/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2008 Tord Romstad (Glaurung author)
- Copyright (C) 2008-2012 Marco Costalba, Joona Kiiski, Tord Romstad
+ Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, 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 <algorithm> // For std::count
#include <cassert>
-#include <iostream>
#include "movegen.h"
#include "search.h"
#include "thread.h"
-#include "ucioption.h"
+#include "uci.h"
using namespace Search;
ThreadPool Threads; // Global object
-namespace { extern "C" {
-
- // start_routine() is the C function which is called when a new thread
- // is launched. It is a wrapper to member function pointed by start_fn.
-
- long start_routine(Thread* th) { (th->*(th->start_fn))(); return 0; }
-
-} }
-
-
-// Thread c'tor starts a newly-created thread of execution that will call
-// the idle loop function pointed by start_fn going immediately to sleep.
-
-Thread::Thread(Fn fn) {
-
- is_searching = do_exit = false;
- maxPly = splitPointsCnt = 0;
- curSplitPoint = NULL;
- start_fn = fn;
- idx = Threads.size();
-
- do_sleep = (fn != &Thread::main_loop); // Avoid a race with start_searching()
-
- lock_init(sleepLock);
- cond_init(sleepCond);
-
- for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++)
- lock_init(splitPoints[j].lock);
-
- if (!thread_create(handle, start_routine, this))
- {
- std::cerr << "Failed to create thread number " << idx << std::endl;
- ::exit(EXIT_FAILURE);
- }
-}
-
-
-// Thread d'tor waits for thread termination before to return.
-
-Thread::~Thread() {
-
- assert(do_sleep);
-
- do_exit = true; // Search must be already finished
- wake_up();
-
- thread_join(handle); // Wait for thread termination
-
- lock_destroy(sleepLock);
- cond_destroy(sleepCond);
-
- for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++)
- lock_destroy(splitPoints[j].lock);
-}
-
-
-// Thread::timer_loop() is where the timer thread waits maxPly milliseconds and
-// then calls check_time(). If maxPly is 0 thread sleeps until is woken up.
extern void check_time();
-void Thread::timer_loop() {
-
- while (!do_exit)
- {
- lock_grab(sleepLock);
- timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX);
- lock_release(sleepLock);
- check_time();
- }
-}
+namespace {
+ // Helpers to launch a thread after creation and joining before delete. Must be
+ // outside Thread c'tor and d'tor because the object must be fully initialized
+ // when start_routine (and hence virtual idle_loop) is called and when joining.
-// Thread::main_loop() is where the main thread is parked waiting to be started
-// when there is a new search. Main thread will launch all the slave threads.
+ template<typename T> T* new_thread() {
+ T* th = new T();
+ th->nativeThread = std::thread(&ThreadBase::idle_loop, th); // Will go to sleep
+ return th;
+ }
-void Thread::main_loop() {
+ void delete_thread(ThreadBase* th) {
- while (true)
- {
- lock_grab(sleepLock);
+ th->mutex.lock();
+ th->exit = true; // Search must be already finished
+ th->mutex.unlock();
- do_sleep = true; // Always return to sleep after a search
- is_searching = false;
+ th->notify_one();
+ th->nativeThread.join(); // Wait for thread termination
+ delete th;
+ }
- while (do_sleep && !do_exit)
- {
- cond_signal(Threads.sleepCond); // Wake up UI thread if needed
- cond_wait(sleepCond, sleepLock);
- }
+}
- lock_release(sleepLock);
- if (do_exit)
- return;
+// ThreadBase::notify_one() wakes up the thread when there is some work to do
- is_searching = true;
+void ThreadBase::notify_one() {
- Search::think();
- }
+ std::unique_lock<std::mutex>(this->mutex);
+ sleepCondition.notify_one();
}
-// Thread::wake_up() wakes up the thread, normally at the beginning of the search
-// or, if "sleeping threads" is used at split time.
