X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=src%2Fthread.cpp;h=6c1d729987d90bb237f0b0c9034d78d19aa2d7f0;hb=9050eac59564fe96b3f24d2889bbef7336b28100;hp=5d2b39f57c8c4791dd6b28975448a55d89fa2622;hpb=9a542d96981e6cb45b6b01f17258a078cf27da36;p=stockfish diff --git a/src/thread.cpp b/src/thread.cpp index 5d2b39f5..6c1d7299 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -1,7 +1,8 @@ /* 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) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2015-2019 Marco Costalba, Joona Kiiski, Gary Linscott, 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 @@ -17,333 +18,186 @@ along with this program. If not, see . */ -#include +#include // For std::count +#include +#include "movegen.h" +#include "search.h" #include "thread.h" -#include "ucioption.h" +#include "uci.h" +#include "syzygy/tbprobe.h" +#include "tt.h" -ThreadsManager Threads; // Global object definition +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() with the supplied threadID. - // There are two versions of this function; one for POSIX threads and - // one for Windows threads. +/// Thread constructor launches the thread and waits until it goes to sleep +/// in idle_loop(). Note that 'searching' and 'exit' should be alredy set. -#if defined(_MSC_VER) +Thread::Thread(size_t n) : idx(n), stdThread(&Thread::idle_loop, this) { - DWORD WINAPI start_routine(LPVOID threadID) { - - Threads.idle_loop(*(int*)threadID, NULL); - return 0; - } - -#else - - void* start_routine(void* threadID) { + wait_for_search_finished(); +} - Threads.idle_loop(*(int*)threadID, NULL); - return NULL; - } -#endif +/// Thread destructor wakes up the thread in idle_loop() and waits +/// for its termination. Thread should be already waiting. -} } +Thread::~Thread() { + assert(!searching); -// 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. + exit = true; + start_searching(); + stdThread.join(); +} -void Thread::wake_up() { - lock_grab(&sleepLock); - cond_signal(&sleepCond); - lock_release(&sleepLock); -} +/// Thread::clear() reset histories, usually before a new game +void Thread::clear() { -// 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. + counterMoves.fill(MOVE_NONE); + mainHistory.fill(0); + captureHistory.fill(0); -bool Thread::cutoff_occurred() const { + for (auto& to : continuationHistory) + for (auto& h : to) + h->fill(0); - for (SplitPoint* sp = splitPoint; sp; sp = sp->parent) - if (sp->is_betaCutoff) - return true; - return false; + continuationHistory[NO_PIECE][0]->fill(Search::CounterMovePruneThreshold - 1); } +/// Thread::start_searching() wakes up the thread that will start the search -// 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). +void Thread::start_searching() { -bool Thread::is_available_to(int master) const { + std::lock_guard lk(mutex); + searching = true; + cv.notify_one(); // Wake up the thread in idle_loop() +} - if (state != AVAILABLE) - 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; +/// Thread::wait_for_search_finished() blocks on the condition variable +/// until the thread has finished searching. - // 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; +void Thread::wait_for_search_finished() { - return false; + std::unique_lock lk(mutex); + cv.wait(lk, [&]{ return !searching; }); } -// 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. +/// Thread::idle_loop() is where the thread is parked, blocked on the +/// condition variable, when it has no work to do. -void ThreadsManager::read_uci_options() { +void Thread::idle_loop() { - 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(); -} + // If OS already scheduled us on a different group than 0 then don't overwrite + // the choice, eventually we are one of many one-threaded processes running on + // some Windows NUMA hardware, for instance in fishtest. To make it simple, + // just check if running threads are below a threshold, in this case all this + // NUMA machinery is not needed. + if (Options["Threads"] > 8) + WinProcGroup::bindThisThread(idx); + while (true) + { + std::unique_lock lk(mutex); + searching = false; + cv.notify_one(); // Wake up anyone waiting for search finished + cv.wait(lk, [&]{ return searching; }); -// init() is called during startup. Initializes locks and condition variables -// and launches all threads sending them immediately to sleep. + if (exit) + return; -void ThreadsManager::init() { + lk.unlock(); - int threadID[MAX_THREADS]; + search(); + } +} - // This flag is needed to properly end the threads when program exits - allThreadsShouldExit = false; +/// ThreadPool::set() creates/destroys threads to match the requested number. +/// Created and launched threads will immediately go to sleep in idle_loop. +/// Upon resizing, threads are recreated to allow for binding if necessary. - // Threads will sent to sleep as soon as created, only main thread is kept alive - activeThreads = 1; - threads[0].state = Thread::SEARCHING; +void ThreadPool::set(size_t requested) { - // Allocate pawn and material hash tables for main thread - init_hash_tables(); + if (size() > 0) { // destroy any existing thread(s) + main()->wait_for_search_finished(); - lock_init(&threadsLock); + while (size() > 0) + delete back(), pop_back(); + } - // Initialize thread and split point locks - for (int i = 0; i < MAX_THREADS; i++) - { - lock_init(&threads[i].sleepLock); - cond_init(&threads[i].sleepCond); + if (requested > 0) { // create new thread(s) + push_back(new MainThread(0)); - for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_init(&(threads[i].splitPoints[j].lock)); - } + while (size() < requested) + push_back(new Thread(size())); + clear(); - // Create and startup all the threads but the main that is already running - for (int i = 1; i < MAX_THREADS; i++) - { - threads[i].state = Thread::INITIALIZING; - threadID[i] = 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; - ::exit(EXIT_FAILURE); - } - - // Wait until the thread has finished launching and is gone to sleep - while (threads[i].state == Thread::INITIALIZING) {} + // Reallocate the hash with the new threadpool size + TT.resize(Options["Hash"]); } } +/// ThreadPool::clear() sets threadPool data to initial values. -// exit() is called to cleanly exit the threads when the program finishes - -void ThreadsManager::exit() { +void ThreadPool::clear() { - // Force the woken up threads to exit idle_loop() and hence terminate - allThreadsShouldExit = true; - - 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) {} - } - - // Now we can safely destroy the 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)); - } + for (Thread* th : *this) + th->clear(); - lock_destroy(&threadsLock); + main()->callsCnt = 0; + main()->previousScore = VALUE_INFINITE; + main()->previousTimeReduction = 1.0; } +/// ThreadPool::start_thinking() wakes up main thread waiting in idle_loop() and +/// returns immediately. Main thread will wake up other threads and start the search. -// 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 ThreadPool::start_thinking(Position& pos, StateListPtr& states, + const Search::LimitsType& limits, bool ponderMode) { -void ThreadsManager::init_hash_tables() { + main()->wait_for_search_finished(); - for (int i = 0; i < activeThreads; i++) - { - threads[i].pawnTable.init(); - threads[i].materialTable.init(); - } -} + main()->stopOnPonderhit = stop = false; + main()->ponder = ponderMode; + Search::Limits = limits; + Search::RootMoves rootMoves; + for (const auto& m : MoveList(pos)) + if ( limits.searchmoves.empty() + || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m)) + rootMoves.emplace_back(m); -// available_slave_exists() tries to find an idle thread which is available as -// a slave for the thread with threadID "master". + if (!rootMoves.empty()) + Tablebases::rank_root_moves(pos, rootMoves); -bool ThreadsManager::available_slave_exists(int master) const { + // After ownership transfer 'states' becomes empty, so if we stop the search + // and call 'go' again without setting a new position states.get() == NULL. + assert(states.get() || setupStates.get()); - assert(master >= 0 && master < activeThreads); + if (states.get()) + setupStates = std::move(states); // Ownership transfer, states is now empty - for (int i = 0; i < activeThreads; i++) - if (i != master && threads[i].is_available_to(master)) - return true; + // We use Position::set() to set root position across threads. But there are + // some StateInfo fields (previous, pliesFromNull, capturedPiece) that cannot + // be deduced from a fen string, so set() clears them and to not lose the info + // we need to backup and later restore setupStates->back(). Note that setupStates + // is shared by threads but is accessed in read-only mode. + StateInfo tmp = setupStates->back(); - return false; -} + for (Thread* th : *this) + { + th->nodes = th->tbHits = th->nmpMinPly = 0; + th->rootDepth = th->completedDepth = DEPTH_ZERO; + th->rootMoves = rootMoves; + th->rootPos.set(pos.fen(), pos.is_chess960(), &setupStates->back(), th); + } + setupStates->back() = tmp; -// 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. - -template -void ThreadsManager::split(Position& pos, SearchStack* ss, Value* alpha, const 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(beta <= VALUE_INFINITE); - assert(depth > DEPTH_ZERO); - assert(pos.thread() >= 0 && pos.thread() < activeThreads); - assert(activeThreads > 1); - - int i, master = pos.thread(); - Thread& masterThread = threads[master]; - - // If we already have too many active split points, don't split - if (masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS) - return; - - // 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; - for (i = 0; i < activeThreads; i++) - splitPoint.is_slave[i] = false; - - // If we are here it means we are not available - assert(masterThread.state == Thread::SEARCHING); - - int booked = 0; - - // 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. - lock_grab(&threadsLock); - - for (i = 0; !Fake && i < activeThreads && booked < 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++; - } - - lock_release(&threadsLock); - - // We failed to allocate even one slave, return - if (!Fake && !booked) - return; - - masterThread.activeSplitPoints++; - masterThread.splitPoint = &splitPoint; - - // 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.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 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); - - // We have returned from the idle loop, which means that all threads are - // finished. Update alpha and bestValue, and return. 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.activeSplitPoints--; - masterThread.splitPoint = splitPoint.parent; - - lock_release(&threadsLock); - - *alpha = splitPoint.alpha; - *bestValue = splitPoint.bestValue; - pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes); + main()->start_searching(); } - -// Explicit template instantiations -template void ThreadsManager::split(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, int); -template void ThreadsManager::split(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, int);