X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fsearch.cpp;h=6715d6808f3f26dc28261108247203b62eb1823f;hp=aa577c79b7c4fdc2eab25bb2663170defb68e91f;hb=5f3c660d5d768ea5132439c5a8916673b909c0c9;hpb=276513c19f333eebb2405cb4eea0bb6f8cc31a7b diff --git a/src/search.cpp b/src/search.cpp index aa577c79..6715d680 100644 --- a/src/search.cpp +++ b/src/search.cpp @@ -54,6 +54,10 @@ namespace { /// Types enum NodeType { NonPV, PV }; + // Set to true to force running with one thread. + // Used for debugging SMP code. + const bool FakeSplit = false; + // ThreadsManager class is used to handle all the threads related stuff in search, // init, starting, parking and, the most important, launching a slave thread at a // split point are what this class does. All the access to shared thread data is @@ -83,8 +87,10 @@ namespace { void wake_sleeping_threads(); void put_threads_to_sleep(); void idle_loop(int threadID, SplitPoint* sp); - bool split(const Position& pos, SearchStack* ss, int ply, Value* alpha, const Value beta, Value* bestValue, - Depth depth, bool mateThreat, int* moves, MovePicker* mp, int master, bool pvNode); + + template + void split(const Position& pos, SearchStack* ss, Value* alpha, const Value beta, Value* bestValue, + Depth depth, bool mateThreat, int* moveCount, MovePicker* mp, bool pvNode); private: friend void poll(); @@ -116,7 +122,7 @@ namespace { // RootMove::operator<() is the comparison function used when // sorting the moves. A move m1 is considered to be better // than a move m2 if it has a higher score, or if the moves - // have equal score but m1 has the higher node count. + // have equal score but m1 has the higher beta cut-off count. bool operator<(const RootMove& m) const { return score != m.score ? score < m.score : theirBeta <= m.theirBeta; @@ -182,7 +188,7 @@ namespace { const Depth IIDDepth[2] = { 8 * OnePly /* non-PV */, 5 * OnePly /* PV */}; // At Non-PV nodes we do an internal iterative deepening search - // when the static evaluation is at most IIDMargin below beta. + // when the static evaluation is bigger then beta - IIDMargin. const Value IIDMargin = Value(0x100); // Step 11. Decide the new search depth @@ -276,29 +282,30 @@ namespace { /// Local functions Value id_loop(const Position& pos, Move searchMoves[]); - Value root_search(Position& pos, SearchStack ss[], RootMoveList& rml, Value* alphaPtr, Value* betaPtr); + Value root_search(Position& pos, SearchStack* ss, RootMoveList& rml, Value* alphaPtr, Value* betaPtr); template - Value search(Position& pos, SearchStack ss[], Value alpha, Value beta, Depth depth, int ply, bool allowNullmove, int threadID, Move excludedMove = MOVE_NONE); + Value search(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth); template - Depth extension(const Position& pos, Move m, bool captureOrPromotion, bool moveIsCheck, bool singleEvasion, bool mateThreat, bool* dangerous); + Value qsearch(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth); - Value qsearch(Position& pos, SearchStack ss[], Value alpha, Value beta, Depth depth, int ply, int threadID); + template void sp_search(SplitPoint* sp, int threadID); - void sp_search_pv(SplitPoint* sp, int threadID); - void init_node(SearchStack ss[], int ply, int threadID); - void update_pv(SearchStack ss[], int ply); - void sp_update_pv(SearchStack* pss, SearchStack ss[], int ply); + + template + Depth extension(const Position& pos, Move m, bool captureOrPromotion, bool moveIsCheck, bool singleEvasion, bool mateThreat, bool* dangerous); + + void update_pv(SearchStack* ss, int ply); + void sp_update_pv(SearchStack* pss, SearchStack* ss, int ply); bool connected_moves(const Position& pos, Move m1, Move m2); bool value_is_mate(Value value); - bool move_is_killer(Move m, const SearchStack& ss); - bool ok_to_do_nullmove(const Position& pos); - bool ok_to_prune(const Position& pos, Move m, Move threat); + bool move_is_killer(Move m, SearchStack* ss); bool ok_to_use_TT(const TTEntry* tte, Depth depth, Value beta, int ply); + bool connected_threat(const Position& pos, Move m, Move threat); Value refine_eval(const TTEntry* tte, Value defaultEval, int ply); void update_history(const Position& pos, Move move, Depth depth, Move movesSearched[], int moveCount); - void update_killers(Move m, SearchStack& ss); + void update_killers(Move m, SearchStack* ss); void update_gains(const Position& pos, Move move, Value before, Value after); int current_search_time(); @@ -306,8 +313,8 @@ namespace { void poll(); void ponderhit(); void wait_for_stop_or_ponderhit(); - void init_ss_array(SearchStack ss[]); - void print_pv_info(const Position& pos, SearchStack ss[], Value alpha, Value beta, Value value); + void init_ss_array(SearchStack* ss, int size); + void print_pv_info(const Position& pos, SearchStack* ss, Value alpha, Value beta, Value value); #if !defined(_MSC_VER) void *init_thread(void *threadID); @@ -330,6 +337,51 @@ void exit_threads() { TM.exit_threads(); } int64_t nodes_searched() { return TM.nodes_searched(); } +/// init_search() is called during startup. It initializes various lookup tables + +void init_search() { + + int d; // depth (OnePly == 2) + int hd; // half depth (OnePly == 1) + int mc; // moveCount + + // Init reductions array + for (hd = 1; hd < 64; hd++) for (mc = 1; mc < 64; mc++) + { + double pvRed = log(double(hd)) * log(double(mc)) / 3.0; + double nonPVRed = log(double(hd)) * log(double(mc)) / 1.5; + ReductionMatrix[PV][hd][mc] = (int8_t) ( pvRed >= 1.0 ? floor( pvRed * int(OnePly)) : 0); + ReductionMatrix[NonPV][hd][mc] = (int8_t) (nonPVRed >= 1.0 ? floor(nonPVRed * int(OnePly)) : 0); + } + + // Init futility margins array + for (d = 0; d < 16; d++) for (mc = 0; mc < 64; mc++) + FutilityMarginsMatrix[d][mc] = 112 * int(log(double(d * d) / 2) / log(2.0) + 1) - 8 * mc + 45; + + // Init futility move count array + for (d = 0; d < 32; d++) + FutilityMoveCountArray[d] = 3 + (1 << (3 * d / 8)); +} + + +// SearchStack::init() initializes a search stack. Used at the beginning of a +// new search from the root. +void SearchStack::init(int ply) { + + pv[ply] = pv[ply + 1] = MOVE_NONE; + currentMove = threatMove = MOVE_NONE; + reduction = Depth(0); + eval = VALUE_NONE; +} + +void SearchStack::initKillers() { + + mateKiller = MOVE_NONE; + for (int i = 0; i < KILLER_MAX; i++) + killers[i] = MOVE_NONE; +} + + /// perft() is our utility to verify move generation is bug free. All the legal /// moves up to given depth are generated and counted and the sum returned. @@ -542,51 +594,6 @@ bool think(const Position& pos, bool infinite, bool ponder, int side_to_move, } -/// init_search() is called during startup. It initializes various lookup tables - -void init_search() { - - // Init our reduction lookup tables - for (int i = 1; i < 64; i++) // i == depth (OnePly = 1) - for (int j = 1; j < 64; j++) // j == moveNumber - { - double pvRed = log(double(i)) * log(double(j)) / 3.0; - double nonPVRed = log(double(i)) * log(double(j)) / 1.5; - ReductionMatrix[PV][i][j] = (int8_t) ( pvRed >= 1.0 ? floor( pvRed * int(OnePly)) : 0); - ReductionMatrix[NonPV][i][j] = (int8_t) (nonPVRed >= 1.0 ? floor(nonPVRed * int(OnePly)) : 0); - } - - // Init futility margins array - for (int i = 0; i < 16; i++) // i == depth (OnePly = 2) - for (int j = 0; j < 64; j++) // j == moveNumber - { - // FIXME: test using log instead of BSR - FutilityMarginsMatrix[i][j] = (i < 2 ? 