/
tbprobe.cpp
1740 lines (1406 loc) · 62.9 KB
/
tbprobe.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/*
Stockfish, a UCI chess playing engine derived from Glaurung 2.1
Copyright (C) 2004-2024 The Stockfish developers (see AUTHORS file)
Stockfish is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
Stockfish is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
*/
#include "tbprobe.h"
#include <algorithm>
#include <atomic>
#include <cassert>
#include <cstdint>
#include <cstdlib>
#include <cstring>
#include <deque>
#include <fstream>
#include <initializer_list>
#include <iostream>
#include <mutex>
#include <sstream>
#include <string_view>
#include <sys/stat.h>
#include <type_traits>
#include <utility>
#include <vector>
#include "../bitboard.h"
#include "../misc.h"
#include "../movegen.h"
#include "../position.h"
#include "../search.h"
#include "../types.h"
#include "../ucioption.h"
#ifndef _WIN32
#include <fcntl.h>
#include <sys/mman.h>
#include <unistd.h>
#else
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
#define NOMINMAX // Disable macros min() and max()
#endif
#include <windows.h>
#endif
using namespace Stockfish::Tablebases;
int Stockfish::Tablebases::MaxCardinality;
namespace Stockfish {
namespace {
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
// 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
};
inline WDLScore operator-(WDLScore d) { return WDLScore(-int(d)); }
inline Square operator^(Square s, int i) { return Square(int(s) ^ i); }
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 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};
template<typename T, int Half = sizeof(T) / 2, int End = sizeof(T) - 1>
inline void swap_endian(T& x) {
static_assert(std::is_unsigned_v<T>, "Argument of swap_endian not unsigned");
uint8_t tmp, *c = (uint8_t*) &x;
for (int i = 0; i < Half; ++i)
tmp = c[i], c[i] = c[End - i], c[End - i] = tmp;
}
template<>
inline void swap_endian<uint8_t>(uint8_t&) {}
template<typename T, int LE>
T number(void* addr) {
T v;
if (uintptr_t(addr) & (alignof(T) - 1)) // Unaligned pointer (very rare)
std::memcpy(&v, addr, sizeof(T));
else
v = *((T*) addr);
if (LE != IsLittleEndian)
swap_endian(v);
return v;
}
// DTZ tables don't store valid scores for moves that reset the rule50 counter
// like captures and pawn moves but we can easily recover the correct dtz of the
// previous move if we know the position's WDL score.
int dtz_before_zeroing(WDLScore wdl) {
return wdl == WDLWin ? 1
: wdl == WDLCursedWin ? 101
: wdl == WDLBlessedLoss ? -101
: wdl == WDLLoss ? -1
: 0;
}
// Return the sign of a number (-1, 0, 1)
template<typename T>
int sign_of(T val) {
return (T(0) < val) - (val < T(0));
}
// Numbers in little-endian used by sparseIndex[] to point into blockLength[]
struct SparseEntry {
char block[4]; // Number of block
char offset[2]; // Offset within the block
};
static_assert(sizeof(SparseEntry) == 6, "SparseEntry must be 6 bytes");
using Sym = uint16_t; // Huffman symbol
struct LR {
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 the symbol has length 1,
// then the left-hand symbol is the stored value.
template<Side S>
Sym get() {
return S == Left ? ((lr[1] & 0xF) << 8) | lr[0]
: S == Right ? (lr[2] << 4) | (lr[1] >> 4)
: (assert(false), Sym(-1));
}
};
static_assert(sizeof(LR) == 3, "LR tree entry must be 3 bytes");
// Tablebases data layout is structured as following:
//
// TBFile: memory maps/unmaps the physical .rtbw and .rtbz files
// TBTable: one object for each file with corresponding indexing information
// TBTables: has ownership of TBTable objects, keeping a list and a hash
// class TBFile memory maps/unmaps the single .rtbw and .rtbz files. Files are
// memory mapped for best performance. Files are mapped at first access: at init
// time only existence of the file is checked.
class TBFile: public std::ifstream {
std::string fname;
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.
