/*
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;
+
+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() of the supplied thread.
- // There are two versions of this function; one for POSIX threads and
- // one for Windows threads.
-
-#if defined(_MSC_VER)
+ // 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().
+#if defined(_WIN32) || defined(_WIN64)
DWORD WINAPI start_routine(LPVOID thread) {
+#else
+ void* start_routine(void* thread) {
+#endif
+
+ Thread* th = (Thread*)thread;
+
+ if (th->threadID == 0)
+ th->main_loop();
+
+ else if (th->threadID == MAX_THREADS)
+ th->timer_loop();
+
+ else
+ th->idle_loop(NULL);
- ((Thread*)thread)->idle_loop(NULL);
return 0;
}
-#else
+} }
- void* start_routine(void* thread) {
- ((Thread*)thread)->idle_loop(NULL);
- return NULL;
+// 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() {
+
+ 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;
+
+ is_searching = true;
+
+ Search::think();
+ }
+}
-// 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.
+// 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.
void Thread::wake_up() {
- lock_grab(&sleepLock);
- cond_signal(&sleepCond);
- lock_release(&sleepLock);
+ lock_grab(sleepLock);
+ cond_signal(sleepCond);
+ lock_release(sleepLock);
+}
+
+
+// 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);
}
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;
}
// 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));
}
-// 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.
+// read_uci_options() updates internal threads parameters from the corresponding
+// UCI options. It is called before to start a new search.
void ThreadsManager::read_uci_options() {
- 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>();
-
- set_size(Options["Threads"].value<int>());
+ maxThreadsPerSplitPoint = Options["Max Threads per Split Point"];
+ minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
+ useSleepingThreads = Options["Use Sleeping Threads"];
}
void ThreadsManager::set_size(int cnt) {
- assert(cnt > 0 && cnt <= MAX_THREADS);
+ assert(cnt > 0 && cnt < MAX_THREADS);
activeThreads = cnt;
{
// 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.
+ // possible threads if only few are used.
threads[i].pawnTable.init();
threads[i].materialTable.init();
+ threads[i].maxPly = 0;
threads[i].do_sleep = false;
+
+ if (!useSleepingThreads)
+ threads[i].wake_up();
}
else
threads[i].do_sleep = true;
void ThreadsManager::init() {
- // Initialize threads lock, used when allocating slaves during splitting
- lock_init(&threadsLock);
+ cond_init(sleepCond);
+ lock_init(splitLock);
- // Initialize sleep and split point locks
- for (int i = 0; i < MAX_THREADS; i++)
+ for (int i = 0; i <= MAX_THREADS; i++)
{
- lock_init(&threads[i].sleepLock);
- cond_init(&threads[i].sleepCond);
+ lock_init(threads[i].sleepLock);
+ cond_init(threads[i].sleepCond);
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_init(&(threads[i].splitPoints[j].lock));
+ for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++)
+ lock_init(threads[i].splitPoints[j].lock);
}
- // Initialize main thread's associated data
- threads[0].is_searching = true;
- threads[0].threadID = 0;
- set_size(1); // This makes all the threads but the main to go to sleep
+ // Allocate main thread tables to call evaluate() also when not searching
+ threads[0].pawnTable.init();
+ threads[0].materialTable.init();
- // 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++)
+ // Create and launch all the threads, threads will go immediately to sleep
+ for (int i = 0; i <= MAX_THREADS; i++)
{
threads[i].is_searching = false;
+ threads[i].do_sleep = (i != 0); // Avoid a race with start_thinking()
threads[i].threadID = i;
-#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
-
- if (!ok)
+ if (!thread_create(threads[i].handle, start_routine, threads[i]))
{
std::cerr << "Failed to create thread number " << i << std::endl;
::exit(EXIT_FAILURE);
void ThreadsManager::exit() {
- for (int i = 0; i < MAX_THREADS; i++)
+ 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();
+ assert(threads[i].do_sleep);
-#if defined(_MSC_VER)
- WaitForSingleObject(threads[i].handle, 0);
- CloseHandle(threads[i].handle);
-#else
- pthread_join(threads[i].handle, NULL);
-#endif
- }
+ threads[i].do_exit = true; // Search must be already finished
+ threads[i].wake_up();
+
+ thread_join(threads[i].handle); // Wait for thread termination
- // Now we can safely destroy locks and wait conditions
- lock_destroy(&threads[i].sleepLock);
- cond_destroy(&threads[i].sleepCond);
+ 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));
+ for (int j = 0; j < MAX_SPLITPOINTS_PER_THREAD; j++)
+ lock_destroy(threads[i].splitPoints[j].lock);
}
- lock_destroy(&threadsLock);
+ lock_destroy(splitLock);
+ cond_destroy(sleepCond);
}
// 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 with threadID 'master'.