0 : 112 * bitScanReverse32(i * i / 2)) - 8 * j + 45; - } - - // Init futility move count array - for (int i = 0; i < 32; i++) // i == depth (OnePly = 2) - FutilityMoveCountArray[i] = 3 + (1 << (3 * i / 8)); -} - - -// SearchStack::init() initializes a search stack. Used at the beginning of a -// new search from the root. -void SearchStack::init(int ply) { - - pv[ply] = pv[ply + 1] = MOVE_NONE; - currentMove = threatMove = MOVE_NONE; - reduction = Depth(0); - eval = VALUE_NONE; -} - -void SearchStack::initKillers() { - - mateKiller = MOVE_NONE; - for (int i = 0; i < KILLER_MAX; i++) - killers[i] = MOVE_NONE; -} - namespace { // id_loop() is the main iterative deepening loop. It calls root_search @@ -596,7 +603,7 @@ namespace { Value id_loop(const Position& pos, Move searchMoves[]) { - Position p(pos); + Position p(pos, pos.thread()); SearchStack ss[PLY_MAX_PLUS_2]; Move EasyMove = MOVE_NONE; Value value, alpha = -VALUE_INFINITE, beta = VALUE_INFINITE; @@ -626,8 +633,9 @@ namespace { // Initialize TT.new_search(); H.clear(); - init_ss_array(ss); + init_ss_array(ss, PLY_MAX_PLUS_2); ValueByIteration[1] = rml.get_move_score(0); + p.reset_ply(); Iteration = 1; // Is one move significantly better than others after initial scoring ? @@ -662,7 +670,7 @@ namespace { // Write PV to transposition table, in case the relevant entries have // been overwritten during the search. - TT.insert_pv(p, ss[0].pv); + TT.insert_pv(p, ss->pv); if (AbortSearch) break; // Value cannot be trusted. Break out immediately! @@ -671,7 +679,7 @@ namespace { ValueByIteration[Iteration] = value; // Drop the easy move if differs from the new best move - if (ss[0].pv[0] != EasyMove) + if (ss->pv[0] != EasyMove) EasyMove = MOVE_NONE; if (UseTimeManagement) @@ -693,7 +701,7 @@ namespace { // Stop search early if one move seems to be much better than the others int64_t nodes = TM.nodes_searched(); if ( Iteration >= 8 - && EasyMove == ss[0].pv[0] + && EasyMove == ss->pv[0] && ( ( rml.get_move_cumulative_nodes(0) > (nodes * 85) / 100 && current_search_time() > MaxSearchTime / 16) ||( rml.get_move_cumulative_nodes(0) > (nodes * 98) / 100 @@ -736,18 +744,18 @@ namespace { << " hashfull " << TT.full() << endl; // Print the best move and the ponder move to the standard output - if (ss[0].pv[0] == MOVE_NONE) + if (ss->pv[0] == MOVE_NONE) { - ss[0].pv[0] = rml.get_move(0); - ss[0].pv[1] = MOVE_NONE; + ss->pv[0] = rml.get_move(0); + ss->pv[1] = MOVE_NONE; } - assert(ss[0].pv[0] != MOVE_NONE); + assert(ss->pv[0] != MOVE_NONE); - cout << "bestmove " << ss[0].pv[0]; + cout << "bestmove " << ss->pv[0]; - if (ss[0].pv[1] != MOVE_NONE) - cout << " ponder " << ss[0].pv[1]; + if (ss->pv[1] != MOVE_NONE) + cout << " ponder " << ss->pv[1]; cout << endl; @@ -761,12 +769,12 @@ namespace { LogFile << "\nNodes: " << TM.nodes_searched() << "\nNodes/second: " << nps() - << "\nBest move: " << move_to_san(p, ss[0].pv[0]); + << "\nBest move: " << move_to_san(p, ss->pv[0]); StateInfo st; - p.do_move(ss[0].pv[0], st); + p.do_move(ss->pv[0], st); LogFile << "\nPonder move: " - << move_to_san(p, ss[0].pv[1]) // Works also with MOVE_NONE + << move_to_san(p, ss->pv[1]) // Works also with MOVE_NONE << endl; } return rml.get_move_score(0); @@ -778,7 +786,7 @@ namespace { // scheme, prints some information to the standard output and handles // the fail low/high loops. - Value root_search(Position& pos, SearchStack ss[], RootMoveList& rml, Value* alphaPtr, Value* betaPtr) { + Value root_search(Position& pos, SearchStack* ss, RootMoveList& rml, Value* alphaPtr, Value* betaPtr) { EvalInfo ei; StateInfo st; @@ -795,17 +803,15 @@ namespace { beta = *betaPtr; isCheck = pos.is_check(); - // Step 1. Initialize node and poll (omitted at root, but I can see no good reason for this, FIXME) - // Step 2. Check for aborted search (omitted at root, because we do not initialize root node) + // Step 1. Initialize node and poll (omitted at root, init_ss_array() has already initialized root node) + // Step 2. Check for aborted search (omitted at root) // Step 3. Mate distance pruning (omitted at root) // Step 4. Transposition table lookup (omitted at root) // Step 5. Evaluate the position statically // At root we do this only to get reference value for child nodes if (!isCheck) - ss[0].eval = evaluate(pos, ei, 0); - else - ss[0].eval = VALUE_NONE; // HACK because we do not initialize root node + ss->eval = evaluate(pos, ei); // Step 6. Razoring (omitted at root) // Step 7. Static null move pruning (omitted at root) @@ -834,7 +840,7 @@ namespace { // Pick the next root move, and print the move and the move number to // the standard output. - move = ss[0].currentMove = rml.get_move(i); + move = ss->currentMove = rml.get_move(i); if (current_search_time() >= 1000) cout << "info currmove " << move @@ -870,7 +876,7 @@ namespace { alpha = -VALUE_INFINITE; // Full depth PV search, done on first move or after a fail high - value = -search(pos, ss, -beta, -alpha, newDepth, 1, false, 0); + value = -search(pos, ss+1, -beta, -alpha, newDepth); } else { @@ -883,11 +889,11 @@ namespace { && !captureOrPromotion && !move_is_castle(move)) { - ss[0].reduction = reduction(depth, i - MultiPV + 2); - if (ss[0].reduction) + ss->reduction = reduction(depth, i - MultiPV + 2); + if (ss->reduction) { // Reduced depth non-pv search using alpha as upperbound - value = -search(pos, ss, -(alpha+1), -alpha, newDepth-ss[0].reduction, 1, true, 0); + value = -search(pos, ss+1, -(alpha+1), -alpha, newDepth-ss->reduction); doFullDepthSearch = (value > alpha); } } @@ -896,13 +902,13 @@ namespace { if (doFullDepthSearch) { // Full depth non-pv search using alpha as upperbound - ss[0].reduction = Depth(0); - value = -search(pos, ss, -(alpha+1), -alpha, newDepth, 1, true, 0); + ss->reduction = Depth(0); + value = -search(pos, ss+1, -(alpha+1), -alpha, newDepth); // If we are above alpha then research at same depth but as PV // to get a correct score or eventually a fail high above beta. if (value > alpha) - value = -search(pos, ss, -beta, -alpha, newDepth, 1, false, 0); + value = -search(pos, ss+1, -beta, -alpha, newDepth); } } @@ -917,8 +923,8 @@ namespace { // the score before research in case we run out of time while researching. rml.set_move_score(i, value); update_pv(ss, 0); - TT.extract_pv(pos, ss[0].pv, PLY_MAX); - rml.set_move_pv(i, ss[0].pv); + TT.extract_pv(pos, ss->pv, PLY_MAX); + rml.set_move_pv(i, ss->pv); // Print information to the standard output print_pv_info(pos, ss, alpha, beta, value); @@ -957,8 +963,8 @@ namespace { // Update PV rml.set_move_score(i, value); update_pv(ss, 0); - TT.extract_pv(pos, ss[0].pv, PLY_MAX); - rml.set_move_pv(i, ss[0].pv); + TT.extract_pv(pos, ss->pv, PLY_MAX); + rml.set_move_pv(i, ss->pv); if (MultiPV == 1) { @@ -1025,34 +1031,41 @@ namespace { // search<>() is the main search function for both PV and non-PV nodes template - Value search(Position& pos, SearchStack ss[], Value alpha, Value beta, Depth depth, - int ply, bool allowNullmove, int threadID, Move excludedMove) { + Value search(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth) { assert(alpha >= -VALUE_INFINITE && alpha <= VALUE_INFINITE); assert(beta > alpha && beta <= VALUE_INFINITE); - assert(ply >= 0 && ply < PLY_MAX); - assert(threadID >= 0 && threadID < TM.active_threads()); + assert(PvNode || alpha == beta - 1); + assert(pos.ply() > 0 && pos.ply() < PLY_MAX); + assert(pos.thread() >= 0 && pos.thread() < TM.active_threads()); Move movesSearched[256]; EvalInfo ei; StateInfo st; const TTEntry* tte; - Move ttMove, move; + Key posKey; + Move ttMove, move, excludedMove; Depth ext, newDepth; Value bestValue, value, oldAlpha; Value refinedValue, nullValue, futilityValueScaled; // Non-PV specific bool isCheck, singleEvasion, moveIsCheck, captureOrPromotion, dangerous; bool mateThreat = false; int moveCount = 0; + int threadID = pos.thread(); + int ply = pos.ply(); refinedValue = bestValue = value = -VALUE_INFINITE; oldAlpha = alpha; - if (depth < OnePly) - return qsearch(pos, ss, alpha, beta, Depth(0), ply, threadID); + // Step 1. Initialize node and poll. Polling can abort search + TM.incrementNodeCounter(threadID); + ss->init(ply); + (ss + 2)->initKillers(); - // Step 1. Initialize node and poll - // Polling can abort search. - init_node(ss, ply, threadID); + if (threadID == 0 && ++NodesSincePoll > NodesBetweenPolls) + { + NodesSincePoll = 0; + poll(); + } // Step 2. Check for aborted search and immediate draw if (AbortSearch || TM.thread_should_stop(threadID)) @@ -1071,7 +1084,8 @@ namespace { // We don't want the score of a partial search to overwrite a previous full search // TT value, so we use a different position key in case of an excluded move exists. - Key posKey = excludedMove ? pos.get_exclusion_key() : pos.get_key(); + excludedMove = ss->excludedMove; + posKey = excludedMove ? pos.get_exclusion_key() : pos.get_key(); tte = TT.retrieve(posKey); ttMove = (tte ? tte->move() : MOVE_NONE); @@ -1087,9 +1101,9 @@ namespace { if (!PvNode && tte && ok_to_use_TT(tte, depth, beta, ply)) { // Refresh tte entry to avoid aging - TT.store(posKey, tte->value(), tte->type(), tte->depth(), ttMove); + TT.store(posKey, tte->value(), tte->type(), tte->depth(), ttMove, tte->static_value(), tte->king_danger()); - ss[ply].currentMove = ttMove; // Can be MOVE_NONE + ss->currentMove = ttMove; // Can be MOVE_NONE return value_from_tt(tte->value(), ply); } @@ -1098,27 +1112,30 @@ namespace { isCheck = pos.is_check(); if (!isCheck) { - if (tte && (tte->type() & VALUE_TYPE_EVAL)) - ss[ply].eval = value_from_tt(tte->value(), ply); + if (tte && tte->static_value() != VALUE_NONE) + { + ss->eval = tte->static_value(); + ei.kingDanger[pos.side_to_move()] = tte->king_danger(); + } else - ss[ply].eval = evaluate(pos, ei, threadID); + ss->eval = evaluate(pos, ei); - refinedValue = refine_eval(tte, ss[ply].eval, ply); // Enhance accuracy with TT value if possible - update_gains(pos, ss[ply - 1].currentMove, ss[ply - 1].eval, ss[ply].eval); + refinedValue = refine_eval(tte, ss->eval, ply); // Enhance accuracy with TT value if possible + update_gains(pos, (ss-1)->currentMove, (ss-1)->eval, ss->eval); } // Step 6. Razoring (is omitted in PV nodes) if ( !PvNode - && refinedValue < beta - razor_margin(depth) - && ttMove == MOVE_NONE - && ss[ply - 1].currentMove != MOVE_NULL && depth < RazorDepth && !isCheck + && refinedValue < beta - razor_margin(depth) + && ttMove == MOVE_NONE + && (ss-1)->currentMove != MOVE_NULL && !value_is_mate(beta) && !pos.has_pawn_on_7th(pos.side_to_move())) { Value rbeta = beta - razor_margin(depth); - Value v = qsearch(pos, ss, rbeta-1, rbeta, Depth(0), ply, threadID); + Value v = qsearch(pos, ss, rbeta-1, rbeta, Depth(0)); if (v < rbeta) // Logically we should return (v + razor_margin(depth)), but // surprisingly this did slightly weaker in tests. @@ -1129,12 +1146,12 @@ namespace { // We're betting that the opponent doesn't have a move that will reduce // the score by more than futility_margin(depth) if we do a null move. if ( !PvNode - && allowNullmove + && !ss->skipNullMove && depth < RazorDepth + && refinedValue >= beta + futility_margin(depth, 0) && !isCheck && !value_is_mate(beta) - && ok_to_do_nullmove(pos) - && refinedValue >= beta + futility_margin(depth, 0)) + && pos.non_pawn_material(pos.side_to_move())) return refinedValue - futility_margin(depth, 0); // Step 8. Null move search with verification search (is omitted in PV nodes) @@ -1142,14 +1159,14 @@ namespace { // at least beta. Otherwise we do a null move if static value is not more than // NullMoveMargin under beta. if ( !PvNode - && allowNullmove + && !ss->skipNullMove && depth > OnePly + && refinedValue >= beta - (depth >= 4 * OnePly ? NullMoveMargin : 0) && !isCheck && !value_is_mate(beta) - && ok_to_do_nullmove(pos) - && refinedValue >= beta - (depth >= 4 * OnePly ? NullMoveMargin : 0)) + && pos.non_pawn_material(pos.side_to_move())) { - ss[ply].currentMove = MOVE_NULL; + ss->currentMove = MOVE_NULL; // Null move dynamic reduction based on depth int R = 3 + (depth >= 5 * OnePly ? depth / 8 : 0); @@ -1159,9 +1176,11 @@ namespace { R++; pos.do_null_move(st); + (ss+1)->skipNullMove = true; - nullValue = -search(pos, ss, -beta, -alpha, depth-R*OnePly, ply+1, false, threadID); - + nullValue = depth-R*OnePly < OnePly ? -qsearch(pos, ss+1, -beta, -alpha, Depth(0)) + : - search(pos, ss+1, -beta, -alpha, depth-R*OnePly); + (ss+1)->skipNullMove = false; pos.undo_null_move(); if (nullValue >= beta) @@ -1170,14 +1189,19 @@ namespace { if (nullValue >= value_mate_in(PLY_MAX)) nullValue = beta; + // Do zugzwang verification search at high depths if (depth < 6 * OnePly) return nullValue; - // Do zugzwang verification search - Value v = search(pos, ss, alpha, beta, depth-5*OnePly, ply, false, threadID); + ss->skipNullMove = true; + Value v = search(pos, ss, alpha, beta, depth-5*OnePly); + ss->skipNullMove = false; + if (v >= beta) return nullValue; - } else { + } + else + { // The null move failed low, which means that we may be faced with // some kind of threat. If the previous move was reduced, check if // the move that refuted the null move was somehow connected to the @@ -1187,22 +1211,26 @@ namespace { if (nullValue == value_mated_in(ply + 2)) mateThreat = true; - ss[ply].threatMove = ss[ply + 1].currentMove; + ss->threatMove = (ss+1)->currentMove; if ( depth < ThreatDepth - && ss[ply - 1].reduction - && connected_moves(pos, ss[ply - 1].currentMove, ss[ply].threatMove)) + && (ss-1)->reduction + && connected_moves(pos, (ss-1)->currentMove, ss->threatMove)) return beta - 1; } } // Step 9. Internal iterative deepening - if ( depth >= IIDDepth[PvNode] - && ttMove == MOVE_NONE - && (PvNode || (!isCheck && ss[ply].eval >= beta - IIDMargin))) + if ( depth >= IIDDepth[PvNode] + && (ttMove == MOVE_NONE || (PvNode && tte->depth() <= depth - 4 * OnePly)) + && (PvNode || (!isCheck && ss->eval >= beta - IIDMargin))) { Depth d = (PvNode ? depth - 2 * OnePly : depth / 2); - search(pos, ss, alpha, beta, d, ply, false, threadID); - ttMove = ss[ply].pv[ply]; + + ss->skipNullMove = true; + search(pos, ss, alpha, beta, d); + ss->skipNullMove = false; + + ttMove = ss->pv[ply]; tte = TT.retrieve(posKey); } @@ -1211,8 +1239,13 @@ namespace { mateThreat = pos.has_mate_threat(opposite_color(pos.side_to_move())); // Initialize a MovePicker object for the current position - MovePicker mp = MovePicker(pos, ttMove, depth, H, &ss[ply], (PvNode ? -VALUE_INFINITE : beta)); + MovePicker mp = MovePicker(pos, ttMove, depth, H, ss, (PvNode ? -VALUE_INFINITE : beta)); CheckInfo ci(pos); + bool