| 1 | /* |
| 2 | * Copyright (c) Meta Platforms, Inc. and affiliates. |
| 3 | * All rights reserved. |
| 4 | * |
| 5 | * This source code is licensed under both the BSD-style license (found in the |
| 6 | * LICENSE file in the root directory of this source tree) and the GPLv2 (found |
| 7 | * in the COPYING file in the root directory of this source tree). |
| 8 | * You may select, at your option, one of the above-listed licenses. |
| 9 | */ |
| 10 | |
| 11 | /* ***************************************************************************** |
| 12 | * Constructs a dictionary using a heuristic based on the following paper: |
| 13 | * |
| 14 | * Liao, Petri, Moffat, Wirth |
| 15 | * Effective Construction of Relative Lempel-Ziv Dictionaries |
| 16 | * Published in WWW 2016. |
| 17 | * |
| 18 | * Adapted from code originally written by @ot (Giuseppe Ottaviano). |
| 19 | ******************************************************************************/ |
| 20 | |
| 21 | /*-************************************* |
| 22 | * Dependencies |
| 23 | ***************************************/ |
| 24 | #include <stdio.h> /* fprintf */ |
| 25 | #include <stdlib.h> /* malloc, free, qsort */ |
| 26 | #include <string.h> /* memset */ |
| 27 | #include <time.h> /* clock */ |
| 28 | |
| 29 | #ifndef ZDICT_STATIC_LINKING_ONLY |
| 30 | # define ZDICT_STATIC_LINKING_ONLY |
| 31 | #endif |
| 32 | |
| 33 | #include "../common/mem.h" /* read */ |
| 34 | #include "../common/pool.h" |
| 35 | #include "../common/threading.h" |
| 36 | #include "../common/zstd_internal.h" /* includes zstd.h */ |
| 37 | #include "../common/bits.h" /* ZSTD_highbit32 */ |
| 38 | #include "../zdict.h" |
| 39 | #include "cover.h" |
| 40 | |
| 41 | /*-************************************* |
| 42 | * Constants |
| 43 | ***************************************/ |
| 44 | /** |
| 45 | * There are 32bit indexes used to ref samples, so limit samples size to 4GB |
| 46 | * on 64bit builds. |
| 47 | * For 32bit builds we choose 1 GB. |
| 48 | * Most 32bit platforms have 2GB user-mode addressable space and we allocate a large |
| 49 | * contiguous buffer, so 1GB is already a high limit. |
| 50 | */ |
| 51 | #define COVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((unsigned)-1) : ((unsigned)1 GB)) |
| 52 | #define COVER_DEFAULT_SPLITPOINT 1.0 |
| 53 | |
| 54 | /*-************************************* |
| 55 | * Console display |
| 56 | ***************************************/ |
| 57 | #ifndef LOCALDISPLAYLEVEL |
| 58 | static int g_displayLevel = 0; |
| 59 | #endif |
| 60 | #undef DISPLAY |
| 61 | #define DISPLAY(...) \ |
| 62 | { \ |
| 63 | fprintf(stderr, __VA_ARGS__); \ |
| 64 | fflush(stderr); \ |
| 65 | } |
| 66 | #undef LOCALDISPLAYLEVEL |
| 67 | #define LOCALDISPLAYLEVEL(displayLevel, l, ...) \ |
| 68 | if (displayLevel >= l) { \ |
| 69 | DISPLAY(__VA_ARGS__); \ |
| 70 | } /* 0 : no display; 1: errors; 2: default; 3: details; 4: debug */ |
| 71 | #undef DISPLAYLEVEL |
| 72 | #define DISPLAYLEVEL(l, ...) LOCALDISPLAYLEVEL(g_displayLevel, l, __VA_ARGS__) |
| 73 | |
| 74 | #ifndef LOCALDISPLAYUPDATE |
| 75 | static const clock_t g_refreshRate = CLOCKS_PER_SEC * 15 / 100; |
