X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fsearch.cpp;h=e853b6ee73e2518490c590283ef10b926efea42c;hp=4a5473fcd96b5b92a844e82ab5fe64f05b40c2f0;hb=dee878082960be198fdb1493940b3d8a2be0bd58;hpb=5b5b496a6d7ac70e9f6a8717f00db7d6c27d3686 diff --git a/src/search.cpp b/src/search.cpp index 4a5473fc..e853b6ee 100644 --- a/src/search.cpp +++ b/src/search.cpp @@ -52,9 +52,6 @@ using std::endl; namespace { - // Maximum number of allowed moves per position - const int MOVES_MAX = 256; - // Types enum NodeType { NonPV, PV }; @@ -62,6 +59,10 @@ namespace { // Used for debugging SMP code. const bool FakeSplit = false; + // Fast lookup table of sliding pieces indexed by Piece + const bool Slidings[18] = { 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1 }; + inline bool piece_is_slider(Piece p) { return Slidings[p]; } + // 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 @@ -76,38 +77,30 @@ namespace { void init_threads(); void exit_threads(); - int active_threads() const { return ActiveThreads; } - void set_active_threads(int newActiveThreads) { ActiveThreads = newActiveThreads; } - void incrementNodeCounter(int threadID) { threads[threadID].nodes++; } + int min_split_depth() const { return minimumSplitDepth; } + int active_threads() const { return activeThreads; } + void set_active_threads(int cnt) { activeThreads = cnt; } - void resetNodeCounters(); - int64_t nodes_searched() const; + void read_uci_options(); bool available_thread_exists(int master) const; bool thread_is_available(int slave, int master) const; - bool thread_should_stop(int threadID) const; - void wake_sleeping_threads(); - void put_threads_to_sleep(); + bool cutoff_at_splitpoint(int threadID) const; + void wake_sleeping_thread(int threadID); void idle_loop(int threadID, SplitPoint* sp); template - void split(const Position& pos, SearchStack* ss, int ply, Value* alpha, const Value beta, Value* bestValue, - Depth depth, Move threatMove, bool mateThreat, int* moveCount, MovePicker* mp, bool pvNode); + void split(Position& pos, SearchStack* ss, int ply, Value* alpha, const Value beta, Value* bestValue, + Depth depth, Move threatMove, bool mateThreat, int moveCount, MovePicker* mp, bool pvNode); private: - friend void poll(); - - int ActiveThreads; - volatile bool AllThreadsShouldExit, AllThreadsShouldSleep; + Depth minimumSplitDepth; + int maxThreadsPerSplitPoint; + bool useSleepingThreads; + int activeThreads; + volatile bool allThreadsShouldExit; Thread threads[MAX_THREADS]; - - Lock MPLock, WaitLock; - -#if !defined(_MSC_VER) - pthread_cond_t WaitCond; -#else - HANDLE SitIdleEvent[MAX_THREADS]; -#endif - + Lock mpLock, sleepLock[MAX_THREADS]; + WaitCondition sleepCond[MAX_THREADS]; }; @@ -188,12 +181,6 @@ namespace { // Dynamic razoring margin based on depth inline Value razor_margin(Depth d) { return Value(0x200 + 0x10 * int(d)); } - // Step 8. Null move search with verification search - - // Null move margin. A null move search will not be done if the static - // evaluation of the position is more than NullMoveMargin below beta. - const Value NullMoveMargin = Value(0x200); - // Maximum depth for use of dynamic threat detection when null move fails low const Depth ThreatDepth = 5 * ONE_PLY; @@ -250,7 +237,10 @@ namespace { const Value EasyMoveMargin = Value(0x200); - /// Global variables + /// Namespace variables + + // Book object + Book OpeningBook; // Iteration counter int Iteration; @@ -275,9 +265,7 @@ namespace { bool UseLogFile; std::ofstream LogFile; - // Multi-threads related variables - Depth MinimumSplitDepth; - int MaxThreadsPerSplitPoint; + // Multi-threads manager object ThreadsManager ThreadsMgr; // Node counters, used only by thread[0] but try to keep in different cache @@ -290,26 +278,30 @@ namespace { /// Local functions - Value id_loop(const Position& pos, Move searchMoves[]); + Value id_loop(Position& pos, Move searchMoves[]); Value root_search(Position& pos, SearchStack* ss, Move* pv, RootMoveList& rml, Value* alphaPtr, Value* betaPtr); - template + template Value search(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth, int ply); template Value qsearch(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth, int ply); template - void sp_search(SplitPoint* sp, int threadID); + inline Value search(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth, int ply) { + + return depth < ONE_PLY ? qsearch(pos, ss, alpha, beta, DEPTH_ZERO, ply) + : search(pos, ss, alpha, beta, depth, ply); + } template Depth extension(const Position& pos, Move m, bool captureOrPromotion, bool moveIsCheck, bool singleEvasion, bool mateThreat, bool* dangerous); + bool check_is_dangerous(Position &pos, Move move, Value futilityBase, Value beta, Value *bValue); bool connected_moves(const Position& pos, Move m1, Move m2); bool value_is_mate(Value value); Value value_to_tt(Value v, int ply); Value value_from_tt(Value v, int ply); - 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); @@ -319,8 +311,8 @@ namespace { int current_search_time(); std::string value_to_uci(Value v); - int nps(); - void poll(); + int nps(const Position& pos); + void poll(const Position& pos); void ponderhit(); void wait_for_stop_or_ponderhit(); void init_ss_array(SearchStack* ss, int size); @@ -329,7 +321,7 @@ namespace { void extract_pv_from_tt(const Position& pos, Move bestMove, Move pv[]); #if !defined(_MSC_VER) - void *init_thread(void *threadID); + void* init_thread(void* threadID); #else DWORD WINAPI init_thread(LPVOID threadID); #endif @@ -346,7 +338,6 @@ namespace { void init_threads() { ThreadsMgr.init_threads(); } void exit_threads() { ThreadsMgr.exit_threads(); } -int64_t nodes_searched() { return ThreadsMgr.nodes_searched(); } /// init_search() is called during startup. It initializes various lookup tables @@ -360,7 +351,7 @@ void init_search() { // Init reductions array for (hd = 1; hd < 64; hd++) for (mc = 1; mc < 64; mc++) { - double pvRed = 0.33 + log(double(hd)) * log(double(mc)) / 4.5; + double pvRed = log(double(hd)) * log(double(mc)) / 3.0; double nonPVRed = 0.33 + log(double(hd)) * log(double(mc)) / 2.25; ReductionMatrix[PV][hd][mc] = (int8_t) ( pvRed >= 1.0 ? floor( pvRed * int(ONE_PLY)) : 0); ReductionMatrix[NonPV][hd][mc] = (int8_t) (nonPVRed >= 1.0 ? floor(nonPVRed * int(ONE_PLY)) : 0); @@ -372,7 +363,7 @@ void init_search() { // Init futility move count array for (d = 0; d < 32; d++) - FutilityMoveCountArray[d] = 3 + (1 << (3 * d / 8)); + FutilityMoveCountArray[d] = int(3.001 + 0.25 * pow(d, 2.0)); } @@ -412,13 +403,12 @@ int perft(Position& pos, Depth depth) /// search-related global variables, and calls root_search(). It returns false /// when a quit command is received during the search. -bool think(const Position& pos, bool infinite, bool ponder, int time[], int increment[], +bool think(Position& pos, bool infinite, bool ponder, int time[], int increment[], int movesToGo, int maxDepth, int maxNodes, int maxTime, Move searchMoves[]) { // Initialize global search variables StopOnPonderhit = AbortSearch = Quit = AspirationFailLow = false; NodesSincePoll = 0; - ThreadsMgr.resetNodeCounters(); SearchStartTime = get_system_time(); ExactMaxTime = maxTime; MaxDepth = maxDepth; @@ -428,12 +418,12 @@ bool think(const Position& pos, bool infinite, bool ponder, int time[], int incr UseTimeManagement = !ExactMaxTime && !MaxDepth && !MaxNodes && !InfiniteSearch; // Look for a book move, only during games, not tests - if (UseTimeManagement && get_option_value_bool("OwnBook")) + if (UseTimeManagement && Options["OwnBook"].value()) { - if (get_option_value_string("Book