2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2008 Tord Romstad (Glaurung author)
4 Copyright (C) 2008-2010 Marco Costalba, Joona Kiiski, Tord Romstad
6 Stockfish is free software: you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation, either version 3 of the License, or
9 (at your option) any later version.
11 Stockfish is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with this program. If not, see <http://www.gnu.org/licenses/>.
40 #include "ucioption.h"
47 // Different node types, used as template parameter
48 enum NodeType { NonPV, PV };
50 // Set to true to force running with one thread. Used for debugging.
51 const bool FakeSplit = false;
53 // Lookup table to check if a Piece is a slider and its access function
54 const bool Slidings[18] = { 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1 };
55 inline bool piece_is_slider(Piece p) { return Slidings[p]; }
57 // ThreadsManager class is used to handle all the threads related stuff like init,
58 // starting, parking and, the most important, launching a slave thread at a split
59 // point. All the access to shared thread data is done through this class.
61 class ThreadsManager {
62 /* As long as the single ThreadsManager object is defined as a global we don't
63 need to explicitly initialize to zero its data members because variables with
64 static storage duration are automatically set to zero before enter main()
67 Thread& operator[](int threadID) { return threads[threadID]; }
71 int min_split_depth() const { return minimumSplitDepth; }
72 int active_threads() const { return activeThreads; }
73 void set_active_threads(int cnt) { activeThreads = cnt; }
75 void read_uci_options();
76 bool available_thread_exists(int master) const;
77 bool thread_is_available(int slave, int master) const;
78 bool cutoff_at_splitpoint(int threadID) const;
79 void idle_loop(int threadID, SplitPoint* sp);
82 void split(Position& pos, SearchStack* ss, Value* alpha, const Value beta, Value* bestValue,
83 Depth depth, Move threatMove, int moveCount, MovePicker* mp, bool pvNode);
86 Depth minimumSplitDepth;
87 int maxThreadsPerSplitPoint;
88 bool useSleepingThreads;
90 volatile bool allThreadsShouldExit;
91 Thread threads[MAX_THREADS];
95 // RootMove struct is used for moves at the root of the tree. For each root
96 // move, we store two scores, a node count, and a PV (really a refutation
97 // in the case of moves which fail low). Value pv_score is normally set at
98 // -VALUE_INFINITE for all non-pv moves, while non_pv_score is computed
99 // according to the order in which moves are returned by MovePicker.
104 RootMove(const RootMove& rm) { *this = rm; }
105 RootMove& operator=(const RootMove& rm);
107 // RootMove::operator<() is the comparison function used when
108 // sorting the moves. A move m1 is considered to be better
109 // than a move m2 if it has an higher pv_score, or if it has
110 // equal pv_score but m1 has the higher non_pv_score. In this way
111 // we are guaranteed that PV moves are always sorted as first.
112 bool operator<(const RootMove& m) const {
113 return pv_score != m.pv_score ? pv_score < m.pv_score
114 : non_pv_score < m.non_pv_score;
117 void extract_pv_from_tt(Position& pos);
118 void insert_pv_in_tt(Position& pos);
119 std::string pv_info_to_uci(Position& pos, int depth, int selDepth,
120 Value alpha, Value beta, int pvIdx);
124 Move pv[PLY_MAX_PLUS_2];
128 // RootMoveList struct is just a vector of RootMove objects,
129 // with an handful of methods above the standard ones.
131 struct RootMoveList : public std::vector<RootMove> {
133 typedef std::vector<RootMove> Base;
135 void init(Position& pos, Move searchMoves[]);
136 void sort() { insertion_sort<RootMove, Base::iterator>(begin(), end()); }
137 void sort_multipv(int n) { insertion_sort<RootMove, Base::iterator>(begin(), begin() + n); }
143 // Overload operator<<() to make it easier to print moves in a coordinate
144 // notation compatible with UCI protocol.
145 std::ostream& operator<<(std::ostream& os, Move m) {
147 bool chess960 = (os.iword(0) != 0); // See set960()
148 return os << move_to_uci(m, chess960);
152 // When formatting a move for std::cout we must know if we are in Chess960
153 // or not. To keep using the handy operator<<() on the move the trick is to
154 // embed this flag in the stream itself. Function-like named enum set960 is
155 // used as a custom manipulator and the stream internal general-purpose array,
156 // accessed through ios_base::iword(), is used to pass the flag to the move's
157 // operator<<() that will read it to properly format castling moves.
160 std::ostream& operator<< (std::ostream& os, const set960& f) {
162 os.iword(0) = int(f);
171 // Maximum depth for razoring
172 const Depth RazorDepth = 4 * ONE_PLY;
174 // Dynamic razoring margin based on depth
175 inline Value razor_margin(Depth d) { return Value(0x200 + 0x10 * int(d)); }
177 // Maximum depth for use of dynamic threat detection when null move fails low
178 const Depth ThreatDepth = 5 * ONE_PLY;
180 // Step 9. Internal iterative deepening
182 // Minimum depth for use of internal iterative deepening
183 const Depth IIDDepth[] = { 8 * ONE_PLY, 5 * ONE_PLY };
185 // At Non-PV nodes we do an internal iterative deepening search
186 // when the static evaluation is bigger then beta - IIDMargin.
187 const Value IIDMargin = Value(0x100);
189 // Step 11. Decide the new search depth
191 // Extensions. Array index 0 is used for non-PV nodes, index 1 for PV nodes
192 const Depth CheckExtension[] = { ONE_PLY / 2, ONE_PLY / 1 };
193 const Depth PawnEndgameExtension[] = { ONE_PLY / 1, ONE_PLY / 1 };
194 const Depth PawnPushTo7thExtension[] = { ONE_PLY / 2, ONE_PLY / 2 };
195 const Depth PassedPawnExtension[] = { DEPTH_ZERO, ONE_PLY / 2 };
197 // Minimum depth for use of singular extension
198 const Depth SingularExtensionDepth[] = { 8 * ONE_PLY, 6 * ONE_PLY };
200 // Step 12. Futility pruning
202 // Futility margin for quiescence search
203 const Value FutilityMarginQS = Value(0x80);
205 // Futility lookup tables (initialized at startup) and their access functions
206 Value FutilityMarginsMatrix[16][64]; // [depth][moveNumber]
207 int FutilityMoveCountArray[32]; // [depth]
209 inline Value futility_margin(Depth d, int mn) {
211 return d < 7 * ONE_PLY ? FutilityMarginsMatrix[Max(d, 1)][Min(mn, 63)]
212 : 2 * VALUE_INFINITE;
215 inline int futility_move_count(Depth d) {
217 return d < 16 * ONE_PLY ? FutilityMoveCountArray[d] : MAX_MOVES;
220 // Step 14. Reduced search
222 // Reduction lookup tables (initialized at startup) and their access function
223 int8_t ReductionMatrix[2][64][64]; // [pv][depth][moveNumber]
225 template <NodeType PV>
226 inline Depth reduction(Depth d, int mn) { return (Depth) ReductionMatrix[PV][Min(d / ONE_PLY, 63)][Min(mn, 63)]; }
228 // Easy move margin. An easy move candidate must be at least this much
229 // better than the second best move.
230 const Value EasyMoveMargin = Value(0x200);
233 /// Namespace variables
242 int MultiPV, UCIMultiPV;
244 // Time management variables
245 bool StopOnPonderhit, FirstRootMove, StopRequest, QuitRequest, AspirationFailLow;
250 std::ofstream LogFile;
252 // Skill level adjustment
254 bool SkillLevelEnabled;
257 // Multi-threads manager
258 ThreadsManager ThreadsMgr;
260 // Node counters, used only by thread[0] but try to keep in different cache
261 // lines (64 bytes each) from the heavy multi-thread read accessed variables.
262 bool SendSearchedNodes;
264 int NodesBetweenPolls = 30000;
272 Move id_loop(Position& pos, Move searchMoves[], Move* ponderMove);
274 template <NodeType PvNode, bool SpNode, bool Root>
275 Value search(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth);
277 template <NodeType PvNode>
278 Value qsearch(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth);
280 template <NodeType PvNode>
281 inline Value search(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth) {
283 return depth < ONE_PLY ? qsearch<PvNode>(pos, ss, alpha, beta, DEPTH_ZERO)
284 : search<PvNode, false, false>(pos, ss, alpha, beta, depth);
287 template <NodeType PvNode>
288 Depth extension(const Position& pos, Move m, bool captureOrPromotion, bool moveIsCheck, bool* dangerous);
290 bool check_is_dangerous(Position &pos, Move move, Value futilityBase, Value beta, Value *bValue);
291 bool connected_moves(const Position& pos, Move m1, Move m2);
292 Value value_to_tt(Value v, int ply);
293 Value value_from_tt(Value v, int ply);
294 bool ok_to_use_TT(const TTEntry* tte, Depth depth, Value beta, int ply);
295 bool connected_threat(const Position& pos, Move m, Move threat);
296 Value refine_eval(const TTEntry* tte, Value defaultEval, int ply);
297 void update_history(const Position& pos, Move move, Depth depth, Move movesSearched[], int moveCount);
298 void update_gains(const Position& pos, Move move, Value before, Value after);
299 void do_skill_level(Move* best, Move* ponder);
301 int current_search_time(int set = 0);
302 std::string value_to_uci(Value v);
303 std::string speed_to_uci(int64_t nodes);
304 void poll(const Position& pos);
305 void wait_for_stop_or_ponderhit();
307 #if !defined(_MSC_VER)
308 void* init_thread(void* threadID);
310 DWORD WINAPI init_thread(LPVOID threadID);
314 // MovePickerExt is an extended MovePicker class used to choose at compile time
315 // the proper move source according to the type of node.
316 template<bool SpNode, bool Root> struct MovePickerExt;
318 // In Root nodes use RootMoveList as source. Score and sort the root moves
319 // before to search them.
320 template<> struct MovePickerExt<false, true> : public MovePicker {
322 MovePickerExt(const Position& p, Move ttm, Depth d, const History& h, SearchStack* ss, Value b)
323 : MovePicker(p, ttm, d, h, ss, b), firstCall(true) {
325 Value score = VALUE_ZERO;
327 // Score root moves using standard ordering used in main search, the moves
328 // are scored according to the order in which they are returned by MovePicker.