singularExtensionNode = depth >= SingularExtensionDepth[PvNode] + && tte && tte->move() + && !excludedMove // Do not allow recursive singular extension search + && is_lower_bound(tte->type()) + && tte->depth() >= depth - 3 * OnePly; // Step 10. Loop through moves // Loop through all legal moves until no moves remain or a beta cutoff occurs @@ -1235,21 +1268,22 @@ namespace { // Singular extension search. We extend the TT move if its value is much better than // its siblings. To verify this we do a reduced search on all the other moves but the // ttMove, if result is lower then ttValue minus a margin then we extend ttMove. - if ( depth >= SingularExtensionDepth[PvNode] - && tte + if ( singularExtensionNode && move == tte->move() - && !excludedMove // Do not allow recursive singular extension search - && ext < OnePly - && is_lower_bound(tte->type()) - && tte->depth() >= depth - 3 * OnePly) + && ext < OnePly) { Value ttValue = value_from_tt(tte->value(), ply); if (abs(ttValue) < VALUE_KNOWN_WIN) { - Value excValue = search(pos, ss, ttValue - SingularExtensionMargin - 1, ttValue - SingularExtensionMargin, depth / 2, ply, false, threadID, move); - - if (excValue < ttValue - SingularExtensionMargin) + Value b = ttValue - SingularExtensionMargin; + ss->excludedMove = move; + ss->skipNullMove = true; + Value v = search(pos, ss, b - 1, b, depth / 2); + ss->skipNullMove = false; + ss->excludedMove = MOVE_NONE; + + if (v < ttValue - SingularExtensionMargin) ext = OnePly; } } @@ -1257,25 +1291,27 @@ namespace { newDepth = depth - OnePly + ext; // Update current move (this must be done after singular extension search) - movesSearched[moveCount++] = ss[ply].currentMove = move; + movesSearched[moveCount++] = ss->currentMove = move; // Step 12. Futility pruning (is omitted in PV nodes) if ( !PvNode + && !captureOrPromotion && !isCheck && !dangerous - && !captureOrPromotion - && !move_is_castle(move) - && move != ttMove) + && move != ttMove + && !move_is_castle(move)) { // Move count based pruning if ( moveCount >= futility_move_count(depth) - && ok_to_prune(pos, move, ss[ply].threatMove) + && !(ss->threatMove && connected_threat(pos, move, ss->threatMove)) && bestValue > value_mated_in(PLY_MAX)) continue; // Value based pruning - Depth predictedDepth = newDepth - reduction(depth, moveCount); // We illogically ignore reduction condition depth >= 3*OnePly - futilityValueScaled = ss[ply].eval + futility_margin(predictedDepth, moveCount) + // We illogically ignore reduction condition depth >= 3*OnePly for predicted depth, + // but fixing this made program slightly weaker. + Depth predictedDepth = newDepth - reduction(depth, moveCount); + futilityValueScaled = ss->eval + futility_margin(predictedDepth, moveCount) + H.gain(pos.piece_on(move_from(move)), move_to(move)); if (futilityValueScaled < beta) @@ -1292,37 +1328,57 @@ namespace { // Step extra. pv search (only in PV nodes) // The first move in list is the expected PV if (PvNode && moveCount == 1) - value = -search(pos, ss, -beta, -alpha, newDepth, ply+1, false, threadID); + value = newDepth < OnePly ? -qsearch(pos, ss+1, -beta, -alpha, Depth(0)) + : - search(pos, ss+1, -beta, -alpha, newDepth); else { - // Step 14. Reduced search - // if the move fails high will be re-searched at full depth. - bool doFullDepthSearch = true; - - if ( depth >= 3 * OnePly - && !dangerous - && !captureOrPromotion - && !move_is_castle(move) - && !move_is_killer(move, ss[ply])) - { - ss[ply].reduction = reduction(depth, moveCount); - if (ss[ply].reduction) - { - value = -search(pos, ss, -(alpha+1), -alpha, newDepth-ss[ply].reduction, ply+1, true, threadID); - doFullDepthSearch = (value > alpha); - } - } - - // Step 15. Full depth search - if (doFullDepthSearch) - { - ss[ply].reduction = Depth(0); - value = -search(pos, ss, -(alpha+1), -alpha, newDepth, ply+1, true, threadID); + // Step 14. Reduced depth search + // If the move fails high will be re-searched at full depth. + bool doFullDepthSearch = true; + + if ( depth >= 3 * OnePly + && !captureOrPromotion + && !dangerous + && !move_is_castle(move) + && !move_is_killer(move, ss)) + { + ss->reduction = reduction(depth, moveCount); + if (ss->reduction) + { + Depth d = newDepth - ss->reduction; + value = d < OnePly ? -qsearch(pos, ss+1, -(alpha+1), -alpha, Depth(0)) + : - search(pos, ss+1, -(alpha+1), -alpha, d); + + doFullDepthSearch = (value > alpha); + } + + // The move failed high, but if reduction is very big we could + // face a false positive, retry with a less aggressive reduction, + // if the move fails high again then go with full depth search. + if (doFullDepthSearch && ss->reduction > 2 * OnePly) + { + assert(newDepth - OnePly >= OnePly); + + ss->reduction = OnePly; + value = -search(pos, ss+1, -(alpha+1), -alpha, newDepth-ss->reduction); + doFullDepthSearch = (value > alpha); + } + ss->reduction = Depth(0); // Restore original reduction + } - // Step extra. pv search (only in PV nodes) - if (PvNode && value > alpha && value < beta) - value = -search(pos, ss, -beta, -alpha, newDepth, ply+1, false, threadID); - } + // Step 15. Full depth search + if (doFullDepthSearch) + { + value = newDepth < OnePly ? -qsearch(pos, ss+1, -(alpha+1), -alpha, Depth(0)) + : - search(pos, ss+1, -(alpha+1), -alpha, newDepth); + + // Step extra. pv search (only in PV nodes) + // Search only for possible new PV nodes, if instead value >= beta then + // parent node fails low with value <= alpha and tries another move. + if (PvNode && value > alpha && value < beta) + value = newDepth < OnePly ? -qsearch(pos, ss+1, -beta, -alpha, Depth(0)) + : - search(pos, ss+1, -beta, -alpha, newDepth); + } } // Step 16. Undo move @@ -1336,24 +1392,26 @@ namespace { bestValue = value; if (value > alpha) { - alpha = value; + if (PvNode && value < beta) // This guarantees that always: alpha < beta + alpha = value; + update_pv(ss, ply); + if (value == value_mate_in(ply + 1)) - ss[ply].mateKiller = move; + ss->mateKiller = move; } } // Step 18. Check for split - if ( TM.active_threads() > 1 + if ( depth >= MinimumSplitDepth + && TM.active_threads() > 1 && bestValue < beta - && depth >= MinimumSplitDepth - && Iteration <= 99 && TM.available_thread_exists(threadID) && !AbortSearch && !TM.thread_should_stop(threadID) - && TM.split(pos, ss, ply, &alpha, beta, &bestValue, - depth, mateThreat, &moveCount, &mp, threadID, PvNode)) - break; + && Iteration <= 99) + TM.split(pos, ss, &alpha, beta, &bestValue, depth, + mateThreat, &moveCount, &mp, PvNode); } // Step 19. Check for mate and stalemate @@ -1370,21 +1428,21 @@ namespace { return bestValue; if (bestValue <= oldAlpha) - TT.store(posKey, value_to_tt(bestValue, ply), VALUE_TYPE_UPPER, depth, MOVE_NONE); + TT.store(posKey, value_to_tt(bestValue, ply), VALUE_TYPE_UPPER, depth, MOVE_NONE, ss->eval, ei.kingDanger[pos.side_to_move()]); else if (bestValue >= beta) { TM.incrementBetaCounter(pos.side_to_move(), depth, threadID); - move = ss[ply].pv[ply]; - TT.store(posKey, value_to_tt(bestValue, ply), VALUE_TYPE_LOWER, depth, move); + move = ss->pv[ply]; + TT.store(posKey, value_to_tt(bestValue, ply), VALUE_TYPE_LOWER, depth, move, ss->eval, ei.kingDanger[pos.side_to_move()]); if (!pos.move_is_capture_or_promotion(move)) { update_history(pos, move, depth, movesSearched, moveCount); - update_killers(move, ss[ply]); + update_killers(move, ss); } } else - TT.store(posKey, value_to_tt(bestValue, ply), VALUE_TYPE_EXACT, depth, ss[ply].pv[ply]); + TT.store(posKey, value_to_tt(bestValue, ply), VALUE_TYPE_EXACT, depth, ss->pv[ply], ss->eval, ei.kingDanger[pos.side_to_move()]); assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE); @@ -1396,33 +1454,31 @@ namespace { // search function when the remaining depth is zero (or, to be more precise, // less than OnePly). - Value qsearch(Position& pos, SearchStack ss[], Value alpha, Value beta, - Depth depth, int ply, int threadID) { + template + Value qsearch(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth) { assert(alpha >= -VALUE_INFINITE && alpha <= VALUE_INFINITE); assert(beta >= -VALUE_INFINITE && beta <= VALUE_INFINITE); + assert(PvNode || alpha == beta - 1); assert(depth <= 0); - assert(ply >= 0 && ply < PLY_MAX); - assert(threadID >= 0 && threadID < TM.active_threads()); + assert(pos.ply() > 0 && pos.ply() < PLY_MAX); + assert(pos.thread() >= 0 && pos.thread() < TM.active_threads()); EvalInfo ei; StateInfo st; Move ttMove, move; - Value staticValue, bestValue, value, futilityBase, futilityValue; + Value staticValue, bestValue, value, futilityBase; bool isCheck, enoughMaterial, moveIsCheck, evasionPrunable; const TTEntry* tte = NULL; int moveCount = 0; - bool pvNode = (beta - alpha != 1); + int ply = pos.ply(); Value oldAlpha = alpha; + Value futilityValue = VALUE_INFINITE; - // Initialize, and make an early exit in case of an aborted search, - // an instant draw, maximum ply reached, etc. - init_node(ss, ply, threadID); - - // After init_node() that calls poll() - if (AbortSearch || TM.thread_should_stop(threadID)) - return Value(0); + TM.incrementNodeCounter(pos.thread()); + ss->init(ply); + // Check for an instant draw or maximum ply reached if (pos.is_draw() || ply >= PLY_MAX - 1) return VALUE_DRAW; @@ -1431,11 +1487,9 @@ namespace { tte = TT.retrieve(pos.get_key()); ttMove = (tte ? tte->move() : MOVE_NONE); - if (!pvNode && tte && ok_to_use_TT(tte, depth, beta, ply)) + if (!PvNode && tte && ok_to_use_TT(tte, depth, beta, ply)) { - assert(tte->type() != VALUE_TYPE_EVAL); - - ss[ply].currentMove = ttMove; // Can be MOVE_NONE + ss->currentMove = ttMove; // Can be MOVE_NONE return value_from_tt(tte->value(), ply); } @@ -1444,15 +1498,18 @@ namespace { // Evaluate the position statically if (isCheck) staticValue = -VALUE_INFINITE; - else if (tte && (tte->type() & VALUE_TYPE_EVAL)) - staticValue = value_from_tt(tte->value(), ply); + else if (tte && tte->static_value() != VALUE_NONE) + { + staticValue = tte->static_value(); + ei.kingDanger[pos.side_to_move()] = tte->king_danger(); + } else - staticValue = evaluate(pos, ei, threadID); + staticValue = evaluate(pos, ei); if (!isCheck) { - ss[ply].eval = staticValue; - update_gains(pos, ss[ply - 1].currentMove, ss[ply - 1].eval, ss[ply].eval); + ss->eval = staticValue; + update_gains(pos, (ss-1)->currentMove, (ss-1)->eval, ss->eval); } // Initialize "stand pat score", and return it immediately if it is @@ -1462,13 +1519,13 @@ namespace { if (bestValue >= beta) { // Store the score to avoid a future costly evaluation() call - if (!isCheck && !tte && ei.kingDanger[pos.side_to_move()] == 0) - TT.store(pos.get_key(), value_to_tt(bestValue, ply), VALUE_TYPE_EV_LO, Depth(-127*OnePly), MOVE_NONE); + if (!isCheck && !tte) + TT.store(pos.get_key(), value_to_tt(bestValue, ply), VALUE_TYPE_LOWER, Depth(-127*OnePly), MOVE_NONE, ss->eval, ei.kingDanger[pos.side_to_move()]); return bestValue; } - if (bestValue > alpha) + if (PvNode && bestValue > alpha) alpha = bestValue; // If we are near beta then try to get a cutoff pushing checks a bit further @@ -1493,17 +1550,22 @@ namespace { // Update current move moveCount++; - ss[ply].currentMove = move; + ss->currentMove = move; // Futility pruning - if ( enoughMaterial + if ( !PvNode + && enoughMaterial && !isCheck - && !pvNode && !moveIsCheck && move != ttMove && !move_is_promotion(move) && !pos.move_is_passed_pawn_push(move)) { + // Can only decrease from previous move because of + // MVV ordering so we don't need to recheck. + if (futilityValue < alpha) + continue; + futilityValue = futilityBase + pos.endgame_value_of_piece_on(move_to(move)) + (move_is_ep(move) ? PawnValueEndgame : Value(0)); @@ -1524,8 +1586,8 @@ namespace { && !pos.can_castle(pos.side_to_move()); // Don't search moves with negative SEE values - if ( (!isCheck || evasionPrunable) - && !pvNode + if ( !PvNode + && (!isCheck || evasionPrunable) && move != ttMove && !move_is_promotion(move) && pos.see_sign(move) < 0) @@ -1533,7 +1595,7 @@ namespace { // Make and search the move pos.do_move(move, st, ci, moveIsCheck); - value = -qsearch(pos, ss, -beta, -alpha, depth-OnePly, ply+1, threadID); + value = -qsearch(pos, ss+1, -beta, -alpha, depth-OnePly); pos.undo_move(move); assert(value > -VALUE_INFINITE && value < VALUE_INFINITE); @@ -1561,20 +1623,19 @@ namespace { { // If bestValue isn't changed it means it is still the static evaluation // of the node, so keep this info to avoid a future evaluation() call. - ValueType type = (bestValue == staticValue && !ei.kingDanger[pos.side_to_move()] ? VALUE_TYPE_EV_UP : VALUE_TYPE_UPPER); - TT.store(pos.get_key(), value_to_tt(bestValue, ply), type, d, MOVE_NONE); + TT.store(pos.get_key(), value_to_tt(bestValue, ply), VALUE_TYPE_UPPER, d, MOVE_NONE, ss->eval, ei.kingDanger[pos.side_to_move()]); } else if (bestValue >= beta) { - move = ss[ply].pv[ply]; - TT.store(pos.get_key(), value_to_tt(bestValue, ply), VALUE_TYPE_LOWER, d, move); + move = ss->pv[ply]; + TT.store(pos.get_key(), value_to_tt(bestValue, ply), VALUE_TYPE_LOWER, d, move, ss->eval, ei.kingDanger[pos.side_to_move()]); // Update killers only for good checking moves if (!pos.move_is_capture_or_promotion(move)) - update_killers(move, ss[ply]); + update_killers(move, ss); } else - TT.store(pos.get_key(), value_to_tt(bestValue, ply), VALUE_TYPE_EXACT, d, ss[ply].pv[ply]); + TT.store(pos.get_key(), value_to_tt(bestValue, ply), VALUE_TYPE_EXACT, d, ss->pv[ply], ss->eval, ei.kingDanger[pos.side_to_move()]); assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE); @@ -1590,6 +1651,7 @@ namespace { // also don't need to store anything to the hash table here: This is taken // care of after we return from the split point. + template void sp_search(SplitPoint* sp, int threadID) { assert(threadID >= 0 && threadID < TM.active_threads()); @@ -1598,14 +1660,16 @@ namespace { StateInfo st; Move move; Depth ext, newDepth; - Value value, futilityValueScaled; + Value value; + Value futilityValueScaled; // NonPV specific bool isCheck, moveIsCheck, captureOrPromotion, dangerous; int moveCount; value = -VALUE_INFINITE; - Position pos(*sp->pos); + Position pos(*sp->pos, threadID); CheckInfo ci(pos); - SearchStack* ss = sp->sstack[threadID]; + int ply = pos.ply(); + SearchStack* ss = sp->sstack[threadID] + 1; isCheck = pos.is_check(); // Step 10. Loop through moves @@ -1613,10 +1677,10 @@ namespace { lock_grab(&(sp->lock)); while ( sp->bestValue < sp->beta - && !TM.thread_should_stop(threadID) - && (move = sp->mp->get_next_move()) != MOVE_NONE) + && (move = sp->mp->get_next_move()) != MOVE_NONE + && !TM.thread_should_stop(threadID)) { - moveCount = ++sp->moves; + moveCount = ++sp->moveCount; lock_release(&(sp->lock)); assert(move_is_ok(move)); @@ -1625,21 +1689,22 @@ namespace { captureOrPromotion = pos.move_is_capture_or_promotion(move); // Step 11. Decide the new search depth - ext = extension(pos, move, captureOrPromotion, moveIsCheck, false, sp->mateThreat, &dangerous); + ext = extension(pos, move, captureOrPromotion, moveIsCheck, false, sp->mateThreat, &dangerous); newDepth = sp->depth - OnePly + ext; // Update current move - ss[sp->ply].currentMove = move; + ss->currentMove = move; - // Step 12. Futility pruning - if ( !isCheck - && !dangerous + // Step 12. Futility pruning (is omitted in PV nodes) + if ( !PvNode && !captureOrPromotion + && !isCheck + && !dangerous && !move_is_castle(move)) { // Move count based pruning if ( moveCount >= futility_move_count(sp->depth) - && ok_to_prune(pos, move, ss[sp->ply].threatMove) + && !