X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fsearch.cpp;h=a50b639ae49283dbec0b2f3e6fb2aa4eb768602d;hp=e6903f1d69cab6d053d999ebeda1e145c3f102e6;hb=09d01ee9dc16f20962d5f5ffcb61ade56d4fa579;hpb=ccd5ccbcdbfd7bc1521fe194a970ab1c6a4b55fe diff --git a/src/search.cpp b/src/search.cpp index e6903f1d..a50b639a 100644 --- a/src/search.cpp +++ b/src/search.cpp @@ -32,7 +32,6 @@ #include "move.h" #include "movegen.h" #include "movepick.h" -#include "lock.h" #include "search.h" #include "timeman.h" #include "thread.h" @@ -54,44 +53,6 @@ namespace { 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 like init, - // starting, parking and, the most important, launching a slave thread at a split - // point. All the access to shared thread data is done through this class. - - class ThreadsManager { - /* As long as the single ThreadsManager object is defined as a global we don't - need to explicitly initialize to zero its data members because variables with - static storage duration are automatically set to zero before enter main() - */ - public: - Thread& operator[](int threadID) { return threads[threadID]; } - void init_threads(); - void exit_threads(); - - int min_split_depth() const { return minimumSplitDepth; } - int active_threads() const { return activeThreads; } - void set_active_threads(int cnt) { activeThreads = cnt; } - - void read_uci_options(); - bool available_thread_exists(int master) const; - bool thread_is_available(int slave, int master) const; - bool cutoff_at_splitpoint(int threadID) const; - void idle_loop(int threadID, SplitPoint* sp); - - template - void split(Position& pos, SearchStack* ss, Value* alpha, const Value beta, Value* bestValue, - Depth depth, Move threatMove, int moveCount, MovePicker* mp, bool pvNode); - private: - Lock mpLock; - Depth minimumSplitDepth; - int maxThreadsPerSplitPoint; - bool useSleepingThreads; - int activeThreads; - volatile bool allThreadsShouldExit; - Thread threads[MAX_THREADS]; - }; - - // RootMove struct is used for moves at the root of the tree. For each root // move, we store two scores, a node count, and a PV (really a refutation // in the case of moves which fail low). Value pv_score is normally set at @@ -203,27 +164,29 @@ namespace { const Value FutilityMarginQS = Value(0x80); // Futility lookup tables (initialized at startup) and their access functions - Value FutilityMarginsMatrix[16][64]; // [depth][moveNumber] - int FutilityMoveCountArray[32]; // [depth] + Value FutilityMargins[16][64]; // [depth][moveNumber] + int FutilityMoveCounts[32]; // [depth] inline Value futility_margin(Depth d, int mn) { - return d < 7 * ONE_PLY ? FutilityMarginsMatrix[Max(d, 1)][Min(mn, 63)] - : 2 * VALUE_INFINITE; + return d < 7 * ONE_PLY ? FutilityMargins[Max(d, 1)][Min(mn, 63)] + : 2 * VALUE_INFINITE; } inline int futility_move_count(Depth d) { - return d < 16 * ONE_PLY ? FutilityMoveCountArray[d] : MOVES_MAX; + return d < 16 * ONE_PLY ? FutilityMoveCounts[d] : MAX_MOVES; } // Step 14. Reduced search // Reduction lookup tables (initialized at startup) and their access function - int8_t ReductionMatrix[2][64][64]; // [pv][depth][moveNumber] + int8_t Reductions[2][64][64]; // [pv][depth][moveNumber] - template - inline Depth reduction(Depth d, int mn) { return (Depth) ReductionMatrix[PV][Min(d / ONE_PLY, 63)][Min(mn, 63)]; } + template inline Depth reduction(Depth d, int mn) { + + return (Depth) Reductions[PV][Min(d / ONE_PLY, 63)][Min(mn, 63)]; + } // Easy move margin. An easy move candidate must be at least this much // better than the second best move. @@ -254,9 +217,6 @@ namespace { bool SkillLevelEnabled; RKISS RK; - // Multi-threads manager - ThreadsManager ThreadsMgr; - // Node counters, used only by thread[0] but try to keep in different cache // lines (64 bytes each) from the heavy multi-thread read accessed variables. bool SendSearchedNodes; @@ -304,12 +264,6 @@ namespace { void poll(const Position& pos); void wait_for_stop_or_ponderhit(); -#if !defined(_MSC_VER) - void* init_thread(void* threadID); -#else - DWORD WINAPI init_thread(LPVOID threadID); -#endif - // MovePickerExt is an extended MovePicker