2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (C) 2004-2008 Tord Romstad (Glaurung author)
4 Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad
5 Copyright (C) 2015-2018 Marco Costalba, Joona Kiiski, Gary Linscott, Tord Romstad
7 Stockfish is free software: you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation, either version 3 of the License, or
10 (at your option) any later version.
12 Stockfish is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with this program. If not, see <http://www.gnu.org/licenses/>.
24 #include <cstring> // For std::memset
38 #include "syzygy/tbprobe.h"
45 namespace Tablebases {
54 namespace TB = Tablebases;
58 using namespace Search;
62 // Different node types, used as a template parameter
63 enum NodeType { NonPV, PV };
65 // Sizes and phases of the skip-blocks, used for distributing search depths across the threads
66 const int skipSize[] = { 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4 };
67 const int skipPhase[] = { 0, 1, 0, 1, 2, 3, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 6, 7 };
69 // Razoring and futility margin based on depth
70 const int razor_margin = 600;
71 Value futility_margin(Depth d) { return Value(150 * d / ONE_PLY); }
73 // Futility and reductions lookup tables, initialized at startup
74 int FutilityMoveCounts[2][16]; // [improving][depth]
75 int Reductions[2][2][64][64]; // [pv][improving][depth][moveNumber]
77 template <bool PvNode> Depth reduction(bool i, Depth d, int mn) {
78 return Reductions[PvNode][i][std::min(d / ONE_PLY, 63)][std::min(mn, 63)] * ONE_PLY;
81 // History and stats update bonus, based on depth
82 int stat_bonus(Depth depth) {
83 int d = depth / ONE_PLY;
84 return d > 17 ? 0 : d * d + 2 * d - 2;
87 // Skill structure is used to implement strength limit
89 explicit Skill(int l) : level(l) {}
90 bool enabled() const { return level < 20; }
91 bool time_to_pick(Depth depth) const { return depth / ONE_PLY == 1 + level; }
92 Move pick_best(size_t multiPV);
95 Move best = MOVE_NONE;
98 template <NodeType NT>
99 Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, bool cutNode, bool skipEarlyPruning);
101 template <NodeType NT, bool InCheck>
102 Value qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth = DEPTH_ZERO);
104 Value value_to_tt(Value v, int ply);
105 Value value_from_tt(Value v, int ply);
106 void update_pv(Move* pv, Move move, Move* childPv);
107 void update_continuation_histories(Stack* ss, Piece pc, Square to, int bonus);
108 void update_stats(const Position& pos, Stack* ss, Move move, Move* quiets, int quietsCnt, int bonus);
109 void update_capture_stats(const Position& pos, Move move, Move* captures, int captureCnt, int bonus);
110 bool pv_is_draw(Position& pos);
112 // perft() is our utility to verify move generation. All the leaf nodes up
113 // to the given depth are generated and counted, and the sum is returned.
115 uint64_t perft(Position& pos, Depth depth) {
118 uint64_t cnt, nodes = 0;
119 const bool leaf = (depth == 2 * ONE_PLY);
121 for (const auto& m : MoveList<LEGAL>(pos))
123 if (Root && depth <= ONE_PLY)
128 cnt = leaf ? MoveList<LEGAL>(pos).size() : perft<false>(pos, depth - ONE_PLY);
133 sync_cout << UCI::move(m, pos.is_chess960()) << ": " << cnt << sync_endl;
141 /// Search::init() is called during startup to initialize various lookup tables
143 void Search::init() {
145 for (int imp = 0; imp <= 1; ++imp)
146 for (int d = 1; d < 64; ++d)
147 for (int mc = 1; mc < 64; ++mc)
149 double r = log(d) * log(mc) / 1.95;
151 Reductions[NonPV][imp][d][mc] = int(std::round(r));
152 Reductions[PV][imp][d][mc] = std::max(Reductions[NonPV][imp][d][mc] - 1, 0);
154 // Increase reduction for non-PV nodes when eval is not improving
155 if (!imp && Reductions[NonPV][imp][d][mc] >= 2)
156 Reductions[NonPV][imp][d][mc]++;
159 for (int d = 0; d < 16; ++d)
161 FutilityMoveCounts[0][d] = int(2.4 + 0.74 * pow(d, 1.78));
162 FutilityMoveCounts[1][d] = int(5.0 + 1.00 * pow(d, 2.00));
167 /// Search::clear() resets search state to its initial value
169 void Search::clear() {
171 Threads.main()->wait_for_search_finished();
173 Time.availableNodes = 0;
179 /// MainThread::search() is called by the main thread when the program receives
180 /// the UCI 'go' command. It searches from the root position and outputs the "bestmove".
182 void MainThread::search() {
186 nodes = perft<true>(rootPos, Limits.perft * ONE_PLY);
187 sync_cout << "\nNodes searched: " << nodes << "\n" << sync_endl;
191 Color us = rootPos.side_to_move();
192 Time.init(Limits, us, rootPos.game_ply());
195 int contempt = Options["Contempt"] * PawnValueEg / 100; // From centipawns
197 Eval::Contempt = (us == WHITE ? make_score(contempt, contempt / 2)
198 : -make_score(contempt, contempt / 2));
200 if (rootMoves.empty())
202 rootMoves.emplace_back(MOVE_NONE);
203 sync_cout << "info depth 0 score "
204 << UCI::value(rootPos.checkers() ? -VALUE_MATE : VALUE_DRAW)
209 for (Thread* th : Threads)
211 th->start_searching();
213 Thread::search(); // Let's start searching!
216 // When we reach the maximum depth, we can arrive here without a raise of
217 // Threads.stop. However, if we are pondering or in an infinite search,
218 // the UCI protocol states that we shouldn't print the best move before the
219 // GUI sends a "stop" or "ponderhit" command. We therefore simply wait here
220 // until the GUI sends one of those commands (which also raises Threads.stop).
221 Threads.stopOnPonderhit = true;
223 while (!Threads.stop && (Threads.ponder || Limits.infinite))
224 {} // Busy wait for a stop or a ponder reset
226 // Stop the threads if not already stopped (also raise the stop if
227 // "ponderhit" just reset Threads.ponder).
230 // Wait until all threads have finished
231 for (Thread* th : Threads)
233 th->wait_for_search_finished();
235 // When playing in 'nodes as time' mode, subtract the searched nodes from
236 // the available ones before exiting.
