2 Stockfish, a UCI chess playing engine derived from Glaurung 2.1
3 Copyright (c) 2013 Ronald de Man
4 Copyright (C) 2016-2018 Marco Costalba, Lucas Braesch
6 Stockfish is free software: you can redistribute it and/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation, either version 3 of the License, or
9 (at your option) any later version.
11 Stockfish is distributed in the hope that it will be useful,
12 but WITHOUT ANY WARRANTY; without even the implied warranty of
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License
17 along with this program. If not, see <http://www.gnu.org/licenses/>.
23 #include <cstring> // For std::memset and std::memcpy
29 #include <type_traits>
31 #include "../bitboard.h"
32 #include "../movegen.h"
33 #include "../position.h"
34 #include "../search.h"
35 #include "../thread_win32.h"
47 #define WIN32_LEAN_AND_MEAN
52 using namespace Tablebases;
54 int Tablebases::MaxCardinality;
58 constexpr int TBPIECES = 7; // Max number of supported pieces
60 enum { BigEndian, LittleEndian };
61 enum TBType { KEY, WDL, DTZ }; // Used as template parameter
63 // Each table has a set of flags: all of them refer to DTZ tables, the last one to WDL tables
64 enum TBFlag { STM = 1, Mapped = 2, WinPlies = 4, LossPlies = 8, Wide = 16, SingleValue = 128 };
66 inline WDLScore operator-(WDLScore d) { return WDLScore(-int(d)); }
67 inline Square operator^=(Square& s, int i) { return s = Square(int(s) ^ i); }
68 inline Square operator^(Square s, int i) { return Square(int(s) ^ i); }
70 const std::string PieceToChar = " PNBRQK pnbrqk";
72 int MapPawns[SQUARE_NB];
73 int MapB1H1H7[SQUARE_NB];
74 int MapA1D1D4[SQUARE_NB];
75 int MapKK[10][SQUARE_NB]; // [MapA1D1D4][SQUARE_NB]
77 int Binomial[6][SQUARE_NB]; // [k][n] k elements from a set of n elements
78 int LeadPawnIdx[6][SQUARE_NB]; // [leadPawnsCnt][SQUARE_NB]
79 int LeadPawnsSize[6][4]; // [leadPawnsCnt][FILE_A..FILE_D]
81 // Comparison function to sort leading pawns in ascending MapPawns[] order
82 bool pawns_comp(Square i, Square j) { return MapPawns[i] < MapPawns[j]; }
83 int off_A1H8(Square sq) { return int(rank_of(sq)) - file_of(sq); }
85 constexpr Value WDL_to_value[] = {
86 -VALUE_MATE + MAX_PLY + 1,
90 VALUE_MATE - MAX_PLY - 1
93 template<typename T, int Half = sizeof(T) / 2, int End = sizeof(T) - 1>
94 inline void swap_endian(T& x)
96 static_assert(std::is_unsigned<T>::value, "Argument of swap_endian not unsigned");
98 uint8_t tmp, *c = (uint8_t*)&x;
99 for (int i = 0; i < Half; ++i)
100 tmp = c[i], c[i] = c[End - i], c[End - i] = tmp;
102 template<> inline void swap_endian<uint8_t>(uint8_t&) {}
104 template<typename T, int LE> T number(void* addr)
106 static const union { uint32_t i; char c[4]; } Le = { 0x01020304 };
107 static const bool IsLittleEndian = (Le.c[0] == 4);
111 if ((uintptr_t)addr & (alignof(T) - 1)) // Unaligned pointer (very rare)
112 std::memcpy(&v, addr, sizeof(T));
116 if (LE != IsLittleEndian)
121 // DTZ tables don't store valid scores for moves that reset the rule50 counter
122 // like captures and pawn moves but we can easily recover the correct dtz of the
123 // previous move if we know the position's WDL score.
124 int dtz_before_zeroing(WDLScore wdl) {
125 return wdl == WDLWin ? 1 :
126 wdl == WDLCursedWin ? 101 :
127 wdl == WDLBlessedLoss ? -101 :
128 wdl == WDLLoss ? -1 : 0;
131 // Return the sign of a number (-1, 0, 1)
132 template <typename T> int sign_of(T val) {
133 return (T(0) < val) - (val < T(0));
136 // Numbers in little endian used by sparseIndex[] to point into blockLength[]
138 char block[4]; // Number of block
139 char offset[2]; // Offset within the block
142 static_assert(sizeof(SparseEntry) == 6, "SparseEntry must be 6 bytes");
144 typedef uint16_t Sym; // Huffman symbol
147 enum Side { Left, Right };
149 uint8_t lr[3]; // The first 12 bits is the left-hand symbol, the second 12
150 // bits is the right-hand symbol. If symbol has length 1,
151 // then the left-hand symbol is the stored value.
154 return S == Left ? ((lr[1] & 0xF) << 8) | lr[0] :
155 S == Right ? (lr[2] << 4) | (lr[1] >> 4) : (assert(false), Sym(-1));
159 static_assert(sizeof(LR) == 3, "LR tree entry must be 3 bytes");
161 // Tablebases data layout is structured as following:
163 // TBFile: memory maps/unmaps the physical .rtbw and .rtbz files
164 // TBTable: one object for each file with corresponding indexing information
165 // TBTables: has ownership of TBTable objects, keeping a list and a hash
167 // class TBFile memory maps/unmaps the single .rtbw and .rtbz files. Files are
168 // memory mapped for best performance. Files are mapped at first access: at init
169 // time only existence of the file is checked.
170 class TBFile : public std::ifstream {
175 // Look for and open the file among the Paths directories where the .rtbw
176 // and .rtbz files can be found. Multiple directories are separated by ";"
177 // on Windows and by ":" on Unix-based operating systems.
180 // C:\tb\wdl345;C:\tb\wdl6;D:\tb\dtz345;D:\tb\dtz6
181 static std::string Paths;
183 TBFile(const std::string& f) {
186 constexpr char SepChar = ':';
188 constexpr char SepChar = ';';
190 std::stringstream ss(Paths);
193 while (std::getline(ss, path, SepChar)) {
194 fname = path + "/" + f;
195 std::ifstream::open(fname);
201 // Memory map the file and check it. File should be already open and will be
202 // closed after mapping.
203 uint8_t* map(void** baseAddress, uint64_t* mapping, TBType type) {
207 close(); // Need to re-open to get native file descriptor
211 int fd = ::open(fname.c_str(), O_RDONLY);
214 return *baseAddress = nullptr, nullptr;
217 *mapping = statbuf.st_size;
218 *baseAddress = mmap(nullptr, statbuf.st_size, PROT_READ, MAP_SHARED, fd, 0);
221 if (*baseAddress == MAP_FAILED) {
222 std::cerr << "Could not mmap() " << fname << std::endl;
226 HANDLE fd = CreateFile(fname.c_str(), GENERIC_READ, FILE_SHARE_READ, nullptr,
227 OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, nullptr);
229 if (fd == INVALID_HANDLE_VALUE)
230 return *baseAddress = nullptr, nullptr;
233 DWORD size_low = GetFileSize(fd, &size_high);
234 HANDLE mmap = CreateFileMapping(fd, nullptr, PAGE_READONLY, size_high, size_low, nullptr);
238 std::cerr << "CreateFileMapping() failed" << std::endl;
242 *mapping = (uint64_t)mmap;
243 *baseAddress = MapViewOfFile(mmap, FILE_MAP_READ, 0, 0, 0);
246 std::cerr << "MapViewOfFile() failed, name = " << fname
247 << ", error = " << GetLastError() << std::endl;
251 uint8_t* data = (uint8_t*)*baseAddress;
253 constexpr uint8_t Magics[][4] = { { 0xD7, 0x66, 0x0C, 0xA5 },
254 { 0x71, 0xE8, 0x23, 0x5D } };
256 if (memcmp(data, Magics[type == WDL], 4)) {
257 std::cerr << "Corrupted table in file " << fname << std::endl;
258 unmap(*baseAddress, *mapping);
259 return *baseAddress = nullptr, nullptr;
262 return data + 4; // Skip Magics's header
265 static void unmap(void* baseAddress, uint64_t mapping) {
268 munmap(baseAddress, mapping);
270 UnmapViewOfFile(baseAddress);
271 CloseHandle((HANDLE)mapping);
276 std::string TBFile::Paths;
278 // struct PairsData contains low level indexing information to access TB data.
