X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=4a37d393edff4ca6f83eaa9a7f86917ec8dcc6ef;hp=e1330396d0007f30815c3ba95d2142c9ad8bb1d7;hb=44c78fdb7ad4c82656f25e4da11f3f9d56a4c6aa;hpb=5b35c149e833e365c2afb8039ca5c658abc53081 diff --git a/src/thread.cpp b/src/thread.cpp index e1330396..4a37d393 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -19,6 +19,7 @@ #include +#include "search.h" #include "thread.h" #include "ucioption.h" @@ -27,27 +28,26 @@ ThreadsManager Threads; // Global object definition 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 + // last two threads are dedicated to read input from GUI and to mimic a + // timer, so they run in listener_loop() and timer_loop() respectively. #if defined(_MSC_VER) - - DWORD WINAPI start_routine(LPVOID threadID) { - - Threads.idle_loop(*(int*)threadID, NULL); - return 0; - } - + DWORD WINAPI start_routine(LPVOID thread) { #else + void* start_routine(void* thread) { +#endif - void* start_routine(void* threadID) { + if (((Thread*)thread)->threadID == MAX_THREADS) + ((Thread*)thread)->listener_loop(); - Threads.idle_loop(*(int*)threadID, NULL); - return NULL; - } + else if (((Thread*)thread)->threadID == MAX_THREADS + 1) + ((Thread*)thread)->timer_loop(); + else + ((Thread*)thread)->idle_loop(NULL); -#endif + return 0; + } } } @@ -63,9 +63,8 @@ void Thread::wake_up() { } -// 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 { @@ -85,7 +84,7 @@ 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 @@ -111,31 +110,51 @@ 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(); + + set_size(Options["Threads"].value()); } -// 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) { + + assert(cnt > 0 && cnt <= MAX_THREADS); - int threadID[MAX_THREADS]; + activeThreads = cnt; - // This flag is needed to properly end the threads when program exits - allThreadsShouldExit = false; + 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 as, for instance, on mobile + // devices where memory is scarce and allocating for MAX_THREADS could + // even result in a crash. + threads[i].pawnTable.init(); + threads[i].materialTable.init(); + + threads[i].do_sleep = false; + } + else + threads[i].do_sleep = true; +} - // Threads will sent to sleep as soon as created, only main thread is kept alive - activeThreads = 1; - threads[0].state = Thread::SEARCHING; - // Allocate pawn and material hash tables for main thread - init_hash_tables(); +// init() is called during startup. Initializes locks and condition variables +// and launches all threads sending them immediately to sleep. + +void ThreadsManager::init() { + // Initialize sleep condition used to block waiting for GUI input + cond_init(&sleepCond); + + // Initialize threads lock, used when allocating slaves during splitting lock_init(&threadsLock); - // Initialize thread and split point locks - for (int i = 0; i < MAX_THREADS; i++) + // Initialize sleep and split point locks + for (int i = 0; i < MAX_THREADS + 2; i++) { lock_init(&threads[i].sleepLock); cond_init(&threads[i].sleepCond); @@ -144,48 +163,55 @@ void ThreadsManager::init() { 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++) + // Initialize main thread's associated data + threads[0].is_searching = true; + threads[0].threadID = 0; + set_size(1); // This makes all the threads but the main to go to sleep + + // Create and launch all the threads but the main that is already running, + // threads will go immediately to sleep. + for (int i = 1; i < MAX_THREADS + 2; i++) { - threads[i].state = Thread::INITIALIZING; - threadID[i] = i; + threads[i].is_searching = false; + threads[i].threadID = i; #if defined(_MSC_VER) - bool ok = (CreateThread(NULL, 0, start_routine, (LPVOID)&threadID[i], 0, NULL) != NULL); + threads[i].handle = CreateThread(NULL, 0, start_routine, (LPVOID)&threads[i], 0, NULL); + bool ok = (threads[i].handle != NULL); #else - pthread_t pthreadID; - bool ok = (pthread_create(&pthreadID, NULL, start_routine, (void*)&threadID[i]) == 0); - pthread_detach(pthreadID); + bool ok = (pthread_create(&threads[i].handle, NULL, start_routine, (void*)&threads[i]) == 0); #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 + 2; i++) { - // Wake up all the threads and waits for termination if (i != 0) { + threads[i].do_terminate = true; threads[i].wake_up(); - while (threads[i].state != Thread::TERMINATED) {} + + // Wait for slave termination +#if defined(_MSC_VER) + WaitForSingleObject(threads[i].handle, 0); + CloseHandle(threads[i].handle); +#else + pthread_join(threads[i].handle, NULL); +#endif } - // Now we can safely destroy the locks and wait conditions + // Now we can safely destroy locks and wait conditions lock_destroy(&threads[i].sleepLock); cond_destroy(&threads[i].sleepCond); @@ -194,22 +220,7 @@ void ThreadsManager::exit() { } 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(); - } + cond_destroy(&sleepCond); } @@ -228,6 +239,19 @@ bool ThreadsManager::available_slave_exists(int master) const { } +// 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 @@ -241,7 +265,7 @@ template Value ThreadsManager::split(Position& pos, SearchStack* ss, Value alpha, Value beta, Value bestValue, Depth depth, Move threatMove, int moveCount, MovePicker* mp, int nodeType) { - assert(pos.is_ok()); + assert(pos.pos_is_ok()); assert(bestValue >= -VALUE_INFINITE); assert(bestValue <= alpha); assert(alpha < beta); @@ -258,90 +282,227 @@ Value ThreadsManager::split(Position& pos, SearchStack* ss, Value alpha, Value b return bestValue; // Pick the next available split point object from the split point stack - SplitPoint& splitPoint = masterThread.splitPoints[masterThread.activeSplitPoints]; + SplitPoint* sp = 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; + 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 booked = 0; + int workersCnt = 1; // At least the master is included - // Try to allocate available threads setting state to