1 /* Copyright 2013 Google Inc. All Rights Reserved.
2
3 Distributed under MIT license.
4 See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
5 */
6
7 /* Block split point selection utilities. */
8
9 #include "block_splitter.h"
10
11 #include <string.h> /* memcpy, memset */
12
13 #include "../common/platform.h"
14 #include "bit_cost.h"
15 #include "cluster.h"
16 #include "command.h"
17 #include "fast_log.h"
18 #include "histogram.h"
19 #include "memory.h"
20 #include "quality.h"
21
22 #if defined(__cplusplus) || defined(c_plusplus)
23 extern "C" {
24 #endif
25
26 static const size_t kMaxLiteralHistograms = 100;
27 static const size_t kMaxCommandHistograms = 50;
28 static const double kLiteralBlockSwitchCost = 28.1;
29 static const double kCommandBlockSwitchCost = 13.5;
30 static const double kDistanceBlockSwitchCost = 14.6;
31 static const size_t kLiteralStrideLength = 70;
32 static const size_t kCommandStrideLength = 40;
33 static const size_t kDistanceStrideLength = 40;
34 static const size_t kSymbolsPerLiteralHistogram = 544;
35 static const size_t kSymbolsPerCommandHistogram = 530;
36 static const size_t kSymbolsPerDistanceHistogram = 544;
37 static const size_t kMinLengthForBlockSplitting = 128;
38 static const size_t kIterMulForRefining = 2;
39 static const size_t kMinItersForRefining = 100;
40
CountLiterals(const Command * cmds,const size_t num_commands)41 static size_t CountLiterals(const Command* cmds, const size_t num_commands) {
42 /* Count how many we have. */
43 size_t total_length = 0;
44 size_t i;
45 for (i = 0; i < num_commands; ++i) {
46 total_length += cmds[i].insert_len_;
47 }
48 return total_length;
49 }
50
CopyLiteralsToByteArray(const Command * cmds,const size_t num_commands,const uint8_t * data,const size_t offset,const size_t mask,uint8_t * literals)51 static void CopyLiteralsToByteArray(const Command* cmds,
52 const size_t num_commands,
53 const uint8_t* data,
54 const size_t offset,
55 const size_t mask,
56 uint8_t* literals) {
57 size_t pos = 0;
58 size_t from_pos = offset & mask;
59 size_t i;
60 for (i = 0; i < num_commands; ++i) {
61 size_t insert_len = cmds[i].insert_len_;
62 if (from_pos + insert_len > mask) {
63 size_t head_size = mask + 1 - from_pos;
64 memcpy(literals + pos, data + from_pos, head_size);
65 from_pos = 0;
66 pos += head_size;
67 insert_len -= head_size;
68 }
69 if (insert_len > 0) {
70 memcpy(literals + pos, data + from_pos, insert_len);
71 pos += insert_len;
72 }
73 from_pos = (from_pos + insert_len + CommandCopyLen(&cmds[i])) & mask;
74 }
75 }
76
MyRand(uint32_t * seed)77 static BROTLI_INLINE uint32_t MyRand(uint32_t* seed) {
78 /* Initial seed should be 7. In this case, loop length is (1 << 29). */
79 *seed *= 16807U;
80 return *seed;
81 }
82
BitCost(size_t count)83 static BROTLI_INLINE double BitCost(size_t count) {
84 return count == 0 ? -2.0 : FastLog2(count);
85 }
86
87 #define HISTOGRAMS_PER_BATCH 64
88 #define CLUSTERS_PER_BATCH 16
89
90 #define FN(X) X ## Literal
91 #define DataType uint8_t
92 /* NOLINTNEXTLINE(build/include) */
93 #include "block_splitter_inc.h"
94 #undef DataType
95 #undef FN
96
97 #define FN(X) X ## Command
98 #define DataType uint16_t
99 /* NOLINTNEXTLINE(build/include) */
100 #include "block_splitter_inc.h"
101 #undef FN
102
103 #define FN(X) X ## Distance
104 /* NOLINTNEXTLINE(build/include) */
105 #include "block_splitter_inc.h"
106 #undef DataType
107 #undef FN
108
BrotliInitBlockSplit(BlockSplit * self)109 void BrotliInitBlockSplit(BlockSplit* self) {
110 self->num_types = 0;
111 self->num_blocks = 0;
112 self->types = 0;
113 self->lengths = 0;
114 self->types_alloc_size = 0;
115 self->lengths_alloc_size = 0;
116 }
117
BrotliDestroyBlockSplit(MemoryManager * m,BlockSplit * self)118 void BrotliDestroyBlockSplit(MemoryManager* m, BlockSplit* self) {
119 BROTLI_FREE(m, self->types);
120 BROTLI_FREE(m, self->lengths);
121 }
122
123 /* Extracts literals, command distance and prefix codes, then applies
124 * SplitByteVector to create partitioning. */
BrotliSplitBlock(MemoryManager * m,const Command * cmds,const size_t num_commands,const uint8_t * data,const size_t pos,const size_t mask,const BrotliEncoderParams * params,BlockSplit * literal_split,BlockSplit * insert_and_copy_split,BlockSplit * dist_split)125 void BrotliSplitBlock(MemoryManager* m,
126 const Command* cmds,
127 const size_t num_commands,
128 const uint8_t* data,
129 const size_t pos,
130 const size_t mask,
131 const BrotliEncoderParams* params,
132 BlockSplit* literal_split,
133 BlockSplit* insert_and_copy_split,
134 BlockSplit* dist_split) {
135 {
136 size_t literals_count = CountLiterals(cmds, num_commands);
137 uint8_t* literals = BROTLI_ALLOC(m, uint8_t, literals_count);
138 if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(literals)) return;
139 /* Create a continuous array of literals. */
140 CopyLiteralsToByteArray(cmds, num_commands, data, pos, mask, literals);
141 /* Create the block split on the array of literals.