//
// Example:
// C:\tb\wdl345;C:\tb\wdl6;D:\tb\dtz345;D:\tb\dtz6
static std::string Paths;
TBFile(const std::string& f) {
#ifndef _WIN32
constexpr char SepChar = ':';
#else
constexpr char SepChar = ';';
#endif
std::stringstream ss(Paths);
std::string path;
while (std::getline(ss, path, SepChar))
{
fname = path + "/" + f;
std::ifstream::open(fname);
if (is_open())
return;
}
}
// Memory map the file and check it.
uint8_t* map(void** baseAddress, uint64_t* mapping, TBType type) {
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);
if (fd == -1)
return *baseAddress = nullptr, nullptr;
fstat(fd, &statbuf);
if (statbuf.st_size % 64 != 16)
{
std::cerr << "Corrupt tablebase file " << fname << std::endl;
exit(EXIT_FAILURE);
}
*mapping = statbuf.st_size;
*baseAddress = mmap(nullptr, statbuf.st_size, PROT_READ, MAP_SHARED, fd, 0);
#if defined(MADV_RANDOM)
madvise(*baseAddress, statbuf.st_size, MADV_RANDOM);
#endif
::close(fd);
if (*baseAddress == MAP_FAILED)
{
std::cerr << "Could not mmap() " << fname << std::endl;
exit(EXIT_FAILURE);
}
#else
// Note FILE_FLAG_RANDOM_ACCESS is only a hint to Windows and as such may get ignored.
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;
DWORD size_high;
DWORD size_low = GetFileSize(fd, &size_high);
if (size_low % 64 != 16)
{
std::cerr << "Corrupt tablebase file " << fname << std::endl;
exit(EXIT_FAILURE);
}
HANDLE mmap = CreateFileMapping(fd, nullptr, PAGE_READONLY, size_high, size_low, nullptr);
CloseHandle(fd);
if (!mmap)
{
std::cerr << "CreateFileMapping() failed" << std::endl;
exit(EXIT_FAILURE);
}
*mapping = uint64_t(mmap);
*baseAddress = MapViewOfFile(mmap, FILE_MAP_READ, 0, 0, 0);
if (!*baseAddress)
{
std::cerr << "MapViewOfFile() failed, name = " << fname
<< ", error = " << GetLastError() << std::endl;
exit(EXIT_FAILURE);
}
#endif
uint8_t* data = (uint8_t*) *baseAddress;
constexpr uint8_t Magics[][4] = {{0xD7, 0x66, 0x0C, 0xA5}, {0x71, 0xE8, 0x23, 0x5D}};
if (memcmp(data, Magics[type == WDL], 4))
{
std::cerr << "Corrupted table in file " << fname << std::endl;
unmap(*baseAddress, *mapping);
return *baseAddress = nullptr, nullptr;
}
return data + 4; // Skip Magics's header
}
static void unmap(void* baseAddress, uint64_t mapping) {
#ifndef _WIN32
munmap(baseAddress, mapping);
#else
UnmapViewOfFile(baseAddress);
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 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<uint64_t>
base64; // base64[l - min_sym_len] is the 64bit-padded lowest symbol of length l
std::vector<uint8_t>
symlen; // Number of values (-1) represented by a given Huffman symbol: 1..256
Piece pieces[TBPIECES]; // Position pieces: the order of pieces defines the groups
uint64_t groupIdx[TBPIECES + 1]; // Start index used for the encoding of the group's pieces
int groupLen[TBPIECES + 1]; // Number of pieces in a given group: KRKN -> (3, 1)
uint16_t map_idx[4]; // WDLWin, WDLLoss, WDLCursedWin, WDLBlessedLoss (used in DTZ)
};
// struct TBTable contains indexing information to access the corresponding TBFile.
