From: Ondrej Mosnáček Date: Sat, 10 Mar 2018 13:37:42 +0000 (+0100) Subject: Refactor tbprobe.cpp X-Git-Url: https://git.sesse.net/?p=stockfish;a=commitdiff_plain;h=8ff2fcf299ac621779dc167ce750ba73aaae90a4;hp=280022baf78a52e9a5c34ef2aec6927aace3e83f Refactor tbprobe.cpp This involves: * replacing the union hacks with simply reusing the EntryPiece arrays for the no-pawns case * merging the PairsData structure with the EntryPiece/-Pawn structs (with credit to Marco: @mcostalba) * simplifying some HashTable functions * thanks to previous changes, removing the ugly memsets * simplifying the template logic for WDL/DTZ distinction (now we distinguish based on an enum type, not the entry classes) * removing the unneeded Atomic wrapper ----------------------------- For reference, here is a manual way to check that patches concerning table bases code are non-functional changes: 0) Download the Syzygy table bases (up to 6 men). 1) Make sure you have branches master and the pull request pointing to the right commits. 2) Download the bench calculation scripts from the following URL: https://gist.github.com/WOnder93/b5fcf9c989b4a1715684d5c82367cdbe and copy into src inside your Stockfish repo. 3) Make the scripts executable (chmod +x *.sh). 4) Run the following command to use TBs located at : export SYZYGY_PATH='' 5) After that, run this (it will take a long time, this is a deep bench): BENCH_ARGS='128 1 22' ./check_benches.sh master tbprobe_cleanup 2>/dev/null` ==> You should see two equal numbers printed. (Of course, now we have to trust that the script itself is correct :) ----------------------------- Closes https://github.com/official-stockfish/Stockfish/pull/1477 No functional change. --- diff --git a/src/syzygy/tbprobe.cpp b/src/syzygy/tbprobe.cpp index 722679fd..9b00bac2 100644 --- a/src/syzygy/tbprobe.cpp +++ b/src/syzygy/tbprobe.cpp @@ -20,7 +20,7 @@ #include #include #include -#include // For std::memset +#include // For std::memset and std::memcpy #include #include #include @@ -54,6 +54,10 @@ int Tablebases::MaxCardinality; namespace { +constexpr int TBPIECES = 6; // Max number of supported pieces + +enum TBType { WDL, DTZ }; // Used as template parameter + // Each table has a set of flags: all of them refer to DTZ tables, the last one to WDL tables enum TBFlag { STM = 1, Mapped = 2, WinPlies = 4, LossPlies = 8, SingleValue = 128 }; @@ -102,8 +106,6 @@ struct LR { static_assert(sizeof(LR) == 3, "LR tree entry must be 3 bytes"); -constexpr int TBPIECES = 6; - struct PairsData { int flags; size_t sizeofBlock; // Block size in bytes @@ -123,81 +125,81 @@ struct PairsData { 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) }; -// Helper struct to avoid manually defining entry copy constructor as we -// should because the default one is not compatible with std::atomic_bool. -struct Atomic { - Atomic() = default; - Atomic(const Atomic& e) { ready = e.ready.load(); } // MSVC 2013 wants assignment within body - std::atomic_bool ready; -}; - -// We define types for the different parts of the WDLEntry and DTZEntry with -// corresponding specializations for pieces or pawns. - -struct WDLEntryPiece { - PairsData* precomp; -}; +template +struct TBEntry { + typedef typename std::conditional::type Result; -struct WDLEntryPawn { - uint8_t pawnCount[2]; // [Lead color / other color] - WDLEntryPiece file[2][4]; // [wtm / btm][FILE_A..FILE_D] -}; - -struct DTZEntryPiece { - PairsData* precomp; - uint16_t map_idx[4]; // WDLWin, WDLLoss, WDLCursedWin, WDLBlessedLoss - uint8_t* map; -}; - -struct DTZEntryPawn { - uint8_t pawnCount[2]; - DTZEntryPiece file[4]; - uint8_t* map; -}; + static constexpr int Sides = Type == WDL ? 