| 76 | static clock_t g_time = 0; |
| 77 | #endif |
| 78 | #undef LOCALDISPLAYUPDATE |
| 79 | #define LOCALDISPLAYUPDATE(displayLevel, l, ...) \ |
| 80 | if (displayLevel >= l) { \ |
| 81 | if ((clock() - g_time > g_refreshRate) || (displayLevel >= 4)) { \ |
| 82 | g_time = clock(); \ |
| 83 | DISPLAY(__VA_ARGS__); \ |
| 84 | } \ |
| 85 | } |
| 86 | #undef DISPLAYUPDATE |
| 87 | #define DISPLAYUPDATE(l, ...) LOCALDISPLAYUPDATE(g_displayLevel, l, __VA_ARGS__) |
| 88 | |
| 89 | /*-************************************* |
| 90 | * Hash table |
| 91 | *************************************** |
| 92 | * A small specialized hash map for storing activeDmers. |
| 93 | * The map does not resize, so if it becomes full it will loop forever. |
| 94 | * Thus, the map must be large enough to store every value. |
| 95 | * The map implements linear probing and keeps its load less than 0.5. |
| 96 | */ |
| 97 | |
| 98 | #define MAP_EMPTY_VALUE ((U32)-1) |
| 99 | typedef struct COVER_map_pair_t_s { |
| 100 | U32 key; |
| 101 | U32 value; |
| 102 | } COVER_map_pair_t; |
| 103 | |
| 104 | typedef struct COVER_map_s { |
| 105 | COVER_map_pair_t *data; |
| 106 | U32 sizeLog; |
| 107 | U32 size; |
| 108 | U32 sizeMask; |
| 109 | } COVER_map_t; |
| 110 | |
| 111 | /** |
| 112 | * Clear the map. |
| 113 | */ |
| 114 | static void COVER_map_clear(COVER_map_t *map) { |
| 115 | memset(map->data, MAP_EMPTY_VALUE, map->size * sizeof(COVER_map_pair_t)); |
| 116 | } |
| 117 | |
| 118 | /** |
| 119 | * Initializes a map of the given size. |
| 120 | * Returns 1 on success and 0 on failure. |
| 121 | * The map must be destroyed with COVER_map_destroy(). |
| 122 | * The map is only guaranteed to be large enough to hold size elements. |
| 123 | */ |
| 124 | static int COVER_map_init(COVER_map_t *map, U32 size) { |
| 125 | map->sizeLog = ZSTD_highbit32(size) + 2; |
| 126 | map->size = (U32)1 << map->sizeLog; |
| 127 | map->sizeMask = map->size - 1; |
| 128 | map->data = (COVER_map_pair_t *)malloc(map->size * sizeof(COVER_map_pair_t)); |
| 129 | if (!map->data) { |
| 130 | map->sizeLog = 0; |
| 131 | map->size = 0; |
| 132 | return 0; |
| 133 | } |
| 134 | COVER_map_clear(map); |
| 135 | return 1; |
| 136 | } |
| 137 | |
| 138 | /** |
| 139 | * Internal hash function |
| 140 | */ |
| 141 | static const U32 COVER_prime4bytes = 2654435761U; |
| 142 | static U32 COVER_map_hash(COVER_map_t *map, U32 key) { |
| 143 | return (key * COVER_prime4bytes) >> (32 - map->sizeLog); |
| 144 | } |
| 145 | |
| 146 | /** |
| 147 | * Helper function that returns the index that a key should be placed into. |
| 148 | */ |
| 149 | static U32 COVER_map_index(COVER_map_t *map, U32 key) { |
| 150 | const U32 hash = COVER_map_hash(map, key); |
| 151 | U32 i; |
| 152 | for (i = hash;; i = (i + 1) & map->sizeMask) { |
| 153 | COVER_map_pair_t *pos = &map->data[i]; |
| 154 | if (pos->value == MAP_EMPTY_VALUE) { |
| 155 | return i; |
| 156 | } |
| 157 | if (pos->key == key) { |
| 158 | return i; |
| 159 | } |
| 160 | } |
| 161 | } |
| 162 | |
| 163 | /** |
| 164 | * Returns the pointer to the value for key. |
| 165 | * If key is not in the map, it is inserted and the value is set to 0. |
