/*
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) 2008-2013 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 <iostream>
+#include <algorithm> // For std::count
+#include <cassert>
+#include "movegen.h"
+#include "search.h"
#include "thread.h"
#include "ucioption.h"
-ThreadsManager Threads; // Global object definition
+using namespace Search;
-namespace { extern "C" {
+ThreadPool Threads; // Global object
+
+namespace {
// 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.
+ // is launched. It is a wrapper to the virtual function idle_loop().
-#if defined(_MSC_VER)
+ extern "C" { long start_routine(ThreadBase* th) { th->idle_loop(); return 0; } }
- DWORD WINAPI start_routine(LPVOID threadID) {
- Threads[*(int*)threadID].idle_loop(NULL);
- return 0;
- }
+ // Helpers to launch a thread after creation and joining before delete. Must be
+ // outside Thread c'tor and d'tor because object shall be fully initialized
+ // when start_routine (and hence virtual idle_loop) is called and when joining.
-#else
+ template<typename T> T* new_thread() {
+ T* th = new T();
+ thread_create(th->handle, start_routine, th); // Will go to sleep
+ return th;
+ }
- void* start_routine(void* threadID) {
+ void delete_thread(ThreadBase* th) {
+ th->exit = true; // Search must be already finished
+ th->notify_one();
+ thread_join(th->handle); // Wait for thread termination
+ delete th;
+ }
- Threads[*(int*)threadID].idle_loop(NULL);
- return NULL;
- }
+}
-#endif
-} }
+// ThreadBase::notify_one() wakes up the thread when there is some work to do
+void ThreadBase::notify_one() {
-// 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.
+ mutex.lock();
+ sleepCondition.notify_one();
+ mutex.unlock();
+}
-void Thread::wake_up() {
- lock_grab(&sleepLock);
- cond_signal(&sleepCond);
- lock_release(&sleepLock);
+// ThreadBase::wait_for() set the thread to sleep until condition 'b' turns true
+
+void ThreadBase::wait_for(volatile const bool& b) {
+
+ mutex.lock();
+ while (!b) sleepCondition.wait(mutex);
+ mutex.unlock();
}
-// 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 c'tor just inits data but does not launch any thread of execution that
+// instead will be started only upon c'tor returns.
-bool Thread::cutoff_occurred() const {
+Thread::Thread() /* : splitPoints() */ { // Value-initialization bug in MSVC
- for (SplitPoint* sp = splitPoint; sp; sp = sp->parent)
- if (sp->is_betaCutoff)
- return true;
- return false;
+ searching = false;
+ maxPly = splitPointsSize = 0;
+ activeSplitPoint = NULL;
+ activePosition = NULL;
+ idx = Threads.size();
}
-// 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).
+// TimerThread::idle_loop() is where the timer thread waits msec milliseconds
+// and then calls check_time(). If msec is 0 thread sleeps until is woken up.
+extern void check_time();
-bool Thread::is_available_to(int master) const {
+void TimerThread::idle_loop() {
- if (state != AVAILABLE)
- return false;
+ while (!exit)
+ {
+ mutex.lock();
- // 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;
+ if (!exit)
+ sleepCondition.wait_for(mutex, msec ? msec : INT_MAX);
- // 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;
+ mutex.unlock();
- return false;
+ if (msec)
+ check_time();
+ }
}
-// 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.
+// MainThread::idle_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.
+
+void MainThread::idle_loop() {
-void ThreadsManager::read_uci_options() {
+ while (true)
+ {
+ mutex.lock();
- 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>();
+ thinking = false;
- set_size(Options["Threads"].value<int>());
-}
+ while (!thinking && !exit)
+ {
+ Threads.sleepCondition.notify_one(); // Wake up UI thread if needed
+ sleepCondition.wait(mutex);
+ }
+ mutex.unlock();
-// set_size() changes the number of active threads and raises do_sleep flag for
-// all the unused threads that will go immediately to sleep.
+ if (exit)
+ return;
-void ThreadsManager::set_size(int cnt) {
+ searching = true;
- assert(cnt > 0 && cnt <= MAX_THREADS);
+ Search::think();
- activeThreads = cnt;
+ assert(searching);
- for (int i = 0; i < MAX_THREADS; i++)
- threads[i].do_sleep = !(i < activeThreads);
+ searching = false;
+ }
}
-// init() is called during startup. Initializes locks and condition variables
-// and launches all threads sending them immediately to sleep.
