X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=faa3a5716aa1570945c5dcaea7cb345ca0e75455;hp=4ef145781511aacbf61930844dcf5153b8942ba4;hb=5a2d525048ca5eec1d5d0effc8868cb2fdc701bb;hpb=ffa150bec3af540fbaff57fdbb657940225e69e6
diff --git a/src/thread.cpp b/src/thread.cpp
index 4ef14578..faa3a571 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) 2008-2012 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
@@ -17,211 +17,277 @@
along with this program. If not, see .
*/
+#include
#include
+#include "movegen.h"
+#include "search.h"
#include "thread.h"
#include "ucioption.h"
-ThreadsManager Threads; // Global object definition
+using namespace Search;
+
+ThreadsManager 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() 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 member function pointed by start_fn.
+
+ long start_routine(Thread* th) {
+
+ Threads.set_this_thread(th); // Save pointer into thread local storage
+ (th->*(th->start_fn))();
+ return 0;
+ }
+
+} }
+
+// Thread c'tor starts a newly-created thread of execution that will call
+// the idle loop function pointed by start_fn going immediately to sleep.
-#if defined(_MSC_VER)
+Thread::Thread(Fn fn) {
- DWORD WINAPI start_routine(LPVOID threadID) {
+ is_searching = do_exit = false;
+ maxPly = splitPointsCnt = 0;
+ curSplitPoint = NULL;
+ start_fn = fn;
+ idx = Threads.size();
- Threads.idle_loop(*(int*)threadID, NULL);
- return 0;
+ do_sleep = (fn != &Thread::main_loop); // Avoid a race with start_searching()
+
+ 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 " << idx << std::endl;
+ ::exit(EXIT_FAILURE);
}
+}
+
-#else
+// Thread d'tor waits for thread termination before to return.
- void* start_routine(void* threadID) {
+Thread::~Thread() {
- Threads.idle_loop(*(int*)threadID, NULL);
- return NULL;
+ 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 check_time(). If maxPly is 0 thread sleeps until is woken up.
+extern void check_time();
+
+void Thread::timer_loop() {
+
+ while (!do_exit)
+ {
+ lock_grab(sleepLock);
+ timed_wait(sleepCond, sleepLock, maxPly ? maxPly : INT_MAX);
+ lock_release(sleepLock);
+ check_time();
}
+}
-#endif
-} }
+// 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 Thread::main_loop() {
+
+ while (true)
+ {
+ lock_grab(sleepLock);
+
+ do_sleep = true; // Always return to sleep after a search
+ is_searching = false;
+
+ 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;
-// 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.
+ is_searching = true;
+
+ Search::think();
+ }
+}
+
+
+// Thread::wake_up() wakes up the thread, normally at the beginning of the search
+// or, if "sleeping threads" is used at split time.
void Thread::wake_up() {
- lock_grab(&sleepLock);
- cond_signal(&sleepCond);
- lock_release(&sleepLock);
+ lock_grab(sleepLock);
+ cond_signal(sleepCond);
+ lock_release(sleepLock);
}
-// 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::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.
+
+void Thread::wait_for_stop_or_ponderhit() {
+
+ Signals.stopOnPonderhit = true;
+
+ lock_grab(sleepLock);
+ while (!Signals.stop) cond_wait(sleepCond, sleepLock);
+ lock_release(sleepLock);
+}
+
+
+// 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 = curSplitPoint; sp; sp = sp->parent)
+ 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
+// 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).
-bool Thread::is_available_to(int master) const {
+bool Thread::is_available_to(Thread* master) const {
- if (state != AVAILABLE)
+ if (is_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;
+ int spCnt = splitPointsCnt;
// 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;
-
- return false;
+ return !spCnt || (splitPoints[spCnt - 1].slavesMask & (1ULL << 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.
+// 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.
-void ThreadsManager::read_uci_options() {
+void ThreadsManager::init() {
- maxThreadsPerSplitPoint = Options["Maximum Number of Threads per Split Point"].value();
- minimumSplitDepth = Options["Minimum Split Depth"].value() * ONE_PLY;
- useSleepingThreads = Options["Use Sleeping Threads"].value();
- activeThreads = Options["Threads"].value();
+ tls_init(tlsKey);
+ cond_init(sleepCond);
+ lock_init(splitLock);
+ timer = new Thread(&Thread::timer_loop);
+ threads.push_back(new Thread(&Thread::main_loop));
+ set_this_thread(main_thread()); // Use main thread's resources
+ read_uci_options();
}
-// init() is called during startup. Initializes locks and condition variables
-// and launches all threads sending them immediately to sleep.
+// d'tor cleanly terminates the threads when the program exits.