238 Time.availableNodes += Limits.inc[us] - Threads.nodes_searched();
240 // Check if there are threads with a better score than main thread
241 Thread* bestThread = this;
242 if ( Options["MultiPV"] == 1
244 && !Skill(Options["Skill Level"]).enabled()
245 && rootMoves[0].pv[0] != MOVE_NONE)
247 for (Thread* th : Threads)
249 Depth depthDiff = th->completedDepth - bestThread->completedDepth;
250 Value scoreDiff = th->rootMoves[0].score - bestThread->rootMoves[0].score;
252 // Select the thread with the best score, always if it is a mate
254 && (depthDiff >= 0 || th->rootMoves[0].score >= VALUE_MATE_IN_MAX_PLY))
259 previousScore = bestThread->rootMoves[0].score;
261 // Send new PV when needed
262 if (bestThread != this)
263 sync_cout << UCI::pv(bestThread->rootPos, bestThread->completedDepth, -VALUE_INFINITE, VALUE_INFINITE) << sync_endl;
265 sync_cout << "bestmove " << UCI::move(bestThread->rootMoves[0].pv[0], rootPos.is_chess960());
267 if (bestThread->rootMoves[0].pv.size() > 1 || bestThread->rootMoves[0].extract_ponder_from_tt(rootPos))
268 std::cout << " ponder " << UCI::move(bestThread->rootMoves[0].pv[1], rootPos.is_chess960());
270 std::cout << sync_endl;
274 /// Thread::search() is the main iterative deepening loop. It calls search()
275 /// repeatedly with increasing depth until the allocated thinking time has been
276 /// consumed, the user stops the search, or the maximum search depth is reached.
278 void Thread::search() {
280 Stack stack[MAX_PLY+7], *ss = stack+4; // To reference from (ss-4) to (ss+2)
281 Value bestValue, alpha, beta, delta;
282 Move lastBestMove = MOVE_NONE;
283 Depth lastBestMoveDepth = DEPTH_ZERO;
284 MainThread* mainThread = (this == Threads.main() ? Threads.main() : nullptr);
285 double timeReduction = 1.0;
287 std::memset(ss-4, 0, 7 * sizeof(Stack));
288 for (int i = 4; i > 0; i--)
289 (ss-i)->contHistory = &this->contHistory[NO_PIECE][0]; // Use as sentinel
291 bestValue = delta = alpha = -VALUE_INFINITE;
292 beta = VALUE_INFINITE;
296 mainThread->failedLow = false;
297 mainThread->bestMoveChanges = 0;
300 size_t multiPV = Options["MultiPV"];
301 Skill skill(Options["Skill Level"]);
303 // When playing with strength handicap enable MultiPV search that we will
304 // use behind the scenes to retrieve a set of possible moves.
306 multiPV = std::max(multiPV, (size_t)4);
308 multiPV = std::min(multiPV, rootMoves.size());
310 // Iterative deepening loop until requested to stop or the target depth is reached
311 while ( (rootDepth += ONE_PLY) < DEPTH_MAX
313 && !(Limits.depth && mainThread && rootDepth / ONE_PLY > Limits.depth))
315 // Distribute search depths across the threads
318 int i = (idx - 1) % 20;
319 if (((rootDepth / ONE_PLY + rootPos.game_ply() + skipPhase[i]) / skipSize[i]) % 2)
323 // Age out PV variability metric
325 mainThread->bestMoveChanges *= 0.505, mainThread->failedLow = false;
327 // Save the last iteration's scores before first PV line is searched and
328 // all the move scores except the (new) PV are set to -VALUE_INFINITE.
329 for (RootMove& rm : rootMoves)
330 rm.previousScore = rm.score;
332 // MultiPV loop. We perform a full root search for each PV line
333 for (PVIdx = 0; PVIdx < multiPV && !Threads.stop; ++PVIdx)
335 // Reset UCI info selDepth for each depth and each PV line
338 // Reset aspiration window starting size
339 if (rootDepth >= 5 * ONE_PLY)
342 alpha = std::max(rootMoves[PVIdx].previousScore - delta,-VALUE_INFINITE);
343 beta = std::min(rootMoves[PVIdx].previousScore + delta, VALUE_INFINITE);
346 // Start with a small aspiration window and, in the case of a fail
347 // high/low, re-search with a bigger window until we're not failing
351 bestValue = ::search<PV>(rootPos, ss, alpha, beta, rootDepth, false, false);
353 // Bring the best move to the front. It is critical that sorting
354 // is done with a stable algorithm because all the values but the
355 // first and eventually the new best one are set to -VALUE_INFINITE
356 // and we want to keep the same order for all the moves except the
357 // new PV that goes to the front. Note that in case of MultiPV
358 // search the already searched PV lines are preserved.
359 std::stable_sort(rootMoves.begin() + PVIdx, rootMoves.end());
361 // If search has been stopped, we break immediately. Sorting and
362 // writing PV back to TT is safe because RootMoves is still
363 // valid, although it refers to the previous iteration.
367 // When failing high/low give some update (without cluttering
368 // the UI) before a re-search.
371 && (bestValue <= alpha || bestValue >= beta)
372 && Time.elapsed() > 3000)
373 sync_cout << UCI::pv(rootPos, rootDepth, alpha, beta) << sync_endl;
375 // In case of failing low/high increase aspiration window and
376 // re-search, otherwise exit the loop.
377 if (bestValue <= alpha)
379 beta = (alpha + beta) / 2;
380 alpha = std::max(bestValue - delta, -VALUE_INFINITE);
384 mainThread->failedLow = true;
385 Threads.stopOnPonderhit = false;
388 else if (bestValue >= beta)
389 beta = std::min(bestValue + delta, VALUE_INFINITE);
393 delta += delta / 4 + 5;
395 assert(alpha >= -VALUE_INFINITE && beta <= VALUE_INFINITE);
398 // Sort the PV lines searched so far and update the GUI
399 std::stable_sort(rootMoves.begin(), rootMoves.begin() + PVIdx + 1);
402 && (Threads.stop || PVIdx + 1 == multiPV || Time.elapsed() > 3000))
403 sync_cout << UCI::pv(rootPos, rootDepth, alpha, beta) << sync_endl;
407 completedDepth = rootDepth;
409 if (rootMoves[0].pv[0] != lastBestMove) {
410 lastBestMove = rootMoves[0].pv[0];
411 lastBestMoveDepth = rootDepth;
414 // Have we found a "mate in x"?
416 && bestValue >= VALUE_MATE_IN_MAX_PLY
417 && VALUE_MATE - bestValue <= 2 * Limits.mate)
423 // If skill level is enabled and time is up, pick a sub-optimal best move
424 if (skill.enabled() && skill.time_to_pick(rootDepth))
425 skill.pick_best(multiPV);
427 // Do we have time for the next iteration? Can we stop searching now?
428 if (Limits.use_time_management())
430 if (!Threads.stop && !Threads.stopOnPonderhit)
432 // Stop the search if only one legal move is available, or if all
433 // of the available time has been used
434 const int F[] = { mainThread->failedLow,
435 bestValue - mainThread->previousScore };
436 int improvingFactor = std::max(229, std::min(715, 357 + 119 * F[0] - 6 * F[1]));
438 Color us = rootPos.side_to_move();
439 bool thinkHard = bestValue == VALUE_DRAW
440 && Limits.time[us] - Time.elapsed() > Limits.time[~us]
441 && ::pv_is_draw(rootPos);
443 double unstablePvFactor = 1 + mainThread->bestMoveChanges + thinkHard;
445 // if the bestMove is stable over several iterations, reduce time for this move,
446 // the longer the move has been stable, the more.
447 // Use part of the gained time from a previous stable move for the current move.
