X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=2c40ef6f8c14ecc7ccc218a01d06dada354d8874;hp=f5bd50ea5cfa28ec157d3082ba440058c51189e5;hb=3d937e1e901c59c04cd5d602c05f72892222ded8;hpb=4a71c862702c05b3c56e902f4675fdf68041710b diff --git a/src/thread.cpp b/src/thread.cpp index f5bd50ea..2c40ef6f 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) 2008-2010 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2008-2012 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 @@ -17,38 +17,45 @@ along with this program. If not, see . */ +#include #include +#include "movegen.h" +#include "search.h" #include "thread.h" #include "ucioption.h" -ThreadsManager Threads; // Global object definition +using namespace Search; + +ThreadsManager 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() with the supplied threadID. - // There are two versions of this function; one for POSIX threads and - // one for Windows threads. + // 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(). #if defined(_MSC_VER) + DWORD WINAPI start_routine(LPVOID thread) { +#else + void* start_routine(void* thread) { +#endif - DWORD WINAPI start_routine(LPVOID threadID) { + Thread* th = (Thread*)thread; - Threads.idle_loop(*(int*)threadID, NULL); - return 0; - } + if (th->threadID == 0) + th->main_loop(); -#else + else if (th->threadID == MAX_THREADS) + th->timer_loop(); - void* start_routine(void* threadID) { + else + th->idle_loop(NULL); - Threads.idle_loop(*(int*)threadID, NULL); - return NULL; + return 0; } -#endif - } } @@ -57,21 +64,21 @@ namespace { extern "C" { void Thread::wake_up() { - lock_grab(&sleepLock); - cond_signal(&sleepCond); - lock_release(&sleepLock); + lock_grab(sleepLock); + cond_signal(sleepCond); + lock_release(sleepLock); } -// cutoff_occurred() checks whether a beta cutoff has occurred in -// the thread's currently active split point, or in some ancestor of -// the current split point. +// 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 = splitPoint; sp; sp = sp->parent) if (sp->is_betaCutoff) return true; + return false; } @@ -85,7 +92,7 @@ bool Thread::cutoff_occurred() const { bool Thread::is_available_to(int master) const { - if (state != AVAILABLE) + if (is_searching) return false; // Make a local copy to be sure doesn't become zero under our feet while @@ -102,147 +109,152 @@ bool Thread::is_available_to(int master) const { } -// read_uci_options() updates number of active threads and other internal -// parameters according to the UCI options values. It is called before -// to start a new search. +// read_uci_options() updates number of active threads and other parameters +// according to the UCI options values. It is called before to start a new search. void ThreadsManager::read_uci_options() { - maxThreadsPerSplitPoint = Options["Maximum Number of Threads per Split Point"].value(); - minimumSplitDepth = Options["Minimum Split Depth"].value() * ONE_PLY; - useSleepingThreads = Options["Use Sleeping Threads"].value(); - activeThreads = Options["Threads"].value(); + maxThreadsPerSplitPoint = Options["Max Threads per Split Point"]; + minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; + useSleepingThreads = Options["Use Sleeping Threads"]; + + set_size(Options["Threads"]); } -// init() is called during startup. Initializes locks and condition variables -// and launches all threads sending them immediately to sleep. +// 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::init() { +void ThreadsManager::set_size(int cnt) { - int threadID[MAX_THREADS]; + assert(cnt > 0 && cnt <= MAX_THREADS); - // This flag is needed to properly end the threads when program exits - allThreadsShouldExit = false; + activeThreads = cnt; - // Threads will sent to sleep as soon as created, only main thread is kept alive - activeThreads = 1; - threads[0].state = Thread::SEARCHING; + for (int i = 1; i < MAX_THREADS; i++) // Ignore main thread + 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].do_sleep = false; + } + else + threads[i].do_sleep = true; +} + + +// init() is called during startup. Initializes locks and condition variables +// and launches all threads sending them immediately to sleep. - // Allocate pawn and material hash tables for main thread - init_hash_tables(); +void ThreadsManager::init() { - lock_init(&threadsLock); + // Initialize sleep condition and lock used by thread manager + cond_init(sleepCond); + lock_init(threadsLock); - // Initialize thread and split point locks - for (int i = 0; i < MAX_THREADS; i++) + // Initialize thread's sleep conditions and split point locks + for (int i = 0; i <= MAX_THREADS; i++) { - lock_init(&threads[i].sleepLock); - cond_init(&threads[i].sleepCond); + lock_init(threads[i].sleepLock); + cond_init(threads[i].sleepCond); for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_init(&(threads[i].splitPoints[j].lock)); + lock_init(threads[i].splitPoints[j].lock); } - // Create and startup all the threads but the main that is already running - for (int i = 1; i < MAX_THREADS; i++) + // 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].state = Thread::INITIALIZING; - threadID[i] = i; + threads[i].is_searching = false; + threads[i].do_sleep = (i != 0); // Avoid a race with start_thinking() + threads[i].threadID = i; + + bool ok = thread_create(threads[i].handle, start_routine, threads[i]); -#if defined(_MSC_VER) - bool ok = (CreateThread(NULL, 0, start_routine, (LPVOID)&threadID[i], 0, NULL) != NULL); -#else - pthread_t pthreadID; - bool ok = (pthread_create(&pthreadID, NULL, start_routine, (void*)&threadID[i]) == 0); - pthread_detach(pthreadID); -#endif if (!ok) { - std::cout << "Failed to create thread number " << i << std::endl; + std::cerr << "Failed to create thread number " << i << std::endl; ::exit(EXIT_FAILURE); } - - // Wait until the thread has finished launching and is gone to sleep - while (threads[i].state == Thread::INITIALIZING) {} } } -// exit() is called to cleanly exit the threads when the program finishes +// exit() is called to cleanly terminate the threads when the program finishes void ThreadsManager::exit() { - // Force the woken up threads to exit idle_loop() and hence terminate - allThreadsShouldExit = true; - - for (int i = 0; i < MAX_THREADS; i++) + for (int i = 0; i <= MAX_THREADS; i++) { - // Wake up all the threads and waits for termination - if (i != 0) - { - threads[i].wake_up(); - while (threads[i].state != Thread::TERMINATED) {} - } + threads[i].do_terminate = true; // Search must be already finished + threads[i].wake_up(); - // Now we can safely destroy the locks and wait conditions - lock_destroy(&threads[i].sleepLock); - cond_destroy(&threads[i].sleepCond); + thread_join(threads[i].handle); // Wait for thread termination + + // Now we can safely destroy associated locks and wait conditions + lock_destroy(threads[i].sleepLock); + cond_destroy(threads[i].sleepCond); for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_destroy(&(threads[i].splitPoints[j].lock)); + lock_destroy(threads[i].splitPoints[j].lock); } - lock_destroy(&threadsLock); -} - - -// init_hash_tables() dynamically allocates 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 as, for instance, -// on mobile devices where memory is scarce and allocating for MAX_THREADS -// threads could even result in a crash. - -void ThreadsManager::init_hash_tables() { - - for (int i = 0; i < activeThreads; i++) - { - threads[i].pawnTable.init(); - threads[i].materialTable.init(); - } + lock_destroy(threadsLock); + cond_destroy(sleepCond); } // available_slave_exists() tries to find an idle thread which is available as -// a slave for the thread with threadID "master". +// a slave for the thread with threadID 'master'. bool ThreadsManager::available_slave_exists(int master) const { assert(master >= 0 && master < activeThreads); for (int i = 0; i < activeThreads; i++) - if (i != master && threads[i].is_available_to(master)) + if (threads[i].is_available_to(master)) return true; return false; } +// split_point_finished() checks if all the slave threads of a given split +// point have finished searching. + +bool ThreadsManager::split_point_finished(SplitPoint* sp) const { + + for (int i = 0; i < activeThreads; i++) + if (sp->is_slave[i]) + return false; + + return true; +} + + // 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 we tell our helper threads 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. +// 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. template -Value ThreadsManager::split(Position& pos, SearchStack* ss, Value alpha, Value beta, +Value ThreadsManager::split(Position& pos, Stack* ss, Value alpha, Value beta, Value bestValue, Depth depth, Move threatMove, int moveCount, MovePicker* mp, int nodeType) { - assert(pos.is_ok()); - assert(bestValue >= -VALUE_INFINITE); + assert(pos.pos_is_ok()); + assert(bestValue > -VALUE_INFINITE); assert(bestValue <= alpha); assert(alpha < beta); assert(beta <= VALUE_INFINITE); @@ -257,83 +269,232 @@ Value ThreadsManager::split(Position& pos, SearchStack* ss, Value alpha, Value b if (masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS) return bestValue; - // Pick the next available split point object from the split point stack - SplitPoint& splitPoint = masterThread.splitPoints[masterThread.activeSplitPoints]; - - // Initialize the split point object - splitPoint.parent = masterThread.splitPoint; - splitPoint.master = master; - splitPoint.is_betaCutoff = false; - splitPoint.depth = depth; - splitPoint.threatMove = threatMove; - splitPoint.alpha = alpha; - splitPoint.beta = beta; - splitPoint.nodeType = nodeType; - splitPoint.bestValue = bestValue; - splitPoint.mp = mp; - splitPoint.moveCount = moveCount; - splitPoint.pos = &pos; - splitPoint.nodes = 0; - splitPoint.ss = ss; + // Pick the next available split point from the split point stack + SplitPoint* sp = &masterThread.splitPoints[masterThread.activeSplitPoints]; + + // Initialize the split point + sp->parent = masterThread.splitPoint; + sp->master = master; + sp->is_betaCutoff = false; + sp->depth = depth; + 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; + for (i = 0; i < activeThreads; i++) - splitPoint.is_slave[i] = false; + sp->is_slave[i] = false; // If we are here it means we are not