brotli 0.1.1 → 0.1.2
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/ext/brotli/brotli.cc +114 -24
- data/ext/brotli/brotli.h +0 -1
- data/ext/brotli/extconf.rb +30 -23
- data/lib/brotli/version.rb +1 -1
- data/vendor/brotli/LICENSE +1 -1
- data/vendor/brotli/dec/Makefile +1 -1
- data/vendor/brotli/dec/bit_reader.c +3 -3
- data/vendor/brotli/dec/bit_reader.h +25 -27
- data/vendor/brotli/dec/context.h +4 -4
- data/vendor/brotli/dec/decode.c +410 -486
- data/vendor/brotli/dec/decode.h +101 -105
- data/vendor/brotli/dec/dictionary.c +1 -1
- data/vendor/brotli/dec/dictionary.h +7 -8
- data/vendor/brotli/dec/huffman.c +103 -105
- data/vendor/brotli/dec/huffman.h +18 -18
- data/vendor/brotli/dec/port.h +52 -40
- data/vendor/brotli/dec/prefix.h +2 -0
- data/vendor/brotli/dec/state.c +13 -19
- data/vendor/brotli/dec/state.h +25 -39
- data/vendor/brotli/dec/transform.h +38 -44
- data/vendor/brotli/dec/types.h +2 -2
- data/vendor/brotli/enc/Makefile +1 -1
- data/vendor/brotli/enc/backward_references.cc +455 -359
- data/vendor/brotli/enc/backward_references.h +79 -3
- data/vendor/brotli/enc/bit_cost.h +54 -32
- data/vendor/brotli/enc/block_splitter.cc +285 -193
- data/vendor/brotli/enc/block_splitter.h +4 -12
- data/vendor/brotli/enc/brotli_bit_stream.cc +623 -324
- data/vendor/brotli/enc/brotli_bit_stream.h +76 -37
- data/vendor/brotli/enc/cluster.h +161 -120
- data/vendor/brotli/enc/command.h +60 -37
- data/vendor/brotli/enc/compress_fragment.cc +701 -0
- data/vendor/brotli/enc/compress_fragment.h +47 -0
- data/vendor/brotli/enc/compress_fragment_two_pass.cc +524 -0
- data/vendor/brotli/enc/compress_fragment_two_pass.h +40 -0
- data/vendor/brotli/enc/compressor.h +15 -0
- data/vendor/brotli/enc/context.h +1 -1
- data/vendor/brotli/enc/dictionary.h +2 -2
- data/vendor/brotli/enc/encode.cc +819 -286
- data/vendor/brotli/enc/encode.h +38 -15
- data/vendor/brotli/enc/encode_parallel.cc +40 -42
- data/vendor/brotli/enc/entropy_encode.cc +144 -147
- data/vendor/brotli/enc/entropy_encode.h +32 -8
- data/vendor/brotli/enc/entropy_encode_static.h +572 -0
- data/vendor/brotli/enc/fast_log.h +7 -40
- data/vendor/brotli/enc/find_match_length.h +9 -9
- data/vendor/brotli/enc/hash.h +462 -154
- data/vendor/brotli/enc/histogram.cc +6 -6
- data/vendor/brotli/enc/histogram.h +13 -13
- data/vendor/brotli/enc/literal_cost.cc +45 -45
- data/vendor/brotli/enc/metablock.cc +92 -89
- data/vendor/brotli/enc/metablock.h +12 -12
- data/vendor/brotli/enc/port.h +7 -16
- data/vendor/brotli/enc/prefix.h +23 -22
- data/vendor/brotli/enc/ringbuffer.h +75 -29
- data/vendor/brotli/enc/static_dict.cc +56 -48
- data/vendor/brotli/enc/static_dict.h +5 -5
- data/vendor/brotli/enc/streams.cc +1 -1
- data/vendor/brotli/enc/streams.h +5 -5
- data/vendor/brotli/enc/transform.h +40 -35
- data/vendor/brotli/enc/types.h +2 -0
- data/vendor/brotli/enc/utf8_util.cc +3 -2
- data/vendor/brotli/enc/write_bits.h +6 -6
- metadata +9 -5
- data/vendor/brotli/dec/streams.c +0 -102
- data/vendor/brotli/dec/streams.h +0 -95
@@ -10,12 +10,10 @@
|
|
10
10
|
|
11
11
|
#include <assert.h>
|
12
12
|
#include <math.h>
|
13
|
-
#include <stdio.h>
|
14
|
-
#include <stdlib.h>
|
15
|
-
#include <string.h>
|
16
13
|
|
17
14
|
#include <algorithm>
|
18
|
-
#include <
|
15
|
+
#include <cstring>
|
16
|
+
#include <vector>
|
19
17
|
|
20
18
|
#include "./cluster.h"
|
21
