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) 2015-2019 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 {
53 namespace TB = Tablebases;
57 using namespace Search;
61 // Different node types, used as a template parameter
62 enum NodeType { NonPV, PV };
64 // Sizes and phases of the skip-blocks, used for distributing search depths across the threads
65 constexpr int SkipSize[] = { 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4 };
66 constexpr int SkipPhase[] = { 0, 1, 0, 1, 2, 3, 0, 1, 2, 3, 4, 5, 0, 1, 2, 3, 4, 5, 6, 7 };
68 // Razor and futility margins
69 constexpr int RazorMargin = 600;
70 Value futility_margin(Depth d, bool improving) {
71 return Value((175 - 50 * improving) * d / ONE_PLY);
74 // Futility and reductions lookup tables, initialized at startup
75 int FutilityMoveCounts[2][16]; // [improving][depth]
76 int Reductions[2][2][64][64]; // [pv][improving][depth][moveNumber]
78 template <bool PvNode> Depth reduction(bool i, Depth d, int mn) {
79 return Reductions[PvNode][i][std::min(d / ONE_PLY, 63)][std::min(mn, 63)] * ONE_PLY;
82 // History and stats update bonus, based on depth
83 int stat_bonus(Depth depth) {
84 int d = depth / ONE_PLY;
85 return d > 17 ? 0 : 29 * d * d + 138 * d - 134;
88 // Add a small random component to draw evaluations to keep search dynamic
89 // and to avoid 3fold-blindness.
90 Value value_draw(Depth depth, Thread* thisThread) {
91 return depth < 4 ? VALUE_DRAW
92 : VALUE_DRAW + Value(2 * (thisThread->nodes.load(std::memory_order_relaxed) % 2) - 1);
95 // Skill structure is used to implement strength limit
97 explicit Skill(int l) : level(l) {}
98 bool enabled() const { return level < 20; }
99 bool time_to_pick(Depth depth) const { return depth / ONE_PLY == 1 + level; }
100 Move pick_best(size_t multiPV);
103 Move best = MOVE_NONE;
106 template <NodeType NT>
107 Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, bool cutNode);
109 template <NodeType NT>
110 Value qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth = DEPTH_ZERO);
112 Value value_to_tt(Value v, int ply);
113 Value value_from_tt(Value v, int ply);
114 void update_pv(Move* pv, Move move, Move* childPv);
115 void update_continuation_histories(Stack* ss, Piece pc, Square to, int bonus);
116 void update_quiet_stats(const Position& pos, Stack* ss, Move move, Move* quiets, int quietsCnt, int bonus);
117 void update_capture_stats(const Position& pos, Move move, Move* captures, int captureCnt, int bonus);
119 inline bool gives_check(const Position& pos, Move move) {
120 Color us = pos.side_to_move();
121 return type_of(move) == NORMAL && !(pos.blockers_for_king(~us) & pos.pieces(us))
122 ? pos.check_squares(type_of(pos.moved_piece(move))) & to_sq(move)
123 : pos.gives_check(move);
126 // perft() is our utility to verify move generation. All the leaf nodes up
127 // to the given depth are generated and counted, and the sum is returned.
129 uint64_t perft(Position& pos, Depth depth) {
132 uint64_t cnt, nodes = 0;
133 const bool leaf = (depth == 2 * ONE_PLY);
135 for (const auto& m : MoveList<LEGAL>(pos))
137 if (Root && depth <= ONE_PLY)
142 cnt = leaf ? MoveList<LEGAL>(pos).size() : perft<false>(pos, depth - ONE_PLY);
147 sync_cout << UCI::move(m, pos.is_chess960()) << ": " << cnt << sync_endl;
155 /// Search::init() is called at startup to initialize various lookup tables
157 void Search::init() {
159 for (int imp = 0; imp <= 1; ++imp)
160 for (int d = 1; d < 64; ++d)
161 for (int mc = 1; mc < 64; ++mc)
163 double r = log(d) * log(mc) / 1.95;
165 Reductions[NonPV][imp][d][mc] = int(std::round(r));
166 Reductions[PV][imp][d][mc] = std::max(Reductions[NonPV][imp][d][mc] - 1, 0);
168 // Increase reduction for non-PV nodes when eval is not improving
170 Reductions[NonPV][imp][d][mc]++;
173 for (int d = 0; d < 16; ++d)
175 FutilityMoveCounts[0][d] = int(2.4 + 0.74 * pow(d, 1.78));
176 FutilityMoveCounts[1][d] = int(5.0 + 1.00 * pow(d, 2.00));
181 /// Search::clear() resets search state to its initial value
183 void Search::clear() {
185 Threads.main()->wait_for_search_finished();
187 Time.availableNodes = 0;
190 Tablebases::init(Options["SyzygyPath"]); // Free up mapped files
194 /// MainThread::search() is called by the main thread when the program receives
195 /// the UCI 'go' command. It searches from the root position and outputs the "bestmove".
197 void MainThread::search() {
201 nodes = perft<true>(rootPos, Limits.perft * ONE_PLY);
202 sync_cout << "\nNodes searched: " << nodes << "\n" << sync_endl;
206 Color us = rootPos.side_to_move();
207 Time.init(Limits, us, rootPos.game_ply());
210 if (rootMoves.empty())
212 rootMoves.emplace_back(MOVE_NONE);
213 sync_cout << "info depth 0 score "
214 << UCI::value(rootPos.checkers() ? -VALUE_MATE : VALUE_DRAW)
219 for (Thread* th : Threads)
221 th->start_searching();
223 Thread::search(); // Let's start searching!
226 // When we reach the maximum depth, we can arrive here without a raise of
227 // Threads.stop. However, if we are pondering or in an infinite search,
228 // the UCI protocol states that we shouldn't print the best move before the
229 // GUI sends a "stop" or "ponderhit" command. We therefore simply wait here
230 // until the GUI sends one of those commands (which also raises Threads.stop).
231 Threads.stopOnPonderhit = true;
233 while (!Threads.stop && (Threads.ponder || Limits.infinite))
234 {} // Busy wait for a stop or a ponder reset
236 // Stop the threads if not already stopped (also raise the stop if
237 // "ponderhit" just reset Threads.ponder).
240 // Wait until all threads have finished
241 for (Thread* th : Threads)
243 th->wait_for_search_finished();
245 // When playing in 'nodes as time' mode, subtract the searched nodes from
246 // the available ones before exiting.