| 166 | * The map must not be full. |
| 167 | */ |
| 168 | static U32 *COVER_map_at(COVER_map_t *map, U32 key) { |
| 169 | COVER_map_pair_t *pos = &map->data[COVER_map_index(map, key)]; |
| 170 | if (pos->value == MAP_EMPTY_VALUE) { |
| 171 | pos->key = key; |
| 172 | pos->value = 0; |
| 173 | } |
| 174 | return &pos->value; |
| 175 | } |
| 176 | |
| 177 | /** |
| 178 | * Deletes key from the map if present. |
| 179 | */ |
| 180 | static void COVER_map_remove(COVER_map_t *map, U32 key) { |
| 181 | U32 i = COVER_map_index(map, key); |
| 182 | COVER_map_pair_t *del = &map->data[i]; |
| 183 | U32 shift = 1; |
| 184 | if (del->value == MAP_EMPTY_VALUE) { |
| 185 | return; |
| 186 | } |
| 187 | for (i = (i + 1) & map->sizeMask;; i = (i + 1) & map->sizeMask) { |
| 188 | COVER_map_pair_t *const pos = &map->data[i]; |
| 189 | /* If the position is empty we are done */ |
| 190 | if (pos->value == MAP_EMPTY_VALUE) { |
| 191 | del->value = MAP_EMPTY_VALUE; |
| 192 | return; |
| 193 | } |
| 194 | /* If pos can be moved to del do so */ |
| 195 | if (((i - COVER_map_hash(map, pos->key)) & map->sizeMask) >= shift) { |
| 196 | del->key = pos->key; |
| 197 | del->value = pos->value; |
| 198 | del = pos; |
| 199 | shift = 1; |
| 200 | } else { |
| 201 | ++shift; |
| 202 | } |
| 203 | } |
| 204 | } |
| 205 | |
| 206 | /** |
| 207 | * Destroys a map that is inited with COVER_map_init(). |
| 208 | */ |
| 209 | static void COVER_map_destroy(COVER_map_t *map) { |
| 210 | if (map->data) { |
| 211 | free(map->data); |
| 212 | } |
| 213 | map->data = NULL; |
| 214 | map->size = 0; |
| 215 | } |
| 216 | |
| 217 | /*-************************************* |
| 218 | * Context |
| 219 | ***************************************/ |
| 220 | |
| 221 | typedef struct { |
| 222 | const BYTE *samples; |
| 223 | size_t *offsets; |
| 224 | const size_t *samplesSizes; |
| 225 | size_t nbSamples; |
| 226 | size_t nbTrainSamples; |
| 227 | size_t nbTestSamples; |
| 228 | U32 *suffix; |
| 229 | size_t suffixSize; |
| 230 | U32 *freqs; |
| 231 | U32 *dmerAt; |
| 232 | unsigned d; |
| 233 | } COVER_ctx_t; |
| 234 | |
| 235 | /* We need a global context for qsort... */ |
| 236 | static COVER_ctx_t *g_coverCtx = NULL; |
| 237 | |
| 238 | /*-************************************* |
| 239 | * Helper functions |
| 240 | ***************************************/ |
| 241 | |
| 242 | /** |
| 243 | * Returns the sum of the sample sizes. |
| 244 | */ |
| 245 | size_t COVER_sum(const size_t *samplesSizes, unsigned nbSamples) { |
| 246 | size_t sum = 0; |
| 247 | unsigned i; |
| 248 | for (i = 0; i < nbSamples; ++i) { |
| 249 | sum += samplesSizes[i]; |
| 250 | } |
| 251 | return sum; |
| 252 | } |
| 253 | |
| 254 | /** |
| 255 | * Returns -1 if the dmer at lp is less than the dmer at rp. |
| 256 | * Return 0 if the dmers at lp and rp are equal. |
| 257 | * Returns 1 if the dmer at lp is greater than the dmer at rp. |
| 258 | */ |
| 259 | static int COVER_cmp(COVER_ctx_t *ctx, const void *lp, const void *rp) { |
| 260 | U32 const lhs = *(U32 const *)lp; |
| 261 | U32 const rhs = *(U32 const *)rp; |
| 262 | return memcmp(ctx->samples + lhs, ctx->samples + rhs, ctx->d); |
