X-Git-Url: https://git.sesse.net/?p=stockfish;a=blobdiff_plain;f=src%2Fsyzygy%2Ftbprobe.cpp;h=61ba8121959ae61b0b8de296b1022730f424386d;hp=764574f75e2976758bbeb2fc415afe3b47978840;hb=HEAD;hpb=a24f28be8567c2527b154ef981090368a2bd8f76 diff --git a/src/syzygy/tbprobe.cpp b/src/syzygy/tbprobe.cpp index 764574f7..e2363157 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) 2016-2020 Marco Costalba, Lucas Braesch + Copyright (C) 2004-2023 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 @@ -17,102 +16,117 @@ along with this program. If not, see . */ +#include "tbprobe.h" + +#include #include #include +#include #include -#include // For std::memset and std::memcpy +#include +#include #include #include +#include #include -#include +#include #include +#include #include -#include +#include +#include #include "../bitboard.h" +#include "../misc.h" #include "../movegen.h" #include "../position.h" #include "../search.h" #include "../types.h" #include "../uci.h" -#include "tbprobe.h" - #ifndef _WIN32 -#include -#include -#include -#include + #include + #include + #include #else -#define WIN32_LEAN_AND_MEAN -#ifndef NOMINMAX -# define NOMINMAX // Disable macros min() and max() -#endif -#include + #define WIN32_LEAN_AND_MEAN + #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 +constexpr int TBPIECES = 7; // Max number of supported pieces +constexpr int MAX_DTZ = + 1 << 18; // Max DTZ supported, large enough to deal with the syzygy TB limit. -enum { BigEndian, LittleEndian }; -enum TBType { WDL, DTZ }; // Used as template parameter +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, Wide = 16, 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 Square(int(s) ^ i); } +inline Square operator^(Square s, int i) { return Square(int(s) ^ i); } -const std::string PieceToChar = " PNBRQK pnbrqk"; +constexpr std::string_view PieceToChar = " PNBRQK pnbrqk"; int MapPawns[SQUARE_NB]; int MapB1H1H7[SQUARE_NB]; int MapA1D1D4[SQUARE_NB]; -int MapKK[10][SQUARE_NB]; // [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] +int Binomial[6][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); } - -constexpr Value WDL_to_value[] = { - -VALUE_MATE + MAX_PLY + 1, - VALUE_DRAW - 2, - VALUE_DRAW, - VALUE_DRAW + 2, - VALUE_MATE - MAX_PLY - 1 -}; +int off_A1H8(Square sq) { return int(rank_of(sq)) - file_of(sq); } + +constexpr Value WDL_to_value[] = {-VALUE_MATE + MAX_PLY + 1, VALUE_DRAW - 2, VALUE_DRAW, + VALUE_DRAW + 2, VALUE_MATE - MAX_PLY - 1}; template -inline void swap_endian(T& x) -{ - static_assert(std::is_unsigned::value, "Argument of swap_endian not unsigned"); +inline void swap_endian(T& x) { + static_assert(std::is_unsigned_v, "Argument of swap_endian not unsigned"); - uint8_t tmp, *c = (uint8_t*)&x; + 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_endian(uint8_t&) {} - -template T number(void* addr) -{ - static const union { uint32_t i; char c[4]; } Le = { 0x01020304 }; - static const bool IsLittleEndian = (Le.c[0] == 4); +template<> +inline void swap_endian(uint8_t&) {} +template +T number(void* addr) { T v; - if ((uintptr_t)addr & (alignof(T) - 1)) // Unaligned pointer (very rare) + if (uintptr_t(addr) & (alignof(T) - 1)) // Unaligned pointer (very rare) std::memcpy(&v, addr, sizeof(T)); else - v = *((T*)addr); + v = *((T*) addr); if (LE != IsLittleEndian) swap_endian(v); @@ -123,18 +137,20 @@ template T number(void* addr) // 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 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) { +template +int sign_of(T val) { return (T(0) < val) - (val < T(0)); } -// Numbers in little endian used by sparseIndex[] to point into blockLength[] +// 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 @@ -142,18 +158,22 @@ struct SparseEntry { static_assert(sizeof(SparseEntry) == 6, "SparseEntry must be 6 bytes"); -typedef uint16_t Sym; // Huffman symbol +using Sym = uint16_t; // Huffman symbol struct LR { - enum Side { Left, Right }; + enum Side { + Left, + Right + }; - 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. + uint8_t lr[3]; // The first 12 bits is the left-hand symbol, the second 12 + // bits is the right-hand symbol. If the 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)); + return S == Left ? ((lr[1] & 0xF) << 8) | lr[0] + : S == Right ? (lr[2] << 4) | (lr[1] >> 4) + : (assert(false), Sym(-1)); } }; @@ -168,11 +188,11 @@ static_assert(sizeof(LR) == 3, "LR tree entry must be 3 bytes"); // 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 { +class TBFile: public std::ifstream { std::string fname; -public: + public: // Look for and open the file among the Paths directories where the .rtbw // and .rtbz files can be found. Multiple directories are separated by ";" // on Windows and by ":" on Unix-based operating systems. @@ -189,9 +209,10 @@ public: constexpr char SepChar = ';'; #endif std::stringstream ss(Paths); - std::string path; + std::string path; - while (std::getline(ss, path, SepChar)) { + while (std::getline(ss, path, SepChar)) + { fname = path + "/" + f; std::ifstream::open(fname); if (is_open()) @@ -199,17 +220,14 @@ public: } } - // Memory map the file and check it. File should be already open and will be - // closed after mapping. + // Memory map the file and check it. uint8_t* map(void** baseAddress, uint64_t* mapping, TBType type) { - - assert(is_open()); - - close(); // Need to re-open to get native file descriptor + if (is_open()) + close(); // Need to re-open to get native file descriptor #ifndef _WIN32 struct stat statbuf; - int fd = ::open(fname.c_str(), O_RDONLY); + int fd = ::open(fname.c_str(), O_RDONLY); if (fd == -1) return *baseAddress = nullptr, nullptr; @@ -222,9 +240,11 @@ public: exit(EXIT_FAILURE); } - *mapping = statbuf.st_size; + *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) @@ -234,8 +254,8 @@ public: } #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_FLAG_RANDOM_ACCESS, nullptr); + HANDLE fd = CreateFileA(fname.c_str(), GENERIC_READ, FILE_SHARE_READ, nullptr, + OPEN_EXISTING, FILE_FLAG_RANDOM_ACCESS, nullptr); if (fd == INVALID_HANDLE_VALUE) return *baseAddress = nullptr, nullptr; @@ -258,7 +278,7 @@ public: exit(EXIT_FAILURE); } - *mapping = (uint64_t)mmap; + *mapping = uint64_t(mmap); *baseAddress = MapViewOfFile(mmap, FILE_MAP_READ, 0, 0, 0); if (!