248 Time.availableNodes += Limits.inc[us] - Threads.nodes_searched();
250 // Check if there are threads with a better score than main thread
251 Thread* bestThread = this;
252 if ( Options["MultiPV"] == 1
254 && !Skill(Options["Skill Level"]).enabled()
255 && rootMoves[0].pv[0] != MOVE_NONE)
257 std::map<Move, int> votes;
258 Value minScore = this->rootMoves[0].score;
260 // Find out minimum score and reset votes for moves which can be voted
261 for (Thread* th: Threads)
263 minScore = std::min(minScore, th->rootMoves[0].score);
264 votes[th->rootMoves[0].pv[0]] = 0;
267 // Vote according to score and depth
268 for (Thread* th : Threads)
269 votes[th->rootMoves[0].pv[0]] += int(th->rootMoves[0].score - minScore)
270 + int(th->completedDepth);
272 // Select best thread
273 int bestVote = votes[this->rootMoves[0].pv[0]];
274 for (Thread* th : Threads)
276 if (votes[th->rootMoves[0].pv[0]] > bestVote)
278 bestVote = votes[th->rootMoves[0].pv[0]];
284 previousScore = bestThread->rootMoves[0].score;
286 // Send again PV info if we have a new best thread
287 if (bestThread != this)
288 sync_cout << UCI::pv(bestThread->rootPos, bestThread->completedDepth, -VALUE_INFINITE, VALUE_INFINITE) << sync_endl;
290 sync_cout << "bestmove " << UCI::move(bestThread->rootMoves[0].pv[0], rootPos.is_chess960());
292 if (bestThread->rootMoves[0].pv.size() > 1 || bestThread->rootMoves[0].extract_ponder_from_tt(rootPos))
293 std::cout << " ponder " << UCI::move(bestThread->rootMoves[0].pv[1], rootPos.is_chess960());
295 std::cout << sync_endl;
299 /// Thread::search() is the main iterative deepening loop. It calls search()
300 /// repeatedly with increasing depth until the allocated thinking time has been
301 /// consumed, the user stops the search, or the maximum search depth is reached.
303 void Thread::search() {
305 Stack stack[MAX_PLY+7], *ss = stack+4; // To reference from (ss-4) to (ss+2)
306 Value bestValue, alpha, beta, delta;
307 Move lastBestMove = MOVE_NONE;
308 Depth lastBestMoveDepth = DEPTH_ZERO;
309 MainThread* mainThread = (this == Threads.main() ? Threads.main() : nullptr);
310 double timeReduction = 1.0;
311 Color us = rootPos.side_to_move();
314 std::memset(ss-4, 0, 7 * sizeof(Stack));
315 for (int i = 4; i > 0; i--)
316 (ss-i)->continuationHistory = &this->continuationHistory[NO_PIECE][0]; // Use as sentinel
318 bestValue = delta = alpha = -VALUE_INFINITE;
319 beta = VALUE_INFINITE;
322 mainThread->bestMoveChanges = 0, failedLow = false;
324 size_t multiPV = Options["MultiPV"];
325 Skill skill(Options["Skill Level"]);
327 // When playing with strength handicap enable MultiPV search that we will
328 // use behind the scenes to retrieve a set of possible moves.
330 multiPV = std::max(multiPV, (size_t)4);
332 multiPV = std::min(multiPV, rootMoves.size());
334 int ct = int(Options["Contempt"]) * PawnValueEg / 100; // From centipawns
336 // In analysis mode, adjust contempt in accordance with user preference
337 if (Limits.infinite || Options["UCI_AnalyseMode"])
338 ct = Options["Analysis Contempt"] == "Off" ? 0
339 : Options["Analysis Contempt"] == "Both" ? ct
340 : Options["Analysis Contempt"] == "White" && us == BLACK ? -ct
341 : Options["Analysis Contempt"] == "Black" && us == WHITE ? -ct
344 // In evaluate.cpp the evaluation is from the white point of view
345 contempt = (us == WHITE ? make_score(ct, ct / 2)
346 : -make_score(ct, ct / 2));
348 // Iterative deepening loop until requested to stop or the target depth is reached
349 while ( (rootDepth += ONE_PLY) < DEPTH_MAX
351 && !(Limits.depth && mainThread && rootDepth / ONE_PLY > Limits.depth))
353 // Distribute search depths across the helper threads
356 int i = (idx - 1) % 20;
357 if (((rootDepth / ONE_PLY + SkipPhase[i]) / SkipSize[i]) % 2)
358 continue; // Retry with an incremented rootDepth
361 // Age out PV variability metric
363 mainThread->bestMoveChanges *= 0.517, failedLow = false;
365 // Save the last iteration's scores before first PV line is searched and
366 // all the move scores except the (new) PV are set to -VALUE_INFINITE.
367 for (RootMove& rm : rootMoves)
368 rm.previousScore = rm.score;
373 // MultiPV loop. We perform a full root search for each PV line
374 for (pvIdx = 0; pvIdx < multiPV && !Threads.stop; ++pvIdx)
379 for (pvLast++; pvLast < rootMoves.size(); pvLast++)
380 if (rootMoves[pvLast].tbRank != rootMoves[pvFirst].tbRank)
384 // Reset UCI info selDepth for each depth and each PV line
387 // Reset aspiration window starting size
388 if (rootDepth >= 5 * ONE_PLY)
390 Value previousScore = rootMoves[pvIdx].previousScore;
392 alpha = std::max(previousScore - delta,-VALUE_INFINITE);
393 beta = std::min(previousScore + delta, VALUE_INFINITE);
395 // Adjust contempt based on root move's previousScore (dynamic contempt)
396 int dct = ct + 88 * previousScore / (abs(previousScore) + 200);
398 contempt = (us == WHITE ? make_score(dct, dct / 2)
399 : -make_score(dct, dct / 2));
402 // Start with a small aspiration window and, in the case of a fail
403 // high/low, re-search with a bigger window until we don't fail
405 int failedHighCnt = 0;
408 Depth adjustedDepth = std::max(ONE_PLY, rootDepth - failedHighCnt * ONE_PLY);
409 bestValue = ::search<PV>(rootPos, ss, alpha, beta, adjustedDepth, false);
411 // Bring the best move to the front. It is critical that sorting
412 // is done with a stable algorithm because all the values but the
413 // first and eventually the new best one are set to -VALUE_INFINITE
414 // and we want to keep the same order for all the moves except the
415 // new PV that goes to the front. Note that in case of MultiPV
416 // search the already searched PV lines are preserved.
417 std::stable_sort(rootMoves.begin() + pvIdx, rootMoves.begin() + pvLast);
419 // If search has been stopped, we break immediately. Sorting is
420 // safe because RootMoves is still valid, although it refers to
421 // the previous iteration.
425 // When failing high/low give some update (without cluttering
426 // the UI) before a re-search.
429 && (bestValue <= alpha || bestValue >= beta)
430 && Time.elapsed() > 3000)
431 sync_cout << UCI::pv(rootPos, rootDepth, alpha, beta) << sync_endl;
433 // In case of failing low/high increase aspiration window and
434 // re-search, otherwise exit the loop.
435 if (bestValue <= alpha)
437 beta = (alpha + beta) / 2;
438 alpha = std::max(bestValue - delta, -VALUE_INFINITE);
444 Threads.stopOnPonderhit = false;
447 else if (bestValue >= beta)
449 beta = std::min(bestValue + delta, VALUE_INFINITE);
456 delta += delta / 4 + 5;
458 assert(alpha >= -VALUE_INFINITE && beta <= VALUE_INFINITE);
461 // Sort the PV lines searched so far and update the GUI
462 std::stable_sort(rootMoves.begin() + pvFirst, rootMoves.begin() + pvIdx + 1);
465 && (Threads.stop || pvIdx + 1 == multiPV || Time.elapsed() > 3000))
466 sync_cout << UCI::pv(rootPos, rootDepth, alpha, beta) << sync_endl;
470 completedDepth = rootDepth;
472 if (rootMoves[0].pv[0] != lastBestMove) {
473 lastBestMove = rootMoves[0].pv[0];
474 lastBestMoveDepth = rootDepth;
477 // Have we found a "mate in x"?