| 263 | } |
| 264 | /** |
| 265 | * Faster version for d <= 8. |
| 266 | */ |
| 267 | static int COVER_cmp8(COVER_ctx_t *ctx, const void *lp, const void *rp) { |
| 268 | U64 const mask = (ctx->d == 8) ? (U64)-1 : (((U64)1 << (8 * ctx->d)) - 1); |
| 269 | U64 const lhs = MEM_readLE64(ctx->samples + *(U32 const *)lp) & mask; |
| 270 | U64 const rhs = MEM_readLE64(ctx->samples + *(U32 const *)rp) & mask; |
| 271 | if (lhs < rhs) { |
| 272 | return -1; |
| 273 | } |
| 274 | return (lhs > rhs); |
| 275 | } |
| 276 | |
| 277 | /** |
| 278 | * Same as COVER_cmp() except ties are broken by pointer value |
| 279 | * NOTE: g_coverCtx must be set to call this function. A global is required because |
| 280 | * qsort doesn't take an opaque pointer. |
| 281 | */ |
| 282 | static int WIN_CDECL COVER_strict_cmp(const void *lp, const void *rp) { |
| 283 | int result = COVER_cmp(g_coverCtx, lp, rp); |
| 284 | if (result == 0) { |
| 285 | result = lp < rp ? -1 : 1; |
| 286 | } |
| 287 | return result; |
| 288 | } |
| 289 | /** |
| 290 | * Faster version for d <= 8. |
| 291 | */ |
| 292 | static int WIN_CDECL COVER_strict_cmp8(const void *lp, const void *rp) { |
| 293 | int result = COVER_cmp8(g_coverCtx, lp, rp); |
| 294 | if (result == 0) { |
| 295 | result = lp < rp ? -1 : 1; |
| 296 | } |
| 297 | return result; |
| 298 | } |
| 299 | |
| 300 | /** |
| 301 | * Returns the first pointer in [first, last) whose element does not compare |
| 302 | * less than value. If no such element exists it returns last. |
| 303 | */ |
| 304 | static const size_t *COVER_lower_bound(const size_t *first, const size_t *last, |
| 305 | size_t value) { |
| 306 | size_t count = last - first; |
| 307 | while (count != 0) { |
| 308 | size_t step = count / 2; |
| 309 | const size_t *ptr = first; |
| 310 | ptr += step; |
| 311 | if (*ptr < value) { |
| 312 | first = ++ptr; |
| 313 | count -= step + 1; |
| 314 | } else { |
| 315 | count = step; |
| 316 | } |
| 317 | } |
| 318 | return first; |
| 319 | } |
| 320 | |
| 321 | /** |
| 322 | * Generic groupBy function. |
| 323 | * Groups an array sorted by cmp into groups with equivalent values. |
| 324 | * Calls grp for each group. |
| 325 | */ |
| 326 | static void |
| 327 | COVER_groupBy(const void *data, size_t count, size_t size, COVER_ctx_t *ctx, |
| 328 | int (*cmp)(COVER_ctx_t *, const void *, const void *), |
| 329 | void (*grp)(COVER_ctx_t *, const void *, const void *)) { |
| 330 | const BYTE *ptr = (const BYTE *)data; |
| 331 | size_t num = 0; |
| 332 | while (num < count) { |
| 333 | const BYTE *grpEnd = ptr + size; |
| 334 | ++num; |
| 335 | while (num < count && cmp(ctx, ptr, grpEnd) == 0) { |
| 336 | grpEnd += size; |
| 337 | ++num; |
| 338 | } |
| 339 | grp(ctx, ptr, grpEnd); |
| 340 | ptr = grpEnd; |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | /*-************************************* |
| 345 | * Cover functions |
| 346 | ***************************************/ |
| 347 | |
| 348 | /** |
| 349 | * Called on each group of positions with the same dmer. |
| 350 | * Counts the frequency of each dmer and saves it in the suffix array. |