*baseAddress) @@ -268,10 +288,9 @@ public: exit(EXIT_FAILURE); } #endif - uint8_t* data = (uint8_t*)*baseAddress; + uint8_t* data = (uint8_t*) *baseAddress; - constexpr uint8_t Magics[][4] = { { 0xD7, 0x66, 0x0C, 0xA5 }, - { 0x71, 0xE8, 0x23, 0x5D } }; + constexpr uint8_t Magics[][4] = {{0xD7, 0x66, 0x0C, 0xA5}, {0x71, 0xE8, 0x23, 0x5D}}; if (memcmp(data, Magics[type == WDL], 4)) { @@ -280,7 +299,7 @@ public: return *baseAddress = nullptr, nullptr; } - return data + 4; // Skip Magics's header + return data + 4; // Skip Magics's header } static void unmap(void* baseAddress, uint64_t mapping) { @@ -289,36 +308,38 @@ public: munmap(baseAddress, mapping); #else UnmapViewOfFile(baseAddress); - CloseHandle((HANDLE)mapping); + CloseHandle((HANDLE) mapping); #endif } }; std::string TBFile::Paths; -// 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 contains low-level indexing information to access TB data. +// There are 8, 4, or 2 PairsData records for each TBTable, according to the 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) + 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. @@ -327,27 +348,27 @@ struct PairsData { // first access, when the corresponding file is memory mapped. template struct TBTable { - typedef typename std::conditional::type Ret; + using Ret = std::conditional_t; 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) {} + 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); @@ -358,26 +379,26 @@ struct TBTable { }; template<> -TBTable::TBTable(const std::string& code) : TBTable() { +TBTable::TBTable(const std::string& code) : + TBTable() { StateInfo st; - Position pos; + Position pos; - key = pos.set(code, WHITE, &st).material_key(); + key = pos.set(code, WHITE, &st).material_key(); pieceCount = pos.count(); - hasPawns = pos.pieces(PAWN); + hasPawns = pos.pieces(PAWN); hasUniquePieces = false; - for (Color c : { WHITE, BLACK }) + for (Color c : {WHITE, BLACK}) for (PieceType pt = PAWN; pt < KING; ++pt) if (popcount(pos.pieces(c, pt)) == 1) hasUniquePieces = true; // 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)); + // is the side with fewer pawns because this leads to better compression. + bool c = !pos.count(BLACK) + || (pos.count(WHITE) && pos.count(BLACK) >= pos.count(WHITE)); pawnCount[0] = pos.count(c ? WHITE : BLACK); pawnCount[1] = pos.count(c ? BLACK : WHITE); @@ -386,36 +407,36 @@ TBTable::TBTable(const std::string& code) : TBTable() { } template<> -TBTable::TBTable(const TBTable& wdl) : TBTable() { +TBTable::TBTable(const TBTable& wdl) : + TBTable() { // Use the corresponding WDL table to avoid recalculating all from scratch - key = wdl.key; - key2 = wdl.key2; - pieceCount = wdl.pieceCount; - hasPawns = wdl.hasPawns; + 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]; + 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 +// 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; + struct Entry { + Key key; TBTable* wdl; TBTable* dtz; - template + template TBTable* get() const { - return (TBTable*)(Type == WDL ? (void*)wdl : (void*)dtz); + 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 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]; @@ -424,23 +445,26 @@ class TBTables { std::deque> dtzTable; void insert(Key key, TBTable* wdl, TBTable* dtz) { - uint32_t homeBucket = (uint32_t)key & (Size - 1); - Entry entry{ key, wdl, dtz }; + uint32_t homeBucket = uint32_t(key) & (Size - 1); + Entry entry{key, wdl, dtz}; // Ensure last element is empty to avoid overflow when looking up - for (uint32_t bucket = homeBucket; bucket < Size + Overflow - 1; ++bucket) { + for (uint32_t bucket = homeBucket; bucket < Size + Overflow - 1; ++bucket) + { Key otherKey = hashTable[bucket].key; - if (otherKey == key || !hashTable[bucket].get()) { + 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) { + uint32_t otherHomeBucket = uint32_t(otherKey) & (Size - 1); + if (otherHomeBucket > homeBucket) + { std::swap(entry, hashTable[bucket]); - key = otherKey; + key = otherKey; homeBucket = otherHomeBucket; } } @@ -448,10 +472,11 @@ class TBTables { exit(EXIT_FAILURE); } -public: + public: template TBTable* get(Key key) { - for (const Entry* entry = &hashTable[(uint32_t)key & (Size - 1)]; ; ++entry) { + for (const Entry* entry = &hashTable[uint32_t(key) & (Size - 1)];; ++entry) + { if (entry->key == key || !entry->get()) return entry->get(); } @@ -463,7 +488,7 @@ public: dtzTable.clear(); } size_t size() const { return wdlTable.size(); } - void add(const std::vector& pieces); + void add(const std::vector& pieces); }; TBTables TBTables; @@ -477,20 +502,20 @@ void TBTables::add(const std::vector& pieces) { for (PieceType pt : pieces) code += PieceToChar[pt]; - TBFile file(code.insert(code.find('K', 1), "v") + ".rtbw"); // KRK -> KRvK + TBFile file(code.insert(code.find('K', 1), "v") + ".rtbw"); // KRK -> KRvK - if (!file.is_open()) // Only WDL file is checked + if (!file.is_open()) // Only WDL file is checked return; file.close(); - MaxCardinality = std::max((int)pieces.size(), MaxCardinality); + MaxCardinality = std::max(int(pieces.size()), MaxCardinality); 