329 // This is the second order score that is used to compare the moves when
330 // the first orders pv_score of both moves are equal.
331 while ((move = MovePicker::get_next_move()) != MOVE_NONE)
332 for (rm = Rml.begin(); rm != Rml.end(); ++rm)
333 if (rm->pv[0] == move)
335 rm->non_pv_score = score--;
343 Move get_next_move() {
350 return rm != Rml.end() ? rm->pv[0] : MOVE_NONE;
353 RootMoveList::iterator rm;
357 // In SpNodes use split point's shared MovePicker object as move source
358 template<> struct MovePickerExt<true, false> : public MovePicker {
360 MovePickerExt(const Position& p, Move ttm, Depth d, const History& h, SearchStack* ss, Value b)
361 : MovePicker(p, ttm, d, h, ss, b), mp(ss->sp->mp) {}
363 Move get_next_move() { return mp->get_next_move(); }
365 RootMoveList::iterator rm; // Dummy, needed to compile
369 // Default case, create and use a MovePicker object as source
370 template<> struct MovePickerExt<false, false> : public MovePicker {
372 MovePickerExt(const Position& p, Move ttm, Depth d, const History& h, SearchStack* ss, Value b)
373 : MovePicker(p, ttm, d, h, ss, b) {}
375 RootMoveList::iterator rm; // Dummy, needed to compile
381 /// init_threads() is called during startup. It initializes various lookup tables
382 /// and creates and launches search threads.
384 void init_threads() {
386 int d; // depth (ONE_PLY == 2)
387 int hd; // half depth (ONE_PLY == 1)
390 // Init reductions array
391 for (hd = 1; hd < 64; hd++) for (mc = 1; mc < 64; mc++)
393 double pvRed = log(double(hd)) * log(double(mc)) / 3.0;
394 double nonPVRed = 0.33 + log(double(hd)) * log(double(mc)) / 2.25;
395 ReductionMatrix[PV][hd][mc] = (int8_t) ( pvRed >= 1.0 ? floor( pvRed * int(ONE_PLY)) : 0);
396 ReductionMatrix[NonPV][hd][mc] = (int8_t) (nonPVRed >= 1.0 ? floor(nonPVRed * int(ONE_PLY)) : 0);
399 // Init futility margins array
400 for (d = 1; d < 16; d++) for (mc = 0; mc < 64; mc++)
401 FutilityMarginsMatrix[d][mc] = Value(112 * int(log(double(d * d) / 2) / log(2.0) + 1.001) - 8 * mc + 45);
403 // Init futility move count array
404 for (d = 0; d < 32; d++)
405 FutilityMoveCountArray[d] = int(3.001 + 0.25 * pow(d, 2.0));
407 // Create and startup threads
408 ThreadsMgr.init_threads();
412 /// exit_threads() is a trampoline to access ThreadsMgr from outside of current file
413 void exit_threads() { ThreadsMgr.exit_threads(); }
416 /// perft() is our utility to verify move generation. All the legal moves up to
417 /// given depth are generated and counted and the sum returned.
419 int64_t perft(Position& pos, Depth depth) {
421 MoveStack mlist[MAX_MOVES];
426 // Generate all legal moves
427 MoveStack* last = generate<MV_LEGAL>(pos, mlist);
429 // If we are at the last ply we don't need to do and undo
430 // the moves, just to count them.
431 if (depth <= ONE_PLY)
432 return int(last - mlist);
434 // Loop through all legal moves
436 for (MoveStack* cur = mlist; cur != last; cur++)
439 pos.do_move(m, st, ci, pos.move_is_check(m, ci));
440 sum += perft(pos, depth - ONE_PLY);
447 /// think() is the external interface to Stockfish's search, and is called when
448 /// the program receives the UCI 'go' command. It initializes various global
449 /// variables, and calls id_loop(). It returns false when a "quit" command is
450 /// received during the search.
452 bool think(Position& pos, const SearchLimits& limits, Move searchMoves[]) {
454 // Initialize global search-related variables
455 StopOnPonderhit = StopRequest = QuitRequest = AspirationFailLow = SendSearchedNodes = false;
457 current_search_time(get_system_time());
459 TimeMgr.init(Limits, pos.startpos_ply_counter());
461 // Set best NodesBetweenPolls interval to avoid lagging under time pressure
463 NodesBetweenPolls = Min(Limits.maxNodes, 30000);
464 else if (Limits.time && Limits.time < 1000)
465 NodesBetweenPolls = 1000;
466 else if (Limits.time && Limits.time < 5000)
467 NodesBetweenPolls = 5000;
469 NodesBetweenPolls = 30000;
471 // Look for a book move, only during games, not tests
472 if (Limits.useTimeManagement() && Options["OwnBook"].value<bool>())
474 if (Options["Book File"].value<std::string>() != OpeningBook.name())
475 OpeningBook.open(Options["Book File"].value<std::string>());
477 Move bookMove = OpeningBook.get_move(pos, Options["Best Book Move"].value<bool>());
478 if (bookMove != MOVE_NONE)
481 wait_for_stop_or_ponderhit();
483 cout << "bestmove " << bookMove << endl;
489 UCIMultiPV = Options["MultiPV"].value<int>();
490 SkillLevel = Options["Skill level"].value<int>();
492 read_evaluation_uci_options(pos.side_to_move());
494 if (Options["Clear Hash"].value<bool>())
496 Options["Clear Hash"].set_value("false");
499 TT.set_size(Options["Hash"].value<int>());
501 // Do we have to play with skill handicap? In this case enable MultiPV that
502 // we will use behind the scenes to retrieve a set of possible moves.
503 SkillLevelEnabled = (SkillLevel < 20);
504 MultiPV = (SkillLevelEnabled ? Max(UCIMultiPV, 4) : UCIMultiPV);
506 // Set the number of active threads
507 ThreadsMgr.read_uci_options();
508 init_eval(ThreadsMgr.active_threads());
510 // Wake up needed threads and reset maxPly counter
511 for (int i = 0; i < ThreadsMgr.active_threads(); i++)
513 ThreadsMgr[i].wake_up();
514 ThreadsMgr[i].maxPly = 0;
517 // Write to log file and keep it open to be accessed during the search
518 if (Options["Use Search Log"].value<bool>())
520 std::string name = Options["Search Log Filename"].value<std::string>();
521 LogFile.open(name.c_str(), std::ios::out | std::ios::app);
523 if (LogFile.is_open())
524 LogFile << "\nSearching: " << pos.to_fen()
525 << "\ninfinite: " << Limits.infinite
526 << " ponder: " << Limits.ponder
527 << " time: " << Limits.time
528 << " increment: " << Limits.increment
529 << " moves to go: " << Limits.movesToGo
533 // We're ready to start thinking. Call the iterative deepening loop function
534 Move ponderMove = MOVE_NONE;
535 Move bestMove = id_loop(pos, searchMoves, &ponderMove);
537 cout << "info" << speed_to_uci(pos.nodes_searched()) << endl;
539 // Write final search statistics and close log file
540 if (LogFile.is_open())
542 int t = current_search_time();
544 LogFile << "Nodes: " << pos.nodes_searched()
545 << "\nNodes/second: " << (t > 0 ? pos.nodes_searched() * 1000 / t : 0)
546 << "\nBest move: " << move_to_san(pos, bestMove);
549 pos.do_move(bestMove, st);
550 LogFile << "\nPonder move: " << move_to_san(pos, ponderMove) << endl;
551 pos.undo_move(bestMove); // Return from think() with unchanged position
555 // This makes all the threads to go to sleep
556 ThreadsMgr.set_active_threads(1);
558 // If we are pondering or in infinite search, we shouldn't print the
559 // best move before we are told to do so.
560 if (!StopRequest && (Limits.ponder || Limits.infinite))
561 wait_for_stop_or_ponderhit();
563 // Could be MOVE_NONE when searching on a stalemate position
564 cout << "bestmove " << bestMove;
566 // UCI protol is not clear on allowing sending an empty ponder move, instead
567 // it is clear that ponder move is optional. So skip it if empty.
568 if (ponderMove != MOVE_NONE)
569 cout << " ponder " << ponderMove;
579 // id_loop() is the main iterative deepening loop. It calls search() repeatedly
580 // with increasing depth until the allocated thinking time has been consumed,
581 // user stops the search, or the maximum search depth is reached.
583 Move id_loop(Position& pos, Move searchMoves[], Move* ponderMove) {
585 SearchStack ss[PLY_MAX_PLUS_2];
586 Value bestValues[PLY_MAX_PLUS_2];
587 int bestMoveChanges[PLY_MAX_PLUS_2];
588 int depth, selDepth, aspirationDelta;
589 Value value, alpha, beta;
590 Move bestMove, easyMove, skillBest, skillPonder;
592 // Initialize stuff before a new search
593 memset(ss, 0, 4 * sizeof(SearchStack));
596 *ponderMove = bestMove = easyMove = skillBest = skillPonder = MOVE_NONE;
597 depth = aspirationDelta = 0;
598 alpha = -VALUE_INFINITE, beta = VALUE_INFINITE;
599 ss->currentMove = MOVE_NULL; // Hack to skip update_gains()
601 // Moves to search are verified and copied
602 Rml.init(pos, searchMoves);
604 // Handle special case of searching on a mate/stalemate position
607 cout << "info depth 0 score "
608 << value_to_uci(pos.is_check() ? -VALUE_MATE : VALUE_DRAW)
614 // Iterative deepening loop until requested to stop or target depth reached
615 while (!StopRequest && ++depth <= PLY_MAX && (!Limits.maxDepth || depth <= Limits.maxDepth))
617 Rml.bestMoveChanges = 0;
618 cout << set960(pos.is_chess960()) << "info depth " << depth << endl;
620 // Calculate dynamic aspiration window based on previous iterations
621 if (MultiPV == 1 && depth >= 5 && abs(bestValues[depth - 1]) < VALUE_KNOWN_WIN)
623 int prevDelta1 = bestValues[depth - 1] - bestValues[depth - 2];
624 int prevDelta2 = bestValues[depth - 2] - bestValues[depth - 3];
626 aspirationDelta = Min(Max(abs(prevDelta1) + abs(prevDelta2) / 2, 16), 24);
627 aspirationDelta = (aspirationDelta + 7) / 8 * 8; // Round to match grainSize
629 alpha = Max(bestValues[depth - 1] - aspirationDelta, -VALUE_INFINITE);
630 beta = Min(bestValues[depth - 1] + aspirationDelta, VALUE_INFINITE);
633 // Start with a small aspiration window and, in case of fail high/low,
634 // research with bigger window until not failing high/low anymore.