+// Thread::cutoff_occurred() checks whether a beta cutoff has occurred in the
+// current active split point, or in some ancestor of the split point.
-void ThreadsManager::init() {
+bool Thread::cutoff_occurred() const {
- // Threads will sent to sleep as soon as created, only main thread is kept alive
- set_size(1);
- threads[0].state = Thread::SEARCHING;
- threads[0].threadID = 0;
+ for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint)
+ if (sp->cutoff)
+ return true;
- // Allocate pawn and material hash tables for main thread
- init_hash_tables();
+ return false;
+}
- // 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++)
- {
- lock_init(&threads[i].sleepLock);
- cond_init(&threads[i].sleepCond);
+// 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 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).
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_init(&(threads[i].splitPoints[j].lock));
- }
+bool Thread::available_to(const Thread* master) const {
- // 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;
- 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);
- }
+ if (searching)
+ return false;
- // Wait until the thread has finished launching and is gone to sleep
- while (threads[i].state == Thread::INITIALIZING) {}
- }
+ // 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 size = splitPointsSize;
+
+ // 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 !size || (splitPoints[size - 1].slavesMask & (1ULL << master->idx));
}
-// exit() is called to cleanly terminate the threads when the program finishes
+// init() is called at startup to create and launch requested threads, that will
+// go immediately to sleep due to 'sleepWhileIdle' set to true. 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.
-void ThreadsManager::exit() {
+void ThreadPool::init() {
- 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) {}
- }
+ sleepWhileIdle = true;
+ timer = new_thread<TimerThread>();
+ push_back(new_thread<MainThread>());
+ read_uci_options();
+}
- // 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));
- }
+// exit() cleanly terminates the threads before the program exits
+
+void ThreadPool::exit() {
+
+ delete_thread(timer); // As first because check_time() accesses threads data
- lock_destroy(&threadsLock);
+ for (iterator it = begin(); it != end(); ++it)
+ delete_thread(*it);
}
-// 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.
+// 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.
-void ThreadsManager::init_hash_tables() {
+void ThreadPool::read_uci_options() {
- for (int i = 0; i < activeThreads; i++)
+ maxThreadsPerSplitPoint = Options["Max Threads per Split Point"];
+ minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
+ size_t requested = Options["Threads"];
+
+ assert(requested > 0);
+
+ // Value 0 has a special meaning: We determine the optimal minimum split depth
+ // automatically. Anyhow the minimumSplitDepth should never be under 4 plies.
+ if (!minimumSplitDepth)
+ minimumSplitDepth = (requested < 8 ? 4 : 7) * ONE_PLY;
+ else
+ minimumSplitDepth = std::max(4 * ONE_PLY, minimumSplitDepth);
+
+ while (size() < requested)
+ push_back(new_thread<Thread>());
+
+ while (size() > requested)
{
- threads[i].pawnTable.init();
- threads[i].materialTable.init();
+ delete_thread(back());
+ pop_back();
}
}
-// available_slave_exists() tries to find an idle thread which is available as
-// a slave for the thread with threadID "master".
+// slave_available() tries to find an idle thread which is available as a slave
+// for the thread 'master'.
-bool ThreadsManager::available_slave_exists(int master) const {
+Thread* ThreadPool::available_slave(const Thread* master) const {
- assert(master >= 0 && master < activeThreads);
+ for (const_iterator it = begin(); it != end(); ++it)
+ if ((*it)->available_to(master))
+ return *it;
- for (int i = 0; i < activeThreads; i++)
- if (i != master && threads[i].is_available_to(master))
- return true;
-
- return false;
+ return NULL;
}
// 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.