479 && bestValue >= VALUE_MATE_IN_MAX_PLY
480 && VALUE_MATE - bestValue <= 2 * Limits.mate)
486 // If skill level is enabled and time is up, pick a sub-optimal best move
487 if (skill.enabled() && skill.time_to_pick(rootDepth))
488 skill.pick_best(multiPV);
490 // Do we have time for the next iteration? Can we stop searching now?
491 if ( Limits.use_time_management()
493 && !Threads.stopOnPonderhit)
495 const int F[] = { failedLow,
496 bestValue - mainThread->previousScore };
498 int improvingFactor = std::max(246, std::min(832, 306 + 119 * F[0] - 6 * F[1]));
500 // If the bestMove is stable over several iterations, reduce time accordingly
502 for (int i : {3, 4, 5})
503 if (lastBestMoveDepth * i < completedDepth)
504 timeReduction *= 1.25;
506 // Use part of the gained time from a previous stable move for the current move
507 double bestMoveInstability = 1.0 + mainThread->bestMoveChanges;
508 bestMoveInstability *= std::pow(mainThread->previousTimeReduction, 0.528) / timeReduction;
510 // Stop the search if we have only one legal move, or if available time elapsed
511 if ( rootMoves.size() == 1
512 || Time.elapsed() > Time.optimum() * bestMoveInstability * improvingFactor / 581)
514 // If we are allowed to ponder do not stop the search now but
515 // keep pondering until the GUI sends "ponderhit" or "stop".
517 Threads.stopOnPonderhit = true;
527 mainThread->previousTimeReduction = timeReduction;
529 // If skill level is enabled, swap best PV line with the sub-optimal one
531 std::swap(rootMoves[0], *std::find(rootMoves.begin(), rootMoves.end(),
532 skill.best ? skill.best : skill.pick_best(multiPV)));
538 // search<>() is the main search function for both PV and non-PV nodes
540 template <NodeType NT>
541 Value search(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth, bool cutNode) {
543 constexpr bool PvNode = NT == PV;
544 const bool rootNode = PvNode && ss->ply == 0;
546 // Check if we have an upcoming move which draws by repetition, or
547 // if the opponent had an alternative move earlier to this position.
548 if ( pos.rule50_count() >= 3
549 && alpha < VALUE_DRAW
551 && pos.has_game_cycle(ss->ply))
553 alpha = value_draw(depth, pos.this_thread());
558 // Dive into quiescence search when the depth reaches zero
560 return qsearch<NT>(pos, ss, alpha, beta);
562 assert(-VALUE_INFINITE <= alpha && alpha < beta && beta <= VALUE_INFINITE);
563 assert(PvNode || (alpha == beta - 1));
564 assert(DEPTH_ZERO < depth && depth < DEPTH_MAX);
565 assert(!(PvNode && cutNode));
566 assert(depth / ONE_PLY * ONE_PLY == depth);
568 Move pv[MAX_PLY+1], capturesSearched[32], quietsSearched[64];
572 Move ttMove, move, excludedMove, bestMove;
573 Depth extension, newDepth;
574 Value bestValue, value, ttValue, eval, maxValue, pureStaticEval;
575 bool ttHit, inCheck, givesCheck, improving;
576 bool captureOrPromotion, doFullDepthSearch, moveCountPruning, skipQuiets, ttCapture, pvExact;
578 int moveCount, captureCount, quietCount;
580 // Step 1. Initialize node
581 Thread* thisThread = pos.this_thread();
582 inCheck = pos.checkers();
583 Color us = pos.side_to_move();
584 moveCount = captureCount = quietCount = ss->moveCount = 0;
585 bestValue = -VALUE_INFINITE;
586 maxValue = VALUE_INFINITE;
588 // Check for the available remaining time
589 if (thisThread == Threads.main())
590 static_cast<MainThread*>(thisThread)->check_time();
592 // Used to send selDepth info to GUI (selDepth counts from 1, ply from 0)
593 if (PvNode && thisThread->selDepth < ss->ply + 1)
594 thisThread->selDepth = ss->ply + 1;
598 // Step 2. Check for aborted search and immediate draw
599 if ( Threads.stop.load(std::memory_order_relaxed)
600 || pos.is_draw(ss->ply)
601 || ss->ply >= MAX_PLY)
602 return (ss->ply >= MAX_PLY && !inCheck) ? evaluate(pos)
603 : value_draw(depth, pos.this_thread());
605 // Step 3. Mate distance pruning. Even if we mate at the next move our score
606 // would be at best mate_in(ss->ply+1), but if alpha is already bigger because
607 // a shorter mate was found upward in the tree then there is no need to search
608 // because we will never beat the current alpha. Same logic but with reversed
609 // signs applies also in the opposite condition of being mated instead of giving
610 // mate. In this case return a fail-high score.
611 alpha = std::max(mated_in(ss->ply), alpha);
612 beta = std::min(mate_in(ss->ply+1), beta);
617 assert(0 <= ss->ply && ss->ply < MAX_PLY);
619 (ss+1)->ply = ss->ply + 1;
620 ss->currentMove = (ss+1)->excludedMove = bestMove = MOVE_NONE;
621 ss->continuationHistory = &thisThread->continuationHistory[NO_PIECE][0];
622 (ss+2)->killers[0] = (ss+2)->killers[1] = MOVE_NONE;
623 Square prevSq = to_sq((ss-1)->currentMove);
625 // Initialize statScore to zero for the grandchildren of the current position.
626 // So statScore is shared between all grandchildren and only the first grandchild
627 // starts with statScore = 0. Later grandchildren start with the last calculated
628 // statScore of the previous grandchild. This influences the reduction rules in
629 // LMR which are based on the statScore of parent position.
630 (ss+2)->statScore = 0;
632 // Step 4. Transposition table lookup. We don't want the score of a partial
633 // search to overwrite a previous full search TT value, so we use a different
634 // position key in case of an excluded move.