636 // Search starting from ss+1 to allow calling update_gains()
637 value = search<PV, false, true>(pos, ss+1, alpha, beta, depth * ONE_PLY);
639 // Write PV back to transposition table in case the relevant entries
640 // have been overwritten during the search.
641 for (int i = 0; i < Min(MultiPV, (int)Rml.size()); i++)
642 Rml[i].insert_pv_in_tt(pos);
644 // Value cannot be trusted. Break out immediately!
648 assert(value >= alpha);
650 // In case of failing high/low increase aspiration window and research,
651 // otherwise exit the fail high/low loop.
654 beta = Min(beta + aspirationDelta, VALUE_INFINITE);
655 aspirationDelta += aspirationDelta / 2;
657 else if (value <= alpha)
659 AspirationFailLow = true;
660 StopOnPonderhit = false;
662 alpha = Max(alpha - aspirationDelta, -VALUE_INFINITE);
663 aspirationDelta += aspirationDelta / 2;
668 } while (abs(value) < VALUE_KNOWN_WIN);
670 // Collect info about search result
671 bestMove = Rml[0].pv[0];
672 *ponderMove = Rml[0].pv[1];
673 bestValues[depth] = value;
674 bestMoveChanges[depth] = Rml.bestMoveChanges;
676 // Do we need to pick now the best and the ponder moves ?
677 if (SkillLevelEnabled && depth == 1 + SkillLevel)
678 do_skill_level(&skillBest, &skillPonder);
680 // Retrieve max searched depth among threads
682 for (int i = 0; i < ThreadsMgr.active_threads(); i++)
683 if (ThreadsMgr[i].maxPly > selDepth)
684 selDepth = ThreadsMgr[i].maxPly;
686 // Send PV line to GUI and to log file
687 for (int i = 0; i < Min(UCIMultiPV, (int)Rml.size()); i++)
688 cout << Rml[i].pv_info_to_uci(pos, depth, selDepth, alpha, beta, i) << endl;
690 if (LogFile.is_open())
691 LogFile << pretty_pv(pos, depth, value, current_search_time(), Rml[0].pv) << endl;
693 // Init easyMove after first iteration or drop if differs from the best move
694 if (depth == 1 && (Rml.size() == 1 || Rml[0].pv_score > Rml[1].pv_score + EasyMoveMargin))
696 else if (bestMove != easyMove)
697 easyMove = MOVE_NONE;
699 // Check for some early stop condition
700 if (!StopRequest && Limits.useTimeManagement())
702 // Stop search early when the last two iterations returned a mate score
704 && abs(bestValues[depth]) >= VALUE_MATE_IN_PLY_MAX
705 && abs(bestValues[depth - 1]) >= VALUE_MATE_IN_PLY_MAX)
708 // Stop search early if one move seems to be much better than the
709 // others or if there is only a single legal move. Also in the latter
710 // case we search up to some depth anyway to get a proper score.
712 && easyMove == bestMove
714 ||( Rml[0].nodes > (pos.nodes_searched() * 85) / 100
715 && current_search_time() > TimeMgr.available_time() / 16)
716 ||( Rml[0].nodes > (pos.nodes_searched() * 98) / 100
717 && current_search_time() > TimeMgr.available_time() / 32)))
720 // Take in account some extra time if the best move has changed
721 if (depth > 4 && depth < 50)
722 TimeMgr.pv_instability(bestMoveChanges[depth], bestMoveChanges[depth - 1]);
724 // Stop search if most of available time is already consumed. We probably don't
725 // have enough time to search the first move at the next iteration anyway.
726 if (current_search_time() > (TimeMgr.available_time() * 62) / 100)
729 // If we are allowed to ponder do not stop the search now but keep pondering
730 if (StopRequest && Limits.ponder)
733 StopOnPonderhit = true;
738 // When using skills overwrite best and ponder moves with the sub-optimal ones
739 if (SkillLevelEnabled)
741 if (skillBest == MOVE_NONE) // Still unassigned ?
742 do_skill_level(&skillBest, &skillPonder);
744 bestMove = skillBest;
745 *ponderMove = skillPonder;
752 // search<>() is the main search function for both PV and non-PV nodes and for
753 // normal and SplitPoint nodes. When called just after a split point the search
754 // is simpler because we have already probed the hash table, done a null move
755 // search, and searched the first move before splitting, we don't have to repeat
756 // all this work again. We also don't need to store anything to the hash table
757 // here: This is taken care of after we return from the split point.
759 template <NodeType PvNode, bool SpNode, bool Root>
760 Value search(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth) {
762 assert(alpha >= -VALUE_INFINITE && alpha <= VALUE_INFINITE);
763 assert(beta > alpha && beta <= VALUE_INFINITE);
764 assert(PvNode || alpha == beta - 1);
765 assert(pos.thread() >= 0 && pos.thread() < ThreadsMgr.active_threads());
767 Move movesSearched[MAX_MOVES];
772 Move ttMove, move, excludedMove, threatMove;
775 Value bestValue, value, oldAlpha;
776 Value refinedValue, nullValue, futilityBase, futilityValueScaled; // Non-PV specific
777 bool isPvMove, isCheck, singularExtensionNode, moveIsCheck, captureOrPromotion, dangerous, isBadCap;
778 int moveCount = 0, playedMoveCount = 0;
779 int threadID = pos.thread();
780 SplitPoint* sp = NULL;
782 refinedValue = bestValue = value = -VALUE_INFINITE;
784 isCheck = pos.is_check();
785 ss->ply = (ss-1)->ply + 1;
787 // Used to send selDepth info to GUI
788 if (PvNode && ThreadsMgr[threadID].maxPly < ss->ply)
789 ThreadsMgr[threadID].maxPly = ss->ply;
795 ttMove = excludedMove = MOVE_NONE;
796 threatMove = sp->threatMove;
797 goto split_point_start;
802 // Step 1. Initialize node and poll. Polling can abort search
803 ss->currentMove = ss->bestMove = threatMove = (ss+1)->excludedMove = MOVE_NONE;
804 (ss+1)->skipNullMove = false; (ss+1)->reduction = DEPTH_ZERO;
805 (ss+2)->killers[0] = (ss+2)->killers[1] = (ss+2)->mateKiller = MOVE_NONE;
807 if (threadID == 0 && ++NodesSincePoll > NodesBetweenPolls)
813 // Step 2. Check for aborted search and immediate draw
815 || ThreadsMgr.cutoff_at_splitpoint(threadID)
817 || ss->ply > PLY_MAX) && !Root)
820 // Step 3. Mate distance pruning
821 alpha = Max(value_mated_in(ss->ply), alpha);
822 beta = Min(value_mate_in(ss->ply+1), beta);
826 // Step 4. Transposition table lookup
827 // We don't want the score of a partial search to overwrite a previous full search
828 // TT value, so we use a different position key in case of an excluded move.
829 excludedMove = ss->excludedMove;
830 posKey = excludedMove ? pos.get_exclusion_key() : pos.get_key();
832 tte = TT.retrieve(posKey);
833 ttMove = tte ? tte->move() : MOVE_NONE;
835 // At PV nodes we check for exact scores, while at non-PV nodes we check for
836 // a fail high/low. Biggest advantage at probing at PV nodes is to have a
837 // smooth experience in analysis mode.
840 && (PvNode ? tte->depth() >= depth && tte->type() == VALUE_TYPE_EXACT
841 : ok_to_use_TT(tte, depth, beta, ss->ply)))
844 ss->bestMove = ttMove; // Can be MOVE_NONE
845 return value_from_tt(tte->value(), ss->ply);
848 // Step 5. Evaluate the position statically and update parent's gain statistics
850 ss->eval = ss->evalMargin = VALUE_NONE;
853 assert(tte->static_value() != VALUE_NONE);
855 ss->eval = tte->static_value();
856 ss->evalMargin = tte->static_value_margin();
857 refinedValue = refine_eval(tte, ss->eval, ss->ply);
861 refinedValue = ss->eval = evaluate(pos, ss->evalMargin);
862 TT.store(posKey, VALUE_NONE, VALUE_TYPE_NONE, DEPTH_NONE, MOVE_NONE, ss->eval, ss->evalMargin);
865 // Save gain for the parent non-capture move
866 update_gains(pos, (ss-1)->currentMove, (ss-1)->eval, ss->eval);
868 // Step 6. Razoring (is omitted in PV nodes)
870 && depth < RazorDepth
872 && refinedValue + razor_margin(depth) < beta
873 && ttMove == MOVE_NONE
874 && abs(beta) < VALUE_MATE_IN_PLY_MAX
875 && !pos.has_pawn_on_7th(pos.side_to_move()))
877 Value rbeta = beta - razor_margin(depth);
878 Value v = qsearch<NonPV>(pos, ss, rbeta-1, rbeta, DEPTH_ZERO);
880 // Logically we should return (v + razor_margin(depth)), but
881 // surprisingly this did slightly weaker in tests.
885 // Step 7. Static null move pruning (is omitted in PV nodes)
886 // We're betting that the opponent doesn't have a move that will reduce
887 // the score by more than futility_margin(depth) if we do a null move.
