1 | package brotli
|
---|
2 |
|
---|
3 | import (
|
---|
4 | "io"
|
---|
5 | "math"
|
---|
6 | )
|
---|
7 |
|
---|
8 | /* Copyright 2016 Google Inc. All Rights Reserved.
|
---|
9 |
|
---|
10 | Distributed under MIT license.
|
---|
11 | See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
|
---|
12 | */
|
---|
13 |
|
---|
14 | /** Minimal value for ::BROTLI_PARAM_LGWIN parameter. */
|
---|
15 | const minWindowBits = 10
|
---|
16 |
|
---|
17 | /**
|
---|
18 | * Maximal value for ::BROTLI_PARAM_LGWIN parameter.
|
---|
19 | *
|
---|
20 | * @note equal to @c BROTLI_MAX_DISTANCE_BITS constant.
|
---|
21 | */
|
---|
22 | const maxWindowBits = 24
|
---|
23 |
|
---|
24 | /**
|
---|
25 | * Maximal value for ::BROTLI_PARAM_LGWIN parameter
|
---|
26 | * in "Large Window Brotli" (32-bit).
|
---|
27 | */
|
---|
28 | const largeMaxWindowBits = 30
|
---|
29 |
|
---|
30 | /** Minimal value for ::BROTLI_PARAM_LGBLOCK parameter. */
|
---|
31 | const minInputBlockBits = 16
|
---|
32 |
|
---|
33 | /** Maximal value for ::BROTLI_PARAM_LGBLOCK parameter. */
|
---|
34 | const maxInputBlockBits = 24
|
---|
35 |
|
---|
36 | /** Minimal value for ::BROTLI_PARAM_QUALITY parameter. */
|
---|
37 | const minQuality = 0
|
---|
38 |
|
---|
39 | /** Maximal value for ::BROTLI_PARAM_QUALITY parameter. */
|
---|
40 | const maxQuality = 11
|
---|
41 |
|
---|
42 | /** Options for ::BROTLI_PARAM_MODE parameter. */
|
---|
43 | const (
|
---|
44 | modeGeneric = 0
|
---|
45 | modeText = 1
|
---|
46 | modeFont = 2
|
---|
47 | )
|
---|
48 |
|
---|
49 | /** Default value for ::BROTLI_PARAM_QUALITY parameter. */
|
---|
50 | const defaultQuality = 11
|
---|
51 |
|
---|
52 | /** Default value for ::BROTLI_PARAM_LGWIN parameter. */
|
---|
53 | const defaultWindow = 22
|
---|
54 |
|
---|
55 | /** Default value for ::BROTLI_PARAM_MODE parameter. */
|
---|
56 | const defaultMode = modeGeneric
|
---|
57 |
|
---|
58 | /** Operations that can be performed by streaming encoder. */
|
---|
59 | const (
|
---|
60 | operationProcess = 0
|
---|
61 | operationFlush = 1
|
---|
62 | operationFinish = 2
|
---|
63 | operationEmitMetadata = 3
|
---|
64 | )
|
---|
65 |
|
---|
66 | const (
|
---|
67 | streamProcessing = 0
|
---|
68 | streamFlushRequested = 1
|
---|
69 | streamFinished = 2
|
---|
70 | streamMetadataHead = 3
|
---|
71 | streamMetadataBody = 4
|
---|
72 | )
|
---|
73 |
|
---|
74 | type Writer struct {
|
---|
75 | dst io.Writer
|
---|
76 | options WriterOptions
|
---|
77 | err error
|
---|
78 |
|
---|
79 | params encoderParams
|
---|
80 | hasher_ hasherHandle
|
---|
81 | input_pos_ uint64
|
---|
82 | ringbuffer_ ringBuffer
|
---|
83 | commands []command
|
---|
84 | num_literals_ uint
|
---|
85 | last_insert_len_ uint
|
---|
86 | last_flush_pos_ uint64
|
---|
87 | last_processed_pos_ uint64
|
---|
88 | dist_cache_ [numDistanceShortCodes]int
|
---|
89 | saved_dist_cache_ [4]int
|
---|
90 | last_bytes_ uint16
|
---|
91 | last_bytes_bits_ byte
|
---|
92 | prev_byte_ byte
|
---|
93 | prev_byte2_ byte
|
---|
94 | storage []byte
|
---|
95 | small_table_ [1 << 10]int
|
---|
96 | large_table_ []int
|
---|
97 | large_table_size_ uint
|
---|
98 | cmd_depths_ [128]byte
|
---|
99 | cmd_bits_ [128]uint16
|
---|
100 | cmd_code_ [512]byte
|
---|
101 | cmd_code_numbits_ uint
|
---|
102 | command_buf_ []uint32
|
---|
103 | literal_buf_ []byte
|
---|
104 | tiny_buf_ struct {
|
---|
105 | u64 [2]uint64
|
---|
106 | u8 [16]byte
|
---|
107 | }
|
---|
108 | remaining_metadata_bytes_ uint32
|
---|
109 | stream_state_ int
|
---|
110 | is_last_block_emitted_ bool
|
---|
111 | is_initialized_ bool
|
---|
112 | }
|
---|
113 |
|
---|
114 | func inputBlockSize(s *Writer) uint {
|
---|
115 | return uint(1) << uint(s.params.lgblock)
|
---|
116 | }
|
---|
117 |
|
---|
118 | func unprocessedInputSize(s *Writer) uint64 {
|
---|
119 | return s.input_pos_ - s.last_processed_pos_
|
---|
120 | }
|
---|
121 |
|
---|
122 | func remainingInputBlockSize(s *Writer) uint {
|
---|
123 | var delta uint64 = unprocessedInputSize(s)
|
---|
124 | var block_size uint = inputBlockSize(s)
|
---|
125 | if delta >= uint64(block_size) {
|
---|
126 | return 0
|
---|
127 | }
|
---|
128 | return block_size - uint(delta)
|
---|
129 | }
|
---|
130 |
|
---|
131 | /* Wraps 64-bit input position to 32-bit ring-buffer position preserving
|
---|
132 | "not-a-first-lap" feature. */
|
---|
133 | func wrapPosition(position uint64) uint32 {
|
---|
134 | var result uint32 = uint32(position)
|
---|
135 | var gb uint64 = position >> 30
|
---|
136 | if gb > 2 {
|
---|
137 | /* Wrap every 2GiB; The first 3GB are continuous. */
|
---|
138 | result = result&((1<<30)-1) | (uint32((gb-1)&1)+1)<<30
|
---|
139 | }
|
---|
140 |
|
---|
141 | return result
|
---|
142 | }
|
---|
143 |
|
---|
144 | func (s *Writer) getStorage(size int) []byte {
|
---|
145 | if len(s.storage) < size {
|
---|
146 | s.storage = make([]byte, size)
|
---|
147 | }
|
---|
148 |
|
---|
149 | return s.storage
|
---|
150 | }
|
---|
151 |
|
---|
152 | func hashTableSize(max_table_size uint, input_size uint) uint {
|
---|
153 | var htsize uint = 256
|
---|
154 | for htsize < max_table_size && htsize < input_size {
|
---|
155 | htsize <<= 1
|
---|
156 | }
|
---|
157 |
|
---|
158 | return htsize
|
---|
159 | }
|
---|
160 |
|
---|
161 | func getHashTable(s *Writer, quality int, input_size uint, table_size *uint) []int {
|
---|
162 | var max_table_size uint = maxHashTableSize(quality)
|
---|
163 | var htsize uint = hashTableSize(max_table_size, input_size)
|
---|
164 | /* Use smaller hash table when input.size() is smaller, since we
|
---|
165 | fill the table, incurring O(hash table size) overhead for
|
---|
166 | compression, and if the input is short, we won't need that
|
---|
167 | many hash table entries anyway. */
|
---|
168 |
|
---|
169 | var table []int
|
---|
170 | assert(max_table_size >= 256)
|
---|
171 | if quality == fastOnePassCompressionQuality {
|
---|
172 | /* Only odd shifts are supported by fast-one-pass. */
|
---|
173 | if htsize&0xAAAAA == 0 {
|
---|
174 | htsize <<= 1
|
---|
175 | }
|
---|
176 | }
|
---|
177 |
|
---|
178 | if htsize <= uint(len(s.small_table_)) {
|
---|
179 | table = s.small_table_[:]
|
---|
180 | } else {
|
---|
181 | if htsize > s.large_table_size_ {
|
---|
182 | s.large_table_size_ = htsize
|
---|
183 | s.large_table_ = nil
|
---|
184 | s.large_table_ = make([]int, htsize)
|
---|
185 | }
|
---|
186 |
|
---|
187 | table = s.large_table_
|
---|
188 | }
|
---|
189 |
|
---|
190 | *table_size = htsize
|
---|
191 | for i := 0; i < int(htsize); i++ {
|
---|
192 | table[i] = 0
|
---|
193 | }
|
---|
194 | return table
|
---|
195 | }
|
---|
196 |
|
---|
197 | func encodeWindowBits(lgwin int, large_window bool, last_bytes *uint16, last_bytes_bits *byte) {
|
---|
198 | if large_window {
|
---|
199 | *last_bytes = uint16((lgwin&0x3F)<<8 | 0x11)
|
---|
200 | *last_bytes_bits = 14
|
---|
201 | } else {
|
---|
202 | if lgwin == 16 {
|
---|
203 | *last_bytes = 0