// There are 2 types of TBTable, corresponding to a WDL or a DTZ file. TBTable
// is populated at init time but the nested PairsData records are populated at
// first access, when the corresponding file is memory mapped.
template<TBType Type>
struct TBTable {
using Ret = std::conditional_t<Type == WDL, WDLScore, int>;
static constexpr int Sides = Type == WDL ? 2 : 1;
std::atomic_bool ready;
void* baseAddress;
uint8_t* map;
uint64_t mapping;
Key key;
Key key2;
int pieceCount;
bool hasPawns;
bool hasUniquePieces;
uint8_t pawnCount[2]; // [Lead color / other color]
PairsData items[Sides][4]; // [wtm / btm][FILE_A..FILE_D or 0]
PairsData* get(int stm, int f) { return &items[stm % Sides][hasPawns ? f : 0]; }
TBTable() :
ready(false),
baseAddress(nullptr) {}
explicit TBTable(const std::string& code);
explicit TBTable(const TBTable<WDL>& wdl);
~TBTable() {
if (baseAddress)
TBFile::unmap(baseAddress, mapping);
}
};
template<>
TBTable<WDL>::TBTable(const std::string& code) :
TBTable() {
StateInfo st;
Position pos;
key = pos.set(code, WHITE, &st).material_key();
pieceCount = pos.count<ALL_PIECES>();
hasPawns = pos.pieces(PAWN);
hasUniquePieces = false;
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 fewer pawns because this leads to better compression.
bool c = !pos.count<PAWN>(BLACK)
|| (pos.count<PAWN>(WHITE) && pos.count<PAWN>(BLACK) >= pos.count<PAWN>(WHITE));
pawnCount[0] = pos.count<PAWN>(c ? WHITE : BLACK);
pawnCount[1] = pos.count<PAWN>(c ? BLACK : WHITE);
key2 = pos.set(code, BLACK, &st).material_key();
}
template<>
TBTable<DTZ>::TBTable(const TBTable<WDL>& 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;
hasUniquePieces = wdl.hasUniquePieces;
pawnCount[0] = wdl.pawnCount[0];
pawnCount[1] = wdl.pawnCount[1];
}
// class TBTables creates and keeps ownership of the TBTable objects, one for
// each TB file found. It supports a fast, hash-based, table lookup. Populated
// at init time, accessed at probe time.
class TBTables {
struct Entry {
Key key;
TBTable<WDL>* wdl;
TBTable<DTZ>* dtz;
template<TBType Type>
TBTable<Type>* get() const {
return (TBTable<Type>*) (Type == WDL ? (void*) wdl : (void*) dtz);
}
};
static constexpr int Size = 1 << 12; // 4K table, indexed by key's 12 lsb
static constexpr int Overflow = 1; // Number of elements allowed to map to the last bucket
Entry hashTable[Size + Overflow];
std::deque<TBTable<WDL>> wdlTable;
std::deque<TBTable<DTZ>> dtzTable;
void insert(Key key, TBTable<WDL>* wdl, TBTable<DTZ>* 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)
{
Key otherKey = hashTable[bucket].key;
if (otherKey == key || !hashTable[bucket].get<WDL>())
{
hashTable[bucket] = entry;
return;
}
// Robin Hood hashing: If we've probed for longer than this element,
// insert here and search for a new spot for the other element instead.
uint32_t otherHomeBucket = uint32_t(otherKey) & (Size - 1);
if (otherHomeBucket > homeBucket)
{
std::swap(entry, hashTable[bucket]);
key = otherKey;
homeBucket = otherHomeBucket;
}
}
std::cerr << "TB hash table size too low!" << std::endl;
exit(EXIT_FAILURE);
}
public:
template<TBType Type>
TBTable<Type>* get(Key key) {
for (const Entry* entry = &hashTable[uint32_t(key) & (Size - 1)];; ++entry)
{
if (entry->key == key || !entry->get<Type>())
return entry->get<Type>();
}
}
void clear() {
memset(hashTable, 0, sizeof(hashTable));
wdlTable.clear();
dtzTable.clear();
}
size_t size() const { return wdlTable.size(); }
void add(const std::vector<PieceType>& pieces);
};
TBTables TBTables;
// If the corresponding file exists two new objects TBTable<WDL> and TBTable<DTZ>
// are created and added to the lists and hash table. Called at init time.
