X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fsyzygy%2Ftbprobe.cpp;h=c0cd04c1d4d8bd8f30607b616ae02363cb84da69;hp=2dd50a6609427b0ca8e591a3f16315b7d319f1c4;hb=699bae632f283746a3eee15c0950fcdbca8a355e;hpb=d8f683760c9eb6d2c4714ec83e717dd2143de55c diff --git a/src/syzygy/tbprobe.cpp b/src/syzygy/tbprobe.cpp index 2dd50a66..c0cd04c1 100644 --- a/src/syzygy/tbprobe.cpp +++ b/src/syzygy/tbprobe.cpp @@ -1,7 +1,6 @@ /* Stockfish, a UCI chess playing engine derived from Glaurung 2.1 - Copyright (c) 2013 Ronald de Man - Copyright (C) 2016-2017 Marco Costalba, Lucas Braesch + 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 @@ -20,20 +19,21 @@ #include #include #include -#include // For std::memset +#include // For std::memset and std::memcpy #include #include #include #include #include #include +#include #include "../bitboard.h" #include "../movegen.h" #include "../position.h" #include "../search.h" -#include "../thread_win32.h" #include "../types.h" +#include "../uci.h" #include "tbprobe.h" @@ -44,171 +44,47 @@ #include #else #define WIN32_LEAN_AND_MEAN -#define NOMINMAX +#ifndef NOMINMAX +# define NOMINMAX // Disable macros min() and max() +#endif #include #endif -using namespace Tablebases; +using namespace Stockfish::Tablebases; + +int Stockfish::Tablebases::MaxCardinality; -int Tablebases::MaxCardinality; +namespace Stockfish { namespace { +constexpr int TBPIECES = 7; // Max number of supported pieces + +enum { BigEndian, LittleEndian }; +enum TBType { WDL, DTZ }; // Used as template parameter + // Each table has a set of flags: all of them refer to DTZ tables, the last one to WDL tables -enum TBFlag { STM = 1, Mapped = 2, WinPlies = 4, LossPlies = 8, SingleValue = 128 }; +enum TBFlag { STM = 1, Mapped = 2, WinPlies = 4, LossPlies = 8, Wide = 16, SingleValue = 128 }; inline WDLScore operator-(WDLScore d) { return WDLScore(-int(d)); } -inline Square operator^=(Square& s, int i) { return s = Square(int(s) ^ i); } inline Square operator^(Square s, int i) { return Square(int(s) ^ i); } -// DTZ tables don't store valid scores for moves that reset the rule50 counter -// like captures and pawn moves but we can easily recover the correct dtz of the -// previous move if we know the position's WDL score. -int dtz_before_zeroing(WDLScore wdl) { - return wdl == WDLWin ? 1 : - wdl == WDLCursedWin ? 101 : - wdl == WDLBlessedLoss ? -101 : - wdl == WDLLoss ? -1 : 0; -} - -// Return the sign of a number (-1, 0, 1) -template int sign_of(T val) { - return (T(0) < val) - (val < T(0)); -} - -// Numbers in little endian used by sparseIndex[] to point into blockLength[] -struct SparseEntry { - char block[4]; // Number of block - char offset[2]; // Offset within the block -}; - -static_assert(sizeof(SparseEntry) == 6, "SparseEntry must be 6 bytes"); - -typedef uint16_t Sym; // Huffman symbol - -struct LR { - enum Side { Left, Right, Value }; - - uint8_t lr[3]; // The first 12 bits is the left-hand symbol, the second 12 - // bits is the right-hand symbol. If symbol has length 1, - // then the first byte is the stored value. - template - Sym get() { - return S == Left ? ((lr[1] & 0xF) << 8) | lr[0] : - S == Right ? (lr[2] << 4) | (lr[1] >> 4) : - S == Value ? lr[0] : (assert(false), Sym(-1)); - } -}; - -static_assert(sizeof(LR) == 3, "LR tree entry must be 3 bytes"); - -const int TBPIECES = 6; - -struct PairsData { - int flags; - size_t sizeofBlock; // Block size in bytes - size_t span; // About every span values there is a SparseIndex[] entry - int blocksNum; // Number of blocks in the TB file - int maxSymLen; // Maximum length in bits of the Huffman symbols - int minSymLen; // Minimum length in bits of the Huffman symbols - Sym* lowestSym; // lowestSym[l] is the symbol of length l with the lowest value - LR* btree; // btree[sym] stores the left and right symbols that expand sym - uint16_t* blockLength; // Number of stored positions (minus one) for each block: 1..65536 - int blockLengthSize; // Size of blockLength[] table: padded so it's bigger than blocksNum - SparseEntry* sparseIndex; // Partial indices into blockLength[] - size_t sparseIndexSize; // Size of SparseIndex[] table - uint8_t* data; // Start of Huffman compressed data - std::vector base64; // base64[l - min_sym_len] is the 64bit-padded lowest symbol of length l - std::vector symlen; // Number of values (-1) represented by a given Huffman symbol: 1..256 - Piece pieces[TBPIECES]; // Position pieces: the order of pieces defines the groups - uint64_t groupIdx[TBPIECES+1]; // Start index used for the encoding of the group's pieces - int groupLen[TBPIECES+1]; // Number of pieces in a given group: KRKN -> (3, 1) -}; - -// Helper struct to avoid manually defining entry copy constructor as we -// should because the default one is not compatible with std::atomic_bool. -struct Atomic { - Atomic() = default; - Atomic(const Atomic& e) { ready = e.ready.load(); } // MSVC 2013 wants assignment within body - std::atomic_bool ready; -}; - -// We define types for the different parts of the WLDEntry and DTZEntry with -// corresponding specializations for pieces or pawns. - -struct WLDEntryPiece { - PairsData* precomp; -}; - -struct WDLEntryPawn { - uint8_t pawnCount[2]; // [Lead color / other color] - WLDEntryPiece file[2][4]; // [wtm / btm][FILE_A..FILE_D] -}; - -struct DTZEntryPiece { - PairsData* precomp; - uint16_t map_idx[4]; // WDLWin, WDLLoss, WDLCursedWin, WDLBlessedLoss - uint8_t* map; -}; - -struct DTZEntryPawn { - uint8_t pawnCount[2]; - DTZEntryPiece file[4]; - uint8_t* map; -}; - -struct TBEntry : public Atomic { - void* baseAddress; - uint64_t mapping; - Key key; - Key key2; - int pieceCount; - bool hasPawns; - bool hasUniquePieces; -}; - -// Now the main types: WDLEntry and DTZEntry -struct WDLEntry : public TBEntry { - WDLEntry(const std::string& code); - ~WDLEntry(); - union { - WLDEntryPiece pieceTable[2]; // [wtm / btm] - WDLEntryPawn pawnTable; - }; -}; - -struct DTZEntry : public TBEntry { - DTZEntry(const WDLEntry& wdl); - ~DTZEntry(); - union { - DTZEntryPiece pieceTable; - DTZEntryPawn pawnTable; - }; -}; - -typedef decltype(WDLEntry::pieceTable) WDLPieceTable; -typedef decltype(DTZEntry::pieceTable) DTZPieceTable; -typedef decltype(WDLEntry::pawnTable ) WDLPawnTable; -typedef decltype(DTZEntry::pawnTable ) DTZPawnTable; - -auto item(WDLPieceTable& e, int stm, int ) -> decltype(e[stm])& { return e[stm]; } -auto item(DTZPieceTable& e, int , int ) -> decltype(e)& { return e; } -auto item(WDLPawnTable& e, int stm, int f) -> decltype(e.file[stm][f])& { return e.file[stm][f]; } -auto item(DTZPawnTable& e, int , int f) -> decltype(e.file[f])& { return e.file[f]; } - -template struct Ret { typedef int type; }; -template<> struct Ret { typedef WDLScore type; }; +const std::string PieceToChar = " PNBRQK pnbrqk"; int MapPawns[SQUARE_NB]; int MapB1H1H7[SQUARE_NB]; int MapA1D1D4[SQUARE_NB]; int MapKK[10][SQUARE_NB]; // [MapA1D1D4][SQUARE_NB] +int Binomial[7][SQUARE_NB]; // [k][n] k elements from a set of n elements +int LeadPawnIdx[6][SQUARE_NB]; // [leadPawnsCnt][SQUARE_NB] +int LeadPawnsSize[6][4]; // [leadPawnsCnt][FILE_A..FILE_D] + // Comparison