X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=c73f55660330545d300857b2874287fd54b21563;hp=031ee848e6343ea44c33835198a9411320b10a32;hb=2ca142a5b4ca200c56cb99495ec51a804983d07d;hpb=41561c9bb80a176f9fce169975fcb553340499fc
diff --git a/src/thread.cpp b/src/thread.cpp
index 031ee848..c73f5566 100644
--- a/src/thread.cpp
+++ b/src/thread.cpp
@@ -1,7 +1,7 @@
/*
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
@@ -17,169 +17,84 @@
along with this program. If not, see .
*/
+#include // For std::count
#include
-#include
#include "movegen.h"
#include "search.h"
#include "thread.h"
-#include "ucioption.h"
+#include "uci.h"
using namespace Search;
-ThreadsManager Threads; // Global object
+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. The first
- // and last thread are special. First one is the main search thread while the
- // last one mimics a timer, they run in main_loop() and timer_loop().
-
- long start_routine(Thread* th) {
-
- if (th->threadID == 0)
- th->main_loop();
-
- else if (th->threadID == MAX_THREADS)
- th->timer_loop();
-
- else
- th->idle_loop(NULL);
-
- return 0;
- }
-
-} }
-
-
-// Thread c'tor creates and launches the OS thread, that will go immediately to
-// sleep.
-
-Thread::Thread(int id) {
-
- is_searching = do_exit = false;
- maxPly = splitPointsCnt = 0;
- curSplitPoint = NULL;
- threadID = id;
- do_sleep = (id != 0); // Avoid a race with start_thinking()
-
- 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 " << id << std::endl;
- ::exit(EXIT_FAILURE);
- }
-}
-
-
-// Thread d'tor will wait 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 do_timer_event(). If maxPly is 0 thread sleeps until is woken up.
extern void check_time();
-void Thread::timer_loop() {
+namespace {
- while (!do_exit)
- {
- lock_grab(sleepLock);
- timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX);
- lock_release(sleepLock);
- check_time();
- }
-}
+ // 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.
+ template T* new_thread() {
+ std::thread* th = new T;
+ *th = std::thread(&T::idle_loop, (T*)th); // Will go to sleep
+ return (T*)th;
+ }
-// 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.
+ void delete_thread(ThreadBase* th) {
-void Thread::main_loop() {
+ th->mutex.lock();
+ th->exit = true; // Search must be already finished
+ th->mutex.unlock();
- while (true)
- {
- lock_grab(sleepLock);
-
- do_sleep = true; // Always return to sleep after a search
- is_searching = false;
+ th->notify_one();
+ th->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 lk(mutex);
+ sleepCondition.notify_one();
}
-// Thread::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.
+// 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 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
}
-// cutoff_occurred() checks whether a beta cutoff has occurred in the current
-// active split point, or in some ancestor of the split point.
+// Thread::cutoff_occurred() checks whether a beta cutoff has occurred in the
+// current active split point, or in some ancestor of the split point.
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;
@@ -187,296 +102,277 @@ bool Thread::cutoff_occurred() const {
}
-// 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::can_join() checks whether the thread is available to join the split
+// point 'sp'. 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 for the split points below his active
+// one (the "helpful master" concept in YBWC terminology).
-bool Thread::is_available_to(int master) const {
+bool Thread::can_join(const SplitPoint* sp) 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));
+ return !size || splitPoints[size - 1].slavesMask.test(sp->master->idx);
}
-// 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::read_uci_options() {
-
- maxThreadsPerSplitPoint = Options["Max Threads per Split Point"];
- minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
- useSleepingThreads = Options["Use Sleeping Threads"];
- int requested = Options["Threads"];
+// 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.spinlock.acquire(); // No contention here until we don't increment 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;
+ sp.allSlavesSearching = true; // Must be set under lock protection
+
+ ++splitPointsSize;
+ activeSplitPoint = &sp;
+ activePosition = nullptr;
+
+ // Try to allocate available threads
+ Thread* slave;
+
+ while ( sp.slavesMask.count() < MAX_SLAVES_PER_SPLITPOINT
+ && (slave = Threads.available_slave(&sp)) != nullptr)
+ {
+ slave->spinlock.acquire();
- while (size() < requested)
- threads.push_back(new Thread(size()));
+ if (slave->can_join(activeSplitPoint))
+ {
+ activeSplitPoint->slavesMask.set(slave->idx);
+ slave->activeSplitPoint = activeSplitPoint;
+ slave->searching = true;
+ }
- while (size() > requested)
- {
- delete threads.back();
- threads.pop_back();
+ slave->spinlock.release();
}
-}
+ // 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.spinlock.release();
-// wake_up() is called before a new search to start the threads that are waiting
-// on the sleep condition. If useSleepingThreads is set threads will be woken up
-// at split time.