void TBTables::add(const std::vector<PieceType>& pieces) {
std::string code;
for (PieceType pt : pieces)
code += PieceToChar[pt];
TBFile file(code.insert(code.find('K', 1), "v") + ".rtbw"); // KRK -> KRvK
if (!file.is_open()) // Only WDL file is checked
return;
file.close();
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().key2, &wdlTable.back(), &dtzTable.back());
}
// TB tables are compressed with canonical Huffman code. The compressed data is divided into
// blocks of size d->sizeofBlock, and each block stores a variable number of symbols.
// Each symbol represents either a WDL or a (remapped) DTZ value, or a pair of other symbols
// (recursively). If you keep expanding the symbols in a block, you end up with up to 65536
// WDL or DTZ values. Each symbol represents up to 256 values and will correspond after
// Huffman coding to at least 1 bit. So a block of 32 bytes corresponds to at most
// 32 x 8 x 256 = 65536 values. This maximum is only reached for tables that consist mostly
// of draws or mostly of wins, but such tables are actually quite common. In principle, the
// blocks in WDL tables are 64 bytes long (and will be aligned on cache lines). But for
// 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 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.
int decompress_pairs(PairsData* d, uint64_t idx) {
// Special case where all table positions store the same value
if (d->flags & TBFlag::SingleValue)
return d->minSymLen;
// First we need to locate the right block that stores the value at index "idx".
// Because each block n stores blockLength[n] + 1 values, the index i of the block
// that contains the value at position idx is:
//
// for (i = -1, sum = 0; sum <= idx; i++)
// sum += blockLength[i + 1] + 1;
//
// This can be slow, so we use SparseIndex[] populated with a set of SparseEntry that
// point to known indices into blockLength[]. Namely SparseIndex[k] is a SparseEntry
// that stores the blockLength[] index and the offset within that block of the value
// with index I(k), where:
//
// I(k) = k * d->span + d->span / 2 (1)
// First step is to get the 'k' of the I(k) nearest to our idx, using definition (1)
uint32_t k = uint32_t(idx / d->span);
// Then we read the corresponding SparseIndex[] entry
uint32_t block = number<uint32_t, LittleEndian>(&d->sparseIndex[k].block);
int offset = number<uint16_t, LittleEndian>(&d->sparseIndex[k].offset);
// Now compute the difference idx - I(k). From the definition of k, we know that
//
// idx = k * d->span + idx % d->span (2)
//
// So from (1) and (2) we can compute idx - I(K):
int diff = idx % d->span - d->span / 2;
// Sum the above to offset to find the offset corresponding to our idx
offset += diff;
// 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;
while (offset > d->blockLength[block])
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));
// 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-bit sequence.
uint64_t buf64 = number<uint64_t, BigEndian>(ptr);
ptr += 2;
int buf64Size = 64;
Sym sym;
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
// can find the symbol length iterating through base64[].
while (buf64 < d->base64[len])
++len;
// All the symbols of a given length are consecutive integers (numerical
// sequence property), so we can compute the offset of our symbol of
// length len, stored at the beginning of buf64.
sym = Sym((buf64 - d->base64[len]) >> (64 - len - d->minSymLen));
// Now add the value of the lowest symbol of length len to get our symbol
sym += number<Sym, LittleEndian>(&d->lowestSym[len]);
// 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
buf64Size -= len;
if (buf64Size <= 32)
{ // Refill the buffer
buf64Size += 32;
buf64 |= uint64_t(number<uint32_t, BigEndian>(ptr++)) << (64 - buf64Size);
}
}
// 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])
{
Sym left = d->btree[sym].get<LR::Left>();
// 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 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
{
offset -= d->symlen[left] + 1;
sym = d->btree[sym].get<LR::Right>();
}
}
return d->btree[sym].get<LR::Left>();
}
bool check_dtz_stm(TBTable<WDL>*, int, File) { return true; }
bool check_dtz_stm(TBTable<DTZ>* entry, int stm, File f) {
auto flags = entry->get(stm, f)->flags;
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 are stored in the TB file and read during map[] init.
