/*
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) 2008-2012 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 <cassert>
#include <iostream>
+#include "movegen.h"
+#include "search.h"
#include "thread.h"
#include "ucioption.h"
-ThreadsManager Threads; // Global object definition
+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 simply calls idle_loop() of the supplied thread.
- // 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)
+ long start_routine(Thread* th) { th->idle_loop(); return 0; }
- DWORD WINAPI start_routine(LPVOID thread) {
+} }
- ((Thread*)thread)->idle_loop(NULL);
- return 0;
- }
-#else
+// Thread c'tor starts a newly-created thread of execution that will call
+// the the virtual function idle_loop(), going immediately to sleep.
- void* start_routine(void* thread) {
+Thread::Thread() : splitPoints() {
- ((Thread*)thread)->idle_loop(NULL);
- return NULL;
+ searching = exit = false;
+ maxPly = splitPointsCnt = 0;
+ curSplitPoint = NULL;
+ idx = Threads.size();
+
+ if (!thread_create(handle, start_routine, this))
+ {
+ std::cerr << "Failed to create thread number " << idx << std::endl;
+ ::exit(EXIT_FAILURE);
}
+}
-#endif
-} }
+// Thread d'tor waits for thread termination before to return
+
+Thread::~Thread() {
+
+ exit = true; // Search must be already finished
+ notify_one();
+ thread_join(handle); // Wait for thread termination
+}
+
+// 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();
-// 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.
+void TimerThread::idle_loop() {
-void Thread::wake_up() {
+ while (!exit)
+ {
+ mutex.lock();
+
+ if (!exit)
+ sleepCondition.wait_for(mutex, msec ? msec : INT_MAX);
+
+ mutex.unlock();
- lock_grab(&sleepLock);
- cond_signal(&sleepCond);
- lock_release(&sleepLock);
+ if (msec)
+ check_time();
+ }
}
-// cutoff_occurred() checks whether a beta cutoff has occurred in the current
-// active split point, or in some ancestor of the split point.
+// 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.
-bool Thread::cutoff_occurred() const {
+void MainThread::idle_loop() {
- for (SplitPoint* sp = splitPoint; sp; sp = sp->parent)
- if (sp->is_betaCutoff)
- return true;
- return false;
+ while (true)
+ {
+ mutex.lock();
+
+ thinking = false;
+
+ while (!thinking && !exit)
+ {
+ Threads.sleepCondition.notify_one(); // Wake up UI thread if needed
+ sleepCondition.wait(mutex);
+ }
+
+ mutex.unlock();
+
+ if (exit)
+ return;
+
+ searching = true;
+
+ Search::think();
+
+ assert(searching);
+
+ searching = false;
+ }
}
-// 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).
+// Thread::notify_one() wakes up the thread when there is some search to do
-bool Thread::is_available_to(int master) const {
+void Thread::notify_one() {
- if (state != AVAILABLE)
- return false;
+ mutex.lock();
+ sleepCondition.notify_one();
+ mutex.unlock();
+}
- // 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;
- // 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;
+// Thread::wait_for() set the thread to sleep until condition 'b' turns true
- return false;
+void Thread::wait_for(volatile const bool& b) {
+
+ mutex.lock();
+ while (!b) sleepCondition.wait(mutex);
+ mutex.unlock();
}
-// 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::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::read_uci_options() {
+bool Thread::cutoff_occurred() const {
- 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>();
+ for (SplitPoint* sp = curSplitPoint; sp; sp = sp->parent)
+ if (sp->cutoff)
+ return true;
- set_size(Options["Threads"].value<int>());
+ return false;
}
-// 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::is_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).
-void ThreadsManager::set_size(int cnt) {
+bool Thread::is_available_to(Thread* master) const {
- assert(cnt > 0 && cnt <= MAX_THREADS);
+ if (searching)
+ return false;
- activeThreads = cnt;
+ // 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;
- for (int i = 0; i < MAX_THREADS; i++)
- if (i < activeThreads)
- {
- // Dynamically allocate 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 could
- // even result in a crash.
- threads[i].pawnTable.init();
- threads[i].materialTable.init();
-
- threads[i].do_sleep = false;
- }
- else
- threads[i].do_sleep = true;
+ // 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.
+ return !spCnt || (splitPoints[spCnt - 1].slavesMask & (1ULL << master->idx));
}
-// init() is called during startup. Initializes locks and condition variables
-// and launches all threads sending them immediately to sleep.
-
-void ThreadsManager::init() {
+// 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.
- // Initialize threads lock, used when allocating slaves during splitting
- lock_init(&threadsLock);
+void ThreadPool::init() {
- // Initialize sleep and split point locks
- for (int i = 0; i < MAX_THREADS; i++)
- {
- lock_init(&threads[i].sleepLock);
- cond_init(&threads[i].sleepCond);
+ sleepWhileIdle = true;
+ timer = new TimerThread();
+ threads.push_back(new MainThread());
+ read_uci_options();
+}
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_init(&(threads[i].splitPoints[j].lock));
- }
- // Initialize main thread's associated data
- threads[0].state = Thread::SEARCHING;
- threads[0].threadID = 0;
- set_size(1); // This makes all the threads but the main to go to sleep
+// exit() cleanly terminates the threads before the program exits.
- // Create and launch all the threads but the main that is already running,
- // threads will go immediately to sleep.