279 // There are 8, 4 or 2 PairsData records for each TBTable, according to type of
280 // table and if positions have pawns or not. It is populated at first access.
282 uint8_t flags; // Table flags, see enum TBFlag
283 uint8_t maxSymLen; // Maximum length in bits of the Huffman symbols
284 uint8_t minSymLen; // Minimum length in bits of the Huffman symbols
285 uint32_t blocksNum; // Number of blocks in the TB file
286 size_t sizeofBlock; // Block size in bytes
287 size_t span; // About every span values there is a SparseIndex[] entry
288 Sym* lowestSym; // lowestSym[l] is the symbol of length l with the lowest value
289 LR* btree; // btree[sym] stores the left and right symbols that expand sym
290 uint16_t* blockLength; // Number of stored positions (minus one) for each block: 1..65536
291 uint32_t blockLengthSize; // Size of blockLength[] table: padded so it's bigger than blocksNum
292 SparseEntry* sparseIndex; // Partial indices into blockLength[]
293 size_t sparseIndexSize; // Size of SparseIndex[] table
294 uint8_t* data; // Start of Huffman compressed data
295 std::vector<uint64_t> base64; // base64[l - min_sym_len] is the 64bit-padded lowest symbol of length l
296 std::vector<uint8_t> symlen; // Number of values (-1) represented by a given Huffman symbol: 1..256
297 Piece pieces[TBPIECES]; // Position pieces: the order of pieces defines the groups
298 uint64_t groupIdx[TBPIECES+1]; // Start index used for the encoding of the group's pieces
299 int groupLen[TBPIECES+1]; // Number of pieces in a given group: KRKN -> (3, 1)
300 uint16_t map_idx[4]; // WDLWin, WDLLoss, WDLCursedWin, WDLBlessedLoss (used in DTZ)
303 // struct TBTable contains indexing information to access the corresponding TBFile.
304 // There are 2 types of TBTable, corresponding to a WDL or a DTZ file. TBTable
305 // is populated at init time but the nested PairsData records are populated at
306 // first access, when the corresponding file is memory mapped.
307 template<TBType Type>
309 typedef typename std::conditional<Type == WDL, WDLScore, int>::type Ret;
311 static constexpr int Sides = Type == WDL ? 2 : 1;
313 std::atomic_bool ready;
321 bool hasUniquePieces;
322 uint8_t pawnCount[2]; // [Lead color / other color]
323 PairsData items[Sides][4]; // [wtm / btm][FILE_A..FILE_D or 0]
325 PairsData* get(int stm, int f) {
326 return &items[stm % Sides][hasPawns ? f : 0];
329 TBTable() : ready(false), baseAddress(nullptr) {}
330 explicit TBTable(const std::string& code);
331 explicit TBTable(const TBTable<WDL>& wdl);
335 TBFile::unmap(baseAddress, mapping);
340 TBTable<WDL>::TBTable(const std::string& code) : TBTable() {
345 key = pos.set(code, WHITE, &st).material_key();
346 pieceCount = pos.count<ALL_PIECES>();
347 hasPawns = pos.pieces(PAWN);
349 hasUniquePieces = false;
350 for (Color c = WHITE; c <= BLACK; ++c)
351 for (PieceType pt = PAWN; pt < KING; ++pt)
352 if (popcount(pos.pieces(c, pt)) == 1)
353 hasUniquePieces = true;
355 // Set the leading color. In case both sides have pawns the leading color
356 // is the side with less pawns because this leads to better compression.
357 bool c = !pos.count<PAWN>(BLACK)
358 || ( pos.count<PAWN>(WHITE)
359 && pos.count<PAWN>(BLACK) >= pos.count<PAWN>(WHITE));
361 pawnCount[0] = pos.count<PAWN>(c ? WHITE : BLACK);
362 pawnCount[1] = pos.count<PAWN>(c ? BLACK : WHITE);
364 key2 = pos.set(code, BLACK, &st).material_key();
368 TBTable<DTZ>::TBTable(const TBTable<WDL>& wdl) : TBTable() {
370 // Use the corresponding WDL table to avoid recalculating all from scratch
373 pieceCount = wdl.pieceCount;
374 hasPawns = wdl.hasPawns;
375 hasUniquePieces = wdl.hasUniquePieces;
376 pawnCount[0] = wdl.pawnCount[0];
377 pawnCount[1] = wdl.pawnCount[1];
380 // class TBTables creates and keeps ownership of the TBTable objects, one for
381 // each TB file found. It supports a fast, hash based, table lookup. Populated
382 // at init time, accessed at probe time.
385 typedef std::tuple<Key, TBTable<WDL>*, TBTable<DTZ>*> Entry;
387 static constexpr int Size = 1 << 12; // 4K table, indexed by key's 12 lsb
388 static constexpr int Overflow = 1; // Number of elements allowed to map to the last bucket
390 Entry hashTable[Size + Overflow];
392 std::deque<TBTable<WDL>> wdlTable;
393 std::deque<TBTable<DTZ>> dtzTable;
395 void insert(Key key, TBTable<WDL>* wdl, TBTable<DTZ>* dtz) {
396 uint32_t homeBucket = (uint32_t)key & (Size - 1);
397 Entry entry = std::make_tuple(key, wdl, dtz);
399 // Ensure last element is empty to avoid overflow when looking up
400 for (uint32_t bucket = homeBucket; bucket < Size + Overflow - 1; ++bucket) {
401 Key otherKey = std::get<KEY>(hashTable[bucket]);
402 if (otherKey == key || !std::get<WDL>(hashTable[bucket])) {
403 hashTable[bucket] = entry;
407 // Robin Hood hashing: If we've probed for longer than this element,
408 // insert here and search for a new spot for the other element instead.
409 uint32_t otherHomeBucket = (uint32_t)otherKey & (Size - 1);
410 if (otherHomeBucket > homeBucket) {
411 swap(entry, hashTable[bucket]);
413 homeBucket = otherHomeBucket;
416 std::cerr << "TB hash table size too low!" << std::endl;
421 template<TBType Type>
422 TBTable<Type>* get(Key key) {
423 for (const Entry* entry = &hashTable[(uint32_t)key & (Size - 1)]; ; ++entry) {
424 if (std::get<KEY>(*entry) == key || !std::get<Type>(*entry))
425 return std::get<Type>(*entry);
430 memset(hashTable, 0, sizeof(hashTable));
434 size_t size() const { return wdlTable.size(); }
435 void add(const std::vector<PieceType>& pieces);
440 // If the corresponding file exists two new objects TBTable<WDL> and TBTable<DTZ>
441 // are created and added to the lists and hash table. Called at init time.