19
|
#include "./command.h"
|
@@ -24,19 +22,19 @@
|
|
24
22
|
|
25
23
|
namespace brotli {
|
26
24
|
|
27
|
-
static const
|
28
|
-
static const
|
25
|
+
static const size_t kMaxLiteralHistograms = 100;
|
26
|
+
static const size_t kMaxCommandHistograms = 50;
|
29
27
|
static const double kLiteralBlockSwitchCost = 28.1;
|
30
28
|
static const double kCommandBlockSwitchCost = 13.5;
|
31
29
|
static const double kDistanceBlockSwitchCost = 14.6;
|
32
|
-
static const
|
33
|
-
static const
|
34
|
-
static const
|
35
|
-
static const
|
36
|
-
static const
|
37
|
-
static const
|
38
|
-
static const
|
39
|
-
static const
|
30
|
+
static const size_t kLiteralStrideLength = 70;
|
31
|
+
static const size_t kCommandStrideLength = 40;
|
32
|
+
static const size_t kSymbolsPerLiteralHistogram = 544;
|
33
|
+
static const size_t kSymbolsPerCommandHistogram = 530;
|
34
|
+
static const size_t kSymbolsPerDistanceHistogram = 544;
|
35
|
+
static const size_t kMinLengthForBlockSplitting = 128;
|
36
|
+
static const size_t kIterMulForRefining = 2;
|
37
|
+
static const size_t kMinItersForRefining = 100;
|
40
38
|
|
41
39
|
void CopyLiteralsToByteArray(const Command* cmds,
|
42
40
|
const size_t num_commands,
|
@@ -72,20 +70,7 @@ void CopyLiteralsToByteArray(const Command* cmds,
|
|
72
70
|
memcpy(&(*literals)[pos], data + from_pos, insert_len);
|
73
71
|
pos += insert_len;
|
74
72
|
}
|
75
|
-
from_pos = (from_pos + insert_len + cmds[i].
|
76
|
-
}
|
77
|
-
}
|
78
|
-
|
79
|
-
void CopyCommandsToByteArray(const Command* cmds,
|
80
|
-
const size_t num_commands,
|
81
|
-
std::vector<uint16_t>* insert_and_copy_codes,
|
82
|
-
std::vector<uint16_t>* distance_prefixes) {
|
83
|
-
for (size_t i = 0; i < num_commands; ++i) {
|
84
|
-
const Command& cmd = cmds[i];
|
85
|
-
insert_and_copy_codes->push_back(cmd.cmd_prefix_);
|
86
|
-
if (cmd.copy_len_ > 0 && cmd.cmd_prefix_ >= 128) {
|
87
|
-
distance_prefixes->push_back(cmd.dist_prefix_);
|
88
|
-
}
|
73
|
+
from_pos = (from_pos + insert_len + cmds[i].copy_len()) & mask;
|
89
74
|
}
|
90
75
|
}
|
91
76
|
|
@@ -99,27 +84,23 @@ inline static unsigned int MyRand(unsigned int* seed) {
|
|
99
84
|
|
100
85
|
template<typename HistogramType, typename DataType>
|
101
86
|
void InitialEntropyCodes(const DataType* data, size_t length,
|
102
|
-
int literals_per_histogram,
|
103
|
-
int max_histograms,
|
104
87
|
size_t stride,
|
105
|
-
|
106
|
-
|
107
|
-
|
108
|
-
|
88
|
+
size_t num_histograms,
|
89
|
+
HistogramType* histograms) {
|
90
|
+
for (size_t i = 0; i < num_histograms; ++i) {
|
91
|
+
histograms[i].Clear();
|
109
92
|
}
|
110
93
|
unsigned int seed = 7;
|
111
|
-
size_t block_length = length /
|
112
|
-
for (
|
113
|
-
size_t pos = length * i /
|
94
|
+
size_t block_length = length / num_histograms;
|
95
|
+
for (size_t i = 0; i < num_histograms; ++i) {
|
96
|
+
size_t pos = length * i / num_histograms;
|
114
97
|
if (i != 0) {
|
115
98
|
pos += MyRand(&seed) % block_length;
|
116
99
|
}
|
117
100
|
if (pos + stride >= length) {
|
118
101
|
pos = length - stride - 1;
|
119
102
|
}
|
120
|
-
|
121
|
-
histo.Add(data + pos, stride);
|
122
|
-
vec->push_back(histo);
|
103
|
+
histograms[i].Add(data + pos, stride);
|
123
104
|
}
|
124
105
|
}
|
125
106
|
|
@@ -142,50 +123,58 @@ void RandomSample(unsigned int* seed,
|
|
142
123
|
template<typename HistogramType, typename DataType>
