-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathfpnge.cc
1714 lines (1523 loc) · 62.3 KB
/
fpnge.cc
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
// Copyright 2021 Google LLC
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "fpnge.h"
#include <algorithm>
#include <assert.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <vector>
#if defined(_MSC_VER) && !defined(__clang__)
#define FORCE_INLINE_LAMBDA
#define FORCE_INLINE __forceinline
#define __SSE4_1__ 1
#define __PCLMUL__ 1
#ifdef __AVX2__
#define __BMI2__ 1
#endif
#else
#define FORCE_INLINE_LAMBDA __attribute__((always_inline))
#define FORCE_INLINE __attribute__((always_inline)) inline
#endif
#if defined(__x86_64__) || defined(__amd64__) || defined(__LP64) || \
defined(_M_X64) || defined(_M_AMD64) || \
(defined(_WIN64) && !defined(_M_ARM64))
#define PLATFORM_AMD64 1
#endif
#if !defined(FPNGE_USE_PEXT)
#if defined(__BMI2__) && defined(PLATFORM_AMD64) && \
!defined(__tune_znver1__) && !defined(__tune_znver2__) && \
!defined(__tune_bdver4__)
#define FPNGE_USE_PEXT 1
#else
#define FPNGE_USE_PEXT 0
#endif
#endif
#ifdef __AVX2__
#include <immintrin.h>
#define MM(f) _mm256_##f
#define MMSI(f) _mm256_##f##_si256
#define MIVEC __m256i
#define BCAST128 _mm256_broadcastsi128_si256
// workaround for compilers not supporting _mm256_zextsi128_si256
#if (defined(__clang__) && __clang_major__ >= 5 && \
(!defined(__APPLE__) || __clang_major__ >= 7)) || \
(defined(__GNUC__) && __GNUC__ >= 10) || \
(defined(_MSC_VER) && _MSC_VER >= 1910)
#define INT2VEC(v) _mm256_zextsi128_si256(_mm_cvtsi32_si128(v))
#elif defined(__OPTIMIZE__)
// technically incorrect, but should work fine most of the time
#define INT2VEC(v) _mm256_castsi128_si256(_mm_cvtsi32_si128(v))
#else
// _mm256_insert_epi32 is unavailable on MSVC 19.0, so prefer the following
#define INT2VEC(v) \
_mm256_inserti128_si256(_mm256_setzero_si256(), _mm_cvtsi32_si128(v), 0);
#endif
#define SIMD_WIDTH 32
#define SIMD_MASK 0xffffffffU
#elif defined(__SSE4_1__)
#include <nmmintrin.h>
#define MM(f) _mm_##f
#define MMSI(f) _mm_##f##_si128
#define MIVEC __m128i
#define BCAST128(v) (v)
#define INT2VEC _mm_cvtsi32_si128
#define SIMD_WIDTH 16
#define SIMD_MASK 0xffffU
#else
#error Requires SSE4.1 support minium
#endif
namespace {
alignas(16) constexpr uint8_t kBitReverseNibbleLookup[16] = {
0b0000, 0b1000, 0b0100, 0b1100, 0b0010, 0b1010, 0b0110, 0b1110,
0b0001, 0b1001, 0b0101, 0b1101, 0b0011, 0b1011, 0b0111, 0b1111,
};
static constexpr uint8_t kLZ77NBits[29] = {0, 0, 0, 0, 0, 0, 0, 0, 1, 1,
1, 1, 2, 2, 2, 2, 3, 3, 3, 3,
4, 4, 4, 4, 5, 5, 5, 5, 0};
static constexpr uint16_t kLZ77Base[29] = {
3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 23, 27,
31, 35, 43, 51, 59, 67, 83, 99, 115, 131, 163, 195, 227, 258};
static uint16_t BitReverse(size_t nbits, uint16_t bits) {
uint16_t rev16 = (kBitReverseNibbleLookup[bits & 0xF] << 12) |
(kBitReverseNibbleLookup[(bits >> 4) & 0xF] << 8) |
(kBitReverseNibbleLookup[(bits >> 8) & 0xF] << 4) |
(kBitReverseNibbleLookup[bits >> 12]);
return rev16 >> (16 - nbits);
}
struct HuffmanTable {
uint8_t nbits[286];
uint16_t end_bits;
alignas(16) uint8_t approx_nbits[16];
alignas(16) uint8_t first16_nbits[16];
alignas(16) uint8_t first16_bits[16];
alignas(16) uint8_t last16_nbits[16];
alignas(16) uint8_t last16_bits[16];
alignas(16) uint8_t mid_lowbits[16];
uint8_t mid_nbits;
uint32_t lz77_length_nbits[259] = {};
uint32_t lz77_length_bits[259] = {};
uint32_t lz77_length_sym[259] = {};
uint32_t dist_nbits, dist_bits;
// Computes nbits[i] for i <= n, subject to min_limit[i] <= nbits[i] <=
// max_limit[i], so to minimize sum(nbits[i] * freqs[i]).
