summaryrefslogtreecommitdiff
path: root/vendor/github.com/klauspost/compress/huff0/compress.go
blob: dd4f7fefb9fbe7bfc33ba1841f80a5c3efe8e568 (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
package huff0

import (
	"fmt"
	"runtime"
	"sync"
)

// Compress1X will compress the input.
// The output can be decoded using Decompress1X.
// Supply a Scratch object. The scratch object contains state about re-use,
// So when sharing across independent encodes, be sure to set the re-use policy.
func Compress1X(in []byte, s *Scratch) (out []byte, reUsed bool, err error) {
	s, err = s.prepare(in)
	if err != nil {
		return nil, false, err
	}
	return compress(in, s, s.compress1X)
}

// Compress4X will compress the input. The input is split into 4 independent blocks
// and compressed similar to Compress1X.
// The output can be decoded using Decompress4X.
// Supply a Scratch object. The scratch object contains state about re-use,
// So when sharing across independent encodes, be sure to set the re-use policy.
func Compress4X(in []byte, s *Scratch) (out []byte, reUsed bool, err error) {
	s, err = s.prepare(in)
	if err != nil {
		return nil, false, err
	}
	if false {
		// TODO: compress4Xp only slightly faster.
		const parallelThreshold = 8 << 10
		if len(in) < parallelThreshold || runtime.GOMAXPROCS(0) == 1 {
			return compress(in, s, s.compress4X)
		}
		return compress(in, s, s.compress4Xp)
	}
	return compress(in, s, s.compress4X)
}

func compress(in []byte, s *Scratch, compressor func(src []byte) ([]byte, error)) (out []byte, reUsed bool, err error) {
	// Nuke previous table if we cannot reuse anyway.
	if s.Reuse == ReusePolicyNone {
		s.prevTable = s.prevTable[:0]
	}

	// Create histogram, if none was provided.
	maxCount := s.maxCount
	var canReuse = false
	if maxCount == 0 {
		maxCount, canReuse = s.countSimple(in)
	} else {
		canReuse = s.canUseTable(s.prevTable)
	}

	// Reset for next run.
	s.clearCount = true
	s.maxCount = 0
	if maxCount >= len(in) {
		if maxCount > len(in) {
			return nil, false, fmt.Errorf("maxCount (%d) > length (%d)", maxCount, len(in))
		}
		if len(in) == 1 {
			return nil, false, ErrIncompressible
		}
		// One symbol, use RLE
		return nil, false, ErrUseRLE
	}
	if maxCount == 1 || maxCount < (len(in)>>7) {
		// Each symbol present maximum once or too well distributed.
		return nil, false, ErrIncompressible
	}

	if s.Reuse == ReusePolicyPrefer && canReuse {
		keepTable := s.cTable
		s.cTable = s.prevTable
		s.Out, err = compressor(in)
		s.cTable = keepTable
		if err == nil && len(s.Out) < len(in) {
			s.OutData = s.Out
			return s.Out, true, nil
		}
		// Do not attempt to re-use later.
		s.prevTable = s.prevTable[:0]
	}

	// Calculate new table.
	s.optimalTableLog()
	err = s.buildCTable()
	if err != nil {
		return nil, false, err
	}

	if false && !s.canUseTable(s.cTable) {
		panic("invalid table generated")
	}

	if s.Reuse == ReusePolicyAllow && canReuse {
		hSize := len(s.Out)
		oldSize := s.prevTable.estimateSize(s.count[:s.symbolLen])
		newSize := s.cTable.estimateSize(s.count[:s.symbolLen])
		if oldSize <= hSize+newSize || hSize+12 >= len(in) {
			// Retain cTable even if we re-use.
			keepTable := s.cTable
			s.cTable = s.prevTable
			s.Out, err = compressor(in)
			s.cTable = keepTable
			if len(s.Out) >= len(in) {
				return nil, false, ErrIncompressible
			}
			s.OutData = s.Out
			return s.Out, true, nil
		}
	}

	// Use new table
	err = s.cTable.write(s)
	if err != nil {
		s.OutTable = nil
		return nil, false, err
	}
	s.OutTable = s.Out

	// Compress using new table
	s.Out, err = compressor(in)
	if err != nil {
		s.OutTable = nil
		return nil, false, err
	}
	if len(s.Out) >= len(in) {
		s.OutTable = nil
		return nil, false, ErrIncompressible
	}
	// Move current table into previous.
	s.prevTable, s.cTable = s.cTable, s.prevTable[:0]
	s.OutData = s.Out[len(s.OutTable):]
	return s.Out, false, nil
}

func (s *Scratch) compress1X(src []byte) ([]byte, error) {
	return s.compress1xDo(s.Out, src)
}

func (s *Scratch) compress1xDo(dst, src []byte) ([]byte, error) {
	var bw = bitWriter{out: dst}