442 void TBTables::add(const std::vector<PieceType>& pieces) {
446 for (PieceType pt : pieces)
447 code += PieceToChar[pt];
449 TBFile file(code.insert(code.find('K', 1), "v") + ".rtbw"); // KRK -> KRvK
451 if (!file.is_open()) // Only WDL file is checked
456 MaxCardinality = std::max((int)pieces.size(), MaxCardinality);
458 wdlTable.emplace_back(code);
459 dtzTable.emplace_back(wdlTable.back());
461 // Insert into the hash keys for both colors: KRvK with KR white and black
462 insert(wdlTable.back().key , &wdlTable.back(), &dtzTable.back());
463 insert(wdlTable.back().key2, &wdlTable.back(), &dtzTable.back());
466 // TB tables are compressed with canonical Huffman code. The compressed data is divided into
467 // blocks of size d->sizeofBlock, and each block stores a variable number of symbols.
468 // Each symbol represents either a WDL or a (remapped) DTZ value, or a pair of other symbols
469 // (recursively). If you keep expanding the symbols in a block, you end up with up to 65536
470 // WDL or DTZ values. Each symbol represents up to 256 values and will correspond after
471 // Huffman coding to at least 1 bit. So a block of 32 bytes corresponds to at most
472 // 32 x 8 x 256 = 65536 values. This maximum is only reached for tables that consist mostly
473 // of draws or mostly of wins, but such tables are actually quite common. In principle, the
474 // blocks in WDL tables are 64 bytes long (and will be aligned on cache lines). But for
475 // mostly-draw or mostly-win tables this can leave many 64-byte blocks only half-filled, so
476 // in such cases blocks are 32 bytes long. The blocks of DTZ tables are up to 1024 bytes long.
477 // The generator picks the size that leads to the smallest table. The "book" of symbols and
478 // Huffman codes is the same for all blocks in the table. A non-symmetric pawnless TB file
479 // will have one table for wtm and one for btm, a TB file with pawns will have tables per
480 // file a,b,c,d also in this case one set for wtm and one for btm.
481 int decompress_pairs(PairsData* d, uint64_t idx) {
483 // Special case where all table positions store the same value
484 if (d->flags & TBFlag::SingleValue)
487 // First we need to locate the right block that stores the value at index "idx".
488 // Because each block n stores blockLength[n] + 1 values, the index i of the block
489 // that contains the value at position idx is:
491 // for (i = -1, sum = 0; sum <= idx; i++)
492 // sum += blockLength[i + 1] + 1;
494 // This can be slow, so we use SparseIndex[] populated with a set of SparseEntry that
495 // point to known indices into blockLength[]. Namely SparseIndex[k] is a SparseEntry
496 // that stores the blockLength[] index and the offset within that block of the value
497 // with index I(k), where:
499 // I(k) = k * d->span + d->span / 2 (1)
501 // First step is to get the 'k' of the I(k) nearest to our idx, using definition (1)
502 uint32_t k = idx / d->span;
504 // Then we read the corresponding SparseIndex[] entry
505 uint32_t block = number<uint32_t, LittleEndian>(&d->sparseIndex[k].block);
506 int offset = number<uint16_t, LittleEndian>(&d->sparseIndex[k].offset);
508 // Now compute the difference idx - I(k). From definition of k we know that
510 // idx = k * d->span + idx % d->span (2)
512 // So from (1) and (2) we can compute idx - I(K):
513 int diff = idx % d->span - d->span / 2;
515 // Sum the above to offset to find the offset corresponding to our idx
518 // Move to previous/next block, until we reach the correct block that contains idx,
519 // that is when 0 <= offset <= d->blockLength[block]
521 offset += d->blockLength[--block] + 1;
523 while (offset > d->blockLength[block])
524 offset -= d->blockLength[block++] + 1;
526 // Finally, we find the start address of our block of canonical Huffman symbols
527 uint32_t* ptr = (uint32_t*)(d->data + ((uint64_t)block * d->sizeofBlock));
529 // Read the first 64 bits in our block, this is a (truncated) sequence of
530 // unknown number of symbols of unknown length but we know the first one
531 // is at the beginning of this 64 bits sequence.
532 uint64_t buf64 = number<uint64_t, BigEndian>(ptr); ptr += 2;
537 int len = 0; // This is the symbol length - d->min_sym_len
539 // Now get the symbol length. For any symbol s64 of length l right-padded
540 // to 64 bits we know that d->base64[l-1] >= s64 >= d->base64[l] so we
541 // can find the symbol length iterating through base64[].
542 while (buf64 < d->base64[len])
545 // All the symbols of a given length are consecutive integers (numerical
546 // sequence property), so we can compute the offset of our symbol of
547 // length len, stored at the beginning of buf64.
548 sym = (buf64 - d->base64[len]) >> (64 - len - d->minSymLen);
550 // Now add the value of the lowest symbol of length len to get our symbol
551 sym += number<Sym, LittleEndian>(&d->lowestSym[len]);
553 // If our offset is within the number of values represented by symbol sym
555 if (offset < d->symlen[sym] + 1)
558 // ...otherwise update the offset and continue to iterate
559 offset -= d->symlen[sym] + 1;
560 len += d->minSymLen; // Get the real length
561 buf64 <<= len; // Consume the just processed symbol
564 if (buf64Size <= 32) { // Refill the buffer
566 buf64 |= (uint64_t)number<uint32_t, BigEndian>(ptr++) << (64 - buf64Size);
570 // Ok, now we have our symbol that expands into d->symlen[sym] + 1 symbols.
571 // We binary-search for our value recursively expanding into the left and
572 // right child symbols until we reach a leaf node where symlen[sym] + 1 == 1
573 // that will store the value we need.
574 while (d->symlen[sym]) {
576 Sym left = d->btree[sym].get<LR::Left>();
578 // If a symbol contains 36 sub-symbols (d->symlen[sym] + 1 = 36) and
579 // expands in a pair (d->symlen[left] = 23, d->symlen[right] = 11), then
580 // we know that, for instance the ten-th value (offset = 10) will be on
581 // the left side because in Recursive Pairing child symbols are adjacent.
582 if (offset < d->symlen[left] + 1)
585 offset -= d->symlen[left] + 1;
586 sym = d->btree[sym].get<LR::Right>();
590 return d->btree[sym].get<LR::Left>();
593 bool check_dtz_stm(TBTable<WDL>*, int, File) { return true; }
595 bool check_dtz_stm(TBTable<DTZ>* entry, int stm, File f) {
597 auto flags = entry->get(stm, f)->flags;
598 return (flags & TBFlag::STM) == stm
599 || ((entry->key == entry->key2) && !entry->hasPawns);
602 // DTZ scores are sorted by frequency of occurrence and then assigned the
603 // values 0, 1, 2, ... in order of decreasing frequency. This is done for each
604 // of the four WDLScore values. The mapping information necessary to reconstruct
605 // the original values is stored in the TB file and read during map[] init.
606 WDLScore map_score(TBTable<WDL>*, File, int value, WDLScore) { return WDLScore(value - 2); }
608 int map_score(TBTable<DTZ>* entry, File f, int value, WDLScore wdl) {
610 constexpr int WDLMap[] = { 1, 3, 0, 2, 0 };
612 auto flags = entry->get(0, f)->flags;
614 uint8_t* map = entry->map;
615 uint16_t* idx = entry->get(0, f)->map_idx;
616 if (flags & TBFlag::Mapped) {
617 if (flags & TBFlag::Wide)
618 value = ((uint16_t *)map)[idx[WDLMap[wdl + 2]] + value];
620 value = map[idx[WDLMap[wdl + 2]] + value];
623 // DTZ tables store distance to zero in number of moves or plies. We
624 // want to return plies, so we have convert to plies when needed.