| 351 | * Fills `ctx->dmerAt`. |
| 352 | */ |
| 353 | static void COVER_group(COVER_ctx_t *ctx, const void *group, |
| 354 | const void *groupEnd) { |
| 355 | /* The group consists of all the positions with the same first d bytes. */ |
| 356 | const U32 *grpPtr = (const U32 *)group; |
| 357 | const U32 *grpEnd = (const U32 *)groupEnd; |
| 358 | /* The dmerId is how we will reference this dmer. |
| 359 | * This allows us to map the whole dmer space to a much smaller space, the |
| 360 | * size of the suffix array. |
| 361 | */ |
| 362 | const U32 dmerId = (U32)(grpPtr - ctx->suffix); |
| 363 | /* Count the number of samples this dmer shows up in */ |
| 364 | U32 freq = 0; |
| 365 | /* Details */ |
| 366 | const size_t *curOffsetPtr = ctx->offsets; |
| 367 | const size_t *offsetsEnd = ctx->offsets + ctx->nbSamples; |
| 368 | /* Once *grpPtr >= curSampleEnd this occurrence of the dmer is in a |
| 369 | * different sample than the last. |
| 370 | */ |
| 371 | size_t curSampleEnd = ctx->offsets[0]; |
| 372 | for (; grpPtr != grpEnd; ++grpPtr) { |
| 373 | /* Save the dmerId for this position so we can get back to it. */ |
| 374 | ctx->dmerAt[*grpPtr] = dmerId; |
| 375 | /* Dictionaries only help for the first reference to the dmer. |
| 376 | * After that zstd can reference the match from the previous reference. |
| 377 | * So only count each dmer once for each sample it is in. |
| 378 | */ |
| 379 | if (*grpPtr < curSampleEnd) { |
| 380 | continue; |
| 381 | } |
| 382 | freq += 1; |
| 383 | /* Binary search to find the end of the sample *grpPtr is in. |
| 384 | * In the common case that grpPtr + 1 == grpEnd we can skip the binary |
| 385 | * search because the loop is over. |
| 386 | */ |
| 387 | if (grpPtr + 1 != grpEnd) { |
| 388 | const size_t *sampleEndPtr = |
| 389 | COVER_lower_bound(curOffsetPtr, offsetsEnd, *grpPtr); |
| 390 | curSampleEnd = *sampleEndPtr; |
| 391 | curOffsetPtr = sampleEndPtr + 1; |
| 392 | } |
| 393 | } |
| 394 | /* At this point we are never going to look at this segment of the suffix |
| 395 | * array again. We take advantage of this fact to save memory. |
| 396 | * We store the frequency of the dmer in the first position of the group, |
| 397 | * which is dmerId. |
| 398 | */ |
| 399 | ctx->suffix[dmerId] = freq; |
| 400 | } |
| 401 | |
| 402 | |
| 403 | /** |
| 404 | * Selects the best segment in an epoch. |
| 405 | * Segments of are scored according to the function: |
| 406 | * |
| 407 | * Let F(d) be the frequency of dmer d. |
| 408 | * Let S_i be the dmer at position i of segment S which has length k. |
| 409 | * |
| 410 | * Score(S) = F(S_1) + F(S_2) + ... + F(S_{k-d+1}) |
| 411 | * |
| 412 | * Once the dmer d is in the dictionary we set F(d) = 0. |
| 413 | */ |
| 414 | static COVER_segment_t COVER_selectSegment(const COVER_ctx_t *ctx, U32 *freqs, |
| 415 | COVER_map_t *activeDmers, U32 begin, |
| 416 | U32 end, |
| 417 | ZDICT_cover_params_t parameters) { |
| 418 | /* Constants */ |
| 419 | const U32 k = parameters.k; |
| 420 | const U32 d = parameters.d; |
| 421 | const U32 dmersInK = k - d + 1; |
| 422 | /* Try each segment (activeSegment) and save the best (bestSegment) */ |