(ss->threatMove && connected_threat(pos, move, ss->threatMove)) && sp->bestValue > value_mated_in(PLY_MAX)) { lock_grab(&(sp->lock)); @@ -1648,7 +1713,7 @@ namespace { // Value based pruning Depth predictedDepth = newDepth - reduction(sp->depth, moveCount); - futilityValueScaled = ss[sp->ply].eval + futility_margin(predictedDepth, moveCount) + futilityValueScaled = ss->eval + futility_margin(predictedDepth, moveCount) + H.gain(pos.piece_on(move_from(move)), move_to(move)); if (futilityValueScaled < sp->beta) @@ -1665,143 +1730,52 @@ namespace { pos.do_move(move, st, ci, moveIsCheck); // Step 14. Reduced search - // if the move fails high will be re-searched at full depth. + // If the move fails high will be re-searched at full depth. bool doFullDepthSearch = true; - if ( !dangerous - && !captureOrPromotion + if ( !captureOrPromotion + && !dangerous && !move_is_castle(move) - && !move_is_killer(move, ss[sp->ply])) + && !move_is_killer(move, ss)) { - ss[sp->ply].reduction = reduction(sp->depth, moveCount); - if (ss[sp->ply].reduction) + ss->reduction = reduction(sp->depth, moveCount); + if (ss->reduction) { - value = -search(pos, ss, -(sp->alpha+1), -(sp->alpha), newDepth-ss[sp->ply].reduction, sp->ply+1, true, threadID); - doFullDepthSearch = (value >= sp->beta && !TM.thread_should_stop(threadID)); + Value localAlpha = sp->alpha; + Depth d = newDepth - ss->reduction; + value = d < OnePly ? -qsearch(pos, ss+1, -(localAlpha+1), -localAlpha, Depth(0)) + : - search(pos, ss+1, -(localAlpha+1), -localAlpha, d); + doFullDepthSearch = (value > localAlpha); } - } - - // Step 15. Full depth search - if (doFullDepthSearch) - { - ss[sp->ply].reduction = Depth(0); - value = -search(pos, ss, -(sp->alpha+1), -(sp->alpha), newDepth, sp->ply+1, true, threadID); - } - - // Step 16. Undo move - pos.undo_move(move); - - assert(value > -VALUE_INFINITE && value < VALUE_INFINITE); - - // Step 17. Check for new best move - lock_grab(&(sp->lock)); - if (value > sp->bestValue && !TM.thread_should_stop(threadID)) - { - sp->bestValue = value; - if (sp->bestValue >= sp->beta) + // The move failed high, but if reduction is very big we could + // face a false positive, retry with a less aggressive reduction, + // if the move fails high again then go with full depth search. + if (doFullDepthSearch && ss->reduction > 2 * OnePly) { - sp->stopRequest = true; - sp_update_pv(sp->parentSstack, ss, sp->ply); - } - } - } - - /* Here we have the lock still grabbed */ - - sp->slaves[threadID] = 0; - sp->cpus--; - - lock_release(&(sp->lock)); - } - - - // sp_search_pv() is used to search from a PV split point. This function - // is called by each thread working at the split point. It is similar to - // the normal search_pv() function, but simpler. Because we have already - // probed the hash table and searched the first move before splitting, we - // don't have to repeat all this work in sp_search_pv(). We also don't - // need to store anything to the hash table here: This is taken care of - // after we return from the split point. - - void sp_search_pv(SplitPoint* sp, int threadID) { - - assert(threadID >= 0 && threadID < TM.active_threads()); - assert(TM.active_threads() > 1); - - StateInfo st; - Move move; - Depth ext, newDepth; - Value value; - bool moveIsCheck, captureOrPromotion, dangerous; - int moveCount; - value = -VALUE_INFINITE; - - Position pos(*sp->pos); - CheckInfo ci(pos); - SearchStack* ss = sp->sstack[threadID]; - - // Step 10. Loop through moves - // Loop through all legal moves until no moves remain or a beta cutoff occurs - lock_grab(&(sp->lock)); - - while ( sp->alpha < sp->beta - && !TM.thread_should_stop(threadID) - && (move = sp->mp->get_next_move()) != MOVE_NONE) - { - moveCount = ++sp->moves; - lock_release(&(sp->lock)); + assert(newDepth - OnePly >= OnePly); - assert(move_is_ok(move)); - - moveIsCheck = pos.move_is_check(move, ci); - captureOrPromotion = pos.move_is_capture_or_promotion(move); - - // Step 11. Decide the new search depth - ext = extension(pos, move, captureOrPromotion, moveIsCheck, false, sp->mateThreat, &dangerous); - newDepth = sp->depth - OnePly + ext; - - // Update current move - ss[sp->ply].currentMove = move; - - // Step 12. Futility pruning (is omitted in PV nodes) - - // Step 13. Make the move - pos.do_move(move, st, ci, moveIsCheck); - - // Step 14. Reduced search - // if the move fails high will be re-searched at full depth. - bool doFullDepthSearch = true; - - if ( !dangerous - && !captureOrPromotion - && !move_is_castle(move) - && !move_is_killer(move, ss[sp->ply])) - { - ss[sp->ply].reduction = reduction(sp->depth, moveCount); - if (ss[sp->ply].reduction) - { + ss->reduction = OnePly; Value localAlpha = sp->alpha; - value = -search(pos, ss, -(localAlpha+1), -localAlpha, newDepth-ss[sp->ply].reduction, sp->ply+1, true, threadID); - doFullDepthSearch = (value > localAlpha && !TM.thread_should_stop(threadID)); + value = -search(pos, ss+1, -(localAlpha+1), -localAlpha, newDepth-ss->reduction); + doFullDepthSearch = (value > localAlpha); } + ss->reduction = Depth(0); // Restore original reduction } // Step 15. Full depth search if (doFullDepthSearch) { Value localAlpha = sp->alpha; - ss[sp->ply].reduction = Depth(0); - value = -search(pos, ss, -(localAlpha+1), -localAlpha, newDepth, sp->ply+1, true, threadID); - - if (value > localAlpha && value < sp->beta && !TM.thread_should_stop(threadID)) - { - // If another thread has failed high then sp->alpha has been increased - // to be higher or equal then beta, if so, avoid to start a PV search. - localAlpha = sp->alpha; - if (localAlpha < sp->beta) - value = -search(pos, ss, -sp->beta, -localAlpha, newDepth, sp->ply+1, false, threadID); - } + value = newDepth < OnePly ? -qsearch(pos, ss+1, -(localAlpha+1), -localAlpha, Depth(0)) + : - search(pos, ss+1, -(localAlpha+1), -localAlpha, newDepth); + + // Step extra. pv search (only in PV nodes) + // Search only for possible new PV nodes, if instead value >= beta then + // parent node fails low with value <= alpha and tries another move. + if (PvNode && value > localAlpha && value < sp->beta) + value = newDepth < OnePly ? -qsearch(pos, ss+1, -sp->beta, -sp->alpha, Depth(0)) + : - search(pos, ss+1, -sp->beta, -sp->alpha, newDepth); } // Step 16. Undo move @@ -1815,17 +1789,16 @@ namespace { if (value > sp->bestValue && !TM.thread_should_stop(threadID)) { sp->bestValue = value; - if (value > sp->alpha) + + if (sp->bestValue > sp->alpha) { - // Ask