449 for (int i : {3, 4, 5})
450 if (lastBestMoveDepth * i < completedDepth && !thinkHard)
451 timeReduction *= 1.3;
452 unstablePvFactor *= std::pow(mainThread->previousTimeReduction, 0.51) / timeReduction;
454 if ( rootMoves.size() == 1
455 || Time.elapsed() > Time.optimum() * unstablePvFactor * improvingFactor / 628)
457 // If we are allowed to ponder do not stop the search now but
458 // keep pondering until the GUI sends "ponderhit" or "stop".
460 Threads.stopOnPonderhit = true;
471 mainThread->previousTimeReduction = timeReduction;
473 // If skill level is enabled, swap best PV line with the sub-optimal one
475 std::swap(rootMoves[0], *std::find(rootMoves.begin(), rootMoves.end(),
476 skill.best ? skill.best : skill.pick_best(multiPV)));
482 // search<>() is the main search function for both PV and non-PV nodes
484 template <NodeType NT>
485 Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, bool cutNode, bool skipEarlyPruning) {
487 const bool PvNode = NT == PV;
488 const bool rootNode = PvNode && ss->ply == 0;
490 assert(-VALUE_INFINITE <= alpha && alpha < beta && beta <= VALUE_INFINITE);
491 assert(PvNode || (alpha == beta - 1));
492 assert(DEPTH_ZERO < depth && depth < DEPTH_MAX);
493 assert(!(PvNode && cutNode));
494 assert(depth / ONE_PLY * ONE_PLY == depth);
496 Move pv[MAX_PLY+1], capturesSearched[32], quietsSearched[64];
500 Move ttMove, move, excludedMove, bestMove;
501 Depth extension, newDepth;
502 Value bestValue, value, ttValue, eval, maxValue;
503 bool ttHit, inCheck, givesCheck, singularExtensionNode, improving;
504 bool captureOrPromotion, doFullDepthSearch, moveCountPruning, skipQuiets, ttCapture, pvExact;
506 int moveCount, captureCount, quietCount;
508 // Step 1. Initialize node
509 Thread* thisThread = pos.this_thread();
510 inCheck = pos.checkers();
511 moveCount = captureCount = quietCount = ss->moveCount = 0;
513 bestValue = -VALUE_INFINITE;
514 maxValue = VALUE_INFINITE;
516 // Check for the available remaining time
517 if (thisThread == Threads.main())
518 static_cast<MainThread*>(thisThread)->check_time();
520 // Used to send selDepth info to GUI (selDepth counts from 1, ply from 0)
521 if (PvNode && thisThread->selDepth < ss->ply + 1)
522 thisThread->selDepth = ss->ply + 1;
526 // Step 2. Check for aborted search and immediate draw
527 if (Threads.stop.load(std::memory_order_relaxed) || pos.is_draw(ss->ply) || ss->ply >= MAX_PLY)
528 return ss->ply >= MAX_PLY && !inCheck ? evaluate(pos) : VALUE_DRAW;
530 // Step 3. Mate distance pruning. Even if we mate at the next move our score
531 // would be at best mate_in(ss->ply+1), but if alpha is already bigger because
532 // a shorter mate was found upward in the tree then there is no need to search
533 // because we will never beat the current alpha. Same logic but with reversed
534 // signs applies also in the opposite condition of being mated instead of giving
535 // mate. In this case return a fail-high score.
536 alpha = std::max(mated_in(ss->ply), alpha);
537 beta = std::min(mate_in(ss->ply+1), beta);
542 assert(0 <= ss->ply && ss->ply < MAX_PLY);
544 (ss+1)->ply = ss->ply + 1;
545 ss->currentMove = (ss+1)->excludedMove = bestMove = MOVE_NONE;
546 ss->contHistory = &thisThread->contHistory[NO_PIECE][0];
547 (ss+2)->killers[0] = (ss+2)->killers[1] = MOVE_NONE;
548 Square prevSq = to_sq((ss-1)->currentMove);
550 // Step 4. Transposition table lookup. We don't want the score of a partial
551 // search to overwrite a previous full search TT value, so we use a different
552 // position key in case of an excluded move.
553 excludedMove = ss->excludedMove;
554 posKey = pos.key() ^ Key(excludedMove << 16); // isn't a very good hash
555 tte = TT.probe(posKey, ttHit);
556 ttValue = ttHit ? value_from_tt(tte->value(), ss->ply) : VALUE_NONE;
557 ttMove = rootNode ? thisThread->rootMoves[thisThread->PVIdx].pv[0]
558 : ttHit ? tte->move() : MOVE_NONE;
560 // At non-PV nodes we check for an early TT cutoff
563 && tte->depth() >= depth
564 && ttValue != VALUE_NONE // Possible in case of TT access race
565 && (ttValue >= beta ? (tte->bound() & BOUND_LOWER)
566 : (tte->bound() & BOUND_UPPER)))
568 // If ttMove is quiet, update move sorting heuristics on TT hit
573 if (!pos.capture_or_promotion(ttMove))
574 update_stats(pos, ss, ttMove, nullptr, 0, stat_bonus(depth));
576 // Extra penalty for a quiet TT move in previous ply when it gets refuted
577 if ((ss-1)->moveCount == 1 && !pos.captured_piece())
578 update_continuation_histories(ss-1, pos.piece_on(prevSq), prevSq, -stat_bonus(depth + ONE_PLY));
580 // Penalty for a quiet ttMove that fails low
581 else if (!pos.capture_or_promotion(ttMove))
583 int penalty = -stat_bonus(depth);
584 thisThread->mainHistory.update(pos.side_to_move(), ttMove, penalty);
585 update_continuation_histories(ss, pos.moved_piece(ttMove), to_sq(ttMove), penalty);
591 // Step 4a. Tablebase probe
592 if (!rootNode && TB::Cardinality)
594 int piecesCount = pos.count<ALL_PIECES>();
596 if ( piecesCount <= TB::Cardinality
597 && (piecesCount < TB::Cardinality || depth >= TB::ProbeDepth)
598 && pos.rule50_count() == 0
599 && !pos.can_castle(ANY_CASTLING))
602 TB::WDLScore wdl = Tablebases::probe_wdl(pos, &err);
604 if (err != TB::ProbeState::FAIL)
606 thisThread->tbHits.fetch_add(1, std::memory_order_relaxed);
608 int drawScore = TB::UseRule50 ? 1 : 0;
610 value = wdl < -drawScore ? -VALUE_MATE + MAX_PLY + ss->ply + 1
611 : wdl > drawScore ? VALUE_MATE - MAX_PLY - ss->ply - 1
612 : VALUE_DRAW + 2 * wdl * drawScore;
614 Bound b = wdl < -drawScore ? BOUND_UPPER
615 : wdl > drawScore ? BOUND_LOWER : BOUND_EXACT;
617 if ( b == BOUND_EXACT
618 || (b == BOUND_LOWER ? value >= beta : value <= alpha))
620 tte->save(posKey, value_to_tt(value, ss->ply), b,
621 std::min(DEPTH_MAX - ONE_PLY, depth + 6 * ONE_PLY),
622 MOVE_NONE, VALUE_NONE, TT.generation());
629 if (b == BOUND_LOWER)
630 bestValue = value, alpha = std::max(alpha, bestValue);
638 // Step 5. Evaluate the position statically
641 ss->staticEval = eval = VALUE_NONE;
647 // Never assume anything on values stored in TT
648 if ((ss->staticEval = eval = tte->eval()) == VALUE_NONE)
649 eval = ss->staticEval = evaluate(pos);
651 // Can ttValue be used as a better position evaluation?