	// N is length divisible by 4.
	n := len(src)
	n -= n & 3
	cTable := s.cTable[:256]

	// Encode last bytes.
	for i := len(src) & 3; i > 0; i-- {
		bw.encSymbol(cTable, src[n+i-1])
	}
	if s.actualTableLog <= 8 {
		n -= 4
		for ; n >= 0; n -= 4 {
			tmp := src[n : n+4]
			// tmp should be len 4
			bw.flush32()
			bw.encSymbol(cTable, tmp[3])
			bw.encSymbol(cTable, tmp[2])
			bw.encSymbol(cTable, tmp[1])
			bw.encSymbol(cTable, tmp[0])
		}
	} else {
		n -= 4
		for ; n >= 0; n -= 4 {
			tmp := src[n : n+4]
			// tmp should be len 4
			bw.flush32()
			bw.encSymbol(cTable, tmp[3])
			bw.encSymbol(cTable, tmp[2])
			bw.flush32()
			bw.encSymbol(cTable, tmp[1])
			bw.encSymbol(cTable, tmp[0])
		}
	}
	err := bw.close()
	return bw.out, err
}

var sixZeros [6]byte

func (s *Scratch) compress4X(src []byte) ([]byte, error) {
	if len(src) < 12 {
		return nil, ErrIncompressible
	}
	segmentSize := (len(src) + 3) / 4

	// Add placeholder for output length
	offsetIdx := len(s.Out)
	s.Out = append(s.Out, sixZeros[:]...)

	for i := 0; i < 4; i++ {
		toDo := src
		if len(toDo) > segmentSize {
			toDo = toDo[:segmentSize]
		}
		src = src[len(toDo):]

		var err error
		idx := len(s.Out)
		s.Out, err = s.compress1xDo(s.Out, toDo)
		if err != nil {
			return nil, err
		}
		// Write compressed length as little endian before block.
		if i < 3 {
			// Last length is not written.
			length := len(s.Out) - idx
			s.Out[i*2+offsetIdx] = byte(length)
			s.Out[i*2+offsetIdx+1] = byte(length >> 8)
		}
	}

	return s.Out, nil
}

// compress4Xp will compress 4 streams using separate goroutines.
func (s *Scratch) compress4Xp(src []byte) ([]byte, error) {
	if len(src) < 12 {
		return nil, ErrIncompressible
	}
	// Add placeholder for output length
	s.Out = s.Out[:6]

	segmentSize := (len(src) + 3) / 4
	var wg sync.WaitGroup
	var errs [4]error
	wg.Add(4)
	for i := 0; i < 4; i++ {
		toDo := src
		if len(toDo) > segmentSize {
			toDo = toDo[:segmentSize]
		}
		src = src[len(toDo):]

		// Separate goroutine for each block.
		go func(i int) {
			s.tmpOut[i], errs[i] = s.compress1xDo(s.tmpOut[i][:0], toDo)
			wg.Done()
		}(i)
	}
	wg.Wait()
	for i := 0; i < 4; i++ {
		if errs[i] != nil {
			return nil, errs[i]
		}
		o := s.tmpOut[i]
		// Write compressed length as little endian before block.
		if i < 3 {
			// Last length is not written.
			s.Out[i*2] = byte(len(o))
			s.Out[i*2+1] = byte(len(o) >> 8)
		}

		// Write output.
		s.Out = append(s.Out, o...)
	}
	return s.Out, 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, reuse bool) {
	reuse = true
	for _, v := range in {
		s.count[v]++
	}
	m := uint32(0)
	if len(s.prevTable) > 0 {
		for i, v := range s.count[:] {
			if v > m {
				m = v
			}
			if v > 0 {
				s.symbolLen = uint16(i) + 1
				if i >= len(s.prevTable) {
					reuse = false
				} else {
					if s.prevTable[i].nBits == 0 {
						reuse = false
					}
				}
			}
		}
		return int(m), reuse
	}
	for i, v := range s.count[:] {
		if v > m {
			m = v
		}
		if v > 0 {
			s.symbolLen = uint16(i) + 1
		}
	}
	return int(m), false
}

func (s *Scratch) canUseTable(c cTable) bool {
	if len(c) < int(s.symbolLen) {
		return false
	}
	for i, v := range s.count[:s.symbolLen] {
		if v != 0 && c[i].nBits == 0 {
			return false
		}
	}
	return true
}

