/*
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 the virtual function idle_loop().
+extern void check_time();
- long start_routine(Thread* th) { th->idle_loop(); return 0; }
+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.
+ template<typename T> T* new_thread() {
+ T* th = new T();
+ th->nativeThread = std::thread(&ThreadBase::idle_loop, th); // Will go to sleep
+ return th;
+ }
-// Thread c'tor starts a newly-created thread of execution that will call
-// the the virtual function idle_loop(), going immediately to sleep.
+ void delete_thread(ThreadBase* th) {
-Thread::Thread() : splitPoints() {
+ th->mutex.lock();
+ th->exit = true; // Search must be already finished
+ th->mutex.unlock();
- searching = exit = false;
- maxPly = splitPointsSize = 0;
- activeSplitPoint = NULL;
- idx = Threads.size();
+ th->notify_one();
+ th->nativeThread.join(); // Wait for thread termination
+ delete th;
+ }
- 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
+// ThreadBase::notify_one() wakes up the thread when there is some work to do
-Thread::~Thread() {
+void ThreadBase::notify_one() {
- exit = true; // Search must be already finished
- notify_one();
- thread_join(handle); // Wait for thread termination
+ std::unique_lock<std::mutex>(this->mutex);
+ sleepCondition.notify_one();
}
-// 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();
+// ThreadBase::wait_for() set the thread to sleep until 'condition' turns true
-void TimerThread::idle_loop() {
+void ThreadBase::wait_for(volatile const bool& condition) {
- while (!exit)
- {
- mutex.lock();
+ std::unique_lock<std::mutex> lk(mutex);
+ sleepCondition.wait(lk, [&]{ return condition; });
+}
- if (!exit)
- sleepCondition.wait_for(mutex, msec ? msec : INT_MAX);
- mutex.unlock();
+// Thread c'tor makes some init but does not launch any execution thread that
+// will be started only when c'tor returns.
- if (msec)
- check_time();
- }
-}
+Thread::Thread() /* : splitPoints() */ { // Initialization of non POD broken in MSVC
+ searching = false;
+ maxPly = 0;
+ splitPointsSize = 0;
+ activeSplitPoint = nullptr;
+ activePosition = nullptr;
+ idx = Threads.size(); // Starts from 0
+}
-// 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() {
+// Thread::cutoff_occurred() checks whether a beta cutoff has occurred in the
+// current active split point, or in some ancestor of the split point.
- while (true)
- {
- mutex.lock();
+bool Thread::cutoff_occurred() const {
- thinking = false;
+ for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint)
+ if (sp->cutoff)
+ return true;
- while (!thinking && !exit)
- {
- Threads.sleepCondition.notify_one(); // Wake up UI thread if needed
- sleepCondition.wait(mutex);
- }
+ return false;
+}
- mutex.unlock();
- if (exit)
- return;
+// 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).
- searching = true;
+bool Thread::can_join(const SplitPoint* sp) const {
- Search::think();
+ if (searching)
+ return false;
- assert(searching);
+ // 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;
- searching = false;
- }
+ // 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.test(sp->master->idx);
}
-// Thread::notify_one() wakes up the thread when there is some search to do
-
-void Thread::notify_one() {
+// 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.
- mutex.lock();
- sleepCondition.notify_one();
- mutex.unlock();
-}
+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);
-// Thread::wait_for() set the thread to sleep until condition 'b' turns true
+ // Pick and init the next available split point
+ SplitPoint& sp = splitPoints[splitPointsSize];
-void Thread::wait_for(volatile const bool& b) {
+ 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;
- mutex.lock();
- while (!b) sleepCondition.wait(mutex);
- mutex.unlock();
-}
+ // 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.spinlock.acquire();
+ sp.spinlock.acquire();
+ sp.allSlavesSearching = true; // Must be set under lock protection
+ ++splitPointsSize;
+ activeSplitPoint = &sp;
+ activePosition = nullptr;
-// Thread::cutoff_occurred() checks whether a beta cutoff has occurred in the
-// current active split point, or in some ancestor of the split point.