890 && depth < RazorDepth
892 && refinedValue - futility_margin(depth, 0) >= beta
893 && abs(beta) < VALUE_MATE_IN_PLY_MAX
894 && pos.non_pawn_material(pos.side_to_move()))
895 return refinedValue - futility_margin(depth, 0);
897 // Step 8. Null move search with verification search (is omitted in PV nodes)
902 && refinedValue >= beta
903 && abs(beta) < VALUE_MATE_IN_PLY_MAX
904 && pos.non_pawn_material(pos.side_to_move()))
906 ss->currentMove = MOVE_NULL;
908 // Null move dynamic reduction based on depth
909 int R = 3 + (depth >= 5 * ONE_PLY ? depth / 8 : 0);
911 // Null move dynamic reduction based on value
912 if (refinedValue - PawnValueMidgame > beta)
915 pos.do_null_move(st);
916 (ss+1)->skipNullMove = true;
917 nullValue = -search<NonPV>(pos, ss+1, -beta, -alpha, depth-R*ONE_PLY);
918 (ss+1)->skipNullMove = false;
919 pos.undo_null_move();
921 if (nullValue >= beta)
923 // Do not return unproven mate scores
924 if (nullValue >= VALUE_MATE_IN_PLY_MAX)
927 if (depth < 6 * ONE_PLY)
930 // Do verification search at high depths
931 ss->skipNullMove = true;
932 Value v = search<NonPV>(pos, ss, alpha, beta, depth-R*ONE_PLY);
933 ss->skipNullMove = false;
940 // The null move failed low, which means that we may be faced with
941 // some kind of threat. If the previous move was reduced, check if
942 // the move that refuted the null move was somehow connected to the
943 // move which was reduced. If a connection is found, return a fail
944 // low score (which will cause the reduced move to fail high in the
945 // parent node, which will trigger a re-search with full depth).
946 threatMove = (ss+1)->bestMove;
948 if ( depth < ThreatDepth
950 && threatMove != MOVE_NONE
951 && connected_moves(pos, (ss-1)->currentMove, threatMove))
956 // Step 9. Internal iterative deepening
957 if ( depth >= IIDDepth[PvNode]
958 && ttMove == MOVE_NONE
959 && (PvNode || (!isCheck && ss->eval + IIDMargin >= beta)))
961 Depth d = (PvNode ? depth - 2 * ONE_PLY : depth / 2);
963 ss->skipNullMove = true;
964 search<PvNode>(pos, ss, alpha, beta, d);
965 ss->skipNullMove = false;
967 ttMove = ss->bestMove;
968 tte = TT.retrieve(posKey);
971 split_point_start: // At split points actual search starts from here
973 // Initialize a MovePicker object for the current position
974 MovePickerExt<SpNode, Root> mp(pos, ttMove, depth, H, ss, (PvNode ? -VALUE_INFINITE : beta));
976 ss->bestMove = MOVE_NONE;
977 futilityBase = ss->eval + ss->evalMargin;
978 singularExtensionNode = !Root
980 && depth >= SingularExtensionDepth[PvNode]
983 && !excludedMove // Do not allow recursive singular extension search
984 && (tte->type() & VALUE_TYPE_LOWER)
985 && tte->depth() >= depth - 3 * ONE_PLY;
988 lock_grab(&(sp->lock));
989 bestValue = sp->bestValue;
992 // Step 10. Loop through moves
993 // Loop through all legal moves until no moves remain or a beta cutoff occurs
994 while ( bestValue < beta
995 && (move = mp.get_next_move()) != MOVE_NONE
996 && !ThreadsMgr.cutoff_at_splitpoint(threadID))
998 assert(move_is_ok(move));
1002 moveCount = ++sp->moveCount;
1003 lock_release(&(sp->lock));
1005 else if (move == excludedMove)
1012 // This is used by time management
1013 FirstRootMove = (moveCount == 1);
1015 // Save the current node count before the move is searched
1016 nodes = pos.nodes_searched();
1018 // If it's time to send nodes info, do it here where we have the
1019 // correct accumulated node counts searched by each thread.
1020 if (SendSearchedNodes)
1022 SendSearchedNodes = false;
1023 cout << "info" << speed_to_uci(pos.nodes_searched()) << endl;
1026 if (current_search_time() > 2000)
1027 cout << "info currmove " << move
1028 << " currmovenumber " << moveCount << endl;
1031 // At Root and at first iteration do a PV search on all the moves to score root moves
1032 isPvMove = (PvNode && moveCount <= (Root ? depth <= ONE_PLY ? 1000 : MultiPV : 1));
1033 moveIsCheck = pos.move_is_check(move, ci);
1034 captureOrPromotion = pos.move_is_capture_or_promotion(move);
1036 // Step 11. Decide the new search depth
1037 ext = extension<PvNode>(pos, move, captureOrPromotion, moveIsCheck, &dangerous);
1039 // Singular extension search. If all moves but one fail low on a search of
1040 // (alpha-s, beta-s), and just one fails high on (alpha, beta), then that move
1041 // is singular and should be extended. To verify this we do a reduced search
1042 // on all the other moves but the ttMove, if result is lower than ttValue minus
1043 // a margin then we extend ttMove.
1044 if ( singularExtensionNode
1045 && move == tte->move()
1048 Value ttValue = value_from_tt(tte->value(), ss->ply);
1050 if (abs(ttValue) < VALUE_KNOWN_WIN)
1052 Value rBeta = ttValue - int(depth);
1053 ss->excludedMove = move;
1054 ss->skipNullMove = true;
1055 Value v = search<NonPV>(pos, ss, rBeta - 1, rBeta, depth / 2);
1056 ss->skipNullMove = false;
1057 ss->excludedMove = MOVE_NONE;
1058 ss->bestMove = MOVE_NONE;
1064 // Update current move (this must be done after singular extension search)
1065 ss->currentMove = move;
1066 newDepth = depth - ONE_PLY + ext;
1068 // Step 12. Futility pruning (is omitted in PV nodes)
1070 && !captureOrPromotion
1074 && !move_is_castle(move))
1076 // Move count based pruning
1077 if ( moveCount >= futility_move_count(depth)
1078 && (!threatMove || !connected_threat(pos, move, threatMove))
1079 && bestValue > VALUE_MATED_IN_PLY_MAX) // FIXME bestValue is racy
1082 lock_grab(&(sp->lock));
1087 // Value based pruning
1088 // We illogically ignore reduction condition depth >= 3*ONE_PLY for predicted depth,
1089 // but fixing this made program slightly weaker.
1090 Depth predictedDepth = newDepth - reduction<NonPV>(depth, moveCount);
1091 futilityValueScaled = futilityBase + futility_margin(predictedDepth, moveCount)
1092 + H.gain(pos.piece_on(move_from(move)), move_to(move));
1094 if (futilityValueScaled < beta)
1098 lock_grab(&(sp->lock));
1099 if (futilityValueScaled > sp->bestValue)
1100 sp->bestValue = bestValue = futilityValueScaled;
1102 else if (futilityValueScaled > bestValue)
1103 bestValue = futilityValueScaled;
1108 // Prune moves with negative SEE at low depths
1109 if ( predictedDepth < 2 * ONE_PLY
1110 && bestValue > VALUE_MATED_IN_PLY_MAX
1111 && pos.see_sign(move) < 0)
1114 lock_grab(&(sp->lock));
1120 // Bad capture detection. Will be used by prob-cut search
1121 isBadCap = depth >= 3 * ONE_PLY
1122 && depth < 8 * ONE_PLY
1123 && captureOrPromotion
1126 && !move_is_promotion(move)
1127 && abs(alpha) < VALUE_MATE_IN_PLY_MAX
1128 && pos.see_sign(move) < 0;
1130 // Step 13. Make the move
1131 pos.do_move(move, st, ci, moveIsCheck);
1133 if (!SpNode && !captureOrPromotion)
1134 movesSearched[playedMoveCount++] = move;
1136 // Step extra. pv search (only in PV nodes)
1137 // The first move in list is the expected PV
1140 // Aspiration window is disabled in multi-pv case
1141 if (Root && MultiPV > 1)
1142 alpha = -VALUE_INFINITE;
1144 value = -search<PV>(pos, ss+1, -beta, -alpha, newDepth);
1148 // Step 14. Reduced depth search
1149 // If the move fails high will be re-searched at full depth.
1150 bool doFullDepthSearch = true;
1151 alpha = SpNode ? sp->alpha : alpha;
1153 if ( depth >= 3 * ONE_PLY
1154 && !captureOrPromotion
1156 && !move_is_castle(move)
1157 && ss->killers[0] != move
1158 && ss->killers[1] != move)
1160 ss->reduction = reduction<PvNode>(depth, moveCount);
1163 alpha = SpNode ? sp->alpha : alpha;
1164 Depth d = newDepth - ss->reduction;
1165 value = -search<NonPV>(pos, ss+1, -(alpha+1), -alpha, d);
1167 doFullDepthSearch = (value > alpha);
1169 ss->reduction = DEPTH_ZERO; // Restore original reduction
1172 // Probcut search for bad captures. If a reduced search returns a value
1173 // very below beta then we can (almost) safely prune the bad capture.
1176 ss->reduction = 3 * ONE_PLY;
1177 Value rAlpha = alpha - 300;
1178 Depth d = newDepth - ss->reduction;
1179 value = -search<NonPV>(pos, ss+1, -(rAlpha+1), -rAlpha, d);
1180 doFullDepthSearch = (value > rAlpha);
1181 ss->reduction = DEPTH_ZERO; // Restore original reduction
1184 // Step 15. Full depth search
1185 if (doFullDepthSearch)
1187 alpha = SpNode ? sp->alpha : alpha;
1188 value = -search<NonPV>(pos, ss+1, -(alpha+1), -alpha, newDepth);
1190 // Step extra. pv search (only in PV nodes)
1191 // Search only for possible new PV nodes, if instead value >= beta then
1192 // parent node fails low with value <= alpha and tries another move.
1193 if (PvNode && value > alpha && (Root || value < beta))
1194 value = -search<PV>(pos, ss+1, -beta, -alpha, newDepth);
1198 // Step 16. Undo move
1199 pos.undo_move(move);
1201 assert(value > -VALUE_INFINITE && value < VALUE_INFINITE);
1203 // Step 17. Check for new best move
1206 lock_grab(&(sp->lock));
1207 bestValue = sp->bestValue;
1211 if (value > bestValue && !(SpNode && ThreadsMgr.cutoff_at_splitpoint(threadID)))
1216 sp->bestValue = value;
1218 if (!Root && value > alpha)
1220 if (PvNode && value < beta) // We want always alpha < beta
1228 sp->betaCutoff = true;
1230 if (value == value_mate_in(ss->ply + 1))
1231 ss->mateKiller = move;
1233 ss->bestMove = move;
1236 sp->ss->bestMove = move;
1242 // Finished searching the move. If StopRequest is true, the search
1243 // was aborted because the user interrupted the search or because we
1244 // ran out of time. In this case, the return value of the search cannot
1245 // be trusted, and we break out of the loop without updating the best
1250 // Remember searched nodes counts for this move
1251 mp.rm->nodes += pos.nodes_searched() - nodes;
1253 // PV move or new best move ?