|
143
124
|
void RefineEntropyCodes(const DataType* data, size_t length,
|
144
125
|
size_t stride,
|
145
|
-
|
126
|
+
size_t num_histograms,
|
127
|
+
HistogramType* histograms) {
|
146
128
|
size_t iters =
|
147
129
|
kIterMulForRefining * length / stride + kMinItersForRefining;
|
148
130
|
unsigned int seed = 7;
|
149
|
-
iters = ((iters +
|
131
|
+
iters = ((iters + num_histograms - 1) / num_histograms) * num_histograms;
|
150
132
|
for (size_t iter = 0; iter < iters; ++iter) {
|
151
133
|
HistogramType sample;
|
152
134
|
RandomSample(&seed, data, length, stride, &sample);
|
153
|
-
size_t ix = iter %
|
154
|
-
|
135
|
+
size_t ix = iter % num_histograms;
|
136
|
+
histograms[ix].AddHistogram(sample);
|
155
137
|
}
|
156
138
|
}
|
157
139
|
|
158
|
-
inline static double BitCost(
|
159
|
-
return count == 0 ? -2 : FastLog2(count);
|
140
|
+
inline static double BitCost(size_t count) {
|
141
|
+
return count == 0 ? -2.0 : FastLog2(count);
|
160
142
|
}
|
161
143
|
|
144
|
+
// Assigns a block id from the range [0, vec.size()) to each data element
|
145
|
+
// in data[0..length) and fills in block_id[0..length) with the assigned values.
|
146
|
+
// Returns the number of blocks, i.e. one plus the number of block switches.
|
162
147
|
template<typename DataType, int kSize>
|
163
|
-
|
164
|
-
|
165
|
-
|
166
|
-
|
167
|
-
|
148
|
+
size_t FindBlocks(const DataType* data, const size_t length,
|
149
|
+
const double block_switch_bitcost,
|
150
|
+
const size_t num_histograms,
|
151
|
+
const Histogram<kSize>* histograms,
|
152
|
+
double* insert_cost,
|
153
|
+
double* cost,
|
154
|
+
uint8_t* switch_signal,
|
155
|
+
uint8_t *block_id) {
|
156
|
+
if (num_histograms <= 1) {
|
168
157
|
for (size_t i = 0; i < length; ++i) {
|
169
158
|
block_id[i] = 0;
|
170
159
|
}
|
171
|
-
return;
|
160
|
+
return 1;
|
172
161
|
}
|
173
|
-
|
174
|
-
assert(
|
175
|
-
|
176
|
-
|
177
|
-
|
178
|
-
|
162
|
+
const size_t bitmaplen = (num_histograms + 7) >> 3;
|
163
|
+
assert(num_histograms <= 256);
|
164
|
+
memset(insert_cost, 0, sizeof(insert_cost[0]) * kSize * num_histograms);
|
165
|
+
for (size_t j = 0; j < num_histograms; ++j) {
|
166
|
+
insert_cost[j] = FastLog2(static_cast<uint32_t>(
|
167
|
+
histograms[j].total_count_));
|
179
168
|
}
|
180
|
-
for (
|
181
|
-
|
182
|
-
|
169
|
+
for (size_t i = kSize; i != 0;) {
|
170
|
+
--i;
|
171
|
+
for (size_t j = 0; j < num_histograms; ++j) {
|
172
|
+
insert_cost[i * num_histograms + j] =
|
173
|
+
insert_cost[j] - BitCost(histograms[j].data_[i]);
|
183
174
|
}
|
184
175
|
}
|
185
|
-
|
186
|
-
memset(
|
187
|
-
bool* switch_signal = new bool[length * vecsize];
|
188
|
-
memset(switch_signal, 0, sizeof(switch_signal[0]) * length * vecsize);
|
176
|
+
memset(cost, 0, sizeof(cost[0]) * num_histograms);
|
177
|
+
memset(switch_signal, 0, sizeof(switch_signal[0]) * length * bitmaplen);
|
189
178
|
// After each iteration of this loop, cost[k] will contain the difference
|
190
179
|
// between the minimum cost of arriving at the current byte position using
|
191
180
|
// entropy code k, and the minimum cost of arriving at the current byte
|
@@ -193,10 +182,10 @@ void FindBlocks(const DataType* data, const size_t length,
|
|
193
182
|
// reaches block switch cost, it means that when we trace back from the last
|
194
183
|
// position, we need to switch here.