class used to choose at compile time // the proper move source according to the type of node. @@ -378,10 +332,9 @@ namespace { } // namespace -/// init_threads() is called during startup. It initializes various lookup tables -/// and creates and launches search threads. +/// init_search() is called during startup to initialize various lookup tables -void init_threads() { +void init_search() { int d; // depth (ONE_PLY == 2) int hd; // half depth (ONE_PLY == 1) @@ -392,33 +345,26 @@ void init_threads() { { 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); + Reductions[PV][hd][mc] = (int8_t) ( pvRed >= 1.0 ? floor( pvRed * int(ONE_PLY)) : 0); + Reductions[NonPV][hd][mc] = (int8_t) (nonPVRed >= 1.0 ? floor(nonPVRed * int(ONE_PLY)) : 0); } // Init futility margins array for (d = 1; d < 16; d++) for (mc = 0; mc < 64; mc++) - FutilityMarginsMatrix[d][mc] = Value(112 * int(log(double(d * d) / 2) / log(2.0) + 1.001) - 8 * mc + 45); + FutilityMargins[d][mc] = Value(112 * int(log(double(d * d) / 2) / log(2.0) + 1.001) - 8 * mc + 45); // Init futility move count array for (d = 0; d < 32; d++) - FutilityMoveCountArray[d] = int(3.001 + 0.25 * pow(d, 2.0)); - - // Create and startup threads - ThreadsMgr.init_threads(); + FutilityMoveCounts[d] = int(3.001 + 0.25 * pow(d, 2.0)); } -/// exit_threads() is a trampoline to access ThreadsMgr from outside of current file -void exit_threads() { ThreadsMgr.exit_threads(); } - - /// perft() is our utility to verify move generation. All the legal moves up to /// given depth are generated and counted and the sum returned. int64_t perft(Position& pos, Depth depth) { - MoveStack mlist[MOVES_MAX]; + MoveStack mlist[MAX_MOVES]; StateInfo st; Move m; int64_t sum = 0; @@ -468,8 +414,8 @@ bool think(Position& pos, const SearchLimits& limits, Move searchMoves[]) { else NodesBetweenPolls = 30000; - // Look for a book move, only during games, not tests - if (Limits.useTimeManagement() && Options["OwnBook"].value()) + // Look for a book move + if (Options["OwnBook"].value()) { if (Options["Book File"].value() != OpeningBook.name()) OpeningBook.open(Options["Book File"].value()); @@ -490,28 +436,28 @@ bool think(Position& pos, const SearchLimits& limits, Move searchMoves[]) { SkillLevel = Options["Skill level"].value(); read_evaluation_uci_options(pos.side_to_move()); + Threads.read_uci_options(); + + // If needed allocate pawn and material hash tables and adjust TT size + Threads.init_hash_tables(); + TT.set_size(Options["Hash"].value()); if (Options["Clear Hash"].value()) { Options["Clear Hash"].set_value("false"); TT.clear(); } - TT.set_size(Options["Hash"].value()); // Do we have to play with skill handicap? In this case enable MultiPV that // we will use behind the scenes to retrieve a set of possible moves. SkillLevelEnabled = (SkillLevel < 20); MultiPV = (SkillLevelEnabled ? Max(UCIMultiPV, 4) : UCIMultiPV); - // Set the number of active threads - ThreadsMgr.read_uci_options(); - init_eval(ThreadsMgr.active_threads()); - // Wake up needed threads and reset maxPly counter - for (int i = 0; i < ThreadsMgr.active_threads(); i++) + for (int i = 0; i < Threads.size(); i++) { - ThreadsMgr[i].wake_up(); - ThreadsMgr[i].maxPly = 0; + Threads[i].wake_up(); + Threads[i].maxPly = 0; } // Write to log file and keep it open to be accessed during the search @@ -553,7 +499,7 @@ bool think(Position& pos, const SearchLimits& limits, Move searchMoves[]) { } // This makes all the threads to go to sleep - ThreadsMgr.set_active_threads(1); + Threads.set_size(1); // If we are pondering or in infinite search, we shouldn't print the // best move before we are told to do so. @@ -679,9 +625,9 @@ namespace { // Retrieve max searched depth among threads selDepth = 0; - for (int i = 0; i < ThreadsMgr.active_threads(); i++) - if (ThreadsMgr[i].maxPly > selDepth) - selDepth = ThreadsMgr[i].maxPly; + for (int i = 0; i < Threads.size(); i++) + if (Threads[i].maxPly > selDepth) + selDepth = Threads[i].maxPly; // Send PV line to GUI and to log file for (int i = 0; i < Min(UCIMultiPV, (int)Rml.size()); i++) @@ -762,9 +708,9 @@ namespace { assert(alpha >= -VALUE_INFINITE && alpha <= VALUE_INFINITE); assert(beta > alpha && beta <= VALUE_INFINITE); assert(PvNode || alpha == beta - 1); - assert(pos.thread() >= 0 && pos.thread() < ThreadsMgr.active_threads()); + assert(pos.thread() >= 0 && pos.thread() < Threads.size()); - Move movesSearched[MOVES_MAX]; + Move movesSearched[MAX_MOVES]; int64_t nodes; StateInfo st; const TTEntry *tte; @@ -785,8 +731,8 @@ namespace { ss->ply = (ss-1)->ply + 1; // Used to send selDepth info to GUI - if (PvNode && ThreadsMgr[threadID].maxPly < ss->ply) - ThreadsMgr[threadID].maxPly = ss->ply; + if (PvNode && Threads[threadID].maxPly < ss->ply) + Threads[threadID].maxPly = ss->ply; if (SpNode) { @@ -812,7 +758,7 @@ namespace { // Step 2. Check for aborted search and immediate draw if (( StopRequest - || ThreadsMgr.cutoff_at_splitpoint(threadID) + || Threads[threadID].cutoff_occurred() || pos.is_draw() || ss->ply > PLY_MAX) && !Root) return VALUE_DRAW; @@ -993,7 +939,7 @@ split_point_start: // At split points actual search starts from here // 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.cutoff_at_splitpoint(threadID)) + && !Threads[threadID].cutoff_occurred()) { assert(move_is_ok(move)); @@ -1208,7 +1154,7 @@ split_point_start: // At split points actual search starts from here alpha = sp->alpha; } - if (value > bestValue && !(SpNode && ThreadsMgr.cutoff_at_splitpoint(threadID))) + if (value > bestValue && !(SpNode && Threads[threadID].cutoff_occurred())) { bestValue = value; @@ -1225,7 +1171,7 @@ split_point_start: // At split points actual search starts from here sp->alpha = value; } else if (SpNode) - sp->betaCutoff = true; + sp->is_betaCutoff = true; if (value == value_mate_in(ss->ply + 1)) ss->mateKiller = move; @@ -1281,14 +1227,13 @@ split_point_start: // At split points actual search starts from here // Step 18. Check for split if ( !Root && !SpNode - && depth >= ThreadsMgr.min_split_depth() - && ThreadsMgr.active_threads() > 1 + && depth >= Threads.min_split_depth() && bestValue < beta - && ThreadsMgr.available_thread_exists(threadID) + && Threads.available_slave_exists(threadID) && !StopRequest - && !ThreadsMgr.cutoff_at_splitpoint(threadID)) - ThreadsMgr.split(pos, ss, &alpha, beta, &bestValue, depth, - threatMove, moveCount, &mp, PvNode); + && !Threads[threadID].cutoff_occurred()) + Threads.split(pos, ss, &alpha, beta, &bestValue, depth, + threatMove, moveCount, &mp, PvNode); } // Step 19. Check for mate and stalemate @@ -1301,7 +1246,7 @@ split_point_start: // At split points actual search starts from here // Step 20. Update tables // If the search is not aborted, update the transposition table, // history counters, and killer moves. - if (!SpNode && !StopRequest && !ThreadsMgr.cutoff_at_splitpoint(threadID)) + if (!SpNode && !StopRequest && !Threads[threadID].cutoff_occurred()) { move = bestValue <= oldAlpha ? MOVE_NONE : ss->bestMove; vt = bestValue <= oldAlpha ? VALUE_TYPE_UPPER @@ -1325,7 +1270,7 @@ split_point_start: // At split points actual search starts from here if (SpNode) { // Here we have the lock still grabbed - sp->slaves[threadID] = 0; + sp->is_slave[threadID] = false; sp->nodes += pos.nodes_searched(); lock_release(&(sp->lock)); } @@ -1346,7 +1291,7 @@ split_point_start: // At split points actual search starts from here assert(beta >= -VALUE_INFINITE && beta <= VALUE_INFINITE); assert(PvNode || alpha == beta - 1); assert(depth <= 0); - assert(pos.thread() >= 0 && pos.thread() < ThreadsMgr.active_threads()); + assert(pos.thread() >= 0 && pos.thread() < Threads.size()); StateInfo st; Move ttMove, move; @@ -1972,414 +1917,45 @@ split_point_start: // At split points actual search starts from here } - // init_thread() is the function which is called when a new thread is - // launched. It simply calls the idle_loop() function with the supplied - // threadID. There are two versions of this function; one for POSIX - // threads and one for Windows threads. - -#if !defined(_MSC_VER) - - void* init_thread(void* threadID) { - - ThreadsMgr.idle_loop(*(int*)threadID, NULL); - return NULL; - } - -#else - - DWORD WINAPI init_thread(LPVOID