static void ComputeCodeLengths(const uint64_t *freqs, size_t n,
uint8_t *min_limit, uint8_t *max_limit,
uint8_t *nbits) {
size_t precision = 0;
uint64_t freqsum = 0;
for (size_t i = 0; i < n; i++) {
assert(freqs[i] != 0);
freqsum += freqs[i];
if (min_limit[i] < 1)
min_limit[i] = 1;
assert(min_limit[i] <= max_limit[i]);
precision = std::max<size_t>(max_limit[i], precision);
}
uint64_t infty = freqsum * precision;
std::vector<uint64_t> dynp(((1U << precision) + 1) * (n + 1), infty);
auto d = [&](size_t sym, size_t off) -> uint64_t & {
return dynp[sym * ((1 << precision) + 1) + off];
};
d(0, 0) = 0;
for (size_t sym = 0; sym < n; sym++) {
for (size_t bits = min_limit[sym]; bits <= max_limit[sym]; bits++) {
size_t off_delta = 1U << (precision - bits);
for (size_t off = 0; off + off_delta <= (1U << precision); off++) {
d(sym + 1, off + off_delta) = std::min(
d(sym, off) + freqs[sym] * bits, d(sym + 1, off + off_delta));
}
}
}
size_t sym = n;
size_t off = 1U << precision;
while (sym-- > 0) {
assert(off > 0);
for (size_t bits = min_limit[sym]; bits <= max_limit[sym]; bits++) {
size_t off_delta = 1U << (precision - bits);
if (off_delta <= off &&
d(sym + 1, off) == d(sym, off - off_delta) + freqs[sym] * bits) {
off -= off_delta;
nbits[sym] = bits;
break;
}
}
}
}
void ComputeNBits(const uint64_t *collected_data) {
constexpr uint64_t kBaselineData[286] = {
113, 54, 28, 18, 12, 9, 7, 6, 5, 4, 3, 3, 2, 2, 2, 2, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 3, 3, 4, 5, 6, 7, 9,
12, 18, 29, 54, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
};
uint64_t data[286];
for (size_t i = 0; i < 286; i++) {
data[i] = collected_data[i] + kBaselineData[i];
}
// Compute Huffman code length ensuring that all the "fake" symbols for [16,
// 240) and [255, 285) have their maximum length.
uint64_t collapsed_data[16 + 14 + 16 + 2] = {};
uint8_t collapsed_min_limit[16 + 14 + 16 + 2] = {};
uint8_t collapsed_max_limit[16 + 14 + 16 + 2];
for (size_t i = 0; i < 48; i++) {
collapsed_max_limit[i] = 8;
}
for (size_t i = 0; i < 16; i++) {
collapsed_data[i] = data[i];
}
for (size_t i = 0; i < 14; i++) {
collapsed_data[16 + i] = 1;
collapsed_min_limit[16 + i] = 8;
}
for (size_t j = 0; j < 16; j++) {
collapsed_data[16 + 14 + j] += data[240 + j];
}
collapsed_data[16 + 14 + 16] = 1;
collapsed_min_limit[16 + 14 + 16] = 8;
collapsed_data[16 + 14 + 16 + 1] = data[285];
uint8_t collapsed_nbits[48] = {};
ComputeCodeLengths(collapsed_data, 48, collapsed_min_limit,
collapsed_max_limit, collapsed_nbits);
// Compute "extra" code lengths for symbols >= 256, except 285.
uint8_t tail_nbits[29] = {};
uint8_t tail_min_limit[29] = {};
uint8_t tail_max_limit[29] = {};
for (size_t i = 0; i < 29; i++) {
tail_min_limit[i] = 4;
tail_max_limit[i] = 7;
}
ComputeCodeLengths(data + 256, 29, tail_min_limit, tail_max_limit,
tail_nbits);
for (size_t i = 0; i < 16; i++) {
nbits[i] = collapsed_nbits[i];
}
for (size_t i = 0; i < 14; i++) {
for (size_t j = 0; j < 16; j++) {
nbits[(i + 1) * 16 + j] = collapsed_nbits[16 + i] + 4;
}
}
for (size_t i = 0; i < 16; i++) {
nbits[240 + i] = collapsed_nbits[30 + i];
}
for (size_t i = 0; i < 29; i++) {
nbits[256 + i] = collapsed_nbits[46] + tail_nbits[i];
}
nbits[285] = collapsed_nbits[47];
}
void ComputeCanonicalCode(const uint8_t *nbits, uint16_t *bits) {
uint8_t code_length_counts[16] = {};
for (size_t i = 0; i < 286; i++) {
code_length_counts[nbits[i]]++;
}
uint16_t next_code[16] = {};
uint16_t code = 0;
for (size_t i = 1; i < 16; i++) {
code = (code + code_length_counts[i - 1]) << 1;
next_code[i] = code;
}
for (size_t i = 0; i < 286; i++) {
bits[i] = BitReverse(nbits[i], next_code[nbits[i]]++);
}
}
void FillNBits() {
for (size_t i = 0; i < 16; i++) {
first16_nbits[i] = nbits[i];
last16_nbits[i] = nbits[240 + i];
}
mid_nbits = nbits[16];
for (size_t i = 16; i < 240; i++) {
assert(nbits[i] == mid_nbits);
}
// Construct lz77 lookup tables.
