X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=src%2Ftt.cpp;h=53e78595045548bed917e75132dfc084c60ed88a;hb=HEAD;hp=d0e2d729c8c766e883092277bfc5fd97b2bb58e2;hpb=40548c9153ea89c0b27b198efb443c5bb9b9c490;p=stockfish
diff --git a/src/tt.cpp b/src/tt.cpp
index d0e2d729..4b55e53f 100644
--- a/src/tt.cpp
+++ b/src/tt.cpp
@@ -1,7 +1,6 @@
/*
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) 2004-2024 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
@@ -17,97 +16,234 @@
along with this program. If not, see .
*/
-#include // For std::memset
+#include "tt.h"
+
+#include
+#include
+#include
+#include
#include
-#include "bitboard.h"
-#include "tt.h"
+#include "memory.h"
+#include "misc.h"
+#include "syzygy/tbprobe.h"
+#include "thread.h"
+
+namespace Stockfish {
+
+
+// TTEntry struct is the 10 bytes transposition table entry, defined as below:
+//
+// key 16 bit
+// depth 8 bit
+// generation 5 bit
+// pv node 1 bit
+// bound type 2 bit
+// move 16 bit
+// value 16 bit
+// evaluation 16 bit
+//
+// These fields are in the same order as accessed by TT::probe(), since memory is fastest sequentially.
+// Equally, the store order in save() matches this order.
+
+struct TTEntry {
+
+ // Convert internal bitfields to external types
+ TTData read() const {
+ return TTData{Move(move16), Value(value16),
+ Value(eval16), Depth(depth8 + DEPTH_ENTRY_OFFSET),
+ Bound(genBound8 & 0x3), bool(genBound8 & 0x4)};
+ }
+
+ bool is_occupied() const;
+ void save(Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev, uint8_t generation8);
+ // The returned age is a multiple of TranspositionTable::GENERATION_DELTA
+ uint8_t relative_age(const uint8_t generation8) const;
+
+ private:
+ friend class TranspositionTable;
+
+ uint16_t key16;
+ uint8_t depth8;
+ uint8_t genBound8;
+ Move move16;
+ int16_t value16;
+ int16_t eval16;
+};
+
+// `genBound8` is where most of the details are. We use the following constants to manipulate 5 leading generation bits
+// and 3 trailing miscellaneous bits.
+
+// These bits are reserved for other things.
+static constexpr unsigned GENERATION_BITS = 3;
+// increment for generation field
+static constexpr int GENERATION_DELTA = (1 << GENERATION_BITS);
+// cycle length
+static constexpr int GENERATION_CYCLE = 255 + GENERATION_DELTA;
+// mask to pull out generation number
+static constexpr int GENERATION_MASK = (0xFF << GENERATION_BITS) & 0xFF;
+
+// DEPTH_ENTRY_OFFSET exists because 1) we use `bool(depth8)` as the occupancy check, but
+// 2) we need to store negative depths for QS. (`depth8` is the only field with "spare bits":
+// we sacrifice the ability to store depths greater than 1<<8 less the offset, as asserted in `save`.)
+bool TTEntry::is_occupied() const { return bool(depth8); }
+
+// Populates the TTEntry with a new node's data, possibly
+// overwriting an old position. The 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, uint8_t generation8) {
+
+ // Preserve the old ttmove if we don't have a new one
+ if (m || uint16_t(k) != key16)
+ move16 = m;
+
+ // Overwrite less valuable entries (cheapest checks first)
+ if (b == BOUND_EXACT || uint16_t(k) != key16 || d - DEPTH_ENTRY_OFFSET + 2 * pv > depth8 - 4
+ || relative_age(generation8))
+ {
+ assert(d > DEPTH_ENTRY_OFFSET);
+ assert(d < 256 + DEPTH_ENTRY_OFFSET);
+
+ key16 = uint16_t(k);
+ depth8 = uint8_t(d - DEPTH_ENTRY_OFFSET);
+ genBound8 = uint8_t(generation8 | uint8_t(pv) << 2 | b);
+ value16 = int16_t(v);
+ eval16 = int16_t(ev);
+ }
+}
+
+
+uint8_t TTEntry::relative_age(const uint8_t generation8) const {
+ // 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.
+ return (GENERATION_CYCLE + generation8 - genBound8) & GENERATION_MASK;
+}
+
+
+// TTWriter is but a very thin wrapper around the pointer
+TTWriter::TTWriter(TTEntry* tte) :
+ entry(tte) {}
+
+void TTWriter::write(
+ Key k, Value v, bool pv, Bound b, Depth d, Move m, Value ev, uint8_t generation8) {
+ entry->save(k, v, pv, b, d, m, ev, generation8);
+}
-TranspositionTable TT; // Our global transposition table
+// A TranspositionTable is an array of Cluster, of size clusterCount. Each cluster consists of ClusterSize number
+// of TTEntry. Each non-empty TTEntry contains information on exactly one position. The size of a Cluster should
+// divide the size of a cache line for best performance, as the cacheline is prefetched when possible.
-/// 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.