1254 if (isPvMove || value > alpha)
1257 ss->bestMove = move;
1258 mp.rm->pv_score = value;
1259 mp.rm->extract_pv_from_tt(pos);
1261 // We record how often the best move has been changed in each
1262 // iteration. This information is used for time management: When
1263 // the best move changes frequently, we allocate some more time.
1264 if (!isPvMove && MultiPV == 1)
1265 Rml.bestMoveChanges++;
1267 Rml.sort_multipv(moveCount);
1269 // Update alpha. In multi-pv we don't use aspiration window, so
1270 // set alpha equal to minimum score among the PV lines.
1272 alpha = Rml[Min(moveCount, MultiPV) - 1].pv_score; // FIXME why moveCount?
1273 else if (value > alpha)
1277 mp.rm->pv_score = -VALUE_INFINITE;
1281 // Step 18. Check for split
1284 && depth >= ThreadsMgr.min_split_depth()
1285 && ThreadsMgr.active_threads() > 1
1287 && ThreadsMgr.available_thread_exists(threadID)
1289 && !ThreadsMgr.cutoff_at_splitpoint(threadID))
1290 ThreadsMgr.split<FakeSplit>(pos, ss, &alpha, beta, &bestValue, depth,
1291 threatMove, moveCount, &mp, PvNode);
1294 // Step 19. Check for mate and stalemate
1295 // All legal moves have been searched and if there are
1296 // no legal moves, it must be mate or stalemate.
1297 // If one move was excluded return fail low score.
1298 if (!SpNode && !moveCount)
1299 return excludedMove ? oldAlpha : isCheck ? value_mated_in(ss->ply) : VALUE_DRAW;
1301 // Step 20. Update tables
1302 // If the search is not aborted, update the transposition table,
1303 // history counters, and killer moves.
1304 if (!SpNode && !StopRequest && !ThreadsMgr.cutoff_at_splitpoint(threadID))
1306 move = bestValue <= oldAlpha ? MOVE_NONE : ss->bestMove;
1307 vt = bestValue <= oldAlpha ? VALUE_TYPE_UPPER
1308 : bestValue >= beta ? VALUE_TYPE_LOWER : VALUE_TYPE_EXACT;
1310 TT.store(posKey, value_to_tt(bestValue, ss->ply), vt, depth, move, ss->eval, ss->evalMargin);
1312 // Update killers and history only for non capture moves that fails high
1313 if ( bestValue >= beta
1314 && !pos.move_is_capture_or_promotion(move))
1316 if (move != ss->killers[0])
1318 ss->killers[1] = ss->killers[0];
1319 ss->killers[0] = move;
1321 update_history(pos, move, depth, movesSearched, playedMoveCount);
1327 // Here we have the lock still grabbed
1328 sp->slaves[threadID] = 0;
1329 sp->nodes += pos.nodes_searched();
1330 lock_release(&(sp->lock));
1333 assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE);
1338 // qsearch() is the quiescence search function, which is called by the main
1339 // search function when the remaining depth is zero (or, to be more precise,
1340 // less than ONE_PLY).
1342 template <NodeType PvNode>
1343 Value qsearch(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth) {
1345 assert(alpha >= -VALUE_INFINITE && alpha <= VALUE_INFINITE);
1346 assert(beta >= -VALUE_INFINITE && beta <= VALUE_INFINITE);
1347 assert(PvNode || alpha == beta - 1);
1349 assert(pos.thread() >= 0 && pos.thread() < ThreadsMgr.active_threads());
1353 Value bestValue, value, evalMargin, futilityValue, futilityBase;
1354 bool isCheck, enoughMaterial, moveIsCheck, evasionPrunable;
1357 Value oldAlpha = alpha;
1359 ss->bestMove = ss->currentMove = MOVE_NONE;
1360 ss->ply = (ss-1)->ply + 1;
1362 // Check for an instant draw or maximum ply reached
1363 if (ss->ply > PLY_MAX || pos.is_draw())
1366 // Decide whether or not to include checks, this fixes also the type of
1367 // TT entry depth that we are going to use. Note that in qsearch we use
1368 // only two types of depth in TT: DEPTH_QS_CHECKS or DEPTH_QS_NO_CHECKS.
1369 isCheck = pos.is_check();
1370 ttDepth = (isCheck || depth >= DEPTH_QS_CHECKS ? DEPTH_QS_CHECKS : DEPTH_QS_NO_CHECKS);
1372 // Transposition table lookup. At PV nodes, we don't use the TT for
1373 // pruning, but only for move ordering.
1374 tte = TT.retrieve(pos.get_key());
1375 ttMove = (tte ? tte->move() : MOVE_NONE);
1377 if (!PvNode && tte && ok_to_use_TT(tte, ttDepth, beta, ss->ply))
1379 ss->bestMove = ttMove; // Can be MOVE_NONE
1380 return value_from_tt(tte->value(), ss->ply);
1383 // Evaluate the position statically
1386 bestValue = futilityBase = -VALUE_INFINITE;
1387 ss->eval = evalMargin = VALUE_NONE;
1388 enoughMaterial = false;
1394 assert(tte->static_value() != VALUE_NONE);
1396 evalMargin = tte->static_value_margin();
1397 ss->eval = bestValue = tte->static_value();
1400 ss->eval = bestValue = evaluate(pos, evalMargin);
1402 update_gains(pos, (ss-1)->currentMove, (ss-1)->eval, ss->eval);
1404 // Stand pat. Return immediately if static value is at least beta
1405 if (bestValue >= beta)
1408 TT.store(pos.get_key(), value_to_tt(bestValue, ss->ply), VALUE_TYPE_LOWER, DEPTH_NONE, MOVE_NONE, ss->eval, evalMargin);
1413 if (PvNode && bestValue > alpha)
1416 // Futility pruning parameters, not needed when in check
1417 futilityBase = ss->eval + evalMargin + FutilityMarginQS;
1418 enoughMaterial = pos.non_pawn_material(pos.side_to_move()) > RookValueMidgame;
1421 // Initialize a MovePicker object for the current position, and prepare
1422 // to search the moves. Because the depth is <= 0 here, only captures,
1423 // queen promotions and checks (only if depth >= DEPTH_QS_CHECKS) will
1425 MovePicker mp(pos, ttMove, depth, H);
1428 // Loop through the moves until no moves remain or a beta cutoff occurs
1429 while ( alpha < beta
1430 && (move = mp.get_next_move()) != MOVE_NONE)
1432 assert(move_is_ok(move));
1434 moveIsCheck = pos.move_is_check(move, ci);
1442 && !move_is_promotion(move)
1443 && !pos.move_is_passed_pawn_push(move))
1445 futilityValue = futilityBase
1446 + pos.endgame_value_of_piece_on(move_to(move))
1447 + (move_is_ep(move) ? PawnValueEndgame : VALUE_ZERO);
1449 if (futilityValue < alpha)
1451 if (futilityValue > bestValue)
1452 bestValue = futilityValue;
1456 // Prune moves with negative or equal SEE
1457 if ( futilityBase < beta
1458 && depth < DEPTH_ZERO
1459 && pos.see(move) <= 0)
1463 // Detect non-capture evasions that are candidate to be pruned
1464 evasionPrunable = isCheck
1465 && bestValue > VALUE_MATED_IN_PLY_MAX
1466 && !pos.move_is_capture(move)
1467 && !pos.can_castle(pos.side_to_move());
1469 // Don't search moves with negative SEE values
1471 && (!isCheck || evasionPrunable)
1473 && !move_is_promotion(move)
1474 && pos.see_sign(move) < 0)
1477 // Don't search useless checks
1482 && !pos.move_is_capture_or_promotion(move)
1483 && ss->eval + PawnValueMidgame / 4 < beta
1484 && !check_is_dangerous(pos, move, futilityBase, beta, &bestValue))
1486 if (ss->eval + PawnValueMidgame / 4 > bestValue)
1487 bestValue = ss->eval + PawnValueMidgame / 4;
1492 // Update current move
1493 ss->currentMove = move;
1495 // Make and search the move
1496 pos.do_move(move, st, ci, moveIsCheck);
1497 value = -qsearch<PvNode>(pos, ss+1, -beta, -alpha, depth-ONE_PLY);
1498 pos.undo_move(move);
1500 assert(value > -VALUE_INFINITE && value < VALUE_INFINITE);
1503 if (value > bestValue)
1509 ss->bestMove = move;
1514 // All legal moves have been searched. A special case: If we're in check
1515 // and no legal moves were found, it is checkmate.
1516 if (isCheck && bestValue == -VALUE_INFINITE)
1517 return value_mated_in(ss->ply);
1519 // Update transposition table
1520 ValueType vt = (bestValue <= oldAlpha ? VALUE_TYPE_UPPER : bestValue >= beta ? VALUE_TYPE_LOWER : VALUE_TYPE_EXACT);
1521 TT.store(pos.get_key(), value_to_tt(bestValue, ss->ply), vt, ttDepth, ss->bestMove, ss->eval, evalMargin);
1523 assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE);
1529 // check_is_dangerous() tests if a checking move can be pruned in qsearch().
1530 // bestValue is updated only when returning false because in that case move
1533 bool check_is_dangerous(Position &pos, Move move, Value futilityBase, Value beta, Value *bestValue)
1535 Bitboard b, occ, oldAtt, newAtt, kingAtt;
1536 Square from, to, ksq, victimSq;
1539 Value futilityValue, bv = *bestValue;
1541 from = move_from(move);
1543 them = opposite_color(pos.side_to_move());
1544 ksq = pos.king_square(them);
1545 kingAtt = pos.attacks_from<KING>(ksq);
1546 pc = pos.piece_on(from);
1548 occ = pos.occupied_squares() & ~(1ULL << from) & ~(1ULL << ksq);
1549 oldAtt = pos.attacks_from(pc, from, occ);
1550 newAtt = pos.attacks_from(pc, to, occ);
1552 // Rule 1. Checks which give opponent's king at most one escape square are dangerous
1553 b = kingAtt & ~pos.pieces_of_color(them) & ~newAtt & ~(1ULL << to);
1555 if (!(b && (b & (b - 1))))
1558 // Rule 2. Queen contact check is very dangerous
1559 if ( type_of_piece(pc) == QUEEN
1560 && bit_is_set(kingAtt, to))
1563 // Rule 3. Creating new double threats with checks
1564 b = pos.pieces_of_color(them) & newAtt & ~oldAtt & ~(1ULL << ksq);
1568 victimSq = pop_1st_bit(&b);
1569 futilityValue = futilityBase + pos.endgame_value_of_piece_on(victimSq);
1571 // Note that here we generate illegal "double move"!