|
---|
204 | *last_bytes_bits = 1
|
---|
205 | } else if lgwin == 17 {
|
---|
206 | *last_bytes = 1
|
---|
207 | *last_bytes_bits = 7
|
---|
208 | } else if lgwin > 17 {
|
---|
209 | *last_bytes = uint16((lgwin-17)<<1 | 0x01)
|
---|
210 | *last_bytes_bits = 4
|
---|
211 | } else {
|
---|
212 | *last_bytes = uint16((lgwin-8)<<4 | 0x01)
|
---|
213 | *last_bytes_bits = 7
|
---|
214 | }
|
---|
215 | }
|
---|
216 | }
|
---|
217 |
|
---|
218 | /* Decide about the context map based on the ability of the prediction
|
---|
219 | ability of the previous byte UTF8-prefix on the next byte. The
|
---|
220 | prediction ability is calculated as Shannon entropy. Here we need
|
---|
221 | Shannon entropy instead of 'BitsEntropy' since the prefix will be
|
---|
222 | encoded with the remaining 6 bits of the following byte, and
|
---|
223 | BitsEntropy will assume that symbol to be stored alone using Huffman
|
---|
224 | coding. */
|
---|
225 |
|
---|
226 | var kStaticContextMapContinuation = [64]uint32{
|
---|
227 | 1, 1, 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
|
---|
228 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
|
---|
229 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
|
---|
230 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
|
---|
231 | }
|
---|
232 | var kStaticContextMapSimpleUTF8 = [64]uint32{
|
---|
233 | 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
|
---|
234 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
|
---|
235 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
|
---|
236 | 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
|
---|
237 | }
|
---|
238 |
|
---|
239 | func chooseContextMap(quality int, bigram_histo []uint32, num_literal_contexts *uint, literal_context_map *[]uint32) {
|
---|
240 | var monogram_histo = [3]uint32{0}
|
---|
241 | var two_prefix_histo = [6]uint32{0}
|
---|
242 | var total uint
|
---|
243 | var i uint
|
---|
244 | var dummy uint
|
---|
245 | var entropy [4]float64
|
---|
246 | for i = 0; i < 9; i++ {
|
---|
247 | monogram_histo[i%3] += bigram_histo[i]
|
---|
248 | two_prefix_histo[i%6] += bigram_histo[i]
|
---|
249 | }
|
---|
250 |
|
---|
251 | entropy[1] = shannonEntropy(monogram_histo[:], 3, &dummy)
|
---|
252 | entropy[2] = (shannonEntropy(two_prefix_histo[:], 3, &dummy) + shannonEntropy(two_prefix_histo[3:], 3, &dummy))
|
---|
253 | entropy[3] = 0
|
---|
254 | for i = 0; i < 3; i++ {
|
---|
255 | entropy[3] += shannonEntropy(bigram_histo[3*i:], 3, &dummy)
|
---|
256 | }
|
---|
257 |
|
---|
258 | total = uint(monogram_histo[0] + monogram_histo[1] + monogram_histo[2])
|
---|
259 | assert(total != 0)
|
---|
260 | entropy[0] = 1.0 / float64(total)
|
---|
261 | entropy[1] *= entropy[0]
|
---|
262 | entropy[2] *= entropy[0]
|
---|
263 | entropy[3] *= entropy[0]
|
---|
264 |
|
---|
265 | if quality < minQualityForHqContextModeling {
|
---|
266 | /* 3 context models is a bit slower, don't use it at lower qualities. */
|
---|
267 | entropy[3] = entropy[1] * 10
|
---|
268 | }
|
---|
269 |
|
---|
270 | /* If expected savings by symbol are less than 0.2 bits, skip the
|
---|
271 | context modeling -- in exchange for faster decoding speed. */
|
---|
272 | if entropy[1]-entropy[2] < 0.2 && entropy[1]-entropy[3] < 0.2 {
|
---|
273 | *num_literal_contexts = 1
|
---|
274 | } else if entropy[2]-entropy[3] < 0.02 {
|
---|
275 | *num_literal_contexts = 2
|
---|
276 | *literal_context_map = kStaticContextMapSimpleUTF8[:]
|
---|
277 | } else {
|
---|
278 | *num_literal_contexts = 3
|
---|
279 | *literal_context_map = kStaticContextMapContinuation[:]
|
---|
280 | }
|
---|
281 | }
|
---|
282 |
|
---|
283 | /* Decide if we want to use a more complex static context map containing 13
|
---|
284 | context values, based on the entropy reduction of histograms over the
|
---|
285 | first 5 bits of literals. */
|
---|
286 |
|
---|
287 | var kStaticContextMapComplexUTF8 = [64]uint32{
|
---|
288 | 11, 11, 12, 12, /* 0 special */
|
---|
289 | 0, 0, 0, 0, /* 4 lf */
|
---|
290 | 1, 1, 9, 9, /* 8 space */
|
---|
291 | 2, 2, 2, 2, /* !, first after space/lf and after something else. */
|
---|
292 | 1, 1, 1, 1, /* " */
|
---|
293 | 8, 3, 3, 3, /* % */
|
---|
294 | 1, 1, 1, 1, /* ({[ */
|
---|
295 | 2, 2, 2, 2, /* }]) */
|
---|
296 | 8, 4, 4, 4, /* :; */
|
---|
297 | 8, 7, 4, 4, /* . */
|
---|
298 | 8, 0, 0, 0, /* > */
|
---|
299 | 3, 3, 3, 3, /* [0..9] */
|
---|
300 | 5, 5, 10, 5, /* [A-Z] */
|
---|
301 | 5, 5, 10, 5,
|
---|
302 | 6, 6, 6, 6, /* [a-z] */
|
---|
303 | 6, 6, 6, 6,
|
---|
304 | }
|
---|
305 |
|
---|
306 | func shouldUseComplexStaticContextMap(input []byte, start_pos uint, length uint, mask uint, quality int, size_hint uint, num_literal_contexts *uint, literal_context_map *[]uint32) bool {
|
---|
307 | /* Try the more complex static context map only for long data. */
|
---|
308 | if size_hint < 1<<20 {
|
---|
309 | return false
|
---|
310 | } else {
|
---|
311 | var end_pos uint = start_pos + length
|
---|
312 | var combined_histo = [32]uint32{0}
|
---|
313 | var context_histo = [13][32]uint32{[32]uint32{0}}
|
---|
314 | var total uint32 = 0
|
---|
315 | var entropy [3]float64
|
---|
316 | var dummy uint
|
---|
317 | var i uint
|
---|
318 | var utf8_lut contextLUT = getContextLUT(contextUTF8)
|
---|
319 | /* To make entropy calculations faster and to fit on the stack, we collect
|
---|
320 | histograms over the 5 most significant bits of literals. One histogram
|
---|
321 | without context and 13 additional histograms for each context value. */
|
---|
322 | for ; start_pos+64 <= end_pos; start_pos += 4096 {
|
---|
323 | var stride_end_pos uint = start_pos + 64
|
---|
324 | var prev2 byte = input[start_pos&mask]
|
---|
325 | var prev1 byte = input[(start_pos+1)&mask]
|
---|
326 | var pos uint
|
---|
327 |
|
---|
328 | /* To make the analysis of the data faster we only examine 64 byte long
|
---|
329 | strides at every 4kB intervals. */
|
---|
330 | for pos = start_pos + 2; pos < stride_end_pos; pos++ {
|
---|
331 | var literal byte = input[pos&mask]
|
---|
332 | var context byte = byte(kStaticContextMapComplexUTF8[getContext(prev1, prev2, utf8_lut)])
|
---|
333 | total++
|
---|
334 | combined_histo[literal>>3]++
|
---|
335 | context_histo[context][literal>>3]++
|
---|
336 | prev2 = prev1
|
---|
337 | prev1 = literal
|
---|
338 | }
|
---|
339 | }
|
---|
340 |
|
---|
341 | entropy[1] = shannonEntropy(combined_histo[:], 32, &dummy)
|
---|
342 | entropy[2] = 0
|
---|
343 | for i = 0; i < 13; i++ {
|
---|
344 | entropy[2] += shannonEntropy(context_histo[i][0:], 32, &dummy)
|
---|
345 | }
|
---|
346 |
|
---|
347 | entropy[0] = 1.0 / float64(total)
|
---|
348 | entropy[1] *= entropy[0]
|
---|
349 | entropy[2] *= entropy[0]
|
---|
350 |
|
---|
351 | /* The triggering heuristics below were tuned by compressing the individual
|
---|
352 | files of the silesia corpus. If we skip this kind of context modeling
|
---|
353 | for not very well compressible input (i.e. entropy using context modeling