File") != OpeningBook.file_name()) - OpeningBook.open(get_option_value_string("Book File")); + if (Options["Book File"].value() != OpeningBook.file_name()) + OpeningBook.open(Options["Book File"].value()); - Move bookMove = OpeningBook.get_move(pos, get_option_value_bool("Best Book Move")); + Move bookMove = OpeningBook.get_move(pos, Options["Best Book Move"].value()); if (bookMove != MOVE_NONE) { if (PonderSearch) @@ -445,43 +435,40 @@ bool think(const Position& pos, bool infinite, bool ponder, int time[], int incr } // Read UCI option values - TT.set_size(get_option_value_int("Hash")); - if (button_was_pressed("Clear Hash")) + TT.set_size(Options["Hash"].value()); + if (Options["Clear Hash"].value()) + { + Options["Clear Hash"].set_value("false"); TT.clear(); + } - CheckExtension[1] = Depth(get_option_value_int("Check Extension (PV nodes)")); - CheckExtension[0] = Depth(get_option_value_int("Check Extension (non-PV nodes)")); - SingleEvasionExtension[1] = Depth(get_option_value_int("Single Evasion Extension (PV nodes)")); - SingleEvasionExtension[0] = Depth(get_option_value_int("Single Evasion Extension (non-PV nodes)")); - PawnPushTo7thExtension[1] = Depth(get_option_value_int("Pawn Push to 7th Extension (PV nodes)")); - PawnPushTo7thExtension[0] = Depth(get_option_value_int("Pawn Push to 7th Extension (non-PV nodes)")); - PassedPawnExtension[1] = Depth(get_option_value_int("Passed Pawn Extension (PV nodes)")); - PassedPawnExtension[0] = Depth(get_option_value_int("Passed Pawn Extension (non-PV nodes)")); - PawnEndgameExtension[1] = Depth(get_option_value_int("Pawn Endgame Extension (PV nodes)")); - PawnEndgameExtension[0] = Depth(get_option_value_int("Pawn Endgame Extension (non-PV nodes)")); - MateThreatExtension[1] = Depth(get_option_value_int("Mate Threat Extension (PV nodes)")); - MateThreatExtension[0] = Depth(get_option_value_int("Mate Threat Extension (non-PV nodes)")); - - MinimumSplitDepth = get_option_value_int("Minimum Split Depth") * ONE_PLY; - MaxThreadsPerSplitPoint = get_option_value_int("Maximum Number of Threads per Split Point"); - MultiPV = get_option_value_int("MultiPV"); - UseLogFile = get_option_value_bool("Use Search Log"); + CheckExtension[1] = Options["Check Extension (PV nodes)"].value(); + CheckExtension[0] = Options["Check Extension (non-PV nodes)"].value(); + SingleEvasionExtension[1] = Options["Single Evasion Extension (PV nodes)"].value(); + SingleEvasionExtension[0] = Options["Single Evasion Extension (non-PV nodes)"].value(); + PawnPushTo7thExtension[1] = Options["Pawn Push to 7th Extension (PV nodes)"].value(); + PawnPushTo7thExtension[0] = Options["Pawn Push to 7th Extension (non-PV nodes)"].value(); + PassedPawnExtension[1] = Options["Passed Pawn Extension (PV nodes)"].value(); + PassedPawnExtension[0] = Options["Passed Pawn Extension (non-PV nodes)"].value(); + PawnEndgameExtension[1] = Options["Pawn Endgame Extension (PV nodes)"].value(); + PawnEndgameExtension[0] = Options["Pawn Endgame Extension (non-PV nodes)"].value(); + MateThreatExtension[1] = Options["Mate Threat Extension (PV nodes)"].value(); + MateThreatExtension[0] = Options["Mate Threat Extension (non-PV nodes)"].value(); + MultiPV = Options["MultiPV"].value(); + UseLogFile = Options["Use Search Log"].value(); if (UseLogFile) - LogFile.open(get_option_value_string("Search Log Filename").c_str(), std::ios::out | std::ios::app); + LogFile.open(Options["Search Log Filename"].value().c_str(), std::ios::out | std::ios::app); read_weights(pos.side_to_move()); // Set the number of active threads - int newActiveThreads = get_option_value_int("Threads"); - if (newActiveThreads != ThreadsMgr.active_threads()) - { - ThreadsMgr.set_active_threads(newActiveThreads); - init_eval(ThreadsMgr.active_threads()); - } + ThreadsMgr.read_uci_options(); + init_eval(ThreadsMgr.active_threads()); - // Wake up sleeping threads - ThreadsMgr.wake_sleeping_threads(); + // Wake up needed threads + for (int i = 1; i < ThreadsMgr.active_threads(); i++) + ThreadsMgr.wake_sleeping_thread(i); // Set thinking time int myTime = time[pos.side_to_move()]; @@ -515,7 +502,8 @@ bool think(const Position& pos, bool infinite, bool ponder, int time[], int incr if (UseLogFile) LogFile.close(); - ThreadsMgr.put_threads_to_sleep(); + // This makes all the threads to go to sleep + ThreadsMgr.set_active_threads(1); return !Quit; } @@ -528,16 +516,15 @@ namespace { // been consumed, the user stops the search, or the maximum search depth is // reached. - Value id_loop(const Position& pos, Move searchMoves[]) { + Value id_loop(Position& pos, Move searchMoves[]) { - Position p(pos, pos.thread()); SearchStack ss[PLY_MAX_PLUS_2]; Move pv[PLY_MAX_PLUS_2]; Move EasyMove = MOVE_NONE; Value value, alpha = -VALUE_INFINITE, beta = VALUE_INFINITE; // Moves to search are verified, copied, scored and sorted - RootMoveList rml(p, searchMoves); + RootMoveList rml(pos, searchMoves); // Handle special case of searching on a mate/stale position if (rml.move_count() == 0) @@ -550,13 +537,13 @@ namespace { // Print RootMoveList startup scoring to the standard output, // so to output information also for iteration 1. - cout << set960(p.is_chess960()) // Is enough to set once at the beginning + cout << set960(pos.is_chess960()) // Is enough to set once at the beginning << "info depth " << 1 << "\ninfo depth " << 1 << " score " << value_to_uci(rml.move_score(0)) << " time " << current_search_time() - << " nodes " << ThreadsMgr.nodes_searched() - << " nps " << nps() + << " nodes " << pos.nodes_searched() + << " nps " << nps(pos) << " pv " << rml.move(0) << "\n"; // Initialize @@ -595,11 +582,11 @@ namespace { } // Search to the current depth, rml is updated and sorted, alpha and beta could change - value = root_search(p, ss, pv, rml, &alpha, &beta); + value = root_search(pos, ss, pv, rml, &alpha, &beta); // Write PV to transposition table, in case the relevant entries have // been overwritten during the search. - insert_pv_in_tt(p, pv); + insert_pv_in_tt(pos, pv); if (AbortSearch) break; // Value cannot be trusted. Break out immediately! @@ -628,18 +615,17 @@ namespace { stopSearch = true; // Stop search early if one move seems to be much better than the others - int64_t nodes = ThreadsMgr.nodes_searched(); if ( Iteration >= 8 && EasyMove == pv[0] - && ( ( rml.move_nodes(0) > (nodes * 85) / 100 + && ( ( rml.move_nodes(0) > (pos.nodes_searched() * 85) / 100 && current_search_time() > TimeMgr.available_time() / 16) - ||( rml.move_nodes(0) > (nodes * 98) / 100 + ||( rml.move_nodes(0) > (pos.nodes_searched() * 98) / 100 && current_search_time() > TimeMgr.available_time() / 32))) stopSearch = true; // Add some extra time if the best move has changed during the last two iterations if (Iteration > 5 && Iteration <= 50) - TimeMgr.pv_unstability(BestMoveChangesByIteration[Iteration], + TimeMgr.pv_instability(BestMoveChangesByIteration[Iteration], BestMoveChangesByIteration[Iteration-1]); // Stop search if most of MaxSearchTime is consumed at the end of the @@ -667,12 +653,12 @@ namespace { wait_for_stop_or_ponderhit(); else // Print final search statistics - cout << "info nodes " << ThreadsMgr.nodes_searched() - << " nps " << nps() + cout << "info nodes " << pos.nodes_searched() + << " nps " << nps(pos) << " time " << current_search_time() << endl; // Print the best move and the ponder move to the standard output - if (pv[0] == MOVE_NONE) + if (pv[0] == MOVE_NONE || MultiPV > 1) { pv[0] = rml.move(0); pv[1] = MOVE_NONE; @@ -695,14 +681,14 @@ namespace { if (dbg_show_hit_rate) dbg_print_hit_rate(LogFile); - LogFile << "\nNodes: " << ThreadsMgr.nodes_searched() - << "\nNodes/second: " << nps() - << "\nBest move: " << move_to_san(p, pv[0]); + LogFile << "\nNodes: " << pos.nodes_searched() + << "\nNodes/second: " << nps(pos) + << "\nBest move: " << move_to_san(pos, pv[0]); StateInfo st; - p.do_move(pv[0], st); + pos.do_move(pv[0], st); LogFile << "\nPonder move: " - << move_to_san(p, pv[1]) // Works also with MOVE_NONE + << move_to_san(pos, pv[1]) // Works also with MOVE_NONE << endl; } return rml.move_score(0); @@ -716,7 +702,6 @@ namespace { Value root_search(Position& pos, SearchStack* ss, Move* pv, RootMoveList& rml, Value* alphaPtr, Value* betaPtr) { - Value margins[2]; StateInfo st; CheckInfo ci(pos); int64_t nodes; @@ -741,7 +726,8 @@ namespace { // Step 5. Evaluate the position statically // At root we do this only to get reference value for child nodes - ss->eval = isCheck ? VALUE_NONE : evaluate(pos, margins); + ss->evalMargin = VALUE_NONE; + ss->eval = isCheck ? VALUE_NONE : evaluate(pos, ss->evalMargin); // Step 6. Razoring (omitted at root) // Step 7. Static null move pruning (omitted at root) @@ -764,7 +750,7 @@ namespace { FirstRootMove = (i == 0); // Save the current node count before the move is searched - nodes = ThreadsMgr.nodes_searched(); + nodes = pos.nodes_searched(); // Pick the next root move, and print the move and the move number to // the standard output. @@ -885,7 +871,7 @@ namespace { break; // Remember searched nodes counts for this move - rml.add_move_nodes(i, ThreadsMgr.nodes_searched() - nodes); + rml.add_move_nodes(i, pos.nodes_searched() - nodes); assert(value >= -VALUE_INFINITE && value <= VALUE_INFINITE); assert(value < beta); @@ -927,8 +913,8 @@ namespace { << " score " << value_to_uci(rml.move_score(j)) << " depth " << (j <= i ? Iteration : Iteration - 1) << " time " << current_search_time() - << " nodes " << ThreadsMgr.nodes_searched() - << " nps " << nps() + << " nodes " << pos.nodes_searched() + << " nps " << nps(pos) << " pv "; for (int k = 0; rml.move_pv(j, k) != MOVE_NONE && k < PLY_MAX; k++) @@ -965,9 +951,14 @@ namespace { } - // search<>() is the main search function for both PV and non-PV nodes + // search<>() is the main search function for both PV and non-PV nodes and for + // normal and SplitPoint nodes. When called just after a split point the search + // is simpler because we have already probed the hash table, done a null move + // search, and searched the first move before splitting, we don't have to repeat + // all this work again. 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. - template + template Value search(Position& pos, SearchStack* ss, Value alpha, Value beta, Depth depth, int ply) { assert(alpha >= -VALUE_INFINITE && alpha <= VALUE_INFINITE); @@ -977,37 +968,49 @@ namespace { assert(pos.thread() >= 0 && pos.thread() < ThreadsMgr.active_threads()); Move movesSearched[MOVES_MAX]; - Value margins[2]; StateInfo st; const TTEntry *tte; Key posKey; Move ttMove, move, excludedMove, threatMove; Depth ext, newDepth; + ValueType vt; Value bestValue, value, oldAlpha; Value refinedValue, nullValue, futilityBase, futilityValueScaled; // Non-PV specific bool isCheck, singleEvasion, singularExtensionNode, moveIsCheck, captureOrPromotion, dangerous; bool mateThreat = false; int moveCount = 0; int threadID = pos.thread(); + SplitPoint* sp = NULL; refinedValue = bestValue = value = -VALUE_INFINITE; oldAlpha = alpha; + isCheck = pos.is_check(); + + if (SpNode) + { + sp = ss->sp; + tte = NULL; + ttMove = excludedMove = MOVE_NONE; + threatMove = sp->threatMove; + mateThreat = sp->mateThreat; + goto split_point_start; + } + else {} // Hack to fix icc's "statement is unreachable" warning // Step 1. Initialize node and poll. Polling can abort search - ThreadsMgr.incrementNodeCounter(threadID); ss->currentMove = ss->bestMove = threatMove = MOVE_NONE; (ss+2)->killers[0] = (ss+2)->killers[1] = (ss+2)->mateKiller = MOVE_NONE; if (threadID == 0 && ++NodesSincePoll > NodesBetweenPolls) { NodesSincePoll = 0; - poll(); + poll(pos); } // Step 2. Check for aborted search and immediate draw - if (AbortSearch || ThreadsMgr.thread_should_stop(threadID)) - return VALUE_ZERO; - - if (pos.is_draw() || ply >= PLY_MAX - 1) + if ( AbortSearch + || ThreadsMgr.cutoff_at_splitpoint(threadID) + || pos.is_draw() + || ply >= PLY_MAX - 1) return VALUE_DRAW; // Step 3. Mate distance pruning @@ -1024,7 +1027,7 @@ namespace { posKey = excludedMove ? pos.get_exclusion_key() : pos.get_key(); tte = TT.retrieve(posKey); - ttMove = (tte ? tte->move() : MOVE_NONE); + ttMove = tte ? tte->move() : MOVE_NONE; // At PV nodes, we don't use the TT for pruning, but only for move ordering. // This is to avoid problems in the following areas: @@ -1033,33 +1036,29 @@ namespace { // * Fifty move rule detection // * Searching for a mate // * Printing of full PV line - 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, tte->static_value(), tte->static_value_margin()); - + TT.refresh(tte); ss->bestMove = ttMove; // Can be MOVE_NONE return value_from_tt(tte->value(), ply); } // Step 5. Evaluate the position statically and // update gain statistics of parent move. - isCheck = pos.is_check(); if (isCheck) - ss->eval = VALUE_NONE; + ss->eval = ss->evalMargin = VALUE_NONE; else if (tte) { assert(tte->static_value() != VALUE_NONE); ss->eval = tte->static_value(); - margins[pos.side_to_move()] = tte->static_value_margin(); + ss->evalMargin = tte->static_value_margin(); refinedValue = refine_eval(tte, ss->eval, ply); } else { - refinedValue = ss->eval = evaluate(pos, margins); - TT.store(posKey, VALUE_NONE, VALUE_TYPE_NONE, DEPTH_NONE, MOVE_NONE, ss->eval, margins[pos.side_to_move()]); + refinedValue = ss->eval = evaluate(pos, ss->evalMargin); + TT.store(posKey, VALUE_NONE, VALUE_TYPE_NONE, DEPTH_NONE, MOVE_NONE, ss->eval, ss->evalMargin); } // Save gain for the parent non-capture move @@ -1071,7 +1070,6 @@ namespace { && !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())) { @@ -1096,14 +1094,11 @@ namespace { return refinedValue - futility_margin(depth, 0); // Step 8. Null move search with verification search (is omitted in PV nodes) - // When we jump directly to qsearch() we do a null move only if static value is - // at least beta. Otherwise we do a null move if static value is not more than - // NullMoveMargin under beta. if ( !PvNode && !ss->skipNullMove && depth > ONE_PLY && !isCheck - && refinedValue >= beta - (depth >= 4 * ONE_PLY ? NullMoveMargin : 0) + && refinedValue >= beta && !value_is_mate(beta) && pos.non_pawn_material(pos.side_to_move())) { @@ -1118,9 +1113,7 @@ namespace { pos.do_null_move(st); (ss+1)->skipNullMove = true; - - nullValue = depth-R*ONE_PLY < ONE_PLY ? -qsearch(pos, ss+1, -beta, -alpha, DEPTH_ZERO, ply+1) - : - search(pos, ss+1, -beta, -alpha, depth-R*ONE_PLY, ply+1); + nullValue = -search(pos, ss+1, -beta, -alpha, depth-R*ONE_PLY, ply+1); (ss+1)->skipNullMove = false; pos.undo_null_move(); @@ -1155,6 +1148,7 @@ namespace { threatMove = (ss+1)->bestMove; if ( depth < ThreatDepth && (ss-1)->reduction + && threatMove != MOVE_NONE && connected_moves(pos, (ss-1)->currentMove, threatMove)) return beta - 1; } @@ -1179,29 +1173,46 @@ namespace { if (PvNode) mateThreat = pos.has_mate_threat(); +split_point_start: // At split points actual search starts from here + // Initialize a MovePicker object for the current position - MovePicker mp = MovePicker(pos, ttMove, depth, H, ss, (PvNode ? -VALUE_INFINITE : beta)); + // FIXME currently MovePicker() c'tor is needless called also in SplitPoint + MovePicker mpBase(pos, ttMove, depth, H, ss, (PvNode ? -VALUE_INFINITE : beta)); + MovePicker& mp = SpNode ? *sp->mp : mpBase; CheckInfo ci(pos); ss->bestMove = MOVE_NONE; - singleEvasion = isCheck && mp.number_of_evasions() == 1; - futilityBase = ss->eval + margins[pos.side_to_move()]; - singularExtensionNode = depth >= SingularExtensionDepth[PvNode] + singleEvasion = !SpNode && isCheck && mp.number_of_evasions() == 1; + futilityBase = ss->eval + ss->evalMargin; + singularExtensionNode = !SpNode + && depth >= SingularExtensionDepth[PvNode] && tte && tte->move() && !excludedMove // Do not allow recursive singular extension search && (tte->type() & VALUE_TYPE_LOWER) && tte->depth() >= depth - 3 * ONE_PLY; + if (SpNode) + { + lock_grab(&(sp->lock)); + bestValue = sp->bestValue; + } // Step 10. Loop through moves // Loop through all legal moves until no moves remain or a beta cutoff occurs while ( bestValue < beta && (move = mp.get_next_move()) != MOVE_NONE - && !ThreadsMgr.thread_should_stop(threadID)) + && !ThreadsMgr.cutoff_at_splitpoint(threadID)) { assert(move_is_ok(move)); - if (move == excludedMove) + if (SpNode) + { + moveCount = ++sp->moveCount; + lock_release(&(sp->lock)); + } + else if (move == excludedMove) continue; + else + movesSearched[moveCount++] = move; moveIsCheck = pos.move_is_check(move, ci); captureOrPromotion = pos.move_is_capture_or_promotion(move); @@ -1233,10 +1244,9 @@ namespace { } } - newDepth = depth - ONE_PLY + ext; - // Update current move (this must be done after singular extension search) - movesSearched[moveCount++] = ss->currentMove = move; + ss->currentMove = move; + newDepth = depth - ONE_PLY + ext; // Step 12. Futility pruning (is omitted in PV nodes) if ( !PvNode @@ -1249,8 +1259,13 @@ namespace { // Move count based pruning if ( moveCount >= futility_move_count(depth) && !(threatMove && connected_threat(pos, move, threatMove)) - && bestValue > value_mated_in(PLY_MAX)) + && bestValue > value_mated_in(PLY_MAX)) // FIXME bestValue is racy + { + if (SpNode) + lock_grab(&(sp->lock)); + continue; + } // Value based pruning // We illogically ignore reduction condition depth >= 3*ONE_PLY for predicted depth, @@ -1261,8 +1276,26 @@ namespace { if (futilityValueScaled < beta) { - if (futilityValueScaled > bestValue) + if (SpNode) + { + lock_grab(&(sp->lock)); + if (futilityValueScaled > sp->bestValue) + sp->bestValue = bestValue = futilityValueScaled; + } + else if (futilityValueScaled > bestValue) bestValue = futilityValueScaled; + + continue; + } + + // Prune moves with negative SEE at low depths + if ( predictedDepth < 2 * ONE_PLY + && bestValue > value_mated_in(PLY_MAX) + && pos.see_sign(move) < 0) + { + if (SpNode) + lock_grab(&(sp->lock)); + continue; } } @@ -1273,8 +1306,7 @@ namespace { // Step extra. pv search (only in PV nodes) // The first move in list is the expected PV if (PvNode && moveCount == 1) - value = newDepth < ONE_PLY ? -qsearch(pos, ss+1, -beta, -alpha, DEPTH_ZERO, ply+1) - : - search(pos, ss+1, -beta, -alpha, newDepth, ply+1); + value = -search(pos, ss+1, -beta, -alpha, newDepth, ply+1); else { // Step 14. Reduced depth search @@ -1285,14 +1317,16 @@ namespace { && !captureOrPromotion && !dangerous && !move_is_castle(move) - && !move_is_killer(move, ss)) + && ss->killers[0] != move + && ss->killers[1] != move) { ss->reduction = reduction(depth, moveCount); + if (ss->reduction) { + alpha = SpNode ? sp->alpha : alpha; Depth d = newDepth - ss->reduction; - value = d < ONE_PLY ? -qsearch(pos, ss+1, -(alpha+1), -alpha, DEPTH_ZERO, ply+1) - : - search(pos, ss+1, -(alpha+1), -alpha, d, ply+1); + value = -search(pos, ss+1, -(alpha+1), -alpha, d, ply+1); doFullDepthSearch = (value > alpha); } @@ -1305,6 +1339,7 @@ namespace { assert(newDepth - ONE_PLY >= ONE_PLY); ss->reduction = ONE_PLY; + alpha = SpNode ? sp->alpha : alpha; value = -search(pos, ss+1, -(alpha+1), -alpha, newDepth-ss->reduction, ply+1); doFullDepthSearch = (value > alpha); } @@ -1314,15 +1349,14 @@ namespace { // Step 15. Full depth search if (doFullDepthSearch) { - value = newDepth < ONE_PLY ? -qsearch(pos, ss+1, -(alpha+1), -alpha, DEPTH_ZERO, ply+1) - : - search(pos, ss+1, -(alpha+1), -alpha, newDepth, ply+1); + alpha = SpNode ? sp->alpha : alpha; + value = -search(pos, ss+1, -(alpha+1), -alpha, newDepth, ply+1); // 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 < ONE_PLY ? -qsearch(pos, ss+1, -beta, -alpha, DEPTH_ZERO, ply+1) - : - search(pos, ss+1, -beta, -alpha, newDepth, ply+1); + value = -search(pos, ss+1, -beta, -alpha, newDepth, ply+1); } } @@ -1332,56 +1366,88 @@ namespace { assert(value > -VALUE_INFINITE && value < VALUE_INFINITE); // Step 17. Check for new best move - if (value > bestValue) + if (SpNode) + { + lock_grab(&(sp->lock)); + bestValue = sp->bestValue; + alpha = sp->alpha; + } + + if (value > bestValue && !(SpNode && ThreadsMgr.cutoff_at_splitpoint(threadID))) { bestValue = value; + + if (SpNode) + sp->bestValue = value; + if (value > alpha) { if (PvNode && value < beta) // We want always alpha < beta + { alpha = value; + if (SpNode) + sp->alpha = value; + } + else if (SpNode) + sp->betaCutoff = true; + if (value == value_mate_in(ply + 1)) ss->mateKiller = move; ss->bestMove = move; + + if (SpNode) + sp->parentSstack->bestMove = move; } } // Step 18. Check for split - if ( depth >= MinimumSplitDepth + if ( !SpNode + && depth >= ThreadsMgr.min_split_depth() && ThreadsMgr.active_threads() > 1 && bestValue < beta && ThreadsMgr.available_thread_exists(threadID) && !AbortSearch - && !ThreadsMgr.thread_should_stop(threadID) + && !ThreadsMgr.cutoff_at_splitpoint(threadID) && Iteration <= 99) ThreadsMgr.split(pos, ss, ply, &alpha, beta, &bestValue, depth, - threatMove, mateThreat, &moveCount, &mp, PvNode); + threatMove, mateThreat, moveCount, &mp, PvNode); } // Step 19. Check for mate and stalemate // All legal moves have been searched and if there are // no legal moves, it must be mate or stalemate. // If one move was excluded return fail low score. - if (!moveCount) + if (!SpNode && !moveCount) return excludedMove ? oldAlpha : isCheck ? value_mated_in(ply) : VALUE_DRAW; // Step 20. Update tables // If the search is not aborted, update the transposition table, // history counters, and killer moves. - if (AbortSearch || ThreadsMgr.thread_should_stop(threadID)) - return bestValue; + if (!SpNode && !AbortSearch && !ThreadsMgr.cutoff_at_splitpoint(threadID)) + { + move = bestValue <= oldAlpha ? MOVE_NONE : ss->bestMove; + vt = bestValue <= oldAlpha ? VALUE_TYPE_UPPER + : bestValue >= beta ? VALUE_TYPE_LOWER : VALUE_TYPE_EXACT; - ValueType vt = (bestValue <= oldAlpha ? VALUE_TYPE_UPPER : bestValue >= beta ? VALUE_TYPE_LOWER : VALUE_TYPE_EXACT); - move = (bestValue <= oldAlpha ? MOVE_NONE : ss->bestMove); - TT.store(posKey, value_to_tt(bestValue, ply), vt, depth, move, ss->eval, margins[pos.side_to_move()]); + TT.store(posKey, value_to_tt(bestValue, ply), vt, depth, move, ss->eval, ss->evalMargin); - // Update killers and history only for non capture moves that fails high - if ( bestValue >= beta - && !pos.move_is_capture_or_promotion(move)) - { + // Update killers and history only for non capture moves that fails high + if ( bestValue >= beta + && !pos.move_is_capture_or_promotion(move)) + { update_history(pos, move, depth, movesSearched, moveCount); update_killers(move, ss); + } + } + + if (SpNode) + { + // Here we have the lock still grabbed + sp->slaves[threadID] = 0; + sp->nodes += pos.nodes_searched(); + lock_release(&(sp->lock)); } assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE); @@ -1389,7 +1455,6 @@ namespace { return