for (size_t i = 0; i < 29; i++) {
for (size_t j = 0; j < (1U << kLZ77NBits[i]); j++) {
lz77_length_nbits[kLZ77Base[i] + j] = nbits[257 + i] + kLZ77NBits[i];
lz77_length_sym[kLZ77Base[i] + j] = 257 + i;
}
}
dist_nbits = 1;
approx_nbits[0] =
nbits[0] - 1; // subtract 1 as a fudge for catering for RLE
for (size_t i = 1; i < 15; i++) {
approx_nbits[i] = (nbits[i] + nbits[256 - i] + 1) / 2;
}
approx_nbits[15] = mid_nbits;
}
void FillBits() {
uint16_t bits[286];
ComputeCanonicalCode(nbits, bits);
for (size_t i = 0; i < 16; i++) {
first16_bits[i] = bits[i];
last16_bits[i] = bits[240 + i];
}
mid_lowbits[0] = mid_lowbits[15] = 0;
for (size_t i = 16; i < 240; i += 16) {
mid_lowbits[i / 16] = bits[i] & ((1U << (mid_nbits - 4)) - 1);
}
for (size_t i = 16; i < 240; i++) {
assert((uint32_t(mid_lowbits[i / 16]) |
(kBitReverseNibbleLookup[i % 16] << (mid_nbits - 4))) == bits[i]);
}
end_bits = bits[256];
// Construct lz77 lookup tables.
for (size_t i = 0; i < 29; i++) {
for (size_t j = 0; j < (1U << kLZ77NBits[i]); j++) {
lz77_length_bits[kLZ77Base[i] + j] =
bits[257 + i] | (j << nbits[257 + i]);
}
}
dist_bits = 0;
}
HuffmanTable(const uint64_t *collected_data) {
ComputeNBits(collected_data);
FillNBits();
FillBits();
}
// estimate for CollectSymbolCounts
// only fills nbits; skips computing actual codes
HuffmanTable() {
// the following is similar to ComputeNBits(0, 0, 0 ...), but much faster
constexpr uint8_t collapsed_nbits[] = {
2, 3, 4, 5, 5, 6, 6, 6, 7, 7, 7, 7, 8, 8, 8, 8,
8, 8, 8, 8, 8, 7, 7, 7, 7, 6, 6, 6, 5, 5, 4, 3,
13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 13,
13, 13, 13, 13, 13, 13, 13, 13, 13, 13, 12, 12, 12, 8};
for (size_t i = 0; i < 16; i++) {
nbits[i] = collapsed_nbits[i];
nbits[240 + i] = collapsed_nbits[16 + i];
}
for (size_t i = 16; i < 240; i++) {
nbits[i] = 12;
}
for (size_t i = 0; i < 30; i++) {
nbits[256 + i] = collapsed_nbits[32 + i];
}
FillNBits();
}
};
struct BitWriter {
void Write(uint32_t count, uint64_t bits) {
buffer |= bits << bits_in_buffer;
bits_in_buffer += count;
memcpy(data + bytes_written, &buffer, 8);
size_t bytes_in_buffer = bits_in_buffer / 8;
bits_in_buffer &= 7;
buffer >>= bytes_in_buffer * 8;
bytes_written += bytes_in_buffer;
}
void ZeroPadToByte() {
if (bits_in_buffer != 0) {
Write(8 - bits_in_buffer, 0);
}
}
void WriteBytes(const char *data, size_t count) {
memcpy(this->data + bytes_written, data, count);
bytes_written += count;
}
unsigned char *data;
size_t bytes_written = 0;
size_t bits_in_buffer = 0;
uint64_t buffer = 0;
};
static void WriteHuffmanCode(const HuffmanTable &table,
BitWriter *__restrict writer) {
constexpr uint8_t kCodeLengthNbits[] = {
4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 0, 0, 0,
};
constexpr uint8_t kCodeLengthOrder[] = {
16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15,
};
writer->Write(5, 29); // all lit/len codes
writer->Write(5, 0); // distance code up to dist, included
writer->Write(4, 15); // all code length codes
for (size_t i = 0; i < 19; i++) {
writer->Write(3, kCodeLengthNbits[kCodeLengthOrder[i]]);
}
for (size_t i = 0; i < 286; i++) {
writer->Write(4, kBitReverseNibbleLookup[table.nbits[i]]);
}
writer->Write(4, 0b1000);
}
#ifdef __PCLMUL__
} // namespace
#include <wmmintrin.h>
namespace {
alignas(32) static const uint8_t pshufb_shf_table[] = {
0x80, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87, 0x88, 0x89, 0x8a,
0x8b, 0x8c, 0x8d, 0x8e, 0x8f, 0x00, 0x01, 0x02, 0x03, 0x04, 0x05,
0x06, 0x07, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f};
class Crc32 {
__m128i x0, x1, x2, x3;
static inline __m128i double_xor(__m128i a, __m128i b, __m128i c) {
#ifdef __AVX512VL__
return _mm_ternarylogic_epi32(a, b, c, 0x96);
#else
return _mm_xor_si128(_mm_xor_si128(a, b), c);
#endif
}
static inline __m128i do_one_fold(__m128i src, __m128i data) {
const auto k1k2 = _mm_set_epi32(1, 0x54442bd4, 1, 0xc6e41596);
return double_xor(_mm_clmulepi64_si128(src, k1k2, 0x01),