2 : 1; -struct TBEntry : public Atomic { + 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] -// Now the main types: WDLEntry and DTZEntry -struct WDLEntry : public TBEntry { - WDLEntry(const std::string& code); - ~WDLEntry(); - union { - WDLEntryPiece pieceTable[2]; // [wtm / btm] - WDLEntryPawn pawnTable; - }; -}; + PairsData* get(int stm, int f) { + return &items[stm % Sides][hasPawns ? f : 0]; + } -struct DTZEntry : public TBEntry { - DTZEntry(const WDLEntry& wdl); - ~DTZEntry(); - union { - DTZEntryPiece pieceTable; - DTZEntryPawn pawnTable; - }; + TBEntry() : ready(false), baseAddress(nullptr) {} + explicit TBEntry(const std::string& code); + explicit TBEntry(const TBEntry& wdl); + ~TBEntry(); }; -typedef decltype(WDLEntry::pieceTable) WDLPieceTable; -typedef decltype(DTZEntry::pieceTable) DTZPieceTable; -typedef decltype(WDLEntry::pawnTable ) WDLPawnTable; -typedef decltype(DTZEntry::pawnTable ) DTZPawnTable; +template<> +TBEntry::TBEntry(const std::string& code) : TBEntry() { + + StateInfo st; + Position pos; + + key = pos.set(code, WHITE, &st).material_key(); + pieceCount = popcount(pos.pieces()); + hasPawns = pos.pieces(PAWN); + + hasUniquePieces = false; + for (Color c = WHITE; c <= BLACK; ++c) + for (PieceType pt = PAWN; pt < KING; ++pt) + if (popcount(pos.pieces(c, pt)) == 1) + hasUniquePieces = true; + + if (hasPawns) { + // Set the leading color. In case both sides have pawns the leading color + // is the side with less pawns because this leads to better compression. + bool c = !pos.count(BLACK) + || ( pos.count(WHITE) + && pos.count(BLACK) >= pos.count(WHITE)); -auto item(WDLPieceTable& e, int stm, int ) -> decltype(e[stm])& { return e[stm]; } -auto item(DTZPieceTable& e, int , int ) -> decltype(e)& { return e; } -auto item(WDLPawnTable& e, int stm, int f) -> decltype(e.file[stm][f])& { return e.file[stm][f]; } -auto item(DTZPawnTable& e, int , int f) -> decltype(e.file[f])& { return e.file[f]; } + pawnCount[0] = pos.count(c ? WHITE : BLACK); + pawnCount[1] = pos.count(c ? BLACK : WHITE); + } -template struct Ret { typedef int type; }; -template<> struct Ret { typedef WDLScore type; }; + key2 = pos.set(code, BLACK, &st).material_key(); +} + +template<> +TBEntry::TBEntry(const TBEntry& wdl) : TBEntry() { + + key = wdl.key; + key2 = wdl.key2; + pieceCount = wdl.pieceCount; + hasPawns = wdl.hasPawns; + hasUniquePieces = wdl.hasUniquePieces; + + if (hasPawns) { + pawnCount[0] = wdl.pawnCount[0]; + pawnCount[1] = wdl.pawnCount[1]; + } +} int MapPawns[SQUARE_NB]; int MapB1H1H7[SQUARE_NB]; @@ -252,7 +254,7 @@ template T number(void* addr) class HashTable { - typedef std::pair EntryPair; + typedef std::pair*, TBEntry*> EntryPair; typedef std::pair Entry; static constexpr int TBHASHBITS = 10; @@ -260,15 +262,13 @@ class HashTable { Entry hashTable[1 << TBHASHBITS][HSHMAX]; - std::deque wdlTable; - std::deque