WDLScore map_score(TBTable<WDL>*, File, int value, WDLScore) { return WDLScore(value - 2); }
int map_score(TBTable<DTZ>* entry, File f, int value, WDLScore wdl) {
constexpr int WDLMap[] = {1, 3, 0, 2, 0};
auto flags = entry->get(0, f)->flags;
uint8_t* map = entry->map;
uint16_t* idx = entry->get(0, f)->map_idx;
if (flags & TBFlag::Mapped)
{
if (flags & TBFlag::Wide)
value = ((uint16_t*) map)[idx[WDLMap[wdl + 2]] + value];
else
value = map[idx[WDLMap[wdl + 2]] + value];
}
// DTZ tables store distance to zero in number of moves or plies. We
// want to return plies, so we have 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 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<typename T, typename Ret = typename T::Ret>
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;
PairsData* d;
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 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 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);
int flipColor = (symmetricBlackToMove || blackStronger) * 8;
int flipSquares = (symmetricBlackToMove || blackStronger) * 56;
int stm = (symmetricBlackToMove || blackStronger) ^ pos.side_to_move();
// For pawns, TB files store 4 separate tables according if leading pawn is on
// 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)
{
// 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(entry->get(0, 0)->pieces[0] ^ flipColor);
assert(type_of(pc) == PAWN);
leadPawns = b = pos.pieces(color_of(pc), PAWN);
do
squares[size++] = pop_lsb(b) ^ flipSquares;
while (b);
leadPawnsCnt = size;
std::swap(squares[0], *std::max_element(squares, squares + leadPawnsCnt, pawns_comp));
tbFile = File(edge_distance(file_of(squares[0])));
}
// DTZ tables are one-sided, i.e. they store positions only for white to
// move or only for black to move, so check for side to move to be stm,
// early exit otherwise.
if (!check_dtz_stm(entry, stm, tbFile))
return *result = CHANGE_STM, Ret();
// Now we are ready to get all the position pieces (but the lead pawns) and
// directly map them to the correct color and square.
b = pos.pieces() ^ leadPawns;
do
{
Square s = pop_lsb(b);
squares[size] = s ^ flipSquares;
pieces[size++] = Piece(pos.piece_on(s) ^ flipColor);
} while (b);
assert(size >= 2);
d = entry->get(stm, tbFile);
// Then we reorder the pieces to have the same sequence as the one stored
// in pieces[i]: the sequence that ensures the best compression.
for (int i = leadPawnsCnt; i < size - 1; ++i)
for (int j = i + 1; j < size; ++j)
if (d->pieces[i] == pieces[j])
{
std::swap(pieces[i], pieces[j]);
std::swap(squares[i], squares[j]);
break;
}
// Now we map again the squares so that the square of the lead piece is in
// the triangle A1-D1-D4.
if (file_of(squares[0]) > FILE_D)
for (int i = 0; i < size; ++i)
squares[i] = flip_file(squares[i]);
// Encode leading pawns starting with the one with minimum MapPawns[] and
// proceeding in ascending order.
if (entry->hasPawns)
{
idx = LeadPawnIdx[leadPawnsCnt][squares[0]];
std::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
}
// 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)
squares[i] = flip_rank(squares[i]);
// Look for the first piece of the leading group not on the A1-D4 diagonal
// and ensure it is mapped below the diagonal.
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
for (int j = i; j < size; ++j)
squares[j] = Square(((squares[j] >> 3) | (squares[j] << 3)) & 63);
break;
}
// Encode the leading group.