_mm_clmulepi64_si128(src, k1k2, 0x10), data);
}
public:
Crc32() {
x0 = _mm_cvtsi32_si128(0x9db42487);
x1 = _mm_setzero_si128();
x2 = _mm_setzero_si128();
x3 = _mm_setzero_si128();
}
size_t update(const unsigned char *__restrict data, size_t len) {
auto amount = len & ~63;
for (size_t i = 0; i < amount; i += 64) {
x0 = do_one_fold(x0, _mm_loadu_si128((__m128i *)(data + i)));
x1 = do_one_fold(x1, _mm_loadu_si128((__m128i *)(data + i + 0x10)));
x2 = do_one_fold(x2, _mm_loadu_si128((__m128i *)(data + i + 0x20)));
x3 = do_one_fold(x3, _mm_loadu_si128((__m128i *)(data + i + 0x30)));
}
return amount;
}
uint32_t update_final(const unsigned char *__restrict data, size_t len) {
if (len >= 64) {
update(data, len);
data += len & ~63;
len &= 63;
}
if (len >= 48) {
auto t3 = x3;
x3 = do_one_fold(x2, _mm_loadu_si128((__m128i *)data + 2));
x2 = do_one_fold(x1, _mm_loadu_si128((__m128i *)data + 1));
x1 = do_one_fold(x0, _mm_loadu_si128((__m128i *)data));
x0 = t3;
} else if (len >= 32) {
auto t2 = x2;
auto t3 = x3;
x3 = do_one_fold(x1, _mm_loadu_si128((__m128i *)data + 1));
x2 = do_one_fold(x0, _mm_loadu_si128((__m128i *)data));
x1 = t3;
x0 = t2;
} else if (len >= 16) {
auto t3 = x3;
x3 = do_one_fold(x0, _mm_loadu_si128((__m128i *)data));
x0 = x1;
x1 = x2;
x2 = t3;
}
data += len & 48;
len &= 15;
if (len > 0) {
auto xmm_shl = _mm_loadu_si128((__m128i *)(pshufb_shf_table + len));
auto xmm_shr = _mm_xor_si128(xmm_shl, _mm_set1_epi8(-128));
auto t0 = _mm_loadu_si128((__m128i *)data);
auto t1 = _mm_shuffle_epi8(x0, xmm_shl);
x0 = _mm_or_si128(_mm_shuffle_epi8(x0, xmm_shr),
_mm_shuffle_epi8(x1, xmm_shl));
x1 = _mm_or_si128(_mm_shuffle_epi8(x1, xmm_shr),
_mm_shuffle_epi8(x2, xmm_shl));
x2 = _mm_or_si128(_mm_shuffle_epi8(x2, xmm_shr),
_mm_shuffle_epi8(x3, xmm_shl));
x3 = _mm_or_si128(_mm_shuffle_epi8(x3, xmm_shr),
_mm_shuffle_epi8(t0, xmm_shl));
x3 = do_one_fold(t1, x3);
}
const auto k3k4 = _mm_set_epi32(1, 0x751997d0, 0, 0xccaa009e);
const auto k5k4 = _mm_set_epi32(1, 0x63cd6124, 0, 0xccaa009e);
const auto poly = _mm_set_epi32(1, 0xdb710640, 0, 0xf7011641);
x0 = double_xor(x1, _mm_clmulepi64_si128(x0, k3k4, 0x10),
_mm_clmulepi64_si128(x0, k3k4, 0x01));
x0 = double_xor(x2, _mm_clmulepi64_si128(x0, k3k4, 0x10),
_mm_clmulepi64_si128(x0, k3k4, 0x01));
x0 = double_xor(x3, _mm_clmulepi64_si128(x0, k3k4, 0x10),
_mm_clmulepi64_si128(x0, k3k4, 0x01));
x1 =
_mm_xor_si128(_mm_clmulepi64_si128(x0, k5k4, 0), _mm_srli_si128(x0, 8));
x0 = _mm_slli_si128(x1, 4);
x0 = _mm_clmulepi64_si128(x0, k5k4, 0x10);
#ifdef __AVX512VL__
x0 = _mm_ternarylogic_epi32(x0, x1, _mm_set_epi32(0, -1, -1, 0), 0x28);
#else
x1 = _mm_and_si128(x1, _mm_set_epi32(0, -1, -1, 0));
x0 = _mm_xor_si128(x0, x1);
#endif
x1 = _mm_clmulepi64_si128(x0, poly, 0);
x1 = _mm_clmulepi64_si128(x1, poly, 0x10);
#ifdef __AVX512VL__
x1 = _mm_ternarylogic_epi32(x1, x0, x0, 0xC3); // NOT(XOR(x1, x0))
#else
x0 = _mm_xor_si128(x0, _mm_set_epi32(0, -1, -1, 0));
x1 = _mm_xor_si128(x1, x0);
#endif
return _mm_extract_epi32(x1, 2);
}
};
#else
} // namespace
#include <array>
#include <cstddef>
#include <utility>
namespace {
// from https://joelfilho.com/blog/2020/compile_time_lookup_tables_in_cpp/
template <std::size_t Length, typename Generator, std::size_t... Indexes>
constexpr auto lut_impl(Generator &&f, std::index_sequence<Indexes...>) {
using content_type = decltype(f(std::size_t{0}));
return std::array<content_type, Length>{{f(Indexes)...}};
}
template <std::size_t Length, typename Generator>
constexpr auto lut(Generator &&f) {
return lut_impl<Length>(std::forward<Generator>(f),
std::make_index_sequence<Length>{});
}
constexpr uint32_t crc32_slice8_gen(unsigned n) {
uint32_t crc = n & 0xff;
for (int i = n >> 8; i >= 0; i--) {
for (int j = 0; j < 8; j++) {
crc = (crc >> 1) ^
((crc & 1) * 0xEDB88320); // 0xEDB88320 = CRC32 polynomial
}
}
return crc;
}
static constexpr auto kCrcSlice8LUT = lut<256 * 8>(crc32_slice8_gen);