dtzTable; + std::deque> wdlTable; + std::deque> dtzTable; - void insert(Key key, WDLEntry* wdl, DTZEntry* dtz) { - Entry* entry = hashTable[key >> (64 - TBHASHBITS)]; - - for (int i = 0; i < HSHMAX; ++i, ++entry) - if (!entry->second.first || entry->first == key) { - *entry = std::make_pair(key, std::make_pair(wdl, dtz)); + void insert(Key key, TBEntry* wdl, TBEntry* dtz) { + for (Entry& entry : hashTable[key >> (64 - TBHASHBITS)]) + if (!entry.second.first || entry.first == key) { + entry = std::make_pair(key, std::make_pair(wdl, dtz)); return; } @@ -277,24 +277,22 @@ class HashTable { } public: - template::value ? 0 : 1> - E* get(Key key) { - Entry* entry = hashTable[key >> (64 - TBHASHBITS)]; - - for (int i = 0; i < HSHMAX; ++i, ++entry) - if (entry->first == key) - return std::get(entry->second); - - return nullptr; - } - - void clear() { - std::memset(hashTable, 0, sizeof(hashTable)); - wdlTable.clear(); - dtzTable.clear(); - } - size_t size() const { return wdlTable.size(); } - void insert(const std::vector& pieces); + template + TBEntry* get(Key key) { + for (Entry& entry : hashTable[key >> (64 - TBHASHBITS)]) + if (entry.first == key) + return std::get(entry.second); + + return nullptr; + } + + void clear() { + memset(hashTable, 0, sizeof(hashTable)); + wdlTable.clear(); + dtzTable.clear(); + } + size_t size() const { return wdlTable.size(); } + void insert(const std::vector& pieces); }; HashTable EntryTable; @@ -407,77 +405,10 @@ public: std::string TBFile::Paths; -WDLEntry::WDLEntry(const std::string& code) { - - StateInfo st; - Position pos; - - memset(this, 0, sizeof(WDLEntry)); - - ready = false; - key = pos.set(code, WHITE, &st).material_key(); - pieceCount = popcount(pos.pieces()); - hasPawns = pos.pieces(PAWN); - - for (Color c = WHITE; c <= BLACK; ++c) - for (PieceType pt = PAWN; pt < KING; ++pt) - if (popcount(pos.pieces(c, pt)) == 1) - hasUniquePieces = true; - - if (hasPawns) { - // Set the leading color. In case both sides have pawns the leading color - // is the side with less pawns because this leads to better compression. - bool c = !pos.count(BLACK) - || ( pos.count(WHITE) - && pos.count(BLACK) >= pos.count(WHITE)); - - pawnTable.pawnCount[0] = pos.count(c ? WHITE : BLACK); - pawnTable.pawnCount[1] = pos.count(c ? BLACK : WHITE); - } - - key2 = pos.set(code, BLACK, &st).material_key(); -} - -WDLEntry::~WDLEntry() { - - if (baseAddress) - TBFile::unmap(baseAddress, mapping); - - for (int i = 0; i < 2; ++i) - if (hasPawns) - for (File f = FILE_A; f <= FILE_D; ++f) - delete pawnTable.file[i][f].precomp; - else - delete pieceTable[i].precomp; -} - -DTZEntry::DTZEntry(const WDLEntry& wdl) { - - memset(this, 0, sizeof(DTZEntry)); - - ready = false; - key = wdl.key; - key2 = wdl.key2; - pieceCount = wdl.pieceCount; - hasPawns = wdl.hasPawns; - hasUniquePieces = wdl.hasUniquePieces; - - if (hasPawns) { - pawnTable.pawnCount[0] = wdl.pawnTable.pawnCount[0]; - pawnTable.pawnCount[1] = wdl.pawnTable.pawnCount[1]; - } -} - -DTZEntry::~DTZEntry() { - +template +TBEntry::~TBEntry() { if (baseAddress) TBFile::unmap(baseAddress, mapping); - - if (hasPawns) - for (File f = FILE_A; f <= FILE_D; ++f) - delete pawnTable.file[f].precomp; - else - delete pieceTable.precomp; } void HashTable::insert(const std::vector& pieces) { @@ -630,13 +561,11 @@ int decompress_pairs(PairsData* d, uint64_t idx) { return d->btree[sym].get(); } -bool check_dtz_stm(WDLEntry*, int, File) { return true; } - -bool check_dtz_stm(DTZEntry* entry, int stm, File f) { +bool check_dtz_stm(TBEntry*, int, File) { return true; } - int flags = entry->hasPawns ? entry->pawnTable.file[f].precomp->flags - : entry->pieceTable.precomp->flags; +bool check_dtz_stm(TBEntry* entry, int stm, File f) { + int flags = entry->get(stm, f)->flags; return (flags & TBFlag::STM) == stm || ((entry->key == entry->key2) && !entry->hasPawns); } @@ -645,20 +574,16 @@ bool check_dtz_stm(DTZEntry* entry, int stm, File f) { // values 0, 1, 2, ... in order of decreasing frequency. This is done for each // of the four WDLScore values. The mapping information necessary to reconstruct // the original values is stored in the TB file and read during map[] init. -WDLScore map_score(WDLEntry*, File, int value, WDLScore) { return WDLScore(value - 2); } +WDLScore map_score(TBEntry*, File, int value, WDLScore) { return WDLScore(value - 2); } -int map_score(DTZEntry* entry, File f, int value, WDLScore wdl) { +int map_score(TBEntry* entry, File f, int value, WDLScore wdl) { constexpr int WDLMap[] = { 1, 3, 0, 2, 0 }; - int flags = entry->hasPawns ? entry->pawnTable.file[f].precomp->flags - : entry->pieceTable.precomp->flags; - - uint8_t* map = entry->hasPawns ? entry->pawnTable.map - : entry->pieceTable.map; + int flags = entry->get(0, f)->flags; - uint16_t* idx = entry->hasPawns ? entry->pawnTable.file[f].map_idx - : entry->pieceTable.map_idx; + uint8_t* map = entry->map; + uint16_t* idx = entry->get(0, f)->map_idx; if (flags & TBFlag::Mapped) value = map[idx[WDLMap[wdl + 2]] + value]; @@ -679,10 +604,8 @@ int map_score(DTZEntry* entry, File f, int value, WDLScore wdl) { // // idx = Binomial[1][s1] + Binomial[2][s2] + ... + Binomial[k][sk] // -template::type> -T do_probe_table(const Position& pos, Entry* entry, WDLScore wdl, ProbeState* result) { - - constexpr bool IsWDL = std::is_same::value; +template::Result> +T do_probe_table(const Position& pos, TBEntry* entry, WDLScore wdl, ProbeState* result) { Square squares[TBPIECES]; Piece pieces[TBPIECES]; @@ -715,7 +638,7 @@ T do_probe_table(const Position& pos, Entry* entry, WDLScore wdl, ProbeState* re // In all the 4 tables, pawns are at the beginning of the piece sequence and // their color is the reference one. So we just pick the first one. - Piece pc = Piece(item(entry->pawnTable, 0, 0).precomp->pieces[0] ^ flipColor); + Piece pc = Piece(entry->get(0, 0)->pieces[0] ^ flipColor); assert(type_of(pc) == PAWN); @@ -731,15 +654,12 @@ T do_probe_table(const Position& pos, Entry* entry, WDLScore wdl, ProbeState* re tbFile = file_of(squares[0]); if (tbFile > FILE_D) tbFile = file_of(squares[0] ^ 7); // Horizontal flip: SQ_H1 -> SQ_A1 - - d = item(entry->pawnTable , stm, tbFile).precomp; - } else - d = item(entry->pieceTable, stm, tbFile).precomp; + } // DTZ tables are one-sided, i.e. they store positions only for white to // move or only for black to move, so check for side to move to be stm, // early exit otherwise. - if (!IsWDL && !check_dtz_stm(entry, stm, tbFile)) + if (Type == DTZ && !check_dtz_stm(entry, stm, tbFile)) return *result = CHANGE_STM, T(); // Now