-void ThreadsManager::init() {
+ThreadsManager::~ThreadsManager() {
- int threadID[MAX_THREADS];
+ for (int i = 0; i < size(); i++)
+ delete threads[i];
- // This flag is needed to properly end the threads when program exits
- allThreadsShouldExit = false;
+ delete timer;
+ lock_destroy(splitLock);
+ cond_destroy(sleepCond);
+ tls_destroy(tlsKey);
+}
- // Threads will sent to sleep as soon as created, only main thread is kept alive
- activeThreads = 1;
- threads[0].state = Thread::SEARCHING;
- // Allocate pawn and material hash tables for main thread
- init_hash_tables();
+// 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.
- lock_init(&mpLock);
+void ThreadsManager::read_uci_options() {
- // Initialize thread and split point locks
- for (int i = 0; i < MAX_THREADS; i++)
- {
- lock_init(&threads[i].sleepLock);
- cond_init(&threads[i].sleepCond);
+ maxThreadsPerSplitPoint = Options["Max Threads per Split Point"];
+ minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
+ useSleepingThreads = Options["Use Sleeping Threads"];
+ int requested = Options["Threads"];
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_init(&(threads[i].splitPoints[j].lock));
- }
+ assert(requested > 0);
- // 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;
- threadID[i] = i;
-
-#if defined(_MSC_VER)
- bool ok = (CreateThread(NULL, 0, start_routine, (LPVOID)&threadID[i], 0, NULL) != NULL);
-#else
- pthread_t pthreadID;
- bool ok = (pthread_create(&pthreadID, NULL, start_routine, (void*)&threadID[i]) == 0);
- pthread_detach(pthreadID);
-#endif
- if (!ok)
- {
- std::cout << "Failed to create thread number " << i << std::endl;
- ::exit(EXIT_FAILURE);
- }
+ while (size() < requested)
+ threads.push_back(new Thread(&Thread::idle_loop));
- // Wait until the thread has finished launching and is gone to sleep
- while (threads[i].state == Thread::INITIALIZING) {}
+ while (size() > requested)
+ {
+ delete threads.back();
+ threads.pop_back();
}
}
-// exit() is called to cleanly exit the threads when the program finishes
-
-void ThreadsManager::exit() {
+// wake_up() is called before a new search to start the threads that are waiting
+// on the sleep condition and to reset maxPly. When useSleepingThreads is set
+// threads will be woken up at split time.
- // Force the woken up threads to exit idle_loop() and hence terminate
- allThreadsShouldExit = true;
+void ThreadsManager::wake_up() const {
- for (int i = 0; i < MAX_THREADS; i++)
+ for (int i = 0; i < size(); i++)
{
- // Wake up all the threads and waits for termination
- if (i != 0)
- {
- threads[i].wake_up();
- while (threads[i].state != Thread::TERMINATED) {}
- }
-
- // Now we can safely destroy the locks and wait conditions
- lock_destroy(&threads[i].sleepLock);
- cond_destroy(&threads[i].sleepCond);
+ threads[i]->maxPly = 0;
+ threads[i]->do_sleep = false;
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_destroy(&(threads[i].splitPoints[j].lock));
+ if (!useSleepingThreads)
+ threads[i]->wake_up();
}
-
- lock_destroy(&mpLock);
}
-// 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.
+// sleep() is called after the search finishes to ask all the threads but the
+// main one to go waiting on a sleep condition.
-void ThreadsManager::init_hash_tables() {
+void ThreadsManager::sleep() const {
- for (int i = 0; i < activeThreads; i++)
- {
- threads[i].pawnTable.init();
- threads[i].materialTable.init();
- }
+ for (int i = 1; i < size(); i++) // Main thread will go to sleep by itself
+ threads[i]->do_sleep = true; // to avoid a race with start_searching()
}
// available_slave_exists() tries to find an idle thread which is available as
-// a slave for the thread with threadID "master".
+// a slave for the thread 'master'.
-bool ThreadsManager::available_slave_exists(int master) const {
+bool ThreadsManager::available_slave_exists(Thread* master) const {
- assert(master >= 0 && master < activeThreads);
-
- for (int i = 0; i < activeThreads; i++)
- if (i != master && threads[i].is_available_to(master))
+ for (int i = 0; i < size(); i++)
+ if (threads[i]->is_available_to(master))
return true;
return false;
@@ -229,117 +295,160 @@ bool ThreadsManager::available_slave_exists(int master) const {
// 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
-void ThreadsManager::split(Position& pos, SearchStack* ss, Value* alpha, const 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);
+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);
- assert(pos.thread() >= 0 && pos.thread() < activeThreads);
- assert(activeThreads > 1);
- int i, master = pos.thread();
- Thread& masterThread = threads[master];
+ Thread* master = this_thread();
+
+ if (master->splitPointsCnt >= MAX_SPLITPOINTS_PER_THREAD)
+ return bestValue;
+
+ // 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 = mp;
+ sp->moveCount = moveCount;
+ sp->pos = &pos;
+ sp->nodes = 0;
+ sp->ss = ss;
+
+ assert(master->is_searching);
+
+ master->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();
+
+ if (++slavesCnt + 1 >= maxThreadsPerSplitPoint) // Master is always included
+ break;
+ }
- lock_grab(&mpLock);
+ lock_release(splitLock);
+ lock_release(sp->lock);
- // If no other thread is available to help us, or if we have too many
- // active split points, don't split.
