X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Ftt.cpp;h=16331b8762ae04c60689a341f46af7c3ad49e9c9;hp=268c21f896ab3c89e962ae6dde36c9001030231b;hb=2f6927ac08887ff3b709cfe9228b27a85bdd1d88;hpb=c2c0ba875f429e497c936b61be9f75dcc88385a9 diff --git a/src/tt.cpp b/src/tt.cpp index 268c21f8..16331b87 100644 --- a/src/tt.cpp +++ b/src/tt.cpp @@ -1,13 +1,14 @@ /* - Glaurung, a UCI chess playing engine. - Copyright (C) 2004-2008 Tord Romstad + Stockfish, a UCI chess playing engine derived from Glaurung 2.1 + Copyright (C) 2004-2008 Tord Romstad (Glaurung author) + Copyright (C) 2008-2010 Marco Costalba, Joona Kiiski, Tord Romstad - Glaurung is free software: you can redistribute it and/or modify + Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. - Glaurung is distributed in the hope that it will be useful, + Stockfish is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. @@ -23,28 +24,28 @@ #include #include +#include +#if !defined(NO_PREFETCH) +# include +#endif +#include "movegen.h" #include "tt.h" +// The main transposition table +TranspositionTable TT; //// //// Functions //// -/// Constructor +TranspositionTable::TranspositionTable() { -TranspositionTable::TranspositionTable(unsigned mbSize) { - - size = 0; - generation = 0; - writes = 0; + size = writes = 0; entries = 0; - set_size(mbSize); + generation = 0; } - -/// Destructor - TranspositionTable::~TranspositionTable() { delete [] entries; @@ -54,124 +55,138 @@ TranspositionTable::~TranspositionTable() { /// TranspositionTable::set_size sets the size of the transposition table, /// measured in megabytes. -void TranspositionTable::set_size(unsigned mbSize) { +void TranspositionTable::set_size(size_t mbSize) { - assert(mbSize >= 4 && mbSize <= 1024); + size_t newSize = 1024; - unsigned newSize = 1024; + // We store a cluster of ClusterSize number of TTEntry for each position + // and newSize is the maximum number of storable positions. + while ((2 * newSize) * sizeof(TTCluster) <= (mbSize << 20)) + newSize *= 2; - // We store a cluster of 4 TTEntry for each position and newSize is - // the maximum number of storable positions - for ( ; newSize * 4 * (sizeof(TTEntry)) <= (mbSize << 20); newSize *= 2); - newSize /= 2; if (newSize != size) { - size = newSize; - delete [] entries; - entries = new TTEntry[size * 4]; - if (!entries) - { - std::cerr << "Failed to allocate " << mbSize - << " MB for transposition table." - << std::endl; - exit(EXIT_FAILURE); - } - clear(); + size = newSize; + delete [] entries; + entries = new TTCluster[size]; + if (!entries) + { + std::cerr << "Failed to allocate " << mbSize + << " MB for transposition table." << std::endl; + Application::exit_with_failure(); + } + clear(); } } /// TranspositionTable::clear overwrites the entire transposition table -/// with zeroes. It is called whenever the table is resized, or when the +/// with zeroes. It is called whenever the table is resized, or when the /// user asks the program to clear the table (from the UCI interface). /// Perhaps we should also clear it when the "ucinewgame" command is recieved? void TranspositionTable::clear() { - memset(entries, 0, size * 4 * sizeof(TTEntry)); + memset(entries, 0, size * sizeof(TTCluster)); +} + + +/// TranspositionTable::first_entry returns a pointer to the first +/// entry of a cluster given a position. The low 32 bits of the key +/// are used to get the index in the table. + +inline TTEntry* TranspositionTable::first_entry(const Key posKey) const { + + return entries[uint32_t(posKey) & (size - 1)].data; } /// TranspositionTable::store writes a new entry containing a position, /// a value, a value type, a search depth, and a best move to the -/// transposition table. The transposition table is organized in clusters -/// of four TTEntry objects, and when a new entry is written, it replaces -/// the least valuable of the four entries in a cluster. A TTEntry t1 is +/// transposition table. Transposition table is organized in clusters of +/// four