142 * Literal histograms can have alphabet size up to 256.
143 * Though, to accomodate context modeling, less than half of maximum size
144 * is allowed. */
145 SplitByteVectorLiteral(
146 m, literals, literals_count,
147 kSymbolsPerLiteralHistogram, kMaxLiteralHistograms,
148 kLiteralStrideLength, kLiteralBlockSwitchCost, params,
149 literal_split);
150 if (BROTLI_IS_OOM(m)) return;
151 BROTLI_FREE(m, literals);
152 /* NB: this might be a good place for injecting extra splitting without
153 * increasing encoder complexity; however, output parition would be less
154 * optimal than one produced with forced splitting inside
155 * SplitByteVector (FindBlocks / ClusterBlocks). */
156 }
157
158 {
159 /* Compute prefix codes for commands. */
160 uint16_t* insert_and_copy_codes = BROTLI_ALLOC(m, uint16_t, num_commands);
161 size_t i;
162 if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(insert_and_copy_codes)) return;
163 for (i = 0; i < num_commands; ++i) {
164 insert_and_copy_codes[i] = cmds[i].cmd_prefix_;
165 }
166 /* Create the block split on the array of command prefixes. */
167 SplitByteVectorCommand(
168 m, insert_and_copy_codes, num_commands,
169 kSymbolsPerCommandHistogram, kMaxCommandHistograms,
170 kCommandStrideLength, kCommandBlockSwitchCost, params,
171 insert_and_copy_split);
172 if (BROTLI_IS_OOM(m)) return;
173 /* TODO(eustas): reuse for distances? */
174 BROTLI_FREE(m, insert_and_copy_codes);
175 }
176
177 {
178 /* Create a continuous array of distance prefixes. */
179 uint16_t* distance_prefixes = BROTLI_ALLOC(m, uint16_t, num_commands);
180 size_t j = 0;
181 size_t i;
182 if (BROTLI_IS_OOM(m) || BROTLI_IS_NULL(distance_prefixes)) return;
183 for (i = 0; i < num_commands; ++i) {
184 const Command* cmd = &cmds[i];
185 if (CommandCopyLen(cmd) && cmd->cmd_prefix_ >= 128) {
186 distance_prefixes[j++] = cmd->dist_prefix_ & 0x3FF;
187 }
188 }
189 /* Create the block split on the array of distance prefixes. */
190 SplitByteVectorDistance(
191 m, distance_prefixes, j,
192 kSymbolsPerDistanceHistogram, kMaxCommandHistograms,
193 kDistanceStrideLength, kDistanceBlockSwitchCost, params,
194 dist_split);
195 if (BROTLI_IS_OOM(m)) return;
196 BROTLI_FREE(m, distance_prefixes);
197 }
198 }
199
200 #if defined(BROTLI_TEST)
201 size_t CountLiteralsForTest(const Command*, const size_t);
CountLiteralsForTest(const Command * cmds,const size_t num_commands)202 size_t CountLiteralsForTest(const Command* cmds, const size_t num_commands) {
203 return CountLiterals(cmds, num_commands);
204 }
205
206 void CopyLiteralsToByteArrayForTest(const Command*,
207 const size_t, const uint8_t*, const size_t, const size_t, uint8_t*);
CopyLiteralsToByteArrayForTest(const Command * cmds,const size_t num_commands,const uint8_t * data,const size_t offset,const size_t mask,uint8_t * literals)208 void CopyLiteralsToByteArrayForTest(const Command* cmds,
209 const size_t num_commands, const uint8_t* data, const size_t offset,
210 const size_t mask, uint8_t* literals) {
211 CopyLiteralsToByteArray(cmds, num_commands, data, offset, mask, literals);
212 }
213 #endif
214
215 #if defined(__cplusplus) || defined(c_plusplus)
216 } /* extern "C" */
217 #endif
218