635 excludedMove = ss->excludedMove;
636 posKey = pos.key() ^ Key(excludedMove << 16); // Isn't a very good hash
637 tte = TT.probe(posKey, ttHit);
638 ttValue = ttHit ? value_from_tt(tte->value(), ss->ply) : VALUE_NONE;
639 ttMove = rootNode ? thisThread->rootMoves[thisThread->pvIdx].pv[0]
640 : ttHit ? tte->move() : MOVE_NONE;
642 // At non-PV nodes we check for an early TT cutoff
645 && tte->depth() >= depth
646 && ttValue != VALUE_NONE // Possible in case of TT access race
647 && (ttValue >= beta ? (tte->bound() & BOUND_LOWER)
648 : (tte->bound() & BOUND_UPPER)))
650 // If ttMove is quiet, update move sorting heuristics on TT hit
655 if (!pos.capture_or_promotion(ttMove))
656 update_quiet_stats(pos, ss, ttMove, nullptr, 0, stat_bonus(depth));
658 // Extra penalty for a quiet TT move in previous ply when it gets refuted
659 if ((ss-1)->moveCount == 1 && !pos.captured_piece())
660 update_continuation_histories(ss-1, pos.piece_on(prevSq), prevSq, -stat_bonus(depth + ONE_PLY));
662 // Penalty for a quiet ttMove that fails low
663 else if (!pos.capture_or_promotion(ttMove))
665 int penalty = -stat_bonus(depth);
666 thisThread->mainHistory[us][from_to(ttMove)] << penalty;
667 update_continuation_histories(ss, pos.moved_piece(ttMove), to_sq(ttMove), penalty);
673 // Step 5. Tablebases probe
674 if (!rootNode && TB::Cardinality)
676 int piecesCount = pos.count<ALL_PIECES>();
678 if ( piecesCount <= TB::Cardinality
679 && (piecesCount < TB::Cardinality || depth >= TB::ProbeDepth)
680 && pos.rule50_count() == 0
681 && !pos.can_castle(ANY_CASTLING))
684 TB::WDLScore wdl = Tablebases::probe_wdl(pos, &err);
686 // Force check of time on the next occasion
687 if (thisThread == Threads.main())
688 static_cast<MainThread*>(thisThread)->callsCnt = 0;
690 if (err != TB::ProbeState::FAIL)
692 thisThread->tbHits.fetch_add(1, std::memory_order_relaxed);
694 int drawScore = TB::UseRule50 ? 1 : 0;
696 value = wdl < -drawScore ? -VALUE_MATE + MAX_PLY + ss->ply + 1
697 : wdl > drawScore ? VALUE_MATE - MAX_PLY - ss->ply - 1
698 : VALUE_DRAW + 2 * wdl * drawScore;
700 Bound b = wdl < -drawScore ? BOUND_UPPER
701 : wdl > drawScore ? BOUND_LOWER : BOUND_EXACT;
703 if ( b == BOUND_EXACT
704 || (b == BOUND_LOWER ? value >= beta : value <= alpha))
706 tte->save(posKey, value_to_tt(value, ss->ply), b,
707 std::min(DEPTH_MAX - ONE_PLY, depth + 6 * ONE_PLY),
708 MOVE_NONE, VALUE_NONE);
715 if (b == BOUND_LOWER)
716 bestValue = value, alpha = std::max(alpha, bestValue);
724 // Step 6. Static evaluation of the position
727 ss->staticEval = eval = pureStaticEval = VALUE_NONE;
729 goto moves_loop; // Skip early pruning when in check
733 // Never assume anything on values stored in TT
734 ss->staticEval = eval = pureStaticEval = tte->eval();
735 if (eval == VALUE_NONE)
736 ss->staticEval = eval = pureStaticEval = evaluate(pos);
738 // Can ttValue be used as a better position evaluation?
739 if ( ttValue != VALUE_NONE
740 && (tte->bound() & (ttValue > eval ? BOUND_LOWER : BOUND_UPPER)))
745 if ((ss-1)->currentMove != MOVE_NULL)
747 int p = (ss-1)->statScore;
748 int bonus = p > 0 ? (-p - 2500) / 512 :
749 p < 0 ? (-p + 2500) / 512 : 0;
751 pureStaticEval = evaluate(pos);
752 ss->staticEval = eval = pureStaticEval + bonus;
755 ss->staticEval = eval = pureStaticEval = -(ss-1)->staticEval + 2 * Eval::Tempo;
757 tte->save(posKey, VALUE_NONE, BOUND_NONE, DEPTH_NONE, MOVE_NONE, pureStaticEval);
760 // Step 7. Razoring (~2 Elo)
761 if ( depth < 2 * ONE_PLY
762 && eval <= alpha - RazorMargin)
763 return qsearch<NT>(pos, ss, alpha, beta);
765 improving = ss->staticEval >= (ss-2)->staticEval
766 || (ss-2)->staticEval == VALUE_NONE;
768 // Step 8. Futility pruning: child node (~30 Elo)
770 && depth < 7 * ONE_PLY
771 && eval - futility_margin(depth, improving) >= beta
772 && eval < VALUE_KNOWN_WIN) // Do not return unproven wins
775 // Step 9. Null move search with verification search (~40 Elo)
777 && (ss-1)->currentMove != MOVE_NULL
778 && (ss-1)->statScore < 23200
780 && pureStaticEval >= beta - 36 * depth / ONE_PLY + 225
782 && pos.non_pawn_material(us)
783 && (ss->ply >= thisThread->nmpMinPly || us != thisThread->nmpColor))
785 assert(eval - beta >= 0);
787 // Null move dynamic reduction based on depth and value
788 Depth R = ((823 + 67 * depth / ONE_PLY) / 256 + std::min(int(eval - beta) / 200, 3)) * ONE_PLY;
790 ss->currentMove = MOVE_NULL;
791 ss->continuationHistory = &thisThread->continuationHistory[NO_PIECE][0];
793 pos.do_null_move(st);
795 Value nullValue = -search<NonPV>(pos, ss+1, -beta, -beta+1, depth-R, !cutNode);
797 pos.undo_null_move();
799 if (nullValue >= beta)
801 // Do not return unproven mate scores
802 if (nullValue >= VALUE_MATE_IN_MAX_PLY)
805 if (thisThread->nmpMinPly || (abs(beta) < VALUE_KNOWN_WIN && depth < 12 * ONE_PLY))
808 assert(!thisThread->nmpMinPly); // Recursive verification is not allowed
810 // Do verification search at high depths, with null move pruning disabled
811 // for us, until ply exceeds nmpMinPly.
812 thisThread->nmpMinPly = ss->ply + 3 * (depth-R) / 4;
813 thisThread->nmpColor = us;
815 Value v = search<NonPV>(pos, ss, beta-1, beta, depth-R, false);
817 thisThread->nmpMinPly = 0;
824 // Step 10. ProbCut (~10 Elo)
825 // If we have a good enough capture and a reduced search returns a value
826 // much above beta, we can (almost) safely prune the previous move.
828 && depth >= 5 * ONE_PLY
829 && abs(beta) < VALUE_MATE_IN_MAX_PLY)
831 Value rbeta = std::min(beta + 216 - 48 * improving, VALUE_INFINITE);
832 MovePicker mp(pos, ttMove, rbeta - ss->staticEval, &thisThread->captureHistory);
833 int probCutCount = 0;
835 while ( (move = mp.next_move()) != MOVE_NONE
837 if (move != excludedMove && pos.legal(move))
841 ss->currentMove = move;
842 ss->continuationHistory = &thisThread->continuationHistory[pos.moved_piece(move)][to_sq(move)];
844 assert(depth >= 5 * ONE_PLY);
846 pos.do_move(move, st);
848 // Perform a preliminary qsearch to verify that the move holds
849 value = -qsearch<NonPV>(pos, ss+1, -rbeta, -rbeta+1);
851 // If the qsearch held perform the regular search
853 value = -search<NonPV>(pos, ss+1, -rbeta, -rbeta+1, depth - 4 * ONE_PLY, !cutNode);
862 // Step 11. Internal iterative deepening (~2 Elo)
863 if ( depth >= 8 * ONE_PLY
866 search<NT>(pos, ss, alpha, beta, depth - 7 * ONE_PLY, cutNode);
868 tte = TT.probe(posKey, ttHit);
869 ttValue = ttHit ? value_from_tt(tte->value(), ss->ply) : VALUE_NONE;
870 ttMove = ttHit ? tte->move() : MOVE_NONE;
873 moves_loop: // When in check, search starts from here
875 const PieceToHistory* contHist[] = { (ss-1)->continuationHistory, (ss-2)->continuationHistory, nullptr, (ss-4)->continuationHistory };
876 Move countermove = thisThread->counterMoves[pos.piece_on(prevSq)][prevSq];
878 MovePicker mp(pos, ttMove, depth, &thisThread->mainHistory,
879 &thisThread->captureHistory,
883 value = bestValue; // Workaround a bogus 'uninitialized' warning under gcc
886 ttCapture = ttMove && pos.capture_or_promotion(ttMove);
887 pvExact = PvNode && ttHit && tte->bound() == BOUND_EXACT;
889 // Step 12. Loop through all pseudo-legal moves until no moves remain
890 // or a beta cutoff occurs.
