes the number of active threads and raises do_sleep flag for -// all the unused threads that will go immediately to sleep. - -void ThreadsManager::set_size(int cnt) { - - assert(cnt > 0 && cnt < MAX_THREADS); +// exit() cleanly terminates the threads before the program exits - activeThreads = cnt; - - for (int i = 0; i < MAX_THREADS; i++) - if (i < activeThreads) - { - // Dynamically allocate pawn and material hash tables according to the - // number of active threads. This avoids preallocating memory for all - // possible threads if only few are used. - threads[i].pawnTable.init(); - threads[i].materialTable.init(); - threads[i].maxPly = 0; +void ThreadPool::exit() { - threads[i].do_sleep = false; + delete timer; // As first because check_time() accesses threads data - if (!useSleepingThreads) - threads[i].wake_up(); - } - else - threads[i].do_sleep = true; + for (iterator it = begin(); it != end(); ++it) + delete *it; } -// init() is called during startup. Initializes locks and condition variables -// and launches all threads sending them immediately to sleep. - -void ThreadsManager::init() { - - read_uci_options(); - - cond_init(sleepCond); - lock_init(splitLock); - - // Allocate main thread tables to call evaluate() also when not searching - threads[0].pawnTable.init(); - threads[0].materialTable.init(); - - // Create and launch all the threads, threads will go immediately to sleep - for (int i = 0; i <= MAX_THREADS; i++) - { - threads[i].is_searching = false; - threads[i].do_sleep = (i != 0); // Avoid a race with start_thinking() - threads[i].threadID = i; - - lock_init(threads[i].sleepLock); - cond_init(threads[i].sleepCond); +// 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. - for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++) - lock_init(threads[i].splitPoints[j].lock); - - if (!thread_create(threads[i].handle, start_routine, threads[i])) - { - std::cerr << "Failed to create thread number " << i << std::endl; - ::exit(EXIT_FAILURE); - } - } -} +void ThreadPool::read_uci_options() { + maxThreadsPerSplitPoint = Options["Max Threads per Split Point"]; + minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; + size_t requested = Options["Threads"]; -// exit() is called to cleanly terminate the threads when the program finishes + assert(requested > 0); -void ThreadsManager::exit() { + while (size() < requested) + push_back(new Thread()); - for (int i = 0; i <= MAX_THREADS; i++) + while (size() > requested) { - assert(threads[i].do_sleep); - - threads[i].do_exit = true; // Search must be already finished - threads[i].wake_up(); - - thread_join(threads[i].handle); // Wait for thread termination - - lock_destroy(threads[i].sleepLock); - cond_destroy(threads[i].sleepCond); - - for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++) - lock_destroy(threads[i].splitPoints[j].lock); + delete back(); + pop_back(); } - - lock_destroy(splitLock); - cond_destroy(sleepCond); } -// available_slave_exists() tries to find an idle thread which is available as -// a slave for the thread with threadID 'master'. +// slave_available() tries to find an idle thread which is available as a slave +// for the thread 'master'. -bool ThreadsManager::available_slave_exists(int master) const { +bool ThreadPool::slave_available(Thread* master) const { - assert(master >= 0 && master < activeThreads); - - for (int i = 0; i < activeThreads; i++) - if (threads[i].is_available_to(master)) + for (const_iterator it = begin(); it != end(); ++it) + if ((*it)->is_available_to(master)) return true; return false; @@ -289,188 +241,140 @@ bool ThreadsManager::available_slave_exists(int master) const { // 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. +// (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 -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); - assert(pos.thread() >= 0 && pos.thread() < activeThreads); - assert(activeThreads > 1); +void Thread::split(Position& pos, Stack* ss, Value alpha, Value beta, Value* bestValue, + Move* bestMove, Depth depth, Move threatMove, int moveCount, + MovePicker* movePicker, int nodeType) { - int master = pos.thread(); - Thread& masterThread = threads[master]; - - if (masterThread.splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD) - return bestValue; + assert(pos.pos_is_ok()); + assert(*bestValue <= alpha && alpha < beta && beta <= VALUE_INFINITE); + assert(*bestValue > -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 = &masterThread.splitPoints[masterThread.splitPointsCnt++]; - - sp->parent = masterThread.curSplitPoint; - sp->master = master; - sp->cutoff = false; - sp->slavesMask = 1ULL << master; - 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(masterThread.is_searching); - - masterThread.curSplitPoint = sp; - int slavesCnt = 0; + SplitPoint& sp = splitPoints[splitPointsSize]; + + sp.masterThread = this; + sp.parentSplitPoint = activeSplitPoint; + sp.slavesMask = 1ULL << idx; + sp.depth = depth; + sp.bestValue = *bestValue; + sp.bestMove = *bestMove; + sp.threatMove = threatMove; + sp.alpha = alpha; + sp.beta = beta; + sp.nodeType = nodeType; + 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 - // is_searching flag. This must be done under lock protection to avoid concurrent + // '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); + Threads.mutex.lock(); + sp.mutex.lock(); - for (int i = 0; i < activeThreads && !