X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Ftt.cpp;h=e8a316c96b7ae5831cb32737ec0df38eb72300b8;hp=49dae31d70ffb5b634709e339c7815d8d5e5184d;hb=b89733b46c9267f61ac41c48204557483658435b;hpb=6f1475b6fcd3e3728d800e622ab7a22265fb8ca4 diff --git a/src/tt.cpp b/src/tt.cpp index 49dae31d..e8a316c9 100644 --- a/src/tt.cpp +++ b/src/tt.cpp @@ -1,7 +1,7 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2009 Marco Costalba + Copyright (C) 2008-2010 Marco Costalba, Joona Kiiski, Tord Romstad Stockfish is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by @@ -25,14 +25,15 @@ #include #include #include +#if defined(__i386__) || defined(_M_IX86) || defined(_X86_) +# include +#endif #include "movegen.h" #include "tt.h" - -/// This is the number of TTEntry slots for each position -static const int ClusterSize = 4; - +// The main transposition table +TranspositionTable TT; //// //// Functions @@ -54,22 +55,20 @@ TranspositionTable::~TranspositionTable() { /// TranspositionTable::set_size sets the size of the transposition table, /// measured in megabytes. -void TranspositionTable::set_size(unsigned mbSize) { - - assert(mbSize >= 4 && mbSize <= 4096); +void TranspositionTable::set_size(size_t mbSize) { - unsigned newSize = 1024; + size_t 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) * ClusterSize * (sizeof(TTEntry)) <= (mbSize << 20)) + while ((2 * newSize) * sizeof(TTCluster) <= (mbSize << 20)) newSize *= 2; if (newSize != size) { size = newSize; delete [] entries; - entries = new TTEntry[size * ClusterSize]; + entries = new TTCluster[size]; if (!entries) { std::cerr << "Failed to allocate " << mbSize @@ -88,7 +87,17 @@ void TranspositionTable::set_size(unsigned mbSize) { void TranspositionTable::clear() { - memset(entries, 0, size * ClusterSize * 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; } @@ -112,10 +121,7 @@ void TranspositionTable::store(const Key posKey, Value v, ValueType t, Depth d, { if (!tte->key() || tte->key() == posKey32) // empty or overwrite old { - // Do not overwrite when new type is VALUE_TYPE_EVAL - if (tte->key() && t == VALUE_TYPE_EVAL) - return; - + // Preserve any exsisting ttMove if (m == MOVE_NONE) m = tte->move(); @@ -144,7 +150,7 @@ void TranspositionTable::store(const Key posKey, Value v, ValueType t, Depth d, TTEntry* TranspositionTable::retrieve(const Key posKey) const { uint32_t posKey32 = posKey >> 32; - TTEntry *tte = first_entry(posKey); + TTEntry* tte = first_entry(posKey); for (int i = 0; i < ClusterSize; i++, tte++) if (tte->key() == posKey32) @@ -154,15 +160,31 @@ TTEntry* TranspositionTable::retrieve(const Key posKey) const { } -/// 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. +/// 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(__i386__) || defined(_M_IX86) || defined(_X86_)) +void TranspositionTable::prefetch(const Key) const {} +#else -inline TTEntry* TranspositionTable::first_entry(const Key posKey) const { +void TranspositionTable::prefetch(const Key posKey) 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 + ((uint32_t(posKey) & (size - 1)) * ClusterSize); + 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 /// distinguish transposition table entries from previous searches from @@ -187,7 +209,9 @@ void TranspositionTable::insert_pv(const Position& pos, Move pv[]) { for (int i = 0; pv[i] != MOVE_NONE; i++) { - store(p.get_key(), VALUE_NONE, VALUE_TYPE_NONE, Depth(-127*OnePly), pv[i]); + 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]); p.do_move(pv[i], st); } } @@ -199,27 +223,26 @@ void TranspositionTable::insert_pv(const Position& pos, Move pv[]) { /// 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. -void TranspositionTable::extract_pv(const Position& pos, Move pv[]) { - - int ply; - Position p(pos); - StateInfo st[100]; - - for (ply = 0; pv[ply] != MOVE_NONE; ply++) - p.do_move(pv[ply], st[ply]); +void TranspositionTable::extract_pv(const Position& pos, Move pv[], const int PLY_MAX) { - bool stop; const TTEntry* tte; - for (stop = false, tte = retrieve(p.get_key()); - tte && tte->move() != MOVE_NONE && !stop; - tte = retrieve(p.get_key()), ply++) + StateInfo st; + Position p(pos); + int ply = 0; + + // Update position to the end of current PV + while (pv[ply] != MOVE_NONE) + p.do_move(pv[ply++], st); + + // 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) { - if (!move_is_legal(p, tte->move())) - break; pv[ply] = tte->move(); - p.do_move(pv[ply], st[ply]); - for (int j = 0; j < ply; j++) - if (st[j].key == p.get_key()) stop = true; + p.do_move(pv[ply++], st); } pv[ply] = MOVE_NONE; }