625 if ( (wdl == WDLWin && !(flags & TBFlag::WinPlies))
626 || (wdl == WDLLoss && !(flags & TBFlag::LossPlies))
627 || wdl == WDLCursedWin
628 || wdl == WDLBlessedLoss)
634 // Compute a unique index out of a position and use it to probe the TB file. To
635 // encode k pieces of same type and color, first sort the pieces by square in
636 // ascending order s1 <= s2 <= ... <= sk then compute the unique index as:
638 // idx = Binomial[1][s1] + Binomial[2][s2] + ... + Binomial[k][sk]
640 template<typename T, typename Ret = typename T::Ret>
641 Ret do_probe_table(const Position& pos, T* entry, WDLScore wdl, ProbeState* result) {
643 Square squares[TBPIECES];
644 Piece pieces[TBPIECES];
646 int next = 0, size = 0, leadPawnsCnt = 0;
648 Bitboard b, leadPawns = 0;
649 File tbFile = FILE_A;
651 // A given TB entry like KRK has associated two material keys: KRvk and Kvkr.
652 // If both sides have the same pieces keys are equal. In this case TB tables
653 // only store the 'white to move' case, so if the position to lookup has black
654 // to move, we need to switch the color and flip the squares before to lookup.
655 bool symmetricBlackToMove = (entry->key == entry->key2 && pos.side_to_move());
657 // TB files are calculated for white as stronger side. For instance we have
658 // KRvK, not KvKR. A position where stronger side is white will have its
659 // material key == entry->key, otherwise we have to switch the color and
660 // flip the squares before to lookup.
661 bool blackStronger = (pos.material_key() != entry->key);
663 int flipColor = (symmetricBlackToMove || blackStronger) * 8;
664 int flipSquares = (symmetricBlackToMove || blackStronger) * 070;
665 int stm = (symmetricBlackToMove || blackStronger) ^ pos.side_to_move();
667 // For pawns, TB files store 4 separate tables according if leading pawn is on
668 // file a, b, c or d after reordering. The leading pawn is the one with maximum
669 // MapPawns[] value, that is the one most toward the edges and with lowest rank.
670 if (entry->hasPawns) {
672 // In all the 4 tables, pawns are at the beginning of the piece sequence and
673 // their color is the reference one. So we just pick the first one.
674 Piece pc = Piece(entry->get(0, 0)->pieces[0] ^ flipColor);
676 assert(type_of(pc) == PAWN);
678 leadPawns = b = pos.pieces(color_of(pc), PAWN);
680 squares[size++] = pop_lsb(&b) ^ flipSquares;
685 std::swap(squares[0], *std::max_element(squares, squares + leadPawnsCnt, pawns_comp));
687 tbFile = file_of(squares[0]);
689 tbFile = file_of(squares[0] ^ 7); // Horizontal flip: SQ_H1 -> SQ_A1
692 // DTZ tables are one-sided, i.e. they store positions only for white to
693 // move or only for black to move, so check for side to move to be stm,
694 // early exit otherwise.
695 if (!check_dtz_stm(entry, stm, tbFile))
696 return *result = CHANGE_STM, Ret();
698 // Now we are ready to get all the position pieces (but the lead pawns) and
699 // directly map them to the correct color and square.
700 b = pos.pieces() ^ leadPawns;
702 Square s = pop_lsb(&b);
703 squares[size] = s ^ flipSquares;
704 pieces[size++] = Piece(pos.piece_on(s) ^ flipColor);
709 d = entry->get(stm, tbFile);
711 // Then we reorder the pieces to have the same sequence as the one stored
712 // in pieces[i]: the sequence that ensures the best compression.
713 for (int i = leadPawnsCnt; i < size; ++i)
714 for (int j = i; j < size; ++j)
715 if (d->pieces[i] == pieces[j])
717 std::swap(pieces[i], pieces[j]);
718 std::swap(squares[i], squares[j]);
722 // Now we map again the squares so that the square of the lead piece is in
723 // the triangle A1-D1-D4.
724 if (file_of(squares[0]) > FILE_D)
725 for (int i = 0; i < size; ++i)
726 squares[i] ^= 7; // Horizontal flip: SQ_H1 -> SQ_A1
728 // Encode leading pawns starting with the one with minimum MapPawns[] and
729 // proceeding in ascending order.
730 if (entry->hasPawns) {
731 idx = LeadPawnIdx[leadPawnsCnt][squares[0]];
733 std::sort(squares + 1, squares + leadPawnsCnt, pawns_comp);
735 for (int i = 1; i < leadPawnsCnt; ++i)
736 idx += Binomial[i][MapPawns[squares[i]]];
738 goto encode_remaining; // With pawns we have finished special treatments
741 // In positions withouth pawns, we further flip the squares to ensure leading
742 // piece is below RANK_5.
743 if (rank_of(squares[0]) > RANK_4)
744 for (int i = 0; i < size; ++i)
745 squares[i] ^= 070; // Vertical flip: SQ_A8 -> SQ_A1
747 // Look for the first piece of the leading group not on the A1-D4 diagonal
748 // and ensure it is mapped below the diagonal.
749 for (int i = 0; i < d->groupLen[0]; ++i) {
750 if (!off_A1H8(squares[i]))
753 if (off_A1H8(squares[i]) > 0) // A1-H8 diagonal flip: SQ_A3 -> SQ_C3
754 for (int j = i; j < size; ++j)
755 squares[j] = Square(((squares[j] >> 3) | (squares[j] << 3)) & 63);
759 // Encode the leading group.
761 // Suppose we have KRvK. Let's say the pieces are on square numbers wK, wR
762 // and bK (each 0...63). The simplest way to map this position to an index
765 // index = wK * 64 * 64 + wR * 64 + bK;
767 // But this way the TB is going to have 64*64*64 = 262144 positions, with
768 // lots of positions being equivalent (because they are mirrors of each
769 // other) and lots of positions being invalid (two pieces on one square,
770 // adjacent kings, etc.).
771 // Usually the first step is to take the wK and bK together. There are just
772 // 462 ways legal and not-mirrored ways to place the wK and bK on the board.
773 // Once we have placed the wK and bK, there are 62 squares left for the wR
774 // Mapping its square from 0..63 to available squares 0..61 can be done like:
776 // wR -= (wR > wK) + (wR > bK);
778 // In words: if wR "comes later" than wK, we deduct 1, and the same if wR
779 // "comes later" than bK. In case of two same pieces like KRRvK we want to
780 // place the two Rs "together". If we have 62 squares left, we can place two
781 // Rs "together" in 62 * 61 / 2 ways (we divide by 2 because rooks can be
782 // swapped and still get the same position.)
784 // In case we have at least 3 unique pieces (inlcuded kings) we encode them
786 if (entry->hasUniquePieces) {
788 int adjust1 = squares[1] > squares[0];
789 int adjust2 = (squares[2] > squares[0]) + (squares[2] > squares[1]);
791 // First piece is below a1-h8 diagonal. MapA1D1D4[] maps the b1-d1-d3
792 // triangle to 0...5. There are 63 squares for second piece and and 62
793 // (mapped to 0...61) for the third.
794 if (off_A1H8(squares[0]))
795 idx = ( MapA1D1D4[squares[0]] * 63
796 + (squares[1] - adjust1)) * 62
797 + squares[2] - adjust2;
799 // First piece is on a1-h8 diagonal, second below: map this occurence to
800 // 6 to differentiate from the above case, rank_of() maps a1-d4 diagonal
801 // to 0...3 and finally MapB1H1H7[] maps the b1-h1-h7 triangle to 0..27.