bestValue; } - // qsearch() is the quiescence search function, which is called by the main // search function when the remaining depth is zero (or, to be more precise, // less than ONE_PLY). @@ -1404,40 +1469,43 @@ namespace { assert(ply > 0 && ply < PLY_MAX); assert(pos.thread() >= 0 && pos.thread() < ThreadsMgr.active_threads()); - Value margins[2]; StateInfo st; Move ttMove, move; - Value bestValue, value, futilityValue, futilityBase; - bool isCheck, deepChecks, enoughMaterial, moveIsCheck, evasionPrunable; + Value bestValue, value, evalMargin, futilityValue, futilityBase; + bool isCheck, enoughMaterial, moveIsCheck, evasionPrunable; const TTEntry* tte; + Depth ttDepth; Value oldAlpha = alpha; - ThreadsMgr.incrementNodeCounter(pos.thread()); ss->bestMove = ss->currentMove = MOVE_NONE; // Check for an instant draw or maximum ply reached if (pos.is_draw() || ply >= PLY_MAX - 1) return VALUE_DRAW; + // Decide whether or not to include checks, this fixes also the type of + // TT entry depth that we are going to use. Note that in qsearch we use + // only two types of depth in TT: DEPTH_QS_CHECKS or DEPTH_QS_NO_CHECKS. + isCheck = pos.is_check(); + ttDepth = (isCheck || depth >= DEPTH_QS_CHECKS ? DEPTH_QS_CHECKS : DEPTH_QS_NO_CHECKS); + // Transposition table lookup. At PV nodes, we don't use the TT for // pruning, but only for move ordering. 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, ttDepth, beta, ply)) { ss->bestMove = ttMove; // Can be MOVE_NONE return value_from_tt(tte->value(), ply); } - isCheck = pos.is_check(); - // Evaluate the position statically if (isCheck) { bestValue = futilityBase = -VALUE_INFINITE; - ss->eval = VALUE_NONE; - deepChecks = enoughMaterial = false; + ss->eval = evalMargin = VALUE_NONE; + enoughMaterial = false; } else { @@ -1445,20 +1513,19 @@ namespace { { assert(tte->static_value() != VALUE_NONE); - margins[pos.side_to_move()] = tte->static_value_margin(); - bestValue = tte->static_value(); + evalMargin = tte->static_value_margin(); + ss->eval = bestValue = tte->static_value(); } else - bestValue = evaluate(pos, margins); + ss->eval = bestValue = evaluate(pos, evalMargin); - ss->eval = bestValue; update_gains(pos, (ss-1)->currentMove, (ss-1)->eval, ss->eval); // Stand pat. Return immediately if static value is at least beta if (bestValue >= beta) { if (!tte) - TT.store(pos.get_key(), value_to_tt(bestValue, ply), VALUE_TYPE_LOWER, DEPTH_NONE, MOVE_NONE, ss->eval, margins[pos.side_to_move()]); + TT.store(pos.get_key(), value_to_tt(bestValue, ply), VALUE_TYPE_LOWER, DEPTH_NONE, MOVE_NONE, ss->eval, evalMargin); return bestValue; } @@ -1466,19 +1533,16 @@ namespace { if (PvNode && bestValue > alpha) alpha = bestValue; - // If we are near beta then try to get a cutoff pushing checks a bit further - deepChecks = (depth == -ONE_PLY && bestValue >= beta - PawnValueMidgame / 8); - // Futility pruning parameters, not needed when in check - futilityBase = bestValue + FutilityMarginQS + margins[pos.side_to_move()]; + futilityBase = ss->eval + evalMargin + FutilityMarginQS; enoughMaterial = pos.non_pawn_material(pos.side_to_move()) > RookValueMidgame; } // Initialize a MovePicker object for the current position, and prepare // to search the moves. Because the depth is <= 0 here, only captures, - // queen promotions and checks (only if depth == 0 or depth == -ONE_PLY - // and we are near beta) will be generated. - MovePicker mp = MovePicker(pos, ttMove, deepChecks ? DEPTH_ZERO : depth, H); + // queen promotions and checks (only if depth >= DEPTH_QS_CHECKS) will + // be generated. + MovePicker mp = MovePicker(pos, ttMove, depth, H); CheckInfo ci(pos); // Loop through the moves until no moves remain or a beta cutoff occurs @@ -1510,11 +1574,10 @@ namespace { } } - // Detect blocking evasions that are candidate to be pruned + // Detect non-capture evasions that are candidate to be pruned evasionPrunable = isCheck && bestValue > value_mated_in(PLY_MAX) && !pos.move_is_capture(move) - && pos.type_of_piece_on(move_from(move)) != KING && !pos.can_castle(pos.side_to_move()); // Don't search moves with negative SEE values @@ -1525,6 +1588,21 @@ namespace { && pos.see_sign(move) < 0) continue; + // Don't search useless checks + if ( !PvNode + && !isCheck + && moveIsCheck + && move != ttMove + && !pos.move_is_capture_or_promotion(move) + && ss->eval + PawnValueMidgame / 4 < beta + && !check_is_dangerous(pos, move, futilityBase, beta, &bestValue)) + { + if (ss->eval + PawnValueMidgame / 4 > bestValue) + bestValue = ss->eval + PawnValueMidgame / 4; + + continue; + } + // Update current move ss->currentMove = move; @@ -1553,14 +1631,8 @@ namespace { return value_mated_in(ply); // Update transposition table - Depth d = (depth == DEPTH_ZERO ? DEPTH_ZERO : DEPTH_ZERO - ONE_PLY); ValueType vt = (bestValue <= oldAlpha ? VALUE_TYPE_UPPER : bestValue >= beta ? VALUE_TYPE_LOWER : VALUE_TYPE_EXACT); - TT.store(pos.get_key(), value_to_tt(bestValue, ply), vt, d, ss->bestMove, ss->eval, margins[pos.side_to_move()]); - - // Update killers only for checking moves that fails high - if ( bestValue >= beta - && !pos.move_is_capture_or_promotion(ss->bestMove)) - update_killers(ss->bestMove, ss); + TT.store(pos.get_key(), value_to_tt(bestValue, ply), vt, ttDepth, ss->bestMove, ss->eval, evalMargin); assert(bestValue > -VALUE_INFINITE && bestValue < VALUE_INFINITE); @@ -1568,171 +1640,60 @@ namespace { } - // sp_search() is used to search from a split point. This function is called - // by each thread working at the split point. It is similar to the normal - // search() function, but simpler. Because we have already probed the hash - // table, done a null move search, and searched the first move before - // splitting, we don't have to repeat all this work in sp_search(). 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. - - template - void sp_search(SplitPoint* sp, int threadID) { - - assert(threadID >= 0 && threadID < ThreadsMgr.active_threads()); - assert(ThreadsMgr.active_threads() > 1); + // check_is_dangerous() tests if a checking move can be pruned in qsearch(). + // bestValue is updated only when returning false because in that case move + // will be pruned. - StateInfo st; - Move move; - Depth ext, newDepth; - Value value; - Value futilityValueScaled; // NonPV specific - bool isCheck, moveIsCheck, captureOrPromotion, dangerous; - int moveCount; - value = -VALUE_INFINITE; + bool check_is_dangerous(Position &pos, Move move, Value futilityBase, Value beta, Value *bestValue) + { + Bitboard b, occ, oldAtt, newAtt, kingAtt; + Square from, to, ksq, victimSq; + Piece pc; + Color them; + Value futilityValue, bv = *bestValue; + + from = move_from(move); + to = move_to(move); + them = opposite_color(pos.side_to_move()); + ksq = pos.king_square(them); + kingAtt = pos.attacks_from(ksq); + pc = pos.piece_on(from); + + occ = pos.occupied_squares() & ~(1ULL << from) & ~(1ULL << ksq); + oldAtt = pos.attacks_from(pc, from, occ); + newAtt = pos.attacks_from(pc, to, occ); + + // Rule 1. Checks which give opponent's king at most one escape square are dangerous + b = kingAtt & ~pos.pieces_of_color(them) & ~newAtt & ~(1ULL << to); + + if (!