|
---|
354 | is 60% of maximal entropy) or if expected savings by symbol are less
|
---|
355 | than 0.2 bits, then in every case when it triggers, the final compression
|
---|
356 | ratio is improved. Note however that this heuristics might be too strict
|
---|
357 | for some cases and could be tuned further. */
|
---|
358 | if entropy[2] > 3.0 || entropy[1]-entropy[2] < 0.2 {
|
---|
359 | return false
|
---|
360 | } else {
|
---|
361 | *num_literal_contexts = 13
|
---|
362 | *literal_context_map = kStaticContextMapComplexUTF8[:]
|
---|
363 | return true
|
---|
364 | }
|
---|
365 | }
|
---|
366 | }
|
---|
367 |
|
---|
368 | func decideOverLiteralContextModeling(input []byte, start_pos uint, length uint, mask uint, quality int, size_hint uint, num_literal_contexts *uint, literal_context_map *[]uint32) {
|
---|
369 | if quality < minQualityForContextModeling || length < 64 {
|
---|
370 | return
|
---|
371 | } else if shouldUseComplexStaticContextMap(input, start_pos, length, mask, quality, size_hint, num_literal_contexts, literal_context_map) {
|
---|
372 | } else /* Context map was already set, nothing else to do. */
|
---|
373 | {
|
---|
374 | var end_pos uint = start_pos + length
|
---|
375 | /* Gather bi-gram data of the UTF8 byte prefixes. To make the analysis of
|
---|
376 | UTF8 data faster we only examine 64 byte long strides at every 4kB
|
---|
377 | intervals. */
|
---|
378 |
|
---|
379 | var bigram_prefix_histo = [9]uint32{0}
|
---|
380 | for ; start_pos+64 <= end_pos; start_pos += 4096 {
|
---|
381 | var lut = [4]int{0, 0, 1, 2}
|
---|
382 | var stride_end_pos uint = start_pos + 64
|
---|
383 | var prev int = lut[input[start_pos&mask]>>6] * 3
|
---|
384 | var pos uint
|
---|
385 | for pos = start_pos + 1; pos < stride_end_pos; pos++ {
|
---|
386 | var literal byte = input[pos&mask]
|
---|
387 | bigram_prefix_histo[prev+lut[literal>>6]]++
|
---|
388 | prev = lut[literal>>6] * 3
|
---|
389 | }
|
---|
390 | }
|
---|
391 |
|
---|
392 | chooseContextMap(quality, bigram_prefix_histo[0:], num_literal_contexts, literal_context_map)
|
---|
393 | }
|
---|
394 | }
|
---|
395 |
|
---|
396 | func shouldCompress_encode(data []byte, mask uint, last_flush_pos uint64, bytes uint, num_literals uint, num_commands uint) bool {
|
---|
397 | /* TODO: find more precise minimal block overhead. */
|
---|
398 | if bytes <= 2 {
|
---|
399 | return false
|
---|
400 | }
|
---|
401 | if num_commands < (bytes>>8)+2 {
|
---|
402 | if float64(num_literals) > 0.99*float64(bytes) {
|
---|
403 | var literal_histo = [256]uint32{0}
|
---|
404 | const kSampleRate uint32 = 13
|
---|
405 | const kMinEntropy float64 = 7.92
|
---|
406 | var bit_cost_threshold float64 = float64(bytes) * kMinEntropy / float64(kSampleRate)
|
---|
407 | var t uint = uint((uint32(bytes) + kSampleRate - 1) / kSampleRate)
|
---|
408 | var pos uint32 = uint32(last_flush_pos)
|
---|
409 | var i uint
|
---|
410 | for i = 0; i < t; i++ {
|
---|
411 | literal_histo[data[pos&uint32(mask)]]++
|
---|
412 | pos += kSampleRate
|
---|
413 | }
|
---|
414 |
|
---|
415 | if bitsEntropy(literal_histo[:], 256) > bit_cost_threshold {
|
---|
416 | return false
|
---|
417 | }
|
---|
418 | }
|
---|
419 | }
|
---|
420 |
|
---|
421 | return true
|
---|
422 | }
|
---|
423 |
|
---|
424 | /* Chooses the literal context mode for a metablock */
|
---|
425 | func chooseContextMode(params *encoderParams, data []byte, pos uint, mask uint, length uint) int {
|
---|
426 | /* We only do the computation for the option of something else than
|
---|
427 | CONTEXT_UTF8 for the highest qualities */
|
---|
428 | if params.quality >= minQualityForHqBlockSplitting && !isMostlyUTF8(data, pos, mask, length, kMinUTF8Ratio) {
|
---|
429 | return contextSigned
|
---|
430 | }
|
---|
431 |
|
---|
432 | return contextUTF8
|
---|
433 | }
|
---|
434 |
|
---|
435 | func writeMetaBlockInternal(data []byte, mask uint, last_flush_pos uint64, bytes uint, is_last bool, literal_context_mode int, params *encoderParams, prev_byte byte, prev_byte2 byte, num_literals uint, commands []command, saved_dist_cache []int, dist_cache []int, storage_ix *uint, storage []byte) {
|
---|
436 | var wrapped_last_flush_pos uint32 = wrapPosition(last_flush_pos)
|
---|
437 | var last_bytes uint16
|
---|
438 | var last_bytes_bits byte
|
---|
439 | var literal_context_lut contextLUT = getContextLUT(literal_context_mode)
|
---|
440 | var block_params encoderParams = *params
|
---|
441 |
|
---|
442 | if bytes == 0 {
|
---|
443 | /* Write the ISLAST and ISEMPTY bits. */
|
---|
444 | writeBits(2, 3, storage_ix, storage)
|
---|
445 |
|
---|
446 | *storage_ix = (*storage_ix + 7) &^ 7
|
---|
447 | return
|
---|
448 | }
|
---|
449 |
|
---|
450 | if !shouldCompress_encode(data, mask, last_flush_pos, bytes, num_literals, uint(len(commands))) {
|
---|
451 | /* Restore the distance cache, as its last update by
|
---|
452 | CreateBackwardReferences is now unused. */
|
---|
453 | copy(dist_cache, saved_dist_cache[:4])
|
---|
454 |
|
---|
455 | storeUncompressedMetaBlock(is_last, data, uint(wrapped_last_flush_pos), mask, bytes, storage_ix, storage)
|
---|
456 | return
|
---|
457 | }
|
---|
458 |
|
---|
459 | assert(*storage_ix <= 14)
|
---|
460 | last_bytes = uint16(storage[1])<<8 | uint16(storage[0])
|
---|
461 | last_bytes_bits = byte(*storage_ix)
|
---|
462 | if params.quality <= maxQualityForStaticEntropyCodes {
|
---|
463 | storeMetaBlockFast(data, uint(wrapped_last_flush_pos), bytes, mask, is_last, params, commands, storage_ix, storage)
|
---|
464 | } else if params.quality < minQualityForBlockSplit {
|
---|
465 | storeMetaBlockTrivial(data, uint(wrapped_last_flush_pos), bytes, mask, is_last, params, commands, storage_ix, storage)
|
---|
466 | } else {
|
---|
467 | mb := getMetaBlockSplit()
|
---|
468 | if params.quality < minQualityForHqBlockSplitting {
|
---|
469 | var num_literal_contexts uint = 1
|
---|
470 | var literal_context_map []uint32 = nil
|
---|
471 | if !params.disable_literal_context_modeling {
|
---|
472 | decideOverLiteralContextModeling(data, uint(wrapped_last_flush_pos), bytes, mask, params.quality, params.size_hint, &num_literal_contexts, &literal_context_map)
|
---|
473 | }
|
---|
474 |
|
---|
475 | buildMetaBlockGreedy(data, uint(wrapped_last_flush_pos), mask, prev_byte, prev_byte2, literal_context_lut, num_literal_contexts, literal_context_map, commands, mb)
|
---|
476 | } else {
|
---|
477 | buildMetaBlock(data, uint(wrapped_last_flush_pos), mask, &block_params, prev_byte, prev_byte2, commands, literal_context_mode, mb)
|
---|
478 | }
|
---|
479 |
|
---|
480 | if params.quality >= minQualityForOptimizeHistograms {
|
---|
481 | /* The number of distance symbols effectively used for distance
|
---|
482 | histograms. It might be less than distance alphabet size
|
---|
483 | for "Large Window Brotli" (32-bit). */
|
---|
484 | var num_effective_dist_codes uint32 = block_params.dist.alphabet_size
|
---|