891 while ((move = mp.next_move(skipQuiets)) != MOVE_NONE)
895 if (move == excludedMove)
898 // At root obey the "searchmoves" option and skip moves not listed in Root
899 // Move List. As a consequence any illegal move is also skipped. In MultiPV
900 // mode we also skip PV moves which have been already searched and those
901 // of lower "TB rank" if we are in a TB root position.
902 if (rootNode && !std::count(thisThread->rootMoves.begin() + thisThread->pvIdx,
903 thisThread->rootMoves.begin() + thisThread->pvLast, move))
906 ss->moveCount = ++moveCount;
908 if (rootNode && thisThread == Threads.main() && Time.elapsed() > 3000)
909 sync_cout << "info depth " << depth / ONE_PLY
910 << " currmove " << UCI::move(move, pos.is_chess960())
911 << " currmovenumber " << moveCount + thisThread->pvIdx << sync_endl;
913 (ss+1)->pv = nullptr;
915 extension = DEPTH_ZERO;
916 captureOrPromotion = pos.capture_or_promotion(move);
917 movedPiece = pos.moved_piece(move);
918 givesCheck = gives_check(pos, move);
920 moveCountPruning = depth < 16 * ONE_PLY
921 && moveCount >= FutilityMoveCounts[improving][depth / ONE_PLY];
923 // Step 13. Extensions (~70 Elo)
925 // Singular extension search (~60 Elo). If all moves but one fail low on a
926 // search of (alpha-s, beta-s), and just one fails high on (alpha, beta),
927 // then that move is singular and should be extended. To verify this we do
928 // a reduced search on all the other moves but the ttMove and if the
929 // result is lower than ttValue minus a margin then we will extend the ttMove.
930 if ( depth >= 8 * ONE_PLY
933 && !excludedMove // Recursive singular search is not allowed
934 && ttValue != VALUE_NONE
935 && (tte->bound() & BOUND_LOWER)
936 && tte->depth() >= depth - 3 * ONE_PLY
939 Value rBeta = std::max(ttValue - 2 * depth / ONE_PLY, -VALUE_MATE);
940 ss->excludedMove = move;
941 value = search<NonPV>(pos, ss, rBeta - 1, rBeta, depth / 2, cutNode);
942 ss->excludedMove = MOVE_NONE;
947 else if ( givesCheck // Check extension (~2 Elo)
951 // Extension if castling
952 else if (type_of(move) == CASTLING)
955 // Calculate new depth for this move
956 newDepth = depth - ONE_PLY + extension;
958 // Step 14. Pruning at shallow depth (~170 Elo)
960 && pos.non_pawn_material(us)
961 && bestValue > VALUE_MATED_IN_MAX_PLY)
963 if ( !captureOrPromotion
965 && (!pos.advanced_pawn_push(move) || pos.non_pawn_material() >= Value(5000)))
967 // Move count based pruning (~30 Elo)
968 if (moveCountPruning)
974 // Reduced depth of the next LMR search
975 int lmrDepth = std::max(newDepth - reduction<PvNode>(improving, depth, moveCount), DEPTH_ZERO) / ONE_PLY;
977 // Countermoves based pruning (~20 Elo)
978 if ( lmrDepth < 3 + ((ss-1)->statScore > 0)
979 && (*contHist[0])[movedPiece][to_sq(move)] < CounterMovePruneThreshold
980 && (*contHist[1])[movedPiece][to_sq(move)] < CounterMovePruneThreshold)
983 // Futility pruning: parent node (~2 Elo)
986 && ss->staticEval + 256 + 200 * lmrDepth <= alpha)
989 // Prune moves with negative SEE (~10 Elo)
990 if (!pos.see_ge(move, Value(-29 * lmrDepth * lmrDepth)))
993 else if ( !extension // (~20 Elo)
994 && !pos.see_ge(move, -PawnValueEg * (depth / ONE_PLY)))
998 // Speculative prefetch as early as possible
999 prefetch(TT.first_entry(pos.key_after(move)));
1001 // Check for legality just before making the move
1002 if (!rootNode && !pos.legal(move))
1004 ss->moveCount = --moveCount;
1008 // Update the current move (this must be done after singular extension search)
1009 ss->currentMove = move;
1010 ss->continuationHistory = &thisThread->continuationHistory[movedPiece][to_sq(move)];
1012 // Step 15. Make the move
1013 pos.do_move(move, st, givesCheck);
1015 // Step 16. Reduced depth search (LMR). If the move fails high it will be
1016 // re-searched at full depth.
1017 if ( depth >= 3 * ONE_PLY
1019 && (!captureOrPromotion || moveCountPruning))
1021 Depth r = reduction<PvNode>(improving, depth, moveCount);
1023 // Decrease reduction if opponent's move count is high (~10 Elo)
1024 if ((ss-1)->moveCount > 15)
1027 if (!captureOrPromotion)
1029 // Decrease reduction for exact PV nodes (~0 Elo)
1033 // Increase reduction if ttMove is a capture (~0 Elo)
1037 // Increase reduction for cut nodes (~5 Elo)
1041 // Decrease reduction for moves that escape a capture. Filter out
1042 // castling moves, because they are coded as "king captures rook" and
1043 // hence break make_move(). (~5 Elo)
1044 else if ( type_of(move) == NORMAL
1045 && !pos.see_ge(make_move(to_sq(move), from_sq(move))))
1048 ss->statScore = thisThread->mainHistory[us][from_to(move)]
1049 + (*contHist[0])[movedPiece][to_sq(move)]
1050 + (*contHist[1])[movedPiece][to_sq(move)]
1051 + (*contHist[3])[movedPiece][to_sq(move)]
1054 // Decrease/increase reduction by comparing opponent's stat score (~10 Elo)
1055 if (ss->statScore >= 0 && (ss-1)->statScore < 0)
1058 else if ((ss-1)->statScore >= 0 && ss->statScore < 0)
1061 // Decrease/increase reduction for moves with a good/bad history (~30 Elo)
1062 r -= ss->statScore / 20000 * ONE_PLY;
1065 Depth d = std::max(newDepth - std::max(r, DEPTH_ZERO), ONE_PLY);
1067 value = -search<NonPV>(pos, ss+1, -(alpha+1), -alpha, d, true);
1069 doFullDepthSearch = (value > alpha && d != newDepth);
1072 doFullDepthSearch = !PvNode || moveCount > 1;
1074 // Step 17. Full depth search when LMR is skipped or fails high
1075 if (doFullDepthSearch)
1076 value = -search<NonPV>(pos, ss+1, -(alpha+1), -alpha, newDepth, !cutNode);
1078 // For PV nodes only, do a full PV search on the first move or after a fail
1079 // high (in the latter case search only if value < beta), otherwise let the
1080 // parent node fail low with value <= alpha and try another move.
