X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=src%2Fthread.cpp;h=30177a3915a84c9245644d779ca6fe0d99bc5a2b;hb=HEAD;hp=c232b849babacc66ef6e262322678b18883a37e5;hpb=f6f1d2422303923927c0c088dee1d6df22dc4b98;p=stockfish

diff --git a/src/thread.cpp b/src/thread.cpp
index c232b849..de8de87d 100644
--- a/src/thread.cpp
+++ b/src/thread.cpp
@@ -1,7 +1,6 @@
 /*
   Stockfish, a UCI chess playing engine derived from Glaurung 2.1
-  Copyright (C) 2004-2008 Tord Romstad (Glaurung author)
-  Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad
+  Copyright (C) 2004-2023 The Stockfish developers (see AUTHORS file)
 
   Stockfish is free software: you can redistribute it and/or modify
   it under the terms of the GNU General Public License as published by
@@ -17,359 +16,262 @@
   along with this program.  If not, see <http://www.gnu.org/licenses/>.
 */
 
-#include <algorithm> // For std::count
-#include <cassert>
+#include "thread.h"
 
+#include <algorithm>
+#include <cassert>
+#include <cmath>
+#include <cstdlib>
+#include <deque>
+#include <initializer_list>
+#include <map>
+#include <memory>
+#include <utility>
+
+#include "evaluate.h"
+#include "misc.h"
 #include "movegen.h"
 #include "search.h"
-#include "thread.h"
+#include "syzygy/tbprobe.h"
+#include "tt.h"
 #include "uci.h"
 
-using namespace Search;
-
-ThreadPool Threads; // Global object
-
-extern void check_time();
+namespace Stockfish {
 
-namespace {
+ThreadPool Threads;  // Global object
 
- // 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.
 
- template<typename T> T* new_thread() {
-   T* th = new T();
-   th->nativeThread = std::thread(&ThreadBase::idle_loop, th); // Will go to sleep
-   return th;
- }
-
- void delete_thread(ThreadBase* th) {
-
-   th->mutex.lock();
-   th->exit = true; // Search must be already finished
-   th->mutex.unlock();
-
-   th->notify_one();
-   th->nativeThread.join(); // Wait for thread termination
-   delete th;
- }
+// Constructor launches the thread and waits until it goes to sleep
+// in idle_loop(). Note that 'searching' and 'exit' should be already set.
+Thread::Thread(size_t n) :
+    idx(n),
+    stdThread(&Thread::idle_loop, this) {
 
+    wait_for_search_finished();
 }
 
 
-// ThreadBase::notify_one() wakes up the thread when there is some work to do
+// Destructor wakes up the thread in idle_loop() and waits
+// for its termination. Thread should be already waiting.
+Thread::~Thread() {
 
-void ThreadBase::notify_one() {
+    assert(!searching);
 
-  std::unique_lock<std::mutex>(this->mutex);
-  sleepCondition.notify_one();
+    exit = true;
+    start_searching();
+    stdThread.join();
 }
 
 
-// ThreadBase::wait_for() set the thread to sleep until 'condition' turns true
+// Reset histories, usually before a new game
+void Thread::clear() {
 
-void ThreadBase::wait_for(volatile const bool& condition) {
+    counterMoves.fill(MOVE_NONE);
+    mainHistory.fill(0);
+    captureHistory.fill(0);
+    pawnHistory.fill(0);
 
-  std::unique_lock<std::mutex> lk(mutex);
-  sleepCondition.wait(lk, [&]{ return condition; });
+    for (bool inCheck : {false, true})
+        for (StatsType c : {NoCaptures, Captures})
+            for (auto& to : continuationHistory[inCheck][c])
+                for (auto& h : to)
+                    h->fill(-71);
 }
 
 
-// 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 = splitPointsSize = 0;
-  activeSplitPoint = nullptr;
-  activePosition = nullptr;
-  idx = Threads.size(); // Starts from 0
+// Wakes up the thread that will start the search
+void Thread::start_searching() {
+    mutex.lock();
+    searching = true;
+    mutex.unlock();   // Unlock before notifying saves a few CPU-cycles
+    cv.notify_one();  // Wake up the thread in idle_loop()
 }
 
 
-// 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 {
+// Blocks on the condition variable
+// until the thread has finished searching.
+void Thread::wait_for_search_finished() {
 
-  for (SplitPoint* sp = activeSplitPoint; sp; sp = sp->parentSplitPoint)
-      if (sp->cutoff)
-          return true;
-
-  return false;
+    std::unique_lock<std::mutex> lk(mutex);
+    cv.wait(lk, [&] { return !searching; });
 }
 
 
-// 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).
+// Thread gets parked here, blocked on the
+// condition variable, when it has no work to do.
 