485 | if num_effective_dist_codes > numHistogramDistanceSymbols {
|
---|
486 | num_effective_dist_codes = numHistogramDistanceSymbols
|
---|
487 | }
|
---|
488 |
|
---|
489 | optimizeHistograms(num_effective_dist_codes, mb)
|
---|
490 | }
|
---|
491 |
|
---|
492 | storeMetaBlock(data, uint(wrapped_last_flush_pos), bytes, mask, prev_byte, prev_byte2, is_last, &block_params, literal_context_mode, commands, mb, storage_ix, storage)
|
---|
493 | freeMetaBlockSplit(mb)
|
---|
494 | }
|
---|
495 |
|
---|
496 | if bytes+4 < *storage_ix>>3 {
|
---|
497 | /* Restore the distance cache and last byte. */
|
---|
498 | copy(dist_cache, saved_dist_cache[:4])
|
---|
499 |
|
---|
500 | storage[0] = byte(last_bytes)
|
---|
501 | storage[1] = byte(last_bytes >> 8)
|
---|
502 | *storage_ix = uint(last_bytes_bits)
|
---|
503 | storeUncompressedMetaBlock(is_last, data, uint(wrapped_last_flush_pos), mask, bytes, storage_ix, storage)
|
---|
504 | }
|
---|
505 | }
|
---|
506 |
|
---|
507 | func chooseDistanceParams(params *encoderParams) {
|
---|
508 | var distance_postfix_bits uint32 = 0
|
---|
509 | var num_direct_distance_codes uint32 = 0
|
---|
510 |
|
---|
511 | if params.quality >= minQualityForNonzeroDistanceParams {
|
---|
512 | var ndirect_msb uint32
|
---|
513 | if params.mode == modeFont {
|
---|
514 | distance_postfix_bits = 1
|
---|
515 | num_direct_distance_codes = 12
|
---|
516 | } else {
|
---|
517 | distance_postfix_bits = params.dist.distance_postfix_bits
|
---|
518 | num_direct_distance_codes = params.dist.num_direct_distance_codes
|
---|
519 | }
|
---|
520 |
|
---|
521 | ndirect_msb = (num_direct_distance_codes >> distance_postfix_bits) & 0x0F
|
---|
522 | if distance_postfix_bits > maxNpostfix || num_direct_distance_codes > maxNdirect || ndirect_msb<<distance_postfix_bits != num_direct_distance_codes {
|
---|
523 | distance_postfix_bits = 0
|
---|
524 | num_direct_distance_codes = 0
|
---|
525 | }
|
---|
526 | }
|
---|
527 |
|
---|
528 | initDistanceParams(params, distance_postfix_bits, num_direct_distance_codes)
|
---|
529 | }
|
---|
530 |
|
---|
531 | func ensureInitialized(s *Writer) bool {
|
---|
532 | if s.is_initialized_ {
|
---|
533 | return true
|
---|
534 | }
|
---|
535 |
|
---|
536 | s.last_bytes_bits_ = 0
|
---|
537 | s.last_bytes_ = 0
|
---|
538 | s.remaining_metadata_bytes_ = math.MaxUint32
|
---|
539 |
|
---|
540 | sanitizeParams(&s.params)
|
---|
541 | s.params.lgblock = computeLgBlock(&s.params)
|
---|
542 | chooseDistanceParams(&s.params)
|
---|
543 |
|
---|
544 | ringBufferSetup(&s.params, &s.ringbuffer_)
|
---|
545 |
|
---|
546 | /* Initialize last byte with stream header. */
|
---|
547 | {
|
---|
548 | var lgwin int = int(s.params.lgwin)
|
---|
549 | if s.params.quality == fastOnePassCompressionQuality || s.params.quality == fastTwoPassCompressionQuality {
|
---|
550 | lgwin = brotli_max_int(lgwin, 18)
|
---|
551 | }
|
---|
552 |
|
---|
553 | encodeWindowBits(lgwin, s.params.large_window, &s.last_bytes_, &s.last_bytes_bits_)
|
---|
554 | }
|
---|
555 |
|
---|
556 | if s.params.quality == fastOnePassCompressionQuality {
|
---|
557 | s.cmd_depths_ = [128]byte{
|
---|
558 | 0, 4, 4, 5, 6, 6, 7, 7, 7, 7, 7, 8, 8, 8, 8, 8,
|
---|
559 | 0, 0, 0, 4, 4, 4, 4, 4, 5, 5, 6, 6, 6, 6, 7, 7,
|
---|
560 | 7, 7, 10, 10, 10, 10, 10, 10, 0, 4, 4, 5, 5, 5, 6, 6,
|
---|
561 | 7, 8, 8, 9, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,
|
---|
562 | 5, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
|
---|
563 | 6, 6, 6, 6, 6, 6, 5, 5, 5, 5, 5, 5, 4, 4, 4, 4,
|
---|
564 | 4, 4, 4, 5, 5, 5, 5, 5, 5, 6, 6, 7, 7, 7, 8, 10,
|
---|
565 | 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12, 12,
|
---|
566 | }
|
---|
567 | s.cmd_bits_ = [128]uint16{
|
---|
568 | 0, 0, 8, 9, 3, 35, 7, 71,
|
---|
569 | 39, 103, 23, 47, 175, 111, 239, 31,
|
---|
570 | 0, 0, 0, 4, 12, 2, 10, 6,
|
---|
571 | 13, 29, 11, 43, 27, 59, 87, 55,
|
---|
572 | 15, 79, 319, 831, 191, 703, 447, 959,
|
---|
573 | 0, 14, 1, 25, 5, 21, 19, 51,
|
---|
574 | 119, 159, 95, 223, 479, 991, 63, 575,
|
---|
575 | 127, 639, 383, 895, 255, 767, 511, 1023,
|
---|
576 | 14, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
|
---|
577 | 27, 59, 7, 39, 23, 55, 30, 1, 17, 9, 25, 5, 0, 8, 4, 12,
|
---|
578 | 2, 10, 6, 21, 13, 29, 3, 19, 11, 15, 47, 31, 95, 63, 127, 255,
|
---|
579 | 767, 2815, 1791, 3839, 511, 2559, 1535, 3583, 1023, 3071, 2047, 4095,
|
---|
580 | }
|
---|
581 | s.cmd_code_ = [512]byte{
|
---|
582 | 0xff, 0x77, 0xd5, 0xbf, 0xe7, 0xde, 0xea, 0x9e, 0x51, 0x5d, 0xde, 0xc6,
|
---|
583 | 0x70, 0x57, 0xbc, 0x58, 0x58, 0x58, 0xd8, 0xd8, 0x58, 0xd5, 0xcb, 0x8c,
|
---|
584 | 0xea, 0xe0, 0xc3, 0x87, 0x1f, 0x83, 0xc1, 0x60, 0x1c, 0x67, 0xb2, 0xaa,
|
---|
585 | 0x06, 0x83, 0xc1, 0x60, 0x30, 0x18, 0xcc, 0xa1, 0xce, 0x88, 0x54, 0x94,
|
---|
586 | 0x46, 0xe1, 0xb0, 0xd0, 0x4e, 0xb2, 0xf7, 0x04, 0x00,
|
---|
587 | }
|
---|
588 | s.cmd_code_numbits_ = 448
|
---|
589 | }
|
---|
590 |
|
---|
591 | s.is_initialized_ = true
|
---|
592 | return true
|
---|
593 | }
|
---|
594 |
|
---|
595 | func encoderInitParams(params *encoderParams) {
|
---|
596 | params.mode = defaultMode
|
---|
597 | params.large_window = false
|
---|
598 | params.quality = defaultQuality
|
---|
599 | params.lgwin = defaultWindow
|
---|
600 | params.lgblock = 0
|
---|
601 | params.size_hint = 0
|
---|
602 | params.disable_literal_context_modeling = false
|
---|
603 | initEncoderDictionary(¶ms.dictionary)
|
---|
604 | params.dist.distance_postfix_bits = 0
|
---|
605 | params.dist.num_direct_distance_codes = 0
|
---|
606 | params.dist.alphabet_size = uint32(distanceAlphabetSize(0, 0, maxDistanceBits))
|
---|
607 | params.dist.max_distance = maxDistance
|
---|
608 | }
|
---|
609 |
|
---|
610 | func encoderInitState(s *Writer) {
|
---|
611 | encoderInitParams(&s.params)
|
---|
612 | s.input_pos_ = 0
|
---|
613 | s.commands = s.commands[:0]
|
---|
614 | s.num_literals_ = 0
|
---|
615 | s.last_insert_len_ = 0
|
---|
616 | s.last_flush_pos_ = 0
|
---|
617 | s.last_processed_pos_ = 0
|
---|
618 | s.prev_byte_ = 0
|
---|
619 | s.prev_byte2_ = 0
|
---|
620 | if s.hasher_ != nil {
|
---|
621 | s.hasher_.Common().is_prepared_ = false
|
---|
622 | }
|
---|
623 | s.cmd_code_numbits_ = 0
|
---|
624 | s.stream_state_ = streamProcessing
|
---|
625 | s.is_last_block_emitted_ = false
|
---|
626 | s.is_initialized_ = false
|
---|
627 |
|
---|
628 | ringBufferInit(&s.ringbuffer_)
|
---|
629 |
|
---|
630 | /* Initialize distance cache. */
|
---|
631 | s.dist_cache_[0] = 4
|
---|
632 |
|
---|
633 | s.dist_cache_[1] = 11
|
---|
634 | s.dist_cache_[2] = 15
|
---|
635 | s.dist_cache_[3] = 16
|
---|
636 |
|
---|
637 | /* Save the state of the distance cache in case we need to restore it for
|
---|
638 | emitting an uncompressed block. */
|
---|
639 | copy(s.saved_dist_cache_[:], s.dist_cache_[:])
|
---|
640 | }
|
---|
641 |
|
---|
642 | /*
|
---|
643 | Copies the given input data to the internal ring buffer of the compressor.