652 if ( ttValue != VALUE_NONE
653 && (tte->bound() & (ttValue > eval ? BOUND_LOWER : BOUND_UPPER)))
658 eval = ss->staticEval =
659 (ss-1)->currentMove != MOVE_NULL ? evaluate(pos)
660 : -(ss-1)->staticEval + 2 * Eval::Tempo;
662 tte->save(posKey, VALUE_NONE, BOUND_NONE, DEPTH_NONE, MOVE_NONE,
663 ss->staticEval, TT.generation());
666 if (skipEarlyPruning || !pos.non_pawn_material(pos.side_to_move()))
669 // Step 6. Razoring (skipped when in check)
671 && depth < 4 * ONE_PLY
672 && eval + razor_margin <= alpha)
674 if (depth <= ONE_PLY)
675 return qsearch<NonPV, false>(pos, ss, alpha, alpha+1);
677 Value ralpha = alpha - razor_margin;
678 Value v = qsearch<NonPV, false>(pos, ss, ralpha, ralpha+1);
683 // Step 7. Futility pruning: child node (skipped when in check)
685 && depth < 7 * ONE_PLY
686 && eval - futility_margin(depth) >= beta
687 && eval < VALUE_KNOWN_WIN) // Do not return unproven wins
690 // Step 8. Null move search with verification search (is omitted in PV nodes)
693 && ss->staticEval >= beta - 36 * depth / ONE_PLY + 225
694 && (ss->ply >= thisThread->nmp_ply || ss->ply % 2 != thisThread->nmp_odd))
697 assert(eval - beta >= 0);
699 // Null move dynamic reduction based on depth and value
700 Depth R = ((823 + 67 * depth / ONE_PLY) / 256 + std::min((eval - beta) / PawnValueMg, 3)) * ONE_PLY;
702 ss->currentMove = MOVE_NULL;
703 ss->contHistory = &thisThread->contHistory[NO_PIECE][0];
705 pos.do_null_move(st);
706 Value nullValue = depth-R < ONE_PLY ? -qsearch<NonPV, false>(pos, ss+1, -beta, -beta+1)
707 : - search<NonPV>(pos, ss+1, -beta, -beta+1, depth-R, !cutNode, true);
708 pos.undo_null_move();
710 if (nullValue >= beta)
712 // Do not return unproven mate scores
713 if (nullValue >= VALUE_MATE_IN_MAX_PLY)
716 if (abs(beta) < VALUE_KNOWN_WIN && (depth < 12 * ONE_PLY || thisThread->nmp_ply))
719 // Do verification search at high depths
720 // disable null move pruning for side to move for the first part of the remaining search tree
721 thisThread->nmp_ply = ss->ply + 3 * (depth-R) / 4;
722 thisThread->nmp_odd = ss->ply % 2;
724 Value v = depth-R < ONE_PLY ? qsearch<NonPV, false>(pos, ss, beta-1, beta)
725 : search<NonPV>(pos, ss, beta-1, beta, depth-R, false, true);
727 thisThread->nmp_odd = thisThread->nmp_ply = 0;
734 // Step 9. ProbCut (skipped when in check)
735 // If we have a good enough capture and a reduced search returns a value
736 // much above beta, we can (almost) safely prune the previous move.
738 && depth >= 5 * ONE_PLY
739 && abs(beta) < VALUE_MATE_IN_MAX_PLY)
741 Value rbeta = std::min(beta + 200, VALUE_INFINITE);
743 assert(is_ok((ss-1)->currentMove));
745 MovePicker mp(pos, ttMove, rbeta - ss->staticEval, &thisThread->captureHistory);
747 while ((move = mp.next_move()) != MOVE_NONE)
750 ss->currentMove = move;
751 ss->contHistory = &thisThread->contHistory[pos.moved_piece(move)][to_sq(move)];
753 assert(depth >= 5 * ONE_PLY);
754 pos.do_move(move, st);
755 value = -search<NonPV>(pos, ss+1, -rbeta, -rbeta+1, depth - 4 * ONE_PLY, !cutNode, false);
762 // Step 10. Internal iterative deepening (skipped when in check)
763 if ( depth >= 6 * ONE_PLY
765 && (PvNode || ss->staticEval + 256 >= beta))
767 Depth d = (3 * depth / (4 * ONE_PLY) - 2) * ONE_PLY;
768 search<NT>(pos, ss, alpha, beta, d, cutNode, true);
770 tte = TT.probe(posKey, ttHit);
771 ttMove = ttHit ? tte->move() : MOVE_NONE;
774 moves_loop: // When in check search starts from here
776 const PieceToHistory* contHist[] = { (ss-1)->contHistory, (ss-2)->contHistory, nullptr, (ss-4)->contHistory };
777 Move countermove = thisThread->counterMoves[pos.piece_on(prevSq)][prevSq];
779 MovePicker mp(pos, ttMove, depth, &thisThread->mainHistory, &thisThread->captureHistory, contHist, countermove, ss->killers);
780 value = bestValue; // Workaround a bogus 'uninitialized' warning under gcc
781 improving = ss->staticEval >= (ss-2)->staticEval
782 /* || ss->staticEval == VALUE_NONE Already implicit in the previous condition */
783 ||(ss-2)->staticEval == VALUE_NONE;
785 singularExtensionNode = !rootNode
786 && depth >= 8 * ONE_PLY
787 && ttMove != MOVE_NONE
788 && ttValue != VALUE_NONE
789 && !excludedMove // Recursive singular search is not allowed
790 && (tte->bound() & BOUND_LOWER)
791 && tte->depth() >= depth - 3 * ONE_PLY;
794 pvExact = PvNode && ttHit && tte->bound() == BOUND_EXACT;
796 // Step 11. Loop through moves
797 // Loop through all pseudo-legal moves until no moves remain or a beta cutoff occurs
798 while ((move = mp.next_move(skipQuiets)) != MOVE_NONE)
802 if (move == excludedMove)
805 // At root obey the "searchmoves" option and skip moves not listed in Root
806 // Move List. As a consequence any illegal move is also skipped. In MultiPV
807 // mode we also skip PV moves which have been already searched.