// minTableLog provides the minimum logSize to safely represent a distribution.
func (s *Scratch) minTableLog() uint8 {
	minBitsSrc := highBit32(uint32(s.br.remain()-1)) + 1
	minBitsSymbols := highBit32(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(highBit32(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 > tableLogMax {
		tableLog = tableLogMax
	}
	s.actualTableLog = tableLog
}

type cTableEntry struct {
	val   uint16
	nBits uint8
	// We have 8 bits extra
}

const huffNodesMask = huffNodesLen - 1

func (s *Scratch) buildCTable() error {
	s.huffSort()
	if cap(s.cTable) < maxSymbolValue+1 {
		s.cTable = make([]cTableEntry, s.symbolLen, maxSymbolValue+1)
	} else {
		s.cTable = s.cTable[:s.symbolLen]
		for i := range s.cTable {
			s.cTable[i] = cTableEntry{}
		}
	}

	var startNode = int16(s.symbolLen)
	nonNullRank := s.symbolLen - 1

	nodeNb := int16(startNode)
	huffNode := s.nodes[1 : huffNodesLen+1]

	// This overlays the slice above, but allows "-1" index lookups.
	// Different from reference implementation.
	huffNode0 := s.nodes[0 : huffNodesLen+1]

	for huffNode[nonNullRank].count == 0 {
		nonNullRank--
	}

	lowS := int16(nonNullRank)
	nodeRoot := nodeNb + lowS - 1
	lowN := nodeNb
	huffNode[nodeNb].count = huffNode[lowS].count + huffNode[lowS-1].count
	huffNode[lowS].parent, huffNode[lowS-1].parent = uint16(nodeNb), uint16(nodeNb)
	nodeNb++
	lowS -= 2
	for n := nodeNb; n <= nodeRoot; n++ {
		huffNode[n].count = 1 << 30
	}
	// fake entry, strong barrier
	huffNode0[0].count = 1 << 31

	// create parents
	for nodeNb <= nodeRoot {
		var n1, n2 int16
		if huffNode0[lowS+1].count < huffNode0[lowN+1].count {
			n1 = lowS
			lowS--
		} else {
			n1 = lowN
			lowN++
		}
		if huffNode0[lowS+1].count < huffNode0[lowN+1].count {
			n2 = lowS
			lowS--
		} else {
			n2 = lowN
			lowN++
		}

		huffNode[nodeNb].count = huffNode0[n1+1].count + huffNode0[n2+1].count
		huffNode0[n1+1].parent, huffNode0[n2+1].parent = uint16(nodeNb), uint16(nodeNb)
		nodeNb++
	}

	// distribute weights (unlimited tree height)
	huffNode[nodeRoot].nbBits = 0
	for n := nodeRoot - 1; n >= startNode; n-- {
		huffNode[n].nbBits = huffNode[huffNode[n].parent].nbBits + 1
	}
	for n := uint16(0); n <= nonNullRank; n++ {
		huffNode[n].nbBits = huffNode[huffNode[n].parent].nbBits + 1
	}
	s.actualTableLog = s.setMaxHeight(int(nonNullRank))
	maxNbBits := s.actualTableLog

	// fill result into tree (val, nbBits)
	if maxNbBits > tableLogMax {
		return fmt.Errorf("internal error: maxNbBits (%d) > tableLogMax (%d)", maxNbBits, tableLogMax)
	}
	var nbPerRank [tableLogMax + 1]uint16
	var valPerRank [tableLogMax + 1]uint16
	for _, v := range huffNode[:nonNullRank+1] {
		nbPerRank[v.nbBits]++
	}
	// determine stating value per rank
	{
		min := uint16(0)
		for n := maxNbBits; n > 0; n-- {
			// get starting value within each rank
			valPerRank[n] = min
			min += nbPerRank[n]
			min >>= 1
		}
	}

	// push nbBits per symbol, symbol order
	// TODO: changed `s.symbolLen` -> `nonNullRank+1` (micro-opt)
	for _, v := range huffNode[:nonNullRank+1] {
		s.cTable[v.symbol].nBits = v.nbBits
	}

	// assign value within rank, symbol order
	for n, val := range s.cTable[:s.symbolLen] {
		v := valPerRank[val.nBits]
		s.cTable[n].val = v
		valPerRank[val.nBits] = v + 1
	}

	return nil
}

// huffSort will sort symbols, decreasing order.
func (s *Scratch) huffSort() {
	type rankPos struct {
		base    uint32
		current uint32
	}

	// Clear nodes
	nodes := s.nodes[:huffNodesLen+1]
	s.nodes = nodes
	nodes = nodes[1 : huffNodesLen+1]