- for (int i = 1; i < MAX_THREADS; i++)
- {
- threads[i].state = Thread::AVAILABLE;
- threads[i].threadID = i;
+void ThreadPool::exit() {
-#if defined(_MSC_VER)
- threads[i].handle = CreateThread(NULL, 0, start_routine, (LPVOID)&threads[i], 0, NULL);
- bool ok = (threads[i].handle != NULL);
-#else
- bool ok = (pthread_create(&threads[i].handle, NULL, start_routine, (void*)&threads[i]) == 0);
-#endif
+ delete timer; // As first becuase check_time() accesses threads data
- if (!ok)
- {
- std::cerr << "Failed to create thread number " << i << std::endl;
- ::exit(EXIT_FAILURE);
- }
- }
+ for (size_t i = 0; i < threads.size(); i++)
+ delete threads[i];
}
-// exit() is called to cleanly terminate the threads when the program finishes
+// 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::exit() {
+void ThreadPool::read_uci_options() {
- 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();
-
-#if defined(_MSC_VER)
- WaitForSingleObject(threads[i].handle, 0);
- CloseHandle(threads[i].handle);
-#else
- pthread_join(threads[i].handle, NULL);
-#endif
- }
+ maxThreadsPerSplitPoint = Options["Max Threads per Split Point"];
+ minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
+ size_t requested = Options["Threads"];
- // Now we can safely destroy locks and wait conditions
- lock_destroy(&threads[i].sleepLock);
- cond_destroy(&threads[i].sleepCond);
+ assert(requested > 0);
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_destroy(&(threads[i].splitPoints[j].lock));
- }
+ while (threads.size() < requested)
+ threads.push_back(new Thread());
- lock_destroy(&threadsLock);
+ while (threads.size() > requested)
+ {
+ delete threads.back();
+ threads.pop_back();
+ }
}
// available_slave_exists() tries to find an idle thread which is available as
-// a slave for the thread with threadID "master".
-
-bool ThreadsManager::available_slave_exists(int master) const {
+// a slave for the thread 'master'.
- assert(master >= 0 && master < activeThreads);
+bool ThreadPool::available_slave_exists(Thread* master) const {
- for (int i = 0; i < activeThreads; i++)
- if (i != master && threads[i].is_available_to(master))
+ for (size_t i = 0; i < threads.size(); i++)
+ if (threads[i]->is_available_to(master))
return true;
return false;
// 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, 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 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);
+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);
- assert(pos.thread() >= 0 && pos.thread() < activeThreads);
- assert(activeThreads > 1);
- int i, master = pos.thread();
- Thread& masterThread = threads[master];
+ Thread* master = pos.this_thread();
- // If we already have too many active split points, don't split
- if (masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS)
+ if (master->splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD)
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
+ // 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 = ∓
+ sp.moveCount = moveCount;
+ sp.pos = &pos;
+ sp.nodes = 0;
+ sp.ss = ss;
+
+ assert(master->searching);
+
+ master->curSplitPoint = &sp;
+ int slavesCnt = 0;
// 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);
+ // is_searching flag. This must be done under lock protection to avoid concurrent
+ // allocation of the same slave by another master.
+ mutex.lock();
+ sp.mutex.lock();
- for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++)
- if (i != master && threads[i].is_available_to(master))
+ for (size_t i = 0; i < threads.size() && !Fake; ++i)
+ if (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::SEARCHING;
+ sp.slavesMask |= 1ULL << i;
+ threads[i]->curSplitPoint = &sp;
+ threads[i]->searching = true; // Slave leaves idle_loop()
+ threads[i]->notify_one(); // Could be sleeping
- if (useSleepingThreads)
- threads[i].wake_up();
+ if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included
+ break;
}
- lock_release(&threadsLock);
-
- // We failed to allocate even one slave, return
- if (!Fake && workersCnt == 1)
- return bestValue;
+ master->splitPointsCnt++;
- masterThread.splitPoint = sp;
- masterThread.activeSplitPoints++;
+ sp.mutex.unlock();
+ mutex.unlock();
- // 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);
+ // 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->Thread::idle_loop(); // Force a call to base class idle_loop()
- // 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);
+ // 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->searching);
+ }
// 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);
+ // finished. Note that setting is_searching and decreasing splitPointsCnt is
+ // done under lock protection to avoid a race with Thread::is_available_to().
+ mutex.lock();
+ sp.mutex.lock();
- masterThread.state = Thread::SEARCHING;
- masterThread.activeSplitPoints--;
+ master->searching = true;
+ master->splitPointsCnt--;
+ master->curSplitPoint = sp.parent;
+ pos.set_nodes_searched(pos.nodes_searched() + sp.nodes);
+ *bestMove = sp.bestMove;
- lock_release(&threadsLock);
+ sp.mutex.unlock();
+ mutex.unlock();
- masterThread.splitPoint = sp->parent;
- pos.set_nodes_searched(pos.nodes_searched() + sp->nodes);
-
- return sp->bestValue;
+ return sp.bestValue;
}
// 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);
+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);
+
+
+// wait_for_think_finished() waits for main thread to go to sleep then returns
+
+void ThreadPool::wait_for_think_finished() {
+
+ MainThread* t = main_thread();
+ t->mutex.lock();
+ while (t->thinking) sleepCondition.wait(t->mutex);
+ t->mutex.unlock();
+}
+
+
+// start_thinking() wakes up the main thread sleeping in main_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;
+
+ RootPos = pos;
+ Limits = limits;
+ SetupStates = states; // Ownership transfer here
+ RootMoves.clear();
+
+ for (MoveList<LEGAL> ml(pos); !ml.end(); ++ml)
+ if (searchMoves.empty() || count(searchMoves.begin(), searchMoves.end(), ml.move()))
+ RootMoves.push_back(RootMove(ml.move()));
+
+ main_thread()->thinking = true;
+ main_thread()->notify_one(); // Starts main thread
+}
projects / stockfish / blobdiff