Hash :
20481890
Author :
Date :
2016-07-26T14:41:59
Update encoder: * booleanification * integer BR scores, may improve performance if FPU is slow * condense speed-quality constants in quality.h * code massage to calm down CoverityScan * hashers refactoring * new hasher - improved speed, compression and reduced memory usage for q:5-9 w:10-16 * reduced static recources -> binary size
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
/* NOLINT(build/header_guard) */
/* Copyright 2013 Google Inc. All Rights Reserved.
Distributed under MIT license.
See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
*/
/* template parameters: FN, DataType */
#define HistogramType FN(Histogram)
static void FN(InitialEntropyCodes)(const DataType* data, size_t length,
size_t stride,
size_t num_histograms,
HistogramType* histograms) {
unsigned int seed = 7;
size_t block_length = length / num_histograms;
size_t i;
FN(ClearHistograms)(histograms, num_histograms);
for (i = 0; i < num_histograms; ++i) {
size_t pos = length * i / num_histograms;
if (i != 0) {
pos += MyRand(&seed) % block_length;
}
if (pos + stride >= length) {
pos = length - stride - 1;
}
FN(HistogramAddVector)(&histograms[i], data + pos, stride);
}
}
static void FN(RandomSample)(unsigned int* seed,
const DataType* data,
size_t length,
size_t stride,
HistogramType* sample) {
size_t pos = 0;
if (stride >= length) {
pos = 0;
stride = length;
} else {
pos = MyRand(seed) % (length - stride + 1);
}
FN(HistogramAddVector)(sample, data + pos, stride);
}
static void FN(RefineEntropyCodes)(const DataType* data, size_t length,
size_t stride,
size_t num_histograms,
HistogramType* histograms) {
size_t iters =
kIterMulForRefining * length / stride + kMinItersForRefining;
unsigned int seed = 7;
size_t iter;
iters = ((iters + num_histograms - 1) / num_histograms) * num_histograms;
for (iter = 0; iter < iters; ++iter) {
HistogramType sample;
FN(HistogramClear)(&sample);
FN(RandomSample)(&seed, data, length, stride, &sample);
FN(HistogramAddHistogram)(&histograms[iter % num_histograms], &sample);
}
}
/* Assigns a block id from the range [0, vec.size()) to each data element
in data[0..length) and fills in block_id[0..length) with the assigned values.
Returns the number of blocks, i.e. one plus the number of block switches. */
static size_t FN(FindBlocks)(const DataType* data, const size_t length,
const double block_switch_bitcost,
const size_t num_histograms,
const HistogramType* histograms,
double* insert_cost,
double* cost,
uint8_t* switch_signal,
uint8_t *block_id) {
const size_t data_size = FN(HistogramDataSize)();
const size_t bitmaplen = (num_histograms + 7) >> 3;
size_t num_blocks = 1;
size_t i;
size_t j;
assert(num_histograms <= 256);
if (num_histograms <= 1) {
for (i = 0; i < length; ++i) {
block_id[i] = 0;
}
return 1;
}
memset(insert_cost, 0, sizeof(insert_cost[0]) * data_size * num_histograms);
for (i = 0; i < num_histograms; ++i) {
insert_cost[i] = FastLog2((uint32_t)histograms[i].total_count_);
}
for (i = data_size; i != 0;) {
--i;
for (j = 0; j < num_histograms; ++j) {
insert_cost[i * num_histograms + j] =
insert_cost[j] - BitCost(histograms[j].data_[i]);
}
}
memset(cost, 0, sizeof(cost[0]) * num_histograms);
memset(switch_signal, 0, sizeof(switch_signal[0]) * length * bitmaplen);
/* After each iteration of this loop, cost[k] will contain the difference
between the minimum cost of arriving at the current byte position using
entropy code k, and the minimum cost of arriving at the current byte
