summaryrefslogtreecommitdiff
path: root/vendor/github.com/klauspost/compress/fse/compress.go
blob: b69237c9b8f56e86f5c2e49111d86f907300ecb1 (plain)
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
// Copyright 2018 Klaus Post. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Based on work Copyright (c) 2013, Yann Collet, released under BSD License.

package fse

import (
	"errors"
	"fmt"
)

// Compress the input bytes. Input must be < 2GB.
// Provide a Scratch buffer to avoid memory allocations.
// Note that the output is also kept in the scratch buffer.
// If input is too hard to compress, ErrIncompressible is returned.
// If input is a single byte value repeated ErrUseRLE is returned.
func Compress(in []byte, s *Scratch) ([]byte, error) {
	if len(in) <= 1 {
		return nil, ErrIncompressible
	}
	if len(in) > (2<<30)-1 {
		return nil, errors.New("input too big, must be < 2GB")
	}
	s, err := s.prepare(in)
	if err != nil {
		return nil, err
	}

	// Create histogram, if none was provided.
	maxCount := s.maxCount
	if maxCount == 0 {
		maxCount = s.countSimple(in)
	}
	// Reset for next run.
	s.clearCount = true
	s.maxCount = 0
	if maxCount == len(in) {
		// One symbol, use RLE
		return nil, ErrUseRLE
	}
	if maxCount == 1 || maxCount < (len(in)>>7) {
		// Each symbol present maximum once or too well distributed.
		return nil, ErrIncompressible
	}
	s.optimalTableLog()
	err = s.normalizeCount()
	if err != nil {
		return nil, err
	}
	err = s.writeCount()
	if err != nil {
		return nil, err
	}

	if false {
		err = s.validateNorm()
		if err != nil {
			return nil, err
		}
	}

	err = s.buildCTable()
	if err != nil {
		return nil, err
	}
	err = s.compress(in)
	if err != nil {
		return nil, err
	}
	s.Out = s.bw.out
	// Check if we compressed.
	if len(s.Out) >= len(in) {
		return nil, ErrIncompressible
	}
	return s.Out, nil
}

// cState contains the compression state of a stream.
type cState struct {
	bw         *bitWriter
	stateTable []uint16
	state      uint16
}

// init will initialize the compression state to the first symbol of the stream.
func (c *cState) init(bw *bitWriter, ct *cTable, tableLog uint8, first symbolTransform) {
	c.bw = bw
	c.stateTable = ct.stateTable

	nbBitsOut := (first.deltaNbBits + (1 << 15)) >> 16
	im := int32((nbBitsOut << 16) - first.deltaNbBits)
	lu := (im >> nbBitsOut) + first.deltaFindState
	c.state = c.stateTable[lu]
	return
}

// encode the output symbol provided and write it to the bitstream.
func (c *cState) encode(symbolTT symbolTransform) {
	nbBitsOut := (uint32(c.state) + symbolTT.deltaNbBits) >> 16
	dstState := int32(c.state>>(nbBitsOut&15)) + symbolTT.deltaFindState
	c.bw.addBits16NC(c.state, uint8(nbBitsOut))
	c.state = c.stateTable[dstState]
}

// encode the output symbol provided and write it to the bitstream.
func (c *cState) encodeZero(symbolTT symbolTransform) {
	nbBitsOut := (uint32(c.state) + symbolTT.deltaNbBits) >> 16
	dstState := int32(c.state>>(nbBitsOut&15)) + symbolTT.deltaFindState
	c.bw.addBits16ZeroNC(c.state, uint8(nbBitsOut))
	c.state = c.stateTable[dstState]
}

// flush will write the tablelog to the output and flush the remaining full bytes.
func (c *cState) flush(tableLog uint8) {
	c.bw.flush32()
	c.bw.addBits16NC(c.state, tableLog)
	c.bw.flush()
}

// compress is the main compression loop that will encode the input from the last byte to the first.
func (s *Scratch) compress(src []byte) error {
	if len(src) <= 2 {
		return errors.New("compress: src too small")
	}
	tt := s.ct.symbolTT[:256]
	s.bw.reset(s.Out)

	// Our two states each encodes every second byte.
	// Last byte encoded (first byte decoded) will always be encoded by c1.
	var c1, c2 cState

