/*
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) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad
+ Copyright (C) 2015-2018 Marco Costalba, Joona Kiiski, Gary Linscott, 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 <iostream>
+#include <algorithm> // For std::count
+#include <cassert>
+#include "movegen.h"
+#include "search.h"
#include "thread.h"
-#include "ucioption.h"
+#include "uci.h"
+#include "syzygy/tbprobe.h"
+#include "tt.h"
-ThreadsManager ThreadsMgr; // Global object definition
+ThreadPool Threads; // Global object
-namespace {
- // init_thread() is the function which is called when a new thread is
- // launched. It simply calls the idle_loop() function with the supplied
- // threadID. There are two versions of this function; one for POSIX
- // threads and one for Windows threads.
+/// Thread constructor launches the thread and waits until it goes to sleep
+/// in idle_loop(). Note that 'searching' and 'exit' should be alredy set.
-#if !defined(_MSC_VER)
+Thread::Thread(size_t n) : idx(n), stdThread(&Thread::idle_loop, this) {
- void* init_thread(void* threadID) {
-
- ThreadsMgr.idle_loop(*(int*)threadID, NULL);
- return NULL;
- }
+ wait_for_search_finished();
+}
-#else
- DWORD WINAPI init_thread(LPVOID threadID) {
+/// Thread destructor wakes up the thread in idle_loop() and waits
+/// for its termination. Thread should be already waiting.
- ThreadsMgr.idle_loop(*(int*)threadID, NULL);
- return 0;
- }
+Thread::~Thread() {
-#endif
+ assert(!searching);
+ exit = true;
+ start_searching();
+ stdThread.join();
}
-// 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.
+/// Thread::clear() reset histories, usually before a new game
-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>();
- activeThreads = Options["Threads"].value<int>();
-}
+void Thread::clear() {
+ counterMoves.fill(MOVE_NONE);
+ mainHistory.fill(0);
+ captureHistory.fill(0);
-// init_threads() is called during startup. Initializes locks and condition
-// variables and launches all threads sending them immediately to sleep.
+ for (auto& to : continuationHistory)
+ for (auto& h : to)
+ h->fill(0);
-void ThreadsManager::init_threads() {
-
- int arg[MAX_THREADS];
+ continuationHistory[NO_PIECE][0]->fill(Search::CounterMovePruneThreshold - 1);
+}
- // This flag is needed to properly end the threads when program exits
- allThreadsShouldExit = false;
+/// Thread::start_searching() wakes up the thread that will start the search
- // Threads will sent to sleep as soon as created, only main thread is kept alive
- activeThreads = 1;
- threads[0].state = THREAD_SEARCHING;
+void Thread::start_searching() {
- // Allocate pawn and material hash tables for main thread
- init_hash_tables();
+ std::lock_guard<Mutex> lk(mutex);
+ searching = true;
+ cv.notify_one(); // Wake up the thread in idle_loop()
+}
- lock_init(&mpLock);
- // Initialize thread and split point locks
- for (int i = 0; i < MAX_THREADS; i++)
- {
- lock_init(&threads[i].sleepLock);
- cond_init(&threads[i].sleepCond);
+/// Thread::wait_for_search_finished() blocks on the condition variable
+/// until the thread has finished searching.
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_init(&(threads[i].splitPoints[j].lock));
- }
+void Thread::wait_for_search_finished() {
- // 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;
- arg[i] = i;
-
-#if !defined(_MSC_VER)
- pthread_t pthread[1];
- bool ok = (pthread_create(pthread, NULL, init_thread, (void*)(&arg[i])) == 0);
- pthread_detach(pthread[0]);
-#else
- bool ok = (CreateThread(NULL, 0, init_thread, (LPVOID)(&arg[i]), 0, NULL) != NULL);
-#endif
- if (!ok)
- {
- std::cout << "Failed to create thread number " << i << std::endl;
- exit(EXIT_FAILURE);
- }
-
- // Wait until the thread has finished launching and is gone to sleep
- while (threads[i].state == THREAD_INITIALIZING) {}
- }
+ std::unique_lock<Mutex> lk(mutex);
+ cv.wait(lk, [&]{ return !searching; });
}
-// exit_threads() is called when the program exits. It makes all the
-// helper threads exit cleanly.
