From 62f531254e03f946c92ad307fb68c7faa2806d16 Mon Sep 17 00:00:00 2001 From: Marco Costalba Date: Thu, 1 Jan 2015 10:46:31 +0100 Subject: [PATCH] Fix comments in thread.cpp And reshuffle a bit the functions to place them in a consistent order. To be on the safe side, patch has been validated for no regression/crashes with a small 8K games test with 3 threads: ELO: 3.98 +-4.4 (95%) LOS: 96.3% Total: 8388 W: 1500 L: 1404 D: 5484 No functional change. --- src/thread.cpp | 256 +++++++++++++++++++++++++------------------------ 1 file changed, 131 insertions(+), 125 deletions(-) diff --git a/src/thread.cpp b/src/thread.cpp index ca87e1dd..9400b38d 100644 --- a/src/thread.cpp +++ b/src/thread.cpp @@ -40,7 +40,7 @@ namespace { // Helpers to launch a thread after creation and joining before delete. Must be - // outside Thread c'tor and d'tor because the object will be fully initialized + // 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() { @@ -50,7 +50,11 @@ namespace { } void delete_thread(ThreadBase* th) { + + th->mutex.lock(); th->exit = true; // Search must be already finished + th->mutex.unlock(); + th->notify_one(); thread_join(th->handle); // Wait for thread termination delete th; @@ -59,7 +63,7 @@ namespace { } -// notify_one() wakes up the thread when there is some work to do +// ThreadBase::notify_one() wakes up the thread when there is some work to do void ThreadBase::notify_one() { @@ -69,7 +73,7 @@ void ThreadBase::notify_one() { } -// wait_for() set the thread to sleep until 'condition' turns true +// ThreadBase::wait_for() set the thread to sleep until 'condition' turns true void ThreadBase::wait_for(volatile const bool& condition) { @@ -79,8 +83,8 @@ void ThreadBase::wait_for(volatile const bool& condition) { } -// Thread c'tor just inits data and does not launch any execution thread. -// Such a thread will only be started when c'tor returns. +// 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() */ { // Value-initialization bug in MSVC @@ -92,7 +96,7 @@ Thread::Thread() /* : splitPoints() */ { // Value-initialization bug in MSVC } -// cutoff_occurred() checks whether a beta cutoff has occurred in the +// 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 { @@ -127,8 +131,101 @@ bool Thread::available_to(const Thread* master) const { } -// TimerThread::idle_loop() is where the timer thread waits msec milliseconds -// and then calls check_time(). If msec is 0 thread sleeps until it's woken up. +// 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.masterThread = 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; + + // 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.mutex.lock(); + sp.mutex.lock(); + + sp.allSlavesSearching = true; // Must be set under lock protection + ++splitPointsSize; + activeSplitPoint = &sp; + activePosition = NULL; + + Thread* slave; + + while ((slave = Threads.available_slave(this)) != NULL) + { + sp.slavesMask.set(slave->idx); + slave->activeSplitPoint = &sp; + slave->searching = true; // Slave leaves idle_loop() + slave->notify_one(); // Could be sleeping + } + + // 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.mutex.unlock(); + Threads.mutex.unlock(); + + 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); + + // 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.mutex.lock(); + sp.mutex.lock(); + + searching = true; + --splitPointsSize; + activeSplitPoint = sp.parentSplitPoint; + activePosition = &pos; + pos.set_nodes_searched(pos.nodes_searched() + sp.nodes); + *bestMove = sp.bestMove; + *bestValue = sp.bestValue; + + sp.mutex.unlock(); + Threads.mutex.unlock(); +} + + +// 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 TimerThread::idle_loop() { @@ -152,7 +249,7 @@ void TimerThread::idle_loop() { void MainThread::idle_loop() { - while (true) + while (!exit) { mutex.lock(); @@ -166,24 +263,24 @@ void MainThread::idle_loop() { mutex.unlock(); - if (exit) - return; - - searching = true; + if (!exit) + { + searching = true; - Search::think(); + Search::think(); - assert(searching); + assert(searching); - searching = false; + searching = false; + } } } -// 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. +// 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() { @@ -193,8 +290,8 @@ void ThreadPool::init() { } -// exit() cleanly terminates the threads before the program exits. Cannot be done in -// d'tor because we have to terminate the threads before to free ThreadPool object. +// ThreadPool::exit() terminates the threads before the program exits. Cannot be +// done in d'tor because threads must be terminated before freeing us. void ThreadPool::exit() { @@ -205,11 +302,11 @@ void ThreadPool::exit() { } -// 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. +// 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 ThreadPool::read_uci_options() { @@ -233,8 +330,8 @@ void ThreadPool::read_uci_options() { } -// available_slave() tries to find an idle thread which is available as a slave -// for the thread 'master'. +// ThreadPool::available_slave() tries to find an idle thread which is available +// as a slave for the thread 'master'. Thread* ThreadPool::available_slave(const Thread* master) const { @@ -246,98 +343,7 @@ Thread* ThreadPool::available_slave(const Thread* 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), 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. - -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(pos.pos_is_ok()); - assert(-VALUE_INFINITE < *bestValue && *bestValue <= alpha && alpha < beta && beta <= VALUE_INFINITE); - assert(depth >= Threads.minimumSplitDepth); - assert(searching); - assert(splitPointsSize < MAX_SPLITPOINTS_PER_THREAD); - - // Pick the next available split point from the split point stack - SplitPoint& sp = splitPoints[splitPointsSize]; - - sp.masterThread = 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; - - // 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.mutex.lock(); - sp.mutex.lock(); - - sp.allSlavesSearching = true; // Must be set under lock protection - ++splitPointsSize; - activeSplitPoint = &sp; - activePosition = NULL; - - for (Thread* slave; (slave = Threads.available_slave(this)) != NULL; ) - { - sp.slavesMask.set(slave->idx); - slave->activeSplitPoint = &sp; - slave->searching = true; // Slave leaves idle_loop() - slave->notify_one(); // Could be sleeping - } - - // 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.mutex.unlock(); - Threads.mutex.unlock(); - - 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); - - // 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::available_to(). - Threads.mutex.lock(); - sp.mutex.lock(); - - searching = true; - --splitPointsSize; - activeSplitPoint = sp.parentSplitPoint; - activePosition = &pos; - pos.set_nodes_searched(pos.nodes_searched() + sp.nodes); - *bestMove = sp.bestMove; - *bestValue = sp.bestValue; - - sp.mutex.unlock(); - Threads.mutex.unlock(); -} - -// 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() { @@ -348,11 +354,11 @@ void ThreadPool::wait_for_think_finished() { } -// start_thinking() wakes up the main thread sleeping in MainThread::idle_loop() -// so to start a new search, then returns immediately. - -void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits, StateStackPtr& states) { +// 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, + StateStackPtr& states) { wait_for_think_finished(); SearchTime = Time::now(); // As early as possible -- 2.39.2