/*
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) 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> // For std::memset
+#include <iostream>
+#include <thread>
-////
-//// Includes
-////
-
-#include <cassert>
-#include <cmath>
-#include <cstring>
-#if !defined(__hpux)
-# include <xmmintrin.h>
-#endif
-
-#include "movegen.h"
+#include "bitboard.h"
+#include "misc.h"
+#include "thread.h"
#include "tt.h"
+#include "uci.h"
-// The main transposition table
-TranspositionTable TT;
-
-////
-//// Functions
-////
-
-TranspositionTable::TranspositionTable() {
-
- size = writes = 0;
- entries = 0;
- generation = 0;
-}
-
-TranspositionTable::~TranspositionTable() {
-
- delete [] entries;
-}
-
+namespace Stockfish {
-/// TranspositionTable::set_size sets the size of the transposition table,
-/// measured in megabytes.
+TranspositionTable TT; // Our global transposition table
-void TranspositionTable::set_size(size_t mbSize) {
+/// 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_t newSize = 1024;
+void TTEntry::save(Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev) {
- // We store a cluster of ClusterSize number of TTEntry for each position
- // and newSize is the maximum number of storable positions.
- while ((2 * newSize) * sizeof(TTCluster) <= (mbSize << 20))
- newSize *= 2;
+ // Preserve any existing move for the same position
+ if (m || (uint16_t)k != key16)
+ move16 = (uint16_t)m;
- if (newSize != size)
+ // Overwrite less valuable entries (cheapest checks first)
+ if (b == BOUND_EXACT
+ || (uint16_t)k != key16
+ || d - DEPTH_OFFSET > depth8 - 4)
{
- size = newSize;
- delete [] entries;
- entries = new TTCluster[size];
- if (!entries)
- {
- std::cerr << "Failed to allocate " << mbSize
- << " MB for transposition table." << std::endl;
- Application::exit_with_failure();
- }
- clear();
+ 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::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).
-/// Perhaps we should also clear it when the "ucinewgame" command is recieved?
+/// 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::clear() {
+void TranspositionTable::resize(size_t mbSize) {
- memset(entries, 0, size * sizeof(TTCluster));
-}
+ Threads.main()->wait_for_search_finished();
+ aligned_large_pages_free(table);
-/// 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.
+ clusterCount = mbSize * 1024 * 1024 / sizeof(Cluster);
-inline TTEntry* TranspositionTable::first_entry(const Key posKey) const {
+ 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);
+ }
- return entries[uint32_t(posKey) & (size - 1)].data;
+ clear();
}
-/// TranspositionTable::store writes a new entry containing a position,
-/// a value, a value type, a search depth, and a best move to the
-/// transposition table. Transposition table is organized in clusters of
-/// four TTEntry objects, and when a new entry is written, it replaces
-/// the least valuable of the four entries in a cluster. 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. A TTEntry of type VALUE_TYPE_EVAL
-/// never replaces another entry for the same position.
+/// TranspositionTable::clear() initializes the entire transposition table to zero,
+// in a multi-threaded way.
-void TranspositionTable::store(const Key posKey, Value v, ValueType t, Depth d, Move m) {
+void TranspositionTable::clear() {
- TTEntry *tte, *replace;
- uint32_t posKey32 = posKey >> 32; // Use the high 32 bits as key
+ 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
- {
- // Do not overwrite when new type is VALUE_TYPE_EV_LO
- if (tte->key() && t == VALUE_TYPE_EV_LO)
- return;
+ threads.emplace_back([this, idx]() {
- if (m == MOVE_NONE)
- m = tte->move();
+ // Thread binding gives faster search on systems with a first-touch policy
+ if (Options["Threads"] > 8)
+ WinProcGroup::bindThisThread(idx);
- *tte = TTEntry(posKey32, v, t, d, m, generation);
- return;
- }
- else if (i == 0) // replace would be a no-op in this common case
- continue;
-
- int c1 = (replace->generation() == generation ? 2 : 0);
- int c2 = (tte->generation() == generation ? -2 : 0);
- int 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 = TTEntry(posKey32, v, t, d, m, generation);
- writes++;
-}
-
-
-/// TranspositionTable::retrieve looks up the current position in the
-/// transposition table. Returns a pointer to the TTEntry or NULL
-/// if position is not found.
-TTEntry* TranspositionTable::retrieve(const Key posKey) const {
-
- uint32_t posKey32 = posKey >> 32;
- TTEntry* tte = first_entry(posKey);
-
- for (int i = 0; i < ClusterSize; i++, tte++)
- if (tte->key() == posKey32)
- return tte;
-
- return NULL;
-}
-
-
-/// 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(__hpux)
-void TranspositionTable::prefetch(const Key) const {} // Not supported on HP UX
-#else
-
-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
-
- char const* addr = (char*)first_entry(posKey);
- _mm_prefetch(addr, _MM_HINT_T2);
- _mm_prefetch(addr+64, _MM_HINT_T2); // 64 bytes ahead
+ for (std::thread& th : threads)
+ th.join();
}
-#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
-/// entries from the current search.
-
-void TranspositionTable::new_search() {
-
- generation++;
- writes = 0;
-}
+/// 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 {
-/// TranspositionTable::insert_pv() is called at the end of a search
-/// iteration, and inserts the PV back into the PV. This makes sure
-/// the old PV moves are searched first, even if the old TT entries
-/// have been overwritten.
+ TTEntry* const tte = first_entry(key);
+ const uint16_t key16 = (uint16_t)key; // Use the low 16 bits as key inside the cluster
-void TranspositionTable::insert_pv(const Position& pos, Move pv[]) {
+ 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
- StateInfo st;
- Position p(pos);
+ return found = (bool)tte[i].depth8, &tte[i];
+ }
- for (int i = 0; pv[i] != MOVE_NONE; 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);
- }
+ // 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::extract_pv() extends a PV by adding moves from the
-/// transposition table at the end. This should ensure that the PV is almost
-/// always at least two plies long, which is important, because otherwise we
-/// 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[], const int PLY_MAX) {
+/// TranspositionTable::hashfull() returns an approximation of the hashtable
+/// occupation during a search. The hash is x permill full, as per UCI protocol.
- const TTEntry* tte;
- StateInfo st;
- Position p(pos);
- int ply = 0;
+int TranspositionTable::hashfull() const {
- // Update position to the end of current PV
- while (pv[ply] != MOVE_NONE)
- p.do_move(pv[ply++], st);
+ 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;
- // 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)
- {
- pv[ply] = tte->move();
- p.do_move(pv[ply++], st);
- }
- pv[ply] = MOVE_NONE;
+ return cnt / ClusterSize;
}
-
-/// TranspositionTable::full() returns the permill of all transposition table
-/// entries which have received at least one write during the current search.
-/// It is used to display the "info hashfull ..." information in UCI.
-
-int TranspositionTable::full() const {
-
- double N = double(size) * ClusterSize;
- return int(1000 * (1 - exp(writes * log(1.0 - 1.0/N))));
-}
+} // namespace Stockfish