class Crc32 {
uint32_t state;
// this is based off Fast CRC32 slice-by-8:
// https://create.stephan-brumme.com/crc32/
static inline uint32_t crc_process_iter(uint32_t crc,
const uint32_t *current) {
uint32_t one = *current++ ^ crc;
uint32_t two = *current;
return kCrcSlice8LUT[(two >> 24) & 0xFF] ^
kCrcSlice8LUT[0x100 + ((two >> 16) & 0xFF)] ^
kCrcSlice8LUT[0x200 + ((two >> 8) & 0xFF)] ^
kCrcSlice8LUT[0x300 + (two & 0xFF)] ^
kCrcSlice8LUT[0x400 + ((one >> 24) & 0xFF)] ^
kCrcSlice8LUT[0x500 + ((one >> 16) & 0xFF)] ^
kCrcSlice8LUT[0x600 + ((one >> 8) & 0xFF)] ^
kCrcSlice8LUT[0x700 + (one & 0xFF)];
}
public:
Crc32() : state(0xffffffff) {}
size_t update(const unsigned char *__restrict data, size_t len) {
auto amount = len & ~7;
for (size_t i = 0; i < amount; i += 8) {
state = crc_process_iter(state, (uint32_t *)(data + i));
}
return amount;
}
uint32_t update_final(const unsigned char *__restrict data, size_t len) {
auto i = update(data, len);
for (; i < len; i++) {
state = (state >> 8) ^ kCrcSlice8LUT[(state & 0xFF) ^ data[i]];
}
return ~state;
}
};
#endif
constexpr unsigned kAdler32Mod = 65521;
static void UpdateAdler32(uint32_t &s1, uint32_t &s2, uint8_t byte) {
s1 += byte;
s2 += s1;
s1 %= kAdler32Mod;
s2 %= kAdler32Mod;
}
static uint32_t hadd(MIVEC v) {
auto sum =
#ifdef __AVX2__
_mm_add_epi32(_mm256_castsi256_si128(v), _mm256_extracti128_si256(v, 1));
#else
v;
#endif
sum = _mm_hadd_epi32(sum, sum);
sum = _mm_hadd_epi32(sum, sum);
return _mm_cvtsi128_si32(sum);
}
template <size_t predictor>
static FORCE_INLINE MIVEC PredictVec(const unsigned char *current_buf,
const unsigned char *top_buf,
const unsigned char *left_buf,
const unsigned char *topleft_buf) {
auto data = MMSI(load)((MIVEC *)(current_buf));
if (predictor == 0) {
return data;
} else if (predictor == 1) {
auto pred = MMSI(loadu)((MIVEC *)(left_buf));
return MM(sub_epi8)(data, pred);
} else if (predictor == 2) {
auto pred = MMSI(load)((MIVEC *)(top_buf));
return MM(sub_epi8)(data, pred);
} else if (predictor == 3) {
auto left = MMSI(loadu)((MIVEC *)(left_buf));
auto top = MMSI(load)((MIVEC *)(top_buf));
auto pred = MM(sub_epi8)(MM(add_epi8)(top, left), MM(avg_epu8)(top, left));
return MM(sub_epi8)(data, pred);
} else {
auto left = MMSI(loadu)((MIVEC *)(left_buf));
auto top = MMSI(load)((MIVEC *)(top_buf));
auto c = MMSI(loadu)((MIVEC *)(topleft_buf));
auto a = MM(min_epu8)(left, top);
auto b = MM(max_epu8)(left, top);
auto pa = MM(subs_epu8)(b, c);
auto pb = MM(subs_epu8)(c, a);
auto min_pab = MM(min_epu8)(pa, pb);
auto pc = MM(sub_epi8)(MM(max_epu8)(pa, pb), min_pab);
auto min_pabc = MM(min_epu8)(min_pab, pc);
auto use_a = MM(cmpeq_epi8)(min_pabc, pa);
auto use_b = MM(cmpeq_epi8)(min_pabc, pb);
auto pred = MM(blendv_epi8)(MM(blendv_epi8)(c, b, use_b), a, use_a);
return MM(sub_epi8)(data, pred);
}
}
alignas(SIMD_WIDTH) constexpr int32_t _kMaskVec[] = {0, 0, 0, 0,
#if SIMD_WIDTH == 32
0, 0, 0, 0,
-1, -1, -1, -1,
#endif
-1, -1, -1, -1};
static const uint8_t *kMaskVec =
reinterpret_cast<const uint8_t *>(_kMaskVec) + SIMD_WIDTH;
template <size_t predictor, typename CB, typename CB_ADL, typename CB_RLE>
static void
ProcessRow(size_t bytes_per_line, const unsigned char *current_row_buf,
const unsigned char *top_buf, const unsigned char *left_buf,
const unsigned char *topleft_buf, CB &&cb, CB_ADL &&cb_adl,
CB_RLE &&cb_rle) {
size_t run = 0;
size_t i = 0;
for (; i + SIMD_WIDTH <= bytes_per_line; i += SIMD_WIDTH) {
auto pdata = PredictVec<predictor>(current_row_buf + i, top_buf + i,
left_buf + i, topleft_buf + i);
unsigned pdatais0 =
MM(movemask_epi8)(MM(cmpeq_epi8)(pdata, MMSI(setzero)()));
if (pdatais0 == SIMD_MASK) {
run += SIMD_WIDTH;
} else {
if (run != 0) {
cb_rle(run);
}
run = 0;
cb(pdata, SIMD_WIDTH);
}
cb_adl(pdata, SIMD_WIDTH, i);
}
size_t bytes_remaining =
bytes_per_line ^ i; // equivalent to `bytes_per_line - i`
if (bytes_remaining) {
auto pdata = PredictVec<predictor>(current_row_buf + i, top_buf + i,
