X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=d845fdf2e4183f2fdaa10bb656b66d1ab6439749;hp=30c921d1808ebd4716dee58bab6deeefe0e54279;hb=83d8fe2d59bad718de70b00ac2c8fddfadda76b5;hpb=fecefbb99cb0147f37d6895765a315f34c935786 diff --git a/src/thread.cpp b/src/thread.cpp index 30c921d1..d845fdf2 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -22,33 +22,82 @@ #include "thread.h" #include "ucioption.h" -ThreadsManager ThreadsMgr; // Global object definition +ThreadsManager Threads; // Global object definition -namespace { +namespace { extern "C" { - // init_thread() is the function which is called when a new thread is - // launched. It simply calls the idle_loop() function with the supplied - // threadID. There are two versions of this function; one for POSIX - // threads and one for Windows threads. + // start_routine() is the C function which is called when a new thread + // is launched. It simply calls idle_loop() of the supplied thread. + // There are two versions of this function; one for POSIX threads and + // one for Windows threads. -#if !defined(_MSC_VER) +#if defined(_MSC_VER) - void* init_thread(void* threadID) { + DWORD WINAPI start_routine(LPVOID thread) { - ThreadsMgr.idle_loop(*(int*)threadID, NULL); - return NULL; + ((Thread*)thread)->idle_loop(NULL); + return 0; } #else - DWORD WINAPI init_thread(LPVOID threadID) { + void* start_routine(void* thread) { - ThreadsMgr.idle_loop(*(int*)threadID, NULL); - return 0; + ((Thread*)thread)->idle_loop(NULL); + return NULL; } #endif +} } + + +// 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() { + + lock_grab(&sleepLock); + cond_signal(&sleepCond); + 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) + return true; + return false; +} + + +// 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). + +bool Thread::is_available_to(int master) const { + + if (is_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 localActiveSplitPoints = activeSplitPoints; + + // 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. + if ( !localActiveSplitPoints + || splitPoints[localActiveSplitPoints - 1].is_slave[master]) + return true; + + return false; } @@ -61,84 +110,110 @@ 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_threads() is called during startup. Initializes locks and condition -// variables and launches all threads sending them immediately to sleep. -void ThreadsManager::init_threads() { +// set_size() changes the number of active threads and raises do_sleep flag for +// all the unused threads that will go immediately to sleep. - int i, arg[MAX_THREADS]; - bool ok; +void ThreadsManager::set_size(int cnt) { - // This flag is needed to properly end the threads when program exits - allThreadsShouldExit = false; + assert(cnt > 0 && cnt <= MAX_THREADS); - // Threads will sent to sleep as soon as created, only main thread is kept alive - activeThreads = 1; + 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 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; +} - lock_init(&mpLock); - for (i = 0; i < MAX_THREADS; i++) +// init() is called during startup. Initializes locks and condition variables +// and launches all threads sending them immediately to sleep. + +void ThreadsManager::init() { + + // Initialize threads lock, used when allocating slaves during splitting + lock_init(&threadsLock); + + // Initialize sleep and split point locks + for (int i = 0; i < MAX_THREADS; i++) { - // Initialize thread and split point locks 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)); - - // All threads but first should be set to THREAD_INITIALIZING - threads[i].state = (i == 0 ? THREAD_SEARCHING : THREAD_INITIALIZING); - - // Not in Threads c'tor to avoid global initialization order issues - threads[i].pawnTable.init(); - threads[i].materialTable.init(); } - // Create and startup the threads - for (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; i++) { - arg[i] = i; + threads[i].is_searching = false; + threads[i].threadID = i; -#if !defined(_MSC_VER) - pthread_t pthread[1]; - ok = (pthread_create(pthread, NULL, init_thread, (void*)(&arg[i])) == 0); - pthread_detach(pthread[0]); +#if defined(_MSC_VER) + threads[i].handle = CreateThread(NULL, 0, start_routine, (LPVOID)&threads[i], 0, NULL); + bool ok = (threads[i].handle != NULL); #else - ok = (CreateThread(NULL, 0, init_thread, (LPVOID)(&arg[i]), 0, NULL) != NULL); + 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; - exit(EXIT_FAILURE); + 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_threads() is called when the program exits. It makes all the -// helper threads exit cleanly. +// exit() is called to cleanly terminate the threads when the program finishes -void ThreadsManager::exit_threads() { +void ThreadsManager::exit() { - // Force the woken up threads to exit idle_loop() and hence terminate - allThreadsShouldExit = true; + // Wake up all the slave threads at once. This is faster than "wake and wait" + // for each thread and avoids a rare crash once every 10K games under Linux. + for (int i = 1; i < MAX_THREADS; i++) + { + threads[i].do_terminate = true; + threads[i].wake_up(); + } 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) {} + // 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); @@ -146,70 +221,19 @@ void ThreadsManager::exit_threads() { lock_destroy(&(threads[i].splitPoints[j].lock)); } - lock_destroy(&mpLock); -} - - -// cutoff_at_splitpoint() checks whether a beta cutoff has occurred in -// the thread's currently active split point, or in some ancestor of -// the current split point. - -bool ThreadsManager::cutoff_at_splitpoint(int threadID) const { - - assert(threadID >= 0 && threadID < activeThreads); - - SplitPoint* sp = threads[threadID].splitPoint; - - for ( ; sp && !sp->betaCutoff; sp = sp->parent) {} - return sp != NULL; -} - - -// thread_is_available() checks whether the thread with threadID "slave" is -// available to help the thread with threadID "master" at a split point. An -// obvious requirement is that "slave" must be idle. With more than two -// threads, this is not by itself sufficient: If "slave" is the master of -// some active split point, it is only available as a slave to the other -// threads which are busy searching the split point at the top of "slave"'s -// split point stack (the "helpful master concept" in YBWC terminology). - -bool ThreadsManager::thread_is_available(int slave, int master) const { - - assert(slave >= 0 && slave < activeThreads); - assert(master >= 0 && master < activeThreads); - assert(activeThreads > 1); - - if (threads[slave].state != THREAD_AVAILABLE || slave == master) - return false; - - // Make a local copy to be sure doesn't change under our feet - int localActiveSplitPoints = threads[slave].activeSplitPoints; - - // No active split points means that the thread is available as - // a slave for any other thread. - if (localActiveSplitPoints == 0 || activeThreads == 2) - return true; - - // Apply the "helpful master" concept if possible. Use localActiveSplitPoints - // that is known to be > 0, instead of threads[slave].activeSplitPoints that - // could have been set to 0 by another thread leading to an out of bound access. - if (threads[slave].splitPoints[localActiveSplitPoints - 1].slaves[master]) - return true; - - return false; + lock_destroy(&threadsLock); } -// available_thread_exists() tries to find an idle thread which is available as +// available_slave_exists() tries to find an idle thread which is available as // a slave for the thread with threadID "master". -bool ThreadsManager::available_thread_exists(int master) const { +bool ThreadsManager::available_slave_exists(int master) const { assert(master >= 0 && master < activeThreads); - assert(activeThreads > 1); for (int i = 0; i < activeThreads; i++) - if (thread_is_available(i, master)) + if (i != master && threads[i].is_available_to(master)) return true; return false; @@ -226,13 +250,13 @@ bool ThreadsManager::available_thread_exists(int master) const { // call search().When all threads have returned from search() then split() returns. template -void ThreadsManager::split(Position& pos, SearchStack* ss, Value* alpha, const Value beta, - Value* bestValue, Depth depth, Move threatMove, - int moveCount, MovePicker* mp, bool pvNode) { +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(*bestValue >= -VALUE_INFINITE); - assert(*bestValue <= *alpha); - assert(*alpha < beta); + 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); @@ -241,93 +265,91 @@ void ThreadsManager::split(Position& pos, SearchStack* ss, Value* alpha, const V int i, master = pos.thread(); Thread& masterThread = threads[master]; - lock_grab(&mpLock); - - // If no other thread is available to help us, or if we have too many - // active split points, don't split. - if ( !available_thread_exists(master) - || masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS) - { - lock_release(&mpLock); - return; - } + // If we already have too many active split points, don't split + 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++]; + SplitPoint* sp = masterThread.splitPoints + masterThread.activeSplitPoints; // Initialize the split point object - splitPoint.parent = masterThread.splitPoint; - splitPoint.master = master; - splitPoint.betaCutoff = false; - splitPoint.depth = depth; - splitPoint.threatMove = threatMove; - splitPoint.alpha = *alpha; - splitPoint.beta = beta; - splitPoint.pvNode = pvNode; - 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.slaves[i] = 0; - - masterThread.splitPoint = &splitPoint; + sp->is_slave[i] = false; // If we are here it means we are not available - assert(masterThread.state != THREAD_AVAILABLE); + assert(masterThread.is_searching); int workersCnt = 1; // At least the master is included - // Allocate available threads setting state to THREAD_BOOKED + // 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 && workersCnt < maxThreadsPerSplitPoint; i++) - if (thread_is_available(i, master)) + if (i != master && threads[i].is_available_to(master)) { - threads[i].state = THREAD_BOOKED; - threads[i].splitPoint = &splitPoint; - splitPoint.slaves[i] = 1; workersCnt++; - } + sp->is_slave[i] = true; + threads[i].splitPoint = sp; - assert(Fake || workersCnt > 1); + // This makes the slave to exit from idle_loop() + threads[i].is_searching = true; - // We can release the lock because slave threads are already booked and master is not available - lock_release(&mpLock); + if (useSleepingThreads) + threads[i].wake_up(); + } - // Tell the threads that they have work to do. This will make them leave - // their idle loop. - for (i = 0; i < activeThreads; i++) - if (i == master || splitPoint.slaves[i]) - { - assert(i == master || threads[i].state == THREAD_BOOKED); + lock_release(&threadsLock); - threads[i].state = THREAD_WORKISWAITING; // This makes the slave to exit from idle_loop() + // We failed to allocate even one slave, return + if (!Fake && workersCnt == 1) + return bestValue; - if (useSleepingThreads && i != master) - threads[i].wake_up(); - } + masterThread.splitPoint = sp; + masterThread.activeSplitPoints++; + + // 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. Update alpha and bestValue, and return. - lock_grab(&mpLock); + // 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); - *alpha = splitPoint.alpha; - *bestValue = splitPoint.bestValue; + masterThread.is_searching = true; masterThread.activeSplitPoints--; - masterThread.splitPoint = splitPoint.parent; - pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes); - lock_release(&mpLock); + lock_release(&threadsLock); + + masterThread.splitPoint = sp->parent; + pos.set_nodes_searched(pos.nodes_searched() + sp->nodes); + + return sp->bestValue; } // Explicit template instantiations -template void ThreadsManager::split<0>(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, bool); -template void ThreadsManager::split<1>(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, bool); +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);