1572 if ( futilityValue >= beta
1573 && pos.see_sign(make_move(from, victimSq)) >= 0)
1576 if (futilityValue > bv)
1580 // Update bestValue only if check is not dangerous (because we will prune the move)
1586 // connected_moves() tests whether two moves are 'connected' in the sense
1587 // that the first move somehow made the second move possible (for instance
1588 // if the moving piece is the same in both moves). The first move is assumed
1589 // to be the move that was made to reach the current position, while the
1590 // second move is assumed to be a move from the current position.
1592 bool connected_moves(const Position& pos, Move m1, Move m2) {
1594 Square f1, t1, f2, t2;
1597 assert(m1 && move_is_ok(m1));
1598 assert(m2 && move_is_ok(m2));
1600 // Case 1: The moving piece is the same in both moves
1606 // Case 2: The destination square for m2 was vacated by m1
1612 // Case 3: Moving through the vacated square
1613 if ( piece_is_slider(pos.piece_on(f2))
1614 && bit_is_set(squares_between(f2, t2), f1))
1617 // Case 4: The destination square for m2 is defended by the moving piece in m1
1618 p = pos.piece_on(t1);
1619 if (bit_is_set(pos.attacks_from(p, t1), t2))
1622 // Case 5: Discovered check, checking piece is the piece moved in m1
1623 if ( piece_is_slider(p)
1624 && bit_is_set(squares_between(t1, pos.king_square(pos.side_to_move())), f2)
1625 && !bit_is_set(squares_between(t1, pos.king_square(pos.side_to_move())), t2))
1627 // discovered_check_candidates() works also if the Position's side to
1628 // move is the opposite of the checking piece.
1629 Color them = opposite_color(pos.side_to_move());
1630 Bitboard dcCandidates = pos.discovered_check_candidates(them);
1632 if (bit_is_set(dcCandidates, f2))
1639 // value_to_tt() adjusts a mate score from "plies to mate from the root" to
1640 // "plies to mate from the current ply". Non-mate scores are unchanged.
1641 // The function is called before storing a value to the transposition table.
1643 Value value_to_tt(Value v, int ply) {
1645 if (v >= VALUE_MATE_IN_PLY_MAX)
1648 if (v <= VALUE_MATED_IN_PLY_MAX)
1655 // value_from_tt() is the inverse of value_to_tt(): It adjusts a mate score from
1656 // the transposition table to a mate score corrected for the current ply.
1658 Value value_from_tt(Value v, int ply) {
1660 if (v >= VALUE_MATE_IN_PLY_MAX)
1663 if (v <= VALUE_MATED_IN_PLY_MAX)
1670 // extension() decides whether a move should be searched with normal depth,
1671 // or with extended depth. Certain classes of moves (checking moves, in
1672 // particular) are searched with bigger depth than ordinary moves and in
1673 // any case are marked as 'dangerous'. Note that also if a move is not
1674 // extended, as example because the corresponding UCI option is set to zero,
1675 // the move is marked as 'dangerous' so, at least, we avoid to prune it.
1676 template <NodeType PvNode>
1677 Depth extension(const Position& pos, Move m, bool captureOrPromotion,
1678 bool moveIsCheck, bool* dangerous) {
1680 assert(m != MOVE_NONE);
1682 Depth result = DEPTH_ZERO;
1683 *dangerous = moveIsCheck;
1685 if (moveIsCheck && pos.see_sign(m) >= 0)
1686 result += CheckExtension[PvNode];
1688 if (pos.type_of_piece_on(move_from(m)) == PAWN)
1690 Color c = pos.side_to_move();
1691 if (relative_rank(c, move_to(m)) == RANK_7)
1693 result += PawnPushTo7thExtension[PvNode];
1696 if (pos.pawn_is_passed(c, move_to(m)))
1698 result += PassedPawnExtension[PvNode];
1703 if ( captureOrPromotion
1704 && pos.type_of_piece_on(move_to(m)) != PAWN
1705 && ( pos.non_pawn_material(WHITE) + pos.non_pawn_material(BLACK)
1706 - pos.midgame_value_of_piece_on(move_to(m)) == VALUE_ZERO)
1707 && !move_is_special(m))
1709 result += PawnEndgameExtension[PvNode];
1713 return Min(result, ONE_PLY);
1717 // connected_threat() tests whether it is safe to forward prune a move or if
1718 // is somehow connected to the threat move returned by null search.
1720 bool connected_threat(const Position& pos, Move m, Move threat) {
1722 assert(move_is_ok(m));
1723 assert(threat && move_is_ok(threat));
1724 assert(!pos.move_is_check(m));
1725 assert(!pos.move_is_capture_or_promotion(m));
1726 assert(!pos.move_is_passed_pawn_push(m));
1728 Square mfrom, mto, tfrom, tto;
1730 mfrom = move_from(m);
1732 tfrom = move_from(threat);
1733 tto = move_to(threat);
1735 // Case 1: Don't prune moves which move the threatened piece
1739 // Case 2: If the threatened piece has value less than or equal to the
1740 // value of the threatening piece, don't prune moves which defend it.
1741 if ( pos.move_is_capture(threat)
1742 && ( pos.midgame_value_of_piece_on(tfrom) >= pos.midgame_value_of_piece_on(tto)
1743 || pos.type_of_piece_on(tfrom) == KING)
1744 && pos.move_attacks_square(m, tto))
1747 // Case 3: If the moving piece in the threatened move is a slider, don't
1748 // prune safe moves which block its ray.
1749 if ( piece_is_slider(pos.piece_on(tfrom))
1750 && bit_is_set(squares_between(tfrom, tto), mto)
1751 && pos.see_sign(m) >= 0)
1758 // ok_to_use_TT() returns true if a transposition table score
1759 // can be used at a given point in search.
1761 bool ok_to_use_TT(const TTEntry* tte, Depth depth, Value beta, int ply) {
1763 Value v = value_from_tt(tte->value(), ply);
1765 return ( tte->depth() >= depth
1766 || v >= Max(VALUE_MATE_IN_PLY_MAX, beta)
1767 || v < Min(VALUE_MATED_IN_PLY_MAX, beta))
1769 && ( ((tte->type() & VALUE_TYPE_LOWER) && v >= beta)
1770 || ((tte->type() & VALUE_TYPE_UPPER) && v < beta));
1774 // refine_eval() returns the transposition table score if
1775 // possible otherwise falls back on static position evaluation.
1777 Value refine_eval(const TTEntry* tte, Value defaultEval, int ply) {
1781 Value v = value_from_tt(tte->value(), ply);
1783 if ( ((tte->type() & VALUE_TYPE_LOWER) && v >= defaultEval)
1784 || ((tte->type() & VALUE_TYPE_UPPER) && v < defaultEval))
1791 // update_history() registers a good move that produced a beta-cutoff
1792 // in history and marks as failures all the other moves of that ply.
1794 void update_history(const Position& pos, Move move, Depth depth,
1795 Move movesSearched[], int moveCount) {
1797 Value bonus = Value(int(depth) * int(depth));
1799 H.update(pos.piece_on(move_from(move)), move_to(move), bonus);
1801 for (int i = 0; i < moveCount - 1; i++)
1803 m = movesSearched[i];
1807 H.update(pos.piece_on(move_from(m)), move_to(m), -bonus);
1812 // update_gains() updates the gains table of a non-capture move given
1813 // the static position evaluation before and after the move.
1815 void update_gains(const Position& pos, Move m, Value before, Value after) {
1818 && before != VALUE_NONE
1819 && after != VALUE_NONE
1820 && pos.captured_piece_type() == PIECE_TYPE_NONE
1821 && !move_is_special(m))
1822 H.update_gain(pos.piece_on(move_to(m)), move_to(m), -(before + after));
1826 // current_search_time() returns the number of milliseconds which have passed
1827 // since the beginning of the current search.
1829 int current_search_time(int set) {
1831 static int searchStartTime;
1834 searchStartTime = set;
1836 return get_system_time() - searchStartTime;
1840 // value_to_uci() converts a value to a string suitable for use with the UCI
1841 // protocol specifications:
1843 // cp <x> The score from the engine's point of view in centipawns.
1844 // mate <y> Mate in y moves, not plies. If the engine is getting mated
1845 // use negative values for y.
1847 std::string value_to_uci(Value v) {
1849 std::stringstream s;
1851 if (abs(v) < VALUE_MATE - PLY_MAX * ONE_PLY)
1852 s << "cp " << int(v) * 100 / int(PawnValueMidgame); // Scale to centipawns
1854 s << "mate " << (v > 0 ? VALUE_MATE - v + 1 : -VALUE_MATE - v) / 2;
1860 // speed_to_uci() returns a string with time stats of current search suitable
1861 // to be sent to UCI gui.
1863 std::string speed_to_uci(int64_t nodes) {
1865 std::stringstream s;
1866 int t = current_search_time();
1868 s << " nodes " << nodes
1869 << " nps " << (t > 0 ? int(nodes * 1000 / t) : 0)
1876 // poll() performs two different functions: It polls for user input, and it
1877 // looks at the time consumed so far and decides if it's time to abort the
1880 void poll(const Position& pos) {
1882 static int lastInfoTime;
1883 int t = current_search_time();
1886 if (input_available())
1888 // We are line oriented, don't read single chars
1889 std::string command;
1891 if (!std::getline(std::cin, command) || command == "quit")
1893 // Quit the program as soon as possible
1894 Limits.ponder = false;
1895 QuitRequest = StopRequest = true;
1898 else if (command == "stop")
1900 // Stop calculating as soon as possible, but still send the "bestmove"
1901 // and possibly the "ponder" token when finishing the search.
1902 Limits.ponder = false;
1905 else if (command == "ponderhit")
1907 // The opponent has played the expected move. GUI sends "ponderhit" if
1908 // we were told to ponder on the same move the opponent has played. We
1909 // should continue searching but switching from pondering to normal search.