|
195
184
|
for (size_t byte_ix = 0; byte_ix < length; ++byte_ix) {
|
196
|
-
size_t ix = byte_ix *
|
197
|
-
|
185
|
+
size_t ix = byte_ix * bitmaplen;
|
186
|
+
size_t insert_cost_ix = data[byte_ix] * num_histograms;
|
198
187
|
double min_cost = 1e99;
|
199
|
-
for (
|
188
|
+
for (size_t k = 0; k < num_histograms; ++k) {
|
200
189
|
// We are coding the symbol in data[byte_ix] with entropy code k.
|
201
190
|
cost[k] += insert_cost[insert_cost_ix + k];
|
202
191
|
if (cost[k] < min_cost) {
|
@@ -207,116 +196,206 @@ void FindBlocks(const DataType* data, const size_t length,
|
|
207
196
|
double block_switch_cost = block_switch_bitcost;
|
208
197
|
// More blocks for the beginning.
|
209
198
|
if (byte_ix < 2000) {
|
210
|
-
block_switch_cost *= 0.77 + 0.07 * byte_ix / 2000;
|
199
|
+
block_switch_cost *= 0.77 + 0.07 * static_cast<double>(byte_ix) / 2000;
|
211
200
|
}
|
212
|
-
for (
|
201
|
+
for (size_t k = 0; k < num_histograms; ++k) {
|
213
202
|
cost[k] -= min_cost;
|
214
203
|
if (cost[k] >= block_switch_cost) {
|
215
204
|
cost[k] = block_switch_cost;
|
216
|
-
|
205
|
+
const uint8_t mask = static_cast<uint8_t>(1u << (k & 7));
|
206
|
+
assert((k >> 3) < bitmaplen);
|
207
|
+
switch_signal[ix + (k >> 3)] |= mask;
|
217
208
|
}
|
218
209
|
}
|
219
210
|
}
|
220
211
|
// Now trace back from the last position and switch at the marked places.
|
221
212
|
size_t byte_ix = length - 1;
|
222
|
-
size_t ix = byte_ix *
|
213
|
+
size_t ix = byte_ix * bitmaplen;
|
223
214
|
uint8_t cur_id = block_id[byte_ix];
|
215
|
+
size_t num_blocks = 1;
|
224
216
|
while (byte_ix > 0) {
|
225
217
|
--byte_ix;
|
226
|
-
ix -=
|
227
|
-
|
228
|
-
|
218
|
+
ix -= bitmaplen;
|
219
|
+
const uint8_t mask = static_cast<uint8_t>(1u << (cur_id & 7));
|
220
|
+
assert((static_cast<size_t>(cur_id) >> 3) < bitmaplen);
|
221
|
+
if (switch_signal[ix + (cur_id >> 3)] & mask) {
|
222
|
+
if (cur_id != block_id[byte_ix]) {
|
223
|
+
cur_id = block_id[byte_ix];
|
224
|
+
++num_blocks;
|
225
|
+
}
|
229
226
|
}
|
230
227
|
block_id[byte_ix] = cur_id;
|
231
228
|
}
|
232
|
-
|
233
|
-
delete[] cost;
|
234
|
-
delete[] switch_signal;
|
229
|
+
return num_blocks;
|
235
230
|
}
|
236
231
|
|
237
|
-
|
238
|
-
|
239
|
-
|
232
|
+
static size_t RemapBlockIds(uint8_t* block_ids, const size_t length,
|
233
|
+
uint16_t* new_id, const size_t num_histograms) {
|
234
|
+
static const uint16_t kInvalidId = 256;
|
235
|
+
for (size_t i = 0; i < num_histograms; ++i) {
|
236
|
+
new_id[i] = kInvalidId;
|
237
|
+
}
|
238
|
+
uint16_t next_id = 0;
|
240
239
|
for (size_t i = 0; i < length; ++i) {
|
241
|
-
|
242
|
-
|
243
|
-
|
240
|
+
assert(block_ids[i] < num_histograms);
|
241
|
+
if (new_id[block_ids[i]] == kInvalidId) {
|
242
|
+
new_id[block_ids[i]] = next_id++;
|
244
243
|
}
|
245
244
|
}
|
246
245
|
for (size_t i = 0; i < length; ++i) {
|
247
|
-
block_ids[i] = new_id[block_ids[i]];
|
246
|
+
block_ids[i] = static_cast<uint8_t>(new_id[block_ids[i]]);
|
247
|
+
assert(block_ids[i] < num_histograms);
|
248
248
|
}
|
249
|
+
assert(next_id <= num_histograms);
|
249
250
|
return next_id;
|
250
251
|
}
|
251
252
|
|
252
253
|
template<typename HistogramType, typename DataType>