left_buf + i, topleft_buf + i);
unsigned pdatais0 =
MM(movemask_epi8)(MM(cmpeq_epi8)(pdata, MMSI(setzero)()));
auto mask = (1UL << bytes_remaining) - 1;
if ((pdatais0 & mask) == mask && run + bytes_remaining >= 16) {
run += bytes_remaining;
} else {
if (run != 0) {
cb_rle(run);
}
run = 0;
cb(pdata, bytes_remaining);
}
cb_adl(pdata, bytes_remaining, i);
}
if (run != 0) {
cb_rle(run);
}
}
template <typename CB, typename CB_ADL, typename CB_RLE>
static void
ProcessRow(uint8_t predictor, size_t bytes_per_line,
const unsigned char *current_row_buf, const unsigned char *top_buf,
const unsigned char *left_buf, const unsigned char *topleft_buf,
CB &&cb, CB_ADL &&cb_adl, CB_RLE &&cb_rle) {
if (predictor == 1) {
ProcessRow<1>(bytes_per_line, current_row_buf, top_buf, left_buf,
topleft_buf, cb, cb_adl, cb_rle);
} else if (predictor == 2) {
ProcessRow<2>(bytes_per_line, current_row_buf, top_buf, left_buf,
topleft_buf, cb, cb_adl, cb_rle);
} else if (predictor == 3) {
ProcessRow<3>(bytes_per_line, current_row_buf, top_buf, left_buf,
topleft_buf, cb, cb_adl, cb_rle);
} else if (predictor == 4) {
ProcessRow<4>(bytes_per_line, current_row_buf, top_buf, left_buf,
topleft_buf, cb, cb_adl, cb_rle);
} else {
assert(predictor == 0);
ProcessRow<0>(bytes_per_line, current_row_buf, top_buf, left_buf,
topleft_buf, cb, cb_adl, cb_rle);
}
}
template <typename CB> static void ForAllRLESymbols(size_t length, CB &&cb) {
assert(length >= 4);
length -= 1;
if (length < 258) {
// fast path if long sequences are rare in the image
cb(length, 1);
} else {
auto runs = length / 258;
auto remain = length % 258;
if (remain == 1 || remain == 2) {
remain += 258 - 3;
runs--;
cb(3, 1);
}
if (runs) {
cb(258, runs);
}
if (remain) {
cb(remain, 1);
}
}
}
template <size_t pred, bool store_pred>
static void
TryPredictor(size_t bytes_per_line, const unsigned char *current_row_buf,
const unsigned char *top_buf, const unsigned char *left_buf,
const unsigned char *topleft_buf, unsigned char *predicted_data,
const HuffmanTable &table, size_t &best_cost, uint8_t &predictor) {
size_t cost_rle = 0;
MIVEC cost_direct = MMSI(setzero)();
auto cost_chunk_cb = [&](const MIVEC bytes,
const size_t bytes_in_vec) FORCE_INLINE_LAMBDA {
auto data_for_lut = MMSI(and)(MM(set1_epi8)(0xF), bytes);
// get a mask of `bytes` that are between -16 and 15 inclusive
// (`-16 <= bytes <= 15` is equivalent to `bytes + 112 > 95`)
auto use_lowhi = MM(cmpgt_epi8)(MM(add_epi8)(bytes, MM(set1_epi8)(112)),
MM(set1_epi8)(95));
auto nbits_low16 = MM(shuffle_epi8)(
BCAST128(_mm_load_si128((__m128i *)table.first16_nbits)), data_for_lut);
auto nbits_hi16 = MM(shuffle_epi8)(
BCAST128(_mm_load_si128((__m128i *)table.last16_nbits)), data_for_lut);
auto nbits = MM(blendv_epi8)(nbits_low16, nbits_hi16, bytes);
nbits = MM(blendv_epi8)(MM(set1_epi8)(table.mid_nbits), nbits, use_lowhi);
auto nbits_discard =
MMSI(and)(nbits, MMSI(loadu)((MIVEC *)(kMaskVec - bytes_in_vec)));
cost_direct =
MM(add_epi32)(cost_direct, MM(sad_epu8)(nbits, nbits_discard));
};
auto rle_cost_cb = [&](size_t run) {
cost_rle += table.first16_nbits[0];
ForAllRLESymbols(run, [&](size_t len, size_t count) {
cost_rle += (table.dist_nbits + table.lz77_length_nbits[len]) * count;
});
};
if (store_pred) {
ProcessRow<pred>(
bytes_per_line, current_row_buf, top_buf, left_buf, topleft_buf,
cost_chunk_cb,
[=](const MIVEC pdata, size_t, size_t i) {
MMSI(store)((MIVEC *)(predicted_data + i), pdata);
},
rle_cost_cb);
} else {
ProcessRow<pred>(
bytes_per_line, current_row_buf, top_buf, left_buf, topleft_buf,
cost_chunk_cb, [](const MIVEC, size_t, size_t) {}, rle_cost_cb);
}
size_t cost = cost_rle + hadd(cost_direct);
if (cost < best_cost) {
best_cost = cost;
predictor = pred;
}
}
static FORCE_INLINE void WriteBitsLong(MIVEC nbits, MIVEC bits_lo,
MIVEC bits_hi, size_t mid_lo_nbits,
BitWriter *__restrict writer) {
// Merge bits_lo and bits_hi in 16-bit "bits".
