X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Ftt.cpp;h=0b4a59de55915962e7ed9409ae462be2556366f3;hp=9dbfcb5ac92e5fada772fd8c20a74a4a41921c63;hb=384bff4264f199ded8fa28d241ce0e7dc021a97c;hpb=6b5322ce000d6a8a6f845beda2d7e149e1baea0c diff --git a/src/tt.cpp b/src/tt.cpp index 9dbfcb5a..0b4a59de 100644 --- a/src/tt.cpp +++ b/src/tt.cpp @@ -1,7 +1,8 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 Copyright (C) 2004-2008 Tord Romstad (Glaurung author) - Copyright (C) 2008-2012 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2015-2020 Marco Costalba, Joona Kiiski, Gary Linscott, 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 @@ -17,125 +18,141 @@ along with this program. If not, see . */ -#include +#include // For std::memset #include +#include #include "bitboard.h" +#include "misc.h" +#include "thread.h" #include "tt.h" +#include "uci.h" TranspositionTable TT; // Our global transposition table -TranspositionTable::TranspositionTable() { +/// TTEntry::save populates the TTEntry with a new node's data, possibly +/// overwriting an old position. Update is not atomic and can be racy. - size = generation = 0; - entries = NULL; -} +void TTEntry::save(Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev) { + + // Preserve any existing move for the same position + if (m || (k >> 48) != key16) + move16 = (uint16_t)m; -TranspositionTable::~TranspositionTable() { + // Overwrite less valuable entries + if ( (k >> 48) != key16 + || d - DEPTH_OFFSET > depth8 - 4 + || b == BOUND_EXACT) + { + assert(d >= DEPTH_OFFSET); - delete [] entries; + key16 = (uint16_t)(k >> 48); + value16 = (int16_t)v; + eval16 = (int16_t)ev; + genBound8 = (uint8_t)(TT.generation8 | uint8_t(pv) << 2 | b); + depth8 = (uint8_t)(d - DEPTH_OFFSET); + } } -/// TranspositionTable::set_size() sets the size of the transposition table, -/// measured in megabytes. Transposition table consists of a power of 2 number of -/// TTCluster and each cluster consists of ClusterSize number of TTEntries. Each -/// non-empty entry contains information of exactly one position. +/// TranspositionTable::resize() sets the size of the transposition table, +/// measured in megabytes. Transposition table consists of a power of 2 number +/// of clusters and each cluster consists of ClusterSize number of TTEntry. -void TranspositionTable::set_size(size_t mbSize) { +void TranspositionTable::resize(size_t mbSize) { - size_t newSize = 1ULL << msb((mbSize << 20) / sizeof(TTCluster)); + Threads.main()->wait_for_search_finished(); - if (newSize == size) - return; + clusterCount = mbSize * 1024 * 1024 / sizeof(Cluster); - size = newSize; - delete [] entries; - entries = new (std::nothrow) TTCluster[size]; + free(mem); + mem = malloc(clusterCount * sizeof(Cluster) + CacheLineSize - 1); - if (!entries) + if (!mem) { std::cerr << "Failed to allocate " << mbSize << "MB for transposition table." << std::endl; exit(EXIT_FAILURE); } - clear(); // Operator new is not guaranteed to initialize memory to zero + table = (Cluster*)((uintptr_t(mem) + CacheLineSize - 1) & ~(CacheLineSize - 1)); + clear(); } -/// TranspositionTable::clear() overwrites the entire transposition table -/// 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). +/// TranspositionTable::clear() initializes the entire transposition table to zero, +// in a multi-threaded way. void TranspositionTable::clear() { - memset(entries, 0, size * sizeof(TTCluster)); -} - - -/// TranspositionTable::store() writes a new entry containing position key and -/// valuable information of current position. The lowest order bits of position -/// key are used to decide on which cluster the position will be placed. -/// When a new entry is written and there are no empty entries available in cluster, -/// it replaces the least valuable of entries. 