X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Ftt.cpp;h=6121b3ad771b24bd40410e8e5f6bff6004fdccc7;hp=a9f8f037ab92925f80a62c7a26d5dfc2be7849d7;hb=d4dca9187e83dde29be8d76ca50ff53d14199ce9;hpb=9dbb3ae8b2c5ef4ad2b399a42707ad1fd851c68b diff --git a/src/tt.cpp b/src/tt.cpp index a9f8f037..6121b3ad 100644 --- a/src/tt.cpp +++ b/src/tt.cpp @@ -2,6 +2,7 @@ Stockfish, a UCI chess playing engine derived from Glaurung 2.1 Copyright (C) 2004-2008 Tord Romstad (Glaurung author) Copyright (C) 2008-2015 Marco Costalba, Joona Kiiski, Tord Romstad + Copyright (C) 2015-2019 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 @@ -19,12 +20,42 @@ #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 +/// TTEntry::save populates the TTEntry with a new node's data, possibly +/// overwriting an old position. Update is not atomic and can be racy. + +void TTEntry::save(Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev) { + + assert(d / ONE_PLY * ONE_PLY == d); + + // Preserve any existing move for the same position + if (m || (k >> 48) != key16) + move16 = (uint16_t)m; + + // Overwrite less valuable entries + if ( (k >> 48) != key16 + ||(d - DEPTH_OFFSET) / ONE_PLY > depth8 - 4 + || b == BOUND_EXACT) + { + assert((d - DEPTH_OFFSET) / ONE_PLY >= 0); + + 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) / ONE_PLY); + } +} + /// TranspositionTable::resize() sets the size of the transposition table, /// measured in megabytes. Transposition table consists of a power of 2 number @@ -32,15 +63,12 @@ TranspositionTable TT; // Our global transposition table void TranspositionTable::resize(size_t mbSize) { - size_t newClusterCount = size_t(1) << msb((mbSize * 1024 * 1024) / sizeof(Cluster)); + Threads.main()->wait_for_search_finished(); - if (newClusterCount == clusterCount) - return; - - clusterCount = newClusterCount; + clusterCount = mbSize * 1024 * 1024 / sizeof(Cluster); free(mem); - mem = calloc(clusterCount * sizeof(Cluster) + CacheLineSize - 1, 1); + mem = malloc(clusterCount * sizeof(Cluster) + CacheLineSize - 1); if (!mem) { @@ -50,18 +78,38 @@ void TranspositionTable::resize(size_t mbSize) { } table = (Cluster*)((uintptr_t(mem) + CacheLineSize - 1) & ~(CacheLineSize - 1)); + clear(); } -/// TranspositionTable::clear() overwrites the entire transposition table -/// with zeros. 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() { - std::memset(table, 0, clusterCount * sizeof(Cluster)); -} + std::vector threads; + for (size_t idx = 0; idx < Options["Threads"]; ++idx) + { + threads.emplace_back([this, idx]() { + + // Thread binding gives faster search on systems with a first-touch policy + if (Options["Threads"] > 8) + WinProcGroup::bindThisThread(idx); + + // 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; + + std::memset(&table[start], 0, len * sizeof(Cluster)); + }); + } + + 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. @@ -78,8 +126,7 @@ TTEntry* TranspositionTable::probe(const Key key, bool& found) const { for (int i = 0; i < ClusterSize; ++i) if (!tte[i].key16 || tte[i].key16 == key16) { - if (tte[i].key16) - tte[i].genBound8 = uint8_t(generation8 | tte[i].bound()); // Refresh + tte[i].genBound8 = uint8_t(generation8 | (tte[i].genBound8 & 0x7)); // Refresh return found = (bool)tte[i].key16, &tte[i]; } @@ -87,26 +134,27 @@ TTEntry* TranspositionTable::probe(const Key key, bool& found) const { // Find an entry to be replaced according to the replacement strategy TTEntry* replace = tte; for (int i = 1; i < ClusterSize; ++i) - if ( replace->depth8 - ((259 + generation8 - replace->genBound8) & 0xFC) * 2 * ONE_PLY - > tte[i].depth8 - ((259 + generation8 - tte[i].genBound8) & 0xFC) * 2 * ONE_PLY) + // 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; } -/// Returns an approximation of the hashtable occupation during a search. The -/// hash is x permill full, as per UCI protocol. +/// 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 TranspositionTable::hashfull() const -{ int cnt = 0; - for (int i = 0; i < 1000 / ClusterSize; i++) - { - const TTEntry* tte = &table[i].entry[0]; - for (int j = 0; j < ClusterSize; j++) - if ((tte[j].genBound8 & 0xFC) == generation8) - cnt++; - } - return cnt; + for (int i = 0; i < 1000 / ClusterSize; ++i) + for (int j = 0; j < ClusterSize; ++j) + cnt += (table[i].entry[j].genBound8 & 0xF8) == generation8; + + return cnt * 1000 / (ClusterSize * (1000 / ClusterSize)); }