position. This difference is capped at the block switch cost, and if it
reaches block switch cost, it means that when we trace back from the last
position, we need to switch here. */
for (i = 0; i < length; ++i) {
const size_t byte_ix = i;
size_t ix = byte_ix * bitmaplen;
size_t insert_cost_ix = data[byte_ix] * num_histograms;
double min_cost = 1e99;
double block_switch_cost = block_switch_bitcost;
size_t k;
for (k = 0; k < num_histograms; ++k) {
/* We are coding the symbol in data[byte_ix] with entropy code k. */
cost[k] += insert_cost[insert_cost_ix + k];
if (cost[k] < min_cost) {
min_cost = cost[k];
block_id[byte_ix] = (uint8_t)k;
}
}
/* More blocks for the beginning. */
if (byte_ix < 2000) {
block_switch_cost *= 0.77 + 0.07 * (double)byte_ix / 2000;
}
for (k = 0; k < num_histograms; ++k) {
cost[k] -= min_cost;
if (cost[k] >= block_switch_cost) {
const uint8_t mask = (uint8_t)(1u << (k & 7));
cost[k] = block_switch_cost;
assert((k >> 3) < bitmaplen);
switch_signal[ix + (k >> 3)] |= mask;
}
}
}
{ /* Trace back from the last position and switch at the marked places. */
size_t byte_ix = length - 1;
size_t ix = byte_ix * bitmaplen;
uint8_t cur_id = block_id[byte_ix];
while (byte_ix > 0) {
const uint8_t mask = (uint8_t)(1u << (cur_id & 7));
assert(((size_t)cur_id >> 3) < bitmaplen);
--byte_ix;
ix -= bitmaplen;
if (switch_signal[ix + (cur_id >> 3)] & mask) {
if (cur_id != block_id[byte_ix]) {
cur_id = block_id[byte_ix];
++num_blocks;
}
}
block_id[byte_ix] = cur_id;
}
}
return num_blocks;
}
static size_t FN(RemapBlockIds)(uint8_t* block_ids, const size_t length,
uint16_t* new_id, const size_t num_histograms) {
static const uint16_t kInvalidId = 256;
uint16_t next_id = 0;
size_t i;
for (i = 0; i < num_histograms; ++i) {
new_id[i] = kInvalidId;
}
for (i = 0; i < length; ++i) {
assert(block_ids[i] < num_histograms);
if (new_id[block_ids[i]] == kInvalidId) {
new_id[block_ids[i]] = next_id++;
}
}
for (i = 0; i < length; ++i) {
block_ids[i] = (uint8_t)new_id[block_ids[i]];
assert(block_ids[i] < num_histograms);
}
assert(next_id <= num_histograms);
return next_id;
}
static void FN(BuildBlockHistograms)(const DataType* data, const size_t length,
const uint8_t* block_ids,
const size_t num_histograms,
HistogramType* histograms) {
size_t i;
FN(ClearHistograms)(histograms, num_histograms);
for (i = 0; i < length; ++i) {
FN(HistogramAdd)(&histograms[block_ids[i]], data[i]);
}
}
static void FN(ClusterBlocks)(MemoryManager* m,
const DataType* data, const size_t length,
const size_t num_blocks,
uint8_t* block_ids,
BlockSplit* split) {
uint32_t* histogram_symbols = BROTLI_ALLOC(m, uint32_t, num_blocks);
uint32_t* block_lengths = BROTLI_ALLOC(m, uint32_t, num_blocks);
const size_t expected_num_clusters = CLUSTERS_PER_BATCH *
(num_blocks + HISTOGRAMS_PER_BATCH - 1) / HISTOGRAMS_PER_BATCH;
size_t all_histograms_size = 0;
size_t all_histograms_capacity = expected_num_clusters;
HistogramType* all_histograms =
BROTLI_ALLOC(m, HistogramType, all_histograms_capacity);
size_t cluster_size_size = 0;
size_t cluster_size_capacity = expected_num_clusters;
uint32_t* cluster_size = BROTLI_ALLOC(m, uint32_t, cluster_size_capacity);
size_t num_clusters = 0;
HistogramType* histograms = BROTLI_ALLOC(m, HistogramType,
BROTLI_MIN(size_t, num_blocks, HISTOGRAMS_PER_BATCH));
size_t max_num_pairs =
HISTOGRAMS_PER_BATCH * HISTOGRAMS_PER_BATCH / 2;
size_t pairs_capacity = max_num_pairs + 1;
HistogramPair* pairs = BROTLI_ALLOC(m, HistogramPair, pairs_capacity);