+/// Thread::idle_loop() is where the thread is parked, blocked on the
+/// condition variable, when it has no work to do.
-void ThreadsManager::exit_threads() {
+void Thread::idle_loop() {
- // Force the woken up threads to exit idle_loop() and hence terminate
- allThreadsShouldExit = true;
+ // 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);
- for (int i = 0; i < MAX_THREADS; i++)
+ while (true)
{
- // 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);
-
- for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
- lock_destroy(&(threads[i].splitPoints[j].lock));
- }
+ std::unique_lock<Mutex> lk(mutex);
+ searching = false;
+ cv.notify_one(); // Wake up anyone waiting for search finished
+ cv.wait(lk, [&]{ return searching; });
- lock_destroy(&mpLock);
-}
+ if (exit)
+ return;
+ lk.unlock();
-// 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.
-
-void ThreadsManager::init_hash_tables() {
-
- for (int i = 0; i < activeThreads; i++)
- {
- threads[i].pawnTable.init();
- threads[i].materialTable.init();
+ search();
}
}
+/// ThreadPool::set() creates/destroys threads to match the requested number.
+/// Created and launched threads will go immediately to sleep in idle_loop.
+/// Upon resizing, threads are recreated to allow for binding if necessary.
-// cutoff_at_splitpoint() checks whether a beta cutoff has occurred in
-// the thread's currently active split point, or in some ancestor of
-// the current split point.
-
-bool ThreadsManager::cutoff_at_splitpoint(int threadID) const {
+void ThreadPool::set(size_t requested) {
- assert(threadID >= 0 && threadID < activeThreads);
+ if (size() > 0) { // destroy any existing thread(s)
+ main()->wait_for_search_finished();
- SplitPoint* sp = threads[threadID].splitPoint;
-
- for ( ; sp && !sp->betaCutoff; sp = sp->parent) {}
- return sp != NULL;
-}
+ while (size() > 0)
+ delete back(), pop_back();
+ }
+ if (requested > 0) { // create new thread(s)
+ push_back(new MainThread(0));
-// thread_is_available() checks whether the thread with threadID "slave" is
-// available to help the thread with threadID "master" at a split point. An
-// obvious requirement is that "slave" must be idle. With more than two
-// threads, this is not by itself sufficient: If "slave" is the master of
-// some active split point, it is only available as a slave to the other
-// threads which are busy searching the split point at the top of "slave"'s
-// split point stack (the "helpful master concept" in YBWC terminology).
+ while (size() < requested)
+ push_back(new Thread(size()));
+ clear();
+ }
-bool ThreadsManager::thread_is_available(int slave, int master) const {
+ // Reallocate the hash with the new threadpool size
+ TT.resize(Options["Hash"]);
+}
- assert(slave >= 0 && slave < activeThreads);
- assert(master >= 0 && master < activeThreads);
- assert(activeThreads > 1);
+/// ThreadPool::clear() sets threadPool data to initial values.
- if (threads[slave].state != THREAD_AVAILABLE || slave == master)
- return false;
+void ThreadPool::clear() {
- // Make a local copy to be sure doesn't change under our feet
- int localActiveSplitPoints = threads[slave].activeSplitPoints;
+ for (Thread* th : *this)
+ th->clear();
- // No active split points means that the thread is available as
- // a slave for any other thread.
- if (localActiveSplitPoints == 0 || activeThreads == 2)
- return true;
+ main()->callsCnt = 0;
+ main()->previousScore = VALUE_INFINITE;
+ main()->previousTimeReduction = 1.0;
+}
- // Apply the "helpful master" concept if possible. Use localActiveSplitPoints
- // that is known to be > 0, instead of threads[slave].activeSplitPoints that
- // could have been set to 0 by another thread leading to an out of bound access.