//
// Suppose we have KRvK. Let's say the pieces are on square numbers wK, wR
// and bK (each 0...63). The simplest way to map this position to an index
// is like this:
//
// index = wK * 64 * 64 + wR * 64 + bK;
//
// But this way the TB is going to have 64*64*64 = 262144 positions, with
// lots of positions being equivalent (because they are mirrors of each
// other) and lots of positions being invalid (two pieces on one square,
// adjacent kings, etc.).
// Usually the first step is to take the wK and bK together. There are just
// 462 ways legal and not-mirrored ways to place the wK and bK on the board.
// Once we have placed the wK and bK, there are 62 squares left for the wR
// Mapping its square from 0..63 to available squares 0..61 can be done like:
//
// wR -= (wR > wK) + (wR > bK);
//
// In words: if wR "comes later" than wK, we deduct 1, and the same if wR
// "comes later" than bK. In case of two same pieces like KRRvK we want to
// place the two Rs "together". If we have 62 squares left, we can place two
// 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 (including kings) we encode them
// together.
if (entry->hasUniquePieces)
{
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 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;
// 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;
// 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]];
// 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
// We don't have at least 3 unique pieces, like in KRRvKBB, just map
// the kings.
idx = MapKK[MapA1D1D4[squares[0]]][squares[1]];
encode_remaining:
idx *= d->groupIdx[0];
Square* groupSq = squares + d->groupLen[0];
// Encode remaining pawns and then pieces according to square, in ascending order
bool remainingPawns = entry->hasPawns && entry->pawnCount[1];
while (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 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 adjust = std::count_if(squares, groupSq, f);
n += Binomial[i + 1][groupSq[i] - adjust - 8 * remainingPawns];
}
remainingPawns = false;
idx += n * d->groupIdx[next];
groupSq += d->groupLen[next];
}
// Now that we have the index, decompress the pair and get the score
return map_score(entry, tbFile, decompress_pairs(d, idx), wdl);
}
// Group together pieces that will be encoded together. The general rule is that
// 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[].
//
// As example KRKN -> KRK + N, KNNK -> KK + NN, KPPKP -> P + PP + K + K
//
// The actual grouping depends on the TB generator and can be inferred from the
// sequence of pieces in piece[] array.
template<typename T>
void set_groups(T& e, PairsData* d, int order[], File f) {
int n = 0, firstLen = e.hasPawns ? 0 : e.hasUniquePieces ? 3 : 2;
d->groupLen[n] = 1;
// Number of pieces per group is stored in groupLen[], for instance in KRKN
// the encoder will default on '111', so groupLen[] will be (3, 1).
for (int i = 1; i < e.pieceCount; ++i)
if (--firstLen > 0 || d->pieces[i] == d->pieces[i - 1])
d->groupLen[n]++;
else
d->groupLen[++n] = 1;
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
// in N(g) different ways, then the position encoding will be of the form:
//
// g1 * N(g2) * N(g3) + g2 * N(g3) + g3
//
// 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 -> 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;
for (int k = 0; next < n || k == order[0] || k == order[1]; ++k)
if (k == order[0]) // Leading pawns or pieces
{
d->groupIdx[0] = idx;
idx *= e.hasPawns ? LeadPawnsSize[d->groupLen[0]][f] : e.hasUniquePieces ? 31332 : 462;
}
else if (k == order[1]) // Remaining pawns
{
d->groupIdx[1] = idx;
idx *= Binomial[d->groupLen[1]][48 - d->groupLen[0]];
}
else // Remaining pieces
{
d->groupIdx[next] = idx;
idx *= Binomial[d->groupLen[next]][freeSquares];
freeSquares -= d->groupLen[next++];
}
d->groupIdx[n] = idx;
}
// 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 leaves that represent the symbol value.
uint8_t set_symlen(PairsData* d, Sym s, std::vector<bool>& visited) {
visited[s] = true; // We can set it now because tree is acyclic
Sym sr = d->btree[s].get<LR::Right>();
if (sr == 0xFFF)
return 0;
Sym sl = d->btree[s].get<LR::Left>();
if (!visited[sl])
d->symlen[sl] = set_symlen(d, sl, visited);