1081 if (PvNode && (moveCount == 1 || (value > alpha && (rootNode || value < beta))))
1084 (ss+1)->pv[0] = MOVE_NONE;
1086 value = -search<PV>(pos, ss+1, -beta, -alpha, newDepth, false);
1089 // Step 18. Undo move
1090 pos.undo_move(move);
1092 assert(value > -VALUE_INFINITE && value < VALUE_INFINITE);
1094 // Step 19. Check for a new best move
1095 // Finished searching the move. If a stop occurred, the return value of
1096 // the search cannot be trusted, and we return immediately without
1097 // updating best move, PV and TT.
1098 if (Threads.stop.load(std::memory_order_relaxed))
1103 RootMove& rm = *std::find(thisThread->rootMoves.begin(),
1104 thisThread->rootMoves.end(), move);
1106 // PV move or new best move?
1107 if (moveCount == 1 || value > alpha)
1110 rm.selDepth = thisThread->selDepth;
1115 for (Move* m = (ss+1)->pv; *m != MOVE_NONE; ++m)
1116 rm.pv.push_back(*m);
1118 // We record how often the best move has been changed in each
1119 // iteration. This information is used for time management: When
1120 // the best move changes frequently, we allocate some more time.
1121 if (moveCount > 1 && thisThread == Threads.main())
1122 ++static_cast<MainThread*>(thisThread)->bestMoveChanges;
1125 // All other moves but the PV are set to the lowest value: this
1126 // is not a problem when sorting because the sort is stable and the
1127 // move position in the list is preserved - just the PV is pushed up.
1128 rm.score = -VALUE_INFINITE;
1131 if (value > bestValue)
1139 if (PvNode && !rootNode) // Update pv even in fail-high case
1140 update_pv(ss->pv, move, (ss+1)->pv);
1142 if (PvNode && value < beta) // Update alpha! Always alpha < beta
1146 assert(value >= beta); // Fail high
1153 if (move != bestMove)
1155 if (captureOrPromotion && captureCount < 32)
1156 capturesSearched[captureCount++] = move;
1158 else if (!captureOrPromotion && quietCount < 64)
1159 quietsSearched[quietCount++] = move;
1163 // The following condition would detect a stop only after move loop has been
1164 // completed. But in this case bestValue is valid because we have fully
1165 // searched our subtree, and we can anyhow save the result in TT.
1171 // Step 20. Check for mate and stalemate
1172 // All legal moves have been searched and if there are no legal moves, it
1173 // must be a mate or a stalemate. If we are in a singular extension search then
1174 // return a fail low score.
1176 assert(moveCount || !inCheck || excludedMove || !MoveList<LEGAL>(pos).size());
1179 bestValue = excludedMove ? alpha
1180 : inCheck ? mated_in(ss->ply) : VALUE_DRAW;
1183 // Quiet best move: update move sorting heuristics
1184 if (!pos.capture_or_promotion(bestMove))
1185 update_quiet_stats(pos, ss, bestMove, quietsSearched, quietCount,
1186 stat_bonus(depth + (bestValue > beta + PawnValueMg ? ONE_PLY : DEPTH_ZERO)));
1188 update_capture_stats(pos, bestMove, capturesSearched, captureCount, stat_bonus(depth + ONE_PLY));
1190 // Extra penalty for a quiet TT move in previous ply when it gets refuted
1191 if ((ss-1)->moveCount == 1 && !pos.captured_piece())
1192 update_continuation_histories(ss-1, pos.piece_on(prevSq), prevSq, -stat_bonus(depth + ONE_PLY));
1194 // Bonus for prior countermove that caused the fail low
1195 else if ( (depth >= 3 * ONE_PLY || PvNode)
1196 && !pos.captured_piece()
1197 && is_ok((ss-1)->currentMove))
1198 update_continuation_histories(ss-1, pos.piece_on(prevSq), prevSq, stat_bonus(depth));
1201 bestValue = std::min(bestValue, maxValue);
1204 tte->save(posKey, value_to_tt(bestValue, ss->ply),
1205 bestValue >= beta ? BOUND_LOWER :
1206 PvNode && bestMove ? BOUND_EXACT : BOUND_UPPER,
1207 depth, bestMove, pureStaticEval);
1209 assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE);
1215 // qsearch() is the quiescence search function, which is called by the main
1216 // search function with depth zero, or recursively with depth less than ONE_PLY.
1217 template <NodeType NT>
1218 Value qsearch(Position& pos, Stack* ss, Value alpha, Value beta, Depth depth) {
1220 constexpr bool PvNode = NT == PV;
1222 assert(alpha >= -VALUE_INFINITE && alpha < beta && beta <= VALUE_INFINITE);
1223 assert(PvNode || (alpha == beta - 1));
1224 assert(depth <= DEPTH_ZERO);
1225 assert(depth / ONE_PLY * ONE_PLY == depth);
1231 Move ttMove, move, bestMove;
1233 Value bestValue, value, ttValue, futilityValue, futilityBase, oldAlpha;
1234 bool ttHit, inCheck, givesCheck, evasionPrunable;
1239 oldAlpha = alpha; // To flag BOUND_EXACT when eval above alpha and no available moves
1241 ss->pv[0] = MOVE_NONE;
1244 Thread* thisThread = pos.this_thread();
1245 (ss+1)->ply = ss->ply + 1;
1246 ss->currentMove = bestMove = MOVE_NONE;
1247 ss->continuationHistory = &thisThread->continuationHistory[NO_PIECE][0];
1248 inCheck = pos.checkers();
1251 // Check for an immediate draw or maximum ply reached
1252 if ( pos.is_draw(ss->ply)
1253 || ss->ply >= MAX_PLY)
1254 return (ss->ply >= MAX_PLY && !inCheck) ? evaluate(pos) : VALUE_DRAW;
1256 assert(0 <= ss->ply && ss->ply < MAX_PLY);
1258 // Decide whether or not to include checks: this fixes also the type of
1259 // TT entry depth that we are going to use. Note that in qsearch we use
1260 // only two types of depth in TT: DEPTH_QS_CHECKS or DEPTH_QS_NO_CHECKS.
1261 ttDepth = inCheck || depth >= DEPTH_QS_CHECKS ? DEPTH_QS_CHECKS
1262 : DEPTH_QS_NO_CHECKS;
1263 // Transposition table lookup
1265 tte = TT.probe(posKey, ttHit);
1266 ttValue = ttHit ? value_from_tt(tte->value(), ss->ply) : VALUE_NONE;
1267 ttMove = ttHit ? tte->move() : MOVE_NONE;
1271 && tte->depth() >= ttDepth
1272 && ttValue != VALUE_NONE // Only in case of TT access race
1273 && (ttValue >= beta ? (tte->bound() & BOUND_LOWER)
1274 : (tte->bound() & BOUND_UPPER)))
1277 // Evaluate the position statically
1280 ss->staticEval = VALUE_NONE;
1281 bestValue = futilityBase = -VALUE_INFINITE;
1287 // Never assume anything on values stored in TT
1288 if ((ss->staticEval = bestValue = tte->eval()) == VALUE_NONE)
1289 ss->staticEval = bestValue = evaluate(pos);
1291 // Can ttValue be used as a better position evaluation?