|
---|
644 | No processing of the data occurs at this time and this function can be
|
---|
645 | called multiple times before calling WriteBrotliData() to process the
|
---|
646 | accumulated input. At most input_block_size() bytes of input data can be
|
---|
647 | copied to the ring buffer, otherwise the next WriteBrotliData() will fail.
|
---|
648 | */
|
---|
649 | func copyInputToRingBuffer(s *Writer, input_size uint, input_buffer []byte) {
|
---|
650 | var ringbuffer_ *ringBuffer = &s.ringbuffer_
|
---|
651 | ringBufferWrite(input_buffer, input_size, ringbuffer_)
|
---|
652 | s.input_pos_ += uint64(input_size)
|
---|
653 |
|
---|
654 | /* TL;DR: If needed, initialize 7 more bytes in the ring buffer to make the
|
---|
655 | hashing not depend on uninitialized data. This makes compression
|
---|
656 | deterministic and it prevents uninitialized memory warnings in Valgrind.
|
---|
657 | Even without erasing, the output would be valid (but nondeterministic).
|
---|
658 |
|
---|
659 | Background information: The compressor stores short (at most 8 bytes)
|
---|
660 | substrings of the input already read in a hash table, and detects
|
---|
661 | repetitions by looking up such substrings in the hash table. If it
|
---|
662 | can find a substring, it checks whether the substring is really there
|
---|
663 | in the ring buffer (or it's just a hash collision). Should the hash
|
---|
664 | table become corrupt, this check makes sure that the output is
|
---|
665 | still valid, albeit the compression ratio would be bad.
|
---|
666 |
|
---|
667 | The compressor populates the hash table from the ring buffer as it's
|
---|
668 | reading new bytes from the input. However, at the last few indexes of
|
---|
669 | the ring buffer, there are not enough bytes to build full-length
|
---|
670 | substrings from. Since the hash table always contains full-length
|
---|
671 | substrings, we erase with dummy zeros here to make sure that those
|
---|
672 | substrings will contain zeros at the end instead of uninitialized
|
---|
673 | data.
|
---|
674 |
|
---|
675 | Please note that erasing is not necessary (because the
|
---|
676 | memory region is already initialized since he ring buffer
|
---|
677 | has a `tail' that holds a copy of the beginning,) so we
|
---|
678 | skip erasing if we have already gone around at least once in
|
---|
679 | the ring buffer.
|
---|
680 |
|
---|
681 | Only clear during the first round of ring-buffer writes. On
|
---|
682 | subsequent rounds data in the ring-buffer would be affected. */
|
---|
683 | if ringbuffer_.pos_ <= ringbuffer_.mask_ {
|
---|
684 | /* This is the first time when the ring buffer is being written.
|
---|
685 | We clear 7 bytes just after the bytes that have been copied from
|
---|
686 | the input buffer.
|
---|
687 |
|
---|
688 | The ring-buffer has a "tail" that holds a copy of the beginning,
|
---|
689 | but only once the ring buffer has been fully written once, i.e.,
|
---|
690 | pos <= mask. For the first time, we need to write values
|
---|
691 | in this tail (where index may be larger than mask), so that
|
---|
692 | we have exactly defined behavior and don't read uninitialized
|
---|
693 | memory. Due to performance reasons, hashing reads data using a
|
---|
694 | LOAD64, which can go 7 bytes beyond the bytes written in the
|
---|
695 | ring-buffer. */
|
---|
696 | for i := 0; i < int(7); i++ {
|
---|
697 | ringbuffer_.buffer_[ringbuffer_.pos_:][i] = 0
|
---|
698 | }
|
---|
699 | }
|
---|
700 | }
|
---|
701 |
|
---|
702 | /* Marks all input as processed.
|
---|
703 | Returns true if position wrapping occurs. */
|
---|
704 | func updateLastProcessedPos(s *Writer) bool {
|
---|
705 | var wrapped_last_processed_pos uint32 = wrapPosition(s.last_processed_pos_)
|
---|
706 | var wrapped_input_pos uint32 = wrapPosition(s.input_pos_)
|
---|
707 | s.last_processed_pos_ = s.input_pos_
|
---|
708 | return wrapped_input_pos < wrapped_last_processed_pos
|
---|
709 | }
|
---|
710 |
|
---|
711 | func extendLastCommand(s *Writer, bytes *uint32, wrapped_last_processed_pos *uint32) {
|
---|
712 | var last_command *command = &s.commands[len(s.commands)-1]
|
---|
713 | var data []byte = s.ringbuffer_.buffer_
|
---|
714 | var mask uint32 = s.ringbuffer_.mask_
|
---|
715 | var max_backward_distance uint64 = ((uint64(1)) << s.params.lgwin) - windowGap
|
---|
716 | var last_copy_len uint64 = uint64(last_command.copy_len_) & 0x1FFFFFF
|
---|
717 | var last_processed_pos uint64 = s.last_processed_pos_ - last_copy_len
|
---|
718 | var max_distance uint64
|
---|
719 | if last_processed_pos < max_backward_distance {
|
---|
720 | max_distance = last_processed_pos
|
---|
721 | } else {
|
---|
722 | max_distance = max_backward_distance
|
---|
723 | }
|
---|
724 | var cmd_dist uint64 = uint64(s.dist_cache_[0])
|
---|
725 | var distance_code uint32 = commandRestoreDistanceCode(last_command, &s.params.dist)
|
---|
726 | if distance_code < numDistanceShortCodes || uint64(distance_code-(numDistanceShortCodes-1)) == cmd_dist {
|
---|
727 | if cmd_dist <= max_distance {
|
---|
728 | for *bytes != 0 && data[*wrapped_last_processed_pos&mask] == data[(uint64(*wrapped_last_processed_pos)-cmd_dist)&uint64(mask)] {
|
---|
729 | last_command.copy_len_++
|
---|
730 | (*bytes)--
|
---|
731 | (*wrapped_last_processed_pos)++
|
---|
732 | }
|
---|
733 | }
|
---|
734 |
|
---|
735 | /* The copy length is at most the metablock size, and thus expressible. */
|
---|
736 | getLengthCode(uint(last_command.insert_len_), uint(int(last_command.copy_len_&0x1FFFFFF)+int(last_command.copy_len_>>25)), (last_command.dist_prefix_&0x3FF == 0), &last_command.cmd_prefix_)
|
---|
737 | }
|
---|
738 | }
|
---|
739 |
|
---|
740 | /*
|
---|
741 | Processes the accumulated input data and writes
|
---|
742 | the new output meta-block to s.dest, if one has been
|
---|
743 | created (otherwise the processed input data is buffered internally).
|
---|
744 | If |is_last| or |force_flush| is true, an output meta-block is
|
---|
745 | always created. However, until |is_last| is true encoder may retain up
|
---|
746 | to 7 bits of the last byte of output. To force encoder to dump the remaining
|
---|
747 | bits use WriteMetadata() to append an empty meta-data block.
|
---|
748 | Returns false if the size of the input data is larger than
|
---|
749 | input_block_size().
