X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fthread.cpp;h=c73f55660330545d300857b2874287fd54b21563;hp=1c5c67bacfc2b587a5bac6f6f56fd84316aaa1bf;hb=2ca142a5b4ca200c56cb99495ec51a804983d07d;hpb=1a414cd9cb274d9572c60fe4067671cd993ffb36

diff --git a/src/thread.cpp b/src/thread.cpp
index 1c5c67ba..c73f5566 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) 2008-2012 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
@@ -19,126 +19,73 @@
 
 #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().
-
- long start_routine(Thread* th) { th->idle_loop(); return 0; }
-
-} }
-
-
-// Thread c'tor starts a newly-created thread of execution that will call
-// the the virtual function idle_loop(), going immediately to sleep.
-
-Thread::Thread() : splitPoints() {
-
-  searching = exit = false;
-  maxPly = splitPointsSize = 0;
-  activeSplitPoint = NULL;
-  idx = Threads.size();
-
-  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
-
-Thread::~Thread() {
-
-  exit = true; // Search must be already finished
-  notify_one();
-  thread_join(handle); // Wait for thread termination
-}
-
-
-// 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();
 
-void TimerThread::idle_loop() {
+namespace {
 
-  while (!exit)
-  {
-      mutex.lock();
+ // 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.
 
-      if (!exit)
-          sleepCondition.wait_for(mutex, msec ? msec : INT_MAX);
+ template<typename T> T* new_thread() {
+   std::thread* th = new T;
+   *th = std::thread(&T::idle_loop, (T*)th); // Will go to sleep
+   return (T*)th;
+ }
 
-      mutex.unlock();
+ void delete_thread(ThreadBase* th) {
 
-      if (msec)
-          check_time();
-  }
-}
-
-
-// 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.
+   th->mutex.lock();
+   th->exit = true; // Search must be already finished
+   th->mutex.unlock();
 
-void MainThread::idle_loop() {
-
-  while (true)
-  {
-      mutex.lock();
-
-      thinking = false;
+   th->notify_one();
+   th->join(); // Wait for thread termination
+   delete th;
+ }
 
-      while (!thinking && !exit)
-      {
-          Threads.sleepCondition.notify_one(); // Wake up UI thread if needed
-          sleepCondition.wait(mutex);
-      }
-
-      mutex.unlock();
-
-      if (exit)
-          return;
+}
 
-      searching = true;
 
-      Search::think();
+// ThreadBase::notify_one() wakes up the thread when there is some work to do
 
-      assert(searching);
+void ThreadBase::notify_one() {
 
-      searching = false;
-  }
+  std::unique_lock<Mutex> lk(mutex);
+  sleepCondition.notify_one();
 }
 
 
-// Thread::notify_one() wakes up the thread when there is some search to do
+// ThreadBase::wait_for() set the thread to sleep until 'condition' turns true
 
-void Thread::notify_one() {
+void ThreadBase::wait_for(volatile const bool& condition) {
 
-  mutex.lock();
-  sleepCondition.notify_one();
-  mutex.unlock();
+  std::unique_lock<Mutex> lk(mutex);
+  sleepCondition.wait(lk, [&]{ return condition; });
 }
 
 
-// Thread::wait_for() set the thread to sleep until condition 'b' turns true
+// Thread c'tor makes some init but does not launch any execution thread that
+// will be started only when c'tor returns.
 
-void Thread::wait_for(volatile const bool& b) {
+Thread::Thread() /* : splitPoints() */ { // Initialization of non POD broken in MSVC
 
-  mutex.lock();
-  while (!b) sleepCondition.wait(mutex);
-  mutex.unlock();
+  searching = false;
+  maxPly = 0;
+  splitPointsSize = 0;
+  activeSplitPoint = nullptr;
+  activePosition = nullptr;
+  idx = Threads.size(); // Starts from 0
 }
 
 
@@ -155,228 +102,277 @@ bool Thread::cutoff_occurred() const {
 }
 
 
-// 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).
+// 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).
 