|
253
254
|
void BuildBlockHistograms(const DataType* data, const size_t length,
|
254
|
-
uint8_t* block_ids,
|
255
|
-
|
256
|
-
|
257
|
-
|
258
|
-
|
259
|
-
|
255
|
+
const uint8_t* block_ids,
|
256
|
+
const size_t num_histograms,
|
257
|
+
HistogramType* histograms) {
|
258
|
+
for (size_t i = 0; i < num_histograms; ++i) {
|
259
|
+
histograms[i].Clear();
|
260
|
+
}
|
260
261
|
for (size_t i = 0; i < length; ++i) {
|
261
|
-
|
262
|
+
histograms[block_ids[i]].Add(data[i]);
|
262
263
|
}
|
263
264
|
}
|
264
265
|
|
265
266
|
template<typename HistogramType, typename DataType>
|
266
267
|
void ClusterBlocks(const DataType* data, const size_t length,
|
267
|
-
|
268
|
-
|
269
|
-
|
270
|
-
|
271
|
-
|
268
|
+
const size_t num_blocks,
|
269
|
+
uint8_t* block_ids,
|
270
|
+
BlockSplit* split) {
|
271
|
+
static const size_t kMaxNumberOfBlockTypes = 256;
|
272
|
+
static const size_t kHistogramsPerBatch = 64;
|
273
|
+
static const size_t kClustersPerBatch = 16;
|
274
|
+
std::vector<uint32_t> histogram_symbols(num_blocks);
|
275
|
+
std::vector<uint32_t> block_lengths(num_blocks);
|
276
|
+
|
277
|
+
size_t block_idx = 0;
|
272
278
|
for (size_t i = 0; i < length; ++i) {
|
273
|
-
|
274
|
-
|
275
|
-
|
276
|
-
|
277
|
-
histograms.push_back(cur_histogram);
|
278
|
-
cur_histogram.Clear();
|
279
|
-
++cur_idx;
|
279
|
+
assert(block_idx < num_blocks);
|
280
|
+
++block_lengths[block_idx];
|
281
|
+
if (i + 1 == length || block_ids[i] != block_ids[i + 1]) {
|
282
|
+
++block_idx;
|
280
283
|
}
|
281
284
|
}
|
282
|
-
|
283
|
-
|
284
|
-
|
285
|
-
|
286
|
-
|
287
|
-
|
288
|
-
|
289
|
-
|
290
|
-
|
291
|
-
|
285
|
+
assert(block_idx == num_blocks);
|
286
|
+
|
287
|
+
const size_t expected_num_clusters =
|
288
|
+
kClustersPerBatch *
|
289
|
+
(num_blocks + kHistogramsPerBatch - 1) / kHistogramsPerBatch;
|
290
|
+
std::vector<HistogramType> all_histograms;
|
291
|
+
std::vector<uint32_t> cluster_size;
|
292
|
+
all_histograms.reserve(expected_num_clusters);
|
293
|
+
cluster_size.reserve(expected_num_clusters);
|
294
|
+
size_t num_clusters = 0;
|
295
|
+
std::vector<HistogramType> histograms(
|
296
|
+
std::min(num_blocks, kHistogramsPerBatch));
|
297
|
+
size_t max_num_pairs = kHistogramsPerBatch * kHistogramsPerBatch / 2;
|
298
|
+
std::vector<HistogramPair> pairs(max_num_pairs + 1);
|
299
|
+
size_t pos = 0;
|
300
|
+
for (size_t i = 0; i < num_blocks; i += kHistogramsPerBatch) {
|
301
|
+
const size_t num_to_combine = std::min(num_blocks - i, kHistogramsPerBatch);
|
302
|
+
uint32_t sizes[kHistogramsPerBatch];
|
303
|
+
uint32_t clusters[kHistogramsPerBatch];
|
304
|
+
uint32_t symbols[kHistogramsPerBatch];
|
305
|
+
uint32_t remap[kHistogramsPerBatch];
|
306
|
+
for (size_t j = 0; j < num_to_combine; ++j) {
|
307
|
+
histograms[j].Clear();
|
308
|
+
for (size_t k = 0; k < block_lengths[i + j]; ++k) {
|
309
|
+
histograms[j].Add(data[pos++]);
|
310
|
+
}
|
311
|
+
histograms[j].bit_cost_ = PopulationCost(histograms[j]);
|
312
|
+
symbols[j] = clusters[j] = static_cast<uint32_t>(j);
|
313
|
+
sizes[j] = 1;
|
314
|
+
}
|
315
|
+
size_t num_new_clusters = HistogramCombine(
|
316
|
+