808 if (rootNode && !std::count(thisThread->rootMoves.begin() + thisThread->PVIdx,
809 thisThread->rootMoves.end(), move))
812 ss->moveCount = ++moveCount;
814 if (rootNode && thisThread == Threads.main() && Time.elapsed() > 3000)
815 sync_cout << "info depth " << depth / ONE_PLY
816 << " currmove " << UCI::move(move, pos.is_chess960())
817 << " currmovenumber " << moveCount + thisThread->PVIdx << sync_endl;
820 (ss+1)->pv = nullptr;
822 extension = DEPTH_ZERO;
823 captureOrPromotion = pos.capture_or_promotion(move);
824 movedPiece = pos.moved_piece(move);
826 givesCheck = type_of(move) == NORMAL && !pos.discovered_check_candidates()
827 ? pos.check_squares(type_of(movedPiece)) & to_sq(move)
828 : pos.gives_check(move);
830 moveCountPruning = depth < 16 * ONE_PLY
831 && moveCount >= FutilityMoveCounts[improving][depth / ONE_PLY];
833 // Step 12. Singular and Gives Check Extensions
835 // Singular extension search. If all moves but one fail low on a search of
836 // (alpha-s, beta-s), and just one fails high on (alpha, beta), then that move
837 // is singular and should be extended. To verify this we do a reduced search
838 // on all the other moves but the ttMove and if the result is lower than
839 // ttValue minus a margin then we will extend the ttMove.
840 if ( singularExtensionNode
844 Value rBeta = std::max(ttValue - 2 * depth / ONE_PLY, -VALUE_MATE);
845 Depth d = (depth / (2 * ONE_PLY)) * ONE_PLY;
846 ss->excludedMove = move;
847 value = search<NonPV>(pos, ss, rBeta - 1, rBeta, d, cutNode, true);
848 ss->excludedMove = MOVE_NONE;
858 // Calculate new depth for this move
859 newDepth = depth - ONE_PLY + extension;
861 // Step 13. Pruning at shallow depth
863 && pos.non_pawn_material(pos.side_to_move())
864 && bestValue > VALUE_MATED_IN_MAX_PLY)
866 if ( !captureOrPromotion
868 && (!pos.advanced_pawn_push(move) || pos.non_pawn_material() >= Value(5000)))
870 // Move count based pruning
871 if (moveCountPruning)
877 // Reduced depth of the next LMR search
878 int lmrDepth = std::max(newDepth - reduction<PvNode>(improving, depth, moveCount), DEPTH_ZERO) / ONE_PLY;
880 // Countermoves based pruning
882 && (*contHist[0])[movedPiece][to_sq(move)] < CounterMovePruneThreshold
883 && (*contHist[1])[movedPiece][to_sq(move)] < CounterMovePruneThreshold)
886 // Futility pruning: parent node
889 && ss->staticEval + 256 + 200 * lmrDepth <= alpha)
892 // Prune moves with negative SEE
894 && !pos.see_ge(move, Value(-35 * lmrDepth * lmrDepth)))
897 else if ( depth < 7 * ONE_PLY
899 && !pos.see_ge(move, -PawnValueEg * (depth / ONE_PLY)))
903 // Speculative prefetch as early as possible
904 prefetch(TT.first_entry(pos.key_after(move)));
906 // Check for legality just before making the move
907 if (!rootNode && !pos.legal(move))
909 ss->moveCount = --moveCount;
913 if (move == ttMove && captureOrPromotion)
916 // Update the current move (this must be done after singular extension search)
917 ss->currentMove = move;
918 ss->contHistory = &thisThread->contHistory[movedPiece][to_sq(move)];
920 // Step 14. Make the move
921 pos.do_move(move, st, givesCheck);
923 // Step 15. Reduced depth search (LMR). If the move fails high it will be
924 // re-searched at full depth.
925 if ( depth >= 3 * ONE_PLY
927 && (!captureOrPromotion || moveCountPruning))
929 Depth r = reduction<PvNode>(improving, depth, moveCount);
931 if (captureOrPromotion)
932 r -= r ? ONE_PLY : DEPTH_ZERO;
935 // Decrease reduction if opponent's move count is high
936 if ((ss-1)->moveCount > 15)
939 // Decrease reduction for exact PV nodes
943 // Increase reduction if ttMove is a capture
947 // Increase reduction for cut nodes
951 // Decrease reduction for moves that escape a capture. Filter out
952 // castling moves, because they are coded as "king captures rook" and
953 // hence break make_move().
954 else if ( type_of(move) == NORMAL
955 && !pos.see_ge(make_move(to_sq(move), from_sq(move))))
958 ss->statScore = thisThread->mainHistory[~pos.side_to_move()][from_to(move)]
959 + (*contHist[0])[movedPiece][to_sq(move)]
960 + (*contHist[1])[movedPiece][to_sq(move)]
961 + (*contHist[3])[movedPiece][to_sq(move)]
964 // Decrease/increase reduction by comparing opponent's stat score
965 if (ss->statScore >= 0 && (ss-1)->statScore < 0)
968 else if ((ss-1)->statScore >= 0 && ss->statScore < 0)
971 // Decrease/increase reduction for moves with a good/bad history
972 r = std::max(DEPTH_ZERO, (r / ONE_PLY - ss->statScore / 20000) * ONE_PLY);
975 Depth d = std::max(newDepth - r, ONE_PLY);
977 value = -search<NonPV>(pos, ss+1, -(alpha+1), -alpha, d, true, false);
979 doFullDepthSearch = (value > alpha && d != newDepth);
982 doFullDepthSearch = !PvNode || moveCount > 1;
984 // Step 16. Full depth search when LMR is skipped or fails high
985 if (doFullDepthSearch)
986 value = newDepth < ONE_PLY ?
987 givesCheck ? -qsearch<NonPV, true>(pos, ss+1, -(alpha+1), -alpha)
988 : -qsearch<NonPV, false>(pos, ss+1, -(alpha+1), -alpha)
989 : - search<NonPV>(pos, ss+1, -(alpha+1), -alpha, newDepth, !cutNode, false);
991 // For PV nodes only, do a full PV search on the first move or after a fail
992 // high (in the latter case search only if value < beta), otherwise let the
993 // parent node fail low with value <= alpha and try another move.
994 if (PvNode && (moveCount == 1 || (value > alpha && (rootNode || value < beta))))
997 (ss+1)->pv[0] = MOVE_NONE;
999 value = newDepth < ONE_PLY ?
1000 givesCheck ? -qsearch<PV, true>(pos, ss+1, -beta, -alpha)
1001 : -qsearch<PV, false>(pos, ss+1, -beta, -alpha)
1002 : - search<PV>(pos, ss+1, -beta, -alpha, newDepth, false, false);
1005 // Step 17. Undo move
1006 pos.undo_move(move);
1008 assert(value > -VALUE_INFINITE && value < VALUE_INFINITE);
1010 // Step 18. Check for a new best move
1011 // Finished searching the move. If a stop occurred, the return value of
1012 // the search cannot be trusted, and we return immediately without
1013 // updating best move, PV and TT.
1014 if (Threads.stop.load(std::memory_order_relaxed))
1019 RootMove& rm = *std::find(thisThread->rootMoves.begin(),
1020 thisThread->rootMoves.end(), move);
1022 // PV move or new best move ?