-bool Thread::available_to(const Thread* master) const {
+void Thread::idle_loop() {
 
-  if (searching)
-      return false;
+    // If OS already scheduled us on a different group than 0 then don't overwrite
+    // the choice, eventually we are one of many one-threaded processes running on
+    // some Windows NUMA hardware, for instance in fishtest. To make it simple,
+    // just check if running threads are below a threshold, in this case, all this
+    // NUMA machinery is not needed.
+    if (Options["Threads"] > 8)
+        WinProcGroup::bindThisThread(idx);
 
-  // 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 int size = splitPointsSize;
+    while (true)
+    {
+        std::unique_lock<std::mutex> lk(mutex);
+        searching = false;
+        cv.notify_one();  // Wake up anyone waiting for search finished
+        cv.wait(lk, [&] { return searching; });
 
-  // 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.test(master->idx);
-}
+        if (exit)
+            return;
 
+        lk.unlock();
 
-// 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 = nullptr;
-
-  Thread* slave;
-
-  while ((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
-  }
-
-  // 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();
+        search();
+    }
 }
 
+// Creates/destroys threads to match the requested number.
+// Created and launched threads will immediately go to sleep in idle_loop.
+// Upon resizing, threads are recreated to allow for binding if necessary.
+void ThreadPool::set(size_t requested) {
 
-// 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.
+    if (threads.size() > 0)  // destroy any existing thread(s)
+    {
+        main()->wait_for_search_finished();
 
-void TimerThread::idle_loop() {
+        while (threads.size() > 0)
+            delete threads.back(), threads.pop_back();
+    }
 
-  while (!exit)
-  {
-      std::unique_lock<std::mutex> lk(mutex);
+    if (requested > 0)  // create new thread(s)
+    {
+        threads.push_back(new MainThread(0));
 
-      if (!exit)
-          sleepCondition.wait_for(lk, std::chrono::milliseconds(run ? Resolution : INT_MAX));
+        while (threads.size() < requested)
+            threads.push_back(new Thread(threads.size()));
+        clear();
 
-      lk.unlock();
+        // Reallocate the hash with the new threadpool size
+        TT.resize(size_t(Options["Hash"]));
 
-      if (run)
-          check_time();
-  }
+        // Init thread number dependent search params.
+        Search::init();
+    }
 }
 
 
-// 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)
-  {
-      std::unique_lock<std::mutex> lk(mutex);
-
-      thinking = false;
-
-      while (!thinking && !exit)
-      {
-          Threads.sleepCondition.notify_one(); // Wake up the UI thread if needed
-          sleepCondition.wait(lk);
-      }
+// Sets threadPool data to initial values
+void ThreadPool::clear() {
 
-      lk.unlock();
+    for (Thread* th : threads)
+        th->clear();
 
-      if (!exit)
-      {
-          searching = true;
-
-          Search::think();
-
-          assert(searching);
-
-          searching = false;
-      }
-  }
+    main()->callsCnt                 = 0;
+    main()->bestPreviousScore        = VALUE_INFINITE;
+    main()->bestPreviousAverageScore = VALUE_INFINITE;
+    main()->previousTimeReduction    = 1.0;
 }
 
 
-// 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();
+// Wakes up main thread waiting in idle_loop() and
+// returns immediately. Main thread will wake up other threads and start the search.
+void ThreadPool::start_thinking(Position&                 pos,
+                                StateListPtr&             states,
+                                const Search::LimitsType& limits,
+                                bool                      ponderMode) {
+
+    main()->wait_for_search_finished();
+
+    main()->stopOnPonderhit = stop = false;
+    increaseDepth                  = true;
+    main()->ponder                 = ponderMode;
+    Search::Limits                 = limits;
+    Search::RootMoves rootMoves;
+
+    for (const auto& m : MoveList<LEGAL>(pos))
+        if (limits.searchmoves.empty()
+            || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m))
+            rootMoves.emplace_back(m);
+
+    if (!rootMoves.empty())
+        Tablebases::rank_root_moves(pos, rootMoves);
+
+    // After ownership transfer 'states' becomes empty, so if we stop the search
+    // and call 'go' again without setting a new position states.get() == nullptr.
+    assert(states.get() || setupStates.get());
+
+    if (states.get())
+        setupStates = std::move(states);  // Ownership transfer, states is now empty
+
+    // We use Position::set() to set root position across threads. But there are
+    // some StateInfo fields (previous, pliesFromNull, capturedPiece) that cannot
+    // be deduced from a fen string, so set() clears them and they are set from
+    // setupStates->back() later. The rootState is per thread, earlier states are shared
+    // since they are read-only.
+    for (Thread* th : threads)
+    {
+        th->nodes = th->tbHits = th->nmpMinPly = th->bestMoveChanges = 0;
+        th->rootDepth = th->completedDepth = 0;
+        th->rootMoves                      = rootMoves;
+        th->rootPos.set(pos.fen(), pos.is_chess960(), &th->rootState, th);
+        th->rootState      = setupStates->back();
+        th->rootSimpleEval = Eval::simple_eval(pos, pos.side_to_move());
+    }
+
+    main()->start_searching();
 }
 