+// 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
+// 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 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
+void Thread::split(Position& pos, const Stack* ss, Value alpha, Value beta, Value* bestValue,
+ Move* bestMove, Depth depth, Move threatMove, int moveCount,
+ MovePicker* movePicker, int nodeType, bool cutNode) {
+
+ assert(pos.pos_is_ok());
+ assert(*bestValue <= alpha && alpha < beta && beta <= VALUE_INFINITE);
+ assert(*bestValue > -VALUE_INFINITE);
+ assert(depth >= Threads.minimumSplitDepth);
+ assert(searching);
+ assert(splitPointsSize < MAX_SPLITPOINTS_PER_THREAD);
+
+ // Pick the next available split point from the split point stack
+ SplitPoint& sp = splitPoints[splitPointsSize];
+
+ sp.masterThread = this;
+ sp.parentSplitPoint = activeSplitPoint;
+ sp.slavesMask = 1ULL << idx;
+ sp.depth = depth;
+ sp.bestValue = *bestValue;
+ sp.bestMove = *bestMove;
+ sp.threatMove = threatMove;
+ 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;
// 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);
+ // '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();
- 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;
+ ++splitPointsSize;
+ activeSplitPoint = &sp;
+ activePosition = NULL;
- // This makes the slave to exit from idle_loop()
- threads[i].state = Thread::WORKISWAITING;
+ size_t slavesCnt = 1; // This thread is always included
+ Thread* slave;
- if (useSleepingThreads)
- threads[i].wake_up();
- }
+ while ( (slave = Threads.available_slave(this)) != NULL
+ && ++slavesCnt <= Threads.maxThreadsPerSplitPoint && !Fake)
+ {
+ sp.slavesMask |= 1ULL << slave->idx;
+ slave->activeSplitPoint = &sp;
+ slave->searching = true; // Slave leaves idle_loop()
+ slave->notify_one(); // Could be sleeping
+ }
- lock_release(&threadsLock);
+ // 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.
+ if (slavesCnt > 1 || Fake)
+ {
+ sp.mutex.unlock();
+ Threads.mutex.unlock();
- // We failed to allocate even one slave, return
- if (!Fake && workersCnt == 1)
- return bestValue;
+ Thread::idle_loop(); // Force a call to base class idle_loop()
- masterThread.splitPoint = sp;
- masterThread.activeSplitPoints++;
- masterThread.state = Thread::WORKISWAITING;
+ // 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(!searching);
+ assert(!activePosition);
- // 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);
+ // We have returned from the idle loop, which means that all threads are
+ // finished. Note that setting 'searching' and decreasing splitPointsSize is
+ // done under lock protection to avoid a race with Thread::available_to().
+ Threads.mutex.lock();
+ sp.mutex.lock();
+ }
- // 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);
+ searching = true;
+ --splitPointsSize;
+ activeSplitPoint = sp.parentSplitPoint;
+ activePosition = &pos;
+ pos.set_nodes_searched(pos.nodes_searched() + sp.nodes);
+ *bestMove = sp.bestMove;
+ *bestValue = sp.bestValue;
- // 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);
+ sp.mutex.unlock();
+ Threads.mutex.unlock();
+}
+
+// Explicit template instantiations
+template void Thread::split<false>(Position&, const Stack*, Value, Value, Value*, Move*, Depth, Move, int, MovePicker*, int, bool);
+template void Thread::split< true>(Position&, const Stack*, Value, Value, Value*, Move*, Depth, Move, int, MovePicker*, int, bool);
- masterThread.state = Thread::SEARCHING;
- masterThread.activeSplitPoints--;
- lock_release(&threadsLock);
+// wait_for_think_finished() waits for main thread to go to sleep then returns
- masterThread.splitPoint = sp->parent;
- pos.set_nodes_searched(pos.nodes_searched() + sp->nodes);
+void ThreadPool::wait_for_think_finished() {
- return sp->bestValue;
+ MainThread* t = main();
+ t->mutex.lock();
+ while (t->thinking) sleepCondition.wait(t->mutex);
+ t->mutex.unlock();
}
-// 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);
+
+// start_thinking() wakes up the main thread sleeping in MainThread::idle_loop()
+// so to start a new search, then returns immediately.
+
+void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits,
+ const std::vector<Move>& searchMoves, StateStackPtr& states) {
+ wait_for_think_finished();
+
+ SearchTime = Time::now(); // As early as possible
+
+ Signals.stopOnPonderhit = Signals.firstRootMove = false;
+ Signals.stop = Signals.failedLowAtRoot = false;
+
+ RootMoves.clear();
+ RootPos = pos;
+ Limits = limits;
+ if (states.get()) // If we don't set a new position, preserve current state
+ {
+ SetupStates = states; // Ownership transfer here
+ assert(!states.get());
+ }
+
+ for (MoveList<LEGAL> it(pos); *it; ++it)
+ if ( searchMoves.empty()
+ || std::count(searchMoves.begin(), searchMoves.end(), *it))
+ RootMoves.push_back(RootMove(*it));
+
+ main()->thinking = true;
+ main()->notify_one(); // Starts main thread
+}