we are ready to get all the position pieces (but the lead pawns) and @@ -753,8 +673,10 @@ T do_probe_table(const Position& pos, Entry* entry, WDLScore wdl, ProbeState* re assert(size >= 2); + d = entry->get(stm, tbFile); + // Then we reorder the pieces to have the same sequence as the one stored - // in precomp->pieces[i]: the sequence that ensures the best compression. + // in pieces[i]: the sequence that ensures the best compression. for (int i = leadPawnsCnt; i < size; ++i) for (int j = i; j < size; ++j) if (d->pieces[i] == pieces[j]) @@ -872,7 +794,7 @@ encode_remaining: Square* groupSq = squares + d->groupLen[0]; // Encode remainig pawns then pieces according to square, in ascending order - bool remainingPawns = entry->hasPawns && entry->pawnTable.pawnCount[1]; + bool remainingPawns = entry->hasPawns && entry->pawnCount[1]; while (d->groupLen[++next]) { @@ -907,8 +829,8 @@ encode_remaining: // // The actual grouping depends on the TB generator and can be inferred from the // sequence of pieces in piece[] array. -template -void set_groups(T& e, PairsData* d, int order[], File f) { +template +void set_groups(TBEntry& e, PairsData* d, int order[], File f) { int n = 0, firstLen = e.hasPawns ? 0 : e.hasUniquePieces ? 3 : 2; d->groupLen[n] = 1; @@ -934,7 +856,7 @@ void set_groups(T& e, PairsData* d, int order[], File f) { // pawns/pieces -> remainig pawns -> remaining pieces. In particular the // first group is at order[0] position and the remaining pawns, when present, // are at order[1] position. - bool pp = e.hasPawns && e.pawnTable.pawnCount[1]; // Pawns on both sides + bool pp = e.hasPawns && e.pawnCount[1]; // Pawns on both sides int next = pp ? 2 : 1; int freeSquares = 64 - d->groupLen[0] - (pp ? d->groupLen[1] : 0); uint64_t idx = 1; @@ -1048,18 +970,16 @@ uint8_t* set_sizes(PairsData* d, uint8_t* data) { return data + d->symlen.size() * sizeof(LR) + (d->symlen.size() & 1); } -template -uint8_t* set_dtz_map(WDLEntry&, T&, uint8_t*, File) { return nullptr; } +uint8_t* set_dtz_map(TBEntry&, uint8_t*, File) { return nullptr; } -template -uint8_t* set_dtz_map(DTZEntry&, T& p, uint8_t* data, File maxFile) { +uint8_t* set_dtz_map(TBEntry& e, uint8_t* data, File maxFile) { - p.map = data; + e.map = data; for (File f = FILE_A; f <= maxFile; ++f) { - if (item(p, 0, f).precomp->flags & TBFlag::Mapped) + if (e.get(0, f)->flags & TBFlag::Mapped) for (int i = 0; i < 4; ++i) { // Sequence like 3,x,x,x,1,x,0,2,x,x - item(p, 0, f).map_idx[i] = (uint16_t)(data - p.map + 1); + e.get(0, f)->map_idx[i] = (uint16_t)(data - e.map + 1); data += *data + 1; } } @@ -1067,10 +987,8 @@ uint8_t* set_dtz_map(DTZEntry&, T& p, uint8_t* data, File maxFile) { return data += (uintptr_t)data & 1; // Word alignment } -template -void do_init(Entry& e, T& p, uint8_t* data) { - - const bool IsWDL = std::is_same::value; +template +void do_init(TBEntry& e, uint8_t* data) { PairsData* d; @@ -1081,17 +999,17 @@ void do_init(Entry& e, T& p, uint8_t* data) { data++; // First byte stores flags - const int sides = IsWDL && (e.key != e.key2) ? 2 : 1; + const int sides = Type == WDL && (e.key != e.key2) ? 