threadID) { - - ThreadsMgr.idle_loop(*(int*)threadID, NULL); - return 0; - } - -#endif - - - /// The ThreadsManager class - - - // 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. - - void ThreadsManager::read_uci_options() { - - 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(); - } - - - // idle_loop() is where the threads are parked when they have no work to do. - // The parameter 'sp', if non-NULL, is a pointer to an active SplitPoint - // object for which the current thread is the master. - - void ThreadsManager::idle_loop(int threadID, SplitPoint* sp) { - - assert(threadID >= 0 && threadID < MAX_THREADS); - - int i; - bool allFinished; - - while (true) - { - // Slave threads can exit as soon as AllThreadsShouldExit raises, - // master should exit as last one. - if (allThreadsShouldExit) - { - assert(!sp); - threads[threadID].state = THREAD_TERMINATED; - return; - } - - // 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 - || (useSleepingThreads && threads[threadID].state == THREAD_AVAILABLE)) - { - assert(!sp || useSleepingThreads); - assert(threadID != 0 || useSleepingThreads); - - if (threads[threadID].state == THREAD_INITIALIZING) - threads[threadID].state = THREAD_AVAILABLE; - - // Grab the lock to avoid races with Thread::wake_up() - lock_grab(&threads[threadID].sleepLock); - - // 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(&threads[threadID].sleepLock); - break; - } - - // Do sleep here after retesting sleep conditions - if (threadID >= activeThreads || threads[threadID].state == THREAD_AVAILABLE) - cond_wait(&threads[threadID].sleepCond, &threads[threadID].sleepLock); - - lock_release(&threads[threadID].sleepLock); - } - - // If this thread has been assigned work, launch a search - if (threads[threadID].state == THREAD_WORKISWAITING) - { - assert(!allThreadsShouldExit); - - threads[threadID].state = THREAD_SEARCHING; - - // Copy split point position and search stack and call search() - // with SplitPoint template parameter set to true. - SearchStack ss[PLY_MAX_PLUS_2]; - SplitPoint* tsp = threads[threadID].splitPoint; - Position pos(*tsp->pos, threadID); - - memcpy(ss, tsp->ss - 1, 4 * sizeof(SearchStack)); - (ss+1)->sp = tsp; - - if (tsp->pvNode) - search(pos, ss+1, tsp->alpha, tsp->beta, tsp->depth); - else - search(pos, ss+1, tsp->alpha, tsp->beta, tsp->depth); - - 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) - threads[tsp->master].wake_up(); - } - - // 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. - for (i = 0; sp && i < activeThreads && !sp->slaves[i]; i++) {} - allFinished = (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; - return; - } - } - } - - - // init_threads() is called during startup. Initializes locks and condition - // variables and launches all threads sending them immediately to sleep. - - void ThreadsManager::init_threads() { - - int i, arg[MAX_THREADS]; - bool ok; - - // This flag is needed to properly end the threads when program exits - allThreadsShouldExit = false; - - // Threads will sent to sleep as soon as created, only main thread is kept alive - activeThreads = 1; - - lock_init(&mpLock); + // When playing with strength handicap choose best move among the MultiPV set + // using a statistical rule dependent on SkillLevel. Idea by Heinz van Saanen. + void do_skill_level(Move* best, Move* ponder) { - for (i = 0; i < MAX_THREADS; i++) - { - // Initialize thread and split point locks - lock_init(&threads[i].sleepLock); - cond_init(&threads[i].sleepCond); + assert(MultiPV > 1); - for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_init(&(threads[i].splitPoints[j].lock)); + // Rml list is already sorted by pv_score in descending order + int s; + int max_s = -VALUE_INFINITE; + int size = Min(MultiPV, (int)Rml.size()); + int max = Rml[0].pv_score; + int var = Min(max - Rml[size - 1].pv_score, PawnValueMidgame); + int wk = 120 - 2 * SkillLevel; - // All threads but first should be set to THREAD_INITIALIZING - threads[i].state = (i == 0 ? THREAD_SEARCHING : THREAD_INITIALIZING); - } + // PRNG sequence should be non deterministic + for (int i = abs(get_system_time() % 50); i > 