#if FPNGE_USE_PEXT
auto bits0 = MM(unpacklo_epi8)(bits_lo, bits_hi);
auto bits1 = MM(unpackhi_epi8)(bits_lo, bits_hi);
// convert nbits into a mask
auto nbits_hi = MM(sub_epi8)(nbits, MM(set1_epi8)(mid_lo_nbits));
auto nbits0 = MM(unpacklo_epi8)(nbits, nbits_hi);
auto nbits1 = MM(unpackhi_epi8)(nbits, nbits_hi);
const auto nbits_to_mask =
BCAST128(_mm_set_epi32(0xffffffff, 0xffffffff, 0x7f3f1f0f, 0x07030100));
auto bitmask0 = MM(shuffle_epi8)(nbits_to_mask, nbits0);
auto bitmask1 = MM(shuffle_epi8)(nbits_to_mask, nbits1);
// aggregate nbits
alignas(16) uint16_t nbits_a[SIMD_WIDTH / 4];
auto bit_count = MM(maddubs_epi16)(nbits, MM(set1_epi8)(1));
#ifdef __AVX2__
auto bit_count2 = _mm_hadd_epi16(_mm256_castsi256_si128(bit_count),
_mm256_extracti128_si256(bit_count, 1));
_mm_store_si128((__m128i *)nbits_a, bit_count2);
#else
bit_count = _mm_hadd_epi16(bit_count, bit_count);
_mm_storel_epi64((__m128i *)nbits_a, bit_count);
#endif
alignas(SIMD_WIDTH) uint64_t bitmask_a[SIMD_WIDTH / 4];
MMSI(store)((MIVEC *)bitmask_a, bitmask0);
MMSI(store)((MIVEC *)bitmask_a + 1, bitmask1);
#else
auto nbits0 = MM(unpacklo_epi8)(nbits, MMSI(setzero)());
auto nbits1 = MM(unpackhi_epi8)(nbits, MMSI(setzero)());
MIVEC bits0, bits1;
if (mid_lo_nbits == 8) {
bits0 = MM(unpacklo_epi8)(bits_lo, bits_hi);
bits1 = MM(unpackhi_epi8)(bits_lo, bits_hi);
} else {
auto nbits_shift = _mm_cvtsi32_si128(8 - mid_lo_nbits);
auto bits_lo_shifted = MM(sll_epi16)(bits_lo, nbits_shift);
bits0 = MM(unpacklo_epi8)(bits_lo_shifted, bits_hi);
bits1 = MM(unpackhi_epi8)(bits_lo_shifted, bits_hi);
bits0 = MM(srl_epi16)(bits0, nbits_shift);
bits1 = MM(srl_epi16)(bits1, nbits_shift);
}
// 16 -> 32
auto nbits0_32_lo = MMSI(and)(nbits0, MM(set1_epi32)(0xFFFF));
auto nbits1_32_lo = MMSI(and)(nbits1, MM(set1_epi32)(0xFFFF));
auto bits0_32_lo = MMSI(and)(bits0, MM(set1_epi32)(0xFFFF));
auto bits1_32_lo = MMSI(and)(bits1, MM(set1_epi32)(0xFFFF));
#ifdef __AVX2__
auto bits0_32_hi = MM(sllv_epi32)(MM(srli_epi32)(bits0, 16), nbits0_32_lo);
auto bits1_32_hi = MM(sllv_epi32)(MM(srli_epi32)(bits1, 16), nbits1_32_lo);
#else
// emulate variable shift by abusing float exponents
// this works because Huffman symbols are not allowed to exceed 15 bits, so
// will fit within a float's mantissa and (number << 15) won't overflow when
// converted back to a signed int
auto bits0_32_hi =
_mm_castps_si128(MM(cvtepi32_ps)(MM(srli_epi32)(bits0, 16)));
auto bits1_32_hi =
_mm_castps_si128(MM(cvtepi32_ps)(MM(srli_epi32)(bits1, 16)));
// add shift amount to the exponent
bits0_32_hi = MM(add_epi32)(bits0_32_hi, MM(slli_epi32)(nbits0_32_lo, 23));
bits1_32_hi = MM(add_epi32)(bits1_32_hi, MM(slli_epi32)(nbits1_32_lo, 23));
bits0_32_hi = MM(cvtps_epi32)(_mm_castsi128_ps(bits0_32_hi));
bits1_32_hi = MM(cvtps_epi32)(_mm_castsi128_ps(bits1_32_hi));
#endif
nbits0 = MM(madd_epi16)(nbits0, MM(set1_epi16)(1));
nbits1 = MM(madd_epi16)(nbits1, MM(set1_epi16)(1));