(b && (b & (b - 1)))) + return true; - Position pos(*sp->pos, threadID); - CheckInfo ci(pos); - SearchStack* ss = sp->sstack[threadID] + 1; - isCheck = pos.is_check(); + // Rule 2. Queen contact check is very dangerous + if ( type_of_piece(pc) == QUEEN + && bit_is_set(kingAtt, to)) + return true; - // Step 10. Loop through moves - // Loop through all legal moves until no moves remain or a beta cutoff occurs - lock_grab(&(sp->lock)); + // Rule 3. Creating new double threats with checks + b = pos.pieces_of_color(them) & newAtt & ~oldAtt & ~(1ULL << ksq); - while ( sp->bestValue < sp->beta - && (move = sp->mp->get_next_move()) != MOVE_NONE - && !ThreadsMgr.thread_should_stop(threadID)) + while (b) { - moveCount = ++sp->moveCount; - lock_release(&(sp->lock)); - - 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 - ONE_PLY + ext; - - // Update current move - ss->currentMove = move; - - // 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) - && !(sp->threatMove && connected_threat(pos, move, sp->threatMove)) - && sp->bestValue > value_mated_in(PLY_MAX)) - { - lock_grab(&(sp->lock)); - continue; - } - - // Value based pruning - Depth predictedDepth = newDepth - reduction(sp->depth, moveCount); - futilityValueScaled = ss->eval + futility_margin(predictedDepth, moveCount) - + H.gain(pos.piece_on(move_from(move)), move_to(move)); - - if (futilityValueScaled < sp->beta) - { - lock_grab(&(sp->lock)); - - if (futilityValueScaled > sp->bestValue) - sp->bestValue = futilityValueScaled; - continue; - } - } - - // 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 ( !captureOrPromotion - && !dangerous - && !move_is_castle(move) - && !move_is_killer(move, ss)) - { - ss->reduction = reduction(sp->depth, moveCount); - if (ss->reduction) - { - Value localAlpha = sp->alpha; - Depth d = newDepth - ss->reduction; - value = d < ONE_PLY ? -qsearch(pos, ss+1, -(localAlpha+1), -localAlpha, DEPTH_ZERO, sp->ply+1) - : - search(pos, ss+1, -(localAlpha+1), -localAlpha, d, sp->ply+1); - - doFullDepthSearch = (value > localAlpha); - } - - // 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 * ONE_PLY) - { - assert(newDepth - ONE_PLY >= ONE_PLY); + victimSq = pop_1st_bit(&b); + futilityValue = futilityBase + pos.endgame_value_of_piece_on(victimSq); - ss->reduction = ONE_PLY; - Value localAlpha = sp->alpha; - value = -search(pos, ss+1, -(localAlpha+1), -localAlpha, newDepth-ss->reduction, sp->ply+1); - doFullDepthSearch = (value > localAlpha); - } - ss->reduction = DEPTH_ZERO; // Restore original reduction - } - - // Step 15. Full depth search - if (doFullDepthSearch) - { - Value localAlpha = sp->alpha; - value = newDepth < ONE_PLY ? -qsearch(pos, ss+1, -(localAlpha+1), -localAlpha, DEPTH_ZERO, sp->ply+1) - : - search(pos, ss+1, -(localAlpha+1), -localAlpha, newDepth, sp->ply+1); - - // 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 < ONE_PLY ? -qsearch(pos, ss+1, -sp->beta, -sp->alpha, DEPTH_ZERO, sp->ply+1) - : - search(pos, ss+1, -sp->beta, -sp->alpha, newDepth, sp->ply+1); - } - - // 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 && !ThreadsMgr.thread_should_stop(threadID)) - { - sp->bestValue = value; - - if (sp->bestValue > sp->alpha) - { - if (!PvNode || value >= sp->beta) - sp->stopRequest = true; - - if (PvNode && value < sp->beta) // This guarantees that always: sp->alpha < sp->beta - sp->alpha = value; + // Note that here we generate illegal "double move"! + if ( futilityValue >= beta + && pos.see_sign(make_move(from, victimSq)) >= 0) + return true; - sp->parentSstack->bestMove = ss->bestMove = move; - } - } + if (futilityValue > bv) + bv = futilityValue; } - /* Here we have the lock still grabbed */ - - sp->slaves[threadID] = 0; - - lock_release(&(sp->lock)); + // Update bestValue only if check is not dangerous (because we will prune the move) + *bestValue = bv; + return false; } @@ -1747,11 +1708,8 @@ namespace { Square f1, t1, f2, t2; Piece p; - assert(move_is_ok(m1)); - assert(move_is_ok(m2)); - - if (m2 == MOVE_NONE) - return false; + assert(m1 && move_is_ok(m1)); + assert(m2 && move_is_ok(m2)); // Case 1: The moving piece is the same in both moves f2 = move_from(m2); @@ -1835,17 +1793,6 @@ namespace { } - // move_is_killer() checks if the given move is among the killer moves - - bool move_is_killer(Move m, SearchStack* ss) { - - if (ss->killers[0] == m || ss->killers[1] == m) - return true; - - return false; - } - - // extension() decides whether a move should be searched with normal depth, // or with extended depth. Certain classes of moves (checking moves, in // particular) are searched with bigger depth than ordinary moves and in @@ -1991,7 +1938,6 @@ namespace { void update_history(const Position& pos, Move move, Depth depth, Move movesSearched[], int moveCount) { - Move m; H.success(pos.piece_on(move_from(move)), move_to(move), depth); @@ -2060,10 +2006,10 @@ namespace { // nps() computes the current nodes/second count. - int nps() { + int nps(const Position& pos) { int t = current_search_time(); - return (t > 0 ? int((ThreadsMgr.nodes_searched() * 1000) / t) : 0); + return (t > 0 ? int((pos.nodes_searched() * 1000) / t) : 0); } @@ -2071,13 +2017,13 @@ namespace { // looks at the time consumed so far and decides if it's time to abort the // search. - void poll() { + void poll(const Position& pos) { static int lastInfoTime; int t = current_search_time(); // Poll for input - if (Bioskey()) + if (data_available()) { // We are line oriented, don't read single chars std::string command; @@ -2120,7 +2066,7 @@ namespace { if (dbg_show_hit_rate) dbg_print_hit_rate(); - cout << "info nodes " << ThreadsMgr.nodes_searched() << " nps " << nps() + cout << "info nodes " << pos.nodes_searched() << " nps " << nps(pos) << " time " << t << endl; } @@ -2137,7 +2083,7 @@ namespace { if ( (Iteration >= 3 && UseTimeManagement && noMoreTime) || (ExactMaxTime && t >= ExactMaxTime) - || (Iteration >= 3 && MaxNodes && ThreadsMgr.nodes_searched() >= MaxNodes)) + || (Iteration >= 3 && MaxNodes && pos.nodes_searched() >= MaxNodes)) AbortSearch = true; } @@ -2173,6 +2119,7 @@ namespace { ss->excludedMove = MOVE_NONE; ss->skipNullMove = false; ss->reduction = DEPTH_ZERO; + ss->sp = NULL; if (i < 3) ss->killers[0] = ss->killers[1] = ss->mateKiller = MOVE_NONE; @@ -2216,8 +2163,8 @@ namespace { << " score " << value_to_uci(value) << (value >= beta ? " lowerbound" : value <= alpha ? " upperbound" : "") << " time " << current_search_time() - << " nodes " << ThreadsMgr.nodes_searched() - << " nps " << nps() + << " nodes " << pos.nodes_searched() + << " nps " << nps(pos) << " pv "; for (Move* m = pv; *m != MOVE_NONE; m++) @@ -2230,8 +2177,7 @@ namespace { ValueType t = value >= beta ? VALUE_TYPE_LOWER : value <= alpha ? VALUE_TYPE_UPPER : VALUE_TYPE_EXACT; - LogFile << pretty_pv(pos, current_search_time(), Iteration, - ThreadsMgr.nodes_searched(), value, t, pv) << endl; + LogFile << pretty_pv(pos, current_search_time(), Iteration, value, t, pv) << endl; } } @@ -2245,16 +2191,15 @@ namespace { StateInfo st; TTEntry* tte; Position p(pos, pos.thread()); - Value margins[2]; - Value v; + Value v, m = VALUE_NONE; for (int i = 0; pv[i] != MOVE_NONE; i++) { tte = TT.retrieve(p.get_key()); if (!tte || tte->move() != pv[i]) { - v = (p.is_check() ? VALUE_NONE : evaluate(p, margins)); - TT.store(p.get_key(), VALUE_NONE, VALUE_TYPE_NONE, DEPTH_NONE, pv[i], v, margins[pos.side_to_move()]); + v = (p.is_check() ? VALUE_NONE : evaluate(p, m)); + TT.store(p.get_key(), VALUE_NONE, VALUE_TYPE_NONE, DEPTH_NONE, pv[i], v, m); } p.do_move(pv[i], st); } @@ -2298,7 +2243,7 @@ namespace { #if !defined(_MSC_VER) - void* init_thread(void *threadID) { + void* init_thread(void* threadID) { ThreadsMgr.idle_loop(*(int*)threadID, NULL); return NULL; @@ -2317,23 +2262,17 @@ namespace { /// The ThreadsManager class - // resetNodeCounters(), resetBetaCounters(), searched_nodes() and - // get_beta_counters() are getters/setters for the per thread - // counters used to sort the moves at root. - - void