0; i--) + RK.rand(); - // Create and startup the threads - for (i = 1; i < MAX_THREADS; i++) + // Choose best move. For each move's score we add two terms both dependent + // on wk, one deterministic and bigger for weaker moves, and one random, + // then we choose the move with the resulting highest score. + for (int i = 0; i < size; i++) { - arg[i] = i; - -#if !defined(_MSC_VER) - 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)(&arg[i]), 0, NULL) != NULL); -#endif - if (!ok) - { - cout << "Failed to create thread number " << i << endl; - exit(EXIT_FAILURE); - } - - // Wait until the thread has finished launching and is gone to sleep - while (threads[i].state == THREAD_INITIALIZING) {} - } - } - - - // exit_threads() is called when the program exits. It makes all the - // helper threads exit cleanly. + s = Rml[i].pv_score; - void ThreadsManager::exit_threads() { + // Don't allow crazy blunders even at very low skills + if (i > 0 && Rml[i-1].pv_score > s + EasyMoveMargin) + break; - // Force the woken up threads to exit idle_loop() and hence terminate - allThreadsShouldExit = true; + // This is our magical formula + s += ((max - s) * wk + var * (RK.rand() % wk)) / 128; - for (int i = 0; i < MAX_THREADS; i++) - { - // Wake up all the threads and waits for termination - if (i != 0) + if (s > max_s) { - threads[i].wake_up(); - while (threads[i].state != THREAD_TERMINATED) {} + max_s = s; + *best = Rml[i].pv[0]; + *ponder = Rml[i].pv[1]; } - - // Now we can safely destroy the locks and wait conditions - lock_destroy(&threads[i].sleepLock); - cond_destroy(&threads[i].sleepCond); - - for (int j = 0; j < MAX_ACTIVE_SPLIT_POINTS; j++) - lock_destroy(&(threads[i].splitPoints[j].lock)); - } - - lock_destroy(&mpLock); - } - - - // 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::cutoff_at_splitpoint(int threadID) const { - - assert(threadID >= 0 && threadID < activeThreads); - - SplitPoint* sp = threads[threadID].splitPoint; - - for ( ; sp && !sp->betaCutoff; sp = sp->parent) {} - return sp != NULL; - } - - - // thread_is_available() checks whether the thread with threadID "slave" is - // available to help the thread with threadID "master" at a split point. An - // obvious requirement is that "slave" must be idle. With more than two - // threads, this is not by itself sufficient: If "slave" is the master of - // some active split point, it is only available as a slave to the other - // threads which are busy searching the split point at the top of "slave"'s - // split point stack (the "helpful master concept" in YBWC terminology). - - bool ThreadsManager::thread_is_available(int slave, int master) const { - - assert(slave >= 0 && slave < activeThreads); - assert(master >= 0 && master < activeThreads); - assert(activeThreads > 1); - - if (threads[slave].state != THREAD_AVAILABLE || slave == master) - return false; - - // Make a local copy to be sure doesn't change under our feet - int localActiveSplitPoints = threads[slave].activeSplitPoints; - - // 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 - // that is known to be > 0, instead of threads[slave].activeSplitPoints that - // could have been set to 0 by another thread leading to an out of bound access. - if (threads[slave].splitPoints[localActiveSplitPoints - 1].slaves[master]) - return true; - - return false; - } - - - // available_thread_exists() tries to find an idle thread which is available as - // a slave for the thread with threadID "master". - - bool ThreadsManager::available_thread_exists(int master) const { - - assert(master >= 0 && master < activeThreads); - assert(activeThreads > 1); - - for (int i = 0; i < activeThreads; i++) - if (thread_is_available(i, master)) - return true; - - return false; - } - - - // split() does the actual work of distributing the work at a node between - // several available threads. If it does not succeed in splitting the - // node (because no idle threads are available, or because we have no unused - // split point objects), the function immediately returns. 