	// Encode so remaining size is divisible by 4.
	ip := len(src)
	if ip&1 == 1 {
		c1.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-1]])
		c2.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-2]])
		c1.encodeZero(tt[src[ip-3]])
		ip -= 3
	} else {
		c2.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-1]])
		c1.init(&s.bw, &s.ct, s.actualTableLog, tt[src[ip-2]])
		ip -= 2
	}
	if ip&2 != 0 {
		c2.encodeZero(tt[src[ip-1]])
		c1.encodeZero(tt[src[ip-2]])
		ip -= 2
	}

	// Main compression loop.
	switch {
	case !s.zeroBits && s.actualTableLog <= 8:
		// We can encode 4 symbols without requiring a flush.
		// We do not need to check if any output is 0 bits.
		for ip >= 4 {
			s.bw.flush32()
			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
			c2.encode(tt[v0])
			c1.encode(tt[v1])
			c2.encode(tt[v2])
			c1.encode(tt[v3])
			ip -= 4
		}
	case !s.zeroBits:
		// We do not need to check if any output is 0 bits.
		for ip >= 4 {
			s.bw.flush32()
			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
			c2.encode(tt[v0])
			c1.encode(tt[v1])
			s.bw.flush32()
			c2.encode(tt[v2])
			c1.encode(tt[v3])
			ip -= 4
		}
	case s.actualTableLog <= 8:
		// We can encode 4 symbols without requiring a flush
		for ip >= 4 {
			s.bw.flush32()
			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
			c2.encodeZero(tt[v0])
			c1.encodeZero(tt[v1])
			c2.encodeZero(tt[v2])
			c1.encodeZero(tt[v3])
			ip -= 4
		}
	default:
		for ip >= 4 {
			s.bw.flush32()
			v3, v2, v1, v0 := src[ip-4], src[ip-3], src[ip-2], src[ip-1]
			c2.encodeZero(tt[v0])
			c1.encodeZero(tt[v1])
			s.bw.flush32()
			c2.encodeZero(tt[v2])
			c1.encodeZero(tt[v3])
			ip -= 4
		}
	}

	// Flush final state.
	// Used to initialize state when decoding.
	c2.flush(s.actualTableLog)
	c1.flush(s.actualTableLog)

	return s.bw.close()
}

// writeCount will write the normalized histogram count to header.
// This is read back by readNCount.
func (s *Scratch) writeCount() error {
	var (
		tableLog  = s.actualTableLog
		tableSize = 1 << tableLog
		previous0 bool
		charnum   uint16

		maxHeaderSize = ((int(s.symbolLen) * int(tableLog)) >> 3) + 3

		// Write Table Size
		bitStream = uint32(tableLog - minTablelog)
		bitCount  = uint(4)
		remaining = int16(tableSize + 1) /* +1 for extra accuracy */
		threshold = int16(tableSize)
		nbBits    = uint(tableLog + 1)
	)
	if cap(s.Out) < maxHeaderSize {
		s.Out = make([]byte, 0, s.br.remain()+maxHeaderSize)
	}
	outP := uint(0)
	out := s.Out[:maxHeaderSize]

	// stops at 1
	for remaining > 1 {
		if previous0 {
			start := charnum
			for s.norm[charnum] == 0 {
				charnum++
			}
			for charnum >= start+24 {
				start += 24
				bitStream += uint32(0xFFFF) << bitCount
				out[outP] = byte(bitStream)
				out[outP+1] = byte(bitStream >> 8)
				outP += 2
				bitStream >>= 16
			}
			for charnum >= start+3 {
				start += 3
				bitStream += 3 << bitCount
				bitCount += 2
			}
			bitStream += uint32(charnum-start) << bitCount
			bitCount += 2
			if bitCount > 16 {
				out[outP] = byte(bitStream)
				out[outP+1] = byte(bitStream >> 8)
				outP += 2
				bitStream >>= 16
				bitCount -= 16
			}
		}

		count := s.norm[charnum]
		charnum++
		max := (2*threshold - 1) - remaining
		if count < 0 {
			remaining += count
		} else {
			remaining -= count
		}
		count++ // +1 for extra accuracy
		if count >= threshold {
			count += max // [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[
		}
		bitStream += uint32(count) << bitCount
		bitCount += nbBits
		if count < max {
			bitCount--
		}