1023 if (moveCount == 1 || value > alpha)
1026 rm.selDepth = thisThread->selDepth;
1031 for (Move* m = (ss+1)->pv; *m != MOVE_NONE; ++m)
1032 rm.pv.push_back(*m);
1034 // We record how often the best move has been changed in each
1035 // iteration. This information is used for time management: When
1036 // the best move changes frequently, we allocate some more time.
1037 if (moveCount > 1 && thisThread == Threads.main())
1038 ++static_cast<MainThread*>(thisThread)->bestMoveChanges;
1041 // All other moves but the PV are set to the lowest value: this
1042 // is not a problem when sorting because the sort is stable and the
1043 // move position in the list is preserved - just the PV is pushed up.
1044 rm.score = -VALUE_INFINITE;
1047 if (value > bestValue)
1055 if (PvNode && !rootNode) // Update pv even in fail-high case
1056 update_pv(ss->pv, move, (ss+1)->pv);
1058 if (PvNode && value < beta) // Update alpha! Always alpha < beta
1062 assert(value >= beta); // Fail high
1068 if (!captureOrPromotion && move != bestMove && quietCount < 64)
1069 quietsSearched[quietCount++] = move;
1070 else if (captureOrPromotion && move != bestMove && captureCount < 32)
1071 capturesSearched[captureCount++] = move;
1074 // The following condition would detect a stop only after move loop has been
1075 // completed. But in this case bestValue is valid because we have fully
1076 // searched our subtree, and we can anyhow save the result in TT.
1082 // Step 20. Check for mate and stalemate
1083 // All legal moves have been searched and if there are no legal moves, it
1084 // must be a mate or a stalemate. If we are in a singular extension search then
1085 // return a fail low score.
1087 assert(moveCount || !inCheck || excludedMove || !MoveList<LEGAL>(pos).size());
1090 bestValue = excludedMove ? alpha
1091 : inCheck ? mated_in(ss->ply) : VALUE_DRAW;
1094 // Quiet best move: update move sorting heuristics
1095 if (!pos.capture_or_promotion(bestMove))
1096 update_stats(pos, ss, bestMove, quietsSearched, quietCount, stat_bonus(depth));
1098 update_capture_stats(pos, bestMove, capturesSearched, captureCount, stat_bonus(depth));
1100 // Extra penalty for a quiet TT move in previous ply when it gets refuted
1101 if ((ss-1)->moveCount == 1 && !pos.captured_piece())
1102 update_continuation_histories(ss-1, pos.piece_on(prevSq), prevSq, -stat_bonus(depth + ONE_PLY));
1104 // Bonus for prior countermove that caused the fail low
1105 else if ( depth >= 3 * ONE_PLY
1106 && !pos.captured_piece()
1107 && is_ok((ss-1)->currentMove))
1108 update_continuation_histories(ss-1, pos.piece_on(prevSq), prevSq, stat_bonus(depth));
1111 bestValue = std::min(bestValue, maxValue);
1114 tte->save(posKey, value_to_tt(bestValue, ss->ply),
1115 bestValue >= beta ? BOUND_LOWER :
1116 PvNode && bestMove ? BOUND_EXACT : BOUND_UPPER,
1117 depth, bestMove, ss->staticEval, TT.generation());
1119 assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE);
1125 // qsearch() is the quiescence search function, which is called by the main
1126 // search function with depth zero, or recursively with depth less than ONE_PLY.
1128 template <NodeType NT, bool InCheck>
1129 Value qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth) {
1131 const bool PvNode = NT == PV;
1133 assert(InCheck == bool(pos.checkers()));
1134 assert(alpha >= -VALUE_INFINITE && alpha < beta && beta <= VALUE_INFINITE);
1135 assert(PvNode || (alpha == beta - 1));
1136 assert(depth <= DEPTH_ZERO);
1137 assert(depth / ONE_PLY * ONE_PLY == depth);
1143 Move ttMove, move, bestMove;
1144 Value bestValue, value, ttValue, futilityValue, futilityBase, oldAlpha;
1145 bool ttHit, givesCheck, evasionPrunable;
1151 oldAlpha = alpha; // To flag BOUND_EXACT when eval above alpha and no available moves
1153 ss->pv[0] = MOVE_NONE;
1156 ss->currentMove = bestMove = MOVE_NONE;
1157 (ss+1)->ply = ss->ply + 1;
1160 // Check for an instant draw or if the maximum ply has been reached
1161 if (pos.is_draw(ss->ply) || ss->ply >= MAX_PLY)
1162 return ss->ply >= MAX_PLY && !InCheck ? evaluate(pos) : VALUE_DRAW;
1164 assert(0 <= ss->ply && ss->ply < MAX_PLY);
1166 // Decide whether or not to include checks: this fixes also the type of
1167 // TT entry depth that we are going to use. Note that in qsearch we use
1168 // only two types of depth in TT: DEPTH_QS_CHECKS or DEPTH_QS_NO_CHECKS.
1169 ttDepth = InCheck || depth >= DEPTH_QS_CHECKS ? DEPTH_QS_CHECKS
1170 : DEPTH_QS_NO_CHECKS;
1171 // Transposition table lookup
1173 tte = TT.probe(posKey, ttHit);
1174 ttMove = ttHit ? tte->move() : MOVE_NONE;
1175 ttValue = ttHit ? value_from_tt(tte->value(), ss->ply) : VALUE_NONE;
1179 && tte->depth() >= ttDepth
1180 && ttValue != VALUE_NONE // Only in case of TT access race
1181 && (ttValue >= beta ? (tte->bound() & BOUND_LOWER)
1182 : (tte->bound() & BOUND_UPPER)))
1185 // Evaluate the position statically
1188 ss->staticEval = VALUE_NONE;
1189 bestValue = futilityBase = -VALUE_INFINITE;
1195 // Never assume anything on values stored in TT
1196 if ((ss->staticEval = bestValue = tte->eval()) == VALUE_NONE)
1197 ss->staticEval = bestValue = evaluate(pos);
1199 // Can ttValue be used as a better position evaluation?
