/*
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) 2004-2021 The Stockfish developers (see AUTHORS file)
Stockfish is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
-#include <cstring>
+#include <cstring> // For std::memset
#include <iostream>
+#include <thread>
+#include "bitboard.h"
+#include "misc.h"
+#include "thread.h"
#include "tt.h"
+#include "uci.h"
+
+namespace Stockfish {
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) {
-TranspositionTable::~TranspositionTable() {
+ // Preserve any existing move for the same position
+ if (m || (uint16_t)k != key16)
+ move16 = (uint16_t)m;
- delete [] entries;
+ // Overwrite less valuable entries (cheapest checks first)
+ if (b == BOUND_EXACT
+ || (uint16_t)k != key16
+ || d - DEPTH_OFFSET > depth8 - 4)
+ {
+ assert(d > DEPTH_OFFSET);
+ assert(d < 256 + DEPTH_OFFSET);
+
+ key16 = (uint16_t)k;
+ depth8 = (uint8_t)(d - DEPTH_OFFSET);
+ genBound8 = (uint8_t)(TT.generation8 | uint8_t(pv) << 2 | b);
+ value16 = (int16_t)v;
+ eval16 = (int16_t)ev;
+ }
}
-/// TranspositionTable::set_size() sets the size of the transposition table,
-/// measured in megabytes.
+/// 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 = 1024;
+ Threads.main()->wait_for_search_finished();
- // Transposition table consists of clusters and each cluster consists
- // of ClusterSize number of TTEntries. Each non-empty entry contains
- // information of exactly one position and newSize is the number of
- // clusters we are going to allocate.
- while (2ULL * newSize * sizeof(TTCluster) <= (mbSize << 20))
- newSize *= 2;
+ aligned_large_pages_free(table);
- if (newSize == size)
- return;
+ clusterCount = mbSize * 1024 * 1024 / sizeof(Cluster);
- size = newSize;
- delete [] entries;
- entries = new (std::nothrow) TTCluster[size];
- if (!entries)
+ table = static_cast<Cluster*>(aligned_large_pages_alloc(clusterCount * sizeof(Cluster)));
+ if (!table)
{
std::cerr << "Failed to allocate " << mbSize
<< "MB for transposition table." << std::endl;
exit(EXIT_FAILURE);
}
+
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<std::thread> 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 = size_t(clusterCount / Options["Threads"]),
+ start = size_t(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. Returns a pointer to the TTEntry or NULL if
-/// position is not found.
+/// 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.
+
+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 = (uint16_t)key; // Use the low 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 == key16 || !tte[i].depth8)
+ {
+ tte[i].genBound8 = uint8_t(generation8 | (tte[i].genBound8 & (GENERATION_DELTA - 1))); // Refresh
- for (int i = 0; i < ClusterSize; i++, tte++)
- if (tte->key() == posKey32)
- return tte;
+ return found = (bool)tte[i].depth8, &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 GENERATION_CYCLE (256 is the modulus, plus what
+ // is needed to keep the unrelated lowest n bits from affecting
+ // the result) to calculate the entry age correctly even after
+ // generation8 overflows into the next cycle.
+ if ( replace->depth8 - ((GENERATION_CYCLE + generation8 - replace->genBound8) & GENERATION_MASK)
+ > tte[i].depth8 - ((GENERATION_CYCLE + generation8 - tte[i].genBound8) & GENERATION_MASK))
+ 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 {
-void TranspositionTable::new_search() {
- generation++;
+ int cnt = 0;
+ for (int i = 0; i < 1000; ++i)
+ for (int j = 0; j < ClusterSize; ++j)
+ cnt += table[i].entry[j].depth8 && (table[i].entry[j].genBound8 & GENERATION_MASK) == generation8;
+
+ return cnt / ClusterSize;
}
+
+} // namespace Stockfish