-bool Thread::is_available_to(Thread* master) const {
+bool Thread::can_join(const SplitPoint* sp) const {
 
   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;
+  // 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 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));
+  return !size || splitPoints[size - 1].slavesMask.test(sp->master->idx);
 }
 
 
-// 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.
-
-void ThreadPool::init() {
-
-  sleepWhileIdle = true;
-  timer = new TimerThread();
-  push_back(new MainThread());
-  read_uci_options();
-}
-
-
-// exit() cleanly terminates the threads before the program exits
-
-void ThreadPool::exit() {
-
-  delete timer; // As first because check_time() accesses threads data
-
-  for (iterator it = begin(); it != end(); ++it)
-      delete *it;
-}
-
-
-// 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.
-
-void ThreadPool::read_uci_options() {
-
-  maxThreadsPerSplitPoint = Options["Max Threads per Split Point"];
-  minimumSplitDepth       = Options["Min Split Depth"] * ONE_PLY;
-  size_t requested        = Options["Threads"];
-
-  assert(requested > 0);
-
-  while (size() < requested)
-      push_back(new Thread());
-
-  while (size() > requested)
-  {
-      delete back();
-      pop_back();
-  }
-}
-
-
-// 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 {
-
-  for (const_iterator it = begin(); it != end(); ++it)
-      if ((*it)->is_available_to(master))
-          return true;
-
-  return false;
-}
-
-
-// split() does the actual work of distributing the work at a node between
+// 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
-// told that they have been assigned work. This will cause them to instantly
+// 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.
 
-template <bool Fake>
-Value Thread::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 <= alpha && alpha < beta && beta <= VALUE_INFINITE);
-  assert(bestValue > -VALUE_INFINITE);
-  assert(depth >= Threads.minimumSplitDepth);
+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 the next available split point from the split point stack
+  // Pick and init the next available split point
   SplitPoint& sp = splitPoints[splitPointsSize];
 
-  sp.masterThread = this;
+  sp.spinlock.acquire(); // No contention here until we don't increment splitPointsSize
+
+  sp.master = this;
   sp.parentSplitPoint = activeSplitPoint;
-  sp.slavesMask = 1ULL << idx;
+  sp.slavesMask = 0, sp.slavesMask.set(idx);
   sp.depth = depth;
+  sp.bestValue = *bestValue;
   sp.bestMove = *bestMove;
-  sp.threatMove = threatMove;
   sp.alpha = alpha;
   sp.beta = beta;
   sp.nodeType = nodeType;
-  sp.bestValue = bestValue;
-  sp.movePicker = &mp;
+  sp.cutNode = cutNode;
+  sp.movePicker = movePicker;
   sp.moveCount = moveCount;
   sp.pos = &pos;
   sp.nodes = 0;
   sp.cutoff = false;
   sp.ss = ss;
+  sp.allSlavesSearching = true; // Must be set under lock protection
 
-  // 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();
-
-  splitPointsSize++;
+  ++splitPointsSize;
   activeSplitPoint = &sp;
+  activePosition = nullptr;
 
-  size_t slavesCnt = 1; // This thread is always included
+  // Try to allocate available threads
+  Thread* slave;
 
-  for (ThreadPool::iterator it = Threads.begin(); it != Threads.end() && !Fake; ++it)
+  while (    sp.slavesMask.count() < MAX_SLAVES_PER_SPLITPOINT
+         && (slave = Threads.available_slave(&sp)) != nullptr)
   {
-      Thread* slave = *it;
+     slave->spinlock.acquire();
 
-      if (slave->is_available_to(this) && ++slavesCnt <= Threads.maxThreadsPerSplitPoint)
+      if (slave->can_join(activeSplitPoint))
       {
-          sp.slavesMask |= 1ULL << slave->idx;
-          slave->activeSplitPoint = &sp;
-          slave->searching = true; // Slave leaves idle_loop()
-          slave->notify_one(); // Could be sleeping
+          activeSplitPoint->slavesMask.set(slave->idx);
+          slave->activeSplitPoint = activeSplitPoint;
+          slave->searching = true;
       }
-  }
 
-  sp.mutex.unlock();
-  Threads.mutex.unlock();
+      slave->spinlock.release();
+  }
 
   // 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)
-  {
-      Thread::idle_loop(); // Force a call to base class idle_loop()
+  sp.spinlock.release();
 
-      // 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(!searching);
-  }
+  Thread::idle_loop(); // Force a call to base class idle_loop()
 
