X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=src%2Fthread.cpp;h=b46fce5e873933494031eade0330582b48003e11;hb=0405f3540366cc16245d51531881c55d3726c8b5;hp=002819de566ce71a2c91e2566ca481bf1376445c;hpb=339e1b49f619ceffa75019e196adf4de74b32cce;p=stockfish diff --git a/src/thread.cpp b/src/thread.cpp index 002819de..b46fce5e 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -1,7 +1,6 @@ /* 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) 2004-2020 The Stockfish developers (see AUTHORS file) Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -17,331 +16,245 @@ along with this program. If not, see . */ -#include +#include +#include // For std::count +#include "movegen.h" +#include "search.h" #include "thread.h" -#include "ucioption.h" +#include "uci.h" +#include "syzygy/tbprobe.h" +#include "tt.h" -ThreadsManager ThreadsMgr; // Global object definition +ThreadPool Threads; // Global object -namespace { - // 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. +/// Thread constructor launches the thread and waits until it goes to sleep +/// in idle_loop(). Note that 'searching' and 'exit' should be already set. -#if !defined(_MSC_VER) +Thread::Thread(size_t n) : idx(n), stdThread(&Thread::idle_loop, this) { - void* init_thread(void* threadID) { + wait_for_search_finished(); +} - ThreadsMgr.idle_loop(*(int*)threadID, NULL); - return NULL; - } -#else +/// Thread destructor wakes up the thread in idle_loop() and waits +/// for its termination. Thread should be already waiting. - DWORD WINAPI init_thread(LPVOID threadID) { +Thread::~Thread() { - ThreadsMgr.idle_loop(*(int*)threadID, NULL); - return 0; - } - -#endif + assert(!searching); + exit = true; + start_searching(); + stdThread.join(); } -// 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. - -void ThreadsManager::read_uci_options() { +/// Thread::clear() reset histories, usually before a new game - 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(); -} - - -// init_threads() is called during startup. Initializes locks and condition -// variables and launches all threads sending them immediately to sleep. +void Thread::clear() { -void ThreadsManager::init_threads() { + counterMoves.fill(MOVE_NONE); + mainHistory.fill(0); + lowPlyHistory.fill(0); + captureHistory.fill(0); - int arg[MAX_THREADS]; + for (bool inCheck : { false, true }) + for (StatsType c : { NoCaptures, Captures }) + { + for (auto& to : continuationHistory[inCheck][c]) + for (auto& h : to) + h->fill(0); + continuationHistory[inCheck][c][NO_PIECE][0]->fill(Search::CounterMovePruneThreshold - 1); + } +} - // This flag is needed to properly end the threads when program exits - allThreadsShouldExit = false; - // Threads will sent to sleep as soon as created, only main thread is kept alive - activeThreads = 1; - threads[0].state = THREAD_SEARCHING; +/// Thread::start_searching() wakes up the thread that will start the search - // Allocate pawn and material hash tables for main thread - init_hash_tables(); +void Thread::start_searching() { - lock_init(&mpLock); + std::lock_guard lk(mutex); + searching = true; + cv.notify_one(); // Wake up the thread in idle_loop() +} - // Initialize thread and split point locks - 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)); - } +/// Thread::wait_for_search_finished() blocks on the condition variable +/// until the thread has finished searching. - // 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; - arg[i] = i; - -#if !defined(_MSC_VER) - pthread_t pthread[1]; - bool ok = (pthread_create(pthread, NULL, init_thread, (void*)(&arg[i])) == 0); - pthread_detach(pthread[0]); -#else - bool ok = (CreateThread(NULL, 0, init_thread, (LPVOID)(&arg[i]), 0, NULL) != NULL); -#endif - if (!ok) - { - std::cout << "Failed to create thread number " << i << std::endl; - exit(EXIT_FAILURE); - } +void Thread::wait_for_search_finished() { - // Wait until the thread has finished launching and is gone to sleep - while (threads[i].state == THREAD_INITIALIZING) {} - } + std::unique_lock lk(mutex); + cv.wait(lk, [&]{ return !searching; }); } -// exit_threads() is called when the program exits. It makes all the -// helper threads exit cleanly. +/// Thread::idle_loop() is where the thread is parked, blocked on the +/// condition variable, when it has no work to do. -void ThreadsManager::exit_threads() { +void Thread::idle_loop() { - // Force the woken up threads to exit idle_loop() and hence terminate - allThreadsShouldExit = true; + // 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); - for (int i = 0; i < MAX_THREADS; i++) + while (true) { - // Wake up all the threads and waits for termination - if (i != 0) - { - threads[i].wake_up(); - while (threads[i].state != THREAD_TERMINATED) {} - } + std::unique_lock lk(mutex); + searching = false; + cv.notify_one(); // Wake up anyone waiting for search finished + cv.wait(lk, [&]{ return searching; }); - // Now we can safely destroy the locks and wait conditions - lock_destroy(&threads[i].sleepLock); - cond_destroy(&threads[i].sleepCond); + if (exit) + return; - for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_destroy(&(threads[i].splitPoints[j].lock)); - } + lk.unlock(); - lock_destroy(&mpLock); + search(); + } } +/// 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. -// 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::set(size_t requested) { -void ThreadsManager::init_hash_tables() { + if (size() > 0) { // destroy any existing thread(s) + main()->wait_for_search_finished(); - for (int i = 0; i < activeThreads; i++) - { - threads[i].pawnTable.init(); - threads[i].materialTable.init(); + while (size() > 0) + delete back(), pop_back(); } -} - -// 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. + if (requested > 0) { // create