&histograms[0], sizes, symbols, clusters, &pairs[0], num_to_combine,
|
317
|
+
num_to_combine, kHistogramsPerBatch, max_num_pairs);
|
318
|
+
for (size_t j = 0; j < num_new_clusters; ++j) {
|
319
|
+
all_histograms.push_back(histograms[clusters[j]]);
|
320
|
+
cluster_size.push_back(sizes[clusters[j]]);
|
321
|
+
remap[clusters[j]] = static_cast<uint32_t>(j);
|
322
|
+
}
|
323
|
+
for (size_t j = 0; j < num_to_combine; ++j) {
|
324
|
+
histogram_symbols[i + j] =
|
325
|
+
static_cast<uint32_t>(num_clusters) + remap[symbols[j]];
|
326
|
+
}
|
327
|
+
num_clusters += num_new_clusters;
|
328
|
+
assert(num_clusters == cluster_size.size());
|
329
|
+
assert(num_clusters == all_histograms.size());
|
292
330
|
}
|
293
|
-
}
|
294
331
|
|
295
|
-
|
296
|
-
|
297
|
-
|
298
|
-
|
299
|
-
|
300
|
-
|
301
|
-
|
302
|
-
|
303
|
-
|
304
|
-
|
332
|
+
max_num_pairs =
|
333
|
+
std::min(64 * num_clusters, (num_clusters / 2) * num_clusters);
|
334
|
+
pairs.resize(max_num_pairs + 1);
|
335
|
+
|
336
|
+
std::vector<uint32_t> clusters(num_clusters);
|
337
|
+
for (size_t i = 0; i < num_clusters; ++i) {
|
338
|
+
clusters[i] = static_cast<uint32_t>(i);
|
339
|
+
}
|
340
|
+
size_t num_final_clusters =
|
341
|
+
HistogramCombine(&all_histograms[0], &cluster_size[0],
|
342
|
+
&histogram_symbols[0],
|
343
|
+
&clusters[0], &pairs[0], num_clusters,
|
344
|
+
num_blocks, kMaxNumberOfBlockTypes, max_num_pairs);
|
345
|
+
|
346
|
+
static const uint32_t kInvalidIndex = std::numeric_limits<uint32_t>::max();
|
347
|
+
std::vector<uint32_t> new_index(num_clusters, kInvalidIndex);
|
348
|
+
uint32_t next_index = 0;
|
349
|
+
pos = 0;
|
350
|
+
for (size_t i = 0; i < num_blocks; ++i) {
|
351
|
+
HistogramType histo;
|
352
|
+
for (size_t j = 0; j < block_lengths[i]; ++j) {
|
353
|
+
histo.Add(data[pos++]);
|
354
|
+
}
|
355
|
+
uint32_t best_out =
|
356
|
+
i == 0 ? histogram_symbols[0] : histogram_symbols[i - 1];
|
357
|
+
double best_bits = HistogramBitCostDistance(
|
358
|
+
histo, all_histograms[best_out]);
|
359
|
+
for (size_t j = 0; j < num_final_clusters; ++j) {
|
360
|
+
const double cur_bits = HistogramBitCostDistance(
|
361
|
+
histo, all_histograms[clusters[j]]);
|
362
|
+
if (cur_bits < best_bits) {
|
363
|
+
best_bits = cur_bits;
|
364
|
+
best_out = clusters[j];
|
365
|
+
}
|
366
|
+
}
|
367
|
+
histogram_symbols[i] = best_out;
|
368
|
+
if (new_index[best_out] == kInvalidIndex) {
|
369
|
+
new_index[best_out] = next_index++;
|
370
|
+
}
|
371
|
+
}
|
372
|
+
uint8_t max_type = 0;
|
373
|
+
uint32_t cur_length = 0;
|
374
|
+
block_idx = 0;
|
375
|
+
split->types.resize(num_blocks);
|
376
|
+
split->lengths.resize(num_blocks);
|
377
|
+
for (size_t i = 0; i < num_blocks; ++i) {
|
378
|
+
cur_length += block_lengths[i];
|
379
|
+
if (i + 1 == num_blocks ||
|
380
|
+
histogram_symbols[i] != histogram_symbols[i + 1]) {
|
381
|
+
const uint8_t id = static_cast<uint8_t>(new_index[histogram_symbols[i]]);
|
382
|
+
split->types[block_idx] = id;
|
383
|
+
split->lengths[block_idx] = cur_length;
|
384
|
+
max_type = std::max(max_type, id);
|
305
385
|
cur_length = 0;
|
386
|
+
++block_idx;
|
306
387
|
}
|
307
|
-
++cur_length;
|
308
388
|
}
|
309
|
-
split->types.