+ Thread* slave;
-bool Thread::cutoff_occurred() const {
+ while ( sp.slavesMask.count() < MAX_SLAVES_PER_SPLITPOINT
+ && (slave = Threads.available_slave(activeSplitPoint)) != nullptr)
+ {
+ sp.slavesMask.set(slave->idx);
+ slave->activeSplitPoint = activeSplitPoint;
+ slave->searching = true; // Slave leaves idle_loop()
+ slave->notify_one(); // Could be sleeping
+ }
- for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parent)
- if (sp->cutoff)
- return true;
+ // 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();
+ Threads.spinlock.release();
- return false;
-}
+ Thread::idle_loop(); // Force a call to base class idle_loop()
+ // 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);
-// 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 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).
+ // 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.spinlock.acquire();
+ sp.spinlock.acquire();
+
+ searching = true;
+ --splitPointsSize;
+ activeSplitPoint = sp.parentSplitPoint;
+ activePosition = &pos;
+ pos.set_nodes_searched(pos.nodes_searched() + sp.nodes);
+ *bestMove = sp.bestMove;
+ *bestValue = sp.bestValue;
-bool Thread::is_available_to(Thread* master) const {
+ sp.spinlock.release();
+ Threads.spinlock.release();
+}
- 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 size = splitPointsSize;
+// 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.
- // 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));
-}
+void TimerThread::idle_loop() {
+ while (!exit)
+ {
+ std::unique_lock<std::mutex> lk(mutex);
-// 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.
+ if (!exit)
+ sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX));
-void ThreadPool::init() {
+ lk.unlock();
- sleepWhileIdle = true;
- timer = new TimerThread();
- threads.push_back(new MainThread());
- read_uci_options();
+ if (run)
+ check_time();
+ }
}
-// exit() cleanly terminates the threads before the program exits
-
-void ThreadPool::exit() {
+// 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.
- delete timer; // As first because check_time() accesses threads data
+void MainThread::idle_loop() {
- for (size_t i = 0; i < threads.size(); i++)
- delete threads[i];
-}
+ while (!exit)
+ {
+ std::unique_lock<std::mutex> lk(mutex);
+ thinking = false;
-// 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.
+ while (!thinking && !exit)
+ {
+ Threads.sleepCondition.notify_one(); // Wake up the UI thread if needed
+ sleepCondition.wait(lk);
+ }
-void ThreadPool::read_uci_options() {
+ lk.unlock();
- maxThreadsPerSplitPoint = Options["Max Threads per Split Point"];
- minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
- size_t requested = Options["Threads"];
+ if (!exit)
+ {
+ searching = true;
- assert(requested > 0);
+ Search::think();
- while (threads.size() < requested)
- threads.push_back(new Thread());
+ assert(searching);
- while (threads.size() > requested)
- {
- delete threads.back();
- threads.pop_back();
+ searching = false;
+ }
}
}
-// slave_available() tries to find an idle thread which is available as a slave
-// for the thread 'master'.
-
-bool ThreadPool::slave_available(Thread* master) const {
+// 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.
- for (size_t i = 0; i < threads.size(); i++)
- if (threads[i]->is_available_to(master))
- return true;
+void ThreadPool::init() {
- return false;
+ timer = new_thread<TimerThread>();
+ push_back(new_thread<MainThread>());
+ read_uci_options();
}
-// 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
-// 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) {
+// ThreadPool::exit() terminates the threads before the program exits. Cannot be
+// done in d'tor because threads must be terminated before freeing us.