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. - -void TranspositionTable::store(const Key posKey, Value v, Bound t, Depth d, Move m, Value statV, Value kingD) { - - int c1, c2, c3; - TTEntry *tte, *replace; - uint32_t posKey32 = posKey >> 32; // Use the high 32 bits as key inside the cluster + std::vector threads; - tte = replace = first_entry(posKey); - - for (int i = 0; i < ClusterSize; i++, tte++) + for (size_t idx = 0; idx < Options["Threads"]; ++idx) { - if (!tte->key() || tte->key() == posKey32) // Empty or overwrite old - { - // Preserve any existing ttMove - if (m == MOVE_NONE) - m = tte->move(); + threads.emplace_back([this, idx]() { - tte->save(posKey32, v, t, d, m, generation, statV, kingD); - return; - } + // Thread binding gives faster search on systems with a first-touch policy + if (Options["Threads"] > 8) + WinProcGroup::bindThisThread(idx); - // Implement replace strategy - c1 = (replace->generation() == generation ? 2 : 0); - c2 = (tte->generation() == generation || tte->type() == BOUND_EXACT ? -2 : 0); - c3 = (tte->depth() < replace->depth() ? 1 : 0); + // Each thread will zero its part of the hash table + const size_t stride = clusterCount / Options["Threads"], + start = stride * idx, + len = idx != Options["Threads"] - 1 ? + stride : clusterCount - start; - if (c1 + c2 + c3 > 0) - replace = tte; + std::memset(&table[start], 0, len * sizeof(Cluster)); + }); } - replace->save(posKey32, v, t, d, m, generation, statV, kingD); + + for (std::thread& th: threads) + th.join(); } +/// TranspositionTable::probe() looks up the current position in the transposition +/// table. It returns true and a pointer to the TTEntry if the position is found. +/// Otherwise, it returns false and a pointer to an empty or least valuable TTEntry +/// to be replaced later. The replace value of an entry is calculated as its depth +/// minus 8 times its relative age. TTEntry t1 is considered more valuable than +/// TTEntry t2 if its replace value is greater than that of t2. -/// TranspositionTable::probe() looks up the current position in the -/// transposition table. Returns a pointer to the TTEntry or NULL if -/// position is not found. +TTEntry* TranspositionTable::probe(const Key key, bool& found) const { -TTEntry* TranspositionTable::probe(const Key posKey) const { + TTEntry* const tte = first_entry(key); + const uint16_t key16 = key >> 48; // Use the high 16 bits as key inside the cluster - uint32_t posKey32 = posKey >> 32; - TTEntry* tte = first_entry(posKey); + for (int i = 0; i < ClusterSize; ++i) + if (!tte[i].key16 || tte[i].key16 == key16) + { + tte[i].genBound8 = uint8_t(generation8 | (tte[i].genBound8 & 0x7)); // Refresh - for (int i = 0; i < ClusterSize; i++, tte++) - if (tte->key() == posKey32) - return tte; + return found = (bool)tte[i].key16, &tte[i]; + } - return NULL; + // Find an entry to be replaced according to the replacement strategy + TTEntry* replace = tte; + for (int i = 1; i < ClusterSize; ++i) + // Due to our packed storage format for generation and its cyclic + // nature we add 263 (256 is the modulus plus 7 to keep the unrelated + // lowest three bits from affecting the result) to calculate the entry + // age correctly even after generation8 overflows into the next cycle. + if ( replace->depth8 - ((263 + generation8 - replace->genBound8) & 0xF8) + > tte[i].depth8 - ((263 + generation8 - tte[i].genBound8) & 0xF8)) + replace = &tte[i]; + + return found = false, replace; } -/// 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 -/// entries from the current search. +/// TranspositionTable::hashfull() returns an approximation of the hashtable +/// occupation during a search. The hash is x permill full, as per UCI protocol. + +int TranspositionTable::hashfull() const { + + int cnt = 0; + for (int i = 0; i < 1000 / ClusterSize; ++i) + for (int j = 0; j < ClusterSize; ++j) + cnt += (table[i].entry[j].genBound8 & 0xF8) == generation8; -void TranspositionTable::new_search() { - generation++; + return cnt * 1000 / (ClusterSize * (1000 / ClusterSize)); }