auto bits0_32 = MMSI(or)(bits0_32_lo, bits0_32_hi);
auto bits1_32 = MMSI(or)(bits1_32_lo, bits1_32_hi);
// 32 -> 64
#ifdef __AVX2__
auto nbits_inv0_64_lo = MM(subs_epu8)(MM(set1_epi64x)(32), nbits0);
auto nbits_inv1_64_lo = MM(subs_epu8)(MM(set1_epi64x)(32), nbits1);
bits0 = MM(sllv_epi32)(bits0_32, nbits_inv0_64_lo);
bits1 = MM(sllv_epi32)(bits1_32, nbits_inv1_64_lo);
bits0 = MM(srlv_epi64)(bits0, nbits_inv0_64_lo);
bits1 = MM(srlv_epi64)(bits1, nbits_inv1_64_lo);
#else
auto nbits0_64_lo = MMSI(and)(nbits0, MM(set1_epi64x)(0xFFFFFFFF));
auto nbits1_64_lo = MMSI(and)(nbits1, MM(set1_epi64x)(0xFFFFFFFF));
// just do two shifts for SSE variant
auto bits0_64_lo = MMSI(and)(bits0_32, MM(set1_epi64x)(0xFFFFFFFF));
auto bits1_64_lo = MMSI(and)(bits1_32, MM(set1_epi64x)(0xFFFFFFFF));
auto bits0_64_hi = MM(srli_epi64)(bits0_32, 32);
auto bits1_64_hi = MM(srli_epi64)(bits1_32, 32);
bits0_64_hi = _mm_blend_epi16(
_mm_sll_epi64(bits0_64_hi, nbits0_64_lo),
_mm_sll_epi64(bits0_64_hi,
_mm_unpackhi_epi64(nbits0_64_lo, nbits0_64_lo)),
0xf0);
bits1_64_hi = _mm_blend_epi16(
_mm_sll_epi64(bits1_64_hi, nbits1_64_lo),
_mm_sll_epi64(bits1_64_hi,
_mm_unpackhi_epi64(nbits1_64_lo, nbits1_64_lo)),
0xf0);
bits0 = MMSI(or)(bits0_64_lo, bits0_64_hi);
bits1 = MMSI(or)(bits1_64_lo, bits1_64_hi);
#endif
auto nbits01 = MM(hadd_epi32)(nbits0, nbits1);
// nbits_a <= 40 as we have at most 10 bits per symbol, so the call to the
// writer is safe.
alignas(SIMD_WIDTH) uint32_t nbits_a[SIMD_WIDTH / 4];
MMSI(store)((MIVEC *)nbits_a, nbits01);
#endif
alignas(SIMD_WIDTH) uint64_t bits_a[SIMD_WIDTH / 4];
MMSI(store)((MIVEC *)bits_a, bits0);
MMSI(store)((MIVEC *)bits_a + 1, bits1);
#ifdef __AVX2__
constexpr uint8_t kPerm[] = {0, 1, 4, 5, 2, 3, 6, 7};
#else
constexpr uint8_t kPerm[] = {0, 1, 2, 3};
#endif
for (size_t ii = 0; ii < SIMD_WIDTH / 4; ii++) {
uint64_t bits = bits_a[kPerm[ii]];
#if FPNGE_USE_PEXT
bits = _pext_u64(bits, bitmask_a[kPerm[ii]]);
#endif
auto count = nbits_a[ii];
writer->Write(count, bits);
}
}
// as above, but where nbits <= 8, so we can ignore bits_hi
static FORCE_INLINE void WriteBitsShort(MIVEC nbits, MIVEC bits,
BitWriter *__restrict writer) {
#if FPNGE_USE_PEXT
// convert nbits into a mask
auto bitmask = MM(shuffle_epi8)(
BCAST128(_mm_set_epi32(0xffffffff, 0xffffffff, 0x7f3f1f0f, 0x07030100)),
nbits);
auto bit_count = MM(sad_epu8)(nbits, MMSI(setzero)());
alignas(SIMD_WIDTH) uint64_t nbits_a[SIMD_WIDTH / 8];
MMSI(store)((MIVEC *)nbits_a, bit_count);
alignas(SIMD_WIDTH) uint64_t bits_a[SIMD_WIDTH / 8];
MMSI(store)((MIVEC *)bits_a, bits);
alignas(SIMD_WIDTH) uint64_t bitmask_a[SIMD_WIDTH / 8];
MMSI(store)((MIVEC *)bitmask_a, bitmask);
#else
// 8 -> 16
auto prod = MM(slli_epi16)(
MM(shuffle_epi8)(BCAST128(_mm_set_epi32(
// since we can't handle 8 bits, we'll under-shift
// it and do an extra shift later on
-1, 0xffffff80, 0x40201008, 0x040201ff)),
nbits),
8);
auto bits_hi =
MM(mulhi_epu16)(MMSI(andnot)(MM(set1_epi16)(0xff), bits), prod);