1200 if ( ttValue != VALUE_NONE
1201 && (tte->bound() & (ttValue > bestValue ? BOUND_LOWER : BOUND_UPPER)))
1202 bestValue = ttValue;
1205 ss->staticEval = bestValue =
1206 (ss-1)->currentMove != MOVE_NULL ? evaluate(pos)
1207 : -(ss-1)->staticEval + 2 * Eval::Tempo;
1209 // Stand pat. Return immediately if static value is at least beta
1210 if (bestValue >= beta)
1213 tte->save(posKey, value_to_tt(bestValue, ss->ply), BOUND_LOWER,
1214 DEPTH_NONE, MOVE_NONE, ss->staticEval, TT.generation());
1219 if (PvNode && bestValue > alpha)
1222 futilityBase = bestValue + 128;
1225 // Initialize a MovePicker object for the current position, and prepare
1226 // to search the moves. Because the depth is <= 0 here, only captures,
1227 // queen promotions and checks (only if depth >= DEPTH_QS_CHECKS) will
1229 MovePicker mp(pos, ttMove, depth, &pos.this_thread()->mainHistory, &pos.this_thread()->captureHistory, to_sq((ss-1)->currentMove));
1231 // Loop through the moves until no moves remain or a beta cutoff occurs
1232 while ((move = mp.next_move()) != MOVE_NONE)
1234 assert(is_ok(move));
1236 givesCheck = type_of(move) == NORMAL && !pos.discovered_check_candidates()
1237 ? pos.check_squares(type_of(pos.moved_piece(move))) & to_sq(move)
1238 : pos.gives_check(move);
1245 && futilityBase > -VALUE_KNOWN_WIN
1246 && !pos.advanced_pawn_push(move))
1248 assert(type_of(move) != ENPASSANT); // Due to !pos.advanced_pawn_push
1250 futilityValue = futilityBase + PieceValue[EG][pos.piece_on(to_sq(move))];
1252 if (futilityValue <= alpha)
1254 bestValue = std::max(bestValue, futilityValue);
1258 if (futilityBase <= alpha && !pos.see_ge(move, VALUE_ZERO + 1))
1260 bestValue = std::max(bestValue, futilityBase);
1265 // Detect non-capture evasions that are candidates to be pruned
1266 evasionPrunable = InCheck
1267 && (depth != DEPTH_ZERO || moveCount > 2)
1268 && bestValue > VALUE_MATED_IN_MAX_PLY
1269 && !pos.capture(move);
1271 // Don't search moves with negative SEE values
1272 if ( (!InCheck || evasionPrunable)
1273 && !pos.see_ge(move))
1276 // Speculative prefetch as early as possible
1277 prefetch(TT.first_entry(pos.key_after(move)));
1279 // Check for legality just before making the move
1280 if (!pos.legal(move))
1286 ss->currentMove = move;
1288 // Make and search the move
1289 pos.do_move(move, st, givesCheck);
1290 value = givesCheck ? -qsearch<NT, true>(pos, ss+1, -beta, -alpha, depth - ONE_PLY)
1291 : -qsearch<NT, false>(pos, ss+1, -beta, -alpha, depth - ONE_PLY);
1292 pos.undo_move(move);
1294 assert(value > -VALUE_INFINITE && value < VALUE_INFINITE);
1296 // Check for a new best move
1297 if (value > bestValue)
1303 if (PvNode) // Update pv even in fail-high case
1304 update_pv(ss->pv, move, (ss+1)->pv);
1306 if (PvNode && value < beta) // Update alpha here!
1313 tte->save(posKey, value_to_tt(value, ss->ply), BOUND_LOWER,
1314 ttDepth, move, ss->staticEval, TT.generation());
1322 // All legal moves have been searched. A special case: If we're in check
1323 // and no legal moves were found, it is checkmate.
1324 if (InCheck && bestValue == -VALUE_INFINITE)
1325 return mated_in(ss->ply); // Plies to mate from the root
1327 tte->save(posKey, value_to_tt(bestValue, ss->ply),
1328 PvNode && bestValue > oldAlpha ? BOUND_EXACT : BOUND_UPPER,
1329 ttDepth, bestMove, ss->staticEval, TT.generation());
1331 assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE);
1337 // value_to_tt() adjusts a mate score from "plies to mate from the root" to
1338 // "plies to mate from the current position". Non-mate scores are unchanged.
1339 // The function is called before storing a value in the transposition table.
1341 Value value_to_tt(Value v, int ply) {
1343 assert(v != VALUE_NONE);
1345 return v >= VALUE_MATE_IN_MAX_PLY ? v + ply
1346 : v <= VALUE_MATED_IN_MAX_PLY ? v - ply : v;
1350 // value_from_tt() is the inverse of value_to_tt(): It adjusts a mate score
1351 // from the transposition table (which refers to the plies to mate/be mated
1352 // from current position) to "plies to mate/be mated from the root".
1354 Value value_from_tt(Value v, int ply) {
1356 return v == VALUE_NONE ? VALUE_NONE
1357 : v >= VALUE_MATE_IN_MAX_PLY ? v - ply
1358 : v <= VALUE_MATED_IN_MAX_PLY ? v + ply : v;
1362 // update_pv() adds current move and appends child pv[]
1364 void update_pv(Move* pv, Move move, Move* childPv) {
1366 for (*pv++ = move; childPv && *childPv != MOVE_NONE; )
1372 // update_continuation_histories() updates histories of the move pairs formed
1373 // by moves at ply -1, -2, and -4 with current move.
1375 void update_continuation_histories(Stack* ss, Piece pc, Square to, int bonus) {
1377 for (int i : {1, 2, 4})
1378 if (is_ok((ss-i)->currentMove))
1379 (ss-i)->contHistory->update(pc, to, bonus);
1383 // update_capture_stats() updates move sorting heuristics when a new capture best move is found
1385 void update_capture_stats(const Position& pos, Move move,
1386 Move* captures, int captureCnt, int bonus) {
1388 CapturePieceToHistory& captureHistory = pos.this_thread()->captureHistory;
1389 Piece moved_piece = pos.moved_piece(move);
1390 PieceType captured = type_of(pos.piece_on(to_sq(move)));
1391 captureHistory.update(moved_piece, to_sq(move), captured, bonus);
1393 // Decrease all the other played capture moves
1394 for (int i = 0; i < captureCnt; ++i)
1396 moved_piece = pos.moved_piece(captures[i]);
1397 captured = type_of(pos.piece_on(to_sq(captures[i])));
1398 captureHistory.update(moved_piece, to_sq(captures[i]), captured, -bonus);
1403 // update_stats() updates move sorting heuristics when a new quiet best move is found
1405 void update_stats(const Position& pos, Stack* ss, Move move,
1406 Move* quiets, int quietsCnt, int bonus) {
1408 if (ss->killers[0] != move)
1410 ss->killers[1] = ss->killers[0];
1411 ss->killers[0] = move;
1414 Color c = pos.side_to_move();
1415 Thread* thisThread = pos.this_thread();
1416 thisThread->mainHistory.update(c, move, bonus);
1417 update_continuation_histories(ss, pos.moved_piece(move), to_sq(move), bonus);
1419 if (is_ok((ss-1)->currentMove))
1421 Square prevSq = to_sq((ss-1)->currentMove);
1422 thisThread->counterMoves[pos.piece_on(prevSq)][prevSq] = move;
1425 // Decrease all the other played quiet moves
1426 for (int i = 0; i < quietsCnt; ++i)
1428 thisThread->mainHistory.update(c, quiets[i], -bonus);
1429 update_continuation_histories(ss, pos.moved_piece(quiets[i]), to_sq(quiets[i]), -bonus);
1434 // Is the PV leading to a draw position? Assumes all pv moves are legal
1435 bool pv_is_draw(Position& pos) {
1437 StateInfo st[MAX_PLY];
1438 auto& pv = pos.this_thread()->rootMoves[0].pv;
1440 for (size_t i = 0; i < pv.size(); ++i)
1441 pos.do_move(pv[i], st[i]);
1443 bool isDraw = pos.is_draw(pv.size());
1445 for (size_t i = pv.size(); i > 0; --i)
1446 pos.undo_move(pv[i-1]);
1452 // When playing with strength handicap, choose best move among a set of RootMoves
1453 // using a statistical rule dependent on 'level'. Idea by Heinz van Saanen.
