X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fsearch.cpp;h=b6493837981ca5394a6101f06c372b7479aa4369;hp=dcfedcdbe0b05afa14e2c3b6221ee4cfff67ad61;hb=556b63b6b6acce9f40952e0f792f84da8101f033;hpb=c416133e2fd1b6bb74fda07cd6050227bbe8994a diff --git a/src/search.cpp b/src/search.cpp index dcfedcdb..b6493837 100644 --- a/src/search.cpp +++ b/src/search.cpp @@ -59,6 +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 @@ -73,12 +77,11 @@ 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; @@ -86,17 +89,18 @@ namespace { void idle_loop(int threadID, SplitPoint* sp); template - void split(const Position& pos, SearchStack* ss, int ply, Value* alpha, const Value beta, Value* bestValue, + 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; + Depth minimumSplitDepth; + int maxThreadsPerSplitPoint; + bool useSleepingThreads; + int activeThreads; + volatile bool allThreadsShouldExit; Thread threads[MAX_THREADS]; - Lock MPLock; - WaitCondition WaitCond[MAX_THREADS]; + Lock mpLock, sleepLock[MAX_THREADS]; + WaitCondition sleepCond[MAX_THREADS]; }; @@ -233,7 +237,10 @@ namespace { const Value EasyMoveMargin = Value(0x200); - /// Global variables + /// Namespace variables + + // Book object + Book OpeningBook; // Iteration counter int Iteration; @@ -258,9 +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 @@ -273,7 +278,7 @@ 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 @@ -305,8 +310,8 @@ 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); @@ -315,7 +320,7 @@ 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 @@ -332,7 +337,6 @@ 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 @@ -346,7 +350,7 @@ 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); @@ -398,13 +402,12 @@ 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; @@ -414,12 +417,12 @@ 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) @@ -431,40 +434,40 @@ 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 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()]; @@ -498,6 +501,9 @@ bool think(const Position& pos, bool infinite, bool ponder, int time[], int incr if (UseLogFile) LogFile.close(); + // This makes all the threads to go to sleep + ThreadsMgr.set_active_threads(1); + return !Quit; } @@ -509,16 +515,15 @@ 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) @@ -531,13 +536,13 @@ 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 @@ -576,11 +581,11 @@ 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! @@ -609,12 +614,11 @@ 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; @@ -648,12 +652,12 @@ 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; @@ -676,14 +680,14 @@ 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); @@ -745,7 +749,7 @@ 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. @@ -866,7 +870,7 @@ 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); @@ -908,8 +912,8 @@ 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++) @@ -968,6 +972,7 @@ namespace { 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; @@ -987,17 +992,16 @@ namespace { 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 @@ -1062,7 +1066,6 @@ 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())) { @@ -1194,16 +1197,17 @@ split_point_start: // At split points actual search starts from here && (move = mp.get_next_move()) != MOVE_NONE && !ThreadsMgr.thread_should_stop(threadID)) { + assert(move_is_ok(move)); + if (SpNode) { moveCount = ++sp->moveCount; lock_release(&(sp->lock)); } - - assert(move_is_ok(move)); - - if (move == excludedMove) + else if (move == excludedMove) continue; + else + movesSearched[moveCount++] = move; moveIsCheck = pos.move_is_check(move, ci); captureOrPromotion = pos.move_is_capture_or_promotion(move); @@ -1235,13 +1239,9 @@ split_point_start: // At split points actual search starts from here } } - newDepth = depth - ONE_PLY + ext; - // Update current move (this must be done after singular extension search) - movesSearched[moveCount] = ss->currentMove = move; - - if (!SpNode) - moveCount++; + ss->currentMove = move; + newDepth = depth - ONE_PLY + ext; // Step 12. Futility pruning (is omitted in PV nodes) if ( !PvNode @@ -1282,6 +1282,17 @@ split_point_start: // At split points actual search starts from here continue; } + + // Prune neg. see moves 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; + } } // Step 13. Make the move @@ -1386,7 +1397,7 @@ split_point_start: // At split points actual search starts from here // Step 18. Check for split if ( !SpNode - && depth >= MinimumSplitDepth + && depth >= ThreadsMgr.min_split_depth() && ThreadsMgr.active_threads() > 1 && bestValue < beta && ThreadsMgr.available_thread_exists(threadID) @@ -1397,37 +1408,39 @@ split_point_start: // At split points actual search starts from here threatMove, mateThreat, moveCount, &mp, PvNode); } - if (SpNode) - { - /* Here we have the lock still grabbed */ - sp->slaves[threadID] = 0; - lock_release(&(sp->lock)); - return bestValue; - } - // 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.thread_should_stop(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, ss->evalMargin); + 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); @@ -1457,7 +1470,6 @@ split_point_start: // At split points actual search starts from here const TTEntry* tte; Value oldAlpha = alpha; - ThreadsMgr.incrementNodeCounter(pos.thread()); ss->bestMove = ss->currentMove = MOVE_NONE; // Check for an instant draw or maximum ply reached @@ -1918,10 +1930,10 @@ split_point_start: // At split points actual search starts from here // 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); } @@ -1929,13 +1941,13 @@ split_point_start: // At split points actual search starts from here // 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; @@ -1978,7 +1990,7 @@ split_point_start: // At split points actual search starts from here 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; } @@ -1995,7 +2007,7 @@ split_point_start: // At split points actual search starts from here if ( (Iteration >= 3 && UseTimeManagement && noMoreTime) || (ExactMaxTime && t >= ExactMaxTime) - || (Iteration >= 3 && MaxNodes && ThreadsMgr.nodes_searched() >= MaxNodes)) + || (Iteration >= 3 && MaxNodes && pos.nodes_searched() >= MaxNodes)) AbortSearch = true; } @@ -2075,8 +2087,8 @@ split_point_start: // At split points actual search starts from here << " 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++) @@ -2089,8 +2101,7 @@ split_point_start: // At split points actual search starts from here 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; } } @@ -2156,7 +2167,7 @@ split_point_start: // At split points actual search starts from here #if !defined(_MSC_VER) - void* init_thread(void *threadID) { + void* init_thread(void* threadID) { ThreadsMgr.idle_loop(*(int*)threadID, NULL); return NULL; @@ -2175,23 +2186,17 @@ split_point_start: // At split points actual search starts from here /// 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; - } + // 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 ThreadsManager::nodes_searched() const { + void ThreadsManager::read_uci_options() { - int64_t result = 0ULL; - for (int i = 0; i < ActiveThreads; i++) - result += threads[i].nodes; - - 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(); } @@ -2203,11 +2208,14 @@ split_point_start: // At split points actual search starts from here 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; @@ -2216,37 +2224,39 @@ split_point_start: // At split points actual search starts from here // If we are not thinking, wait for a condition to be signaled // instead of wasting CPU time polling for work. - while ( threadID >= ActiveThreads - || threads[threadID].state == THREAD_INITIALIZING - || (!sp && threads[threadID].state == THREAD_AVAILABLE)) + while ( threadID >= activeThreads || threads[threadID].state == THREAD_INITIALIZING + || (useSleepingThreads && threads[threadID].state == THREAD_AVAILABLE)) { - assert(!sp); - assert(threadID != 0); + assert(!sp || useSleepingThreads); + assert(threadID != 0 || useSleepingThreads); - if (AllThreadsShouldExit) - break; + if (threads[threadID].state == THREAD_INITIALIZING) + threads[threadID].state = THREAD_AVAILABLE; - lock_grab(&MPLock); + // Grab the lock to avoid races with wake_sleeping_thread() + lock_grab(&sleepLock[threadID]); - // Retest condition under lock protection - if (!( threadID >= ActiveThreads - || threads[threadID].state == THREAD_INITIALIZING - || (!sp && threads[threadID].state == THREAD_AVAILABLE))) + // 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(&MPLock); - continue; + lock_release(&sleepLock[threadID]); + break; } - // Put thread to sleep - threads[threadID].state = THREAD_AVAILABLE; - cond_wait(&WaitCond[threadID], &MPLock); - lock_release(&MPLock); + // 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); + assert(!allThreadsShouldExit); threads[threadID].state = THREAD_SEARCHING; @@ -2258,20 +2268,25 @@ split_point_start: // At split points actual search starts from here if (tsp->pvNode) search(pos, ss, tsp->alpha, tsp->beta, tsp->depth, tsp->ply); - else { + else 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. @@ -2295,14 +2310,17 @@ split_point_start: // At split points actual search starts from here void ThreadsManager::init_threads() { - volatile