802 else if (off_A1H8(squares[1]))
803 idx = ( 6 * 63 + rank_of(squares[0]) * 28
804 + MapB1H1H7[squares[1]]) * 62
805 + squares[2] - adjust2;
807 // First two pieces are on a1-h8 diagonal, third below
808 else if (off_A1H8(squares[2]))
809 idx = 6 * 63 * 62 + 4 * 28 * 62
810 + rank_of(squares[0]) * 7 * 28
811 + (rank_of(squares[1]) - adjust1) * 28
812 + MapB1H1H7[squares[2]];
814 // All 3 pieces on the diagonal a1-h8
816 idx = 6 * 63 * 62 + 4 * 28 * 62 + 4 * 7 * 28
817 + rank_of(squares[0]) * 7 * 6
818 + (rank_of(squares[1]) - adjust1) * 6
819 + (rank_of(squares[2]) - adjust2);
821 // We don't have at least 3 unique pieces, like in KRRvKBB, just map
823 idx = MapKK[MapA1D1D4[squares[0]]][squares[1]];
826 idx *= d->groupIdx[0];
827 Square* groupSq = squares + d->groupLen[0];
829 // Encode remainig pawns then pieces according to square, in ascending order
830 bool remainingPawns = entry->hasPawns && entry->pawnCount[1];
832 while (d->groupLen[++next])
834 std::sort(groupSq, groupSq + d->groupLen[next]);
837 // Map down a square if "comes later" than a square in the previous
838 // groups (similar to what done earlier for leading group pieces).
839 for (int i = 0; i < d->groupLen[next]; ++i)
841 auto f = [&](Square s) { return groupSq[i] > s; };
842 auto adjust = std::count_if(squares, groupSq, f);
843 n += Binomial[i + 1][groupSq[i] - adjust - 8 * remainingPawns];
846 remainingPawns = false;
847 idx += n * d->groupIdx[next];
848 groupSq += d->groupLen[next];
851 // Now that we have the index, decompress the pair and get the score
852 return map_score(entry, tbFile, decompress_pairs(d, idx), wdl);
855 // Group together pieces that will be encoded together. The general rule is that
856 // a group contains pieces of same type and color. The exception is the leading
857 // group that, in case of positions withouth pawns, can be formed by 3 different
858 // pieces (default) or by the king pair when there is not a unique piece apart
859 // from the kings. When there are pawns, pawns are always first in pieces[].
861 // As example KRKN -> KRK + N, KNNK -> KK + NN, KPPKP -> P + PP + K + K
863 // The actual grouping depends on the TB generator and can be inferred from the
864 // sequence of pieces in piece[] array.
866 void set_groups(T& e, PairsData* d, int order[], File f) {
868 int n = 0, firstLen = e.hasPawns ? 0 : e.hasUniquePieces ? 3 : 2;
871 // Number of pieces per group is stored in groupLen[], for instance in KRKN
872 // the encoder will default on '111', so groupLen[] will be (3, 1).
873 for (int i = 1; i < e.pieceCount; ++i)
874 if (--firstLen > 0 || d->pieces[i] == d->pieces[i - 1])
877 d->groupLen[++n] = 1;
879 d->groupLen[++n] = 0; // Zero-terminated
881 // The sequence in pieces[] defines the groups, but not the order in which
882 // they are encoded. If the pieces in a group g can be combined on the board
883 // in N(g) different ways, then the position encoding will be of the form:
885 // g1 * N(g2) * N(g3) + g2 * N(g3) + g3
887 // This ensures unique encoding for the whole position. The order of the
888 // groups is a per-table parameter and could not follow the canonical leading
889 // pawns/pieces -> remainig pawns -> remaining pieces. In particular the
890 // first group is at order[0] position and the remaining pawns, when present,
891 // are at order[1] position.
892 bool pp = e.hasPawns && e.pawnCount[1]; // Pawns on both sides
893 int next = pp ? 2 : 1;
894 int freeSquares = 64 - d->groupLen[0] - (pp ? d->groupLen[1] : 0);
897 for (int k = 0; next < n || k == order[0] || k == order[1]; ++k)
898 if (k == order[0]) // Leading pawns or pieces
900 d->groupIdx[0] = idx;
901 idx *= e.hasPawns ? LeadPawnsSize[d->groupLen[0]][f]
902 : e.hasUniquePieces ? 31332 : 462;
904 else if (k == order[1]) // Remaining pawns
906 d->groupIdx[1] = idx;
907 idx *= Binomial[d->groupLen[1]][48 - d->groupLen[0]];
909 else // Remainig pieces
911 d->groupIdx[next] = idx;
912 idx *= Binomial[d->groupLen[next]][freeSquares];
913 freeSquares -= d->groupLen[next++];
916 d->groupIdx[n] = idx;
919 // In Recursive Pairing each symbol represents a pair of childern symbols. So
920 // read d->btree[] symbols data and expand each one in his left and right child
921 // symbol until reaching the leafs that represent the symbol value.
922 uint8_t set_symlen(PairsData* d, Sym s, std::vector<bool>& visited) {
924 visited[s] = true; // We can set it now because tree is acyclic
925 Sym sr = d->btree[s].get<LR::Right>();
930 Sym sl = d->btree[s].get<LR::Left>();
933 d->symlen[sl] = set_symlen(d, sl, visited);
936 d->symlen[sr] = set_symlen(d, sr, visited);
938 return d->symlen[sl] + d->symlen[sr] + 1;
941 uint8_t* set_sizes(PairsData* d, uint8_t* data) {
945 if (d->flags & TBFlag::SingleValue) {
946 d->blocksNum = d->blockLengthSize = 0;
947 d->span = d->sparseIndexSize = 0; // Broken MSVC zero-init
948 d->minSymLen = *data++; // Here we store the single value
952 // groupLen[] is a zero-terminated list of group lengths, the last groupIdx[]
953 // element stores the biggest index that is the tb size.
954 uint64_t tbSize = d->groupIdx[std::find(d->groupLen, d->groupLen + 7, 0) - d->groupLen];
956 d->sizeofBlock = 1ULL << *data++;
957 d->span = 1ULL << *data++;
958 d->sparseIndexSize = (tbSize + d->span - 1) / d->span; // Round up
959 auto padding = number<uint8_t, LittleEndian>(data++);
960 d->blocksNum = number<uint32_t, LittleEndian>(data); data += sizeof(uint32_t);
961 d->blockLengthSize = d->blocksNum + padding; // Padded to ensure SparseIndex[]
962 // does not point out of range.
963 d->maxSymLen = *data++;
964 d->minSymLen = *data++;
965 d->lowestSym = (Sym*)data;
966 d->base64.resize(d->maxSymLen - d->minSymLen + 1);
968 // The canonical code is ordered such that longer symbols (in terms of
969 // the number of bits of their Huffman code) have lower numeric value,
970 // so that d->lowestSym[i] >= d->lowestSym[i+1] (when read as LittleEndian).
971 // Starting from this we compute a base64[] table indexed by symbol length
972 // and containing 64 bit values so that d->base64[i] >= d->base64[i+1].
973 // See http://www.eecs.harvard.edu/~michaelm/E210/huffman.pdf
974 for (int i = d->base64.size() - 2; i >= 0; --i) {
975 d->base64[i] = (d->base64[i + 1] + number<Sym, LittleEndian>(&d->lowestSym[i])
976 - number<Sym, LittleEndian>(&d->lowestSym[i + 1])) / 2;
978 assert(d->base64[i] * 2 >= d->base64[i+1]);
981 // Now left-shift by an amount so that d->base64[i] gets shifted 1 bit more
982 // than d->base64[i+1] and given the above assert condition, we ensure that
983 // d->base64[i] >= d->base64[i+1]. Moreover for any symbol s64 of length i
984 // and right-padded to 64 bits holds d->base64[i-1] >= s64 >= d->base64[i].