|
310
|
-
split->lengths.
|
311
|
-
split->num_types =
|
312
|
-
++split->num_types;
|
389
|
+
split->types.resize(block_idx);
|
390
|
+
split->lengths.resize(block_idx);
|
391
|
+
split->num_types = static_cast<size_t>(max_type) + 1;
|
313
392
|
}
|
314
393
|
|
315
|
-
template<
|
394
|
+
template<int kSize, typename DataType>
|
316
395
|
void SplitByteVector(const std::vector<DataType>& data,
|
317
|
-
const
|
318
|
-
const
|
319
|
-
const
|
396
|
+
const size_t literals_per_histogram,
|
397
|
+
const size_t max_histograms,
|
398
|
+
const size_t sampling_stride_length,
|
320
399
|
const double block_switch_cost,
|
321
400
|
BlockSplit* split) {
|
322
401
|
if (data.empty()) {
|
@@ -325,30 +404,47 @@ void SplitByteVector(const std::vector<DataType>& data,
|
|
325
404
|
} else if (data.size() < kMinLengthForBlockSplitting) {
|
326
405
|
split->num_types = 1;
|
327
406
|
split->types.push_back(0);
|
328
|
-
split->lengths.push_back(static_cast<
|
407
|
+
split->lengths.push_back(static_cast<uint32_t>(data.size()));
|
329
408
|
return;
|
330
409
|
}
|
331
|
-
|
410
|
+
size_t num_histograms = data.size() / literals_per_histogram + 1;
|
411
|
+
if (num_histograms > max_histograms) {
|
412
|
+
num_histograms = max_histograms;
|
413
|
+
}
|
414
|
+
Histogram<kSize>* histograms = new Histogram<kSize>[num_histograms];
|
332
415
|
// Find good entropy codes.
|
333
416
|
InitialEntropyCodes(&data[0], data.size(),
|
334
|
-
literals_per_histogram,
|
335
|
-
max_histograms,
|
336
417
|
sampling_stride_length,
|
337
|
-
|
418
|
+
num_histograms, histograms);
|
338
419
|
RefineEntropyCodes(&data[0], data.size(),
|
339
420
|
sampling_stride_length,
|
340
|
-
|
421
|
+
num_histograms, histograms);
|
341
422
|
// Find a good path through literals with the good entropy codes.
|
342
423
|
std::vector<uint8_t> block_ids(data.size());
|
343
|
-
|
344
|
-
|
345
|
-
|
346
|
-
|
347
|
-
|
348
|
-
|
424
|
+
size_t num_blocks;
|
425
|
+
const size_t bitmaplen = (num_histograms + 7) >> 3;
|
426
|
+
double* insert_cost = new double[kSize * num_histograms];
|
427
|
+
double *cost = new double[num_histograms];
|
428
|
+
uint8_t* switch_signal = new uint8_t[data.size() * bitmaplen];
|
429
|
+
uint16_t* new_id = new uint16_t[num_histograms];
|
430
|
+
for (size_t i = 0; i < 10; ++i) {
|
431
|
+
num_blocks = FindBlocks(&data[0], data.size(),
|
432
|
+
block_switch_cost,
|
433
|
+
num_histograms, histograms,
|
434
|
+
insert_cost, cost, switch_signal,
|
435
|
+
&block_ids[0]);
|
436
|
+
num_histograms = RemapBlockIds(&block_ids[0], data.size(),
|
437
|
+
new_id, num_histograms);
|
438
|
+
BuildBlockHistograms(&data[0], data.size(), &block_ids[0],
|
439
|
+
num_histograms, histograms);
|
349
440
|
}
|
350
|
-
|
351
|
-
|
441
|
+
delete[] insert_cost;
|
442
|
+
delete[] cost;
|
443
|
+
delete[] switch_signal;
|
444
|
+
delete[] new_id;
|
445
|
+
delete[] histograms;
|
446
|
+
ClusterBlocks<Histogram<kSize> >(&data[0], data.size(), num_blocks,
|
447
|
+
&block_ids[0], split);
|
352
448
|
}
|
353
449
|
|
354
450
|
void SplitBlock(const Command* cmds,
|
@@ -359,55 +455,51 @@ void SplitBlock(const Command* cmds,
|
|
359
455
|
BlockSplit* literal_split,
|
360
456
|
BlockSplit* insert_and_copy_split,
|
361
457
|
BlockSplit* dist_split) {
|
362
|
-
|
363
|
-
|
364
|
-
|
365
|
-
|
366
|
-
|
367
|
-
|
368
|
-
|
369
|
-
|
370
|
-
|
371
|
-
|
458
|
+
{
|
459
|
+
// Create a continuous array of literals.