-  // 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().
-  Threads.mutex.lock();
-  sp.mutex.lock();
+  // 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);
 
   searching = true;
-  splitPointsSize--;
+
+  // We have returned from the idle loop, which means that all threads are
+  // finished. Note that decreasing splitPointsSize must be done under lock
+  // protection to avoid a race with Thread::can_join().
+  sp.spinlock.acquire();
+
+  --splitPointsSize;
   activeSplitPoint = sp.parentSplitPoint;
+  activePosition = &pos;
   pos.set_nodes_searched(pos.nodes_searched() + sp.nodes);
   *bestMove = sp.bestMove;
+  *bestValue = sp.bestValue;
+
+  sp.spinlock.release();
+}
+
+
+// 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.
 
-  sp.mutex.unlock();
-  Threads.mutex.unlock();
+void TimerThread::idle_loop() {
+
+  while (!exit)
+  {
+      std::unique_lock<Mutex> lk(mutex);
+
+      if (!exit)
+          sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX));
 
-  return sp.bestValue;
+      lk.unlock();
+
+      if (run)
+          check_time();
+  }
 }
 
-// Explicit template instantiations
-template Value Thread::split<false>(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker&, int);
-template Value Thread::split<true>(Position&, Stack*, Value, Value, Value, Move*, Depth, Move, int, MovePicker&, int);
 
+// 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.
 
-// wait_for_think_finished() waits for main thread to go to sleep then returns
+void MainThread::idle_loop() {
 
-void ThreadPool::wait_for_think_finished() {
+  while (!exit)
+  {
+      std::unique_lock<Mutex> lk(mutex);
+
+      thinking = false;
 
-  MainThread* t = main_thread();
-  t->mutex.lock();
-  while (t->thinking) sleepCondition.wait(t->mutex);
-  t->mutex.unlock();
+      while (!thinking && !exit)
+      {
+          sleepCondition.notify_one(); // Wake up the UI thread if needed
+          sleepCondition.wait(lk);
+      }
+
+      lk.unlock();
+
+      if (!exit)
+      {
+          searching = true;
+
+          Search::think();
+
+          assert(searching);
+
+          searching = false;
+      }
+  }
 }
 
 
-// start_thinking() wakes up the main thread sleeping in MainThread::idle_loop()
-// so to start a new search, then returns immediately.
+// MainThread::join() waits for main thread to finish the search
 
-void ThreadPool::start_thinking(const Position& pos, const LimitsType& limits,
-                                const std::vector<Move>& searchMoves, StateStackPtr& states) {
-  wait_for_think_finished();
+void MainThread::join() {
+
+  std::unique_lock<Mutex> lk(mutex);
+  sleepCondition.wait(lk, [&]{ return !thinking; });
+}
 
-  SearchTime = Time::now(); // As early as possible
+
+// 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() {
+
+  timer = new_thread<TimerThread>();
+  push_back(new_thread<MainThread>());
+  read_uci_options();
+}
+
+
+// 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() {
+
+  delete_thread(timer); // As first because check_time() accesses threads data
+  timer = nullptr;
+
+  for (Thread* th : *this)
+      delete_thread(th);
+
+  clear(); // Get rid of stale pointers
+}
+
+
+// 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() {
+
+  minimumSplitDepth = Options["Min Split Depth"] * ONE_PLY;
+  size_t requested  = Options["Threads"];
+
+  assert(requested > 0);
+
+  // If zero (default) then set best minimum split depth automatically
+  if (!minimumSplitDepth)
+      minimumSplitDepth =  5 * ONE_PLY ;
+
+  while (size() < requested)
+      push_back(new_thread<Thread>());
+
+  while (size() > requested)
+  {
+      delete_thread(back());
+      pop_back();
+  }
+}
+
+
+// ThreadPool::available_slave() tries to find an idle thread which is available
+// to join SplitPoint 'sp'.
+
+Thread* ThreadPool::available_slave(const SplitPoint* sp) const {
+
+  for (Thread* th : *this)
+      if (th->can_join(sp))
+          return th;
+
+  return nullptr;
+}
+
+
+// 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) {
+  main()->join();
 
   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()
-          || std::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(); // Wake up main thread: 'thinking' must be already set
 }