function to sort leading pawns in ascending MapPawns[] order bool pawns_comp(Square i, Square j) { return MapPawns[i] < MapPawns[j]; } int off_A1H8(Square sq) { return int(rank_of(sq)) - file_of(sq); } -const Value WDL_to_value[] = { +constexpr Value WDL_to_value[] = { -VALUE_MATE + MAX_PLY + 1, VALUE_DRAW - 2, VALUE_DRAW, @@ -216,27 +92,21 @@ const Value WDL_to_value[] = { VALUE_MATE - MAX_PLY - 1 }; -const std::string PieceToChar = " PNBRQK pnbrqk"; - -int Binomial[6][SQUARE_NB]; // [k][n] k elements from a set of n elements -int LeadPawnIdx[5][SQUARE_NB]; // [leadPawnsCnt][SQUARE_NB] -int LeadPawnsSize[5][4]; // [leadPawnsCnt][FILE_A..FILE_D] - -enum { BigEndian, LittleEndian }; - template -inline void swap_byte(T& x) +inline void swap_endian(T& x) { - char tmp, *c = (char*)&x; + static_assert(std::is_unsigned::value, "Argument of swap_endian not unsigned"); + + uint8_t tmp, *c = (uint8_t*)&x; for (int i = 0; i < Half; ++i) tmp = c[i], c[i] = c[End - i], c[End - i] = tmp; } -template<> inline void swap_byte(uint8_t&) {} +template<> inline void swap_endian(uint8_t&) {} template T number(void* addr) { - const union { uint32_t i; char c[4]; } Le = { 0x01020304 }; - const bool IsLittleEndian = (Le.c[0] == 4); + static const union { uint32_t i; char c[4]; } Le = { 0x01020304 }; + static const bool IsLittleEndian = (Le.c[0] == 4); T v; @@ -246,59 +116,59 @@ template T number(void* addr) v = *((T*)addr); if (LE != IsLittleEndian) - swap_byte(v); + swap_endian(v); return v; } -class HashTable { - - typedef std::pair EntryPair; - typedef std::pair Entry; +// DTZ tables don't store valid scores for moves that reset the rule50 counter +// like captures and pawn moves but we can easily recover the correct dtz of the +// previous move if we know the position's WDL score. +int dtz_before_zeroing(WDLScore wdl) { + return wdl == WDLWin ? 1 : + wdl == WDLCursedWin ? 101 : + wdl == WDLBlessedLoss ? -101 : + wdl == WDLLoss ? -1 : 0; +} - static const int TBHASHBITS = 10; - static const int HSHMAX = 5; +// Return the sign of a number (-1, 0, 1) +template int sign_of(T val) { + return (T(0) < val) - (val < T(0)); +} - Entry hashTable[1 << TBHASHBITS][HSHMAX]; +// Numbers in little endian used by sparseIndex[] to point into blockLength[] +struct SparseEntry { + char block[4]; // Number of block + char offset[2]; // Offset within the block +}; - std::deque wdlTable; - std::deque dtzTable; +static_assert(sizeof(SparseEntry) == 6, "SparseEntry must be 6 bytes"); - void insert(Key key, WDLEntry* wdl, DTZEntry* dtz) { - Entry* entry = hashTable[key >> (64 - TBHASHBITS)]; +typedef uint16_t Sym; // Huffman symbol - for (int i = 0; i < HSHMAX; ++i, ++entry) - if (!entry->second.first || entry->first == key) { - *entry = std::make_pair(key, std::make_pair(wdl, dtz)); - return; - } +struct LR { + enum Side { Left, Right }; - std::cerr << "HSHMAX too low!" << std::endl; - exit(1); + uint8_t lr[3]; // The first 12 bits is the left-hand symbol, the second 12 + // bits is the right-hand symbol. If symbol has length 1, + // then the left-hand symbol is the stored value. + template + Sym get() { + return S == Left ? ((lr[1] & 0xF) << 8) | lr[0] : + S == Right ? (lr[2] << 4) | (lr[1] >> 4) : (assert(false), Sym(-1)); } - -public: - template::value ? 0 : 1> - E* get(Key key) { - Entry* entry = hashTable[key >> (64 - TBHASHBITS)]; - - for (int i = 0; i < HSHMAX; ++i, ++entry) - if (entry->first == key) - return std::get(entry->second); - - return nullptr; - } - - void clear() { - std::memset(hashTable, 0, sizeof(hashTable)); - wdlTable.clear(); - dtzTable.clear(); - } - size_t size() const { return wdlTable.size(); } - void insert(const std::vector& pieces); }; -HashTable EntryTable; +static_assert(sizeof(LR) == 3, "LR tree entry must be 3 bytes"); + +// Tablebases data layout is structured as following: +// +// TBFile: memory maps/unmaps the physical .rtbw and .rtbz files +// TBTable: one object for each file with corresponding indexing information +// TBTables: has ownership of TBTable objects, keeping a list and a hash +// class TBFile memory maps/unmaps the single .rtbw and .rtbz files. Files are +// memory mapped for best performance. Files are mapped at first access: at init +// time only existence of the file is checked. class TBFile : public std::ifstream { std::string fname; @@ -315,9 +185,9 @@ public: TBFile(const std::string& f) { #ifndef _WIN32 - const char SepChar = ':'; + constexpr char SepChar = ':'; #else - const char SepChar = ';'; + constexpr char SepChar = ';'; #endif std::stringstream ss(Paths); std::string path; @@ -332,7 +202,7 @@ public: // Memory map the file and check it. File should be already open and will be // closed after mapping. - uint8_t* map(void** baseAddress, uint64_t* mapping, const uint8_t* TB_MAGIC) { + uint8_t* map(void** baseAddress, uint64_t* mapping, TBType type) { assert(is_open()); @@ -341,50 +211,79 @@ public: #ifndef _WIN32 struct stat statbuf; int fd = ::open(fname.c_str(), O_RDONLY); + + if (fd == -1) + return *baseAddress = nullptr, nullptr; + fstat(fd, &statbuf); + + if (statbuf.st_size % 64 != 16) + { + std::cerr << "Corrupt tablebase file " << fname << std::endl; + exit(EXIT_FAILURE); + } + *mapping = statbuf.st_size; *baseAddress = mmap(nullptr, statbuf.st_size, PROT_READ, MAP_SHARED, fd, 0); +#if defined(MADV_RANDOM) + madvise(*baseAddress, statbuf.st_size, MADV_RANDOM); +#endif ::close(fd); - if (*baseAddress == MAP_FAILED) { + if (*baseAddress == MAP_FAILED) + { std::cerr << "Could not mmap() " << fname << std::endl; - exit(1); + exit(EXIT_FAILURE); } #else + // Note FILE_FLAG_RANDOM_ACCESS is only a hint to Windows and as such may get ignored. HANDLE fd = CreateFile(fname.c_str(), GENERIC_READ, FILE_SHARE_READ, nullptr, - OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, nullptr); + OPEN_EXISTING, FILE_FLAG_RANDOM_ACCESS, nullptr); + + if (fd == INVALID_HANDLE_VALUE) + return *baseAddress = nullptr, nullptr; + DWORD size_high; DWORD size_low = GetFileSize(fd, &size_high); + + if (size_low % 64 != 16) + { + std::cerr << "Corrupt tablebase file " << fname << std::endl; + exit(EXIT_FAILURE); + } + HANDLE mmap = CreateFileMapping(fd, nullptr, PAGE_READONLY, size_high, size_low, nullptr); CloseHandle(fd); - if (!mmap) { + if (!mmap) + { std::cerr << "CreateFileMapping() failed" << std::endl; - exit(1); + exit(EXIT_FAILURE); } *mapping = (uint64_t)mmap; *baseAddress = MapViewOfFile(mmap, FILE_MAP_READ, 0, 0, 0); - if (!*baseAddress) { + if (!