size_t pos = 0;
uint32_t* clusters;
size_t num_final_clusters;
static const uint32_t kInvalidIndex = BROTLI_UINT32_MAX;
uint32_t* new_index;
uint8_t max_type = 0;
size_t i;
uint32_t sizes[HISTOGRAMS_PER_BATCH] = { 0 };
uint32_t new_clusters[HISTOGRAMS_PER_BATCH] = { 0 };
uint32_t symbols[HISTOGRAMS_PER_BATCH] = { 0 };
uint32_t remap[HISTOGRAMS_PER_BATCH] = { 0 };
if (BROTLI_IS_OOM(m)) return;
memset(block_lengths, 0, num_blocks * sizeof(uint32_t));
{
size_t block_idx = 0;
for (i = 0; i < length; ++i) {
assert(block_idx < num_blocks);
++block_lengths[block_idx];
if (i + 1 == length || block_ids[i] != block_ids[i + 1]) {
++block_idx;
}
}
assert(block_idx == num_blocks);
}
for (i = 0; i < num_blocks; i += HISTOGRAMS_PER_BATCH) {
const size_t num_to_combine =
BROTLI_MIN(size_t, num_blocks - i, HISTOGRAMS_PER_BATCH);
size_t num_new_clusters;
size_t j;
for (j = 0; j < num_to_combine; ++j) {
size_t k;
FN(HistogramClear)(&histograms[j]);
for (k = 0; k < block_lengths[i + j]; ++k) {
FN(HistogramAdd)(&histograms[j], data[pos++]);
}
histograms[j].bit_cost_ = FN(BrotliPopulationCost)(&histograms[j]);
new_clusters[j] = (uint32_t)j;
symbols[j] = (uint32_t)j;
sizes[j] = 1;
}
num_new_clusters = FN(BrotliHistogramCombine)(
histograms, sizes, symbols, new_clusters, pairs, num_to_combine,
num_to_combine, HISTOGRAMS_PER_BATCH, max_num_pairs);
BROTLI_ENSURE_CAPACITY(m, HistogramType, all_histograms,
all_histograms_capacity, all_histograms_size + num_new_clusters);
BROTLI_ENSURE_CAPACITY(m, uint32_t, cluster_size,
cluster_size_capacity, cluster_size_size + num_new_clusters);
if (BROTLI_IS_OOM(m)) return;
for (j = 0; j < num_new_clusters; ++j) {
all_histograms[all_histograms_size++] = histograms[new_clusters[j]];
cluster_size[cluster_size_size++] = sizes[new_clusters[j]];
remap[new_clusters[j]] = (uint32_t)j;
}
for (j = 0; j < num_to_combine; ++j) {
histogram_symbols[i + j] = (uint32_t)num_clusters + remap[symbols[j]];
}
num_clusters += num_new_clusters;
assert(num_clusters == cluster_size_size);
assert(num_clusters == all_histograms_size);
}
BROTLI_FREE(m, histograms);
max_num_pairs =
BROTLI_MIN(size_t, 64 * num_clusters, (num_clusters / 2) * num_clusters);
if (pairs_capacity < max_num_pairs + 1) {
BROTLI_FREE(m, pairs);
pairs = BROTLI_ALLOC(m, HistogramPair, max_num_pairs + 1);
if (BROTLI_IS_OOM(m)) return;
}
clusters = BROTLI_ALLOC(m, uint32_t, num_clusters);
if (BROTLI_IS_OOM(m)) return;
for (i = 0; i < num_clusters; ++i) {
clusters[i] = (uint32_t)i;
}
num_final_clusters = FN(BrotliHistogramCombine)(
all_histograms, cluster_size, histogram_symbols, clusters, pairs,
num_clusters, num_blocks, BROTLI_MAX_NUMBER_OF_BLOCK_TYPES,
max_num_pairs);
BROTLI_FREE(m, pairs);
BROTLI_FREE(m, cluster_size);
new_index = BROTLI_ALLOC(m, uint32_t, num_clusters);
if (BROTLI_IS_OOM(m)) return;
for (i = 0; i < num_clusters; ++i) new_index[i] = kInvalidIndex;
pos = 0;
{
uint32_t next_index = 0;
for (i = 0; i < num_blocks; ++i) {
HistogramType histo;
size_t j;
uint32_t best_out;
double best_bits;
FN(HistogramClear)(&histo);
for (j = 0; j < block_lengths[i]; ++j) {
FN(HistogramAdd)(&histo, data[pos++]);
}
best_out = (i == 0) ? histogram_symbols[0] : histogram_symbols[i - 1];
best_bits =
FN(BrotliHistogramBitCostDistance)(&histo, &all_histograms[best_out]);
for (j = 0; j < num_final_clusters; ++j) {
const double cur_bits = FN(BrotliHistogramBitCostDistance)(
&histo, &all_histograms[clusters[j]]);
if (cur_bits < best_bits) {