		previous0 = count == 1
		if remaining < 1 {
			return errors.New("internal error: remaining<1")
		}
		for remaining < threshold {
			nbBits--
			threshold >>= 1
		}

		if bitCount > 16 {
			out[outP] = byte(bitStream)
			out[outP+1] = byte(bitStream >> 8)
			outP += 2
			bitStream >>= 16
			bitCount -= 16
		}
	}

	out[outP] = byte(bitStream)
	out[outP+1] = byte(bitStream >> 8)
	outP += (bitCount + 7) / 8

	if uint16(charnum) > s.symbolLen {
		return errors.New("internal error: charnum > s.symbolLen")
	}
	s.Out = out[:outP]
	return nil
}

// symbolTransform contains the state transform for a symbol.
type symbolTransform struct {
	deltaFindState int32
	deltaNbBits    uint32
}

// String prints values as a human readable string.
func (s symbolTransform) String() string {
	return fmt.Sprintf("dnbits: %08x, fs:%d", s.deltaNbBits, s.deltaFindState)
}

// cTable contains tables used for compression.
type cTable struct {
	tableSymbol []byte
	stateTable  []uint16
	symbolTT    []symbolTransform
}

// allocCtable will allocate tables needed for compression.
// If existing tables a re big enough, they are simply re-used.
func (s *Scratch) allocCtable() {
	tableSize := 1 << s.actualTableLog
	// get tableSymbol that is big enough.
	if cap(s.ct.tableSymbol) < int(tableSize) {
		s.ct.tableSymbol = make([]byte, tableSize)
	}
	s.ct.tableSymbol = s.ct.tableSymbol[:tableSize]

	ctSize := tableSize
	if cap(s.ct.stateTable) < ctSize {
		s.ct.stateTable = make([]uint16, ctSize)
	}
	s.ct.stateTable = s.ct.stateTable[:ctSize]

	if cap(s.ct.symbolTT) < 256 {
		s.ct.symbolTT = make([]symbolTransform, 256)
	}
	s.ct.symbolTT = s.ct.symbolTT[:256]
}

// buildCTable will populate the compression table so it is ready to be used.
func (s *Scratch) buildCTable() error {
	tableSize := uint32(1 << s.actualTableLog)
	highThreshold := tableSize - 1
	var cumul [maxSymbolValue + 2]int16

	s.allocCtable()
	tableSymbol := s.ct.tableSymbol[:tableSize]
	// symbol start positions
	{
		cumul[0] = 0
		for ui, v := range s.norm[:s.symbolLen-1] {
			u := byte(ui) // one less than reference
			if v == -1 {
				// Low proba symbol
				cumul[u+1] = cumul[u] + 1
				tableSymbol[highThreshold] = u
				highThreshold--
			} else {
				cumul[u+1] = cumul[u] + v
			}
		}
		// Encode last symbol separately to avoid overflowing u
		u := int(s.symbolLen - 1)
		v := s.norm[s.symbolLen-1]
		if v == -1 {
			// Low proba symbol
			cumul[u+1] = cumul[u] + 1
			tableSymbol[highThreshold] = byte(u)
			highThreshold--
		} else {
			cumul[u+1] = cumul[u] + v
		}
		if uint32(cumul[s.symbolLen]) != tableSize {
			return fmt.Errorf("internal error: expected cumul[s.symbolLen] (%d) == tableSize (%d)", cumul[s.symbolLen], tableSize)
		}
		cumul[s.symbolLen] = int16(tableSize) + 1
	}
	// Spread symbols
	s.zeroBits = false
	{
		step := tableStep(tableSize)
		tableMask := tableSize - 1
		var position uint32
		// if any symbol > largeLimit, we may have 0 bits output.
		largeLimit := int16(1 << (s.actualTableLog - 1))
		for ui, v := range s.norm[:s.symbolLen] {
			symbol := byte(ui)
			if v > largeLimit {
				s.zeroBits = true
			}
			for nbOccurrences := int16(0); nbOccurrences < v; nbOccurrences++ {
				tableSymbol[position] = symbol
				position = (position + step) & tableMask
				for position > highThreshold {
					position = (position + step) & tableMask
				} /* Low proba area */
			}
		}

