X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=0a8bacf7d09acebf34d40145697338b8fb970a10;hp=a8e26b9f85c43a69a3f0260db1313285ca3ccee2;hb=0b36ba74fc0a80388cac43a35962ffc73c01b071;hpb=86b95f210508de4c30fb5ee9f86efee6641d45f8
diff --git a/src/thread.cpp b/src/thread.cpp
index a8e26b9f..ba5d13d5 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) 2008-2010 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
@@ -17,342 +17,361 @@
along with this program. If not, see .
*/
-#include
+#include // For std::count
+#include
+#include "movegen.h"
+#include "search.h"
#include "thread.h"
-#include "ucioption.h"
+#include "uci.h"
-ThreadsManager Threads; // Global object definition
+using namespace Search;
-namespace { extern "C" {
+ThreadPool Threads; // Global object
- // 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.
+extern void check_time();
-#if defined(_MSC_VER)
+namespace {
- DWORD WINAPI start_routine(LPVOID threadID) {
+ // 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.
- Threads[*(int*)threadID].idle_loop(NULL);
- return 0;
- }
+ template T* new_thread() {
+ T* th = new T();
+ th->nativeThread = std::thread(&ThreadBase::idle_loop, th); // Will go to sleep
+ return th;
+ }
-#else
+ void delete_thread(ThreadBase* th) {
- void* start_routine(void* threadID) {
+ th->mutex.lock();
+ th->exit = true; // Search must be already finished
+ th->mutex.unlock();
- Threads[*(int*)threadID].idle_loop(NULL);
- return NULL;
- }
+ th->notify_one();
+ th->nativeThread.join(); // Wait for thread termination
+ delete th;
+ }
+
+}
-#endif
-} }
+// ThreadBase::notify_one() wakes up the thread when there is some work to do
+void ThreadBase::notify_one() {
+
+ std::unique_lock(this->mutex);
+ sleepCondition.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.
-void Thread::wake_up() {
+// ThreadBase::wait_for() set the thread to sleep until 'condition' turns true
- lock_grab(&sleepLock);
- cond_signal(&sleepCond);
- lock_release(&sleepLock);
+void ThreadBase::wait_for(volatile const bool& condition) {
+
+ std::unique_lock lk(mutex);
+ sleepCondition.wait(lk, [&]{ return condition; });
}
-// 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 makes some init but does not launch any execution thread that
+// will be started only when c'tor returns.
+
+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
+}
+
+
+// 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 = splitPoint; sp; sp = sp->parent)
- if (sp->is_betaCutoff)
+ for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint)
+ if (sp->cutoff)
return true;
+
return 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::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 slave's split point
+// stack (the "helpful master concept" in YBWC terminology).
-bool Thread::is_available_to(int master) const {
+bool Thread::available_to(const Thread* master) const {
- if (state != AVAILABLE)
+ 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 localActiveSplitPoints = activeSplitPoints;
+ // 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.
- if ( !localActiveSplitPoints
- || splitPoints[localActiveSplitPoints - 1].is_slave[master])
- return true;
-
- return false;
+ return !size || splitPoints[size - 1].slavesMask.test(master->idx);
}
-// 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::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 ThreadsManager::read_uci_options() {
-
- maxThreadsPerSplitPoint = Options["Maximum Number of Threads per Split Point"].value();
- minimumSplitDepth = Options["Minimum Split Depth"].value() * ONE_PLY;
- useSleepingThreads = Options["Use Sleeping Threads"].value();
+ // 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();
- set_size(Options["Threads"].value());
-}
+ sp.allSlavesSearching = true; // Must be set under lock protection
+ ++splitPointsSize;
+ activeSplitPoint = &sp;
+ activePosition = nullptr;
+ Thread* slave;
-// set_size() changes the number of active threads and raises do_sleep flag for
-// all the unused threads that will go immediately to sleep.
+ 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
+ }
-void ThreadsManager::set_size(int cnt) {
+ // 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();
- assert(cnt > 0 && cnt <= MAX_THREADS);
+ Thread::idle_loop(); // Force a call to base class idle_loop()
- activeThreads = cnt;
+ // 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 < MAX_THREADS; i++)
- threads[i].do_sleep = !(i < activeThreads);
+ // 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;
+
+ sp.spinlock.release();
+ Threads.spinlock.release();
}
-// init() is called during startup. Initializes locks and condition variables
-// and launches all threads sending them immediately to sleep.
-
-void ThreadsManager::init() {
+// 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.