|
460
|
+
std::vector<uint8_t> literals;
|
461
|
+
CopyLiteralsToByteArray(cmds, num_commands, data, pos, mask, &literals);
|
462
|
+
// Create the block split on the array of literals.
|
463
|
+
// Literal histograms have alphabet size 256.
|
464
|
+
SplitByteVector<256>(
|
465
|
+
literals,
|
466
|
+
kSymbolsPerLiteralHistogram, kMaxLiteralHistograms,
|
467
|
+
kLiteralStrideLength, kLiteralBlockSwitchCost,
|
468
|
+
literal_split);
|
469
|
+
}
|
372
470
|
|
373
|
-
|
374
|
-
|
375
|
-
|
376
|
-
|
377
|
-
|
378
|
-
|
379
|
-
|
380
|
-
|
381
|
-
|
382
|
-
|
383
|
-
|
384
|
-
|
385
|
-
|
386
|
-
kCommandStrideLength, kDistanceBlockSwitchCost,
|
387
|
-
dist_split);
|
388
|
-
}
|
471
|
+
{
|
472
|
+
// Compute prefix codes for commands.
|
473
|
+
std::vector<uint16_t> insert_and_copy_codes(num_commands);
|
474
|
+
for (size_t i = 0; i < num_commands; ++i) {
|
475
|
+
insert_and_copy_codes[i] = cmds[i].cmd_prefix_;
|
476
|
+
}
|
477
|
+
// Create the block split on the array of command prefixes.
|
478
|
+
SplitByteVector<kNumCommandPrefixes>(
|
479
|
+
insert_and_copy_codes,
|
480
|
+
kSymbolsPerCommandHistogram, kMaxCommandHistograms,
|
481
|
+
kCommandStrideLength, kCommandBlockSwitchCost,
|
482
|
+
insert_and_copy_split);
|
483
|
+
}
|
389
484
|
|
390
|
-
|
391
|
-
|
392
|
-
|
393
|
-
|
394
|
-
|
395
|
-
|
396
|
-
|
397
|
-
|
398
|
-
|
399
|
-
for (size_t i = 0; i < num_commands; ++i) {
|
400
|
-
const Command& cmd = all_commands[i];
|
401
|
-
int cmd_length = cmd.insert_len_ + cmd.copy_len_;
|
402
|
-
if (total_length > length_limit) {
|
403
|
-
blocks->push_back(cur_block);
|
404
|
-
cur_block.clear();
|
405
|
-
total_length = 0;
|
485
|
+
{
|
486
|
+
// Create a continuous array of distance prefixes.
|
487
|
+
std::vector<uint16_t> distance_prefixes(num_commands);
|
488
|
+
size_t pos = 0;
|
489
|
+
for (size_t i = 0; i < num_commands; ++i) {
|
490
|
+
const Command& cmd = cmds[i];
|
491
|
+
if (cmd.copy_len() && cmd.cmd_prefix_ >= 128) {
|
492
|
+
distance_prefixes[pos++] = cmd.dist_prefix_;
|
493
|
+
}
|
406
494
|
}
|
407
|
-
|
408
|
-
|
495
|
+
distance_prefixes.resize(pos);
|
496
|
+
// Create the block split on the array of distance prefixes.
|
497
|
+
SplitByteVector<kNumDistancePrefixes>(
|
498
|
+
distance_prefixes,
|
499
|
+
kSymbolsPerDistanceHistogram, kMaxCommandHistograms,
|
500
|
+
kCommandStrideLength, kDistanceBlockSwitchCost,
|
501
|
+
dist_split);
|
409
502
|
}
|
410
|
-
blocks->push_back(cur_block);
|
411
503
|
}
|
412
504
|
|
413
505
|
} // namespace brotli
|