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 search().When all threads have returned from search() then split() returns. - - template - void ThreadsManager::split(Position& pos, SearchStack* ss, Value* alpha, const Value beta, - Value* bestValue, Depth depth, Move threatMove, - int moveCount, MovePicker* mp, bool pvNode) { - assert(pos.is_ok()); - assert(*bestValue >= -VALUE_INFINITE); - assert(*bestValue <= *alpha); - assert(*alpha < beta); - assert(beta <= VALUE_INFINITE); - assert(depth > DEPTH_ZERO); - assert(pos.thread() >= 0 && pos.thread() < activeThreads); - assert(activeThreads > 1); - - int i, master = pos.thread(); - Thread& masterThread = threads[master]; - - 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); - return; } - - // Pick the next available split point object from the split point stack - SplitPoint& splitPoint = masterThread.splitPoints[masterThread.activeSplitPoints++]; - - // Initialize the split point object - splitPoint.parent = masterThread.splitPoint; - splitPoint.master = master; - splitPoint.betaCutoff = false; - splitPoint.depth = depth; - splitPoint.threatMove = threatMove; - splitPoint.alpha = *alpha; - splitPoint.beta = beta; - splitPoint.pvNode = pvNode; - splitPoint.bestValue = *bestValue; - splitPoint.mp = mp; - splitPoint.moveCount = moveCount; - splitPoint.pos = &pos; - splitPoint.nodes = 0; - splitPoint.ss = ss; - for (i = 0; i < activeThreads; i++) - splitPoint.slaves[i] = 0; - - masterThread.splitPoint = &splitPoint; - - // If we are here it means we are not available - assert(masterThread.state != THREAD_AVAILABLE); - - 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++) - if (thread_is_available(i, master)) - { - threads[i].state = THREAD_BOOKED; - threads[i].splitPoint = &splitPoint; - splitPoint.slaves[i] = 1; - workersCnt++; - } - - assert(Fake || workersCnt > 1); - - // We can release the lock because slave threads are already booked and master is not available - lock_release(&mpLock); - - // Tell the threads that they have work to do. This will make them leave - // their idle loop. - for (i = 0; i < activeThreads; i++) - if (i == master || splitPoint.slaves[i]) - { - 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) - threads[i].wake_up(); - } - - // Everything is set up. The master thread enters the idle loop, from - // which it will instantly launch a search, because its state is - // THREAD_WORKISWAITING. We send the split point as a second parameter to the - // idle loop, which means that the main thread will return from the idle - // loop when all threads have finished their work at this split point. - idle_loop(master, &splitPoint); - - // We have returned from the idle loop, which means that all threads are - // finished. Update alpha and bestValue, and return. - 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); } @@ -2406,6 +1982,33 @@ split_point_start: // At split points actual search starts from here return *this; } + void RootMoveList::init(Position& pos, Move searchMoves[]) { + + MoveStack mlist[MAX_MOVES]; + Move* sm; + + clear(); + bestMoveChanges = 0; + + // Generate all legal moves and add them to RootMoveList + MoveStack* last = generate(pos, mlist); + for (MoveStack* cur = mlist; cur != last; cur++) + { + // If we have a searchMoves[] list then verify cur->move + // is in the list before to add it. + for (sm = searchMoves; *sm && *sm != cur->move; sm++) {} + + if (searchMoves[0] && *sm != cur->move) + continue; + + RootMove rm; + rm.pv[0] = cur->move; + rm.pv[1] = MOVE_NONE; + rm.pv_score = -VALUE_INFINITE; + push_back(rm); + } + } + // extract_pv_from_tt() builds a PV by adding moves from the transposition table. // We consider also failing high nodes and not only VALUE_TYPE_EXACT nodes. This // allow to always have a ponder move even when we fail high at root and also a @@ -2487,74 +2090,114 @@ split_point_start: // At split points actual search starts from here return s.str(); } +} // namespace - void RootMoveList::init(Position& pos, Move searchMoves[]) { - MoveStack mlist[MOVES_MAX]; - Move* sm; +// ThreadsManager::idle_loop() is where the threads are parked when they have no work +// to do. The parameter 'sp', if non-NULL, is a pointer to an