threads to stop before to modify sp->alpha - if (value >= sp->beta) + if (!PvNode || value >= sp->beta) sp->stopRequest = true; - sp->alpha = value; + if (PvNode && value < sp->beta) // This guarantees that always: sp->alpha < sp->beta + sp->alpha = value; - sp_update_pv(sp->parentSstack, ss, sp->ply); - if (value == value_mate_in(sp->ply + 1)) - ss[sp->ply].mateKiller = move; + sp_update_pv(sp->parentSstack, ss, ply); } } } @@ -1833,55 +1806,26 @@ namespace { /* Here we have the lock still grabbed */ sp->slaves[threadID] = 0; - sp->cpus--; lock_release(&(sp->lock)); } - - // init_node() is called at the beginning of all the search functions - // (search() qsearch(), and so on) and initializes the - // search stack object corresponding to the current node. Once every - // NodesBetweenPolls nodes, init_node() also calls poll(), which polls - // for user input and checks whether it is time to stop the search. - - void init_node(SearchStack ss[], int ply, int threadID) { - - assert(ply >= 0 && ply < PLY_MAX); - assert(threadID >= 0 && threadID < TM.active_threads()); - - TM.incrementNodeCounter(threadID); - - if (threadID == 0) - { - NodesSincePoll++; - if (NodesSincePoll >= NodesBetweenPolls) - { - poll(); - NodesSincePoll = 0; - } - } - ss[ply].init(ply); - ss[ply + 2].initKillers(); - } - - // update_pv() is called whenever a search returns a value > alpha. // It updates the PV in the SearchStack object corresponding to the // current node. - void update_pv(SearchStack ss[], int ply) { + void update_pv(SearchStack* ss, int ply) { assert(ply >= 0 && ply < PLY_MAX); int p; - ss[ply].pv[ply] = ss[ply].currentMove; + ss->pv[ply] = ss->currentMove; - for (p = ply + 1; ss[ply + 1].pv[p] != MOVE_NONE; p++) - ss[ply].pv[p] = ss[ply + 1].pv[p]; + for (p = ply + 1; (ss+1)->pv[p] != MOVE_NONE; p++) + ss->pv[p] = (ss+1)->pv[p]; - ss[ply].pv[p] = MOVE_NONE; + ss->pv[p] = MOVE_NONE; } @@ -1889,18 +1833,18 @@ namespace { // difference between the two functions is that sp_update_pv also updates // the PV at the parent node. - void sp_update_pv(SearchStack* pss, SearchStack ss[], int ply) { + void sp_update_pv(SearchStack* pss, SearchStack* ss, int ply) { assert(ply >= 0 && ply < PLY_MAX); int p; - ss[ply].pv[ply] = pss[ply].pv[ply] = ss[ply].currentMove; + ss->pv[ply] = pss->pv[ply] = ss->currentMove; - for (p = ply + 1; ss[ply + 1].pv[p] != MOVE_NONE; p++) - ss[ply].pv[p] = pss[ply].pv[p] = ss[ply + 1].pv[p]; + for (p = ply + 1; (ss+1)->pv[p] != MOVE_NONE; p++) + ss->pv[p] = pss->pv[p] = (ss+1)->pv[p]; - ss[ply].pv[p] = pss[ply].pv[p] = MOVE_NONE; + ss->pv[p] = pss->pv[p] = MOVE_NONE; } @@ -1975,9 +1919,9 @@ namespace { // move_is_killer() checks if the given move is among the // killer moves of that ply. - bool move_is_killer(Move m, const SearchStack& ss) { + bool move_is_killer(Move m, SearchStack* ss) { - const Move* k = ss.killers; + const Move* k = ss->killers; for (int i = 0; i < KILLER_MAX; i++, k++) if (*k == m) return true; @@ -2052,38 +1996,19 @@ namespace { } - // ok_to_do_nullmove() looks at the current position and decides whether - // doing a 'null move' should be allowed. In order to avoid zugzwang - // problems, null moves are not allowed when the side to move has very - // little material left. Currently, the test is a bit too simple: Null - // moves are avoided only when the side to move has only pawns left. - // It's probably a good idea to avoid null moves in at least some more - // complicated endgames, e.g. KQ vs KR. FIXME - - bool ok_to_do_nullmove(const Position& pos) { - - return pos.non_pawn_material(pos.side_to_move()) != Value(0); - } - - - // ok_to_prune() tests whether it is safe to forward prune a move. Only - // non-tactical moves late in the move list close to the leaves are - // candidates for pruning. + // connected_threat() tests whether it is safe to forward prune a move or if + // is somehow coonected to the threat move returned by null search. - bool ok_to_prune(const Position& pos, Move m, Move threat) { + bool connected_threat(const Position& pos, Move m, Move threat) { assert(move_is_ok(m)); - assert(threat == MOVE_NONE || move_is_ok(threat)); + assert(threat && move_is_ok(threat)); assert(!pos.move_is_check(m)); assert(!pos.move_is_capture_or_promotion(m)); assert(!pos.move_is_passed_pawn_push(m)); Square mfrom, mto, tfrom, tto; - // Prune if there isn't any threat move - if (threat == MOVE_NONE) - return true; - mfrom = move_from(m); mto = move_to(m); tfrom = move_from(threat); @@ -2091,7 +2016,7 @@ namespace { // Case 1: Don't prune moves which move the threatened piece if (mfrom == tto) - return false; + return true; // Case 2: If the threatened piece has value less than or equal to the // value of the threatening piece, don't prune move which defend it. @@ -2099,16 +2024,16 @@ namespace { && ( pos.midgame_value_of_piece_on(tfrom) >= pos.midgame_value_of_piece_on(tto) || pos.type_of_piece_on(tfrom) == KING) && pos.move_attacks_square(m, tto)) - return false; + return true; // Case 3: If the moving piece in the threatened move is a slider, don't // prune safe moves which block its ray. if ( piece_is_slider(pos.piece_on(tfrom)) && bit_is_set(squares_between(tfrom, tto), mto) && pos.see_sign(m) >= 0) - return false; + return true; - return true; + return false; } @@ -2171,15 +2096,15 @@ namespace { // update_killers() add a good move that produced a beta-cutoff // among the killer moves of that ply. - void update_killers(Move m, SearchStack& ss) { + void update_killers(Move m, SearchStack* ss) { - if (m == ss.killers[0]) + if (m == ss->killers[0]) return; for (int i = KILLER_MAX - 1; i > 0; i--) - ss.killers[i] = ss.killers[i - 1]; + ss->killers[i] = ss->killers[i - 1]; - ss.killers[0] = m; + ss->killers[0] = m; } @@ -2312,14 +2237,21 @@ namespace { } - // init_ss_array() does a fast reset of the first entries of a SearchStack array + // init_ss_array() does a fast reset of the first entries of a SearchStack + // array and of all the excludedMove and skipNullMove entries. - void init_ss_array(SearchStack ss[]) { + void init_ss_array(SearchStack* ss, int size) { - for (int i = 0; i < 3; i++) + for (int i = 0; i < size; i++, ss++) { - ss[i].init(i); - ss[i].initKillers(); + ss->excludedMove = MOVE_NONE; + ss->skipNullMove = false; + + if (i < 3) + { + ss->init(i); + ss->initKillers(); + } } } @@ -2354,7 +2286,7 @@ namespace { // print_pv_info() prints to standard output and eventually to log file information on // the current PV line. It is called at each iteration or after a new pv is found. - void print_pv_info(const Position& pos, SearchStack ss[], Value alpha, Value beta, Value value) { + void print_pv_info(const Position& pos, SearchStack* ss, Value alpha, Value beta, Value value) { cout << "info depth " << Iteration << " score " << value_to_string(value) @@ -2365,8 +2297,8 @@ namespace { << " nps " << nps() << " pv "; - for (int j = 0; ss[0].pv[j] != MOVE_NONE && j < PLY_MAX; j++) - cout << ss[0].pv[j] << " "; + for (int j = 0; ss->pv[j] != MOVE_NONE && j < PLY_MAX; j++) + cout << ss->pv[j] << " "; cout << endl; @@ -2376,7 +2308,7 @@ namespace { : (value <= alpha ? VALUE_TYPE_UPPER : VALUE_TYPE_EXACT)); LogFile << pretty_pv(pos, current_search_time(), Iteration, - TM.nodes_searched(), value, type, ss[0].pv) << endl; + TM.nodes_searched(), value, type, ss->pv) << endl; } } @@ -2492,21 +2424,24 @@ namespace { threads[threadID].state = THREAD_SEARCHING; if (threads[threadID].splitPoint->pvNode) - sp_search_pv(threads[threadID].splitPoint, threadID); + sp_search(threads[threadID].splitPoint, threadID); else - sp_search(threads[threadID].splitPoint, threadID); + sp_search(threads[threadID].splitPoint, threadID); assert(threads[threadID].state == THREAD_SEARCHING); threads[threadID].state = THREAD_AVAILABLE; } - // If this thread is the master of a split point and all threads have + // If this thread is the master of a split point and all slaves have // finished their work at this split point, return from the idle loop. - if (sp && sp->cpus == 0) + int i = 0; + for ( ; sp && i < ActiveThreads && !sp->slaves[i]; i++) {} + + if (i == ActiveThreads) { - // Because sp->cpus is decremented under lock protection, - // be sure sp->lock has been released before to proceed. + // Because sp->slaves[] is reset under lock protection, + // be sure sp->lock has been released before to return. lock_grab(&(sp->lock)); lock_release(&(sp->lock)); @@ -2546,10 +2481,7 @@ namespace { // Initialize SplitPointStack locks for (i = 0; i < MAX_THREADS; i++) for (int j = 0; j < ACTIVE_SPLIT_POINTS_MAX; j++) - { - SplitPointStack[i][j].parent = NULL; lock_init(&(SplitPointStack[i][j].lock), NULL); - } // Will be set just before program exits to properly end the threads AllThreadsShouldExit = false; @@ -2599,7 +2531,7 @@ namespace { // Wait for thread termination for (int i = 1; i < MAX_THREADS; i++) - while (threads[i].state != THREAD_TERMINATED); + while (threads[i].state != THREAD_TERMINATED) {} // Now we can safely destroy the locks for (int i = 0; i < MAX_THREADS; i++) @@ -2681,34 +2613,29 @@ namespace { // split() does the actual work of distributing the work at a node between - // several threads at PV nodes. If it does not succeed in splitting the + // several available threads. If it does not succeed in splitting the // node (because no idle threads are available, or because we have no unused - // split point objects), the function immediately returns false. If - // splitting is possible, a SplitPoint object is initialized with all the - // data that must be copied to the helper threads (the current position and - // search stack, alpha, beta, the search depth, etc.), and we tell our - // helper threads that they have been assigned work. This will cause them - // to instantly leave their idle loops and call sp_search_pv(). When all - // threads have returned from sp_search_pv (or, equivalently, when - // splitPoint->cpus becomes 0), split() returns true. - - bool ThreadsManager::split(const Position& p, SearchStack* sstck, int ply, - Value* alpha, const Value beta, Value* bestValue, - Depth depth, bool mateThreat, int* moves, MovePicker* mp, int master, bool pvNode) { - + // split point objects), the function immediately returns. If splitting is + // possible, a SplitPoint object is initialized with all the data that must be + // copied to the helper threads and we tell our helper threads that they have + // been assigned work. This will cause them to instantly leave their idle loops + // and call sp_search(). When all threads have returned from sp_search() then + // split() returns. + + template + void ThreadsManager::split(const Position& p, SearchStack* ss, Value* alpha, const Value beta, + Value* bestValue, Depth depth, bool mateThreat, int* moveCount, + MovePicker* mp, bool pvNode) { assert(p.is_ok()); - assert(sstck != NULL); - assert(ply >= 0 && ply < PLY_MAX); assert(*bestValue >= -VALUE_INFINITE); - assert( ( pvNode && *bestValue <= *alpha) - || (!pvNode && *bestValue < beta )); - assert(!pvNode || *alpha < beta); + assert(*bestValue <= *alpha); + assert(*alpha < beta); assert(beta <= VALUE_INFINITE); assert(depth > Depth(0)); - assert(master >= 0 && master < ActiveThreads); + assert(p.thread() >= 0 && p.thread() < ActiveThreads); assert(ActiveThreads > 1); - SplitPoint* splitPoint; + int master = p.thread(); lock_grab(&MPLock); @@ -2718,28 +2645,25 @@ namespace { || threads[master].activeSplitPoints >= ACTIVE_SPLIT_POINTS_MAX) { lock_release(&MPLock); - return false; + return; } // Pick the next available split point object from the split point stack - splitPoint = &SplitPointStack[master][threads[master].activeSplitPoints]; + SplitPoint* splitPoint = &SplitPointStack[master][threads[master].activeSplitPoints]; // Initialize the split point object splitPoint->parent = threads[master].splitPoint; splitPoint->stopRequest = false; - splitPoint->ply = ply; splitPoint->depth = depth; splitPoint->mateThreat = mateThreat; splitPoint->alpha = *alpha; splitPoint->beta = beta; splitPoint->pvNode = pvNode; splitPoint->bestValue = *bestValue; - splitPoint->master = master; splitPoint->mp = mp; - splitPoint->moves = *moves; - splitPoint->cpus = 1; + splitPoint->moveCount = *moveCount; splitPoint->pos = &p; - splitPoint->parentSstack = sstck; + splitPoint->parentSstack = ss; for (int i = 0; i < ActiveThreads; i++) splitPoint->slaves[i] = 0; @@ -2749,17 +2673,19 @@ namespace { // If we are here it means we are not available assert(threads[master].state != THREAD_AVAILABLE); + int workersCnt = 1; // At least the master is included + // Allocate available threads setting state to THREAD_BOOKED - for (int i = 0; i < ActiveThreads && splitPoint->cpus < MaxThreadsPerSplitPoint; i++) + for (int i = 0; !Fake && i < ActiveThreads && workersCnt < MaxThreadsPerSplitPoint; i++) if (thread_is_available(i, master)) { threads[i].state = THREAD_BOOKED; threads[i].splitPoint = splitPoint; splitPoint->slaves[i] = 1; - splitPoint->cpus++; + workersCnt++; } - assert(splitPoint->cpus > 1); + assert(Fake || workersCnt > 1); // We can release the lock because slave threads are already booked and master is not available lock_release(&MPLock); @@ -2769,7 +2695,7 @@ namespace { for (int i = 0; i < ActiveThreads; i++) if (i == master || splitPoint->slaves[i]) { - memcpy(splitPoint->sstack[i] + ply - 1, sstck + ply - 1, 4 * sizeof(SearchStack)); + memcpy(splitPoint->sstack[i], ss - 1, 4 * sizeof(SearchStack)); assert(i == master || threads[i].state == THREAD_BOOKED); @@ -2780,8 +2706,7 @@ namespace { // which it will instantly launch a search, because its state is // THREAD_WORKISWAITING. We send the split point as a second parameter to the // idle loop, which means that the main thread will return from the idle - // loop when all threads have finished their work at this split point - // (i.e. when splitPoint->cpus == 0). + // loop when all threads have finished their work at this split point. idle_loop(master, splitPoint); // We have returned from the idle loop, which means that all threads are @@ -2794,7 +2719,6 @@ namespace { threads[master].splitPoint = splitPoint->parent; lock_release(&MPLock); - return true; } @@ -2861,10 +2785,10 @@ namespace { continue; // Find a quick score for the move - init_ss_array(ss); + init_ss_array(ss, PLY_MAX_PLUS_2); pos.do_move(cur->move, st); moves[count].move = cur->move; - moves[count].score = -qsearch(pos, ss, -VALUE_INFINITE, VALUE_INFINITE, Depth(0), 1, 0); + moves[count].score = -qsearch(pos, ss+1, -VALUE_INFINITE, VALUE_INFINITE, Depth(0)); moves[count].pv[0] = cur->move; moves[count].pv[1] = MOVE_NONE; pos.undo_move(cur->move);