1455 Move Skill::pick_best(size_t multiPV) {
1457 const RootMoves& rootMoves = Threads.main()->rootMoves;
1458 static PRNG rng(now()); // PRNG sequence should be non-deterministic
1460 // RootMoves are already sorted by score in descending order
1461 Value topScore = rootMoves[0].score;
1462 int delta = std::min(topScore - rootMoves[multiPV - 1].score, PawnValueMg);
1463 int weakness = 120 - 2 * level;
1464 int maxScore = -VALUE_INFINITE;
1466 // Choose best move. For each move score we add two terms, both dependent on
1467 // weakness. One is deterministic and bigger for weaker levels, and one is
1468 // random. Then we choose the move with the resulting highest score.
1469 for (size_t i = 0; i < multiPV; ++i)
1471 // This is our magic formula
1472 int push = ( weakness * int(topScore - rootMoves[i].score)
1473 + delta * (rng.rand<unsigned>() % weakness)) / 128;
1475 if (rootMoves[i].score + push >= maxScore)
1477 maxScore = rootMoves[i].score + push;
1478 best = rootMoves[i].pv[0];
1487 // check_time() is used to print debug info and, more importantly, to detect
1488 // when we are out of available time and thus stop the search.
1490 void MainThread::check_time() {
1495 // At low node count increase the checking rate to about 0.1% of nodes
1496 // otherwise use a default value.
1497 callsCnt = Limits.nodes ? std::min(4096, int(Limits.nodes / 1024)) : 4096;
1499 static TimePoint lastInfoTime = now();
1501 int elapsed = Time.elapsed();
1502 TimePoint tick = Limits.startTime + elapsed;
1504 if (tick - lastInfoTime >= 1000)
1506 lastInfoTime = tick;
1510 // An engine may not stop pondering until told so by the GUI
1514 if ( (Limits.use_time_management() && elapsed > Time.maximum() - 10)
1515 || (Limits.movetime && elapsed >= Limits.movetime)
1516 || (Limits.nodes && Threads.nodes_searched() >= (uint64_t)Limits.nodes))
1517 Threads.stop = true;
1521 /// UCI::pv() formats PV information according to the UCI protocol. UCI requires
1522 /// that all (if any) unsearched PV lines are sent using a previous search score.
1524 string UCI::pv(const Position& pos, Depth depth, Value alpha, Value beta) {
1526 std::stringstream ss;
1527 int elapsed = Time.elapsed() + 1;
1528 const RootMoves& rootMoves = pos.this_thread()->rootMoves;
1529 size_t PVIdx = pos.this_thread()->PVIdx;
1530 size_t multiPV = std::min((size_t)Options["MultiPV"], rootMoves.size());
1531 uint64_t nodesSearched = Threads.nodes_searched();
1532 uint64_t tbHits = Threads.tb_hits() + (TB::RootInTB ? rootMoves.size() : 0);
1534 for (size_t i = 0; i < multiPV; ++i)
1536 bool updated = (i <= PVIdx && rootMoves[i].score != -VALUE_INFINITE);
1538 if (depth == ONE_PLY && !updated)
1541 Depth d = updated ? depth : depth - ONE_PLY;
1542 Value v = updated ? rootMoves[i].score : rootMoves[i].previousScore;
1544 bool tb = TB::RootInTB && abs(v) < VALUE_MATE - MAX_PLY;
1545 v = tb ? TB::Score : v;
1547 if (ss.rdbuf()->in_avail()) // Not at first line
1551 << " depth " << d / ONE_PLY
1552 << " seldepth " << rootMoves[i].selDepth
1553 << " multipv " << i + 1
1554 << " score " << UCI::value(v);
1556 if (!tb && i == PVIdx)
1557 ss << (v >= beta ? " lowerbound" : v <= alpha ? " upperbound" : "");
1559 ss << " nodes " << nodesSearched
1560 << " nps " << nodesSearched * 1000 / elapsed;
1562 if (elapsed > 1000) // Earlier makes little sense
1563 ss << " hashfull " << TT.hashfull();
1565 ss << " tbhits " << tbHits
1566 << " time " << elapsed
1569 for (Move m : rootMoves[i].pv)
1570 ss << " " << UCI::move(m, pos.is_chess960());
1577 /// RootMove::extract_ponder_from_tt() is called in case we have no ponder move
1578 /// before exiting the search, for instance, in case we stop the search during a
1579 /// fail high at root. We try hard to have a ponder move to return to the GUI,
1580 /// otherwise in case of 'ponder on' we have nothing to think on.
1582 bool RootMove::extract_ponder_from_tt(Position& pos) {
1587 assert(pv.size() == 1);
1592 pos.do_move(pv[0], st);
1593 TTEntry* tte = TT.probe(pos.key(), ttHit);
1597 Move m = tte->move(); // Local copy to be SMP safe
1598 if (MoveList<LEGAL>(pos).contains(m))
1602 pos.undo_move(pv[0]);
1603 return pv.size() > 1;
1606 void Tablebases::filter_root_moves(Position& pos, Search::RootMoves& rootMoves) {
1609 UseRule50 = Options["Syzygy50MoveRule"];
1610 ProbeDepth = Options["SyzygyProbeDepth"] * ONE_PLY;
1611 Cardinality = Options["SyzygyProbeLimit"];
1613 // Skip TB probing when no TB found: !TBLargest -> !TB::Cardinality
1614 if (Cardinality > MaxCardinality)
1616 Cardinality = MaxCardinality;
1617 ProbeDepth = DEPTH_ZERO;
1620 if (Cardinality < popcount(pos.pieces()) || pos.can_castle(ANY_CASTLING))
1623 // Don't filter any moves if the user requested analysis on multiple
1624 if (Options["MultiPV"] != 1)
1627 // If the current root position is in the tablebases, then RootMoves
1628 // contains only moves that preserve the draw or the win.
1629 RootInTB = root_probe(pos, rootMoves, TB::Score);
1632 Cardinality = 0; // Do not probe tablebases during the search
1634 else // If DTZ tables are missing, use WDL tables as a fallback
1636 // Filter out moves that do not preserve the draw or the win.
1637 RootInTB = root_probe_wdl(pos, rootMoves, TB::Score);
1639 // Only probe during search if winning
1640 if (RootInTB && TB::Score <= VALUE_DRAW)
1644 if (RootInTB && !UseRule50)
1645 TB::Score = TB::Score > VALUE_DRAW ? VALUE_MATE - MAX_PLY - 1
1646 : TB::Score < VALUE_DRAW ? -VALUE_MATE + MAX_PLY + 1
1649 // Since root_probe() and root_probe_wdl() dirty the root move scores,
1650 // we reset them to -VALUE_INFINITE
1651 for (RootMove& rm : rootMoves)
1652 rm.score = -VALUE_INFINITE;