X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=0d11a2d5330d1b3a720bd67faf56e58689e46130;hp=8a8e7199d43e00c42b3c5405ea95ace53af93d40;hb=553655eb073cdd59c726dd77fcf368d499029467;hpb=51e8efdab5f62ebc23e4f7adaea96f619cbca194 diff --git a/src/thread.cpp b/src/thread.cpp index 8a8e7199..0d11a2d5 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -36,13 +36,7 @@ namespace { extern "C" { // 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(_WIN32) || defined(_WIN64) - DWORD WINAPI start_routine(LPVOID thread) { -#else - void* start_routine(void* thread) { -#endif - - Thread* th = (Thread*)thread; + long start_routine(Thread* th) { if (th->threadID == 0) th->main_loop(); @@ -59,8 +53,93 @@ namespace { extern "C" { } } -// 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::Thread(int id) { + + threadID = id; + do_sleep = (id != 0); // Avoid a race with start_thinking() + is_searching = do_exit = false; + maxPly = splitPointsCnt = 0; + curSplitPoint = NULL; + + lock_init(sleepLock); + cond_init(sleepCond); + + for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++) + lock_init(splitPoints[j].lock); + + if (!thread_create(handle, start_routine, this)) + { + std::cerr << "Failed to create thread number " << id << std::endl; + ::exit(EXIT_FAILURE); + } +} + + +Thread::~Thread() { + + assert(do_sleep); + + do_exit = true; // Search must be already finished + wake_up(); + + thread_join(handle); // Wait for thread termination + + lock_destroy(sleepLock); + cond_destroy(sleepCond); + + for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++) + lock_destroy(splitPoints[j].lock); +} + + +// 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_exit) + { + lock_grab(sleepLock); + timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX); + lock_release(sleepLock); + check_time(); + } +} + + +// 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_exit) + { + cond_signal(Threads.sleepCond); // Wake up UI thread if needed + cond_wait(sleepCond, sleepLock); + } + + lock_release(sleepLock); + + if (do_exit) + return; + + is_searching = true; + + Search::think(); + } +} + + +// 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. void Thread::wake_up() { @@ -70,13 +149,33 @@ void Thread::wake_up() { } +// 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() { + + Signals.stopOnPonderhit = true; + + lock_grab(sleepLock); + + while (!Signals.stop) + cond_wait(sleepCond, sleepLock); + + lock_release(sleepLock); +} + + // 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) + for (SplitPoint* sp = curSplitPoint; sp; sp = sp->parent) + if (sp->cutoff) return true; return false; @@ -97,49 +196,52 @@ bool Thread::is_available_to(int master) const { // 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 sp_count = activeSplitPoints; + int spCnt = splitPointsCnt; // No active split points means that the thread is available as a slave for any // other thread otherwise apply the "helpful master" concept if possible. - return !sp_count || (splitPoints[sp_count - 1].slavesMask & (1ULL << master)); + return !spCnt || (splitPoints[spCnt - 1].slavesMask & (1ULL << master)); } -// 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. +// read_uci_options() updates internal threads parameters from the corresponding +// UCI options. It is called before to start a new search. void ThreadsManager::read_uci_options() { maxThreadsPerSplitPoint = Options["Max Threads per Split Point"]; minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY; useSleepingThreads = Options["Use Sleeping Threads"]; - - set_size(Options["Threads"]); + activeThreads = Options["Threads"]; + + // Dynamically allocate Thread object according to the number of + // active threads. This avoids preallocating memory for all possible + // threads if only few are used. + for (int i = 0; i < MAX_THREADS; i++) + if (i < activeThreads && !threads[i]) + threads[i] = new Thread(i); + else if (i >= activeThreads && threads[i]) + { + delete threads[i]; + threads[i] = NULL; + } } -// 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::wake_up() { -void ThreadsManager::set_size(int cnt) { - - assert(cnt > 0 && cnt <= MAX_THREADS); + for (int i = 0; i < activeThreads; i++) + { + threads[i]->do_sleep = false; + threads[i]->wake_up(); + } +} - activeThreads = cnt; - 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(); +void ThreadsManager::sleep() { - threads[i].do_sleep = false; - } - else - threads[i].do_sleep = true; + for (int i = 0; i < activeThreads; i++) + threads[i]->do_sleep = true; } @@ -148,35 +250,10 @@ void ThreadsManager::set_size(int cnt) { void ThreadsManager::init() { - cond_init(sleepCond); - lock_init(splitLock); - - for (int i = 0; i <= MAX_THREADS; i++) - { - 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); - } - - // 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; - - if (!thread_create(threads[i].handle, start_routine, threads[i])) - { - std::cerr << "Failed to create thread number " << i << std::endl; - ::exit(EXIT_FAILURE); - } - } + cond_init(sleepCond); + lock_init(splitLock); + timer = new Thread(MAX_THREADS); + read_uci_options(); // Creates at least main thread } @@ -184,22 +261,11 @@ void ThreadsManager::init() { void ThreadsManager::exit() { - assert(threads[0].is_searching == false); - - for (int i = 0; i <= MAX_THREADS; i++) - { - 