985 for (size_t i = 0; i < d->base64.size(); ++i)
986 d->base64[i] <<= 64 - i - d->minSymLen; // Right-padding to 64 bits
988 data += d->base64.size() * sizeof(Sym);
989 d->symlen.resize(number<uint16_t, LittleEndian>(data)); data += sizeof(uint16_t);
990 d->btree = (LR*)data;
992 // The compression scheme used is "Recursive Pairing", that replaces the most
993 // frequent adjacent pair of symbols in the source message by a new symbol,
994 // reevaluating the frequencies of all of the symbol pairs with respect to
995 // the extended alphabet, and then repeating the process.
996 // See http://www.larsson.dogma.net/dcc99.pdf
997 std::vector<bool> visited(d->symlen.size());
999 for (Sym sym = 0; sym < d->symlen.size(); ++sym)
1001 d->symlen[sym] = set_symlen(d, sym, visited);
1003 return data + d->symlen.size() * sizeof(LR) + (d->symlen.size() & 1);
1006 uint8_t* set_dtz_map(TBTable<WDL>&, uint8_t* data, File) { return data; }
1008 uint8_t* set_dtz_map(TBTable<DTZ>& e, uint8_t* data, File maxFile) {
1012 for (File f = FILE_A; f <= maxFile; ++f) {
1013 auto flags = e.get(0, f)->flags;
1014 if (flags & TBFlag::Mapped) {
1015 if (flags & TBFlag::Wide) {
1016 data += (uintptr_t)data & 1; // Word alignment, we may have a mixed table
1017 for (int i = 0; i < 4; ++i) { // Sequence like 3,x,x,x,1,x,0,2,x,x
1018 e.get(0, f)->map_idx[i] = (uint16_t)((uint16_t *)data - (uint16_t *)e.map + 1);
1019 data += 2 * number<uint16_t, LittleEndian>(data) + 2;
1023 for (int i = 0; i < 4; ++i) {
1024 e.get(0, f)->map_idx[i] = (uint16_t)(data - e.map + 1);
1031 return data += (uintptr_t)data & 1; // Word alignment
1034 // Populate entry's PairsData records with data from the just memory mapped file.
1035 // Called at first access.
1036 template<typename T>
1037 void set(T& e, uint8_t* data) {
1041 enum { Split = 1, HasPawns = 2 };
1043 assert(e.hasPawns == !!(*data & HasPawns));
1044 assert((e.key != e.key2) == !!(*data & Split));
1046 data++; // First byte stores flags
1048 const int sides = T::Sides == 2 && (e.key != e.key2) ? 2 : 1;
1049 const File maxFile = e.hasPawns ? FILE_D : FILE_A;
1051 bool pp = e.hasPawns && e.pawnCount[1]; // Pawns on both sides
1053 assert(!pp || e.pawnCount[0]);
1055 for (File f = FILE_A; f <= maxFile; ++f) {
1057 for (int i = 0; i < sides; i++)
1058 *e.get(i, f) = PairsData();
1060 int order[][2] = { { *data & 0xF, pp ? *(data + 1) & 0xF : 0xF },
1061 { *data >> 4, pp ? *(data + 1) >> 4 : 0xF } };
1064 for (int k = 0; k < e.pieceCount; ++k, ++data)
1065 for (int i = 0; i < sides; i++)
1066 e.get(i, f)->pieces[k] = Piece(i ? *data >> 4 : *data & 0xF);
1068 for (int i = 0; i < sides; ++i)
1069 set_groups(e, e.get(i, f), order[i], f);
1072 data += (uintptr_t)data & 1; // Word alignment
1074 for (File f = FILE_A; f <= maxFile; ++f)
1075 for (int i = 0; i < sides; i++)
1076 data = set_sizes(e.get(i, f), data);
1078 data = set_dtz_map(e, data, maxFile);
1080 for (File f = FILE_A; f <= maxFile; ++f)
1081 for (int i = 0; i < sides; i++) {
1082 (d = e.get(i, f))->sparseIndex = (SparseEntry*)data;
1083 data += d->sparseIndexSize * sizeof(SparseEntry);
1086 for (File f = FILE_A; f <= maxFile; ++f)
1087 for (int i = 0; i < sides; i++) {
1088 (d = e.get(i, f))->blockLength = (uint16_t*)data;
1089 data += d->blockLengthSize * sizeof(uint16_t);
1092 for (File f = FILE_A; f <= maxFile; ++f)
1093 for (int i = 0; i < sides; i++) {
1094 data = (uint8_t*)(((uintptr_t)data + 0x3F) & ~0x3F); // 64 byte alignment
1095 (d = e.get(i, f))->data = data;
1096 data += d->blocksNum * d->sizeofBlock;
1100 // If the TB file corresponding to the given position is already memory mapped
1101 // then return its base address, otherwise try to memory map and init it. Called
1102 // at every probe, memory map and init only at first access. Function is thread
1103 // safe and can be called concurrently.
1104 template<TBType Type>
1105 void* mapped(TBTable<Type>& e, const Position& pos) {
1109 // Use 'aquire' to avoid a thread reads 'ready' == true while another is
1110 // still working, this could happen due to compiler reordering.
1111 if (e.ready.load(std::memory_order_acquire))
1112 return e.baseAddress; // Could be nullptr if file does not exsist
1114 std::unique_lock<Mutex> lk(mutex);
1116 if (e.ready.load(std::memory_order_relaxed)) // Recheck under lock
1117 return e.baseAddress;
1119 // Pieces strings in decreasing order for each color, like ("KPP","KR")
1120 std::string fname, w, b;
1121 for (PieceType pt = KING; pt >= PAWN; --pt) {
1122 w += std::string(popcount(pos.pieces(WHITE, pt)), PieceToChar[pt]);
1123 b += std::string(popcount(pos.pieces(BLACK, pt)), PieceToChar[pt]);
1126 fname = (e.key == pos.material_key() ? w + 'v' + b : b + 'v' + w)
1127 + (Type == WDL ? ".rtbw" : ".rtbz");
1129 uint8_t* data = TBFile(fname).map(&e.baseAddress, &e.mapping, Type);
1134 e.ready.store(true, std::memory_order_release);
1135 return e.baseAddress;
1138 template<TBType Type, typename Ret = typename TBTable<Type>::Ret>
1139 Ret probe_table(const Position& pos, ProbeState* result, WDLScore wdl = WDLDraw) {
1141 if (pos.count<ALL_PIECES>() == 2) // KvK
1142 return Ret(WDLDraw);
1144 TBTable<Type>* entry = TBTables.get<Type>(pos.material_key());
1146 if (!entry || !mapped(*entry, pos))
1147 return *result = FAIL, Ret();
1149 return do_probe_table(pos, entry, wdl, result);
1152 // For a position where the side to move has a winning capture it is not necessary
1153 // to store a winning value so the generator treats such positions as "don't cares"
1154 // and tries to assign to it a value that improves the compression ratio. Similarly,
1155 // if the side to move has a drawing capture, then the position is at least drawn.
1156 // If the position is won, then the TB needs to store a win value. But if the
1157 // position is drawn, the TB may store a loss value if that is better for compression.
1158 // All of this means that during probing, the engine must look at captures and probe
1159 // their results and must probe the position itself. The "best" result of these
1160 // probes is the correct result for the position.
1161 // DTZ tables do not store values when a following move is a zeroing winning move
1162 // (winning capture or winning pawn move). Also DTZ store wrong values for positions
1163 // where the best move is an ep-move (even if losing). So in all these cases set
1164 // the state to ZEROING_BEST_MOVE.