	// Sort into buckets based on length of symbol count.
	var rank [32]rankPos
	for _, v := range s.count[:s.symbolLen] {
		r := highBit32(v+1) & 31
		rank[r].base++
	}
	for n := 30; n > 0; n-- {
		rank[n-1].base += rank[n].base
	}
	for n := range rank[:] {
		rank[n].current = rank[n].base
	}
	for n, c := range s.count[:s.symbolLen] {
		r := (highBit32(c+1) + 1) & 31
		pos := rank[r].current
		rank[r].current++
		prev := nodes[(pos-1)&huffNodesMask]
		for pos > rank[r].base && c > prev.count {
			nodes[pos&huffNodesMask] = prev
			pos--
			prev = nodes[(pos-1)&huffNodesMask]
		}
		nodes[pos&huffNodesMask] = nodeElt{count: c, symbol: byte(n)}
	}
	return
}

func (s *Scratch) setMaxHeight(lastNonNull int) uint8 {
	maxNbBits := s.TableLog
	huffNode := s.nodes[1 : huffNodesLen+1]
	//huffNode = huffNode[: huffNodesLen]

	largestBits := huffNode[lastNonNull].nbBits

	// early exit : no elt > maxNbBits
	if largestBits <= maxNbBits {
		return largestBits
	}
	totalCost := int(0)
	baseCost := int(1) << (largestBits - maxNbBits)
	n := uint32(lastNonNull)

	for huffNode[n].nbBits > maxNbBits {
		totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits))
		huffNode[n].nbBits = maxNbBits
		n--
	}
	// n stops at huffNode[n].nbBits <= maxNbBits

	for huffNode[n].nbBits == maxNbBits {
		n--
	}
	// n end at index of smallest symbol using < maxNbBits

	// renorm totalCost
	totalCost >>= largestBits - maxNbBits /* note : totalCost is necessarily a multiple of baseCost */

	// repay normalized cost
	{
		const noSymbol = 0xF0F0F0F0
		var rankLast [tableLogMax + 2]uint32

		for i := range rankLast[:] {
			rankLast[i] = noSymbol
		}

		// Get pos of last (smallest) symbol per rank
		{
			currentNbBits := uint8(maxNbBits)
			for pos := int(n); pos >= 0; pos-- {
				if huffNode[pos].nbBits >= currentNbBits {
					continue
				}
				currentNbBits = huffNode[pos].nbBits // < maxNbBits
				rankLast[maxNbBits-currentNbBits] = uint32(pos)
			}
		}

		for totalCost > 0 {
			nBitsToDecrease := uint8(highBit32(uint32(totalCost))) + 1

			for ; nBitsToDecrease > 1; nBitsToDecrease-- {
				highPos := rankLast[nBitsToDecrease]
				lowPos := rankLast[nBitsToDecrease-1]
				if highPos == noSymbol {
					continue
				}
				if lowPos == noSymbol {
					break
				}
				highTotal := huffNode[highPos].count
				lowTotal := 2 * huffNode[lowPos].count
				if highTotal <= lowTotal {
					break
				}
			}
			// only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !)
			// HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary
			// FIXME: try to remove
			for (nBitsToDecrease <= tableLogMax) && (rankLast[nBitsToDecrease] == noSymbol) {
				nBitsToDecrease++
			}
			totalCost -= 1 << (nBitsToDecrease - 1)
			if rankLast[nBitsToDecrease-1] == noSymbol {
				// this rank is no longer empty
				rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease]
			}
			huffNode[rankLast[nBitsToDecrease]].nbBits++
			if rankLast[nBitsToDecrease] == 0 {
				/* special case, reached largest symbol */
				rankLast[nBitsToDecrease] = noSymbol
			} else {
				rankLast[nBitsToDecrease]--
				if huffNode[rankLast[nBitsToDecrease]].nbBits != maxNbBits-nBitsToDecrease {
					rankLast[nBitsToDecrease] = noSymbol /* this rank is now empty */
				}
			}
		}

		for totalCost < 0 { /* Sometimes, cost correction overshoot */
			if rankLast[1] == noSymbol { /* special case : no rank 1 symbol (using maxNbBits-1); let's create one from largest rank 0 (using maxNbBits) */
				for huffNode[n].nbBits == maxNbBits {
					n--
				}
				huffNode[n+1].nbBits--
				rankLast[1] = n + 1
				totalCost++
				continue
			}
			huffNode[rankLast[1]+1].nbBits--
			rankLast[1]++
			totalCost++
		}
	}
	return maxNbBits
}

type nodeElt struct {
	count  uint32
	parent uint16
	symbol byte
	nbBits uint8
}