active SplitPoint +// object for which the current thread is the master. - clear(); - bestMoveChanges = 0; +void ThreadsManager::idle_loop(int threadID, SplitPoint* sp) { - // Generate all legal moves and add them to RootMoveList - MoveStack* last = generate(pos, mlist); - for (MoveStack* cur = mlist; cur != last; cur++) - { - // If we have a searchMoves[] list then verify cur->move - // is in the list before to add it. - for (sm = searchMoves; *sm && *sm != cur->move; sm++) {} + assert(threadID >= 0 && threadID < MAX_THREADS); - if (searchMoves[0] && *sm != cur->move) - continue; + int i; + bool allFinished; - RootMove rm; - rm.pv[0] = cur->move; - rm.pv[1] = MOVE_NONE; - rm.pv_score = -VALUE_INFINITE; - push_back(rm); - } - } + while (true) + { + // Slave threads can exit as soon as AllThreadsShouldExit raises, + // master should exit as last one. + if (allThreadsShouldExit) + { + assert(!sp); + threads[threadID].state = Thread::TERMINATED; + return; + } + // 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 + || (useSleepingThreads && threads[threadID].state == Thread::AVAILABLE)) + { + assert(!sp || useSleepingThreads); + assert(threadID != 0 || useSleepingThreads); - // When playing with strength handicap choose best move among the MultiPV set - // using a statistical rule dependent on SkillLevel. Idea by Heinz van Saanen. - void do_skill_level(Move* best, Move* ponder) { + if (threads[threadID].state == Thread::INITIALIZING) + threads[threadID].state = Thread::AVAILABLE; - assert(MultiPV > 1); + // Grab the lock to avoid races with Thread::wake_up() + lock_grab(&threads[threadID].sleepLock); - // Rml list is already sorted by pv_score in descending order - int s; - int max_s = -VALUE_INFINITE; - int size = Min(MultiPV, (int)Rml.size()); - int max = Rml[0].pv_score; - int var = Min(max - Rml[size - 1].pv_score, PawnValueMidgame); - int wk = 120 - 2 * SkillLevel; + // If we are master and all slaves have finished do not go to sleep + for (i = 0; sp && i < activeThreads && !sp->is_slave[i]; i++) {} + allFinished = (i == activeThreads); - // PRNG sequence should be non deterministic - for (int i = abs(get_system_time() % 50); i > 0; i--) - RK.rand(); + if (allFinished || allThreadsShouldExit) + { + lock_release(&threads[threadID].sleepLock); + break; + } - // Choose best move. For each move's score we add two terms both dependent - // on wk, one deterministic and bigger for weaker moves, and one random, - // then we choose the move with the resulting highest score. - for (int i = 0; i < size; i++) - { - s = Rml[i].pv_score; + // Do sleep here after retesting sleep conditions + if (threadID >= activeThreads || threads[threadID].state == Thread::AVAILABLE) + cond_wait(&threads[threadID].sleepCond, &threads[threadID].sleepLock); - // Don't allow crazy blunders even at very low skills - if (i > 0 && Rml[i-1].pv_score > s + EasyMoveMargin) - break; + lock_release(&threads[threadID].sleepLock); + } - // This is our magical formula - s += ((max - s) * wk + var * (RK.rand() % wk)) / 128; + // If this thread has been assigned work, launch a search + if (threads[threadID].state == Thread::WORKISWAITING) + { + assert(!allThreadsShouldExit); - if (s > max_s) - { - max_s = s; - *best = Rml[i].pv[0]; - *ponder = Rml[i].pv[1]; - } - } - } + threads[threadID].state = Thread::SEARCHING; -} // namespace + // Copy split point position and search stack and call search() + // with SplitPoint template parameter set to true. + SearchStack ss[PLY_MAX_PLUS_2]; + SplitPoint* tsp = threads[threadID].splitPoint; + Position pos(*tsp->pos, threadID); + + memcpy(ss, tsp->ss - 1, 4 * sizeof(SearchStack)); + (ss+1)->sp = tsp; + + if (tsp->pvNode) + search(pos, ss+1, tsp->alpha, tsp->beta, tsp->depth); + else + search(pos, ss+1, tsp->alpha, tsp->beta, tsp->depth); + + 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) + threads[tsp->master].wake_up(); + } + + // 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. + for (i = 0; sp && i < activeThreads && !sp->is_slave[i]; i++) {} + allFinished = (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; + return; + } + } +}