ThreadsManager::resetNodeCounters() { - - for (int i = 0; i < MAX_THREADS; i++) - threads[i].nodes = 0ULL; - } - int64_t ThreadsManager::nodes_searched() const { + // read_uci_options() updates number of active threads and other internal + // parameters according to the UCI options values. It is called before + // to start a new search. - int64_t result = 0ULL; - for (int i = 0; i < ActiveThreads; i++) - result += threads[i].nodes; + void ThreadsManager::read_uci_options() { - return result; + maxThreadsPerSplitPoint = Options["Maximum Number of Threads per Split Point"].value(); + minimumSplitDepth = Options["Minimum Split Depth"].value() * ONE_PLY; + useSleepingThreads = Options["Use Sleeping Threads"].value(); + activeThreads = Options["Threads"].value(); } @@ -2345,11 +2284,14 @@ namespace { assert(threadID >= 0 && threadID < MAX_THREADS); + int i; + bool allFinished = false; + while (true) { // Slave threads can exit as soon as AllThreadsShouldExit raises, // master should exit as last one. - if (AllThreadsShouldExit) + if (allThreadsShouldExit) { assert(!sp); threads[threadID].state = THREAD_TERMINATED; @@ -2358,55 +2300,77 @@ namespace { // If we are not thinking, wait for a condition to be signaled // instead of wasting CPU time polling for work. - while (AllThreadsShouldSleep || threadID >= ActiveThreads) + while ( threadID >= activeThreads || threads[threadID].state == THREAD_INITIALIZING + || (useSleepingThreads && threads[threadID].state == THREAD_AVAILABLE)) { - assert(!sp); - assert(threadID != 0); - threads[threadID].state = THREAD_SLEEPING; + assert(!sp || useSleepingThreads); + assert(threadID != 0 || useSleepingThreads); -#if !defined(_MSC_VER) - lock_grab(&WaitLock); - if (AllThreadsShouldSleep || threadID >= ActiveThreads) - pthread_cond_wait(&WaitCond, &WaitLock); - lock_release(&WaitLock); -#else - WaitForSingleObject(SitIdleEvent[threadID], INFINITE); -#endif - } + if (threads[threadID].state == THREAD_INITIALIZING) + threads[threadID].state = THREAD_AVAILABLE; - // If thread has just woken up, mark it as available - if (threads[threadID].state == THREAD_SLEEPING) - threads[threadID].state = THREAD_AVAILABLE; + // Grab the lock to avoid races with wake_sleeping_thread() + lock_grab(&sleepLock[threadID]); + + // If we are master and all slaves have finished do not go to sleep + for (i = 0; sp && i < activeThreads && !sp->slaves[i]; i++) {} + allFinished = (i == activeThreads); + + if (allFinished || allThreadsShouldExit) + { + lock_release(&sleepLock[threadID]); + break; + } + + // Do sleep here after retesting sleep conditions + if (threadID >= activeThreads || threads[threadID].state == THREAD_AVAILABLE) + cond_wait(&sleepCond[threadID], &sleepLock[threadID]); + + lock_release(&sleepLock[threadID]); + } // If this thread has been assigned work, launch a search if (threads[threadID].state == THREAD_WORKISWAITING) { - assert(!AllThreadsShouldExit && !AllThreadsShouldSleep); + assert(!allThreadsShouldExit); threads[threadID].state = THREAD_SEARCHING; - if (threads[threadID].splitPoint->pvNode) - sp_search(threads[threadID].splitPoint, threadID); + // Here we call search() with SplitPoint template parameter set to true + SplitPoint* tsp = threads[threadID].splitPoint; + Position pos(*tsp->pos, threadID); + SearchStack* ss = tsp->sstack[threadID] + 1; + ss->sp = tsp; + + if (tsp->pvNode) + search(pos, ss, tsp->alpha, tsp->beta, tsp->depth, tsp->ply); else - sp_search(threads[threadID].splitPoint, threadID); + search(pos, ss, tsp->alpha, tsp->beta, tsp->depth, tsp->ply); assert(threads[threadID].state == THREAD_SEARCHING); threads[threadID].state = THREAD_AVAILABLE; + + // Wake up master thread so to allow it to return from the idle loop in + // case we are the last slave of the split point. + if (useSleepingThreads && threadID != tsp->master && threads[tsp->master].state == THREAD_AVAILABLE) + wake_sleeping_thread(tsp->master); } // 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. - int i = 0; - for ( ; sp && i < ActiveThreads && !sp->slaves[i]; i++) {} + for (i = 0; sp && i < activeThreads && !sp->slaves[i]; i++) {} + allFinished = (i == activeThreads); - if (i == ActiveThreads) + if (allFinished) { // 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)); + // In helpful master concept a master can help only a sub-tree, and + // because here is all finished is not possible master is booked. assert(threads[threadID].state == THREAD_AVAILABLE); threads[threadID].state = THREAD_SEARCHING; @@ -2422,23 +2386,17 @@ namespace { void ThreadsManager::init_threads() { - volatile int i; + int i, arg[MAX_THREADS]; bool ok; -#if !defined(_MSC_VER) - pthread_t pthread[1]; -#endif - // Initialize global locks - lock_init(&MPLock); - lock_init(&WaitLock); + lock_init(&mpLock); -#if !defined(_MSC_VER) - pthread_cond_init(&WaitCond, NULL); -#else for (i = 0; i < MAX_THREADS; i++) - SitIdleEvent[i] = CreateEvent(0, FALSE, FALSE, 0); -#endif + { + lock_init(&sleepLock[i]); + cond_init(&sleepCond[i]); + } // Initialize splitPoints[] locks for (i = 0; i < MAX_THREADS; i++) @@ -2446,35 +2404,36 @@ namespace { lock_init(&(threads[i].splitPoints[j].lock)); // Will be set just before program exits to properly end the threads - AllThreadsShouldExit = false; + allThreadsShouldExit = false; - // Threads will be put to sleep as soon as created - AllThreadsShouldSleep = true; + // Threads will be put all threads to sleep as soon as created + activeThreads = 1; - // All threads except the main thread should be initialized to THREAD_AVAILABLE - ActiveThreads = 1; + // All threads except the main thread should be initialized to THREAD_INITIALIZING threads[0].state = THREAD_SEARCHING; for (i = 1; i < MAX_THREADS; i++) - threads[i].state = THREAD_AVAILABLE; + threads[i].state = THREAD_INITIALIZING; // Launch the helper threads for (i = 1; i < MAX_THREADS; i++) { + arg[i] = i; #if !defined(_MSC_VER) - ok = (pthread_create(pthread, NULL, init_thread, (void*)(&i)) == 0); + pthread_t pthread[1]; + ok = (pthread_create(pthread, NULL, init_thread, (void*)(&arg[i])) == 0); + pthread_detach(pthread[0]); #else - ok = (CreateThread(NULL, 0, init_thread, (LPVOID)(&i), 0, NULL) != NULL); + ok = (CreateThread(NULL, 0, init_thread, (LPVOID)(&arg[i]), 0, NULL) != NULL); #endif - if (!ok) { cout << "Failed to create thread number " << i << endl; - Application::exit_with_failure(); + exit(EXIT_FAILURE); } // Wait until the thread has finished launching and is gone to sleep - while (threads[i].state != THREAD_SLEEPING) {} + while (threads[i].state == THREAD_INITIALIZING) {} } } @@ -2484,38 +2443,42 @@ namespace { void ThreadsManager::exit_threads() { - ActiveThreads = MAX_THREADS; // HACK - AllThreadsShouldSleep = true; // HACK - wake_sleeping_threads(); - - // This makes the threads to exit idle_loop() - AllThreadsShouldExit = true; + allThreadsShouldExit = true; // Let the woken up threads to exit idle_loop() - // Wait for thread termination + // Wake up all the threads and waits for termination for (int i = 1; i < MAX_THREADS; i++) + { + wake_sleeping_thread(i); while (threads[i].state != THREAD_TERMINATED) {} + } // Now we can safely destroy the locks for (int i = 0; i < MAX_THREADS; i++) for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) lock_destroy(&(threads[i].splitPoints[j].lock)); - lock_destroy(&WaitLock); - lock_destroy(&MPLock); + lock_destroy(&mpLock); + + // Now we can safely destroy the wait conditions + for (int i = 0; i < MAX_THREADS; i++) + { + lock_destroy(&sleepLock[i]); + cond_destroy(&sleepCond[i]); + } } - // thread_should_stop() checks whether the