TTEntry objects, and when a new entry is written, it replaces +/// the least valuable of the four entries in a cluster. A TTEntry t1 is /// considered to be more valuable than a TTEntry t2 if t1 is from the /// current search and t2 is from a previous search, or if the depth of t1 -/// is bigger than the depth of t2. +/// is bigger than the depth of t2. A TTEntry of type VALUE_TYPE_EVAL +/// never replaces another entry for the same position. + +void TranspositionTable::store(const Key posKey, Value v, ValueType t, Depth d, Move m, Value statV, Value kingD) { -void TranspositionTable::store(const Position &pos, Value v, Depth d, - Move m, ValueType type) { TTEntry *tte, *replace; + uint32_t posKey32 = posKey >> 32; // Use the high 32 bits as key - tte = replace = first_entry(pos); - for (int i = 0; i < 4; i++) + tte = replace = first_entry(posKey); + for (int i = 0; i < ClusterSize; i++, tte++) { - if (!(tte+i)->key()) // still empty - { - *(tte+i) = TTEntry(pos.get_key(), v, type, d, m, generation); - writes++; - return; - } - if ((tte+i)->key() == pos.get_key()) // overwrite old - { - if (m == MOVE_NONE) - m = (tte+i)->move(); - - *(tte+i) = TTEntry(pos.get_key(), v, type, d, m, generation); - return; - } - if ( i == 0 // already is (replace == tte+i), common case - || replace->generation() < (tte+i)->generation()) - continue; - - if ( replace->generation() > (tte+i)->generation() - || (tte+i)->depth() < replace->depth()) - replace = tte+i; + if (!tte->key() || tte->key() == posKey32) // empty or overwrite old + { + // Preserve any exsisting ttMove + if (m == MOVE_NONE) + m = tte->move(); + + *tte = TTEntry(posKey32, v, t, d, m, generation, statV, kingD); + return; + } + else if (i == 0) // replace would be a no-op in this common case + continue; + + int c1 = (replace->generation() == generation ? 2 : 0); + int c2 = (tte->generation() == generation ? -2 : 0); + int c3 = (tte->depth() < replace->depth() ? 1 : 0); + + if (c1 + c2 + c3 > 0) + replace = tte; } - *replace = TTEntry(pos.get_key(), v, type, d, m, generation); + *replace = TTEntry(posKey32, v, t, d, m, generation, statV, kingD); writes++; } /// TranspositionTable::retrieve looks up the current position in the -/// transposition table, and extracts the value, value type, depth and -/// best move if the position is found. The return value is true if -/// the position is found, and false if it isn't. +/// transposition table. Returns a pointer to the TTEntry or NULL +/// if position is not found. -bool TranspositionTable::retrieve(const Position &pos, Value *value, - Depth *d, Move *move, - ValueType *type) const { - TTEntry *tte = first_entry(pos); +TTEntry* TranspositionTable::retrieve(const Key posKey) const { - for (int i = 0; i < 4; i++) - { - tte += i; - if (tte->key() == pos.get_key()) - { - *value = tte->value(); - *type = tte->type(); - *d = tte->depth(); - *move = tte->move(); - return true; - } - } - *move = MOVE_NONE; - return false; + uint32_t posKey32 = posKey >> 32; + TTEntry* tte = first_entry(posKey); + + for (int i = 0; i < ClusterSize; i++, tte++) + if (tte->key() == posKey32) + return tte; + + return NULL; } -/// TranspositionTable::first_entry returns a pointer to the first -/// entry of a cluster given a position. +/// TranspositionTable::prefetch looks up the current position in the +/// transposition table and load it in L1/L2 cache. This is a non +/// blocking function and do not stalls the CPU waiting for data +/// to be loaded from RAM, that can be very slow. When we will +/// subsequently call retrieve() the TT data will be already +/// quickly accessible in L1/L2 CPU cache. +#if defined(NO_PREFETCH) +void TranspositionTable::prefetch(const Key) const {} +#else + +void TranspositionTable::prefetch(const Key posKey) const { -inline TTEntry* TranspositionTable::first_entry(const Position &pos) const { +#if defined(__INTEL_COMPILER) || defined(__ICL) + // This hack prevents prefetches to be optimized away by + // Intel compiler. Both