- if (threads[slave].splitPoints[localActiveSplitPoints - 1].slaves[master])
- return true;
+/// ThreadPool::start_thinking() wakes up main thread waiting in idle_loop() and
+/// returns immediately. Main thread will wake up other threads and start the search.
- return false;
-}
+void ThreadPool::start_thinking(Position& pos, StateListPtr& states,
+ const Search::LimitsType& limits, bool ponderMode) {
+ main()->wait_for_search_finished();
-// available_thread_exists() tries to find an idle thread which is available as
-// a slave for the thread with threadID "master".
+ stopOnPonderhit = stop = false;
+ ponder = ponderMode;
+ Search::Limits = limits;
+ Search::RootMoves rootMoves;
-bool ThreadsManager::available_thread_exists(int master) const {
+ for (const auto& m : MoveList<LEGAL>(pos))
+ if ( limits.searchmoves.empty()
+ || std::count(limits.searchmoves.begin(), limits.searchmoves.end(), m))
+ rootMoves.emplace_back(m);
- assert(master >= 0 && master < activeThreads);
- assert(activeThreads > 1);
+ if (!rootMoves.empty())
+ Tablebases::rank_root_moves(pos, rootMoves);
- for (int i = 0; i < activeThreads; i++)
- if (thread_is_available(i, master))
- return true;
+ // After ownership transfer 'states' becomes empty, so if we stop the search
+ // and call 'go' again without setting a new position states.get() == NULL.
+ assert(states.get() || setupStates.get());
- return false;
-}
+ 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 to not lose the info
+ // we need to backup and later restore setupStates->back(). Note that setupStates
+ // is shared by threads but is accessed in read-only mode.
+ StateInfo tmp = setupStates->back();
-// 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.
-
-template <bool Fake>
-void ThreadsManager::split(Position& pos, SearchStack* ss, Value* alpha, const Value beta,
- Value* bestValue, Depth depth, Move threatMove,
- int moveCount, MovePicker* mp, bool pvNode) {
- assert(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];
-
- lock_grab(&mpLock);
-
- // If no other thread is available to help us, or if we have too many
- // active split points, don't split.
- if ( !available_thread_exists(master)
- || masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS)
+ for (Thread* th : *this)
{
- lock_release(&mpLock);
- return;
+ th->nodes = th->tbHits = th->nmpMinPly = 0;
+ th->rootDepth = th->completedDepth = DEPTH_ZERO;
+ th->rootMoves = rootMoves;
+ th->rootPos.set(pos.fen(), pos.is_chess960(), &setupStates->back(), th);
}
- // 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.betaCutoff = false;
- splitPoint.depth = depth;
- splitPoint.threatMove = threatMove;
- splitPoint.alpha = *alpha;
- splitPoint.beta = beta;
- splitPoint.pvNode = pvNode;
- 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.slaves[i] = 0;
-
- 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 (thread_is_available(i, master))
- {
- threads[i].state = THREAD_BOOKED;
- threads[i].splitPoint = &splitPoint;
- splitPoint.slaves[i] = 1;
- workersCnt++;
- }
-
- assert(Fake || workersCnt > 1);
-
- // We can release the lock because slave threads are already booked and master is not available
- lock_release(&mpLock);
-
- // 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.slaves[i])
- {
- assert(i == master || threads[i].state == THREAD_BOOKED);
-
- threads[i].state = THREAD_WORKISWAITING; // This makes the slave to exit from idle_loop()
-
- 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);
-
- // We have returned from the idle loop, which means that all threads are
- // finished. Update alpha and bestValue, and return.
- lock_grab(&mpLock);
-
- *alpha = splitPoint.alpha;
- *bestValue = splitPoint.bestValue;
- masterThread.activeSplitPoints--;
- masterThread.splitPoint = splitPoint.parent;
- pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes);
-
- lock_release(&mpLock);
-}
+ setupStates->back() = tmp;
-// Explicit template instantiations
-template void ThreadsManager::split<0>(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, bool);
-template void ThreadsManager::split<1>(Position&, SearchStack*, Value*, const Value, Value*, Depth, Move, int, MovePicker*, bool);
+ main()->start_searching();
+}