|
---|
750 | */
|
---|
751 | func encodeData(s *Writer, is_last bool, force_flush bool) bool {
|
---|
752 | var delta uint64 = unprocessedInputSize(s)
|
---|
753 | var bytes uint32 = uint32(delta)
|
---|
754 | var wrapped_last_processed_pos uint32 = wrapPosition(s.last_processed_pos_)
|
---|
755 | var data []byte
|
---|
756 | var mask uint32
|
---|
757 | var literal_context_mode int
|
---|
758 |
|
---|
759 | data = s.ringbuffer_.buffer_
|
---|
760 | mask = s.ringbuffer_.mask_
|
---|
761 |
|
---|
762 | /* Adding more blocks after "last" block is forbidden. */
|
---|
763 | if s.is_last_block_emitted_ {
|
---|
764 | return false
|
---|
765 | }
|
---|
766 | if is_last {
|
---|
767 | s.is_last_block_emitted_ = true
|
---|
768 | }
|
---|
769 |
|
---|
770 | if delta > uint64(inputBlockSize(s)) {
|
---|
771 | return false
|
---|
772 | }
|
---|
773 |
|
---|
774 | if s.params.quality == fastTwoPassCompressionQuality {
|
---|
775 | if s.command_buf_ == nil || cap(s.command_buf_) < int(kCompressFragmentTwoPassBlockSize) {
|
---|
776 | s.command_buf_ = make([]uint32, kCompressFragmentTwoPassBlockSize)
|
---|
777 | s.literal_buf_ = make([]byte, kCompressFragmentTwoPassBlockSize)
|
---|
778 | } else {
|
---|
779 | s.command_buf_ = s.command_buf_[:kCompressFragmentTwoPassBlockSize]
|
---|
780 | s.literal_buf_ = s.literal_buf_[:kCompressFragmentTwoPassBlockSize]
|
---|
781 | }
|
---|
782 | }
|
---|
783 |
|
---|
784 | if s.params.quality == fastOnePassCompressionQuality || s.params.quality == fastTwoPassCompressionQuality {
|
---|
785 | var storage []byte
|
---|
786 | var storage_ix uint = uint(s.last_bytes_bits_)
|
---|
787 | var table_size uint
|
---|
788 | var table []int
|
---|
789 |
|
---|
790 | if delta == 0 && !is_last {
|
---|
791 | /* We have no new input data and we don't have to finish the stream, so
|
---|
792 | nothing to do. */
|
---|
793 | return true
|
---|
794 | }
|
---|
795 |
|
---|
796 | storage = s.getStorage(int(2*bytes + 503))
|
---|
797 | storage[0] = byte(s.last_bytes_)
|
---|
798 | storage[1] = byte(s.last_bytes_ >> 8)
|
---|
799 | table = getHashTable(s, s.params.quality, uint(bytes), &table_size)
|
---|
800 | if s.params.quality == fastOnePassCompressionQuality {
|
---|
801 | compressFragmentFast(data[wrapped_last_processed_pos&mask:], uint(bytes), is_last, table, table_size, s.cmd_depths_[:], s.cmd_bits_[:], &s.cmd_code_numbits_, s.cmd_code_[:], &storage_ix, storage)
|
---|
802 | } else {
|
---|
803 | compressFragmentTwoPass(data[wrapped_last_processed_pos&mask:], uint(bytes), is_last, s.command_buf_, s.literal_buf_, table, table_size, &storage_ix, storage)
|
---|
804 | }
|
---|
805 |
|
---|
806 | s.last_bytes_ = uint16(storage[storage_ix>>3])
|
---|
807 | s.last_bytes_bits_ = byte(storage_ix & 7)
|
---|
808 | updateLastProcessedPos(s)
|
---|
809 | s.writeOutput(storage[:storage_ix>>3])
|
---|
810 | return true
|
---|
811 | }
|
---|
812 | {
|
---|
813 | /* Theoretical max number of commands is 1 per 2 bytes. */
|
---|
814 | newsize := len(s.commands) + int(bytes)/2 + 1
|
---|
815 | if newsize > cap(s.commands) {
|
---|
816 | /* Reserve a bit more memory to allow merging with a next block
|
---|
817 | without reallocation: that would impact speed. */
|
---|
818 | newsize += int(bytes/4) + 16
|
---|
819 |
|
---|
820 | new_commands := make([]command, len(s.commands), newsize)
|
---|
821 | if s.commands != nil {
|
---|
822 | copy(new_commands, s.commands)
|
---|
823 | }
|
---|
824 |
|
---|
825 | s.commands = new_commands
|
---|
826 | }
|
---|
827 | }
|
---|
828 |
|
---|
829 | initOrStitchToPreviousBlock(&s.hasher_, data, uint(mask), &s.params, uint(wrapped_last_processed_pos), uint(bytes), is_last)
|
---|
830 |
|
---|
831 | literal_context_mode = chooseContextMode(&s.params, data, uint(wrapPosition(s.last_flush_pos_)), uint(mask), uint(s.input_pos_-s.last_flush_pos_))
|
---|
832 |
|
---|
833 | if len(s.commands) != 0 && s.last_insert_len_ == 0 {
|
---|
834 | extendLastCommand(s, &bytes, &wrapped_last_processed_pos)
|
---|
835 | }
|
---|
836 |
|
---|
837 | if s.params.quality == zopflificationQuality {
|
---|
838 | assert(s.params.hasher.type_ == 10)
|
---|
839 | createZopfliBackwardReferences(uint(bytes), uint(wrapped_last_processed_pos), data, uint(mask), &s.params, s.hasher_.(*h10), s.dist_cache_[:], &s.last_insert_len_, &s.commands, &s.num_literals_)
|
---|
840 | } else if s.params.quality == hqZopflificationQuality {
|
---|
841 | assert(s.params.hasher.type_ == 10)
|
---|
842 | createHqZopfliBackwardReferences(uint(bytes), uint(wrapped_last_processed_pos), data, uint(mask), &s.params, s.hasher_, s.dist_cache_[:], &s.last_insert_len_, &s.commands, &s.num_literals_)
|
---|
843 | } else {
|
---|
844 | createBackwardReferences(uint(bytes), uint(wrapped_last_processed_pos), data, uint(mask), &s.params, s.hasher_, s.dist_cache_[:], &s.last_insert_len_, &s.commands, &s.num_literals_)
|
---|
845 | }
|
---|
846 | {
|
---|
847 | var max_length uint = maxMetablockSize(&s.params)
|
---|
848 | var max_literals uint = max_length / 8
|
---|
849 | max_commands := int(max_length / 8)
|
---|
850 | var processed_bytes uint = uint(s.input_pos_ - s.last_flush_pos_)
|
---|
851 | var next_input_fits_metablock bool = (processed_bytes+inputBlockSize(s) <= max_length)
|
---|
852 | var should_flush bool = (s.params.quality < minQualityForBlockSplit && s.num_literals_+uint(len(s.commands)) >= maxNumDelayedSymbols)
|
---|
853 | /* If maximal possible additional block doesn't fit metablock, flush now. */
|
---|
854 | /* TODO: Postpone decision until next block arrives? */
|
---|
855 |
|
---|
856 | /* If block splitting is not used, then flush as soon as there is some
|
---|
857 | amount of commands / literals produced. */
|
---|
858 | if !is_last && !force_flush && !should_flush && next_input_fits_metablock && s.num_literals_ < max_literals && len(s.commands) < max_commands {
|
---|
859 | /* Merge with next input block. Everything will happen later. */
|
---|
860 | if updateLastProcessedPos(s) {
|
---|
861 | hasherReset(s.hasher_)
|
---|
862 | }
|
---|
863 |
|
---|
864 | return true
|
---|
865 | }
|
---|
866 | }
|
---|
867 |
|
---|
868 | /* Create the last insert-only command. */
|
---|
869 | if s.last_insert_len_ > 0 {
|
---|
870 | s.commands = append(s.commands, makeInsertCommand(s.last_insert_len_))
|
---|
871 | s.num_literals_ += s.last_insert_len_
|
---|
872 | s.last_insert_len_ = 0
|
---|
873 | }
|
---|
874 |
|
---|
875 | if !is_last && s.input_pos_ == s.last_flush_pos_ {
|
---|
876 | /* We have no new input data and we don't have to finish the stream, so
|
---|
877 | nothing to do. */
|
---|
878 | return true
|
---|
879 | }
|
---|
880 |
|
---|
881 | assert(s.input_pos_ >= s.last_flush_pos_)
|
---|
882 | assert(s.input_pos_ > s.last_flush_pos_ || is_last)
|
---|
883 | assert(s.input_pos_-s.last_flush_pos_ <= 1<<24)
|
---|
884 | {
|
---|
885 | var metablock_size uint32 = uint32(s.input_pos_ - s.last_flush_pos_)
|
---|
886 | var storage []byte = s.getStorage(int(2*metablock_size + 503))
|
---|
887 | var storage_ix uint = uint(s.last_bytes_bits_)
|
---|
888 | storage[0] = byte(s.last_bytes_)
|
---|
889 | storage[1] = byte(s.last_bytes_ >> 8)
|
---|
890 | writeMetaBlockInternal(data, uint(mask), s.last_flush_pos_, uint(metablock_size), is_last, literal_context_mode, &s.params, s.prev_byte_, s.prev_byte2_, s.num_literals_, s.commands, s.saved_dist_cache_[:], s.dist_cache_[:], &storage_ix, storage)
|
---|
891 | s.last_bytes_ = uint16(storage[storage_ix>>3])
|
---|
892 | s.last_bytes_bits_ = byte(storage_ix & 7)
|
---|
893 | s.last_flush_pos_ = s.input_pos_
|
---|
894 | if updateLastProcessedPos(s) {
|
---|
895 | hasherReset(s.hasher_)
|
---|
896 | }
|
---|
897 |
|
---|
898 | if s.last_flush_pos_ > 0 {
|
---|
899 | s.prev_byte_ = data[(uint32(s.last_flush_pos_)-1)&mask]
|
---|
900 | }
|
---|
901 |
|
---|
902 | if s.last_flush_pos_ > 1 {
|
---|
903 | s.prev_byte2_ = data[uint32(s.last_flush_pos_-2)&mask]
|
---|
904 | }
|
---|
905 |
|
---|
906 | s.commands = s.commands[:0]
|
---|
907 | s.num_literals_ = 0
|
---|
908 |
|
---|
909 | /* Save the state of the distance cache in case we need to restore it for
|
---|
910 | emitting an uncompressed block. */
|
---|
911 | copy(s.saved_dist_cache_[:], s.dist_cache_[:])
|
---|
912 |
|
---|
913 | s.writeOutput(storage[:storage_ix>>3])
|
---|
914 | return true
|
---|
915 | }
|
---|
916 | }
|
---|
917 |
|
---|
918 | /* Dumps remaining output bits and metadata header to |header|.