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() + splitPointsSize++; + activeSplitPoint = &sp; - if (useSleepingThreads) - threads[i].wake_up(); + size_t slavesCnt = 1; // This thread is always included + + for (ThreadPool::iterator it = Threads.begin(); it != Threads.end() && !Fake; ++it) + { + Thread* slave = *it; - if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included - break; + if (slave->is_available_to(this) && ++slavesCnt <= Threads.maxThreadsPerSplitPoint) + { + sp.slavesMask |= 1ULL << slave->idx; + slave->activeSplitPoint = &sp; + slave->searching = true; // Slave leaves idle_loop() + slave->notify_one(); // Could be sleeping } + } - lock_release(splitLock); - lock_release(sp->lock); + sp.mutex.unlock(); + Threads.mutex.unlock(); // 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 + // 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 || Fake) + if (slavesCnt > 1 || Fake) { - masterThread.idle_loop(sp); + Thread::idle_loop(); // Force a call to base class idle_loop() // 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(!masterThread.is_searching); + assert(!searching); } // We have returned from the idle loop, which means that all threads are - // finished. Note that setting is_searching and decreasing splitPointsCnt is + // finished. Note that setting 'searching' and decreasing splitPointsSize 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); - - masterThread.is_searching = true; - masterThread.splitPointsCnt--; - masterThread.curSplitPoint = sp->parent; - pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); - *bestMove = sp->bestMove; - - lock_release(splitLock); - lock_release(sp->lock); - - return sp->bestValue; + Threads.mutex.lock(); + sp.mutex.lock(); + + searching = true; + splitPointsSize--; + activeSplitPoint = sp.parentSplitPoint; + pos.set_nodes_searched(pos.nodes_searched() + sp.nodes); + *bestMove = sp.bestMove; + *bestValue = sp.bestValue; + + sp.mutex.unlock(); + Threads.mutex.unlock(); } // 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); - +template void Thread::split(Position&, Stack*, Value, Value, Value*, Move*, Depth, Move, int, MovePicker*, int); +template void Thread::split< true>(Position&, Stack*, Value, Value, Value*, Move*, Depth, Move, int, MovePicker*, int); -// ThreadsManager::set_timer() is used to set the timer to trigger after msec -// milliseconds. If msec is 0 then timer is stopped. -void ThreadsManager::set_timer(int msec) { +// wait_for_think_finished() waits for main thread to go to sleep then returns - Thread& timer = threads[MAX_THREADS]; +void ThreadPool::wait_for_think_finished() { - lock_grab(timer.sleepLock); - timer.maxPly = msec; - cond_signal(timer.sleepCond); // Wake up and restart the timer - lock_release(timer.sleepLock); + MainThread* t = main_thread(); + t->mutex.lock(); + while (t->thinking) sleepCondition.wait(t->mutex); + t->mutex.unlock(); } -// ThreadsManager::start_thinking() is used by UI thread to wake up the main -// thread parked in main_loop() and starting a new search. If asyncMode is true -// then function returns immediately, otherwise caller is blocked waiting for -// the search to finish. +// start_thinking() wakes up the main thread sleeping in MainThread::idle_loop() +// so to start a new search, then returns immediately. -void ThreadsManager::start_thinking(const Position& pos, const LimitsType& limits, - const std::set& searchMoves, bool async) { - Thread& main = threads[0]; +void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, + const std::vector& searchMoves, StateStackPtr& states) { + wait_for_think_finished(); - lock_grab(main.sleepLock); + SearchTime = Time::now(); // As early as possible - // Wait main thread has finished before to launch a new search - while (!main.do_sleep) - cond_wait(sleepCond, main.sleepLock); + Signals.stopOnPonderhit = Signals.firstRootMove = false; + Signals.stop = Signals.failedLowAtRoot = false; - // Copy input arguments to initialize the search - RootPosition.copy(pos, 0); + RootPos = pos; Limits = limits; + SetupStates = states; // Ownership transfer here RootMoves.clear(); - // Populate RootMoves with all the legal moves (default) or, if a searchMoves - // set is given, with the subset of legal moves to search. - for (MoveList ml(pos); !ml.end(); ++ml) - if (searchMoves.empty() || searchMoves.count(ml.move())) + for (MoveList ml(pos); !ml.end(); ++ml) + if ( searchMoves.empty() + || std::count(searchMoves.begin(), searchMoves.end(), ml.move())) RootMoves.push_back(RootMove(ml.move())); - // Reset signals before to start the new search - Signals.stopOnPonderhit = Signals.firstRootMove = false; - Signals.stop = Signals.failedLowAtRoot = false; - - main.do_sleep = false; - cond_signal(main.sleepCond); // Wake up main thread and start searching - - if (!async) - while (!main.do_sleep) - cond_wait(sleepCond, main.sleepLock); - - lock_release(main.sleepLock); -} - - -// ThreadsManager::stop_thinking() is used by UI thread to raise a stop request -// and to wait for the main thread finishing the search. Needed to wait exiting -// and terminate the threads after a 'quit' command. - -void ThreadsManager::stop_thinking() { - - Thread& main = threads[0]; - - Search::Signals.stop = true; - - lock_grab(main.sleepLock); - - cond_signal(main.sleepCond); // In case is waiting for stop or ponderhit - - while (!main.do_sleep) - cond_wait(sleepCond, main.sleepLock); - - lock_release(main.sleepLock); + main_thread()->thinking = true; + main_thread()->notify_one(); // Starts main thread }