| 423 | COVER_segment_t bestSegment = {0, 0, 0}; |
| 424 | COVER_segment_t activeSegment; |
| 425 | /* Reset the activeDmers in the segment */ |
| 426 | COVER_map_clear(activeDmers); |
| 427 | /* The activeSegment starts at the beginning of the epoch. */ |
| 428 | activeSegment.begin = begin; |
| 429 | activeSegment.end = begin; |
| 430 | activeSegment.score = 0; |
| 431 | /* Slide the activeSegment through the whole epoch. |
| 432 | * Save the best segment in bestSegment. |
| 433 | */ |
| 434 | while (activeSegment.end < end) { |
| 435 | /* The dmerId for the dmer at the next position */ |
| 436 | U32 newDmer = ctx->dmerAt[activeSegment.end]; |
| 437 | /* The entry in activeDmers for this dmerId */ |
| 438 | U32 *newDmerOcc = COVER_map_at(activeDmers, newDmer); |
| 439 | /* If the dmer isn't already present in the segment add its score. */ |
| 440 | if (*newDmerOcc == 0) { |
| 441 | /* The paper suggest using the L-0.5 norm, but experiments show that it |
| 442 | * doesn't help. |
| 443 | */ |
| 444 | activeSegment.score += freqs[newDmer]; |
| 445 | } |
| 446 | /* Add the dmer to the segment */ |
| 447 | activeSegment.end += 1; |
| 448 | *newDmerOcc += 1; |
| 449 | |
| 450 | /* If the window is now too large, drop the first position */ |
| 451 | if (activeSegment.end - activeSegment.begin == dmersInK + 1) { |
| 452 | U32 delDmer = ctx->dmerAt[activeSegment.begin]; |
| 453 | U32 *delDmerOcc = COVER_map_at(activeDmers, delDmer); |
| 454 | activeSegment.begin += 1; |
| 455 | *delDmerOcc -= 1; |
| 456 | /* If this is the last occurrence of the dmer, subtract its score */ |
| 457 | if (*delDmerOcc == 0) { |
| 458 | COVER_map_remove(activeDmers, delDmer); |
| 459 | activeSegment.score -= freqs[delDmer]; |
| 460 | } |
| 461 | } |
| 462 | |
| 463 | /* If this segment is the best so far save it */ |
| 464 | if (activeSegment.score > bestSegment.score) { |
| 465 | bestSegment = activeSegment; |
| 466 | } |
| 467 | } |
| 468 | { |
| 469 | /* Trim off the zero frequency head and tail from the segment. */ |
| 470 | U32 newBegin = bestSegment.end; |
| 471 | U32 newEnd = bestSegment.begin; |
| 472 | U32 pos; |
| 473 | for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) { |
| 474 | U32 freq = freqs[ctx->dmerAt[pos]]; |
| 475 | if (freq != 0) { |
| 476 | newBegin = MIN(newBegin, pos); |
| 477 | newEnd = pos + 1; |
| 478 | } |
| 479 | } |
| 480 | bestSegment.begin = newBegin; |
| 481 | bestSegment.end = newEnd; |
| 482 | } |
| 483 | { |
| 484 | /* Zero out the frequency of each dmer covered by the chosen segment. */ |
| 485 | U32 pos; |
| 486 | for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) { |
| 487 | freqs[ctx->dmerAt[pos]] = 0; |
| 488 | } |
| 489 | } |
| 490 | return bestSegment; |
| 491 | } |
| 492 | |
| 493 | /** |
| 494 | * Check the validity of the parameters. |
| 495 | * Returns non-zero if the parameters are valid and 0 otherwise. |
| 496 | */ |
| 497 | static int COVER_checkParameters(ZDICT_cover_params_t parameters, |
| 498 | size_t maxDictSize) { |
| 499 | /* k and d are required parameters */ |
| 500 | if (parameters.d == 0 || parameters.k == 0) { |
| 501 | return 0; |
| 502 | } |
| 503 | /* k <= maxDictSize */ |