- if ( !available_slave_exists(master)
- || masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS)
+ // 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)
{
- lock_release(&mpLock);
- return;
+ master->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(!master->is_searching);
}
- // Pick the next available split point object from the split point stack
- SplitPoint& splitPoint = masterThread.splitPoints[masterThread.activeSplitPoints++];
-
- // Initialize the split point object
- splitPoint.parent = masterThread.splitPoint;
- splitPoint.master = master;
- splitPoint.is_betaCutoff = false;
- splitPoint.depth = depth;
- splitPoint.threatMove = threatMove;
- splitPoint.alpha = *alpha;
- splitPoint.beta = beta;
- splitPoint.nodeType = nodeType;
- splitPoint.bestValue = *bestValue;
- splitPoint.mp = mp;
- splitPoint.moveCount = moveCount;
- splitPoint.pos = &pos;
- splitPoint.nodes = 0;
- splitPoint.ss = ss;
- for (i = 0; i < activeThreads; i++)
- splitPoint.is_slave[i] = false;
-
- masterThread.splitPoint = &splitPoint;
-
- // If we are here it means we are not available
- assert(masterThread.state != Thread::AVAILABLE);
-
- int workersCnt = 1; // At least the master is included
-
- // Allocate available threads setting state to THREAD_BOOKED
- for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++)
- if (i != master && threads[i].is_available_to(master))
- {
- threads[i].state = Thread::BOOKED;
- threads[i].splitPoint = &splitPoint;
- splitPoint.is_slave[i] = true;
- workersCnt++;
- }
+ // 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);
- assert(Fake || workersCnt > 1);
+ master->is_searching = true;
+ master->splitPointsCnt--;
+ master->curSplitPoint = sp->parent;
+ pos.set_nodes_searched(pos.nodes_searched() + sp->nodes);
+ *bestMove = sp->bestMove;
- // We can release the lock because slave threads are already booked and master is not available
- lock_release(&mpLock);
+ lock_release(splitLock);
+ lock_release(sp->lock);
- // Tell the threads that they have work to do. This will make them leave
- // their idle loop.
- for (i = 0; i < activeThreads; i++)
- if (i == master || splitPoint.is_slave[i])
- {
- assert(i == master || threads[i].state == Thread::BOOKED);
+ return sp->bestValue;
+}
- threads[i].state = Thread::WORKISWAITING; // This makes the slave to exit from idle_loop()
+// 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);
- if (useSleepingThreads && i != master)
- threads[i].wake_up();
- }
- // 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.
- idle_loop(master, &splitPoint);
+// ThreadsManager::set_timer() is used to set the timer to trigger after msec
+// milliseconds. If msec is 0 then timer is stopped.
- // We have returned from the idle loop, which means that all threads are
- // finished. Update alpha and bestValue, and return.
- lock_grab(&mpLock);
+void ThreadsManager::set_timer(int msec) {
+
+ lock_grab(timer->sleepLock);
+ timer->maxPly = msec;
+ cond_signal(timer->sleepCond); // Wake up and restart the timer
+ lock_release(timer->sleepLock);
+}
- *alpha = splitPoint.alpha;
- *bestValue = splitPoint.bestValue;
- masterThread.activeSplitPoints--;
- masterThread.splitPoint = splitPoint.parent;
- pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes);
- lock_release(&mpLock);
+// ThreadsManager::wait_for_search_finished() waits for main thread to go to
+// sleep, this means search is finished. Then returns.
+
+void ThreadsManager::wait_for_search_finished() {
+
+ Thread* t = main_thread();
+ lock_grab(t->sleepLock);
+ cond_signal(t->sleepCond); // In case is waiting for stop or ponderhit
+ while (!t->do_sleep) cond_wait(sleepCond, t->sleepLock);
+ lock_release(t->sleepLock);
}
-// Explicit template instantiations
-template void ThreadsManager::split(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, int);
-template void ThreadsManager::split(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, int);
+
+// ThreadsManager::start_searching() wakes up the main thread sleeping in
+// main_loop() so to start a new search, then returns immediately.
+
+void ThreadsManager::start_searching(const Position& pos, const LimitsType& limits,
+ const std::vector& searchMoves) {
+ wait_for_search_finished();
+
+ SearchTime.restart(); // As early as possible
+
+ Signals.stopOnPonderhit = Signals.firstRootMove = false;
+ Signals.stop = Signals.failedLowAtRoot = false;
+
+ RootPosition = pos;
+ Limits = limits;
+ RootMoves.clear();
+
+ for (MoveList ml(pos); !ml.end(); ++ml)
+ if (searchMoves.empty() || count(searchMoves.begin(), searchMoves.end(), ml.move()))
+ RootMoves.push_back(RootMove(ml.move()));
+
+ main_thread()->do_sleep = false;
+ main_thread()->wake_up();
+}