+ Thread::idle_loop(); // Force a call to base class idle_loop()
-void ThreadsManager::wake_up() {
+ // 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);
- for (int i = 0; i < size(); i++)
- {
- threads[i]->do_sleep = false;
+ searching = true;
- if (!useSleepingThreads)
- threads[i]->wake_up();
- }
+ // We have returned from the idle loop, which means that all threads are
+ // finished. Note that decreasing splitPointsSize must be done under lock
+ // protection to avoid a race with Thread::can_join().
+ sp.spinlock.acquire();
+
+ --splitPointsSize;
+ activeSplitPoint = sp.parentSplitPoint;
+ activePosition = &pos;
+ pos.set_nodes_searched(pos.nodes_searched() + sp.nodes);
+ *bestMove = sp.bestMove;
+ *bestValue = sp.bestValue;
+
+ sp.spinlock.release();
}
-// sleep() is called after the search to ask threads to wait on sleep condition
+// 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 ThreadsManager::sleep() {
-
- for (int i = 0; i < size(); i++)
- threads[i]->do_sleep = true;
-}
+void TimerThread::idle_loop() {
+ while (!exit)
+ {
+ std::unique_lock lk(mutex);
-// init() is called during startup. Initializes locks and condition variables
-// and launches all threads sending them immediately to sleep.
+ if (!exit)
+ sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX));
-void ThreadsManager::init() {
+ lk.unlock();
- cond_init(sleepCond);
- lock_init(splitLock);
- timer = new Thread(MAX_THREADS);
- read_uci_options(); // Creates at least the main thread
+ if (run)
+ check_time();
+ }
}
-// exit() is called to cleanly terminate the threads before the program finishes
+// 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 ThreadsManager::exit() {
+void MainThread::idle_loop() {
- for (int i = 0; i < size(); i++)
- delete threads[i];
+ while (!exit)
+ {
+ std::unique_lock lk(mutex);
- delete timer;
- lock_destroy(splitLock);
- cond_destroy(sleepCond);
-}
+ thinking = false;
+ while (!thinking && !exit)
+ {
+ sleepCondition.notify_one(); // Wake up the UI thread if needed
+ sleepCondition.wait(lk);
+ }
-// available_slave_exists() tries to find an idle thread which is available as
-// a slave for the thread with threadID 'master'.
+ lk.unlock();
-bool ThreadsManager::available_slave_exists(int master) const {
+ if (!exit)
+ {
+ searching = true;
- assert(master >= 0 && master < size());
+ Search::think();
- for (int i = 0; i < size(); i++)
- if (threads[i]->is_available_to(master))
- return true;
+ assert(searching);
- return false;
+ searching = 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 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
-Value ThreadsManager::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);
-
- int master = pos.thread();
- Thread& masterThread = *threads[master];
-
- if (masterThread.splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD)
- return bestValue;
-
- // Pick the next available split point from the split point stack
- SplitPoint* sp = &masterThread.splitPoints[masterThread.splitPointsCnt++];
-
- sp->parent = masterThread.curSplitPoint;
- sp->master = master;
- sp->cutoff = false;
- sp->slavesMask = 1ULL << master;
- 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;
-
- assert(masterThread.is_searching);
-
- masterThread.curSplitPoint = sp;
- int slavesCnt = 0;
-
- // 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);
-
- for (int 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()
-
- if (useSleepingThreads)
- threads[i]->wake_up();
+// MainThread::join() waits for main thread to finish the search
- if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included
- break;
- }
-
- lock_release(splitLock);
- lock_release(sp->lock);
+void MainThread::join() {
- // 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.
- // We pass the split point as a parameter to the idle loop, which means that
- // the thread will return from the idle loop when all slaves have finished
- // their work at this split point.
- if (slavesCnt || Fake)
- {
- masterThread.idle_loop(sp);
-
- // 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(!masterThread.is_searching);
- }
+ std::unique_lock lk(mutex);
+ sleepCondition.wait(lk, [&]{ return !thinking; });
+}
- // 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);
- masterThread.is_searching = true;
- masterThread.splitPointsCnt--;
- masterThread.curSplitPoint = sp->parent;
- pos.set_nodes_searched(pos.nodes_searched() + sp->nodes);
- *bestMove = sp->bestMove;
+// 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.