int i; + int i, arg[MAX_THREADS]; bool ok; // Initialize global locks - lock_init(&MPLock); + lock_init(&mpLock); for (i = 0; i < MAX_THREADS; i++) - cond_init(&WaitCond[i]); + { + lock_init(&sleepLock[i]); + cond_init(&sleepCond[i]); + } // Initialize splitPoints[] locks for (i = 0; i < MAX_THREADS; i++) @@ -2310,10 +2328,10 @@ split_point_start: // At split points actual search starts from here 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 all threads to sleep as soon as created - ActiveThreads = 1; + activeThreads = 1; // All threads except the main thread should be initialized to THREAD_INITIALIZING threads[0].state = THREAD_SEARCHING; @@ -2323,19 +2341,19 @@ split_point_start: // At split points actual search starts from here // Launch the helper threads for (i = 1; i < MAX_THREADS; i++) { + arg[i] = i; #if !defined(_MSC_VER) pthread_t pthread[1]; - ok = (pthread_create(pthread, NULL, init_thread, (void*)(&i)) == 0); + 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 @@ -2349,7 +2367,7 @@ split_point_start: // At split points actual search starts from here void ThreadsManager::exit_threads() { - AllThreadsShouldExit = true; // Let the woken up threads to exit idle_loop() + allThreadsShouldExit = true; // Let the woken up threads to exit idle_loop() // Wake up all the threads and waits for termination for (int i = 1; i < MAX_THREADS; i++) @@ -2363,11 +2381,14 @@ split_point_start: // At split points actual search starts from here for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) lock_destroy(&(threads[i].splitPoints[j].lock)); - lock_destroy(&MPLock); + lock_destroy(&mpLock); // Now we can safely destroy the wait conditions for (int i = 0; i < MAX_THREADS; i++) - cond_destroy(&WaitCond[i]); + { + lock_destroy(&sleepLock[i]); + cond_destroy(&sleepCond[i]); + } } @@ -2377,7 +2398,7 @@ split_point_start: // At split points actual search starts from here bool ThreadsManager::thread_should_stop(int threadID) const { - assert(threadID >= 0 && threadID < ActiveThreads); + assert(threadID >= 0 && threadID < activeThreads); SplitPoint* sp = threads[threadID].splitPoint; @@ -2396,9 +2417,9 @@ split_point_start: // At split points actual search starts from here 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; @@ -2408,7 +2429,7 @@ split_point_start: // At split points actual search starts from here // No active split points means that the thread is available as // a slave for any other thread. - if (localActiveSplitPoints == 0 || ActiveThreads == 2) + if (localActiveSplitPoints == 0 || activeThreads == 2) return true; // Apply the "helpful master" concept if possible. Use localActiveSplitPoints @@ -2426,10 +2447,10 @@ split_point_start: // At split points actual search starts from here 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; @@ -2447,30 +2468,30 @@ split_point_start: // At split points actual search starts from here // 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()); + 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; } @@ -2479,6 +2500,7 @@ split_point_start: // At split points actual search starts from here // Initialize the split point object splitPoint.parent = masterThread.splitPoint; + splitPoint.master = master; splitPoint.stopRequest = false; splitPoint.ply = ply; splitPoint.depth = depth; @@ -2490,9 +2512,10 @@ split_point_start: // At split points actual search starts from here splitPoint.bestValue = *bestValue; splitPoint.mp = mp; splitPoint.moveCount = moveCount; - splitPoint.pos = &p; + 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; @@ -2503,7 +2526,7 @@ split_point_start: // At split points actual search starts from here 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; @@ -2515,11 +2538,11 @@ split_point_start: // At split points actual search starts from here 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)); @@ -2527,7 +2550,8 @@ split_point_start: // At split points actual search starts from here assert(i == master || threads[i].state == THREAD_BOOKED); threads[i].state = THREAD_WORKISWAITING; // This makes the slave to exit from idle_loop() - if (i != master) + + if (useSleepingThreads && i != master) wake_sleeping_thread(i); } @@ -2540,25 +2564,26 @@ split_point_start: // At split points actual search starts from here // 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_thread() 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_thread(int threadID) { - lock_grab(&MPLock); - cond_signal(&WaitCond[threadID]); - lock_release(&MPLock); + lock_grab(&sleepLock[threadID]); + cond_signal(&sleepCond[threadID]); + lock_release(&sleepLock[threadID]); }