bool ThreadsManager::available_slave_exists(int master) const {
assert(master >= 0 && master < activeThreads);
for (int i = 0; i < activeThreads; i++)
- if (i != master && threads[i].is_available_to(master))
+ 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 ThreadsManager::split(Position& pos, Stack* 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(pos.pos_is_ok());
+ assert(bestValue > -VALUE_INFINITE);
assert(bestValue <= alpha);
assert(alpha < beta);
assert(beta <= VALUE_INFINITE);
assert(pos.thread() >= 0 && pos.thread() < activeThreads);
assert(activeThreads > 1);
- int i, master = pos.thread();
+ int 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)
+ if (masterThread.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;
+ // Pick the next available split point from the split point stack
+ SplitPoint* sp = &masterThread.splitPoints[masterThread.splitPointsCnt];
- // Initialize the split point object
- sp->parent = masterThread.splitPoint;
+ sp->parent = masterThread.curSplitPoint;
sp->master = master;
- sp->is_betaCutoff = false;
+ sp->cutoff = false;
+ sp->slavesMask = 1ULL << master;
sp->depth = depth;
sp->threatMove = threatMove;
sp->alpha = alpha;
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.is_searching);
- int workersCnt = 1; // At least the master is included
+ 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.
+ lock_grab(sp->lock);
+ lock_grab(splitLock);
- for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++)
- if (i != master && threads[i].is_available_to(master))
+ for (int i = 0; i < activeThreads && !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].is_searching = true;
+ 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_release(&threadsLock);
+ masterThread.curSplitPoint = sp;
+ masterThread.splitPointsCnt++;
- // We failed to allocate even one slave, return
- if (!Fake && workersCnt == 1)
- return bestValue;
-
- masterThread.splitPoint = sp;
- masterThread.activeSplitPoints++;
+ lock_release(splitLock);
+ lock_release(sp->lock);
// 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.
- masterThread.idle_loop(sp);
-
- // 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.is_searching);
+ if (slavesCnt || Fake)
+ masterThread.idle_loop(sp);
// 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 activeSplitPoints 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.activeSplitPoints--;
-
- lock_release(&threadsLock);
-
- masterThread.splitPoint = sp->parent;
+ masterThread.splitPointsCnt--;
+ masterThread.curSplitPoint = sp->parent;
pos.set_nodes_searched(pos.nodes_searched() + sp->nodes);
+ lock_release(splitLock);
+ lock_release(sp->lock);
+
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 ThreadsManager::split<false>(Position&, Stack*, Value, Value, Value, Depth, Move, int, MovePicker*, int);
+template Value ThreadsManager::split<true>(Position&, Stack*, Value, Value, Value, Depth, Move, int, MovePicker*, int);
+
+
+// ThreadsManager::set_timer() is used to set the timer to trigger after msec
+// milliseconds. If msec is 0 then timer is stopped.
+
+void ThreadsManager::set_timer(int msec) {
+
+ Thread& timer = threads[MAX_THREADS];
+
+ lock_grab(timer.sleepLock);
+ timer.maxPly = msec;
+ cond_signal(timer.sleepCond); // Wake up and restart the timer
+ lock_release(timer.sleepLock);
+}
+
+
+// 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.
+
+void ThreadsManager::start_thinking(const Position& pos, const LimitsType& limits,
+ const std::set<Move>& searchMoves, bool async) {
+ Thread& main = threads[0];
+
+ lock_grab(main.sleepLock);
+
+ // Wait main thread has finished before to launch a new search
+ while (!main.do_sleep)
+ cond_wait(sleepCond, main.sleepLock);
+
+ // Copy input arguments to initialize the search
+ RootPosition.copy(pos, 0);
+ Limits = limits;
+ RootMoves.clear();
+
+ // 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<MV_LEGAL> ml(pos); !ml.end(); ++ml)
+ if (searchMoves.empty() || searchMoves.count(ml.move()))
+ RootMoves.push_back(RootMove(ml.move()));
+
+ // Reset signals before to start the new search
+ Signals.stopOnPonderhit = Signals.firstRootMove = false;
+ Signals.stop = Signals.failedLowAtRoot = false;
+
+ 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);
+
+ lock_release(main.sleepLock);
+}
+
+
+// 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.
+
+void ThreadsManager::stop_thinking() {
+
+ Thread& main = threads[0];
+
+ Search::Signals.stop = true;
+
+ lock_grab(main.sleepLock);
+
+ cond_signal(main.sleepCond); // In case is waiting for stop or ponderhit
+
+ while (!main.do_sleep)
+ cond_wait(sleepCond, main.sleepLock);
+
+ lock_release(main.sleepLock);
+}