| 504 | if (parameters.k > maxDictSize) { |
| 505 | return 0; |
| 506 | } |
| 507 | /* d <= k */ |
| 508 | if (parameters.d > parameters.k) { |
| 509 | return 0; |
| 510 | } |
| 511 | /* 0 < splitPoint <= 1 */ |
| 512 | if (parameters.splitPoint <= 0 || parameters.splitPoint > 1){ |
| 513 | return 0; |
| 514 | } |
| 515 | return 1; |
| 516 | } |
| 517 | |
| 518 | /** |
| 519 | * Clean up a context initialized with `COVER_ctx_init()`. |
| 520 | */ |
| 521 | static void COVER_ctx_destroy(COVER_ctx_t *ctx) { |
| 522 | if (!ctx) { |
| 523 | return; |
| 524 | } |
| 525 | if (ctx->suffix) { |
| 526 | free(ctx->suffix); |
| 527 | ctx->suffix = NULL; |
| 528 | } |
| 529 | if (ctx->freqs) { |
| 530 | free(ctx->freqs); |
| 531 | ctx->freqs = NULL; |
| 532 | } |
| 533 | if (ctx->dmerAt) { |
| 534 | free(ctx->dmerAt); |
| 535 | ctx->dmerAt = NULL; |
| 536 | } |
| 537 | if (ctx->offsets) { |
| 538 | free(ctx->offsets); |
| 539 | ctx->offsets = NULL; |
| 540 | } |
| 541 | } |
| 542 | |
| 543 | /** |
| 544 | * Prepare a context for dictionary building. |
| 545 | * The context is only dependent on the parameter `d` and can be used multiple |
| 546 | * times. |
| 547 | * Returns 0 on success or error code on error. |
| 548 | * The context must be destroyed with `COVER_ctx_destroy()`. |
| 549 | */ |
| 550 | static size_t COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer, |
| 551 | const size_t *samplesSizes, unsigned nbSamples, |
| 552 | unsigned d, double splitPoint) { |
| 553 | const BYTE *const samples = (const BYTE *)samplesBuffer; |
| 554 | const size_t totalSamplesSize = COVER_sum(samplesSizes, nbSamples); |
| 555 | /* Split samples into testing and training sets */ |
| 556 | const unsigned nbTrainSamples = splitPoint < 1.0 ? (unsigned)((double)nbSamples * splitPoint) : nbSamples; |
| 557 | const unsigned nbTestSamples = splitPoint < 1.0 ? nbSamples - nbTrainSamples : nbSamples; |
| 558 | const size_t trainingSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes, nbTrainSamples) : totalSamplesSize; |
| 559 | const size_t testSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes + nbTrainSamples, nbTestSamples) : totalSamplesSize; |
| 560 | /* Checks */ |
| 561 | if (totalSamplesSize < MAX(d, sizeof(U64)) || |
| 562 | totalSamplesSize >= (size_t)COVER_MAX_SAMPLES_SIZE) { |
| 563 | DISPLAYLEVEL(1, "Total samples size is too large (%u MB), maximum size is %u MB\n", |
| 564 | (unsigned)(totalSamplesSize>>20), (COVER_MAX_SAMPLES_SIZE >> 20)); |
| 565 | return ERROR(srcSize_wrong); |
| 566 | } |
| 567 | /* Check if there are at least 5 training samples */ |
| 568 | if (nbTrainSamples < 5) { |
| 569 | DISPLAYLEVEL(1, "Total number of training samples is %u and is invalid.", nbTrainSamples); |
| 570 | return ERROR(srcSize_wrong); |
| 571 | } |
| 572 | /* Check if there's testing sample */ |
| 573 | if (nbTestSamples < 1) { |
| 574 | DISPLAYLEVEL(1, "Total number of testing samples is %u and is invalid.", nbTestSamples); |
| 575 | return ERROR(srcSize_wrong); |
| 576 | } |
| 577 | /* Zero the context */ |
| 578 | memset(ctx, 0, sizeof(*ctx)); |
| 579 | DISPLAYLEVEL(2, "Training on %u samples of total size %u\n", nbTrainSamples, |
