X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Ftt.cpp;h=92aaee001824276e49f1090c72ec451070900dd0;hp=354ead87adb34dac0a007dcf064067fe9362b9d8;hb=4b10578acbe099482ed40200478df4d775c01af5;hpb=2ba47416cbdd5db2c7c79257072cd8675b61721f diff --git a/src/tt.cpp b/src/tt.cpp index 354ead87..92aaee00 100644 --- a/src/tt.cpp +++ b/src/tt.cpp @@ -2,7 +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) 2015-2017 Marco Costalba, Joona Kiiski, Gary Linscott, 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 @@ -20,12 +20,40 @@ #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) { + + // 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 > depth8 - 4 + || b == BOUND_EXACT) + { + assert(d >= DEPTH_OFFSET); + + 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::resize() sets the size of the transposition table, /// measured in megabytes. Transposition table consists of a power of 2 number @@ -33,16 +61,14 @@ TranspositionTable TT; // Our global transposition table void TranspositionTable::resize(size_t mbSize) { - size_t newClusterCount = mbSize * 1024 * 1024 / sizeof(Cluster); - - if (newClusterCount == clusterCount) - return; + Threads.main()->wait_for_search_finished(); - clusterCount = newClusterCount; + aligned_ttmem_free(mem); - free(mem); - mem = malloc(clusterCount * sizeof(Cluster) + CacheLineSize - 1); + superClusterCount = mbSize * 1024 * 1024 / (sizeof(Cluster) * ClustersPerSuperCluster); + table = static_cast( + aligned_ttmem_alloc(superClusterCount * ClustersPerSuperCluster * sizeof(Cluster), mem)); if (!mem) { std::cerr << "Failed to allocate " << mbSize @@ -50,20 +76,40 @@ void TranspositionTable::resize(size_t mbSize) { exit(EXIT_FAILURE); } - 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]() { + + const size_t clusterCount = superClusterCount * ClustersPerSuperCluster; + + // 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 = size_t(clusterCount / Options["Threads"]), + start = size_t(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. @@ -80,8 +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].genBound8 & 0xFC) != generation8 && 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]; } @@ -90,11 +135,11 @@ TTEntry* TranspositionTable::probe(const Key key, bool& found) const { TTEntry* replace = tte; for (int i = 1; i < ClusterSize; ++i) // Due to our packed storage format for generation and its cyclic - // nature we add 259 (256 is the modulus plus 3 to keep the lowest - // two bound bits from affecting the result) to calculate the entry + // 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 - ((259 + generation8 - replace->genBound8) & 0xFC) * 2 - > tte[i].depth8 - ((259 + generation8 - tte[i].genBound8) & 0xFC) * 2) + if ( replace->depth8 - ((263 + generation8 - replace->genBound8) & 0xF8) + > tte[i].depth8 - ((263 + generation8 - tte[i].genBound8) & 0xF8)) replace = &tte[i]; return found = false, replace; @@ -107,12 +152,9 @@ TTEntry* TranspositionTable::probe(const Key key, bool& found) 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; ++i) + for (int j = 0; j < ClusterSize; ++j) + cnt += (table[i].entry[j].genBound8 & 0xF8) == generation8; + + return cnt / ClusterSize; }