1910 Limits.ponder = false;
1912 if (StopOnPonderhit)
1917 // Print search information
1921 else if (lastInfoTime > t)
1922 // HACK: Must be a new search where we searched less than
1923 // NodesBetweenPolls nodes during the first second of search.
1926 else if (t - lastInfoTime >= 1000)
1931 dbg_print_hit_rate();
1933 // Send info on searched nodes as soon as we return to root
1934 SendSearchedNodes = true;
1937 // Should we stop the search?
1941 bool stillAtFirstMove = FirstRootMove
1942 && !AspirationFailLow
1943 && t > TimeMgr.available_time();
1945 bool noMoreTime = t > TimeMgr.maximum_time()
1946 || stillAtFirstMove;
1948 if ( (Limits.useTimeManagement() && noMoreTime)
1949 || (Limits.maxTime && t >= Limits.maxTime)
1950 || (Limits.maxNodes && pos.nodes_searched() >= Limits.maxNodes)) // FIXME
1955 // wait_for_stop_or_ponderhit() is called when the maximum depth is reached
1956 // while the program is pondering. The point is to work around a wrinkle in
1957 // the UCI protocol: When pondering, the engine is not allowed to give a
1958 // "bestmove" before the GUI sends it a "stop" or "ponderhit" command.
1959 // We simply wait here until one of these commands is sent, and return,
1960 // after which the bestmove and pondermove will be printed.
1962 void wait_for_stop_or_ponderhit() {
1964 std::string command;
1966 // Wait for a command from stdin
1967 while ( std::getline(std::cin, command)
1968 && command != "ponderhit" && command != "stop" && command != "quit") {};
1970 if (command != "ponderhit" && command != "stop")
1971 QuitRequest = true; // Must be "quit" or getline() returned false
1975 // init_thread() is the function which is called when a new thread is
1976 // launched. It simply calls the idle_loop() function with the supplied
1977 // threadID. There are two versions of this function; one for POSIX
1978 // threads and one for Windows threads.
1980 #if !defined(_MSC_VER)
1982 void* init_thread(void* threadID) {
1984 ThreadsMgr.idle_loop(*(int*)threadID, NULL);
1990 DWORD WINAPI init_thread(LPVOID threadID) {
1992 ThreadsMgr.idle_loop(*(int*)threadID, NULL);
1999 /// The ThreadsManager class
2002 // read_uci_options() updates number of active threads and other internal
2003 // parameters according to the UCI options values. It is called before
2004 // to start a new search.
2006 void ThreadsManager::read_uci_options() {
2008 maxThreadsPerSplitPoint = Options["Maximum Number of Threads per Split Point"].value<int>();
2009 minimumSplitDepth = Options["Minimum Split Depth"].value<int>() * ONE_PLY;
2010 useSleepingThreads = Options["Use Sleeping Threads"].value<bool>();
2011 activeThreads = Options["Threads"].value<int>();
2015 // idle_loop() is where the threads are parked when they have no work to do.
2016 // The parameter 'sp', if non-NULL, is a pointer to an active SplitPoint
2017 // object for which the current thread is the master.
2019 void ThreadsManager::idle_loop(int threadID, SplitPoint* sp) {
2021 assert(threadID >= 0 && threadID < MAX_THREADS);
2028 // Slave threads can exit as soon as AllThreadsShouldExit raises,
2029 // master should exit as last one.
2030 if (allThreadsShouldExit)
2033 threads[threadID].state = THREAD_TERMINATED;
2037 // If we are not thinking, wait for a condition to be signaled
2038 // instead of wasting CPU time polling for work.
2039 while ( threadID >= activeThreads
2040 || threads[threadID].state == THREAD_INITIALIZING
2041 || (useSleepingThreads && threads[threadID].state == THREAD_AVAILABLE))
2043 assert(!sp || useSleepingThreads);
2044 assert(threadID != 0 || useSleepingThreads);
2046 if (threads[threadID].state == THREAD_INITIALIZING)
2047 threads[threadID].state = THREAD_AVAILABLE;
2049 // Grab the lock to avoid races with Thread::wake_up()
2050 lock_grab(&threads[threadID].sleepLock);
2052 // If we are master and all slaves have finished do not go to sleep
2053 for (i = 0; sp && i < activeThreads && !sp->slaves[i]; i++) {}
2054 allFinished = (i == activeThreads);
2056 if (allFinished || allThreadsShouldExit)
2058 lock_release(&threads[threadID].sleepLock);
2062 // Do sleep here after retesting sleep conditions
2063 if (threadID >= activeThreads || threads[threadID].state == THREAD_AVAILABLE)
2064 cond_wait(&threads[threadID].sleepCond, &threads[threadID].sleepLock);
2066 lock_release(&threads[threadID].sleepLock);
2069 // If this thread has been assigned work, launch a search
2070 if (threads[threadID].state == THREAD_WORKISWAITING)
2072 assert(!allThreadsShouldExit);
2074 threads[threadID].state = THREAD_SEARCHING;
2076 // Copy split point position and search stack and call search()
2077 // with SplitPoint template parameter set to true.
2078 SearchStack ss[PLY_MAX_PLUS_2];
2079 SplitPoint* tsp = threads[threadID].splitPoint;
2080 Position pos(*tsp->pos, threadID);
2082 memcpy(ss, tsp->ss - 1, 4 * sizeof(SearchStack));
2086 search<PV, true, false>(pos, ss+1, tsp->alpha, tsp->beta, tsp->depth);
2088 search<NonPV, true, false>(pos, ss+1, tsp->alpha, tsp->beta, tsp->depth);
2090 assert(threads[threadID].state == THREAD_SEARCHING);
2092 threads[threadID].state = THREAD_AVAILABLE;
2094 // Wake up master thread so to allow it to return from the idle loop in
2095 // case we are the last slave of the split point.
2096 if ( useSleepingThreads
2097 && threadID != tsp->master
2098 && threads[tsp->master].state == THREAD_AVAILABLE)
2099 threads[tsp->master].wake_up();
2102 // If this thread is the master of a split point and all slaves have
2103 // finished their work at this split point, return from the idle loop.
2104 for (i = 0; sp && i < activeThreads && !sp->slaves[i]; i++) {}
2105 allFinished = (i == activeThreads);
2109 // Because sp->slaves[] is reset under lock protection,
2110 // be sure sp->lock has been released before to return.
2111 lock_grab(&(sp->lock));
2112 lock_release(&(sp->lock));
2114 // In helpful master concept a master can help only a sub-tree, and
2115 // because here is all finished is not possible master is booked.
2116 assert(threads[threadID].state == THREAD_AVAILABLE);
2118 threads[threadID].state = THREAD_SEARCHING;
2125 // init_threads() is called during startup. Initializes locks and condition
2126 // variables and launches all threads sending them immediately to sleep.
2128 void ThreadsManager::init_threads() {
2130 int i, arg[MAX_THREADS];
2133 // This flag is needed to properly end the threads when program exits
2134 allThreadsShouldExit = false;
2136 // Threads will sent to sleep as soon as created, only main thread is kept alive
2141 for (i = 0; i < MAX_THREADS; i++)
2143 // Initialize thread and split point locks
2144 lock_init(&threads[i].sleepLock);
2145 cond_init(&threads[i].sleepCond);
2147 for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
2148 lock_init(&(threads[i].splitPoints[j].lock));
2150 // All threads but first should be set to THREAD_INITIALIZING
2151 threads[i].state = (i == 0 ? THREAD_SEARCHING : THREAD_INITIALIZING);
2154 // Create and startup the threads
2155 for (i = 1; i < MAX_THREADS; i++)
2159 #if !defined(_MSC_VER)
2160 pthread_t pthread[1];
2161 ok = (pthread_create(pthread, NULL, init_thread, (void*)(&arg[i])) == 0);
2162 pthread_detach(pthread[0]);
2164 ok = (CreateThread(NULL, 0, init_thread, (LPVOID)(&arg[i]), 0, NULL) != NULL);
2168 cout << "Failed to create thread number " << i << endl;
2172 // Wait until the thread has finished launching and is gone to sleep
2173 while (threads[i].state == THREAD_INITIALIZING) {}
2178 // exit_threads() is called when the program exits. It makes all the
2179 // helper threads exit cleanly.
2181 void ThreadsManager::exit_threads() {
2183 // Force the woken up threads to exit idle_loop() and hence terminate
2184 allThreadsShouldExit = true;
2186 for (int i = 0; i < MAX_THREADS; i++)
2188 // Wake up all the threads and waits for termination
2191 threads[i].wake_up();
2192 while (threads[i].state != THREAD_TERMINATED) {}
2195 // Now we can safely destroy the locks and wait conditions
2196 lock_destroy(&threads[i].sleepLock);
2197 cond_destroy(&threads[i].sleepCond);
2199 for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++)
2200 lock_destroy(&(threads[i].splitPoints[j].lock));
2203 lock_destroy(&mpLock);
2207 // cutoff_at_splitpoint() checks whether a beta cutoff has occurred in
2208 // the thread's currently active split point, or in some ancestor of
2209 // the current split point.
2211 bool ThreadsManager::cutoff_at_splitpoint(int threadID) const {
2213 assert(threadID >= 0 && threadID < activeThreads);
2215 SplitPoint* sp = threads[threadID].splitPoint;
2217 for ( ; sp && !sp->betaCutoff; sp = sp->parent) {}
2222 // thread_is_available() checks whether the thread with threadID "slave" is
2223 // available to help the thread with threadID "master" at a split point. An
2224 // obvious requirement is that "slave" must be idle. With more than two
2225 // threads, this is not by itself sufficient: If "slave" is the master of
2226 // some active split point, it is only available as a slave to the other
2227 // threads which are busy searching the split point at the top of "slave"'s
2228 // split point stack (the "helpful master concept" in YBWC terminology).
2230 bool ThreadsManager::thread_is_available(int slave, int master) const {
2232 assert(slave >= 0 && slave < activeThreads);
2233 assert(master >= 0 && master < activeThreads);
2234 assert(activeThreads > 1);
2236 if (threads[slave].state != THREAD_AVAILABLE || slave == master)
2239 // Make a local copy to be sure doesn't change under our feet
2240 int localActiveSplitPoints = threads[slave].activeSplitPoints;
2242 // No active split points means that the thread is available as
2243 // a slave for any other thread.