-
-// 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
-
-  for (Thread* th : *this)
-      delete_thread(th);
-}
-
-
-// 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 = requested < 8 ? 4 * ONE_PLY : 7 * ONE_PLY;
-
-  while (size() < requested)
-      push_back(new_thread<Thread>());
-
-  while (size() > requested)
-  {
-      delete_thread(back());
-      pop_back();
-  }
+Thread* ThreadPool::get_best_thread() const {
+
+    Thread*                 bestThread = threads.front();
+    std::map<Move, int64_t> votes;
+    Value                   minScore = VALUE_NONE;
+
+    // Find the minimum score of all threads
+    for (Thread* th : threads)
+        minScore = std::min(minScore, th->rootMoves[0].score);
+
+    // Vote according to score and depth, and select the best thread
+    auto thread_value = [minScore](Thread* th) {
+        return (th->rootMoves[0].score - minScore + 14) * int(th->completedDepth);
+    };
+
+    for (Thread* th : threads)
+        votes[th->rootMoves[0].pv[0]] += thread_value(th);
+
+    for (Thread* th : threads)
+        if (std::abs(bestThread->rootMoves[0].score) >= VALUE_TB_WIN_IN_MAX_PLY)
+        {
+            // Make sure we pick the shortest mate / TB conversion or stave off mate the longest
+            if (th->rootMoves[0].score > bestThread->rootMoves[0].score)
+                bestThread = th;
+        }
+        else if (th->rootMoves[0].score >= VALUE_TB_WIN_IN_MAX_PLY
+                 || (th->rootMoves[0].score > VALUE_TB_LOSS_IN_MAX_PLY
+                     && (votes[th->rootMoves[0].pv[0]] > votes[bestThread->rootMoves[0].pv[0]]
+                         || (votes[th->rootMoves[0].pv[0]] == votes[bestThread->rootMoves[0].pv[0]]
+                             && thread_value(th) * int(th->rootMoves[0].pv.size() > 2)
+                                  > thread_value(bestThread)
+                                      * int(bestThread->rootMoves[0].pv.size() > 2)))))
+            bestThread = th;
+
+    return bestThread;
 }
 
 
-// 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 {
+// Start non-main threads
 
-  for (Thread* th : *this)
-      if (th->available_to(master))
-          return th;
+void ThreadPool::start_searching() {
 
-  return nullptr;
+    for (Thread* th : threads)
+        if (th != threads.front())
+            th->start_searching();
 }
 
 
-// ThreadPool::wait_for_think_finished() waits for main thread to finish the search
+// Wait for non-main threads
 
-void ThreadPool::wait_for_think_finished() {
+void ThreadPool::wait_for_search_finished() const {
 
-  std::unique_lock<std::mutex> lk(main()->mutex);
-  sleepCondition.wait(lk, [&]{ return !main()->thinking; });
+    for (Thread* th : threads)
+        if (th != threads.front())
+            th->wait_for_search_finished();
 }
 
-
-// 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
-
-  Signals.stopOnPonderhit = Signals.firstRootMove = false;
-  Signals.stop = Signals.failedLowAtRoot = false;
-
-  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());
-  }
-
-  for (const ExtMove& ms : MoveList<LEGAL>(pos))
-      if (   limits.searchmoves.empty()
-          || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), ms.move))
-          RootMoves.push_back(RootMove(ms.move));
-
-  main()->thinking = true;
-  main()->notify_one(); // Starts main thread
-}
+}  // namespace Stockfish