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().key, &wdlTable.back(), &dtzTable.back()); insert(wdlTable.back().key2, &wdlTable.back(), &dtzTable.back()); } @@ -506,9 +531,9 @@ void TBTables::add(const std::vector& pieces) { // mostly-draw or mostly-win tables this can leave many 64-byte blocks only half-filled, so // in such cases blocks are 32 bytes long. The blocks of DTZ tables are up to 1024 bytes long. // The generator picks the size that leads to the smallest table. The "book" of symbols and -// Huffman codes is the same for all blocks in the table. A non-symmetric pawnless TB file +// Huffman codes are the same for all blocks in the table. A non-symmetric pawnless TB file // will have one table for wtm and one for btm, a TB file with pawns will have tables per -// file a,b,c,d also in this case one set for wtm and one for btm. +// file a,b,c,d also, in this case, one set for wtm and one for btm. int decompress_pairs(PairsData* d, uint64_t idx) { // Special case where all table positions store the same value @@ -533,10 +558,10 @@ int decompress_pairs(PairsData* d, uint64_t idx) { 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); + uint32_t block = number(&d->sparseIndex[k].block); + int offset = number(&d->sparseIndex[k].offset); - // Now compute the difference idx - I(k). From definition of k we know that + // Now compute the difference idx - I(k). From the definition of k, we know that // // idx = k * d->span + idx % d->span (2) // @@ -546,7 +571,7 @@ int decompress_pairs(PairsData* d, uint64_t idx) { // Sum the above to offset to find the offset corresponding to our idx offset += diff; - // Move to previous/next block, until we reach the correct block that contains idx, + // Move to the previous/next block, until we reach the correct block that contains idx, // that is when 0 <= offset <= d->blockLength[block] while (offset < 0) offset += d->blockLength[--block] + 1; @@ -555,17 +580,19 @@ 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 + ((uint64_t)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 - // is at the beginning of this 64 bits sequence. - uint64_t buf64 = number(ptr); ptr += 2; + // is at the beginning of this 64-bit sequence. + uint64_t buf64 = number(ptr); + ptr += 2; int buf64Size = 64; Sym sym; - while (true) { - int len = 0; // This is the symbol length - d->min_sym_len + while (true) + { + int len = 0; // This is the symbol length - d->min_sym_len // Now get the symbol length. For any symbol s64 of length l right-padded // to 64 bits we know that d->base64[l-1] >= s64 >= d->base64[l] so we @@ -581,38 +608,40 @@ int decompress_pairs(PairsData* d, uint64_t idx) { // Now add the value of the lowest symbol of length len to get our symbol sym += number(&d->lowestSym[len]); - // If our offset is within the number of values represented by symbol sym - // we are done... + // If our offset is within the number of values represented by symbol sym, + // we are done. if (offset < d->symlen[sym] + 1) break; // ...otherwise update the offset and continue to iterate offset -= d->symlen[sym] + 1; - len += d->minSymLen; // Get the real length - buf64 <<= len; // Consume the just processed symbol + len += d->minSymLen; // Get the real length + buf64 <<= len; // Consume the just processed symbol buf64Size -= len; - if (buf64Size <= 32) { // Refill the buffer + if (buf64Size <= 32) + { // Refill the buffer buf64Size += 32; - buf64 |= (uint64_t)number(ptr++) << (64 - buf64Size); + buf64 |= uint64_t(number(ptr++)) << (64 - buf64Size); } } - // Ok, now we have our symbol that expands into d->symlen[sym] + 1 symbols. + // Now we have our symbol that expands into d->symlen[sym] + 1 symbols. // We binary-search for our value recursively expanding into the left and // right child symbols until we reach a leaf node where symlen[sym] + 1 == 1 // that will store the value we need. - while (d->symlen[sym]) { - + while (d->symlen[sym]) + { Sym left = d->btree[sym].get(); // If a symbol contains 36 sub-symbols (d->symlen[sym] + 1 = 36) and // expands in a pair (d->symlen[left] = 23, d->symlen[right] = 11), then - // we know that, for instance the ten-th value (offset = 10) will be on + // we know that, for instance, the tenth value (offset = 10) will be on // the left side because in Recursive Pairing child symbols are adjacent. if (offset < d->symlen[left] + 1) sym = left; - else { + else + { offset -= d->symlen[left] + 1; sym = d->btree[sym].get(); } @@ -626,68 +655,79 @@ bool check_dtz_stm(TBTable*, int, File) { return true; } 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); + return (flags & TBFlag::STM) == stm || ((entry->key == entry->key2) && !entry->hasPawns); } // DTZ scores are sorted by frequency of occurrence and then assigned the // 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. +// the original values are stored in the TB file and read during map[] init. WDLScore map_score(TBTable*, File, int value, WDLScore) { return WDLScore(value - 2); } int map_score(TBTable* entry, File f, int value, WDLScore wdl) { - constexpr int WDLMap[] = { 1, 3, 0, 2, 0 }; + constexpr int WDLMap[] = {1, 3, 0, 2, 0}; auto flags = entry->get(0, f)->flags; - uint8_t* map = entry->map; + uint8_t* map = entry->map; uint16_t* idx = entry->get(0, f)->map_idx; - if (flags & TBFlag::Mapped) { + if (flags & TBFlag::Mapped) + { if (flags & TBFlag::Wide) - value = ((uint16_t *)map)[idx[WDLMap[wdl + 2]] + value]; + 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. - if ( (wdl == WDLWin && !(flags & TBFlag::WinPlies)) - || (wdl == WDLLoss && !(flags & TBFlag::LossPlies)) - || wdl == WDLCursedWin - || wdl == WDLBlessedLoss) + // want to return plies, so we have to convert to plies when needed. + if ((wdl == WDLWin && !