*baseAddress) + { std::cerr << "MapViewOfFile() failed, name = " << fname << ", error = " << GetLastError() << std::endl; - exit(1); + exit(EXIT_FAILURE); } #endif uint8_t* data = (uint8_t*)*baseAddress; - if ( *data++ != *TB_MAGIC++ - || *data++ != *TB_MAGIC++ - || *data++ != *TB_MAGIC++ - || *data++ != *TB_MAGIC) { + constexpr uint8_t Magics[][4] = { { 0xD7, 0x66, 0x0C, 0xA5 }, + { 0x71, 0xE8, 0x23, 0x5D } }; + + if (memcmp(data, Magics[type == WDL], 4)) + { std::cerr << "Corrupted table in file " << fname << std::endl; unmap(*baseAddress, *mapping); - *baseAddress = nullptr; - return nullptr; + return *baseAddress = nullptr, nullptr; } - return data; + return data + 4; // Skip Magics's header } static void unmap(void* baseAddress, uint64_t mapping) { @@ -400,80 +299,181 @@ public: std::string TBFile::Paths; -WDLEntry::WDLEntry(const std::string& code) { +// struct PairsData contains low level indexing information to access TB data. +// There are 8, 4 or 2 PairsData records for each TBTable, according to type of +// table and if positions have pawns or not. It is populated at first access. +struct PairsData { + uint8_t flags; // Table flags, see enum TBFlag + uint8_t maxSymLen; // Maximum length in bits of the Huffman symbols + uint8_t minSymLen; // Minimum length in bits of the Huffman symbols + uint32_t blocksNum; // Number of blocks in the TB file + size_t sizeofBlock; // Block size in bytes + size_t span; // About every span values there is a SparseIndex[] entry + Sym* lowestSym; // lowestSym[l] is the symbol of length l with the lowest value + LR* btree; // btree[sym] stores the left and right symbols that expand sym + uint16_t* blockLength; // Number of stored positions (minus one) for each block: 1..65536 + uint32_t blockLengthSize; // Size of blockLength[] table: padded so it's bigger than blocksNum + SparseEntry* sparseIndex; // Partial indices into blockLength[] + size_t sparseIndexSize; // Size of SparseIndex[] table + uint8_t* data; // Start of Huffman compressed data + std::vector base64; // base64[l - min_sym_len] is the 64bit-padded lowest symbol of length l + std::vector symlen; // Number of values (-1) represented by a given Huffman symbol: 1..256 + Piece pieces[TBPIECES]; // Position pieces: the order of pieces defines the groups + uint64_t groupIdx[TBPIECES+1]; // Start index used for the encoding of the group's pieces + int groupLen[TBPIECES+1]; // Number of pieces in a given group: KRKN -> (3, 1) + uint16_t map_idx[4]; // WDLWin, WDLLoss, WDLCursedWin, WDLBlessedLoss (used in DTZ) +}; + +// struct TBTable contains indexing information to access the corresponding TBFile. +// There are 2 types of TBTable, corresponding to a WDL or a DTZ file. TBTable +// is populated at init time but the nested PairsData records are populated at +// first access, when the corresponding file is memory mapped. +template +struct TBTable { + typedef typename std::conditional::type Ret; + + static constexpr int Sides = Type == WDL ? 2 : 1; + + std::atomic_bool ready; + void* baseAddress; + uint8_t* map; + uint64_t mapping; + Key key; + Key key2; + int pieceCount; + bool hasPawns; + bool hasUniquePieces; + uint8_t pawnCount[2]; // [Lead color / other color] + PairsData items[Sides][4]; // [wtm / btm][FILE_A..FILE_D or 0] + + PairsData* get(int stm, int f) { + return &items[stm % Sides][hasPawns ? f : 0]; + } + + TBTable() : ready(false), baseAddress(nullptr) {} + explicit TBTable(const std::string& code); + explicit TBTable(const TBTable& wdl); + + ~TBTable() { + if (baseAddress) + TBFile::unmap(baseAddress, mapping); + } +}; + +template<> +TBTable::TBTable(const std::string& code) : TBTable() { StateInfo st; Position pos; - memset(this, 0, sizeof(WDLEntry)); - - ready = false; key = pos.set(code, WHITE, &st).material_key(); - pieceCount = popcount(pos.pieces()); + pieceCount = pos.count(); hasPawns = pos.pieces(PAWN); - for (Color c = WHITE; c <= BLACK; ++c) + hasUniquePieces = false; + for (Color c : { WHITE, BLACK }) for (PieceType pt = PAWN; pt < KING; ++pt) if (popcount(pos.pieces(c, pt)) == 1) hasUniquePieces = true; - if (hasPawns) { - // Set the leading color. In case both sides have pawns the leading color - // is the side with less pawns because this leads to better compression. - bool c = !pos.count(BLACK) - || ( pos.count(WHITE) - && pos.count(BLACK) >= pos.count(WHITE)); + // Set the leading color. In case both sides have pawns the leading color + // is the side with less pawns because this leads to better compression. + bool c = !pos.count(BLACK) + || ( pos.count(WHITE) + && pos.count(BLACK) >= pos.count(WHITE)); - pawnTable.pawnCount[0] = pos.count(c ? WHITE : BLACK); - pawnTable.pawnCount[1] = pos.count(c ? BLACK : WHITE); - } + pawnCount[0] = pos.count(c ? WHITE : BLACK); + pawnCount[1] = pos.count(c ? BLACK : WHITE); key2 = pos.set(code, BLACK, &st).material_key(); } -WDLEntry::~WDLEntry() { - - if (baseAddress) - TBFile::unmap(baseAddress, mapping); - - for (int i = 0; i < 2; ++i) - if (hasPawns) - for (File f = FILE_A; f <= FILE_D; ++f) - delete pawnTable.file[i][f].precomp; - else - delete pieceTable[i].precomp; -} - -DTZEntry::DTZEntry(const WDLEntry& wdl) { +template<> +TBTable::TBTable(const TBTable& wdl) : TBTable() { - memset(this, 0, sizeof(DTZEntry)); - - ready = false; + // Use the corresponding WDL table to avoid recalculating all from scratch key = wdl.key; key2 = wdl.key2; pieceCount = wdl.pieceCount; hasPawns = wdl.hasPawns; hasUniquePieces = wdl.hasUniquePieces; + pawnCount[0] = wdl.pawnCount[0]; + pawnCount[1] = wdl.pawnCount[1]; +} + +// class TBTables creates and keeps ownership of the TBTable objects, one for +// each TB file found. It supports a fast, hash based, table lookup. Populated +// at init time, accessed at probe time. +class TBTables { + + struct Entry + { + Key key; + TBTable* wdl; + TBTable* dtz; + + template + TBTable* get() const { + return (TBTable*)(Type == WDL ? (void*)wdl : (void*)dtz); + } + }; + + static constexpr int Size = 1 << 12; // 4K table, indexed by key's 12 lsb + static constexpr int Overflow = 1; // Number of elements allowed to map to the last bucket + + Entry hashTable[Size + Overflow]; + + std::deque> wdlTable; + std::deque> dtzTable; + + void insert(Key key, TBTable* wdl, TBTable* dtz) { + uint32_t homeBucket = (uint32_t)key & (Size - 1); + Entry entry{ key, wdl, dtz }; - if (hasPawns) { - pawnTable.pawnCount[0] = wdl.pawnTable.pawnCount[0]; - pawnTable.pawnCount[1] = wdl.pawnTable.pawnCount[1]; + // Ensure last element is empty to avoid overflow when looking up + for (uint32_t bucket = homeBucket; bucket < Size + Overflow - 1; ++bucket) { + Key otherKey = hashTable[bucket].key; + if (otherKey == key || !hashTable[bucket].get()) { + hashTable[bucket] = entry; + return; + } + + // Robin Hood hashing: If we've probed for longer than this element, + // insert here and search for a new spot for the other element instead. + uint32_t otherHomeBucket = (uint32_t)otherKey & (Size - 1); + if (otherHomeBucket > homeBucket) { + std::swap(entry, hashTable[bucket]); + key = otherKey; + homeBucket = otherHomeBucket; + } + } + std::cerr << "TB hash table size too low!" << std::endl; + exit(EXIT_FAILURE); } -} -DTZEntry::~DTZEntry() { +public: + template + TBTable* get(Key key) { + for (const Entry* entry = &hashTable[(uint32_t)key & (Size - 1)]; ; ++entry) { + if (entry->key == key || !entry->get()) + return entry->get(); + } + } - if (baseAddress) - TBFile::unmap(baseAddress, mapping); + void clear() { + memset(hashTable, 0, sizeof(hashTable)); + wdlTable.clear(); + dtzTable.clear(); + } + size_t size() const { return wdlTable.size(); } + void add(const std::vector& pieces); +}; - if (hasPawns) - for (File f = FILE_A; f <= FILE_D; ++f) - delete pawnTable.file[f].precomp; - else - delete pieceTable.precomp; -} +TBTables TBTables; -void HashTable::insert(const std::vector& pieces) { +// If the corresponding file exists two new objects TBTable and TBTable +// are created and added to the lists and hash table. Called at init time. +void TBTables::add(const std::vector& pieces) { std::string code; @@ -482,16 +482,17 @@ void HashTable::insert(const std::vector& pieces) { TBFile file(code.insert(code.find('K', 1), "v") + ".rtbw"); // KRK -> KRvK - if (!file.is_open()) + if (!file.is_open()) // Only WDL file is checked return; file.close(); MaxCardinality = std::max((int)pieces.size(), MaxCardinality); - wdlTable.push_back(WDLEntry(code)); - dtzTable.push_back(DTZEntry(wdlTable.back())); + wdlTable.emplace_back(code); + dtzTable.emplace_back(wdlTable.back()); + // Insert into the hash keys for both colors: KRvK with KR white and black insert(wdlTable.back().key , &wdlTable.back(), &dtzTable.back()); insert(wdlTable.back().key2, &wdlTable.back(), &dtzTable.back()); } @@ -531,14 +532,14 @@ int decompress_pairs(PairsData* d, uint64_t idx) { // // I(k) = k * d->span + d->span / 2 (1) - // First step is to get the 'k' of the I(k) nearest to our idx, using defintion (1) - uint32_t k = idx / d->span; + // First step is to get the 'k' of the I(k) nearest to our idx, using definition (1) + uint32_t k = uint32_t(idx / d->span); // Then we read the corresponding SparseIndex[] entry uint32_t block = number(&d->sparseIndex[k].block); int offset = number(&d->sparseIndex[k].offset); - // Now compute the difference idx - I(k). From defintion of k we know that + // Now compute the difference idx - I(k). From definition of k we know that // // idx = k * d->span + idx % d->span (2) // @@ -557,7 +558,7 @@ int decompress_pairs(PairsData* d, uint64_t idx) { offset -= d->blockLength[block++] + 1; // Finally, we find the start address of our block of canonical Huffman symbols - uint32_t* ptr = (uint32_t*)(d->data + block * d->sizeofBlock); + uint32_t* ptr = (uint32_t*)(d->data + ((uint64_t)block * d->sizeofBlock)); // Read the first 64 bits in our block, this is a (truncated) sequence of // unknown number of symbols of unknown length but we know the first one @@ -578,7 +579,7 @@ int decompress_pairs(PairsData* d, uint64_t idx) { // All the symbols of a given length are consecutive integers (numerical // sequence property), so we can compute the offset of our symbol of // length len, stored at the beginning of buf64. - sym = (buf64 - d->base64[len]) >> (64 - len - d->minSymLen); + sym = Sym((buf64 - d->base64[len]) >> (64 - len - d->minSymLen)); // Now add the value of the lowest symbol of length len to get our symbol sym += number(&d->lowestSym[len]); @@ -620,16 +621,14 @@ int decompress_pairs(PairsData* d, uint64_t idx) { } } - return d->btree[sym].get(); + return d->btree[sym].get(); } -bool check_dtz_stm(WDLEntry*, int, File) { return true; } - -bool check_dtz_stm(DTZEntry* entry, int stm, File f) { +bool check_dtz_stm(TBTable*, int, File) { return true; } - int flags = entry->hasPawns ? entry->pawnTable.file[f].precomp->flags - : entry->pieceTable.precomp->flags; +bool check_dtz_stm(TBTable* entry, int stm, File f) { + auto flags = entry->get(stm, f)->flags; return (flags & TBFlag::STM) == stm || ((entry->key == entry->key2) && !entry->hasPawns); } @@ -638,22 +637,22 @@ bool check_dtz_stm(DTZEntry* entry, int stm, File f) { // values 0, 1, 2, ... in order of decreasing frequency. This is done for each // of the four WDLScore values. The mapping information necessary to reconstruct // the original values is stored in the TB file and read during map[] init. -WDLScore map_score(WDLEntry*, File, int value, WDLScore) { return WDLScore(value - 2); } +WDLScore map_score(TBTable*, File, int value, WDLScore) { return WDLScore(value - 2); } -int map_score(DTZEntry* entry, File f, int value, WDLScore wdl) { +int map_score(TBTable* entry, File f, int value, WDLScore wdl) { - const int WDLMap[] = { 1, 3, 0, 2, 0 }; + constexpr int WDLMap[] = { 1, 3, 0, 2, 0 }; - int flags = entry->hasPawns ? entry->pawnTable.file[f].precomp->flags - : entry->pieceTable.precomp->flags; + auto flags = entry->get(0, f)->flags; - uint8_t* map = entry->hasPawns ? entry->pawnTable.map - : entry->pieceTable.map; - - uint16_t* idx = entry->hasPawns ? entry->pawnTable.file[f].map_idx - : entry->pieceTable.map_idx; - if (flags & TBFlag::Mapped) - value = map[idx[WDLMap[wdl + 2]] + value]; + uint8_t* map = entry->map; + uint16_t* idx = entry->get(0, f)->map_idx; + if (flags & TBFlag::Mapped) { + if (flags & TBFlag::Wide) + value = ((uint16_t *)map)[idx[WDLMap[wdl + 2]] + value]; + else + value = map[idx[WDLMap[wdl + 2]] + value]; + } // DTZ tables store distance to zero in number of moves or plies. We // want to return plies, so we have convert to plies when needed. @@ -672,10 +671,8 @@ int map_score(DTZEntry* entry, File f, int value, WDLScore wdl) { // // idx = Binomial[1][s1] + Binomial[2][s2] + ... + Binomial[k][sk] // -template::type> -T do_probe_table(const Position& pos, Entry* entry, WDLScore wdl, ProbeState* result) { - - const bool IsWDL = std::is_same::value; +template +Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* result) { Square squares[TBPIECES]; Piece pieces[TBPIECES]; @@ -698,7 +695,7 @@ T do_probe_table(const Position& pos, Entry* entry, WDLScore wdl, ProbeState* r bool blackStronger = (pos.material_key() != entry->key); int flipColor = (symmetricBlackToMove || blackStronger) * 8; - int flipSquares = (symmetricBlackToMove || blackStronger) * 070; + int flipSquares = (symmetricBlackToMove || blackStronger) * 56; int stm = (symmetricBlackToMove || blackStronger) ^ pos.side_to_move(); // For pawns, TB files store 4 separate tables according if leading pawn is on @@ -708,45 +705,45 @@ T do_probe_table(const Position& pos, Entry* entry, WDLScore wdl, ProbeState* r // In all the 4 tables, pawns are at the beginning of the piece sequence and // their color is the reference one. So we just pick the first one. - Piece pc = Piece(item(entry->pawnTable, 0, 0).precomp->pieces[0] ^ flipColor); + Piece pc = Piece(entry->get(0, 0)->pieces[0] ^ flipColor); assert(type_of(pc) == PAWN); leadPawns = b = pos.pieces(color_of(pc), PAWN); - while (b) + do squares[size++] = pop_lsb(&b) ^ flipSquares; + while (b); leadPawnsCnt = size; std::swap(squares[0], *std::max_element(squares, squares + leadPawnsCnt, pawns_comp)); - tbFile = file_of(squares[0]); - if (tbFile > FILE_D) - tbFile = file_of(squares[0] ^ 7); // Horizontal flip: SQ_H1 -> SQ_A1 - - d = item(entry->pawnTable , stm, tbFile).precomp; - } else - d = item(entry->pieceTable, stm, tbFile).precomp; + tbFile = File(edge_distance(file_of(squares[0]))); + } // DTZ tables are one-sided, i.e. they store positions only for white to // move or only for black to move, so check for side to move to be stm, // early exit otherwise. - if (!IsWDL && !check_dtz_stm(entry, stm, tbFile)) - return *result = CHANGE_STM, T(); + if (!check_dtz_stm(entry, stm, tbFile)) + return *result = CHANGE_STM, Ret(); // Now we are ready to get all the position pieces (but the lead pawns) and // directly map them to the correct color and square. b = pos.pieces() ^ leadPawns; - while (b) { + do { Square s = pop_lsb(&b); squares[size] = s ^ flipSquares; pieces[size++] = Piece(pos.piece_on(s) ^ flipColor); - } + } while (b); + + assert(size >= 2); + + d = entry->get(stm, tbFile); // Then we reorder the pieces to have the same sequence as the one stored - // in precomp->pieces[i]: the sequence that ensures the best compression. - for (int i = leadPawnsCnt; i < size; ++i) - for (int j = i; j < size; ++j) + // in pieces[i]: the sequence that ensures the best compression. + for (int i = leadPawnsCnt; i < size - 1; ++i) + for (int j = i + 1; j < size; ++j) if (d->pieces[i] == pieces[j]) { std::swap(pieces[i], pieces[j]); @@ -758,14 +755,14 @@ T do_probe_table(const Position& pos, Entry* entry, WDLScore wdl, ProbeState* r // the triangle A1-D1-D4. if (file_of(squares[0]) > FILE_D) for (int i = 0; i < size; ++i) - squares[i] ^= 7; // Horizontal flip: SQ_H1 -> SQ_A1 + squares[i] = flip_file(squares[i]); // Encode leading pawns starting with the one with minimum MapPawns[] and // proceeding in ascending order. if (entry->hasPawns) { idx = LeadPawnIdx[leadPawnsCnt][squares[0]]; - std::sort(squares + 1, squares + leadPawnsCnt, pawns_comp); + std::stable_sort(squares + 1, squares + leadPawnsCnt, pawns_comp); for (int i = 1; i < leadPawnsCnt; ++i) idx += Binomial[i][MapPawns[squares[i]]]; @@ -777,7 +774,7 @@ T do_probe_table(const Position& pos, Entry* entry, WDLScore wdl, ProbeState* r // piece is below RANK_5. if (rank_of(squares[0]) > RANK_4) for (int i = 0; i < size; ++i) - squares[i] ^= 070; // Vertical flip: SQ_A8 -> SQ_A1 + squares[i] = flip_rank(squares[i]); // Look for the first piece of the leading group not on the A1-D4 diagonal // and ensure it is mapped below the diagonal. @@ -785,7 +782,7 @@ T do_probe_table(const Position& pos, Entry* entry, WDLScore wdl, ProbeState* r if (!off_A1H8(squares[i])) continue; - if (off_A1H8(squares[i]) > 0) // A1-H8 diagonal flip: SQ_A3 -> SQ_C3 + if (off_A1H8(squares[i]) > 0) // A1-H8 diagonal flip: SQ_A3 -> SQ_C1 for (int j = i; j < size; ++j) squares[j] = Square(((squares[j] >> 3) | (squares[j] << 3)) & 63); break; @@ -862,11 +859,11 @@ encode_remaining: Square* groupSq = squares + d->groupLen[0]; // Encode remainig pawns then pieces according to square, in ascending order - bool remainingPawns = entry->hasPawns && entry->pawnTable.pawnCount[1]; + bool remainingPawns = entry->hasPawns && entry->pawnCount[1]; while (d->groupLen[++next]) { - std::sort(groupSq, groupSq + d->groupLen[next]); + std::stable_sort(groupSq, groupSq + d->groupLen[next]); uint64_t n = 0; // Map down a square if "comes later" than a square in the previous @@ -924,7 +921,7 @@ void set_groups(T& e, PairsData* d, int order[], File f) { // pawns/pieces -> remainig pawns -> remaining pieces. In particular the // first group is at order[0] position and the remaining pawns, when present, // are at order[1] position. - bool pp = e.hasPawns && e.pawnTable.pawnCount[1]; // Pawns on both sides + bool pp = e.hasPawns && e.pawnCount[1]; // Pawns on both sides int next = pp ? 2 : 1; int freeSquares = 64 - d->groupLen[0] - (pp ? d->groupLen[1] : 0); uint64_t idx = 1; @@ -978,8 +975,8 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) { d->flags = *data++; if (d->flags & TBFlag::SingleValue) { - d->blocksNum = d->span = - d->blockLengthSize = d->sparseIndexSize = 0; // Broken MSVC zero-init + d->blocksNum = d->blockLengthSize = 0; + d->span = d->sparseIndexSize = 0; // Broken MSVC zero-init d->minSymLen = *data++; // Here we store the single value return data; } @@ -990,8 +987,8 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) { d->sizeofBlock = 1ULL << *data++; d->span = 1ULL << *data++; - d->sparseIndexSize = (tbSize + d->span - 1) / d->span; // Round up - int padding = number(data++); + d->sparseIndexSize = size_t((tbSize + d->span - 1) / d->span); // Round up + auto padding = number(data++); d->blocksNum = number(data); data += sizeof(uint32_t); d->blockLengthSize = d->blocksNum + padding; // Padded to ensure SparseIndex[] // does not point out of range. @@ -1005,7 +1002,7 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) { // so that d->lowestSym[i] >= d->lowestSym[i+1] (when read as LittleEndian). // Starting from this we compute a base64[] table indexed by symbol length // and containing 64 bit values so that d->base64[i] >= d->base64[i+1]. - // See http://www.eecs.harvard.edu/~michaelm/E210/huffman.pdf + // See https://en.wikipedia.org/wiki/Huffman_coding for (int i = d->base64.size() - 2; i >= 0; --i) { d->base64[i] = (d->base64[i + 1] + number(&d->lowestSym[i]) - number(&d->lowestSym[i + 1])) / 2; @@ -1024,7 +1021,7 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) { d->symlen.resize(number(data)); data += sizeof(uint16_t); d->btree = (LR*)data; - // The comrpession scheme used is "Recursive Pairing", that replaces the most + // The compression scheme used is "Recursive Pairing", that replaces the most // frequent adjacent pair of symbols in the source message by a new symbol, // reevaluating the frequencies of all of the symbol pairs with respect to // the extended alphabet, and then repeating the process. @@ -1038,105 +1035,115 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) { return data + d->symlen.size() * sizeof(LR) + (d->symlen.size() & 1); } -template -uint8_t* set_dtz_map(WDLEntry&, T&, uint8_t*, File) { return nullptr; } +uint8_t* set_dtz_map(TBTable&, uint8_t* data, File) { return data; } -template -uint8_t* set_dtz_map(DTZEntry&, T& p, uint8_t* data, File maxFile) { +uint8_t* set_dtz_map(TBTable& e, uint8_t* data, File maxFile) { - p.map = data; + e.map = data; for (File f = FILE_A; f <= maxFile; ++f) { - if (item(p, 0, f).precomp->flags & TBFlag::Mapped) - for (int i = 0; i < 4; ++i) { // Sequence like 3,x,x,x,1,x,0,2,x,x - item(p, 0, f).map_idx[i] = (uint16_t)(data - p.map + 1); - data += *data + 1; + auto flags = e.get(0, f)->flags; + if (flags & TBFlag::Mapped) { + if (flags & TBFlag::Wide) { + data += (uintptr_t)data & 1; // Word alignment, we may have a mixed table + for (int i = 0; i < 4; ++i) { // Sequence like 3,x,x,x,1,x,0,2,x,x + e.get(0, f)->map_idx[i] = (uint16_t)((uint16_t *)data - (uint16_t *)e.map + 1); + data += 2 * number(data) + 2; + } + } + else { + for (int i = 0; i < 4; ++i) { + e.get(0, f)->map_idx[i] = (uint16_t)(data - e.map + 1); + data += *data + 1; + } } + } } return data += (uintptr_t)data & 1; // Word alignment } -template -void do_init(Entry& e, T& p, uint8_t* data) { - - const bool IsWDL = std::is_same::value; +// Populate entry's PairsData records with data from the just memory mapped file. +// Called at first access. +template +void set(T& e, uint8_t* data) { PairsData* d; enum { Split = 1, HasPawns = 2 }; - assert(e.hasPawns == !!(*data & HasPawns)); - assert((e.key != e.key2) == !!