1292 if ( ttValue != VALUE_NONE
1293 && (tte->bound() & (ttValue > bestValue ? BOUND_LOWER : BOUND_UPPER)))
1294 bestValue = ttValue;
1297 ss->staticEval = bestValue =
1298 (ss-1)->currentMove != MOVE_NULL ? evaluate(pos)
1299 : -(ss-1)->staticEval + 2 * Eval::Tempo;
1301 // Stand pat. Return immediately if static value is at least beta
1302 if (bestValue >= beta)
1305 tte->save(posKey, value_to_tt(bestValue, ss->ply), BOUND_LOWER,
1306 DEPTH_NONE, MOVE_NONE, ss->staticEval);
1311 if (PvNode && bestValue > alpha)
1314 futilityBase = bestValue + 128;
1317 const PieceToHistory* contHist[] = { (ss-1)->continuationHistory, (ss-2)->continuationHistory, nullptr, (ss-4)->continuationHistory };
1319 // Initialize a MovePicker object for the current position, and prepare
1320 // to search the moves. Because the depth is <= 0 here, only captures,
1321 // queen promotions and checks (only if depth >= DEPTH_QS_CHECKS) will
1323 MovePicker mp(pos, ttMove, depth, &thisThread->mainHistory,
1324 &thisThread->captureHistory,
1326 to_sq((ss-1)->currentMove));
1328 // Loop through the moves until no moves remain or a beta cutoff occurs
1329 while ((move = mp.next_move()) != MOVE_NONE)
1331 assert(is_ok(move));
1333 givesCheck = gives_check(pos, move);
1340 && futilityBase > -VALUE_KNOWN_WIN
1341 && !pos.advanced_pawn_push(move))
1343 assert(type_of(move) != ENPASSANT); // Due to !pos.advanced_pawn_push
1345 futilityValue = futilityBase + PieceValue[EG][pos.piece_on(to_sq(move))];
1347 if (futilityValue <= alpha)
1349 bestValue = std::max(bestValue, futilityValue);
1353 if (futilityBase <= alpha && !pos.see_ge(move, VALUE_ZERO + 1))
1355 bestValue = std::max(bestValue, futilityBase);
1360 // Detect non-capture evasions that are candidates to be pruned
1361 evasionPrunable = inCheck
1362 && (depth != DEPTH_ZERO || moveCount > 2)
1363 && bestValue > VALUE_MATED_IN_MAX_PLY
1364 && !pos.capture(move);
1366 // Don't search moves with negative SEE values
1367 if ( (!inCheck || evasionPrunable)
1368 && !pos.see_ge(move))
1371 // Speculative prefetch as early as possible
1372 prefetch(TT.first_entry(pos.key_after(move)));
1374 // Check for legality just before making the move
1375 if (!pos.legal(move))
1381 ss->currentMove = move;
1382 ss->continuationHistory = &thisThread->continuationHistory[pos.moved_piece(move)][to_sq(move)];
1384 // Make and search the move
1385 pos.do_move(move, st, givesCheck);
1386 value = -qsearch<NT>(pos, ss+1, -beta, -alpha, depth - ONE_PLY);
1387 pos.undo_move(move);
1389 assert(value > -VALUE_INFINITE && value < VALUE_INFINITE);
1391 // Check for a new best move
1392 if (value > bestValue)
1400 if (PvNode) // Update pv even in fail-high case
1401 update_pv(ss->pv, move, (ss+1)->pv);
1403 if (PvNode && value < beta) // Update alpha here!
1411 // All legal moves have been searched. A special case: If we're in check
1412 // and no legal moves were found, it is checkmate.
1413 if (inCheck && bestValue == -VALUE_INFINITE)
1414 return mated_in(ss->ply); // Plies to mate from the root
1416 tte->save(posKey, value_to_tt(bestValue, ss->ply),
1417 bestValue >= beta ? BOUND_LOWER :
1418 PvNode && bestValue > oldAlpha ? BOUND_EXACT : BOUND_UPPER,
1419 ttDepth, bestMove, ss->staticEval);
1421 assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE);
1427 // value_to_tt() adjusts a mate score from "plies to mate from the root" to
1428 // "plies to mate from the current position". Non-mate scores are unchanged.
1429 // The function is called before storing a value in the transposition table.
1431 Value value_to_tt(Value v, int ply) {
1433 assert(v != VALUE_NONE);
1435 return v >= VALUE_MATE_IN_MAX_PLY ? v + ply
1436 : v <= VALUE_MATED_IN_MAX_PLY ? v - ply : v;
1440 // value_from_tt() is the inverse of value_to_tt(): It adjusts a mate score
1441 // from the transposition table (which refers to the plies to mate/be mated
1442 // from current position) to "plies to mate/be mated from the root".
1444 Value value_from_tt(Value v, int ply) {
1446 return v == VALUE_NONE ? VALUE_NONE
1447 : v >= VALUE_MATE_IN_MAX_PLY ? v - ply
1448 : v <= VALUE_MATED_IN_MAX_PLY ? v + ply : v;
1452 // update_pv() adds current move and appends child pv[]
1454 void update_pv(Move* pv, Move move, Move* childPv) {
1456 for (*pv++ = move; childPv && *childPv != MOVE_NONE; )
1462 // update_continuation_histories() updates histories of the move pairs formed
1463 // by moves at ply -1, -2, and -4 with current move.
1465 void update_continuation_histories(Stack* ss, Piece pc, Square to, int bonus) {
1467 for (int i : {1, 2, 4})
1468 if (is_ok((ss-i)->currentMove))
1469 (*(ss-i)->continuationHistory)[pc][to] << bonus;
1473 // update_capture_stats() updates move sorting heuristics when a new capture best move is found
1475 void update_capture_stats(const Position& pos, Move move,
1476 Move* captures, int captureCnt, int bonus) {
1478 CapturePieceToHistory& captureHistory = pos.this_thread()->captureHistory;
1479 Piece moved_piece = pos.moved_piece(move);
1480 PieceType captured = type_of(pos.piece_on(to_sq(move)));
1482 if (pos.capture_or_promotion(move))
1483 captureHistory[moved_piece][to_sq(move)][captured] << bonus;
1485 // Decrease all the other played capture moves
1486 for (int i = 0; i < captureCnt; ++i)
1488 moved_piece = pos.moved_piece(captures[i]);
1489 captured = type_of(pos.piece_on(to_sq(captures[i])));
1490 captureHistory[moved_piece][to_sq(captures[i])][captured] << -bonus;
1495 // update_quiet_stats() updates move sorting heuristics when a new quiet best move is found
1497 void update_quiet_stats(const Position& pos, Stack* ss, Move move,
1498 Move* quiets, int quietsCnt, int bonus) {
1500 if (ss->killers[0] != move)
1502 ss->killers[1] = ss->killers[0];
1503 ss->killers[0] = move;
1506 Color us = pos.side_to_move();
1507 Thread* thisThread = pos.this_thread();
1508 thisThread->mainHistory[us][from_to(move)] << bonus;
1509 update_continuation_histories(ss, pos.moved_piece(move), to_sq(move), bonus);
1511 if (is_ok((ss-1)->currentMove))
1513 Square prevSq = to_sq((ss-1)->currentMove);
1514 thisThread->counterMoves[pos.piece_on(prevSq)][prevSq] = move;
1517 // Decrease all the other played quiet moves
1518 for (int i = 0; i < quietsCnt; ++i)
1520 thisThread->mainHistory[us][from_to(quiets[i])] << -bonus;
1521 update_continuation_histories(ss, pos.moved_piece(quiets[i]), to_sq(quiets[i]), -bonus);
1525 // When playing with strength handicap, choose best move among a set of RootMoves
1526 // using a statistical rule dependent on 'level'. Idea by Heinz van Saanen.