+// ThreadBase::wait_for() set the thread to sleep until 'condition' turns true
-void Thread::wake_up() {
+void ThreadBase::wait_for(volatile const bool& condition) {
- lock_grab(sleepLock);
- cond_signal(sleepCond);
- lock_release(sleepLock);
+ std::unique_lock<std::mutex> lk(mutex);
+ sleepCondition.wait(lk, [&]{ return condition; });
}
-// Thread::wait_for_stop_or_ponderhit() is called when the maximum depth is
-// reached while the program is pondering. The point is to work around a wrinkle
-// in the UCI protocol: When pondering, the engine is not allowed to give a
-// "bestmove" before the GUI sends it a "stop" or "ponderhit" command. We simply
-// wait here until one of these commands (that raise StopRequest) is sent and
-// then return, after which the bestmove and pondermove will be printed.
+// Thread c'tor makes some init but does not launch any execution thread that
+// will be started only when c'tor returns.
-void Thread::wait_for_stop_or_ponderhit() {
+Thread::Thread() /* : splitPoints() */ { // Initialization of non POD broken in MSVC
- Signals.stopOnPonderhit = true;
-
- lock_grab(sleepLock);
- while (!Signals.stop) cond_wait(sleepCond, sleepLock);
- lock_release(sleepLock);
+ searching = false;
+ maxPly = 0;
+ splitPointsSize = 0;
+ activeSplitPoint = nullptr;
+ activePosition = nullptr;
+ idx = Threads.size(); // Starts from 0
}
bool Thread::cutoff_occurred() const {
- for (SplitPoint* sp = curSplitPoint; sp; sp = sp->parent)
+ for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint)
if (sp->cutoff)
return true;
}
-// Thread::is_available_to() checks whether the thread is available to help the
+// Thread::available_to() checks whether the thread is available to help the
// thread 'master' at a split point. An obvious requirement is that thread must
// be idle. With more than two threads, this is not sufficient: If the thread is
-// the master of some active split point, it is only available as a slave to the
-// slaves which are busy searching the split point at the top of slaves split
-// point stack (the "helpful master concept" in YBWC terminology).
+// the master of some split point, it is only available as a slave to the slaves
+// 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(Thread* master) const {
+bool Thread::available_to(const Thread* master) const {
- if (is_searching)
+ if (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 spCnt = splitPointsCnt;
+ // Make a local copy to be sure it doesn't become zero under our feet while
+ // testing next condition and so leading to an out of bounds access.
+ const size_t size = splitPointsSize;
- // No active split points means that the thread is available as a slave for any
+ // No split points means that the thread is available as a slave for any
// other thread otherwise apply the "helpful master" concept if possible.
- return !spCnt || (splitPoints[spCnt - 1].slavesMask & (1ULL << master->idx));
+ return !size || splitPoints[size - 1].slavesMask.test(master->idx);
}
-// init() is called at startup. Initializes lock and condition variable and
-// launches requested threads sending them immediately to sleep. We cannot use
-// a c'tor becuase Threads is a static object and we need a fully initialized
-// engine at this point due to allocation of endgames in Thread c'tor.
+// Thread::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), 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 then helper threads are
+// informed 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.
+
+void Thread::split(Position& pos, Stack* ss, Value alpha, Value beta, Value* bestValue,
+ Move* bestMove, Depth depth, int moveCount,
+ MovePicker* movePicker, int nodeType, bool cutNode) {
+
+ assert(searching);
+ assert(-VALUE_INFINITE < *bestValue && *bestValue <= alpha && alpha < beta && beta <= VALUE_INFINITE);
+ assert(depth >= Threads.minimumSplitDepth);
+ assert(splitPointsSize < MAX_SPLITPOINTS_PER_THREAD);
+
+ // Pick and init the next available split point
+ SplitPoint& sp = splitPoints[splitPointsSize];
+
+ sp.master = this;
+ sp.parentSplitPoint = activeSplitPoint;
+ sp.slavesMask = 0, sp.slavesMask.set(idx);
+ sp.depth = depth;
+ sp.bestValue = *bestValue;
+ sp.bestMove = *bestMove;
+ sp.alpha = alpha;
+ sp.beta = beta;
+ sp.nodeType = nodeType;
+ sp.cutNode = cutNode;
+ sp.movePicker = movePicker;
+ sp.moveCount = moveCount;
+ sp.pos = &pos;
+ sp.nodes = 0;
+ sp.cutoff = false;
+ sp.ss = ss;
-void ThreadPool::init() {
+ // Try to allocate available threads and ask them to start searching setting
+ // 'searching' flag. This must be done under lock protection to avoid concurrent
+ // allocation of the same slave by another master.