thread should stop its search. - // This can happen if a beta cutoff has occurred in the thread's currently - // active split point, or in some ancestor of the current split point. + // cutoff_at_splitpoint() checks whether a beta cutoff has occurred in + // the thread's currently active split point, or in some ancestor of + // the current split point. - bool ThreadsManager::thread_should_stop(int threadID) const { + bool ThreadsManager::cutoff_at_splitpoint(int threadID) const { - assert(threadID >= 0 && threadID < ActiveThreads); + assert(threadID >= 0 && threadID < activeThreads); - SplitPoint* sp; + SplitPoint* sp = threads[threadID].splitPoint; - for (sp = threads[threadID].splitPoint; sp && !sp->stopRequest; sp = sp->parent) {} + for ( ; sp && !sp->betaCutoff; sp = sp->parent) {} return sp != NULL; } @@ -2530,9 +2493,9 @@ namespace { bool ThreadsManager::thread_is_available(int slave, int master) const { - assert(slave >= 0 && slave < ActiveThreads); - assert(master >= 0 && master < ActiveThreads); - assert(ActiveThreads > 1); + assert(slave >= 0 && slave < activeThreads); + assert(master >= 0 && master < activeThreads); + assert(activeThreads > 1); if (threads[slave].state != THREAD_AVAILABLE || slave == master) return false; @@ -2540,12 +2503,9 @@ namespace { // Make a local copy to be sure doesn't change under our feet int localActiveSplitPoints = threads[slave].activeSplitPoints; - if (localActiveSplitPoints == 0) - // No active split points means that the thread is available as - // a slave for any other thread. - return true; - - if (ActiveThreads == 2) + // No active split points means that the thread is available as + // a slave for any other thread. + if (localActiveSplitPoints == 0 || activeThreads == 2) return true; // Apply the "helpful master" concept if possible. Use localActiveSplitPoints @@ -2563,10 +2523,10 @@ namespace { bool ThreadsManager::available_thread_exists(int master) const { - assert(master >= 0 && master < ActiveThreads); - assert(ActiveThreads > 1); + assert(master >= 0 && master < activeThreads); + assert(activeThreads > 1); - for (int i = 0; i < ActiveThreads; i++) + for (int i = 0; i < activeThreads; i++) if (thread_is_available(i, master)) return true; @@ -2580,35 +2540,34 @@ namespace { // 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. + // been assigned work. This will cause them to instantly leave their idle loops and + // call search().When all threads have returned from search() then split() returns. template - void ThreadsManager::split(const Position& p, SearchStack* ss, int ply, Value* alpha, + void ThreadsManager::split(Position& pos, SearchStack* ss, int ply, Value* alpha, const Value beta, Value* bestValue, Depth depth, Move threatMove, - bool mateThreat, int* moveCount, MovePicker* mp, bool pvNode) { - assert(p.is_ok()); + bool mateThreat, int moveCount, MovePicker* mp, bool pvNode) { + assert(pos.is_ok()); assert(ply > 0 && ply < PLY_MAX); assert(*bestValue >= -VALUE_INFINITE); assert(*bestValue <= *alpha); assert(*alpha < beta); assert(beta <= VALUE_INFINITE); assert(depth > DEPTH_ZERO); - assert(p.thread() >= 0 && p.thread() < ActiveThreads); - assert(ActiveThreads > 1); + assert(pos.thread() >= 0 && pos.thread() < activeThreads); + assert(activeThreads > 1); - int i, master = p.thread(); + int i, master = pos.thread(); Thread& masterThread = threads[master]; - lock_grab(&MPLock); + lock_grab(&mpLock); // If no other thread is available to help us, or if we have too many // active split points, don't split. if ( !available_thread_exists(master) || masterThread.activeSplitPoints >= MAX_ACTIVE_SPLIT_POINTS) { - lock_release(&MPLock); + lock_release(&mpLock); return; } @@ -2617,7 +2576,8 @@ namespace { // Initialize the split point object splitPoint.parent = masterThread.splitPoint; - splitPoint.stopRequest = false; + splitPoint.master = master; + splitPoint.betaCutoff = false; splitPoint.ply = ply; splitPoint.depth = depth; splitPoint.threatMove = threatMove; @@ -2627,10 +2587,11 @@ namespace { splitPoint.pvNode = pvNode; splitPoint.bestValue = *bestValue; splitPoint.mp = mp; - splitPoint.moveCount = *moveCount; - splitPoint.pos = &p; + splitPoint.moveCount = moveCount; + splitPoint.pos = &pos; + splitPoint.nodes = 0; splitPoint.parentSstack = ss; - for (i = 0; i < ActiveThreads; i++) + for (i = 0; i < activeThreads; i++) splitPoint.slaves[i] = 0; masterThread.splitPoint = &splitPoint; @@ -2641,7 +2602,7 @@ namespace { int workersCnt = 1; // At least the master is included // Allocate available threads setting state to THREAD_BOOKED - for (i = 0; !Fake && i < ActiveThreads && workersCnt < MaxThreadsPerSplitPoint; i++) + for (i = 0; !Fake && i < activeThreads && workersCnt < maxThreadsPerSplitPoint; i++) if (thread_is_available(i, master)) { threads[i].state = THREAD_BOOKED; @@ -2653,11 +2614,11 @@ namespace { assert(Fake || workersCnt > 1); // We can release the lock because slave threads are already booked and master is not available - lock_release(&MPLock); + lock_release(&mpLock); // Tell the threads that they have work to do. This will make them leave // their idle loop. But before copy search stack tail for each thread. - for (i = 0; i < ActiveThreads; i++) + for (i = 0; i < activeThreads; i++) if (i == master || splitPoint.slaves[i]) { memcpy(splitPoint.sstack[i], ss - 1, 4 * sizeof(SearchStack)); @@ -2665,6 +2626,9 @@ namespace { assert(i == master || threads[i].state == THREAD_BOOKED); threads[i].state = THREAD_WORKISWAITING; // This makes the slave to exit from idle_loop() + + if (useSleepingThreads && i != master) + wake_sleeping_thread(i); } // Everything is set up. The master thread enters the idle loop, from @@ -2676,54 +2640,29 @@ namespace { // We have returned from the idle loop, which means that all threads are // finished. Update alpha and bestValue, and return. - lock_grab(&MPLock); + lock_grab(&mpLock); *alpha = splitPoint.alpha; *bestValue = splitPoint.bestValue; masterThread.activeSplitPoints--; masterThread.splitPoint = splitPoint.parent; + pos.set_nodes_searched(pos.nodes_searched() + splitPoint.nodes); - lock_release(&MPLock); + lock_release(&mpLock); } - // wake_sleeping_threads() wakes up all sleeping threads when it is time - // to start a new search from the root. + // wake_sleeping_thread() wakes up the thread with the given threadID + // when it is time to start a new search. - void ThreadsManager::wake_sleeping_threads() { - - assert(AllThreadsShouldSleep); - assert(ActiveThreads > 0); - - AllThreadsShouldSleep = false; - - if (ActiveThreads == 1) - return; - -#if !defined(_MSC_VER) - pthread_mutex_lock(&WaitLock); - pthread_cond_broadcast(&WaitCond); - pthread_mutex_unlock(&WaitLock); -#else - for (int i = 1; i < MAX_THREADS; i++) - SetEvent(SitIdleEvent[i]); -#endif + void ThreadsManager::wake_sleeping_thread(int threadID) { + lock_grab(&sleepLock[threadID]); + cond_signal(&sleepCond[threadID]); + lock_release(&sleepLock[threadID]); } - // put_threads_to_sleep() makes all the threads go to sleep just before - // to leave think(), at the end of the search. Threads should have already - // finished the job and should be idle. - - void ThreadsManager::put_threads_to_sleep() { - - assert(!AllThreadsShouldSleep); - - // This makes the threads to go to sleep - AllThreadsShouldSleep = true; - } - /// The RootMoveList class // RootMoveList c'tor @@ -2737,7 +2676,7 @@ namespace { // Initialize search stack init_ss_array(ss, PLY_MAX_PLUS_2); - ss[0].eval = VALUE_NONE; + ss[0].eval = ss[0].evalMargin = VALUE_NONE; count = 0; // Generate all legal moves