- lock_release(splitLock);
- lock_release(sp->lock);
+void ThreadPool::init() {
- return sp->bestValue;
+ timer = new_thread();
+ push_back(new_thread());
+ read_uci_options();
}
-// Explicit template instantiations
-template Value ThreadsManager::split(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker*, int);
-template Value ThreadsManager::split(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker*, int);
+// ThreadPool::exit() terminates the threads before the program exits. Cannot be
+// done in d'tor because threads must be terminated before freeing us.
-// ThreadsManager::set_timer() is used to set the timer to trigger after msec
-// milliseconds. If msec is 0 then timer is stopped.
+void ThreadPool::exit() {
-void ThreadsManager::set_timer(int msec) {
+ delete_thread(timer); // As first because check_time() accesses threads data
+ timer = nullptr;
- lock_grab(timer->sleepLock);
- timer->maxPly = msec;
- cond_signal(timer->sleepCond); // Wake up and restart the timer
- lock_release(timer->sleepLock);
+ for (Thread* th : *this)
+ delete_thread(th);
+
+ clear(); // Get rid of stale pointers
}
-// ThreadsManager::start_thinking() is used by UI thread to wake up the main
-// thread parked in main_loop() and starting a new search. If asyncMode is true
-// then function returns immediately, otherwise caller is blocked waiting for
-// the search to finish.
+// 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.
-void ThreadsManager::start_thinking(const Position& pos, const LimitsType& limits,
- const std::set& searchMoves, bool async) {
- Thread& main = *threads.front();
+void ThreadPool::read_uci_options() {
- lock_grab(main.sleepLock);
+ minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
+ size_t requested = Options["Threads"];
- // Wait main thread has finished before to launch a new search
- while (!main.do_sleep)
- cond_wait(sleepCond, main.sleepLock);
+ assert(requested > 0);
- // Copy input arguments to initialize the search
- RootPosition.copy(pos, 0);
- Limits = limits;
- RootMoves.clear();
+ // If zero (default) then set best minimum split depth automatically
+ if (!minimumSplitDepth)
+ minimumSplitDepth = 5 * ONE_PLY ;
- // Populate RootMoves with all the legal moves (default) or, if a searchMoves
- // set is given, with the subset of legal moves to search.
- for (MoveList ml(pos); !ml.end(); ++ml)
- if (searchMoves.empty() || searchMoves.count(ml.move()))
- RootMoves.push_back(RootMove(ml.move()));
+ while (size() < requested)
+ push_back(new_thread());
- // Reset signals before to start the new search
- Signals.stopOnPonderhit = Signals.firstRootMove = false;
- Signals.stop = Signals.failedLowAtRoot = false;
+ while (size() > requested)
+ {
+ delete_thread(back());
+ pop_back();
+ }
+}
- main.do_sleep = false;
- cond_signal(main.sleepCond); // Wake up main thread and start searching
- if (!async)
- while (!main.do_sleep)
- cond_wait(sleepCond, main.sleepLock);
+// ThreadPool::available_slave() tries to find an idle thread which is available
+// to join SplitPoint 'sp'.
- lock_release(main.sleepLock);
-}
+Thread* ThreadPool::available_slave(const SplitPoint* sp) const {
+ for (Thread* th : *this)
+ if (th->can_join(sp))
+ return th;
-// ThreadsManager::stop_thinking() is used by UI thread to raise a stop request
-// and to wait for the main thread finishing the search. Needed to wait exiting
-// and terminate the threads after a 'quit' command.
+ return nullptr;
+}
-void ThreadsManager::stop_thinking() {
- Thread& main = *threads.front();
+// ThreadPool::start_thinking() wakes up the main thread sleeping in
+// MainThread::idle_loop() and starts a new search, then returns immediately.
- Search::Signals.stop = true;
+void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits,
+ StateStackPtr& states) {
+ main()->join();
- lock_grab(main.sleepLock);
+ Signals.stopOnPonderhit = Signals.firstRootMove = false;
+ Signals.stop = Signals.failedLowAtRoot = false;
- cond_signal(main.sleepCond); // In case is waiting for stop or ponderhit
+ 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());
+ }
- while (!main.do_sleep)
- cond_wait(sleepCond, main.sleepLock);
+ for (const auto& m : MoveList(pos))
+ if ( limits.searchmoves.empty()
+ || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m))
+ RootMoves.push_back(RootMove(m));
- lock_release(main.sleepLock);
+ main()->thinking = true;
+ main()->notify_one(); // Wake up main thread: 'thinking' must be already set
}