+ Threads.mutex.lock();
+ sp.mutex.lock();
- cond_init(sleepCond);
- lock_init(splitLock);
- timer = new Thread(&Thread::timer_loop);
- threads.push_back(new Thread(&Thread::main_loop));
- read_uci_options();
-}
+ sp.allSlavesSearching = true; // Must be set under lock protection
+ ++splitPointsSize;
+ activeSplitPoint = &sp;
+ activePosition = nullptr;
+
+ Thread* slave;
+ while ( sp.slavesMask.count() < MAX_SLAVES_PER_SPLITPOINT
+ && (slave = Threads.available_slave(this)) != nullptr)
+ {
+ sp.slavesMask.set(slave->idx);
+ slave->activeSplitPoint = &sp;
+ slave->searching = true; // Slave leaves idle_loop()
+ slave->notify_one(); // Could be sleeping
+ }
-// d'tor cleanly terminates the threads when the program exits.
+ // Everything is set up. The master thread enters the idle loop, from which
+ // it will instantly launch a search, because its 'searching' flag is set.
+ // The thread will return from the idle loop when all slaves have finished
+ // their work at this split point.
+ sp.mutex.unlock();
+ Threads.mutex.unlock();
-ThreadPool::~ThreadPool() {
+ Thread::idle_loop(); // Force a call to base class idle_loop()
- for (size_t i = 0; i < size(); i++)
- delete threads[i];
+ // In the helpful master concept, a master can help only a sub-tree of its
+ // split point and because everything is finished here, it's not possible
+ // for the master to be booked.
+ assert(!searching);
+ assert(!activePosition);
- delete timer;
- lock_destroy(splitLock);
- cond_destroy(sleepCond);
+ // We have returned from the idle loop, which means that all threads are
+ // finished. Note that setting 'searching' and decreasing splitPointsSize must
+ // be done under lock protection to avoid a race with Thread::available_to().
+ Threads.mutex.lock();
+ sp.mutex.lock();
+
+ searching = true;
+ --splitPointsSize;
+ activeSplitPoint = sp.parentSplitPoint;
+ activePosition = &pos;
+ pos.set_nodes_searched(pos.nodes_searched() + sp.nodes);
+ *bestMove = sp.bestMove;
+ *bestValue = sp.bestValue;
+
+ sp.mutex.unlock();
+ Threads.mutex.unlock();
}
-// read_uci_options() updates internal threads parameters from the corresponding
-// UCI options and creates/destroys threads to match the requested number. Thread
-// objects are dynamically allocated to avoid creating in advance all possible
-// threads, with included pawns and material tables, if only few are used.
+// TimerThread::idle_loop() is where the timer thread waits Resolution milliseconds
+// and then calls check_time(). When not searching, thread sleeps until it's woken up.
-void ThreadPool::read_uci_options() {
+void TimerThread::idle_loop() {
- maxThreadsPerSplitPoint = Options["Max Threads per Split Point"];
- minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
- useSleepingThreads = Options["Use Sleeping Threads"];
- size_t requested = Options["Threads"];
+ while (!exit)
+ {
+ std::unique_lock<std::mutex> lk(mutex);
- assert(requested > 0);
+ if (!exit)
+ sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX));
- while (size() < requested)
- threads.push_back(new Thread(&Thread::idle_loop));
+ lk.unlock();
- while (size() > requested)
- {
- delete threads.back();
- threads.pop_back();
+ if (run)
+ check_time();
}
}
-// wake_up() is called before a new search to start the threads that are waiting
-// on the sleep condition and to reset maxPly. When useSleepingThreads is set
-// threads will be woken up at split time.
+// MainThread::idle_loop() is where the main thread is parked waiting to be started
+// when there is a new search. The main thread will launch all the slave threads.