2244 if (localActiveSplitPoints == 0 || activeThreads == 2)
2247 // Apply the "helpful master" concept if possible. Use localActiveSplitPoints
2248 // that is known to be > 0, instead of threads[slave].activeSplitPoints that
2249 // could have been set to 0 by another thread leading to an out of bound access.
2250 if (threads[slave].splitPoints[localActiveSplitPoints - 1].slaves[master])
2257 // available_thread_exists() tries to find an idle thread which is available as
2258 // a slave for the thread with threadID "master".
2260 bool ThreadsManager::available_thread_exists(int master) const {
2262 assert(master >= 0 && master < activeThreads);
2263 assert(activeThreads > 1);
2265 for (int i = 0; i < activeThreads; i++)
2266 if (thread_is_available(i, master))
2273 // split() does the actual work of distributing the work at a node between
2274 // several available threads. If it does not succeed in splitting the
2275 // node (because no idle threads are available, or because we have no unused
2276 // split point objects), the function immediately returns. If splitting is
2277 // possible, a SplitPoint object is initialized with all the data that must be
2278 // copied to the helper threads and we tell our helper threads that they have
2279 // been assigned work. This will cause them to instantly leave their idle loops and
2280 // call search().When all threads have returned from search() then split() returns.
2282 template <bool Fake>
2283 void ThreadsManager::split(Position& pos, SearchStack* ss, Value* alpha, const Value beta,
2284 Value* bestValue, Depth depth, Move threatMove,
2285 int moveCount, MovePicker* mp, bool pvNode) {
2286 assert(pos.is_ok());
2287 assert(*bestValue >= -VALUE_INFINITE);
2288 assert(*bestValue <= *alpha);
2289 assert(*alpha < beta);
2290 assert(beta <= VALUE_INFINITE);
2291 assert(depth > DEPTH_ZERO);
2292 assert(pos.thread() >= 0 && pos.thread() < activeThreads);
2293 assert(activeThreads > 1);
2295 int i, master = pos.thread();
2296 Thread& masterThread = threads[master];
2300 // If no other thread is available to help us, or if we have too many
2301 // active split points, don't split.
2302 if ( !available_thread_exists(master)
2303 || masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS)
2305 lock_release(&mpLock);
2309 // Pick the next available split point object from the split point stack
2310 SplitPoint& splitPoint = masterThread.splitPoints[masterThread.activeSplitPoints++];
2312 // Initialize the split point object
2313 splitPoint.parent = masterThread.splitPoint;
2314 splitPoint.master = master;
2315 splitPoint.betaCutoff = false;
2316 splitPoint.depth = depth;
2317 splitPoint.threatMove = threatMove;
2318 splitPoint.alpha = *alpha;
2319 splitPoint.beta = beta;
2320 splitPoint.pvNode = pvNode;
2321 splitPoint.bestValue = *bestValue;
2323 splitPoint.moveCount = moveCount;
2324 splitPoint.pos = &pos;
2325 splitPoint.nodes = 0;
2327 for (i = 0; i < activeThreads; i++)
2328 splitPoint.slaves[i] = 0;
2330 masterThread.splitPoint = &splitPoint;
2332 // If we are here it means we are not available
2333 assert(masterThread.state != THREAD_AVAILABLE);
2335 int workersCnt = 1; // At least the master is included
2337 // Allocate available threads setting state to THREAD_BOOKED
2338 for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++)
2339 if (thread_is_available(i, master))
2341 threads[i].state = THREAD_BOOKED;
2342 threads[i].splitPoint = &splitPoint;
2343 splitPoint.slaves[i] = 1;
2347 assert(Fake || workersCnt > 1);
2349 // We can release the lock because slave threads are already booked and master is not available
2350 lock_release(&mpLock);
2352 // Tell the threads that they have work to do. This will make them leave
2354 for (i = 0; i < activeThreads; i++)
2355 if (i == master || splitPoint.slaves[i])
2357 assert(i == master || threads[i].state == THREAD_BOOKED);
2359 threads[i].state = THREAD_WORKISWAITING; // This makes the slave to exit from idle_loop()
2361 if (useSleepingThreads && i != master)
2362 threads[i].wake_up();
2365 // Everything is set up. The master thread enters the idle loop, from
2366 // which it will instantly launch a search, because its state is
2367 // THREAD_WORKISWAITING. We send the split point as a second parameter to the
2368 // idle loop, which means that the main thread will return from the idle
2369 // loop when all threads have finished their work at this split point.
2370 idle_loop(master, &splitPoint);
2372 // We have returned from the idle loop, which means that all threads are
2373 // finished. Update alpha and bestValue, and return.
2376 *alpha = splitPoint.alpha;
2377 *bestValue = splitPoint.bestValue;
2378 masterThread.activeSplitPoints--;
2379 masterThread.splitPoint = splitPoint.parent;
2380 pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes);
2382 lock_release(&mpLock);
2386 /// RootMove and RootMoveList method's definitions
2388 RootMove::RootMove() {
2391 pv_score = non_pv_score = -VALUE_INFINITE;
2395 RootMove& RootMove::operator=(const RootMove& rm) {
2397 const Move* src = rm.pv;
2400 // Avoid a costly full rm.pv[] copy
2401 do *dst++ = *src; while (*src++ != MOVE_NONE);
2404 pv_score = rm.pv_score;
2405 non_pv_score = rm.non_pv_score;
2409 // extract_pv_from_tt() builds a PV by adding moves from the transposition table.
2410 // We consider also failing high nodes and not only VALUE_TYPE_EXACT nodes. This
2411 // allow to always have a ponder move even when we fail high at root and also a
2412 // long PV to print that is important for position analysis.
2414 void RootMove::extract_pv_from_tt(Position& pos) {
2416 StateInfo state[PLY_MAX_PLUS_2], *st = state;
2420 assert(pv[0] != MOVE_NONE && pos.move_is_legal(pv[0]));
2422 pos.do_move(pv[0], *st++);
2424 while ( (tte = TT.retrieve(pos.get_key())) != NULL
2425 && tte->move() != MOVE_NONE
2426 && pos.move_is_legal(tte->move())
2428 && (!pos.is_draw() || ply < 2))
2430 pv[ply] = tte->move();
2431 pos.do_move(pv[ply++], *st++);
2433 pv[ply] = MOVE_NONE;
2435 do pos.undo_move(pv[--ply]); while (ply);
2438 // insert_pv_in_tt() is called at the end of a search iteration, and inserts
2439 // the PV back into the TT. This makes sure the old PV moves are searched
2440 // first, even if the old TT entries have been overwritten.
2442 void RootMove::insert_pv_in_tt(Position& pos) {
2444 StateInfo state[PLY_MAX_PLUS_2], *st = state;
2447 Value v, m = VALUE_NONE;
2450 assert(pv[0] != MOVE_NONE && pos.move_is_legal(pv[0]));
2454 tte = TT.retrieve(k);
2456 // Don't overwrite existing correct entries
2457 if (!tte || tte->move() != pv[ply])
2459 v = (pos.is_check() ? VALUE_NONE : evaluate(pos, m));
2460 TT.store(k, VALUE_NONE, VALUE_TYPE_NONE, DEPTH_NONE, pv[ply], v, m);
2462 pos.do_move(pv[ply], *st++);
2464 } while (pv[++ply] != MOVE_NONE);
2466 do pos.undo_move(pv[--ply]); while (ply);
2469 // pv_info_to_uci() returns a string with information on the current PV line
2470 // formatted according to UCI specification.
2472 std::string RootMove::pv_info_to_uci(Position& pos, int depth, int selDepth, Value alpha,
2473 Value beta, int pvIdx) {
2474 std::stringstream s;
2476 s << "info depth " << depth
2477 << " seldepth " << selDepth
2478 << " multipv " << pvIdx + 1
2479 << " score " << value_to_uci(pv_score)
2480 << (pv_score >= beta ? " lowerbound" : pv_score <= alpha ? " upperbound" : "")
2481 << speed_to_uci(pos.nodes_searched())
2484 for (Move* m = pv; *m != MOVE_NONE; m++)
2491 void RootMoveList::init(Position& pos, Move searchMoves[]) {
2493 MoveStack mlist[MAX_MOVES];
2497 bestMoveChanges = 0;
2499 // Generate all legal moves and add them to RootMoveList
2500 MoveStack* last = generate<MV_LEGAL>(pos, mlist);
2501 for (MoveStack* cur = mlist; cur != last; cur++)
2503 // If we have a searchMoves[] list then verify cur->move
2504 // is in the list before to add it.
2505 for (sm = searchMoves; *sm && *sm != cur->move; sm++) {}
2507 if (searchMoves[0] && *sm != cur->move)
2511 rm.pv[0] = cur->move;
2512 rm.pv[1] = MOVE_NONE;
2513 rm.pv_score = -VALUE_INFINITE;
2519 // When playing with strength handicap choose best move among the MultiPV set
2520 // using a statistical rule dependent on SkillLevel. Idea by Heinz van Saanen.
2521 void do_skill_level(Move* best, Move* ponder) {
2523 assert(MultiPV > 1);
2525 // Rml list is already sorted by pv_score in descending order
2527 int max_s = -VALUE_INFINITE;
2528 int size = Min(MultiPV, (int)Rml.size());
2529 int max = Rml[0].pv_score;
2530 int var = Min(max - Rml[size - 1].pv_score, PawnValueMidgame);
2531 int wk = 120 - 2 * SkillLevel;
2533 // PRNG sequence should be non deterministic
2534 for (int i = abs(get_system_time() % 50); i > 0; i--)
2535 RK.rand<unsigned>();
2537 // Choose best move. For each move's score we add two terms both dependent
2538 // on wk, one deterministic and bigger for weaker moves, and one random,
2539 // then we choose the move with the resulting highest score.
2540 for (int i = 0; i < size; i++)
2542 s = Rml[i].pv_score;
2544 // Don't allow crazy blunders even at very low skills
2545 if (i > 0 && Rml[i-1].pv_score > s + EasyMoveMargin)
2548 // This is our magical formula
2549 s += ((max - s) * wk + var * (RK.rand<unsigned>() % wk)) / 128;
2554 *best = Rml[i].pv[0];
2555 *ponder = Rml[i].pv[1];