(flags & TBFlag::WinPlies)) + || (wdl == WDLLoss && !(flags & TBFlag::LossPlies)) || wdl == WDLCursedWin + || wdl == WDLBlessedLoss) value *= 2; return value + 1; } +// A temporary fix for the compiler bug with AVX-512. (#4450) +#ifdef USE_AVX512 + #if defined(__clang__) && defined(__clang_major__) && __clang_major__ >= 15 + #define CLANG_AVX512_BUG_FIX __attribute__((optnone)) + #endif +#endif + +#ifndef CLANG_AVX512_BUG_FIX + #define CLANG_AVX512_BUG_FIX +#endif + // Compute a unique index out of a position and use it to probe the TB file. To -// encode k pieces of same type and color, first sort the pieces by square in +// encode k pieces of the same type and color, first sort the pieces by square in // ascending order s1 <= s2 <= ... <= sk then compute the unique index as: // // idx = Binomial[1][s1] + Binomial[2][s2] + ... + Binomial[k][sk] // template -Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* result) { +CLANG_AVX512_BUG_FIX Ret +do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* result) { - Square squares[TBPIECES]; - Piece pieces[TBPIECES]; - uint64_t idx; - int next = 0, size = 0, leadPawnsCnt = 0; + Square squares[TBPIECES]; + Piece pieces[TBPIECES]; + uint64_t idx; + int next = 0, size = 0, leadPawnsCnt = 0; PairsData* d; - Bitboard b, leadPawns = 0; - File tbFile = FILE_A; + Bitboard b, leadPawns = 0; + File tbFile = FILE_A; // A given TB entry like KRK has associated two material keys: KRvk and Kvkr. // If both sides have the same pieces keys are equal. In this case TB tables - // only store the 'white to move' case, so if the position to lookup has black + // only stores the 'white to move' case, so if the position to lookup has black // to move, we need to switch the color and flip the squares before to lookup. bool symmetricBlackToMove = (entry->key == entry->key2 && pos.side_to_move()); - // TB files are calculated for white as stronger side. For instance we have - // KRvK, not KvKR. A position where stronger side is white will have its - // material key == entry->key, otherwise we have to switch the color and + // TB files are calculated for white as the stronger side. For instance, we + // have KRvK, not KvKR. A position where the stronger side is white will have + // its material key == entry->key, otherwise we have to switch the color and // flip the squares before to lookup. bool blackStronger = (pos.material_key() != entry->key); @@ -698,7 +738,8 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu // For pawns, TB files store 4 separate tables according if leading pawn is on // file a, b, c or d after reordering. The leading pawn is the one with maximum // MapPawns[] value, that is the one most toward the edges and with lowest rank. - if (entry->hasPawns) { + if (entry->hasPawns) + { // 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. @@ -708,7 +749,7 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu leadPawns = b = pos.pieces(color_of(pc), PAWN); do - squares[size++] = pop_lsb(&b) ^ flipSquares; + squares[size++] = pop_lsb(b) ^ flipSquares; while (b); leadPawnsCnt = size; @@ -727,9 +768,10 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu // 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; - do { - Square s = pop_lsb(&b); - squares[size] = s ^ flipSquares; + do + { + Square s = pop_lsb(b); + squares[size] = s ^ flipSquares; pieces[size++] = Piece(pos.piece_on(s) ^ flipColor); } while (b); @@ -756,18 +798,19 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu // Encode leading pawns starting with the one with minimum MapPawns[] and // proceeding in ascending order. - if (entry->hasPawns) { + 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]]]; - goto encode_remaining; // With pawns we have finished special treatments + goto encode_remaining; // With pawns we have finished special treatments } - // In positions withouth pawns, we further flip the squares to ensure leading + // In positions without pawns, we further flip the squares to ensure leading // piece is below RANK_5. if (rank_of(squares[0]) > RANK_4) for (int i = 0; i < size; ++i) @@ -775,11 +818,12 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu // Look for the first piece of the leading group not on the A1-D4 diagonal // and ensure it is mapped below the diagonal. - for (int i = 0; i < d->groupLen[0]; ++i) { + for (int i = 0; i < d->groupLen[0]; ++i) + { if (!off_A1H8(squares[i])) continue; - if (off_A1H8(squares[i]) > 0) // A1-H8 diagonal flip: SQ_A3 -> SQ_C1 + 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; @@ -810,43 +854,38 @@ Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* resu // Rs "together" in 62 * 61 / 2 ways (we divide by 2 because rooks can be // swapped and still get the same position.) // - // In case we have at least 3 unique pieces (inlcuded kings) we encode them + // In case we have at least 3 unique pieces (including kings) we encode them // together. - if (entry->hasUniquePieces) { + if (entry->hasUniquePieces) + { - int adjust1 = squares[1] > squares[0]; + int adjust1 = squares[1] > squares[0]; int adjust2 = (squares[2] > squares[0]) + (squares[2] > squares[1]); // First piece is below a1-h8 diagonal. MapA1D1D4[] maps the b1-d1-d3 // triangle to 0...5. There are 63 squares for second piece and and 62 // (mapped to 0...61) for the third. if (off_A1H8(squares[0])) - idx = ( MapA1D1D4[squares[0]] * 63 - + (squares[1] - adjust1)) * 62 - + squares[2] - adjust2; + idx = (MapA1D1D4[squares[0]] * 63 + (squares[1] - adjust1)) * 62 + squares[2] - adjust2; - // First piece is on a1-h8 diagonal, second below: map this occurence to + // First piece is on a1-h8 diagonal, second below: map this occurrence to // 6 to differentiate from the above case, rank_of() maps