(*data & Split)); + assert(e.hasPawns == bool(*data & HasPawns)); + assert((e.key != e.key2) == bool(*data & Split)); data++; // First byte stores flags - const int Sides = IsWDL && (e.key != e.key2) ? 2 : 1; - const File MaxFile = e.hasPawns ? FILE_D : FILE_A; + const int sides = T::Sides == 2 && (e.key != e.key2) ? 2 : 1; + const File maxFile = e.hasPawns ? FILE_D : FILE_A; - bool pp = e.hasPawns && e.pawnTable.pawnCount[1]; // Pawns on both sides + bool pp = e.hasPawns && e.pawnCount[1]; // Pawns on both sides - assert(!pp || e.pawnTable.pawnCount[0]); + assert(!pp || e.pawnCount[0]); - for (File f = FILE_A; f <= MaxFile; ++f) { + for (File f = FILE_A; f <= maxFile; ++f) { - for (int i = 0; i < Sides; i++) - item(p, i, f).precomp = new PairsData(); + for (int i = 0; i < sides; i++) + *e.get(i, f) = PairsData(); int order[][2] = { { *data & 0xF, pp ? *(data + 1) & 0xF : 0xF }, { *data >> 4, pp ? *(data + 1) >> 4 : 0xF } }; data += 1 + pp; for (int k = 0; k < e.pieceCount; ++k, ++data) - for (int i = 0; i < Sides; i++) - item(p, i, f).precomp->pieces[k] = Piece(i ? *data >> 4 : *data & 0xF); + for (int i = 0; i < sides; i++) + e.get(i, f)->pieces[k] = Piece(i ? *data >> 4 : *data & 0xF); - for (int i = 0; i < Sides; ++i) - set_groups(e, item(p, i, f).precomp, order[i], f); + for (int i = 0; i < sides; ++i) + set_groups(e, e.get(i, f), order[i], f); } data += (uintptr_t)data & 1; // Word alignment - for (File f = FILE_A; f <= MaxFile; ++f) - for (int i = 0; i < Sides; i++) - data = set_sizes(item(p, i, f).precomp, data); + for (File f = FILE_A; f <= maxFile; ++f) + for (int i = 0; i < sides; i++) + data = set_sizes(e.get(i, f), data); - if (!IsWDL) - data = set_dtz_map(e, p, data, MaxFile); + data = set_dtz_map(e, data, maxFile); - for (File f = FILE_A; f <= MaxFile; ++f) - for (int i = 0; i < Sides; i++) { - (d = item(p, i, f).precomp)->sparseIndex = (SparseEntry*)data; - data += d->sparseIndexSize * sizeof(SparseEntry) ; + for (File f = FILE_A; f <= maxFile; ++f) + for (int i = 0; i < sides; i++) { + (d = e.get(i, f))->sparseIndex = (SparseEntry*)data; + data += d->sparseIndexSize * sizeof(SparseEntry); } - for (File f = FILE_A; f <= MaxFile; ++f) - for (int i = 0; i < Sides; i++) { - (d = item(p, i, f).precomp)->blockLength = (uint16_t*)data; + for (File f = FILE_A; f <= maxFile; ++f) + for (int i = 0; i < sides; i++) { + (d = e.get(i, f))->blockLength = (uint16_t*)data; data += d->blockLengthSize * sizeof(uint16_t); } - for (File f = FILE_A; f <= MaxFile; ++f) - for (int i = 0; i < Sides; i++) { + for (File f = FILE_A; f <= maxFile; ++f) + for (int i = 0; i < sides; i++) { data = (uint8_t*)(((uintptr_t)data + 0x3F) & ~0x3F); // 64 byte alignment - (d = item(p, i, f).precomp)->data = data; + (d = e.get(i, f))->data = data; data += d->blocksNum * d->sizeofBlock; } } -template -void* init(Entry& e, const Position& pos) { - - const bool IsWDL = std::is_same::value; +// If the TB file corresponding to the given position is already memory mapped +// then return its base address, otherwise try to memory map and init it. Called +// at every probe, memory map and init only at first access. Function is thread +// safe and can be called concurrently. +template +void* mapped(TBTable& e, const Position& pos) { - static Mutex mutex; + static std::mutex mutex; - // Avoid a thread reads 'ready' == true while another is still in do_init(), - // this could happen due to compiler reordering. + // Use 'acquire' to avoid a thread reading 'ready' == true while + // another is still working. (compiler reordering may cause this). if (e.ready.load(std::memory_order_acquire)) - return e.baseAddress; + return e.baseAddress; // Could be nullptr if file does not exist - std::unique_lock lk(mutex); + std::scoped_lock lk(mutex); if (e.ready.load(std::memory_order_relaxed)) // Recheck under lock return e.baseAddress; @@ -1148,30 +1155,28 @@ void* init(Entry& e, const Position& pos) { b += std::string(popcount(pos.pieces(BLACK, pt)), PieceToChar[pt]); } - const uint8_t TB_MAGIC[][4] = { { 0xD7, 0x66, 0x0C, 0xA5 }, - { 0x71, 0xE8, 0x23, 0x5D } }; - fname = (e.key == pos.material_key() ? w + 'v' + b : b + 'v' + w) - + (IsWDL ? ".rtbw" : ".rtbz"); + + (Type == WDL ? ".rtbw" : ".rtbz"); + + uint8_t* data = TBFile(fname).map(&e.baseAddress, &e.mapping, Type); - uint8_t* data = TBFile(fname).map(&e.baseAddress, &e.mapping, TB_MAGIC[IsWDL]); if (data) - e.hasPawns ? do_init(e, e.pawnTable, data) : do_init(e, e.pieceTable, data); + set(e, data); e.ready.store(true, std::memory_order_release); return e.baseAddress; } -template::type> -T probe_table(const Position& pos, ProbeState* result, WDLScore wdl = WDLDraw) { +template::Ret> +Ret probe_table(const Position& pos, ProbeState* result, WDLScore wdl = WDLDraw) { - if (!(pos.pieces() ^ pos.pieces(KING))) - return T(WDLDraw); // KvK + if (pos.count() == 2) // KvK + return Ret(WDLDraw); - E* entry = EntryTable.get(pos.material_key()); + TBTable* entry = TBTables.get(pos.material_key()); - if (!entry || !init(*entry, pos)) - return *result = FAIL, T(); + if (!entry || !mapped(*entry, pos)) + return *result = FAIL, Ret(); return do_probe_table(pos, entry, wdl, result); } @@ -1185,11 +1190,11 @@ T probe_table(const Position& pos, ProbeState* result, WDLScore wdl = WDLDraw) { // All of this means that during probing, the engine must look at captures and probe // their results and must probe the position itself. The "best" result of these // probes is the correct result for the position. -// DTZ table don't store values when a following move is a zeroing winning move +// DTZ tables do not store values when a following move is a zeroing winning move // (winning capture or winning pawn move). Also DTZ store wrong values for positions // where the best move is an ep-move (even if losing). So in all these cases set // the state to ZEROING_BEST_MOVE. -template +template WDLScore search(Position& pos, ProbeState* result) { WDLScore value, bestValue = WDLLoss; @@ -1198,7 +1203,7 @@ WDLScore search(Position& pos, ProbeState* result) { auto moveList = MoveList(pos); size_t totalCount = moveList.size(), moveCount = 0; - for (const Move& move : moveList) + for (const Move move : moveList) { if ( !pos.capture(move) && (!CheckZeroingMoves || type_of(pos.moved_piece(move)) != PAWN)) @@ -1207,7 +1212,7 @@ WDLScore search(Position& pos, ProbeState* result) { moveCount++; pos.do_move(move, st); - value = -search(pos, result); + value = -search(pos, result); pos.undo_move(move); if (*result == FAIL) @@ -1237,7 +1242,7 @@ WDLScore search(Position& pos, ProbeState* result) { value = bestValue; else { - value = probe_table(pos, result); + value = probe_table(pos, result); if (*result == FAIL) return WDLDraw; @@ -1253,9 +1258,13 @@ WDLScore search(Position& pos, ProbeState* result) { } // namespace + +/// Tablebases::init() is called at startup and after every change to +/// "SyzygyPath" UCI option to (re)create the various tables. It is not thread +/// safe, nor it needs to be. void Tablebases::init(const std::string& paths) { - EntryTable.clear(); + TBTables.clear(); MaxCardinality = 0; TBFile::Paths = paths; @@ -1292,14 +1301,14 @@ void Tablebases::init(const