Thread::BOOKED, this - // must be done under lock protection to avoid concurrent allocation of - // the same slave by another master. + // 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); - for (i = 0; !Fake && i < activeThreads && booked < maxThreadsPerSplitPoint; i++) + for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++) if (i != master && threads[i].is_available_to(master)) { - threads[i].state = Thread::BOOKED; - threads[i].splitPoint = &splitPoint; - splitPoint.is_slave[i] = true; - booked++; + workersCnt++; + sp->is_slave[i] = true; + threads[i].splitPoint = sp; + + // This makes the slave to exit from idle_loop() + threads[i].is_searching = true; + + if (useSleepingThreads) + threads[i].wake_up(); } lock_release(&threadsLock); // We failed to allocate even one slave, return - if (!Fake && !booked) + if (!Fake && workersCnt == 1) return bestValue; + masterThread.splitPoint = sp; masterThread.activeSplitPoints++; - masterThread.splitPoint = &splitPoint; - // Tell the threads that they have some work to do. This will make them leave - // their idle loop. - for (i = 0; i < activeThreads; i++) - if (i == master || splitPoint.is_slave[i]) - { - assert(i == master || threads[i].state == Thread::BOOKED); - - // This makes the slave to exit from idle_loop() - threads[i].state = Thread::WORKISWAITING; - - if (useSleepingThreads && i != master) - threads[i].wake_up(); - } + // 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); - // 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); + // In helpful master concept a master can help only a sub-tree, 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); - masterThread.state = Thread::SEARCHING; + masterThread.is_searching = true; masterThread.activeSplitPoints--; - masterThread.splitPoint = splitPoint.parent; lock_release(&threadsLock); - pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes); - return splitPoint.bestValue; + masterThread.splitPoint = sp->parent; + pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); + + 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); + + +// Thread::timer_loop() is where the timer thread waits maxPly milliseconds +// and then calls do_timer_event(). + +void Thread::timer_loop() { + + while (!do_terminate) + { + lock_grab(&sleepLock); + timed_wait(&sleepCond, &sleepLock, maxPly ? maxPly : INT_MAX); + lock_release(&sleepLock); + do_timer_event(); + } +} + + +// 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 + 1]; + + lock_grab(&timer.sleepLock); + timer.maxPly = msec; + cond_signal(&timer.sleepCond); // Wake up and restart the timer + lock_release(&timer.sleepLock); +} + + +// Thread::listener_loop() is where the listener thread, used for I/O, waits for +// input. When is_searching is false then input is read in sync with main thread +// (that blocks), otherwise the listener thread reads any input asynchronously +// and processes the input line calling do_uci_async_cmd(). + +void Thread::listener_loop() { + + std::string cmd; + + while (true) + { + lock_grab(&sleepLock); + + Threads.inputLine = cmd; + do_sleep = !is_searching; + + // Here the thread is parked in sync mode after a line has been read + while (do_sleep && !do_terminate) // Catches spurious wake ups + { + cond_signal(&Threads.sleepCond); // Wake up main thread + cond_wait(&sleepCond, &sleepLock); // Sleep here + } + + lock_release(&sleepLock); + + if (do_terminate) + return; + + if (!std::getline(std::cin, cmd)) // Block waiting for input + cmd = "quit"; + + lock_grab(&sleepLock); + + // If we are in async mode then process the command now + if (is_searching) + { + // Command "quit" is the last one received by the GUI, so park the + // thread waiting for exiting. Also, after a "stop", for instance on a + // ponder miss, GUI can immediately send the new position to search, + // so return to in-sync mode to avoid discarding good data. + if (cmd == "quit" || cmd == "stop") + is_searching = false; + + do_uci_async_cmd(cmd); + cmd = ""; // Input has been consumed + } + + lock_release(&sleepLock); + } +} + + +// ThreadsManager::getline() is used by main thread to block and wait for input, +// the behaviour mimics std::getline(). + +void ThreadsManager::getline(std::string& cmd) { + + Thread& listener = threads[MAX_THREADS]; + + lock_grab(&listener.sleepLock); + + listener.is_searching = false; // Set sync mode + + // If there is already some input to grab then skip without to wake up the + // listener. This can happen if after we send the "bestmove", the GUI sends + // a command that the listener buffers in inputLine before going to sleep. + if (inputLine.empty()) + { + listener.do_sleep = false; + cond_signal(&listener.sleepCond); // Wake up listener thread + + while (!listener.do_sleep) + cond_wait(&sleepCond, &listener.sleepLock); // Wait for input + } + + cmd = inputLine; + inputLine = ""; // Input has been consumed + + lock_release(&listener.sleepLock); +} + + +// ThreadsManager::start_listener() is called at the beginning of the search to +// swith from sync behaviour (default) to async and so be able to read from UCI +// while other threads are searching. This avoids main thread polling for input. + +void ThreadsManager::start_listener() { + + Thread& listener = threads[MAX_THREADS]; + + lock_grab(&listener.sleepLock); + listener.is_searching = true; + listener.do_sleep = false; + cond_signal(&listener.sleepCond); // Wake up listener thread + lock_release(&listener.sleepLock); +} + + +// ThreadsManager::stop_listener() is called before to send "bestmove" to GUI to +// return to in-sync behaviour. This is needed because while in async mode any +// command is discarded without being processed (except for a very few ones). + +void ThreadsManager::stop_listener() { + + Thread& listener = threads[MAX_THREADS]; + + lock_grab(&listener.sleepLock); + listener.is_searching = false; + lock_release(&listener.sleepLock); +}