-void ThreadPool::wake_up() const {
+void MainThread::idle_loop() {
- for (size_t i = 0; i < size(); i++)
+ while (!exit)
{
- threads[i]->maxPly = 0;
- threads[i]->do_sleep = false;
+ std::unique_lock<std::mutex> lk(mutex);
- if (!useSleepingThreads)
- threads[i]->wake_up();
- }
-}
+ thinking = false;
+ while (!thinking && !exit)
+ {
+ Threads.sleepCondition.notify_one(); // Wake up the UI thread if needed
+ sleepCondition.wait(lk);
+ }
-// sleep() is called after the search finishes to ask all the threads but the
-// main one to go waiting on a sleep condition.
-
-void ThreadPool::sleep() const {
+ lk.unlock();
- for (size_t i = 1; i < size(); i++) // Main thread will go to sleep by itself
- threads[i]->do_sleep = true; // to avoid a race with start_searching()
-}
+ if (!exit)
+ {
+ searching = true;
+ Search::think();
-// available_slave_exists() tries to find an idle thread which is available as
-// a slave for the thread 'master'.
+ assert(searching);
-bool ThreadPool::available_slave_exists(Thread* master) const {
+ searching = false;
+ }
+ }
+}
- for (size_t i = 0; i < size(); i++)
- if (threads[i]->is_available_to(master))
- return true;
- return false;
-}
+// ThreadPool::init() is called at startup to create and launch requested threads,
+// that will go immediately to sleep. We cannot use a c'tor because Threads is a
+// static object and we need a fully initialized engine at this point due to
+// allocation of Endgames in Thread c'tor.
+void ThreadPool::init() {
-// 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 then helper threads are told 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 ThreadPool::split(Position& pos, Stack* ss, Value alpha, Value beta,
- Value bestValue, Move* bestMove, Depth depth,
- Move threatMove, int moveCount, MovePicker* mp, int nodeType) {
-
- assert(pos.pos_is_ok());
- assert(bestValue > -VALUE_INFINITE);
- assert(bestValue <= alpha);
- assert(alpha < beta);
- assert(beta <= VALUE_INFINITE);
- assert(depth > DEPTH_ZERO);
-
- Thread* master = pos.this_thread();
-
- if (master->splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD)
- return bestValue;
-
- // Pick the next available split point from the split point stack
- SplitPoint& sp = master->splitPoints[master->splitPointsCnt];
-
- sp.parent = master->curSplitPoint;
- sp.master = master;
- sp.cutoff = false;
- sp.slavesMask = 1ULL << master->idx;
- sp.depth = depth;
- sp.bestMove = *bestMove;
- 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;
+ timer = new_thread<TimerThread>();
+ push_back(new_thread<MainThread>());
+ read_uci_options();
+}
- assert(master->is_searching);
- master->curSplitPoint = &sp;
- int slavesCnt = 0;
+// ThreadPool::exit() terminates the threads before the program exits. Cannot be
+// done in d'tor because threads must be terminated before freeing us.
- // Try to allocate available threads and ask them to start searching setting
- // is_searching flag. This must be done under lock protection to avoid concurrent
- // allocation of the same slave by another master.
- lock_grab(sp.lock);
- lock_grab(splitLock);
+void ThreadPool::exit() {
- for (size_t i = 0; i < size() && !Fake; ++i)
- if (threads[i]->is_available_to(master))
- {
- sp.slavesMask |= 1ULL << i;
- threads[i]->curSplitPoint = &sp;
- threads[i]->is_searching = true; // Slave leaves idle_loop()
+ delete_thread(timer); // As first because check_time() accesses threads data
- if (useSleepingThreads)
- threads[i]->wake_up();
+ for (Thread* th : *this)
+ delete_thread(th);
+}
- if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included
- break;
- }
- master->splitPointsCnt++;
+// ThreadPool::read_uci_options() updates internal threads parameters from the
+// corresponding UCI options and creates/destroys threads to match the requested
+// number. Thread objects are dynamically allocated to avoid creating all possible
+// threads in advance (which include pawns and material tables), even if only a
+// few are to be used.
- lock_release(splitLock);
- lock_release(sp.lock);
+void ThreadPool::read_uci_options() {
- // 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.
- // The thread will return from the idle loop when all slaves have finished
- // their work at this split point.
- if (slavesCnt || Fake)
- {
- master->idle_loop();
+ minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
+ size_t requested = Options["Threads"];
- // In helpful master concept a master can help only a sub-tree of its split
- // point, and because here is all finished is not possible master is booked.
