/*
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) 2015-2018 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
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
-#include <iostream>
+#include <algorithm> // For std::count
+#include <cassert>
+#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[*(int*)threadID].idle_loop(NULL);
- return 0;
- }
-
-#else
-
- void* start_routine(void* threadID) {
-
- Threads[*(int*)threadID].idle_loop(NULL);
- return NULL;
- }
-
-#endif
+ wait_for_search_finished();
+}
-} }
+/// Thread destructor wakes up the thread in idle_loop() and waits
+/// for its termination. Thread should be already waiting.
-// 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() {
-void Thread::wake_up() {
+ assert(!searching);
- lock_grab(&sleepLock);
- cond_signal(&sleepCond);
- lock_release(&sleepLock);
+ exit = true;
+ start_searching();
+ stdThread.join();
}
-// 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.
+/// Thread::clear() reset histories, usually before a new game
-bool Thread::cutoff_occurred() const {
-
- for (SplitPoint* sp = splitPoint; sp; sp = sp->parent)
- if (sp->is_betaCutoff)
- return true;
- return false;
-}
+void Thread::clear() {
+ counterMoves.fill(MOVE_NONE);
+ mainHistory.fill(0);
+ captureHistory.fill(0);
-// 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).
+ for (auto& to : contHistory)
+ for (auto& h : to)
+ h.get()->fill(0);
-bool Thread::is_available_to(int master) const {
-
- if (state != AVAILABLE)
- return false;
+ contHistory[NO_PIECE][0].get()->fill(Search::CounterMovePruneThreshold - 1);
+}
- // 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::start_searching() wakes up the thread that will start the search
- // 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::start_searching() {
- return false;
+ std::lock_guard<Mutex> lk(mutex);
+ searching = true;
+ cv.notify_one(); // Wake up the thread in idle_loop()
}
-// 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::wait_for_search_finished() blocks on the condition variable
+/// until the thread has finished searching.
-void ThreadsManager::read_uci_options() {
+void Thread::wait_for_search_finished() {
- maxThreadsPerSplitPoint = Options["Maximum Number of Threads per Split Point"].value<int>();
- minimumSplitDepth = Options["Minimum Split Depth"].value<int>() * ONE_PLY;
- useSleepingThreads = Options["Use Sleeping Threads"].value<bool>();
-
- set_size(Options["Threads"].value<int>());
+ std::unique_lock<Mutex> lk(mutex);
+ cv.wait(lk, [&]{ return !searching; });
}
-// set_size() changes the number of active threads and raises do_sleep flag for
-// all the unused threads that will go immediately to sleep.
+/// Thread::idle_loop() is where the thread is parked, blocked on the
+/// condition variable, when it has no work to do.
-void ThreadsManager::set_size(int cnt) {
+void Thread::idle_loop() {
- assert(cnt > 0 && cnt <= MAX_THREADS);
+ // 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);
- activeThreads = cnt;
+ while (true)
+ {
+ std::unique_lock<Mutex> lk(mutex);
+ searching = false;
+ cv.notify_one(); // Wake up anyone waiting for search finished
+ cv.wait(lk, [&]{ return searching; });
- for (int i = 0; i < MAX_THREADS; i++)
- threads[i].do_sleep = !(i < activeThreads);
-}
+ if (exit)
+ return;
+ lk.unlock();
-// init() is called during startup. Initializes locks and condition variables
-// and launches all threads sending them immediately to sleep.
+ search();
+ }
+}
-void ThreadsManager::init() {
+/// ThreadPool::set() creates/destroys threads to match the requested number.
+/// Created and launched threads will go immediately to sleep in idle_loop.
+/// Upon resizing, threads are recreated to allow for binding if necessary.
- // Threads will go to sleep as soon as created, only main thread is kept alive
- set_size(1);
- threads[0].state = Thread::SEARCHING;
- threads[0].threadID = 0;
+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();
- // Initialize threads lock, used when allocating slaves during splitting
- lock_init(&threadsLock);
+ while (size() > 0)
+ delete back(), pop_back();
+ }
- // Initialize sleep 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::AVAILABLE;
- threads[i].threadID = i;
-
-#if defined(_MSC_VER)
- bool ok = (CreateThread(NULL, 0, start_routine, (LPVOID)&threads[i].threadID , 0, NULL) != NULL);
-#else
- pthread_t pthreadID;
- bool ok = (pthread_create(&pthreadID, NULL, start_routine, (void*)&threads[i].threadID) == 0);
- pthread_detach(pthreadID);
-#endif
- if (!ok)
- {
- std::cout << "Failed to create thread number " << i << std::endl;
- ::exit(EXIT_FAILURE);
- }
- }
+ // 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 terminate the threads when the program finishes
-
-void ThreadsManager::exit() {
+void ThreadPool::clear() {
- for (int i = 0; i < MAX_THREADS; i++)
- {
- // Wake up all the slave threads and wait for termination
- if (i != 0)
- {
- threads[i].do_terminate = true;
- threads[i].wake_up();
- while (threads[i].state != Thread::TERMINATED) {}
- }
-
- // Now we can safely destroy 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();
- }
-}
+ stopOnPonderhit = stop = false;
+ ponder = ponderMode;
+ Search::Limits = limits;
+ Search::RootMoves rootMoves;
+ for (const auto& m : MoveList<LEGAL>(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 <bool Fake>
-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(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 bestValue;
-
- // Pick the next available split point object from the split point stack
- SplitPoint* sp = masterThread.splitPoints + masterThread.activeSplitPoints;
-
- // Initialize the split point object
- 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++)
- sp->is_slave[i] = false;
-
- // If we are here it means we are not available
- assert(masterThread.state == Thread::SEARCHING);
-
- int workersCnt = 1; // At least the master is included
-
- // 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 (i != master && threads[i].is_available_to(master))
- {
- workersCnt++;
- sp->is_slave[i] = true;
- threads[i].splitPoint = sp;
-
- // This makes the slave to exit from idle_loop()
- threads[i].state = Thread::WORKISWAITING;
-
- if (useSleepingThreads)
- threads[i].wake_up();
- }
-
- lock_release(&threadsLock);
-
- // We failed to allocate even one slave, return
- if (!Fake && workersCnt == 1)
- return bestValue;
-
- masterThread.splitPoint = sp;
- masterThread.activeSplitPoints++;
- masterThread.state = Thread::WORKISWAITING;
-
- // 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.
- masterThread.idle_loop(sp);
-
- // 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.state == Thread::AVAILABLE);
-
- // 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.activeSplitPoints--;
-
- lock_release(&threadsLock);
-
- masterThread.splitPoint = sp->parent;
- pos.set_nodes_searched(pos.nodes_searched() + sp->nodes);
-
- return sp->bestValue;
+ main()->start_searching();
}
-
-// Explicit template instantiations
-template Value ThreadsManager::split<false>(Position&, SearchStack*, Value, Value, Value, Depth, Move, int, MovePicker*, int);
-template Value ThreadsManager::split<true>(Position&, SearchStack*, Value, Value, Value, Depth, Move, int, MovePicker*, int);