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_ACTIVE_SPLIT_POINTS; j++) - lock_destroy(threads[i].splitPoints[j].lock); - } + for (int i = 0; i < MAX_THREADS; i++) + if (threads[i]) + delete threads[i]; + delete timer; lock_destroy(splitLock); cond_destroy(sleepCond); } @@ -213,7 +279,7 @@ bool ThreadsManager::available_slave_exists(int master) const { assert(master >= 0 && master < activeThreads); for (int i = 0; i < activeThreads; i++) - if (threads[i].is_available_to(master)) + if (threads[i]->is_available_to(master)) return true; return false; @@ -231,8 +297,8 @@ bool ThreadsManager::available_slave_exists(int master) const { template Value ThreadsManager::split(Position& pos, Stack* ss, Value alpha, Value beta, - Value bestValue, Depth depth, Move threatMove, - int moveCount, MovePicker* mp, int nodeType) { + 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); @@ -243,19 +309,20 @@ Value ThreadsManager::split(Position& pos, Stack* ss, Value alpha, Value beta, assert(activeThreads > 1); int master = pos.thread(); - Thread& masterThread = threads[master]; + Thread& masterThread = *threads[master]; - if (masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS) + if (masterThread.splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD) return bestValue; // Pick the next available split point from the split point stack - SplitPoint* sp = &masterThread.splitPoints[masterThread.activeSplitPoints]; + SplitPoint* sp = &masterThread.splitPoints[masterThread.splitPointsCnt++]; - sp->parent = masterThread.splitPoint; + sp->parent = masterThread.curSplitPoint; sp->master = master; - sp->is_betaCutoff = false; + sp->cutoff = false; sp->slavesMask = 1ULL << master; sp->depth = depth; + sp->bestMove = *bestMove; sp->threatMove = threatMove; sp->alpha = alpha; sp->beta = beta; @@ -269,33 +336,31 @@ Value ThreadsManager::split(Position& pos, Stack* ss, Value alpha, Value beta, assert(masterThread.is_searching); + masterThread.curSplitPoint = sp; int slavesCnt = 0; // 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 // allocation of the same slave by another master. + lock_grab(sp->lock); lock_grab(splitLock); - lock_grab(sp->lock); // To protect sp->slaves_mask for (int i = 0; i < activeThreads && !Fake; i++) - if (threads[i].is_available_to(master)) + if (threads[i]->is_available_to(master)) { sp->slavesMask |= 1ULL << i; - threads[i].splitPoint = sp; - threads[i].is_searching = true; // Slave leaves idle_loop() + threads[i]->curSplitPoint = sp; + threads[i]->is_searching = true; // Slave leaves idle_loop() if (useSleepingThreads) - threads[i].wake_up(); + threads[i]->wake_up(); if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included break; } - masterThread.splitPoint = sp; - masterThread.activeSplitPoints++; - - lock_release(sp->lock); lock_release(splitLock); + lock_release(sp->lock); // 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. @@ -303,44 +368,35 @@ Value ThreadsManager::split(Position& pos, Stack* ss, Value alpha, Value beta, // the thread will return from the idle loop when all slaves have finished // their work at this split point. if (slavesCnt || Fake) + { 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 setting is_searching and decreasing activeSplitPoints is + // finished. Note that setting is_searching and decreasing splitPointsCnt is // done under lock protection to avoid a race with Thread::is_available_to(). - lock_grab(splitLock); lock_grab(sp->lock); // To protect sp->nodes + lock_grab(splitLock); masterThread.is_searching = true; - masterThread.activeSplitPoints--; - masterThread.splitPoint = sp->parent; + masterThread.splitPointsCnt--; + masterThread.curSplitPoint = sp->parent; pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); + *bestMove = sp->bestMove; - lock_release(sp->lock); lock_release(splitLock); + lock_release(sp->lock); return sp->bestValue; } // Explicit template instantiations -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_exit) - { - lock_grab(sleepLock); - timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX); - lock_release(sleepLock); - check_time(); - } -} +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); // ThreadsManager::set_timer() is used to set the timer to trigger after msec @@ -348,42 +404,10 @@ void Thread::timer_loop() { 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_exit) - { - cond_signal(Threads.sleepCond); // Wake up UI thread if needed - cond_wait(sleepCond, sleepLock); - } - - is_searching = true; - - lock_release(sleepLock); - - if (do_exit) - return; - - Search::think(); - } + lock_grab(timer->sleepLock); + timer->maxPly = msec; + cond_signal(timer->sleepCond); // Wake up and restart the timer + lock_release(timer->sleepLock); } @@ -394,7 +418,7 @@ void Thread::main_loop() { void ThreadsManager::start_thinking(const Position& pos, const LimitsType& limits, const std::set& searchMoves, bool async) { - Thread& main = threads[0]; + Thread& main = *threads[0]; lock_grab(main.sleepLock); @@ -434,7 +458,7 @@ void ThreadsManager::start_thinking(const Position& pos, const LimitsType& limit void ThreadsManager::stop_thinking() { - Thread& main = threads[0]; + Thread& main = *threads[0]; Search::Signals.stop = true; @@ -447,25 +471,3 @@ void ThreadsManager::stop_thinking() { 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); -}