std::string& paths) { if (MapA1D1D4[s1] == idx && (idx || s1 == SQ_B1)) // SQ_B1 is mapped to 0 { for (Square s2 = SQ_A1; s2 <= SQ_H8; ++s2) - if ((StepAttacksBB[KING][s1] | s1) & s2) + if ((PseudoAttacks[KING][s1] | s1) & s2) continue; // Illegal position else if (!off_A1H8(s1) && off_A1H8(s2) > 0) continue; // First on diagonal, second above else if (!off_A1H8(s1) && !off_A1H8(s2)) - bothOnDiagonal.push_back(std::make_pair(idx, s2)); + bothOnDiagonal.emplace_back(idx, s2); else MapKK[idx][s2] = code++; @@ -1314,7 +1323,7 @@ void Tablebases::init(const std::string& paths) { Binomial[0][0] = 1; for (int n = 1; n < 64; n++) // Squares - for (int k = 0; k < 6 && k <= n; ++k) // Pieces + for (int k = 0; k < 7 && k <= n; ++k) // Pieces Binomial[k][n] = (k > 0 ? Binomial[k - 1][n - 1] : 0) + (k < n ? Binomial[k ][n - 1] : 0); @@ -1324,9 +1333,9 @@ void Tablebases::init(const std::string& paths) { // among pawns with same file, the one with lowest rank. int availableSquares = 47; // Available squares when lead pawn is in a2 - // Init the tables for the encoding of leading pawns group: with 6-men TB we - // can have up to 4 leading pawns (KPPPPK). - for (int leadPawnsCnt = 1; leadPawnsCnt <= 4; ++leadPawnsCnt) + // Init the tables for the encoding of leading pawns group: with 7-men TB we + // can have up to 5 leading pawns (KPPPPPK). + for (int leadPawnsCnt = 1; leadPawnsCnt <= 5; ++leadPawnsCnt) for (File f = FILE_A; f <= FILE_D; ++f) { // Restart the index at every file because TB table is splitted @@ -1347,7 +1356,7 @@ void Tablebases::init(const std::string& paths) { if (leadPawnsCnt == 1) { MapPawns[sq] = availableSquares--; - MapPawns[sq ^ 7] = availableSquares--; // Horizontal flip + MapPawns[flip_file(sq)] = availableSquares--; } LeadPawnIdx[leadPawnsCnt][sq] = idx; idx += Binomial[leadPawnsCnt - 1][MapPawns[sq]]; @@ -1356,33 +1365,45 @@ void Tablebases::init(const std::string& paths) { LeadPawnsSize[leadPawnsCnt][f] = idx; } + // Add entries in TB tables if the corresponding ".rtbw" file exists for (PieceType p1 = PAWN; p1 < KING; ++p1) { - EntryTable.insert({KING, p1, KING}); + TBTables.add({KING, p1, KING}); for (PieceType p2 = PAWN; p2 <= p1; ++p2) { - EntryTable.insert({KING, p1, p2, KING}); - EntryTable.insert({KING, p1, KING, p2}); + TBTables.add({KING, p1, p2, KING}); + TBTables.add({KING, p1, KING, p2}); for (PieceType p3 = PAWN; p3 < KING; ++p3) - EntryTable.insert({KING, p1, p2, KING, p3}); + TBTables.add({KING, p1, p2, KING, p3}); for (PieceType p3 = PAWN; p3 <= p2; ++p3) { - EntryTable.insert({KING, p1, p2, p3, KING}); + TBTables.add({KING, p1, p2, p3, KING}); + + for (PieceType p4 = PAWN; p4 <= p3; ++p4) { + TBTables.add({KING, p1, p2, p3, p4, KING}); - for (PieceType p4 = PAWN; p4 <= p3; ++p4) - EntryTable.insert({KING, p1, p2, p3, p4, KING}); + for (PieceType p5 = PAWN; p5 <= p4; ++p5) + TBTables.add({KING, p1, p2, p3, p4, p5, KING}); - for (PieceType p4 = PAWN; p4 < KING; ++p4) - EntryTable.insert({KING, p1, p2, p3, KING, p4}); + for (PieceType p5 = PAWN; p5 < KING; ++p5) + TBTables.add({KING, p1, p2, p3, p4, KING, p5}); + } + + for (PieceType p4 = PAWN; p4 < KING; ++p4) { + TBTables.add({KING, p1, p2, p3, KING, p4}); + + for (PieceType p5 = PAWN; p5 <= p4; ++p5) + TBTables.add({KING, p1, p2, p3, KING, p4, p5}); + } } for (PieceType p3 = PAWN; p3 <= p1; ++p3) for (PieceType p4 = PAWN; p4 <= (p1 == p3 ? p2 : p3); ++p4) - EntryTable.insert({KING, p1, p2, KING, p3, p4}); + TBTables.add({KING, p1, p2, KING, p3, p4}); } } - sync_cout << "info string Found " << EntryTable.size() << " tablebases" << sync_endl; + sync_cout << "info string Found " << TBTables.size() << " tablebases" << sync_endl; } // Probe the WDL table for a particular position. @@ -1396,7 +1417,7 @@ void Tablebases::init(const std::string& paths) { WDLScore Tablebases::probe_wdl(Position& pos, ProbeState* result) { *result = OK; - return search(pos, result); + return search(pos, result); } // Probe the DTZ table for a particular position. @@ -1404,6 +1425,7 @@ WDLScore Tablebases::probe_wdl(Position& pos, ProbeState* result) { // The return value is from the point of view of the side to move: // n < -100 : loss, but draw under 50-move rule // -100 <= n < -1 : loss in n ply (assuming 50-move counter == 0) +// -1 : loss, the side to move is mated // 0 : draw // 1 < n <= 100 : win in n ply (assuming 50-move counter == 0) // 100 < n : win, but draw under 50-move rule @@ -1420,7 +1442,7 @@ WDLScore Tablebases::probe_wdl(Position& pos, ProbeState* result) { // If n = 100 immediately after a capture or pawn move, then the position // is also certainly a win, and during the whole phase until the next // capture or pawn move, the inequality to be preserved is -// dtz + 50-movecounter <= 100. +// dtz + 50-move-counter <= 100. // // In short, if a move is available resulting in dtz + 50-move-counter <= 99, // then do not accept moves leading to dtz + 50-move-counter == 100. @@ -1437,7 +1459,7 @@ int Tablebases::probe_dtz(Position& pos, ProbeState* result) { if (*result == ZEROING_BEST_MOVE) return dtz_before_zeroing(wdl); - int dtz = probe_table(pos, result, wdl); + int dtz = probe_table(pos, result, wdl); if (*result == FAIL) return 0; @@ -1450,7 +1472,7 @@ int Tablebases::probe_dtz(Position& pos, ProbeState* result) { StateInfo st; int minDTZ = 0xFFFF; - for (const Move& move : MoveList(pos)) + for (const Move move : MoveList(pos)) { bool zeroing = pos.capture(move) || type_of(pos.moved_piece(move)) == PAWN; @@ -1460,13 +1482,12 @@ int Tablebases::probe_dtz(Position& pos, ProbeState* result) { // otherwise we will get the dtz of the next move sequence. Search the // position after the move to get the score sign (because even in a // winning position we could make a losing capture or going for a draw). - dtz = zeroing ? -dtz_before_zeroing(search(pos, result)) + dtz = zeroing ? -dtz_before_zeroing(search(pos, result)) : -probe_dtz(pos, result); - pos.undo_move(move); - - if (*result == FAIL) - return 0; + // If the move mates, force minDTZ to 1 + if (dtz == 1 && pos.checkers() && MoveList(pos).size() == 0) + minDTZ = 1; // Convert result from 1-ply search. Zeroing moves are already accounted // by dtz_before_zeroing() that returns the DTZ of the previous move. @@ -1476,215 +1497,120 @@ int Tablebases::probe_dtz(Position& pos, ProbeState* result) { // Skip the draws and if we are winning only pick positive dtz if (dtz < minDTZ && sign_of(dtz) == sign_of(wdl)) minDTZ = dtz; - } - - // Special handle a mate position, when there are no legal moves, in this - // case return value is somewhat arbitrary, so stick to the original TB code - // that returns -1 in this case. - return minDTZ == 0xFFFF ? -1 : minDTZ; -} -// Check whether there has been at least one repetition of positions -// since the last capture or pawn move. -static int has_repeated(StateInfo *st) -{ - while (1) { - int i = 4, e = std::min(st->rule50, st->pliesFromNull); + pos.undo_move(move); - if (e < i) + if (*result == FAIL) return 0; - - StateInfo *stp = st->previous->previous; - - do { - stp = stp->previous->previous; - - if (stp->key == st->key) - return 1; - - i += 2; - } while (i <= e); - - st = st->previous; } + + // When there are no legal moves, the position is mate: we return -1 + return minDTZ == 0xFFFF ? -1 : minDTZ; } -// Use the DTZ tables to filter out moves that don't preserve the win or draw. -// If the position is lost, but DTZ is fairly high, only keep moves that -// maximise DTZ. -// -// A return value false indicates that not all probes were successful and that -// no moves were filtered out. -bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves, Value& score) -{ - ProbeState result; - int dtz = probe_dtz(pos, &result); - if (result == FAIL) - return false; +// Use the DTZ tables to rank root moves. +// +// A return value false indicates that not all probes were successful. +bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves) { + ProbeState result; StateInfo st; - // Probe each move - for (size_t i = 0; i < rootMoves.size(); ++i) { - Move move = rootMoves[i].pv[0]; - pos.do_move(move, st); - int v = 0; - - if (pos.checkers() && dtz > 0) { - ExtMove s[MAX_MOVES]; - - if (generate(pos, s) == s) - v = 1; - } + // Obtain 50-move counter for the root position + int cnt50 = pos.rule50_count(); - if (!v) { - if (st.rule50 != 0) { - v = -probe_dtz(pos, &result); - - if (v > 0) - ++v; - else if (v < 0) - --v; - } else { - v = -probe_wdl(pos, &result); - v = dtz_before_zeroing(WDLScore(v)); - } - } + // Check whether a position was repeated since the last zeroing move. + bool rep = pos.has_repeated(); - pos.undo_move(move); + int dtz, bound = Options["Syzygy50MoveRule"] ? 900 : 1; - if (result == FAIL) - return false; - - rootMoves[i].score = (Value)v; - } - - // Obtain 50-move counter for the root position. - // In Stockfish there seems to be no clean way, so we do it like this: - int cnt50 = st.previous->rule50; - - // Use 50-move counter to determine whether the root position is - // won, lost or drawn. - WDLScore wdl = WDLDraw; - - if (dtz > 0) - wdl = (dtz + cnt50 <= 100) ? WDLWin : WDLCursedWin; - else if (dtz < 0) - wdl = (-dtz + cnt50 <= 100) ? WDLLoss : WDLBlessedLoss; - - // Determine the score to report to the user. - score = WDL_to_value[wdl + 2]; - - // If the position is winning or losing, but too few moves left, adjust the - // score to show how close it is to winning or losing. - // NOTE: int(PawnValueEg) is used as scaling factor in score_to_uci(). - if (wdl == WDLCursedWin && dtz <= 100) - score = (Value)(((200 - dtz - cnt50) * int(PawnValueEg)) / 200); - else if (wdl == WDLBlessedLoss && dtz >= -100) - score = -(Value)(((200 + dtz - cnt50) * int(PawnValueEg)) / 200); - - // Now be a bit smart about filtering out moves. - size_t j = 0; - - if (dtz > 0) { // winning (or 50-move rule draw) - int best = 0xffff; - - for (size_t i = 0; i < rootMoves.size(); ++i) { - int v = rootMoves[i].score; + // Probe and rank each move + for (auto& m : rootMoves) + { + pos.do_move(m.pv[0], st); - if (v > 0 && v < best) - best = v; + // Calculate dtz for the current move counting from the root position + if (pos.rule50_count() == 0) + { + // In case of a zeroing move, dtz is one of -101/-1/0/1/101 + WDLScore wdl = -probe_wdl(pos, &result); + dtz = dtz_before_zeroing(wdl); } - - int max = best; - - // If the current phase has not seen repetitions, then try all moves - // that stay safely within the 50-move budget, if there are any. - if (!has_repeated(st.previous) && best + cnt50 <= 99) - max = 99 - cnt50; - - for (size_t i = 0; i < rootMoves.size(); ++i) { - int v = rootMoves[i].score; - - if (v > 0 && v <= max) - rootMoves[j++] = rootMoves[i]; + else + { + // Otherwise, take dtz for the new position and correct by 1 ply + dtz = -probe_dtz(pos, &result); + dtz = dtz > 0 ? dtz + 1 + : dtz < 0 ? dtz - 1 : dtz; } - } else if (dtz < 0) { // losing (or 50-move rule draw) - int best = 0; - for (size_t i = 0; i < rootMoves.size(); ++i) { - int v = rootMoves[i].score; + // Make sure that a mating move is assigned a dtz value of 1 + if ( pos.checkers() + && dtz == 2 + && MoveList(pos).size() == 0) + dtz = 1; - if (v < best) - best = v; - } + pos.undo_move(m.pv[0]); - // Try all moves, unless we approach or have a 50-move rule draw. - if (-best * 2 + cnt50 < 100) - return true; + if (result == FAIL) + return false; - for (size_t i = 0; i < rootMoves.size(); ++i) { - if (rootMoves[i].score == best) - rootMoves[j++] = rootMoves[i]; - } - } else { // drawing - // Try all moves that preserve the draw. - for (size_t i = 0; i < rootMoves.size(); ++i) { - if (rootMoves[i].score == 0) - rootMoves[j++] = rootMoves[i]; - } + // Better moves are ranked higher. Certain wins are ranked equally. + // Losing moves are ranked equally unless a 50-move draw is in sight. + int r = dtz > 0 ? (dtz + cnt50 <= 99 && !rep ? 1000 : 1000 - (dtz + cnt50)) + : dtz < 0 ? (-dtz * 2 + cnt50 < 100 ? -1000 : -1000 + (-dtz + cnt50)) + : 0; + m.tbRank = r; + + // Determine the score to be displayed for this move. Assign at least + // 1 cp to cursed wins and let it grow to 49 cp as the positions gets + // closer to a real win. + m.tbScore = r >= bound ? VALUE_MATE - MAX_PLY - 1 + : r > 0 ? Value((std::max( 3, r - 800) * int(PawnValueEg)) / 200) + : r == 0 ? VALUE_DRAW + : r > -bound ? Value((std::min(-3, r + 800) * int(PawnValueEg)) / 200) + : -VALUE_MATE + MAX_PLY + 1; } - rootMoves.resize(j, Search::RootMove(MOVE_NONE)); - return true; } -// Use the WDL tables to filter out moves that don't preserve the win or draw. + +// Use the WDL tables to rank root moves. // This is a fallback for the case that some or all DTZ tables are missing. // -// A return value false indicates that not all probes were successful and that -// no moves were filtered out. -bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves, Value& score) -{ - ProbeState result; +// A return value false indicates that not all probes were successful. +bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves) { - WDLScore wdl = Tablebases::probe_wdl(pos, &result); + static const int WDL_to_rank[] = { -1000, -899, 0, 899, 1000 }; - if (result == FAIL) - return false; + ProbeState result; + StateInfo st; - score = WDL_to_value[wdl + 2]; + bool rule50 = Options["Syzygy50MoveRule"]; - StateInfo st; + // Probe and rank each move + for (auto& m : rootMoves) + { + pos.do_move(m.pv[0], st); - int best = WDLLoss; + WDLScore wdl = -probe_wdl(pos, &result); - // Probe each move - for (size_t i = 0; i < rootMoves.size(); ++i) { - Move move = rootMoves[i].pv[0]; - pos.do_move(move, st); - WDLScore v = -Tablebases::probe_wdl(pos, &result); - pos.undo_move(move); + pos.undo_move(m.pv[0]); if (result == FAIL) return false; - rootMoves[i].score = (Value)v; + m.tbRank = WDL_to_rank[wdl + 2]; - if (v > best) - best = v; + if (!rule50) + wdl = wdl > WDLDraw ? WDLWin + : wdl < WDLDraw ? WDLLoss : WDLDraw; + m.tbScore = WDL_to_value[wdl + 2]; } - size_t j = 0; - - for (size_t i = 0; i < rootMoves.size(); ++i) { - if (rootMoves[i].score == best) - rootMoves[j++] = rootMoves[i]; - } - - rootMoves.resize(j, Search::RootMove(MOVE_NONE)); - return true; } + +} // namespace Stockfish