1165 template<bool CheckZeroingMoves>
1166 WDLScore search(Position& pos, ProbeState* result) {
1168 WDLScore value, bestValue = WDLLoss;
1171 auto moveList = MoveList<LEGAL>(pos);
1172 size_t totalCount = moveList.size(), moveCount = 0;
1174 for (const Move& move : moveList)
1176 if ( !pos.capture(move)
1177 && (!CheckZeroingMoves || type_of(pos.moved_piece(move)) != PAWN))
1182 pos.do_move(move, st);
1183 value = -search<false>(pos, result);
1184 pos.undo_move(move);
1186 if (*result == FAIL)
1189 if (value > bestValue)
1193 if (value >= WDLWin)
1195 *result = ZEROING_BEST_MOVE; // Winning DTZ-zeroing move
1201 // In case we have already searched all the legal moves we don't have to probe
1202 // the TB because the stored score could be wrong. For instance TB tables
1203 // do not contain information on position with ep rights, so in this case
1204 // the result of probe_wdl_table is wrong. Also in case of only capture
1205 // moves, for instance here 4K3/4q3/6p1/2k5/6p1/8/8/8 w - - 0 7, we have to
1206 // return with ZEROING_BEST_MOVE set.
1207 bool noMoreMoves = (moveCount && moveCount == totalCount);
1213 value = probe_table<WDL>(pos, result);
1215 if (*result == FAIL)
1219 // DTZ stores a "don't care" value if bestValue is a win
1220 if (bestValue >= value)
1221 return *result = ( bestValue > WDLDraw
1222 || noMoreMoves ? ZEROING_BEST_MOVE : OK), bestValue;
1224 return *result = OK, value;
1230 /// Tablebases::init() is called at startup and after every change to
1231 /// "SyzygyPath" UCI option to (re)create the various tables. It is not thread
1232 /// safe, nor it needs to be.
1233 void Tablebases::init(const std::string& paths) {
1237 TBFile::Paths = paths;
1239 if (paths.empty() || paths == "<empty>")
1242 // MapB1H1H7[] encodes a square below a1-h8 diagonal to 0..27
1244 for (Square s = SQ_A1; s <= SQ_H8; ++s)
1245 if (off_A1H8(s) < 0)
1246 MapB1H1H7[s] = code++;
1248 // MapA1D1D4[] encodes a square in the a1-d1-d4 triangle to 0..9
1249 std::vector<Square> diagonal;
1251 for (Square s = SQ_A1; s <= SQ_D4; ++s)
1252 if (off_A1H8(s) < 0 && file_of(s) <= FILE_D)
1253 MapA1D1D4[s] = code++;
1255 else if (!off_A1H8(s) && file_of(s) <= FILE_D)
1256 diagonal.push_back(s);
1258 // Diagonal squares are encoded as last ones
1259 for (auto s : diagonal)
1260 MapA1D1D4[s] = code++;
1262 // MapKK[] encodes all the 461 possible legal positions of two kings where
1263 // the first is in the a1-d1-d4 triangle. If the first king is on the a1-d4
1264 // diagonal, the other one shall not to be above the a1-h8 diagonal.
1265 std::vector<std::pair<int, Square>> bothOnDiagonal;
1267 for (int idx = 0; idx < 10; idx++)
1268 for (Square s1 = SQ_A1; s1 <= SQ_D4; ++s1)
1269 if (MapA1D1D4[s1] == idx && (idx || s1 == SQ_B1)) // SQ_B1 is mapped to 0
1271 for (Square s2 = SQ_A1; s2 <= SQ_H8; ++s2)
1272 if ((PseudoAttacks[KING][s1] | s1) & s2)
1273 continue; // Illegal position
1275 else if (!off_A1H8(s1) && off_A1H8(s2) > 0)
1276 continue; // First on diagonal, second above
1278 else if (!off_A1H8(s1) && !off_A1H8(s2))
1279 bothOnDiagonal.push_back(std::make_pair(idx, s2));
1282 MapKK[idx][s2] = code++;
1285 // Legal positions with both kings on diagonal are encoded as last ones
1286 for (auto p : bothOnDiagonal)
1287 MapKK[p.first][p.second] = code++;
1289 // Binomial[] stores the Binomial Coefficents using Pascal rule. There
1290 // are Binomial[k][n] ways to choose k elements from a set of n elements.
1293 for (int n = 1; n < 64; n++) // Squares
1294 for (int k = 0; k < 6 && k <= n; ++k) // Pieces
1295 Binomial[k][n] = (k > 0 ? Binomial[k - 1][n - 1] : 0)
1296 + (k < n ? Binomial[k ][n - 1] : 0);
1298 // MapPawns[s] encodes squares a2-h7 to 0..47. This is the number of possible
1299 // available squares when the leading one is in 's'. Moreover the pawn with
1300 // highest MapPawns[] is the leading pawn, the one nearest the edge and,
1301 // among pawns with same file, the one with lowest rank.
1302 int availableSquares = 47; // Available squares when lead pawn is in a2
1304 // Init the tables for the encoding of leading pawns group: with 7-men TB we
1305 // can have up to 5 leading pawns (KPPPPPK).
1306 for (int leadPawnsCnt = 1; leadPawnsCnt <= 5; ++leadPawnsCnt)
1307 for (File f = FILE_A; f <= FILE_D; ++f)
1309 // Restart the index at every file because TB table is splitted
1310 // by file, so we can reuse the same index for different files.
1313 // Sum all possible combinations for a given file, starting with
1314 // the leading pawn on rank 2 and increasing the rank.
1315 for (Rank r = RANK_2; r <= RANK_7; ++r)
1317 Square sq = make_square(f, r);
1319 // Compute MapPawns[] at first pass.
1320 // If sq is the leading pawn square, any other pawn cannot be
1321 // below or more toward the edge of sq. There are 47 available
1322 // squares when sq = a2 and reduced by 2 for any rank increase
1323 // due to mirroring: sq == a3 -> no a2, h2, so MapPawns[a3] = 45
1324 if (leadPawnsCnt == 1)
1326 MapPawns[sq] = availableSquares--;
1327 MapPawns[sq ^ 7] = availableSquares--; // Horizontal flip
1329 LeadPawnIdx[leadPawnsCnt][sq] = idx;
1330 idx += Binomial[leadPawnsCnt - 1][MapPawns[sq]];
1332 // After a file is traversed, store the cumulated per-file index
1333 LeadPawnsSize[leadPawnsCnt][f] = idx;
1336 // Add entries in TB tables if the corresponding ".rtbw" file exsists
1337 for (PieceType p1 = PAWN; p1 < KING; ++p1) {
1338 TBTables.add({KING, p1, KING});
1340 for (PieceType p2 = PAWN; p2 <= p1; ++p2) {
1341 TBTables.add({KING, p1, p2, KING});
1342 TBTables.add({KING, p1, KING, p2});
1344 for (PieceType p3 = PAWN; p3 < KING; ++p3)
1345 TBTables.add({KING, p1, p2, KING, p3});
1347 for (PieceType p3 = PAWN; p3 <= p2; ++p3) {
1348 TBTables.add({KING, p1, p2, p3, KING});
1350 for (PieceType p4 = PAWN; p4 <= p3; ++p4) {
1351 TBTables.add({KING, p1, p2, p3, p4, KING});
1353 for (PieceType p5 = PAWN; p5 <= p4; ++p5)
1354 TBTables.add({KING, p1, p2, p3, p4, p5, KING});
1356 for (PieceType p5 = PAWN; p5 < KING; ++p5)
1357 TBTables.add({KING, p1, p2, p3, p4, KING, p5});
1360 for (PieceType p4 = PAWN; p4 < KING; ++p4) {
1361 TBTables.add({KING, p1, p2, p3, KING, p4});
1363 for (PieceType p5 = PAWN; p5 <= p4; ++p5)
1364 TBTables.add({KING, p1, p2, p3, KING, p4, p5});
1368 for (PieceType p3 = PAWN; p3 <= p1; ++p3)
1369 for (PieceType p4 = PAWN; p4 <= (p1 == p3 ? p2 : p3); ++p4)
1370 TBTables.add({KING, p1, p2, KING, p3, p4});
1374 sync_cout << "info string Found " << TBTables.size() << " tablebases" << sync_endl;
1377 // Probe the WDL table for a particular position.