- assert(!master->is_searching);
- }
+ assert(requested > 0);
- // We have returned from the idle loop, which means that all threads are
- // finished. Note that setting is_searching and decreasing splitPointsCnt is
- // done under lock protection to avoid a race with Thread::is_available_to().
- lock_grab(sp.lock); // To protect sp.nodes
- lock_grab(splitLock);
-
- master->is_searching = true;
- master->splitPointsCnt--;
- master->curSplitPoint = sp.parent;
- pos.set_nodes_searched(pos.nodes_searched() + sp.nodes);
- *bestMove = sp.bestMove;
+ // If zero (default) then set best minimum split depth automatically
+ if (!minimumSplitDepth)
+ minimumSplitDepth = requested < 8 ? 4 * ONE_PLY : 7 * ONE_PLY;
- lock_release(splitLock);
- lock_release(sp.lock);
+ while (size() < requested)
+ push_back(new_thread<Thread>());
- return sp.bestValue;
+ while (size() > requested)
+ {
+ delete_thread(back());
+ pop_back();
+ }
}
-// Explicit template instantiations
-template Value ThreadPool::split<false>(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker*, int);
-template Value ThreadPool::split<true>(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker*, int);
+// ThreadPool::available_slave() tries to find an idle thread which is available
+// as a slave for the thread 'master'.
-// set_timer() is used to set the timer to trigger after msec milliseconds.
-// If msec is 0 then timer is stopped.
+Thread* ThreadPool::available_slave(const Thread* master) const {
-void ThreadPool::set_timer(int msec) {
+ for (Thread* th : *this)
+ if (th->available_to(master))
+ return th;
- lock_grab(timer->sleepLock);
- timer->maxPly = msec;
- cond_signal(timer->sleepCond); // Wake up and restart the timer
- lock_release(timer->sleepLock);
+ return nullptr;
}
-// wait_for_search_finished() waits for main thread to go to sleep, this means
-// search is finished. Then returns.
+// ThreadPool::wait_for_think_finished() waits for main thread to finish the search
-void ThreadPool::wait_for_search_finished() {
+void ThreadPool::wait_for_think_finished() {
- Thread* t = main_thread();
- lock_grab(t->sleepLock);
- cond_signal(t->sleepCond); // In case is waiting for stop or ponderhit
- while (!t->do_sleep) cond_wait(sleepCond, t->sleepLock);
- lock_release(t->sleepLock);
+ std::unique_lock<std::mutex> lk(main()->mutex);
+ sleepCondition.wait(lk, [&]{ return !main()->thinking; });
}
-// start_searching() wakes up the main thread sleeping in main_loop() so to start
-// a new search, then returns immediately.
+// ThreadPool::start_thinking() wakes up the main thread sleeping in
+// MainThread::idle_loop() and starts a new search, then returns immediately.
-void ThreadPool::start_searching(const Position& pos, const LimitsType& limits,
- const std::vector<Move>& searchMoves) {
- wait_for_search_finished();
+void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits,
+ StateStackPtr& states) {
+ wait_for_think_finished();
- SearchTime.restart(); // As early as possible
+ SearchTime = Time::now(); // As early as possible
Signals.stopOnPonderhit = Signals.firstRootMove = false;
Signals.stop = Signals.failedLowAtRoot = false;
- RootPosition = pos;
- Limits = limits;
RootMoves.clear();
+ RootPos = pos;
+ Limits = limits;
+ if (states.get()) // If we don't set a new position, preserve current state
+ {
+ SetupStates = std::move(states); // Ownership transfer here
+ assert(!states.get());
+ }
- for (MoveList<LEGAL> ml(pos); !ml.end(); ++ml)
- if (searchMoves.empty() || count(searchMoves.begin(), searchMoves.end(), ml.move()))
- RootMoves.push_back(RootMove(ml.move()));
+ for (const auto& m : MoveList<LEGAL>(pos))
+ if ( limits.searchmoves.empty()
+ || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m))
+ RootMoves.push_back(RootMove(m));
- main_thread()->do_sleep = false;
- main_thread()->wake_up();
+ main()->thinking = true;
+ main()->notify_one(); // Starts main thread
}