a1-d4 diagonal // to 0...3 and finally MapB1H1H7[] maps the b1-h1-h7 triangle to 0..27. else if (off_A1H8(squares[1])) - idx = ( 6 * 63 + rank_of(squares[0]) * 28 - + MapB1H1H7[squares[1]]) * 62 - + squares[2] - adjust2; + idx = (6 * 63 + rank_of(squares[0]) * 28 + MapB1H1H7[squares[1]]) * 62 + squares[2] + - adjust2; // First two pieces are on a1-h8 diagonal, third below else if (off_A1H8(squares[2])) - idx = 6 * 63 * 62 + 4 * 28 * 62 - + rank_of(squares[0]) * 7 * 28 - + (rank_of(squares[1]) - adjust1) * 28 - + MapB1H1H7[squares[2]]; + idx = 6 * 63 * 62 + 4 * 28 * 62 + rank_of(squares[0]) * 7 * 28 + + (rank_of(squares[1]) - adjust1) * 28 + MapB1H1H7[squares[2]]; // All 3 pieces on the diagonal a1-h8 else - idx = 6 * 63 * 62 + 4 * 28 * 62 + 4 * 7 * 28 - + rank_of(squares[0]) * 7 * 6 - + (rank_of(squares[1]) - adjust1) * 6 - + (rank_of(squares[2]) - adjust2); - } else + idx = 6 * 63 * 62 + 4 * 28 * 62 + 4 * 7 * 28 + rank_of(squares[0]) * 7 * 6 + + (rank_of(squares[1]) - adjust1) * 6 + (rank_of(squares[2]) - adjust2); + } + else // We don't have at least 3 unique pieces, like in KRRvKBB, just map // the kings. idx = MapKK[MapA1D1D4[squares[0]]][squares[1]]; @@ -855,19 +894,19 @@ encode_remaining: idx *= d->groupIdx[0]; Square* groupSq = squares + d->groupLen[0]; - // Encode remainig pawns then pieces according to square, in ascending order + // Encode remaining pawns and then pieces according to square, in ascending order 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 - // groups (similar to what done earlier for leading group pieces). + // groups (similar to what was done earlier for leading group pieces). for (int i = 0; i < d->groupLen[next]; ++i) { - auto f = [&](Square s) { return groupSq[i] > s; }; + auto f = [&](Square s) { return groupSq[i] > s; }; auto adjust = std::count_if(squares, groupSq, f); n += Binomial[i + 1][groupSq[i] - adjust - 8 * remainingPawns]; } @@ -882,8 +921,8 @@ encode_remaining: } // Group together pieces that will be encoded together. The general rule is that -// a group contains pieces of same type and color. The exception is the leading -// group that, in case of positions withouth pawns, can be formed by 3 different +// a group contains pieces of the same type and color. The exception is the leading +// group that, in case of positions without pawns, can be formed by 3 different // pieces (default) or by the king pair when there is not a unique piece apart // from the kings. When there are pawns, pawns are always first in pieces[]. // @@ -905,7 +944,7 @@ void set_groups(T& e, PairsData* d, int order[], File f) { else d->groupLen[++n] = 1; - d->groupLen[++n] = 0; // Zero-terminated + d->groupLen[++n] = 0; // Zero-terminated // The sequence in pieces[] defines the groups, but not the order in which // they are encoded. If the pieces in a group g can be combined on the board @@ -915,27 +954,26 @@ void set_groups(T& e, PairsData* d, int order[], File f) { // // This ensures unique encoding for the whole position. The order of the // groups is a per-table parameter and could not follow the canonical leading - // pawns/pieces -> remainig pawns -> remaining pieces. In particular the + // pawns/pieces -> remaining 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.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; + 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; for (int k = 0; next < n || k == order[0] || k == order[1]; ++k) - if (k == order[0]) // Leading pawns or pieces + if (k == order[0]) // Leading pawns or pieces { d->groupIdx[0] = idx; - idx *= e.hasPawns ? LeadPawnsSize[d->groupLen[0]][f] - : e.hasUniquePieces ? 31332 : 462; + idx *= e.hasPawns ? LeadPawnsSize[d->groupLen[0]][f] : e.hasUniquePieces ? 31332 : 462; } - else if (k == order[1]) // Remaining pawns + else if (k == order[1]) // Remaining pawns { d->groupIdx[1] = idx; idx *= Binomial[d->groupLen[1]][48 - d->groupLen[0]]; } - else // Remainig pieces + else // Remaining pieces { d->groupIdx[next] = idx; idx *= Binomial[d->groupLen[next]][freeSquares]; @@ -945,13 +983,13 @@ void set_groups(T& e, PairsData* d, int order[], File f) { d->groupIdx[n] = idx; } -// In Recursive Pairing each symbol represents a pair of childern symbols. So +// In Recursive Pairing each symbol represents a pair of children symbols. So // read d->btree[] symbols data and expand each one in his left and right child -// symbol until reaching the leafs that represent the symbol value. +// symbol until reaching the leaves that represent the symbol value. uint8_t set_symlen(PairsData* d, Sym s, std::vector& visited) { - visited[s] = true; // We can set it now because tree is acyclic - Sym sr = d->btree[s].get(); + visited[s] = true; // We can set it now because tree is acyclic + Sym sr = d->btree[s].get(); if (sr == 0xFFF) return 0; @@ -971,10 +1009,11 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) { d->flags = *data++; - if (d->flags & TBFlag::SingleValue) { + if (d->flags & TBFlag::SingleValue) + { d->blocksNum = d->blockLengthSize = 0; - d->span = d->sparseIndexSize = 0; // Broken MSVC zero-init - d->minSymLen = *data++; // Here we store the single value + d->span = d->sparseIndexSize = 0; // Broken MSVC zero-init + d->minSymLen = *data++; // Here we store the single value return data; } @@ -982,47 +1021,57 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) { // element stores the biggest index that is the tb size. uint64_t tbSize = d->groupIdx[std::find(d->groupLen, d->groupLen + 7, 0) - d->groupLen]; - d->sizeofBlock = 1ULL << *data++; - d->span = 1ULL << *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. + d->sizeofBlock = 1ULL << *data++; + d->span = 1ULL << *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. d->maxSymLen = *data++; d->minSymLen = *data++; - d->lowestSym = (Sym*)data; + d->lowestSym = (Sym*) data; d->base64.resize(d->maxSymLen - d->minSymLen + 1); + // See https://en.wikipedia.org/wiki/Huffman_coding // The canonical code is ordered such that longer symbols (in terms of - // the number of bits of their Huffman code) have lower numeric value, + // the number of bits of their Huffman code) have a lower numeric value, // 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 - for (int i = d->base64.size() - 2; i >= 0; --i) { + + // Implementation note: we first cast the unsigned size_t "base64.size()" + // to a signed int "base64_size" variable and then we are able to subtract 2, + // avoiding unsigned overflow warnings. + + int base64_size = static_cast(d->base64.size()); + for (int i = base64_size - 2; i >= 0; --i) + { d->base64[i] = (d->base64[i + 1] + number(&d->lowestSym[i]) - - number(&d->lowestSym[i + 1])) / 2; + - number(&d->lowestSym[i + 1])) + / 2; - assert(d->base64[i] * 2 >= d->base64[i+1]); + assert(d->base64[i] * 2 >= d->base64[i + 1]); } // Now left-shift by an amount so that d->base64[i] gets shifted 1 bit more // than d->base64[i+1] and given the above assert condition, we ensure that // d->base64[i] >= d->base64[i+1]. Moreover for any symbol s64 of length i // and right-padded to 64 bits holds d->base64[i-1] >= s64 >= d->base64[i]. - for (size_t i = 0; i < d->base64.size(); ++i) - d->base64[i] <<= 64 - i - d->minSymLen; // Right-padding to 64 bits + for (int i = 0; i < base64_size; ++i) + d->base64[i] <<= 64 - i - d->minSymLen; // Right-padding to 64 bits - data += d->base64.size() * sizeof(Sym); - d->symlen.resize(number(data)); data += sizeof(uint16_t); - d->btree = (LR*)data; + data += base64_size * sizeof(Sym); + d->symlen.resize(number(data)); + data += sizeof(uint16_t); + d->btree = (LR*) data; // 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. - // See http://www.larsson.dogma.net/dcc99.pdf + // See https://web.archive.org/web/20201106232444/http://www.larsson.dogma.net/dcc99.pdf std::vector visited(d->symlen.size()); for (Sym sym = 0; sym < d->symlen.size(); ++sym) @@ -1038,67 +1087,77 @@ uint8_t* set_dtz_map(TBTable& e, uint8_t* data, File maxFile) { e.map = data; - for (File f = FILE_A; f <= maxFile; ++f) { + for (File f = FILE_A; f <= maxFile; ++f) + { 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); + 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); + 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 + return data += uintptr_t(data) & 1; // Word alignment } -// Populate entry's PairsData records with data from the just memory mapped file. +// 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 }; + enum { + Split = 1, + HasPawns = 2 + }; - assert(e.hasPawns == bool(*data & HasPawns)); + assert(e.hasPawns == bool(*data & HasPawns)); assert((e.key != e.key2) == bool(*data & Split)); - data++; // First byte stores flags + data++; // First byte stores flags - const int sides = T::Sides == 2 && (e.key != e.key2) ? 2 : 1; + 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.pawnCount[1]; // Pawns on both sides + bool pp = e.hasPawns && e.pawnCount[1]; // Pawns on both sides 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++) *e.get(i, f) = PairsData(); - int order[][2] = { { *data & 0xF, pp ? *(data + 1) & 0xF : 0xF }, - { *data >> 4, pp ? *(data + 1) >> 4 : 0xF } }; + 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++) - e.get(i, f)->pieces[k] = Piece(i ? *data >> 4 : *data & 0xF); + e.get(i, f)->pieces[k] = Piece(i ? *data >> 4 : *data & 0xF); for (int i = 0; i < sides; ++i) set_groups(e, e.get(i, f), order[i], f); } - data += (uintptr_t)data & 1; // Word alignment + data += uintptr_t(data) & 1; // Word alignment for (File f = FILE_A; f <= maxFile; ++f) for (int i = 0; i < sides; i++) @@ -1107,28 +1166,31 @@ void set(T& e, uint8_t* data) { data = set_dtz_map(e, data, maxFile); for (File f = FILE_A; f <= maxFile; ++f) - for (int i = 0; i < sides; i++) { - (d = e.get(i, f))->sparseIndex = (SparseEntry*)data; + 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 = e.get(i, f))->blockLength = (uint16_t*)data; + 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++) { - data = (uint8_t*)(((uintptr_t)data + 0x3F) & ~0x3F); // 64 byte alignment + for (int i = 0; i < sides; i++) + { + data = (uint8_t*) ((uintptr_t(data) + 0x3F) & ~0x3F); // 64 byte alignment (d = e.get(i, f))->data = data; data += d->blocksNum * d->sizeofBlock; } } -// 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 +// 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) { @@ -1138,22 +1200,23 @@ void* mapped(TBTable& e, const Position& pos) { // 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; // Could be nullptr if file does not exist + 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 + if (e.ready.load(std::memory_order_relaxed)) // Recheck under lock return e.baseAddress; // Pieces strings in decreasing order for each color, like ("KPP","KR") std::string fname, w, b; - for (PieceType pt = KING; pt >= PAWN; --pt) { + for (PieceType pt = KING; pt >= PAWN; --pt) + { w += std::string(popcount(pos.pieces(WHITE, pt)), PieceToChar[pt]); b += std::string(popcount(pos.pieces(BLACK, pt)), PieceToChar[pt]); } - fname = (e.key == pos.material_key() ? w + 'v' + b : b + 'v' + w) - + (Type == WDL ? ".rtbw" : ".rtbz"); + fname = + (e.key == pos.material_key() ? w + 'v' + b : b + 'v' + w) + (Type == WDL ? ".rtbw" : ".rtbz"); uint8_t* data = TBFile(fname).map(&e.baseAddress, &e.mapping, Type); @@ -1167,7 +1230,7 @@ void* mapped(TBTable& e, const Position& pos) { template::Ret> Ret probe_table(const Position& pos, ProbeState* result, WDLScore