- // Threads will go to sleep as soon as created, only main thread is kept alive
- set_size(1);
- threads[0].state = Thread::SEARCHING;
- threads[0].threadID = 0;
+void TimerThread::idle_loop() {
- // Allocate pawn and material hash tables for main thread
- init_hash_tables();
+ while (!exit)
+ {
+ std::unique_lock lk(mutex);
- // Initialize threads lock, used when allocating slaves during splitting
- lock_init(&threadsLock);
+ if (!exit)
+ sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX));
- // Initialize sleep and split point locks
- for (int i = 0; i < MAX_THREADS; i++)
- {
- lock_init(&threads[i].sleepLock);
- cond_init(&threads[i].sleepCond);
+ lk.unlock();
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_init(&(threads[i].splitPoints[j].lock));
+ if (run)
+ check_time();
}
+}
+
- // Create and startup all the threads but the main that is already running
- for (int i = 1; i < MAX_THREADS; i++)
+// 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 MainThread::idle_loop() {
+
+ while (!exit)
{
- threads[i].state = Thread::AVAILABLE;
- threads[i].threadID = i;
+ std::unique_lock lk(mutex);
-#if defined(_MSC_VER)
- threads[i].handle = CreateThread(NULL, 0, start_routine, (LPVOID)&threads[i].threadID, 0, NULL);
- bool ok = (threads[i].handle != NULL);
-#else
- bool ok = (pthread_create(&threads[i].handle, NULL, start_routine, (void*)&threads[i].threadID) == 0);
-#endif
+ thinking = false;
- if (!ok)
+ while (!thinking && !exit)
{
- std::cout << "Failed to create thread number " << i << std::endl;
- ::exit(EXIT_FAILURE);
+ Threads.sleepCondition.notify_one(); // Wake up the UI thread if needed
+ sleepCondition.wait(lk);
}
- }
-}
+ lk.unlock();
-// exit() is called to cleanly terminate the threads when the program finishes
+ if (!exit)
+ {
+ searching = true;
-void ThreadsManager::exit() {
+ Search::think();
- 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);
- pthread_detach(threads[i].handle);
-#endif
+ assert(searching);
+
+ searching = false;
}
+ }
+}
- // 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));
- }
+// 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() {
- lock_destroy(&threadsLock);
+ timer = new_thread();
+ push_back(new_thread());
+ read_uci_options();
}
-// 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.
+// ThreadPool::exit() terminates the threads before the program exits. Cannot be
+// done in d'tor because threads must be terminated before freeing us.
-void ThreadsManager::init_hash_tables() {
+void ThreadPool::exit() {
- for (int i = 0; i < activeThreads; i++)
- {
- threads[i].pawnTable.init();
- threads[i].materialTable.init();
- }
+ delete_thread(timer); // As first because check_time() accesses threads data
+
+ for (Thread* th : *this)
+ delete_thread(th);
}
-// available_slave_exists() tries to find an idle thread which is available as
-// a slave for the thread with threadID "master".
+// 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.
-bool ThreadsManager::available_slave_exists(int master) const {
+void ThreadPool::read_uci_options() {
- assert(master >= 0 && master < activeThreads);
+ minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
+ size_t requested = Options["Threads"];
- for (int i = 0; i < activeThreads; i++)
- if (i != master && threads[i].is_available_to(master))
- return true;
+ assert(requested > 0);
- return false;
+ // If zero (default) then set best minimum split depth automatically
+ if (!minimumSplitDepth)
+ minimumSplitDepth = requested < 8 ? 4 * ONE_PLY : 7 * ONE_PLY;
+
+ while (size() < requested)
+ push_back(new_thread());
+
+ while (size() > requested)
+ {
+ delete_thread(back());
+ pop_back();
+ }
}
-// 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.
-
-template
-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
+// ThreadPool::available_slave() tries to find an idle thread which is available
+// as a slave for the thread 'master'.
- // 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);
+Thread* ThreadPool::available_slave(const Thread* master) const {
- 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;
+ for (Thread* th : *this)
+ if (th->available_to(master))
+ return th;
+
+ return nullptr;
+}
- // This makes the slave to exit from idle_loop()
- threads[i].state = Thread::WORKISWAITING;
- if (useSleepingThreads)
- threads[i].wake_up();
- }
+// ThreadPool::wait_for_think_finished() waits for main thread to finish the search
- lock_release(&threadsLock);
+void ThreadPool::wait_for_think_finished() {
- // We failed to allocate even one slave, return
- if (!Fake && workersCnt == 1)
- return bestValue;
+ std::unique_lock lk(main()->mutex);
+ sleepCondition.wait(lk, [&]{ return !main()->thinking; });
+}
- masterThread.splitPoint = sp;
- masterThread.activeSplitPoints++;
- masterThread.state = Thread::WORKISWAITING;
- // 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);
+// ThreadPool::start_thinking() wakes up the main thread sleeping in
+// MainThread::idle_loop() and starts a new search, then returns immediately.
- // 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);
+void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits,
+ StateStackPtr& states) {
+ wait_for_think_finished();
- // 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);
+ SearchTime = now(); // As early as possible
- masterThread.state = Thread::SEARCHING;
- masterThread.activeSplitPoints--;
+ Signals.stopOnPonderhit = Signals.firstRootMove = false;
+ Signals.stop = Signals.failedLowAtRoot = false;
- lock_release(&threadsLock);
+ 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());
+ }
- masterThread.splitPoint = sp->parent;
- pos.set_nodes_searched(pos.nodes_searched() + sp->nodes);
+ for (const auto& m : MoveList(pos))
+ if ( limits.searchmoves.empty()
+ || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m))
+ RootMoves.push_back(RootMove(m));
- return sp->bestValue;
+ main()->thinking = true;
+ main()->notify_one(); // Starts main thread
}
-
-// Explicit template instantiations
-template Value ThreadsManager::split(Position&, SearchStack*, Value, Value, Value, Depth, Move, int, MovePicker*, int);
-template Value ThreadsManager::split(Position&, SearchStack*, Value, Value, Value, Depth, Move, int, MovePicker*, int);