|
---|
919 | Returns number of produced bytes.
|
---|
920 | REQUIRED: |header| should be 8-byte aligned and at least 16 bytes long.
|
---|
921 | REQUIRED: |block_size| <= (1 << 24). */
|
---|
922 | func writeMetadataHeader(s *Writer, block_size uint, header []byte) uint {
|
---|
923 | storage_ix := uint(s.last_bytes_bits_)
|
---|
924 | header[0] = byte(s.last_bytes_)
|
---|
925 | header[1] = byte(s.last_bytes_ >> 8)
|
---|
926 | s.last_bytes_ = 0
|
---|
927 | s.last_bytes_bits_ = 0
|
---|
928 |
|
---|
929 | writeBits(1, 0, &storage_ix, header)
|
---|
930 | writeBits(2, 3, &storage_ix, header)
|
---|
931 | writeBits(1, 0, &storage_ix, header)
|
---|
932 | if block_size == 0 {
|
---|
933 | writeBits(2, 0, &storage_ix, header)
|
---|
934 | } else {
|
---|
935 | var nbits uint32
|
---|
936 | if block_size == 1 {
|
---|
937 | nbits = 0
|
---|
938 | } else {
|
---|
939 | nbits = log2FloorNonZero(uint(uint32(block_size)-1)) + 1
|
---|
940 | }
|
---|
941 | var nbytes uint32 = (nbits + 7) / 8
|
---|
942 | writeBits(2, uint64(nbytes), &storage_ix, header)
|
---|
943 | writeBits(uint(8*nbytes), uint64(block_size)-1, &storage_ix, header)
|
---|
944 | }
|
---|
945 |
|
---|
946 | return (storage_ix + 7) >> 3
|
---|
947 | }
|
---|
948 |
|
---|
949 | func injectBytePaddingBlock(s *Writer) {
|
---|
950 | var seal uint32 = uint32(s.last_bytes_)
|
---|
951 | var seal_bits uint = uint(s.last_bytes_bits_)
|
---|
952 | s.last_bytes_ = 0
|
---|
953 | s.last_bytes_bits_ = 0
|
---|
954 |
|
---|
955 | /* is_last = 0, data_nibbles = 11, reserved = 0, meta_nibbles = 00 */
|
---|
956 | seal |= 0x6 << seal_bits
|
---|
957 |
|
---|
958 | seal_bits += 6
|
---|
959 |
|
---|
960 | destination := s.tiny_buf_.u8[:]
|
---|
961 |
|
---|
962 | destination[0] = byte(seal)
|
---|
963 | if seal_bits > 8 {
|
---|
964 | destination[1] = byte(seal >> 8)
|
---|
965 | }
|
---|
966 | if seal_bits > 16 {
|
---|
967 | destination[2] = byte(seal >> 16)
|
---|
968 | }
|
---|
969 | s.writeOutput(destination[:(seal_bits+7)>>3])
|
---|
970 | }
|
---|
971 |
|
---|
972 | func checkFlushComplete(s *Writer) {
|
---|
973 | if s.stream_state_ == streamFlushRequested && s.err == nil {
|
---|
974 | s.stream_state_ = streamProcessing
|
---|
975 | }
|
---|
976 | }
|
---|
977 |
|
---|
978 | func encoderCompressStreamFast(s *Writer, op int, available_in *uint, next_in *[]byte) bool {
|
---|
979 | var block_size_limit uint = uint(1) << s.params.lgwin
|
---|
980 | var buf_size uint = brotli_min_size_t(kCompressFragmentTwoPassBlockSize, brotli_min_size_t(*available_in, block_size_limit))
|
---|
981 | var command_buf []uint32 = nil
|
---|
982 | var literal_buf []byte = nil
|
---|
983 | if s.params.quality != fastOnePassCompressionQuality && s.params.quality != fastTwoPassCompressionQuality {
|
---|
984 | return false
|
---|
985 | }
|
---|
986 |
|
---|
987 | if s.params.quality == fastTwoPassCompressionQuality {
|
---|
988 | if s.command_buf_ == nil || cap(s.command_buf_) < int(buf_size) {
|
---|
989 | s.command_buf_ = make([]uint32, buf_size)
|
---|
990 | s.literal_buf_ = make([]byte, buf_size)
|
---|
991 | } else {
|
---|
992 | s.command_buf_ = s.command_buf_[:buf_size]
|
---|
993 | s.literal_buf_ = s.literal_buf_[:buf_size]
|
---|
994 | }
|
---|
995 |
|
---|
996 | command_buf = s.command_buf_
|
---|
997 | literal_buf = s.literal_buf_
|
---|
998 | }
|
---|
999 |
|
---|
1000 | for {
|
---|
1001 | if s.stream_state_ == streamFlushRequested && s.last_bytes_bits_ != 0 {
|
---|
1002 | injectBytePaddingBlock(s)
|
---|
1003 | continue
|
---|
1004 | }
|
---|
1005 |
|
---|
1006 | /* Compress block only when stream is not
|
---|
1007 | finished, there is no pending flush request, and there is either
|
---|
1008 | additional input or pending operation. */
|
---|
1009 | if s.stream_state_ == streamProcessing && (*available_in != 0 || op != int(operationProcess)) {
|
---|
1010 | var block_size uint = brotli_min_size_t(block_size_limit, *available_in)
|
---|
1011 | var is_last bool = (*available_in == block_size) && (op == int(operationFinish))
|
---|
1012 | var force_flush bool = (*available_in == block_size) && (op == int(operationFlush))
|
---|
1013 | var max_out_size uint = 2*block_size + 503
|
---|
1014 | var storage []byte = nil
|
---|
1015 | var storage_ix uint = uint(s.last_bytes_bits_)
|
---|
1016 | var table_size uint
|
---|
1017 | var table []int
|
---|
1018 |
|
---|
1019 | if force_flush && block_size == 0 {
|
---|
1020 | s.stream_state_ = streamFlushRequested
|
---|
1021 | continue
|
---|
1022 | }
|
---|
1023 |
|
---|
1024 | storage = s.getStorage(int(max_out_size))
|
---|
1025 |
|
---|
1026 | storage[0] = byte(s.last_bytes_)
|
---|
1027 | storage[1] = byte(s.last_bytes_ >> 8)
|
---|
1028 | table = getHashTable(s, s.params.quality, block_size, &table_size)
|
---|
1029 |
|
---|
1030 | if s.params.quality == fastOnePassCompressionQuality {
|
---|
1031 | compressFragmentFast(*next_in, block_size, is_last, table, table_size, s.cmd_depths_[:], s.cmd_bits_[:], &s.cmd_code_numbits_, s.cmd_code_[:], &storage_ix, storage)
|
---|
1032 | } else {
|
---|
1033 | compressFragmentTwoPass(*next_in, block_size, is_last, command_buf, literal_buf, table, table_size, &storage_ix, storage)
|
---|
1034 | }
|
---|
1035 |
|
---|
1036 | *next_in = (*next_in)[block_size:]
|
---|
1037 | *available_in -= block_size
|
---|
1038 | var out_bytes uint = storage_ix >> 3
|
---|
1039 | s.writeOutput(storage[:out_bytes])
|
---|
1040 |
|
---|
1041 | s.last_bytes_ = uint16(storage[storage_ix>>3])
|
---|
1042 | s.last_bytes_bits_ = byte(storage_ix & 7)
|
---|
1043 |
|
---|
1044 | if force_flush {
|
---|
1045 | s.stream_state_ = streamFlushRequested
|
---|
1046 | }
|
---|
1047 | if is_last {
|
---|
1048 | s.stream_state_ = streamFinished
|
---|
1049 | }
|
---|
1050 | continue
|
---|
1051 | }
|
---|
1052 |
|
---|
1053 | break
|
---|
1054 | }
|
---|
1055 |
|
---|
1056 | checkFlushComplete(s)
|
---|
1057 | return true
|
---|
1058 | }
|
---|
1059 |
|
---|
1060 | func processMetadata(s *Writer, available_in *uint, next_in *[]byte) bool {
|
---|
1061 | if *available_in > 1<<24 {
|
---|
1062 | return false
|
---|
1063 | }
|
---|
1064 |
|
---|
1065 | /* Switch to metadata block workflow, if required. */
|
---|
1066 | if s.stream_state_ == streamProcessing {
|
---|
1067 | s.remaining_metadata_bytes_ = uint32(*available_in)
|
---|
1068 | s.stream_state_ = streamMetadataHead
|
---|
1069 | }
|
---|
1070 |
|
---|
1071 | if s.stream_state_ != streamMetadataHead && s.stream_state_ != streamMetadataBody {
|
---|
1072 | return false
|
---|
1073 | }
|
---|
1074 |
|
---|
1075 | for {
|
---|
1076 | if s.stream_state_ == streamFlushRequested && s.last_bytes_bits_ != 0 {
|
---|
1077 | injectBytePaddingBlock(s)
|
---|
1078 | continue
|
---|
1079 | }
|
---|
1080 |
|
---|
1081 | if s.input_pos_ != s.last_flush_pos_ {
|
---|
1082 | var result bool = encodeData(s, false, true)
|
---|
1083 | if !result {
|
---|
1084 | return false
|
---|
1085 | }
|
---|
1086 | continue
|
---|
1087 | }
|
---|
1088 |
|
---|
1089 | if s.stream_state_ == streamMetadataHead {
|
---|
1090 | n := writeMetadataHeader(s, uint(s.remaining_metadata_bytes_), s.tiny_buf_.u8[:])
|
---|
1091 | s.writeOutput(s.tiny_buf_.u8[:n])
|
---|
1092 | s.stream_state_ = streamMetadataBody
|
---|
1093 | continue
|
---|
1094 | } else {
|
---|
1095 | /* Exit workflow only when there is no more input and no more output.