wdl = WDLDraw) { - if (pos.count() == 2) // KvK + if (pos.count() == 2) // KvK return Ret(WDLDraw); TBTable* entry = TBTables.get(pos.material_key()); @@ -1179,7 +1242,7 @@ Ret probe_table(const Position& pos, ProbeState* result, WDLScore wdl = WDLDraw) } // For a position where the side to move has a winning capture it is not necessary -// to store a winning value so the generator treats such positions as "don't cares" +// to store a winning value so the generator treats such positions as "don't care" // and tries to assign to it a value that improves the compression ratio. Similarly, // if the side to move has a drawing capture, then the position is at least drawn. // If the position is won, then the TB needs to store a win value. But if the @@ -1188,22 +1251,21 @@ Ret probe_table(const Position& pos, ProbeState* result, WDLScore wdl = WDLDraw) // their results and must probe the position itself. The "best" result of these // probes is the correct result for the position. // 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 +// (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 WDLScore search(Position& pos, ProbeState* result) { - WDLScore value, bestValue = WDLLoss; + WDLScore value, bestValue = WDLLoss; StateInfo st; - auto moveList = MoveList(pos); + auto moveList = MoveList(pos); size_t totalCount = moveList.size(), moveCount = 0; for (const Move move : moveList) { - if ( !pos.capture(move) - && (!CheckZeroingMoves || type_of(pos.moved_piece(move)) != PAWN)) + if (!pos.capture(move) && (!CheckZeroingMoves || type_of(pos.moved_piece(move)) != PAWN)) continue; moveCount++; @@ -1221,7 +1283,7 @@ WDLScore search(Position& pos, ProbeState* result) { if (value >= WDLWin) { - *result = ZEROING_BEST_MOVE; // Winning DTZ-zeroing move + *result = ZEROING_BEST_MOVE; // Winning DTZ-zeroing move return value; } } @@ -1247,23 +1309,22 @@ WDLScore search(Position& pos, ProbeState* result) { // DTZ stores a "don't care" value if bestValue is a win if (bestValue >= value) - return *result = ( bestValue > WDLDraw - || noMoreMoves ? ZEROING_BEST_MOVE : OK), bestValue; + return *result = (bestValue > WDLDraw || noMoreMoves ? ZEROING_BEST_MOVE : OK), bestValue; return *result = OK, value; } -} // namespace +} // 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. +// 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) { TBTables.clear(); MaxCardinality = 0; - TBFile::Paths = paths; + TBFile::Paths = paths; if (paths.empty() || paths == "") return; @@ -1288,21 +1349,21 @@ void Tablebases::init(const std::string& paths) { for (auto s : diagonal) MapA1D1D4[s] = code++; - // MapKK[] encodes all the 461 possible legal positions of two kings where + // MapKK[] encodes all the 462 possible legal positions of two kings where // the first is in the a1-d1-d4 triangle. If the first king is on the a1-d4 - // diagonal, the other one shall not to be above the a1-h8 diagonal. + // diagonal, the other one shall not be above the a1-h8 diagonal. std::vector> bothOnDiagonal; code = 0; for (int idx = 0; idx < 10; idx++) for (Square s1 = SQ_A1; s1 <= SQ_D4; ++s1) - if (MapA1D1D4[s1] == idx && (idx || s1 == SQ_B1)) // SQ_B1 is mapped to 0 + if (MapA1D1D4[s1] == idx && (idx || s1 == SQ_B1)) // SQ_B1 is mapped to 0 { for (Square s2 = SQ_A1; s2 <= SQ_H8; ++s2) if ((PseudoAttacks[KING][s1] | s1) & s2) - continue; // Illegal position + continue; // Illegal position else if (!off_A1H8(s1) && off_A1H8(s2) > 0) - continue; // First on diagonal, second above + continue; // First on diagonal, second above else if (!off_A1H8(s1) && !off_A1H8(s2)) bothOnDiagonal.emplace_back(idx, s2); @@ -1311,31 +1372,31 @@ void Tablebases::init(const std::string& paths) { MapKK[idx][s2] = code++; } - // Legal positions with both kings on diagonal are encoded as last ones + // Legal positions with both kings on a diagonal are encoded as last ones for (auto p : bothOnDiagonal) MapKK[p.first][p.second] = code++; - // Binomial[] stores the Binomial Coefficents using Pascal rule. There + // Binomial[] stores the Binomial Coefficients using Pascal rule. There // are Binomial[k][n] ways to choose k elements from a set of n elements. Binomial[0][0] = 1; - for (int n = 1; n < 64; n++) // Squares - 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); + for (int n = 1; n < 64; n++) // Squares + for (int k = 0; k < 6 && k <= n; ++k) // Pieces + Binomial[k][n] = + (k > 0 ? Binomial[k - 1][n - 1] : 0) + (k < n ? Binomial[k][n - 1] : 0); // MapPawns[s] encodes squares a2-h7 to 0..47. This is the number of possible // available squares when the leading one is in 's'. Moreover the pawn with - // highest MapPawns[] is the leading pawn, the one nearest the edge and, - // among pawns with same file, the one with lowest rank. - int availableSquares = 47; // Available squares when lead pawn is in a2 + // highest MapPawns[] is the leading pawn, the one nearest the edge, and + // among pawns with the same file, the one with the lowest rank. + int availableSquares = 47; // Available squares when lead pawn is in a2 // 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 + // Restart the index at every file because TB table is split // by file, so we can reuse the same index for different files. int idx = 0; @@ -1352,7 +1413,7 @@ void Tablebases::init(const std::string& paths) { // due to mirroring: sq == a3 -> no a2, h2, so MapPawns[a3] = 45 if (leadPawnsCnt == 1) { - MapPawns[sq] = availableSquares--; + MapPawns[sq] = availableSquares--; MapPawns[flip_file(sq)] = availableSquares--; } LeadPawnIdx[leadPawnsCnt][sq] = idx; @@ -1363,20 +1424,24 @@ void Tablebases::init(const std::string& paths) { } // Add entries in TB tables if the corresponding ".rtbw" file exists - for (PieceType p1 = PAWN; p1 < KING; ++p1) { + for (PieceType p1 = PAWN; p1 < KING; ++p1) + { TBTables.add({KING, p1, KING}); - for (PieceType p2 = PAWN; p2 <= p1; ++p2) { + for (PieceType p2 = PAWN; p2 <= p1; ++p2) + { TBTables.add({KING, p1, p2, KING}); TBTables.add({KING, p1, KING, p2}); for (PieceType p3 = PAWN; p3 < KING; ++p3) TBTables.add({KING, p1, p2, KING, p3}); - for (PieceType p3 = PAWN; p3 <= p2; ++p3) { + for (PieceType p3 = PAWN; p3 <= p2; ++p3) + { TBTables.add({KING, p1, p2, p3, KING}); - for (PieceType p4 = PAWN; p4 <= p3; ++p4) { + for (PieceType p4 = PAWN; p4 <= p3; ++p4) + { TBTables.add({KING, p1, p2, p3, p4, KING}); for (PieceType p5 = PAWN; p5 <= p4; ++p5) @@ -1386,7 +1451,8 @@ void Tablebases::init(const std::string& paths) { TBTables.add({KING, p1, p2, p3, p4, KING, p5}); } - for (PieceType p4 = PAWN; p4 < KING; ++p4) { + for (PieceType p4 = PAWN; p4 < KING; ++p4) + { TBTables.add({KING, p1, p2, p3, KING, p4}); for (PieceType p5 = PAWN; p5 <= p4; ++p5) @@ -1439,19 +1505,19 @@ 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. int Tablebases::probe_dtz(Position& pos, ProbeState* result) { - *result = OK; + *result = OK; WDLScore wdl = search(pos, result); - if (*result == FAIL || wdl == WDLDraw) // DTZ tables don't store draws + if (*result == FAIL || wdl == WDLDraw) // DTZ tables don't store draws return 0; - // DTZ stores a 'don't care' value in this case, or even a plain wrong + // DTZ stores a 'don't care value in this case, or even a plain wrong // one as in case the best move is a losing ep, so it cannot be probed. if (*result == ZEROING_BEST_MOVE) return dtz_before_zeroing(wdl); @@ -1467,7 +1533,7 @@ int Tablebases::probe_dtz(Position& pos, ProbeState* result) { // DTZ stores results for the other side, so we need to do a 1-ply search and // find the winning move that minimizes DTZ. StateInfo st; - int minDTZ = 0xFFFF; + int minDTZ = 0xFFFF; for (const Move move : MoveList(pos)) { @@ -1478,9 +1544,8 @@ int Tablebases::probe_dtz(Position& pos, ProbeState* result) { // For zeroing moves we want the dtz of the move _before_ doing it, // 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)) - : -probe_dtz(pos, result); + // winning position we could make a losing capture or go for a draw). + dtz = zeroing ? -dtz_before_zeroing(search(pos, result)) : -probe_dtz(pos, result); // If the move mates, force minDTZ to 1 if (dtz == 1 && pos.checkers() && MoveList(pos).size() == 0) @@ -1511,8 +1576,8 @@ int Tablebases::probe_dtz(Position& pos, ProbeState* result) { // A return value false indicates that not all probes were successful. bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves) { - ProbeState result; - StateInfo st; + ProbeState result = OK; + StateInfo st; // Obtain 50-move counter for the root position int cnt50 = pos.rule50_count(); @@ -1520,7 +1585,7 @@ bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves) { // Check whether a position was repeated since the last zeroing move. bool rep = pos.has_repeated(); - int dtz, bound = Options["Syzygy50MoveRule"] ? 900 : 1; + int dtz, bound = Options["Syzygy50MoveRule"] ? (MAX_DTZ - 100) : 1; // Probe and rank each move for (auto& m : rootMoves) @@ -1532,20 +1597,24 @@ bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves) { { // 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); + dtz = dtz_before_zeroing(wdl); + } + else if (pos.is_draw(1)) + { + // In case a root move leads to a draw by repetition or 50-move rule, + // we set dtz to zero. Note: since we are only 1 ply from the root, + // this must be a true 3-fold repetition inside the game history. + dtz = 0; } 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; + dtz = dtz > 0 ? dtz + 1 : dtz < 0 ? dtz - 1 : dtz; } // Make sure that a mating move is assigned a dtz value of 1 - if ( pos.checkers() - && dtz == 2 - && MoveList(pos).size() == 0) + if (pos.checkers() && dtz == 2 && MoveList(pos).size() == 0) dtz = 1; pos.undo_move(m.pv[0]); @@ -1555,19 +1624,19 @@ bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves) { // 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; + int r = dtz > 0 ? (dtz + cnt50 <= 99 && !rep ? MAX_DTZ : MAX_DTZ - (dtz + cnt50)) + : dtz < 0 ? (-dtz * 2 + cnt50 < 100 ? -MAX_DTZ : -MAX_DTZ + (-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; + m.tbScore = r >= bound ? VALUE_MATE - MAX_PLY - 1 + : r > 0 ? Value((std::max(3, r - (MAX_DTZ - 200)) * int(PawnValue)) / 200) + : r == 0 ? VALUE_DRAW + : r > -bound ? Value((std::min(-3, r + (MAX_DTZ - 200)) * int(PawnValue)) / 200) + : -VALUE_MATE + MAX_PLY + 1; } return true; @@ -1580,10 +1649,11 @@ bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves) { // A return value false indicates that not all probes were successful. bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves) { - static const int WDL_to_rank[] = { -1000, -899, 0, 899, 1000 }; + static const int WDL_to_rank[] = {-MAX_DTZ, -MAX_DTZ + 101, 0, MAX_DTZ - 101, MAX_DTZ}; - ProbeState result; - StateInfo st; + ProbeState result = OK; + StateInfo st; + WDLScore wdl; bool rule50 = Options["Syzygy50MoveRule"]; @@ -1592,7 +1662,10 @@ bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves) { { pos.do_move(m.pv[0], st); - WDLScore wdl = -probe_wdl(pos, &result); + if (pos.is_draw(1)) + wdl = WDLDraw; + else + wdl = -probe_wdl(pos, &result); pos.undo_move(m.pv[0]); @@ -1602,10 +1675,11 @@ bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves) { m.tbRank = WDL_to_rank[wdl + 2]; if (!rule50) - wdl = wdl > WDLDraw ? WDLWin - : wdl < WDLDraw ? WDLLoss : WDLDraw; + wdl = wdl > WDLDraw ? WDLWin : wdl < WDLDraw ? WDLLoss : WDLDraw; m.tbScore = WDL_to_value[wdl + 2]; } return true; } + +} // namespace Stockfish