+static constexpr int ClusterSize = 3;
-void TranspositionTable::resize(size_t mbSize) {
+struct Cluster {
+ TTEntry entry[ClusterSize];
+ char padding[2]; // Pad to 32 bytes
+};
- size_t newClusterCount = size_t(1) << msb((mbSize * 1024 * 1024) / sizeof(Cluster));
+static_assert(sizeof(Cluster) == 32, "Suboptimal Cluster size");
- if (newClusterCount == clusterCount)
- return;
- clusterCount = newClusterCount;
+// Sets the size of the transposition table,
+// measured in megabytes. Transposition table consists
+// of clusters and each cluster consists of ClusterSize number of TTEntry.
+void TranspositionTable::resize(size_t mbSize, ThreadPool& threads) {
+ aligned_large_pages_free(table);
- free(mem);
- mem = calloc(clusterCount * sizeof(Cluster) + CacheLineSize - 1, 1);
+ clusterCount = mbSize * 1024 * 1024 / sizeof(Cluster);
- if (!mem)
- {
- std::cerr << "Failed to allocate " << mbSize
- << "MB for transposition table." << std::endl;
- exit(EXIT_FAILURE);
- }
+ table = static_cast(aligned_large_pages_alloc(clusterCount * sizeof(Cluster)));
- table = (Cluster*)((uintptr_t(mem) + CacheLineSize - 1) & ~(CacheLineSize - 1));
+ if (!table)
+ {
+ std::cerr << "Failed to allocate " << mbSize << "MB for transposition table." << std::endl;
+ exit(EXIT_FAILURE);
+ }
+
+ clear(threads);
}
-/// 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).
+// Initializes the entire transposition table to zero,
+// in a multi-threaded way.
+void TranspositionTable::clear(ThreadPool& threads) {
+ generation8 = 0;
+ const size_t threadCount = threads.num_threads();
+
+ for (size_t i = 0; i < threadCount; ++i)
+ {
+ threads.run_on_thread(i, [this, i, threadCount]() {
+ // Each thread will zero its part of the hash table
+ const size_t stride = clusterCount / threadCount;
+ const size_t start = stride * i;
+ const size_t len = i + 1 != threadCount ? stride : clusterCount - start;
-void TranspositionTable::clear() {
+ std::memset(&table[start], 0, len * sizeof(Cluster));
+ });
+ }
- std::memset(table, 0, clusterCount * sizeof(Cluster));
+ for (size_t i = 0; i < threadCount; ++i)
+ threads.wait_on_thread(i);
}
-/// 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. 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.
+// Returns an approximation of the hashtable
+// occupation during a search. The hash is x permill full, as per UCI protocol.
+// Only counts entries which match the current generation.
+int TranspositionTable::hashfull() const {
+
+ int cnt = 0;
+ for (int i = 0; i < 1000; ++i)
+ for (int j = 0; j < ClusterSize; ++j)
+ cnt += table[i].entry[j].is_occupied()
+ && (table[i].entry[j].genBound8 & GENERATION_MASK) == generation8;
+
+ return cnt / ClusterSize;
+}
-TTEntry* TranspositionTable::probe(const Key key, bool& found) const {
- TTEntry* const tte = first_entry(key);
- const uint16_t key16 = key >> 48; // Use the high 16 bits as key inside the cluster
+void TranspositionTable::new_search() {
+ // increment by delta to keep lower bits as is
+ generation8 += GENERATION_DELTA;
+}
- 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
- return found = (bool)tte[i].key16, &tte[i];
- }
+uint8_t TranspositionTable::generation() const { return generation8; }
- // Find an entry to be replaced according to the replacement strategy
- TTEntry* replace = tte;
- for (int i = 1; i < ClusterSize; ++i)
- if ( (( tte[i].genBound8 & 0xFC) == generation8 || tte[i].bound() == BOUND_EXACT)
- - ((replace->genBound8 & 0xFC) == generation8)
- - (tte[i].depth8 < replace->depth8) < 0)
- replace = &tte[i];
- return found = false, replace;
+// Looks up the current position in the transposition
+// table. It returns true 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.
+std::tuple TranspositionTable::probe(const Key key) const {
+
+ TTEntry* const tte = first_entry(key);
+ const uint16_t key16 = uint16_t(key); // Use the low 16 bits as key inside the cluster
+
+ for (int i = 0; i < ClusterSize; ++i)
+ if (tte[i].key16 == key16)
+ // This gap is the main place for read races.
+ // After `read()` completes that copy is final, but may be self-inconsistent.
+ return {tte[i].is_occupied(), tte[i].read(), TTWriter(&tte[i])};
+
+ // Find an entry to be replaced according to the replacement strategy
+ TTEntry* replace = tte;
+ for (int i = 1; i < ClusterSize; ++i)
+ if (replace->depth8 - replace->relative_age(generation8) * 2
+ > tte[i].depth8 - tte[i].relative_age(generation8) * 2)
+ replace = &tte[i];
+
+ return {false, TTData(), TTWriter(replace)};
}
-/// 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++)
- {
- const TTEntry* tte = &table[i].entry[0];
- for (int j = 0; j < ClusterSize; j++)
- if ((tte[j].genBound8 & 0xFC) == generation8)
- cnt++;
- }
- return cnt;
+TTEntry* TranspositionTable::first_entry(const Key key) const {
+ return &table[mul_hi64(key, clusterCount)].entry[0];
}
+
+} // namespace Stockfish