available - assert(masterThread.state == Thread::SEARCHING); + assert(masterThread.is_searching); int workersCnt = 1; // At least the master is included // Try to allocate available threads and ask them to start searching setting - // the state to Thread::WORKISWAITING, this must be done under lock protection - // to avoid concurrent allocation of the same slave by another master. - lock_grab(&threadsLock); + // is_searching flag. This must be done under lock protection to avoid concurrent + // allocation of the same slave by another master. + lock_grab(threadsLock); for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++) - if (i != master && threads[i].is_available_to(master)) + if (threads[i].is_available_to(master)) { workersCnt++; - splitPoint.is_slave[i] = true; - threads[i].splitPoint = &splitPoint; + sp->is_slave[i] = true; + threads[i].splitPoint = sp; // This makes the slave to exit from idle_loop() - threads[i].state = Thread::WORKISWAITING; + threads[i].is_searching = true; if (useSleepingThreads) threads[i].wake_up(); } - lock_release(&threadsLock); + lock_release(threadsLock); // We failed to allocate even one slave, return if (!Fake && workersCnt == 1) return bestValue; - masterThread.splitPoint = &splitPoint; + masterThread.splitPoint = sp; masterThread.activeSplitPoints++; - masterThread.state = Thread::WORKISWAITING; - // Everything is set up. The master thread enters the idle loop, from - // which it will instantly launch a search, because its state is - // Thread::WORKISWAITING. We send the split point as a second parameter to - // the idle loop, which means that the main thread will return from the idle - // loop when all threads have finished their work at this split point. - idle_loop(master, &splitPoint); + // 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 + // their work at this split point. + masterThread.idle_loop(sp); + + // 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); // We have returned from the idle loop, which means that all threads are // finished. Note that changing state and decreasing activeSplitPoints is done // under lock protection to avoid a race with Thread::is_available_to(). - lock_grab(&threadsLock); + lock_grab(threadsLock); - masterThread.state = Thread::SEARCHING; + masterThread.is_searching = true; masterThread.activeSplitPoints--; - masterThread.splitPoint = splitPoint.parent; - lock_release(&threadsLock); + lock_release(threadsLock); + + masterThread.splitPoint = sp->parent; + pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); - pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes); - return splitPoint.bestValue; + return sp->bestValue; } // Explicit template instantiations -template Value ThreadsManager::split(Position&, SearchStack*, Value, Value, Value, Depth, Move, int, MovePicker*, int); -template Value ThreadsManager::split(Position&, SearchStack*, Value, Value, Value, Depth, Move, int, MovePicker*, int); +template Value ThreadsManager::split(Position&, Stack*, Value, Value, Value, Depth, Move, int, MovePicker*, int); +template Value ThreadsManager::split(Position&, Stack*, Value, Value, Value, Depth, Move, int, MovePicker*, int); + + +// 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. +extern void check_time(); + +void Thread::timer_loop() { + + while (!do_terminate) + { + lock_grab(sleepLock); + timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX); + lock_release(sleepLock); + check_time(); + } +} + + +// 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) { + + Thread& timer = threads[MAX_THREADS]; + + lock_grab(timer.sleepLock); + timer.maxPly = msec; + cond_signal(timer.sleepCond); // Wake up and restart the timer + lock_release(timer.sleepLock); +} + + +// Thread::main_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() { + + while (true) + { + lock_grab(sleepLock); + + do_sleep = true; // Always return to sleep after a search + is_searching = false; + + while (do_sleep && !do_terminate) + { + cond_signal(Threads.sleepCond); // Wake up UI thread if needed + cond_wait(sleepCond, sleepLock); + } + + is_searching = true; + + lock_release(sleepLock); + + if (do_terminate) + return; + + Search::think(); + } +} + + +// 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. + +void ThreadsManager::start_thinking(const Position& pos, const LimitsType& limits, + const std::set& searchMoves, bool async) { + Thread& main = threads[0]; + + lock_grab(main.sleepLock); + + // Wait main thread has finished before to launch a new search + while (!main.do_sleep) + cond_wait(sleepCond, main.sleepLock); + + // Copy input arguments to initialize the search + RootPosition.copy(pos, 0); + Limits = limits; + 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())) + 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); +} + + +// ThreadsManager::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 ThreadsManager::wait_for_stop_or_ponderhit() { + + Signals.stopOnPonderhit = true; + + Thread& main = threads[0]; + + lock_grab(main.sleepLock); + + while (!Signals.stop) + cond_wait(main.sleepCond, main.sleepLock); + + lock_release(main.sleepLock); +}