		// Check if we have gone through all positions
		if position != 0 {
			return errors.New("position!=0")
		}
	}

	// Build table
	table := s.ct.stateTable
	{
		tsi := int(tableSize)
		for u, v := range tableSymbol {
			// TableU16 : sorted by symbol order; gives next state value
			table[cumul[v]] = uint16(tsi + u)
			cumul[v]++
		}
	}

	// Build Symbol Transformation Table
	{
		total := int16(0)
		symbolTT := s.ct.symbolTT[:s.symbolLen]
		tableLog := s.actualTableLog
		tl := (uint32(tableLog) << 16) - (1 << tableLog)
		for i, v := range s.norm[:s.symbolLen] {
			switch v {
			case 0:
			case -1, 1:
				symbolTT[i].deltaNbBits = tl
				symbolTT[i].deltaFindState = int32(total - 1)
				total++
			default:
				maxBitsOut := uint32(tableLog) - highBits(uint32(v-1))
				minStatePlus := uint32(v) << maxBitsOut
				symbolTT[i].deltaNbBits = (maxBitsOut << 16) - minStatePlus
				symbolTT[i].deltaFindState = int32(total - v)
				total += v
			}
		}
		if total != int16(tableSize) {
			return fmt.Errorf("total mismatch %d (got) != %d (want)", total, tableSize)
		}
	}
	return nil
}

// countSimple will create a simple histogram in s.count.
// Returns the biggest count.
// Does not update s.clearCount.
func (s *Scratch) countSimple(in []byte) (max int) {
	for _, v := range in {
		s.count[v]++
	}
	m := uint32(0)
	for i, v := range s.count[:] {
		if v > m {
			m = v
		}
		if v > 0 {
			s.symbolLen = uint16(i) + 1
		}
	}
	return int(m)
}

// minTableLog provides the minimum logSize to safely represent a distribution.
func (s *Scratch) minTableLog() uint8 {
	minBitsSrc := highBits(uint32(s.br.remain()-1)) + 1
	minBitsSymbols := highBits(uint32(s.symbolLen-1)) + 2
	if minBitsSrc < minBitsSymbols {
		return uint8(minBitsSrc)
	}
	return uint8(minBitsSymbols)
}

// optimalTableLog calculates and sets the optimal tableLog in s.actualTableLog
func (s *Scratch) optimalTableLog() {
	tableLog := s.TableLog
	minBits := s.minTableLog()
	maxBitsSrc := uint8(highBits(uint32(s.br.remain()-1))) - 2
	if maxBitsSrc < tableLog {
		// Accuracy can be reduced
		tableLog = maxBitsSrc
	}
	if minBits > tableLog {
		tableLog = minBits
	}
	// Need a minimum to safely represent all symbol values
	if tableLog < minTablelog {
		tableLog = minTablelog
	}
	if tableLog > maxTableLog {
		tableLog = maxTableLog
	}
	s.actualTableLog = tableLog
}

var rtbTable = [...]uint32{0, 473195, 504333, 520860, 550000, 700000, 750000, 830000}

// normalizeCount will normalize the count of the symbols so
// the total is equal to the table size.
func (s *Scratch) normalizeCount() error {
	var (
		tableLog          = s.actualTableLog
		scale             = 62 - uint64(tableLog)
		step              = (1 << 62) / uint64(s.br.remain())
		vStep             = uint64(1) << (scale - 20)
		stillToDistribute = int16(1 << tableLog)
		largest           int
		largestP          int16
		lowThreshold      = (uint32)(s.br.remain() >> tableLog)
	)

	for i, cnt := range s.count[:s.symbolLen] {
		// already handled
		// if (count[s] == s.length) return 0;   /* rle special case */

		if cnt == 0 {
			s.norm[i] = 0
			continue
		}
		if cnt <= lowThreshold {
			s.norm[i] = -1
			stillToDistribute--
		} else {
			proba := (int16)((uint64(cnt) * step) >> scale)
			if proba < 8 {
				restToBeat := vStep * uint64(rtbTable[proba])
				v := uint64(cnt)*step - (uint64(proba) << scale)
				if v > restToBeat {
					proba++
				}
			}
			if proba > largestP {
				largestP = proba
				largest = i
			}
			s.norm[i] = proba
			stillToDistribute -= proba
		}
	}

	if -stillToDistribute >= (s.norm[largest] >> 1) {
		// corner case, need another normalization method
		return s.normalizeCount2()
	}
	s.norm[largest] += stillToDistribute
	return nil
}