- assert(pos.pos_is_ok());
- assert(bestValue <= alpha && alpha < beta && beta <= VALUE_INFINITE);
- assert(bestValue > -VALUE_INFINITE);
- assert(depth >= Threads.minimumSplitDepth);
+void ThreadPool::exit() {
- Thread* master = pos.this_thread();
+ delete_thread(timer); // As first because check_time() accesses threads data
- assert(master->searching);
- assert(master->splitPointsSize < MAX_SPLITPOINTS_PER_THREAD);
+ for (Thread* th : *this)
+ delete_thread(th);
+}
- // Pick the next available split point from the split point stack
- SplitPoint& sp = master->splitPoints[master->splitPointsSize];
- sp.master = master;
- sp.parent = master->activeSplitPoint;
- 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.cutoff = false;
- sp.ss = ss;
+// 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.
- // 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.
- mutex.lock();
- sp.mutex.lock();
+void ThreadPool::read_uci_options() {
- master->splitPointsSize++;
- master->activeSplitPoint = &sp;
+ minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
+ size_t requested = Options["Threads"];
- size_t slavesCnt = 1; // Master is always included
+ assert(requested > 0);
- for (size_t i = 0; i < threads.size() && !Fake; ++i)
- if (threads[i]->is_available_to(master) && ++slavesCnt <= maxThreadsPerSplitPoint)
- {
- sp.slavesMask |= 1ULL << i;
- threads[i]->activeSplitPoint = &sp;
- threads[i]->searching = true; // Slave leaves idle_loop()
- threads[i]->notify_one(); // Could be sleeping
- }
+ // If zero (default) then set best minimum split depth automatically
+ if (!minimumSplitDepth)
+ minimumSplitDepth = requested < 8 ? 4 * ONE_PLY : 7 * ONE_PLY;
- sp.mutex.unlock();
- mutex.unlock();
+ while (size() < requested)
+ push_back(new_thread<Thread>());
- // 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)
+ while (size() > requested)
{
- master->Thread::idle_loop(); // Force a call to base class idle_loop()
-
- // 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);
+ delete_thread(back());
+ pop_back();
}
+}
- // 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::is_available_to().
- mutex.lock();
- sp.mutex.lock();
-
- master->searching = true;
- master->splitPointsSize--;
- master->activeSplitPoint = sp.parent;
- pos.set_nodes_searched(pos.nodes_searched() + sp.nodes);
- *bestMove = sp.bestMove;
- sp.mutex.unlock();
- mutex.unlock();
+// ThreadPool::available_slave() tries to find an idle thread which is available
+// to join SplitPoint 'sp'.
- return sp.bestValue;
-}
+Thread* ThreadPool::available_slave(const SplitPoint* sp) const {
-// 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);
+ for (Thread* th : *this)
+ if (th->can_join(sp))
+ return th;
+
+ return nullptr;
+}
-// wait_for_think_finished() waits for main thread to go to sleep then returns
+// ThreadPool::wait_for_think_finished() waits for main thread to finish the search
void ThreadPool::wait_for_think_finished() {
- MainThread* t = main_thread();
- t->mutex.lock();
- while (t->thinking) sleepCondition.wait(t->mutex);
- t->mutex.unlock();
+ std::unique_lock<std::mutex> lk(main()->mutex);
+ sleepCondition.wait(lk, [&]{ return !main()->thinking; });
}
-// start_thinking() 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_thinking(const Position& pos, const LimitsType& limits,
- const std::vector<Move>& searchMoves, StateStackPtr& states) {
+ StateStackPtr& states) {
wait_for_think_finished();
- SearchTime = Time::now(); // As early as possible
+ SearchTime = now(); // As early as possible
Signals.stopOnPonderhit = Signals.firstRootMove = false;
Signals.stop = Signals.failedLowAtRoot = false;
+ RootMoves.clear();
RootPos = pos;
Limits = limits;
- SetupStates = states; // Ownership transfer here
- RootMoves.clear();
+ 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()->thinking = true;
- main_thread()->notify_one(); // Starts main thread
+ main()->thinking = true;
+ main()->notify_one(); // Starts main thread
}
projects / stockfish / blobdiff