2 : 1; const File maxFile = e.hasPawns ? FILE_D : FILE_A; - bool pp = e.hasPawns && e.pawnTable.pawnCount[1]; // Pawns on both sides + bool pp = e.hasPawns && e.pawnCount[1]; // Pawns on both sides - assert(!pp || e.pawnTable.pawnCount[0]); + assert(!pp || e.pawnCount[0]); for (File f = FILE_A; f <= maxFile; ++f) { for (int i = 0; i < sides; i++) - item(p, i, f).precomp = new PairsData(); + *e.get(i, f) = PairsData(); int order[][2] = { { *data & 0xF, pp ? *(data + 1) & 0xF : 0xF }, { *data >> 4, pp ? *(data + 1) >> 4 : 0xF } }; @@ -1099,45 +1017,43 @@ void do_init(Entry& e, T& p, uint8_t* data) { for (int k = 0; k < e.pieceCount; ++k, ++data) for (int i = 0; i < sides; i++) - item(p, i, f).precomp->pieces[k] = Piece(i ? *data >> 4 : *data & 0xF); + e.get(i, f)->pieces[k] = Piece(i ? *data >> 4 : *data & 0xF); for (int i = 0; i < sides; ++i) - set_groups(e, item(p, i, f).precomp, order[i], f); + set_groups(e, e.get(i, f), order[i], f); } data += (uintptr_t)data & 1; // Word alignment for (File f = FILE_A; f <= maxFile; ++f) for (int i = 0; i < sides; i++) - data = set_sizes(item(p, i, f).precomp, data); + data = set_sizes(e.get(i, f), data); - if (!IsWDL) - data = set_dtz_map(e, p, data, maxFile); + if (Type == DTZ) + data = set_dtz_map(e, data, maxFile); for (File f = FILE_A; f <= maxFile; ++f) for (int i = 0; i < sides; i++) { - (d = item(p, i, f).precomp)->sparseIndex = (SparseEntry*)data; + (d = e.get(i, f))->sparseIndex = (SparseEntry*)data; data += d->sparseIndexSize * sizeof(SparseEntry); } for (File f = FILE_A; f <= maxFile; ++f) for (int i = 0; i < sides; i++) { - (d = item(p, i, f).precomp)->blockLength = (uint16_t*)data; + (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 - (d = item(p, i, f).precomp)->data = data; + (d = e.get(i, f))->data = data; data += d->blocksNum * d->sizeofBlock; } } -template -void* init(Entry& e, const Position& pos) { - - constexpr bool IsWDL = std::is_same::value; +template +void* init(TBEntry& e, const Position& pos) { static Mutex mutex; @@ -1162,23 +1078,24 @@ void* init(Entry& e, const Position& pos) { { 0x71, 0xE8, 0x23, 0x5D } }; fname = (e.key == pos.material_key() ? w + 'v' + b : b + 'v' + w) - + (IsWDL ? ".rtbw" : ".rtbz"); + + (Type == WDL ? ".rtbw" : ".rtbz"); - uint8_t* data = TBFile(fname).map(&e.baseAddress, &e.mapping, TB_MAGIC[IsWDL]); + uint8_t* data = TBFile(fname).map(&e.baseAddress, &e.mapping, + TB_MAGIC[Type == WDL]); if (data) - e.hasPawns ? do_init(e, e.pawnTable, data) : do_init(e, e.pieceTable, data); + do_init(e, data); e.ready.store(true, std::memory_order_release); return e.baseAddress; } -template::type> +template::Result> T probe_table(const Position& pos, ProbeState* result, WDLScore wdl = WDLDraw) { if (!(pos.pieces() ^ pos.pieces(KING))) return T(WDLDraw); // KvK - E* entry = EntryTable.get(pos.material_key()); + TBEntry* entry = EntryTable.get(pos.material_key()); if (!entry || !init(*entry, pos)) return *result = FAIL, T(); @@ -1247,7 +1164,7 @@ WDLScore search(Position& pos, ProbeState* result) { value = bestValue; else { - value = probe_table(pos, result); + value = probe_table(pos, result); if (*result == FAIL) return WDLDraw; @@ -1447,7 +1364,7 @@ int Tablebases::probe_dtz(Position& pos, ProbeState* result) { if (*result == ZEROING_BEST_MOVE) return dtz_before_zeroing(wdl); - int dtz = probe_table(pos, result, wdl); + int dtz = probe_table(pos, result, wdl); if (*result == FAIL) return 0;