// Secondary normalization method.
// To be used when primary method fails.
func (s *Scratch) normalizeCount2() error {
	const notYetAssigned = -2
	var (
		distributed  uint32
		total        = uint32(s.br.remain())
		tableLog     = s.actualTableLog
		lowThreshold = uint32(total >> tableLog)
		lowOne       = uint32((total * 3) >> (tableLog + 1))
	)
	for i, cnt := range s.count[:s.symbolLen] {
		if cnt == 0 {
			s.norm[i] = 0
			continue
		}
		if cnt <= lowThreshold {
			s.norm[i] = -1
			distributed++
			total -= cnt
			continue
		}
		if cnt <= lowOne {
			s.norm[i] = 1
			distributed++
			total -= cnt
			continue
		}
		s.norm[i] = notYetAssigned
	}
	toDistribute := (1 << tableLog) - distributed

	if (total / toDistribute) > lowOne {
		// risk of rounding to zero
		lowOne = uint32((total * 3) / (toDistribute * 2))
		for i, cnt := range s.count[:s.symbolLen] {
			if (s.norm[i] == notYetAssigned) && (cnt <= lowOne) {
				s.norm[i] = 1
				distributed++
				total -= cnt
				continue
			}
		}
		toDistribute = (1 << tableLog) - distributed
	}
	if distributed == uint32(s.symbolLen)+1 {
		// all values are pretty poor;
		//   probably incompressible data (should have already been detected);
		//   find max, then give all remaining points to max
		var maxV int
		var maxC uint32
		for i, cnt := range s.count[:s.symbolLen] {
			if cnt > maxC {
				maxV = i
				maxC = cnt
			}
		}
		s.norm[maxV] += int16(toDistribute)
		return nil
	}

	if total == 0 {
		// all of the symbols were low enough for the lowOne or lowThreshold
		for i := uint32(0); toDistribute > 0; i = (i + 1) % (uint32(s.symbolLen)) {
			if s.norm[i] > 0 {
				toDistribute--
				s.norm[i]++
			}
		}
		return nil
	}

	var (
		vStepLog = 62 - uint64(tableLog)
		mid      = uint64((1 << (vStepLog - 1)) - 1)
		rStep    = (((1 << vStepLog) * uint64(toDistribute)) + mid) / uint64(total) // scale on remaining
		tmpTotal = mid
	)
	for i, cnt := range s.count[:s.symbolLen] {
		if s.norm[i] == notYetAssigned {
			var (
				end    = tmpTotal + uint64(cnt)*rStep
				sStart = uint32(tmpTotal >> vStepLog)
				sEnd   = uint32(end >> vStepLog)
				weight = sEnd - sStart
			)
			if weight < 1 {
				return errors.New("weight < 1")
			}
			s.norm[i] = int16(weight)
			tmpTotal = end
		}
	}
	return nil
}

// validateNorm validates the normalized histogram table.
func (s *Scratch) validateNorm() (err error) {
	var total int
	for _, v := range s.norm[:s.symbolLen] {
		if v >= 0 {
			total += int(v)
		} else {
			total -= int(v)
		}
	}
	defer func() {
		if err == nil {
			return
		}
		fmt.Printf("selected TableLog: %d, Symbol length: %d\n", s.actualTableLog, s.symbolLen)
		for i, v := range s.norm[:s.symbolLen] {
			fmt.Printf("%3d: %5d -> %4d \n", i, s.count[i], v)
		}
	}()
	if total != (1 << s.actualTableLog) {
		return fmt.Errorf("warning: Total == %d != %d", total, 1<<s.actualTableLog)
	}
	for i, v := range s.count[s.symbolLen:] {
		if v != 0 {
			return fmt.Errorf("warning: Found symbol out of range, %d after cut", i)
		}
	}
	return nil
}