1528 Move Skill::pick_best(size_t multiPV) {
1530 const RootMoves& rootMoves = Threads.main()->rootMoves;
1531 static PRNG rng(now()); // PRNG sequence should be non-deterministic
1533 // RootMoves are already sorted by score in descending order
1534 Value topScore = rootMoves[0].score;
1535 int delta = std::min(topScore - rootMoves[multiPV - 1].score, PawnValueMg);
1536 int weakness = 120 - 2 * level;
1537 int maxScore = -VALUE_INFINITE;
1539 // Choose best move. For each move score we add two terms, both dependent on
1540 // weakness. One is deterministic and bigger for weaker levels, and one is
1541 // random. Then we choose the move with the resulting highest score.
1542 for (size_t i = 0; i < multiPV; ++i)
1544 // This is our magic formula
1545 int push = ( weakness * int(topScore - rootMoves[i].score)
1546 + delta * (rng.rand<unsigned>() % weakness)) / 128;
1548 if (rootMoves[i].score + push >= maxScore)
1550 maxScore = rootMoves[i].score + push;
1551 best = rootMoves[i].pv[0];
1560 /// MainThread::check_time() is used to print debug info and, more importantly,
1561 /// to detect when we are out of available time and thus stop the search.
1563 void MainThread::check_time() {
1568 // When using nodes, ensure checking rate is not lower than 0.1% of nodes
1569 callsCnt = Limits.nodes ? std::min(1024, int(Limits.nodes / 1024)) : 1024;
1571 static TimePoint lastInfoTime = now();
1573 TimePoint elapsed = Time.elapsed();
1574 TimePoint tick = Limits.startTime + elapsed;
1576 if (tick - lastInfoTime >= 1000)
1578 lastInfoTime = tick;
1582 // We should not stop pondering until told so by the GUI
1586 if ( (Limits.use_time_management() && elapsed > Time.maximum() - 10)
1587 || (Limits.movetime && elapsed >= Limits.movetime)
1588 || (Limits.nodes && Threads.nodes_searched() >= (uint64_t)Limits.nodes))
1589 Threads.stop = true;
1593 /// UCI::pv() formats PV information according to the UCI protocol. UCI requires
1594 /// that all (if any) unsearched PV lines are sent using a previous search score.
1596 string UCI::pv(const Position& pos, Depth depth, Value alpha, Value beta) {
1598 std::stringstream ss;
1599 TimePoint elapsed = Time.elapsed() + 1;
1600 const RootMoves& rootMoves = pos.this_thread()->rootMoves;
1601 size_t pvIdx = pos.this_thread()->pvIdx;
1602 size_t multiPV = std::min((size_t)Options["MultiPV"], rootMoves.size());
1603 uint64_t nodesSearched = Threads.nodes_searched();
1604 uint64_t tbHits = Threads.tb_hits() + (TB::RootInTB ? rootMoves.size() : 0);
1606 for (size_t i = 0; i < multiPV; ++i)
1608 bool updated = (i <= pvIdx && rootMoves[i].score != -VALUE_INFINITE);
1610 if (depth == ONE_PLY && !updated)
1613 Depth d = updated ? depth : depth - ONE_PLY;
1614 Value v = updated ? rootMoves[i].score : rootMoves[i].previousScore;
1616 bool tb = TB::RootInTB && abs(v) < VALUE_MATE - MAX_PLY;
1617 v = tb ? rootMoves[i].tbScore : v;
1619 if (ss.rdbuf()->in_avail()) // Not at first line
1623 << " depth " << d / ONE_PLY
1624 << " seldepth " << rootMoves[i].selDepth
1625 << " multipv " << i + 1
1626 << " score " << UCI::value(v);
1628 if (!tb && i == pvIdx)
1629 ss << (v >= beta ? " lowerbound" : v <= alpha ? " upperbound" : "");
1631 ss << " nodes " << nodesSearched
1632 << " nps " << nodesSearched * 1000 / elapsed;
1634 if (elapsed > 1000) // Earlier makes little sense
1635 ss << " hashfull " << TT.hashfull();
1637 ss << " tbhits " << tbHits
1638 << " time " << elapsed
1641 for (Move m : rootMoves[i].pv)
1642 ss << " " << UCI::move(m, pos.is_chess960());
1649 /// RootMove::extract_ponder_from_tt() is called in case we have no ponder move
1650 /// before exiting the search, for instance, in case we stop the search during a
1651 /// fail high at root. We try hard to have a ponder move to return to the GUI,
1652 /// otherwise in case of 'ponder on' we have nothing to think on.
1654 bool RootMove::extract_ponder_from_tt(Position& pos) {
1659 assert(pv.size() == 1);
1664 pos.do_move(pv[0], st);
1665 TTEntry* tte = TT.probe(pos.key(), ttHit);
1669 Move m = tte->move(); // Local copy to be SMP safe
1670 if (MoveList<LEGAL>(pos).contains(m))
1674 pos.undo_move(pv[0]);
1675 return pv.size() > 1;
1678 void Tablebases::rank_root_moves(Position& pos, Search::RootMoves& rootMoves) {
1681 UseRule50 = bool(Options["Syzygy50MoveRule"]);
1682 ProbeDepth = int(Options["SyzygyProbeDepth"]) * ONE_PLY;
1683 Cardinality = int(Options["SyzygyProbeLimit"]);
1684 bool dtz_available = true;
1686 // Tables with fewer pieces than SyzygyProbeLimit are searched with
1687 // ProbeDepth == DEPTH_ZERO
1688 if (Cardinality > MaxCardinality)
1690 Cardinality = MaxCardinality;
1691 ProbeDepth = DEPTH_ZERO;
1694 if (Cardinality >= popcount(pos.pieces()) && !pos.can_castle(ANY_CASTLING))
1696 // Rank moves using DTZ tables
1697 RootInTB = root_probe(pos, rootMoves);
1701 // DTZ tables are missing; try to rank moves using WDL tables
1702 dtz_available = false;
1703 RootInTB = root_probe_wdl(pos, rootMoves);
1709 // Sort moves according to TB rank
1710 std::sort(rootMoves.begin(), rootMoves.end(),
1711 [](const RootMove &a, const RootMove &b) { return a.tbRank > b.tbRank; } );
1713 // Probe during search only if DTZ is not available and we are winning
1714 if (dtz_available || rootMoves[0].tbScore <= VALUE_DRAW)
1719 // Assign the same rank to all moves
1720 for (auto& m : rootMoves)