MSVC and gcc seems not affected. + __asm__ (""); +#endif - return entries + (int(pos.get_key() & (size - 1)) << 2); + char const* addr = (char*)first_entry(posKey); + _mm_prefetch(addr, _MM_HINT_T2); + _mm_prefetch(addr+64, _MM_HINT_T2); // 64 bytes ahead } +#endif + /// TranspositionTable::new_search() is called at the beginning of every new -/// search. It increments the "generation" variable, which is used to +/// search. It increments the "generation" variable, which is used to /// distinguish transposition table entries from previous searches from /// entries from the current search. @@ -183,67 +198,62 @@ void TranspositionTable::new_search() { /// TranspositionTable::insert_pv() is called at the end of a search -/// iteration, and inserts the PV back into the PV. This makes sure the -/// old PV moves are searched first, even if the old TT entries have been -/// overwritten. +/// iteration, and inserts the PV back into the PV. This makes sure +/// the old PV moves are searched first, even if the old TT entries +/// have been overwritten. -void TranspositionTable::insert_pv(const Position &pos, Move pv[]) { +void TranspositionTable::insert_pv(const Position& pos, Move pv[]) { - UndoInfo u; - Position p(pos); + StateInfo st; + Position p(pos, pos.thread()); for (int i = 0; pv[i] != MOVE_NONE; i++) { - store(p, VALUE_NONE, Depth(0), pv[i], VALUE_TYPE_NONE); - p.do_move(pv[i], u); + TTEntry *tte = retrieve(p.get_key()); + if (!tte || tte->move() != pv[i]) + store(p.get_key(), VALUE_NONE, VALUE_TYPE_NONE, Depth(-127*OnePly), pv[i], VALUE_NONE, VALUE_NONE); + p.do_move(pv[i], st); } } -/// TranspositionTable::full() returns the permill of all transposition table -/// entries which have received at least one write during the current search. -/// It is used to display the "info hashfull ..." information in UCI. - -int TranspositionTable::full() { +/// TranspositionTable::extract_pv() extends a PV by adding moves from the +/// transposition table at the end. This should ensure that the PV is almost +/// always at least two plies long, which is important, because otherwise we +/// will often get single-move PVs when the search stops while failing high, +/// and a single-move PV means that we don't have a ponder move. - double N = double(size) * 4.0; - return int(1000 * (1 - exp(writes * log(1.0 - 1.0/N)))); -} - - -/// Constructors - -TTEntry::TTEntry() { -} +void TranspositionTable::extract_pv(const Position& pos, Move pv[], const int PLY_MAX) { -TTEntry::TTEntry(Key k, Value v, ValueType t, Depth d, Move m, - int generation) : - key_ (k), data((m & 0x7FFFF) | (t << 20) | (generation << 23)), - value_(v), depth_(int16_t(d)) {} + const TTEntry* tte; + StateInfo st; + Position p(pos, pos.thread()); + int ply = 0; + // Update position to the end of current PV + while (pv[ply] != MOVE_NONE) + p.do_move(pv[ply++], st); -/// Functions for extracting data from TTEntry objects. - -inline Key TTEntry::key() const { - return key_; -} - -inline Depth TTEntry::depth() const { - return Depth(depth_); + // Try to add moves from TT while possible + while ( (tte = retrieve(p.get_key())) != NULL + && tte->move() != MOVE_NONE + && move_is_legal(p, tte->move()) + && (!p.is_draw() || ply < 2) + && ply < PLY_MAX) + { + pv[ply] = tte->move(); + p.do_move(pv[ply++], st); + } + pv[ply] = MOVE_NONE; } -inline Move TTEntry::move() const { - return Move(data & 0x7FFFF); -} -inline Value TTEntry::value() const { - return Value(value_); -} +/// TranspositionTable::full() returns the permill of all transposition table +/// entries which have received at least one write during the current search. +/// It is used to display the "info hashfull ..." information in UCI. -inline ValueType TTEntry::type() const { - return ValueType((data >> 20) & 3); -} +int TranspositionTable::full() const { -inline int TTEntry::generation() const { - return (data >> 23); + double N = double(size) * ClusterSize; + return int(1000 * (1 - exp(writes * log(1.0 - 1.0/N)))); }