1378 // If *result != FAIL, the probe was successful.
1379 // The return value is from the point of view of the side to move:
1381 // -1 : loss, but draw under 50-move rule
1383 // 1 : win, but draw under 50-move rule
1385 WDLScore Tablebases::probe_wdl(Position& pos, ProbeState* result) {
1388 return search<false>(pos, result);
1391 // Probe the DTZ table for a particular position.
1392 // If *result != FAIL, the probe was successful.
1393 // The return value is from the point of view of the side to move:
1394 // n < -100 : loss, but draw under 50-move rule
1395 // -100 <= n < -1 : loss in n ply (assuming 50-move counter == 0)
1396 // -1 : loss, the side to move is mated
1398 // 1 < n <= 100 : win in n ply (assuming 50-move counter == 0)
1399 // 100 < n : win, but draw under 50-move rule
1401 // The return value n can be off by 1: a return value -n can mean a loss
1402 // in n+1 ply and a return value +n can mean a win in n+1 ply. This
1403 // cannot happen for tables with positions exactly on the "edge" of
1404 // the 50-move rule.
1406 // This implies that if dtz > 0 is returned, the position is certainly
1407 // a win if dtz + 50-move-counter <= 99. Care must be taken that the engine
1408 // picks moves that preserve dtz + 50-move-counter <= 99.
1410 // If n = 100 immediately after a capture or pawn move, then the position
1411 // is also certainly a win, and during the whole phase until the next
1412 // capture or pawn move, the inequality to be preserved is
1413 // dtz + 50-movecounter <= 100.
1415 // In short, if a move is available resulting in dtz + 50-move-counter <= 99,
1416 // then do not accept moves leading to dtz + 50-move-counter == 100.
1417 int Tablebases::probe_dtz(Position& pos, ProbeState* result) {
1420 WDLScore wdl = search<true>(pos, result);
1422 if (*result == FAIL || wdl == WDLDraw) // DTZ tables don't store draws
1425 // DTZ stores a 'don't care' value in this case, or even a plain wrong
1426 // one as in case the best move is a losing ep, so it cannot be probed.
1427 if (*result == ZEROING_BEST_MOVE)
1428 return dtz_before_zeroing(wdl);
1430 int dtz = probe_table<DTZ>(pos, result, wdl);
1432 if (*result == FAIL)
1435 if (*result != CHANGE_STM)
1436 return (dtz + 100 * (wdl == WDLBlessedLoss || wdl == WDLCursedWin)) * sign_of(wdl);
1438 // DTZ stores results for the other side, so we need to do a 1-ply search and
1439 // find the winning move that minimizes DTZ.
1441 int minDTZ = 0xFFFF;
1443 for (const Move& move : MoveList<LEGAL>(pos))
1445 bool zeroing = pos.capture(move) || type_of(pos.moved_piece(move)) == PAWN;
1447 pos.do_move(move, st);
1449 // For zeroing moves we want the dtz of the move _before_ doing it,
1450 // otherwise we will get the dtz of the next move sequence. Search the
1451 // position after the move to get the score sign (because even in a
1452 // winning position we could make a losing capture or going for a draw).
1453 dtz = zeroing ? -dtz_before_zeroing(search<false>(pos, result))
1454 : -probe_dtz(pos, result);
1456 // If the move mates, force minDTZ to 1
1457 if (dtz == 1 && pos.checkers() && MoveList<LEGAL>(pos).size() == 0)
1460 // Convert result from 1-ply search. Zeroing moves are already accounted
1461 // by dtz_before_zeroing() that returns the DTZ of the previous move.
1463 dtz += sign_of(dtz);
1465 // Skip the draws and if we are winning only pick positive dtz
1466 if (dtz < minDTZ && sign_of(dtz) == sign_of(wdl))
1469 pos.undo_move(move);
1471 if (*result == FAIL)
1475 // When there are no legal moves, the position is mate: we return -1
1476 return minDTZ == 0xFFFF ? -1 : minDTZ;
1480 // Use the DTZ tables to rank root moves.
1482 // A return value false indicates that not all probes were successful.
1483 bool Tablebases::root_probe(Position& pos, Search::RootMoves& rootMoves) {
1488 // Obtain 50-move counter for the root position
1489 int cnt50 = pos.rule50_count();
1491 // Check whether a position was repeated since the last zeroing move.
1492 bool rep = pos.has_repeated();
1494 int dtz, bound = Options["Syzygy50MoveRule"] ? 900 : 1;
1496 // Probe and rank each move
1497 for (auto& m : rootMoves)
1499 pos.do_move(m.pv[0], st);
1501 // Calculate dtz for the current move counting from the root position
1502 if (pos.rule50_count() == 0)
1504 // In case of a zeroing move, dtz is one of -101/-1/0/1/101
1505 WDLScore wdl = -probe_wdl(pos, &result);
1506 dtz = dtz_before_zeroing(wdl);
1510 // Otherwise, take dtz for the new position and correct by 1 ply
1511 dtz = -probe_dtz(pos, &result);
1512 dtz = dtz > 0 ? dtz + 1
1513 : dtz < 0 ? dtz - 1 : dtz;
1516 // Make sure that a mating move is assigned a dtz value of 1
1519 && MoveList<LEGAL>(pos).size() == 0)
1522 pos.undo_move(m.pv[0]);
1527 // Better moves are ranked higher. Certain wins are ranked equally.
1528 // Losing moves are ranked equally unless a 50-move draw is in sight.
1529 int r = dtz > 0 ? (dtz + cnt50 <= 99 && !rep ? 1000 : 1000 - (dtz + cnt50))
1530 : dtz < 0 ? (-dtz * 2 + cnt50 < 100 ? -1000 : -1000 + (-dtz + cnt50))
1534 // Determine the score to be displayed for this move. Assign at least
1535 // 1 cp to cursed wins and let it grow to 49 cp as the positions gets
1536 // closer to a real win.
1537 m.tbScore = r >= bound ? VALUE_MATE - MAX_PLY - 1
1538 : r > 0 ? Value((std::max( 3, r - 800) * int(PawnValueEg)) / 200)
1539 : r == 0 ? VALUE_DRAW
1540 : r > -bound ? Value((std::min(-3, r + 800) * int(PawnValueEg)) / 200)
1541 : -VALUE_MATE + MAX_PLY + 1;
1548 // Use the WDL tables to rank root moves.
1549 // This is a fallback for the case that some or all DTZ tables are missing.
1551 // A return value false indicates that not all probes were successful.
1552 bool Tablebases::root_probe_wdl(Position& pos, Search::RootMoves& rootMoves) {
1554 static const int WDL_to_rank[] = { -1000, -899, 0, 899, 1000 };
1559 bool rule50 = Options["Syzygy50MoveRule"];
1561 // Probe and rank each move
1562 for (auto& m : rootMoves)
1564 pos.do_move(m.pv[0], st);
1566 WDLScore wdl = -probe_wdl(pos, &result);
1568 pos.undo_move(m.pv[0]);
1573 m.tbRank = WDL_to_rank[wdl + 2];
1576 wdl = wdl > WDLDraw ? WDLWin
1577 : wdl < WDLDraw ? WDLLoss : WDLDraw;
1578 m.tbScore = WDL_to_value[wdl + 2];