best_bits = cur_bits;
best_out = clusters[j];
}
}
histogram_symbols[i] = best_out;
if (new_index[best_out] == kInvalidIndex) {
new_index[best_out] = next_index++;
}
}
}
BROTLI_FREE(m, clusters);
BROTLI_FREE(m, all_histograms);
BROTLI_ENSURE_CAPACITY(
m, uint8_t, split->types, split->types_alloc_size, num_blocks);
BROTLI_ENSURE_CAPACITY(
m, uint32_t, split->lengths, split->lengths_alloc_size, num_blocks);
if (BROTLI_IS_OOM(m)) return;
{
uint32_t cur_length = 0;
size_t block_idx = 0;
for (i = 0; i < num_blocks; ++i) {
cur_length += block_lengths[i];
if (i + 1 == num_blocks ||
histogram_symbols[i] != histogram_symbols[i + 1]) {
const uint8_t id = (uint8_t)new_index[histogram_symbols[i]];
split->types[block_idx] = id;
split->lengths[block_idx] = cur_length;
max_type = BROTLI_MAX(uint8_t, max_type, id);
cur_length = 0;
++block_idx;
}
}
split->num_blocks = block_idx;
split->num_types = (size_t)max_type + 1;
}
BROTLI_FREE(m, new_index);
BROTLI_FREE(m, block_lengths);
BROTLI_FREE(m, histogram_symbols);
}
static void FN(SplitByteVector)(MemoryManager* m,
const DataType* data, const size_t length,
const size_t literals_per_histogram,
const size_t max_histograms,
const size_t sampling_stride_length,
const double block_switch_cost,
const BrotliEncoderParams* params,
BlockSplit* split) {
const size_t data_size = FN(HistogramDataSize)();
size_t num_histograms = length / literals_per_histogram + 1;
HistogramType* histograms;
if (num_histograms > max_histograms) {
num_histograms = max_histograms;
}
if (length == 0) {
split->num_types = 1;
return;
} else if (length < kMinLengthForBlockSplitting) {
BROTLI_ENSURE_CAPACITY(m, uint8_t,
split->types, split->types_alloc_size, split->num_blocks + 1);
BROTLI_ENSURE_CAPACITY(m, uint32_t,
split->lengths, split->lengths_alloc_size, split->num_blocks + 1);
if (BROTLI_IS_OOM(m)) return;
split->num_types = 1;
split->types[split->num_blocks] = 0;
split->lengths[split->num_blocks] = (uint32_t)length;
split->num_blocks++;
return;
}
histograms = BROTLI_ALLOC(m, HistogramType, num_histograms);
if (BROTLI_IS_OOM(m)) return;
/* Find good entropy codes. */
FN(InitialEntropyCodes)(data, length,
sampling_stride_length,
num_histograms, histograms);
FN(RefineEntropyCodes)(data, length,
sampling_stride_length,
num_histograms, histograms);
{
/* Find a good path through literals with the good entropy codes. */
uint8_t* block_ids = BROTLI_ALLOC(m, uint8_t, length);
size_t num_blocks;
const size_t bitmaplen = (num_histograms + 7) >> 3;
double* insert_cost = BROTLI_ALLOC(m, double, data_size * num_histograms);
double* cost = BROTLI_ALLOC(m, double, num_histograms);
uint8_t* switch_signal = BROTLI_ALLOC(m, uint8_t, length * bitmaplen);
uint16_t* new_id = BROTLI_ALLOC(m, uint16_t, num_histograms);
const size_t iters = params->quality < HQ_ZOPFLIFICATION_QUALITY ? 3 : 10;
size_t i;
if (BROTLI_IS_OOM(m)) return;
for (i = 0; i < iters; ++i) {
num_blocks = FN(FindBlocks)(data, length,
block_switch_cost,
num_histograms, histograms,
insert_cost, cost, switch_signal,
block_ids);
num_histograms = FN(RemapBlockIds)(block_ids, length,
new_id, num_histograms);
FN(BuildBlockHistograms)(data, length, block_ids,
num_histograms, histograms);
}
BROTLI_FREE(m, insert_cost);
BROTLI_FREE(m, cost);
BROTLI_FREE(m, switch_signal);
BROTLI_FREE(m, new_id);
BROTLI_FREE(m, histograms);
FN(ClusterBlocks)(m, data, length, num_blocks, block_ids, split);
if (BROTLI_IS_OOM(m)) return;
BROTLI_FREE(m, block_ids);
}
}
#undef HistogramType