new thread(s) + push_back(new MainThread(0)); -bool ThreadsManager::cutoff_at_splitpoint(int threadID) const { + while (size() < requested) + push_back(new Thread(size())); + clear(); - assert(threadID >= 0 && threadID < activeThreads); + // Reallocate the hash with the new threadpool size + TT.resize(size_t(Options["Hash"])); - SplitPoint* sp = threads[threadID].splitPoint; - - for ( ; sp && !sp->betaCutoff; sp = sp->parent) {} - return sp != NULL; + // Init thread number dependent search params. + Search::init(); + } } -// 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 { +/// ThreadPool::clear() sets threadPool data to initial values - assert(slave >= 0 && slave < activeThreads); - assert(master >= 0 && master < activeThreads); - assert(activeThreads > 1); +void ThreadPool::clear() { - if (threads[slave].state != THREAD_AVAILABLE || slave == master) - return false; + for (Thread* th : *this) + th->clear(); - // 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; + main()->callsCnt = 0; + main()->bestPreviousScore = VALUE_INFINITE; + main()->previousTimeReduction = 1.0; +} - // 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; -} +/// 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. +void ThreadPool::start_thinking(Position& pos, StateListPtr& states, + const Search::LimitsType& limits, bool ponderMode) { -// available_thread_exists() tries to find an idle thread which is available as -// a slave for the thread with threadID "master". + main()->wait_for_search_finished(); -bool ThreadsManager::available_thread_exists(int master) const { + main()->stopOnPonderhit = stop = false; + increaseDepth = true; + main()->ponder = ponderMode; + Search::Limits = limits; + Search::RootMoves rootMoves; - assert(master >= 0 && master < activeThreads); - assert(activeThreads > 1); + for (const auto& m : MoveList(pos)) + if ( limits.searchmoves.empty() + || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m)) + rootMoves.emplace_back(m); - for (int i = 0; i < activeThreads; i++) - if (thread_is_available(i, master)) - return true; + if (!rootMoves.empty()) + Tablebases::rank_root_moves(pos, rootMoves); - return false; -} + // 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()); + if (states.get()) + setupStates = std::move(states); // Ownership transfer, states is now empty -// 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, bool pvNode) { - 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]; - - 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) + // 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 they are set from + // setupStates->back() later. The rootState is per thread, earlier states are shared + // since they are read-only. + for (Thread* th : *this) { - lock_release(&mpLock); - return; + th->nodes = th->tbHits = th->nmpMinPly = th->bestMoveChanges = 0; + th->rootDepth = th->completedDepth = 0; + th->rootMoves = rootMoves; + th->rootPos.set(pos.fen(), pos.is_chess960(), &th->rootState, th); + th->rootState = setupStates->back(); } - // 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.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; - for (i = 0; i < activeThreads; i++) - splitPoint.slaves[i] = 0; - - masterThread.splitPoint = &splitPoint; - - // If we are here it means we are not available - assert(masterThread.state != THREAD_AVAILABLE); - - int workersCnt = 1; // At least the master is included - - // Allocate available threads setting state to THREAD_BOOKED - for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++) - if (thread_is_available(i, master)) - { - threads[i].state = THREAD_BOOKED; - threads[i].splitPoint = &splitPoint; - splitPoint.slaves[i] = 1; - workersCnt++; - } + main()->start_searching(); +} - assert(Fake || workersCnt > 1); +Thread* ThreadPool::get_best_thread() const { + + Thread* bestThread = front(); + std::map votes; + Value minScore = VALUE_NONE; + + // Find minimum score of all threads + for (Thread* th: *this) + minScore = std::min(minScore, th->rootMoves[0].score); + + // Vote according to score and depth, and select the best thread + for (Thread* th : *this) + { + votes[th->rootMoves[0].pv[0]] += + (th->rootMoves[0].score - minScore + 14) * int(th->completedDepth); + + if (abs(bestThread->rootMoves[0].score) >= VALUE_TB_WIN_IN_MAX_PLY) + { + // Make sure we pick the shortest mate / TB conversion or stave off mate the longest + if (th->rootMoves[0].score > bestThread->rootMoves[0].score) + bestThread = th; + } + else if ( th->rootMoves[0].score >= VALUE_TB_WIN_IN_MAX_PLY + || ( th->rootMoves[0].score > VALUE_TB_LOSS_IN_MAX_PLY + && votes[th->rootMoves[0].pv[0]] > votes[bestThread->rootMoves[0].pv[0]])) + bestThread = th; + } + + return bestThread; +} - // We can release the lock because slave threads are already booked and master is not available - lock_release(&mpLock); - // 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); +/// Start non-main threads - threads[i].state = THREAD_WORKISWAITING; // This makes the slave to exit from idle_loop() +void ThreadPool::start_searching() { - if (useSleepingThreads && i != master) - threads[i].wake_up(); - } + for (Thread* th : *this) + if (th != front()) + th->start_searching(); +} - // 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. - lock_grab(&mpLock); +/// Wait for non-main threads - *alpha = splitPoint.alpha; - *bestValue = splitPoint.bestValue; - masterThread.activeSplitPoints--; - masterThread.splitPoint = splitPoint.parent; - pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes); +void ThreadPool::wait_for_search_finished() const { - lock_release(&mpLock); + for (Thread* th : *this) + if (th != front()) + th->wait_for_search_finished(); } - -// 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);