|
---|
1096 | Otherwise client may continue producing empty metadata blocks. */
|
---|
1097 | if s.remaining_metadata_bytes_ == 0 {
|
---|
1098 | s.remaining_metadata_bytes_ = math.MaxUint32
|
---|
1099 | s.stream_state_ = streamProcessing
|
---|
1100 | break
|
---|
1101 | }
|
---|
1102 |
|
---|
1103 | /* This guarantees progress in "TakeOutput" workflow. */
|
---|
1104 | var c uint32 = brotli_min_uint32_t(s.remaining_metadata_bytes_, 16)
|
---|
1105 | copy(s.tiny_buf_.u8[:], (*next_in)[:c])
|
---|
1106 | *next_in = (*next_in)[c:]
|
---|
1107 | *available_in -= uint(c)
|
---|
1108 | s.remaining_metadata_bytes_ -= c
|
---|
1109 | s.writeOutput(s.tiny_buf_.u8[:c])
|
---|
1110 |
|
---|
1111 | continue
|
---|
1112 | }
|
---|
1113 | }
|
---|
1114 |
|
---|
1115 | return true
|
---|
1116 | }
|
---|
1117 |
|
---|
1118 | func updateSizeHint(s *Writer, available_in uint) {
|
---|
1119 | if s.params.size_hint == 0 {
|
---|
1120 | var delta uint64 = unprocessedInputSize(s)
|
---|
1121 | var tail uint64 = uint64(available_in)
|
---|
1122 | var limit uint32 = 1 << 30
|
---|
1123 | var total uint32
|
---|
1124 | if (delta >= uint64(limit)) || (tail >= uint64(limit)) || ((delta + tail) >= uint64(limit)) {
|
---|
1125 | total = limit
|
---|
1126 | } else {
|
---|
1127 | total = uint32(delta + tail)
|
---|
1128 | }
|
---|
1129 |
|
---|
1130 | s.params.size_hint = uint(total)
|
---|
1131 | }
|
---|
1132 | }
|
---|
1133 |
|
---|
1134 | func encoderCompressStream(s *Writer, op int, available_in *uint, next_in *[]byte) bool {
|
---|
1135 | if !ensureInitialized(s) {
|
---|
1136 | return false
|
---|
1137 | }
|
---|
1138 |
|
---|
1139 | /* Unfinished metadata block; check requirements. */
|
---|
1140 | if s.remaining_metadata_bytes_ != math.MaxUint32 {
|
---|
1141 | if uint32(*available_in) != s.remaining_metadata_bytes_ {
|
---|
1142 | return false
|
---|
1143 | }
|
---|
1144 | if op != int(operationEmitMetadata) {
|
---|
1145 | return false
|
---|
1146 | }
|
---|
1147 | }
|
---|
1148 |
|
---|
1149 | if op == int(operationEmitMetadata) {
|
---|
1150 | updateSizeHint(s, 0) /* First data metablock might be emitted here. */
|
---|
1151 | return processMetadata(s, available_in, next_in)
|
---|
1152 | }
|
---|
1153 |
|
---|
1154 | if s.stream_state_ == streamMetadataHead || s.stream_state_ == streamMetadataBody {
|
---|
1155 | return false
|
---|
1156 | }
|
---|
1157 |
|
---|
1158 | if s.stream_state_ != streamProcessing && *available_in != 0 {
|
---|
1159 | return false
|
---|
1160 | }
|
---|
1161 |
|
---|
1162 | if s.params.quality == fastOnePassCompressionQuality || s.params.quality == fastTwoPassCompressionQuality {
|
---|
1163 | return encoderCompressStreamFast(s, op, available_in, next_in)
|
---|
1164 | }
|
---|
1165 |
|
---|
1166 | for {
|
---|
1167 | var remaining_block_size uint = remainingInputBlockSize(s)
|
---|
1168 |
|
---|
1169 | if remaining_block_size != 0 && *available_in != 0 {
|
---|
1170 | var copy_input_size uint = brotli_min_size_t(remaining_block_size, *available_in)
|
---|
1171 | copyInputToRingBuffer(s, copy_input_size, *next_in)
|
---|
1172 | *next_in = (*next_in)[copy_input_size:]
|
---|
1173 | *available_in -= copy_input_size
|
---|
1174 | continue
|
---|
1175 | }
|
---|
1176 |
|
---|
1177 | if s.stream_state_ == streamFlushRequested && s.last_bytes_bits_ != 0 {
|
---|
1178 | injectBytePaddingBlock(s)
|
---|
1179 | continue
|
---|
1180 | }
|
---|
1181 |
|
---|
1182 | /* Compress data only when stream is not
|
---|
1183 | finished and there is no pending flush request. */
|
---|
1184 | if s.stream_state_ == streamProcessing {
|
---|
1185 | if remaining_block_size == 0 || op != int(operationProcess) {
|
---|
1186 | var is_last bool = ((*available_in == 0) && op == int(operationFinish))
|
---|
1187 | var force_flush bool = ((*available_in == 0) && op == int(operationFlush))
|
---|
1188 | var result bool
|
---|
1189 | updateSizeHint(s, *available_in)
|
---|
1190 | result = encodeData(s, is_last, force_flush)
|
---|
1191 | if !result {
|
---|
1192 | return false
|
---|
1193 | }
|
---|
1194 | if force_flush {
|
---|
1195 | s.stream_state_ = streamFlushRequested
|
---|
1196 | }
|
---|
1197 | if is_last {
|
---|
1198 | s.stream_state_ = streamFinished
|
---|
1199 | }
|
---|
1200 | continue
|
---|
1201 | }
|
---|
1202 | }
|
---|
1203 |
|
---|
1204 | break
|
---|
1205 | }
|
---|
1206 |
|
---|
1207 | checkFlushComplete(s)
|
---|
1208 | return true
|
---|
1209 | }
|
---|
1210 |
|
---|
1211 | func (w *Writer) writeOutput(data []byte) {
|
---|
1212 | if w.err != nil {
|
---|
1213 | return
|
---|
1214 | }
|
---|
1215 |
|
---|
1216 | _, w.err = w.dst.Write(data)
|
---|
1217 | if w.err == nil {
|
---|
1218 | checkFlushComplete(w)
|
---|
1219 | }
|
---|
1220 | }
|
---|