| 580 | (unsigned)trainingSamplesSize); |
| 581 | DISPLAYLEVEL(2, "Testing on %u samples of total size %u\n", nbTestSamples, |
| 582 | (unsigned)testSamplesSize); |
| 583 | ctx->samples = samples; |
| 584 | ctx->samplesSizes = samplesSizes; |
| 585 | ctx->nbSamples = nbSamples; |
| 586 | ctx->nbTrainSamples = nbTrainSamples; |
| 587 | ctx->nbTestSamples = nbTestSamples; |
| 588 | /* Partial suffix array */ |
| 589 | ctx->suffixSize = trainingSamplesSize - MAX(d, sizeof(U64)) + 1; |
| 590 | ctx->suffix = (U32 *)malloc(ctx->suffixSize * sizeof(U32)); |
| 591 | /* Maps index to the dmerID */ |
| 592 | ctx->dmerAt = (U32 *)malloc(ctx->suffixSize * sizeof(U32)); |
| 593 | /* The offsets of each file */ |
| 594 | ctx->offsets = (size_t *)malloc((nbSamples + 1) * sizeof(size_t)); |
| 595 | if (!ctx->suffix || !ctx->dmerAt || !ctx->offsets) { |
| 596 | DISPLAYLEVEL(1, "Failed to allocate scratch buffers\n"); |
| 597 | COVER_ctx_destroy(ctx); |
| 598 | return ERROR(memory_allocation); |
| 599 | } |
| 600 | ctx->freqs = NULL; |
| 601 | ctx->d = d; |
| 602 | |
| 603 | /* Fill offsets from the samplesSizes */ |
| 604 | { |
| 605 | U32 i; |
| 606 | ctx->offsets[0] = 0; |
| 607 | for (i = 1; i <= nbSamples; ++i) { |
| 608 | ctx->offsets[i] = ctx->offsets[i - 1] + samplesSizes[i - 1]; |
| 609 | } |
| 610 | } |
| 611 | DISPLAYLEVEL(2, "Constructing partial suffix array\n"); |
| 612 | { |
| 613 | /* suffix is a partial suffix array. |
| 614 | * It only sorts suffixes by their first parameters.d bytes. |
| 615 | * The sort is stable, so each dmer group is sorted by position in input. |
| 616 | */ |
| 617 | U32 i; |
| 618 | for (i = 0; i < ctx->suffixSize; ++i) { |
| 619 | ctx->suffix[i] = i; |
| 620 | } |
| 621 | /* qsort doesn't take an opaque pointer, so pass as a global. |
| 622 | * On OpenBSD qsort() is not guaranteed to be stable, their mergesort() is. |
| 623 | */ |
| 624 | g_coverCtx = ctx; |
| 625 | #if defined(__OpenBSD__) |
| 626 | mergesort(ctx->suffix, ctx->suffixSize, sizeof(U32), |
| 627 | (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp)); |
| 628 | #else |
| 629 | qsort(ctx->suffix, ctx->suffixSize, sizeof(U32), |
| 630 | (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp)); |
| 631 | #endif |
| 632 | } |
| 633 | DISPLAYLEVEL(2, "Computing frequencies\n"); |
| 634 | /* For each dmer group (group of positions with the same first d bytes): |
| 635 | * 1. For each position we set dmerAt[position] = dmerID. The dmerID is |
| 636 | * (groupBeginPtr - suffix). This allows us to go from position to |
| 637 | * dmerID so we can look up values in freq. |
| 638 | * 2. We calculate how many samples the dmer occurs in and save it in |
| 639 | * freqs[dmerId]. |
| 640 | */ |
| 641 | COVER_groupBy(ctx->suffix, ctx->suffixSize, sizeof(U32), ctx, |
| 642 | (ctx->d <= 8 ? &COVER_cmp8 : &COVER_cmp), &COVER_group); |
| 643 | ctx->freqs = ctx->suffix; |
| 644 | ctx->suffix = NULL; |
| 645 | return 0; |
| 646 | } |
| 647 | |
| 648 | void COVER_warnOnSmallCorpus(size_t maxDictSize, size_t nbDmers, int displayLevel) |
| 649 | { |
| 650 | const double ratio = (double)nbDmers / (double)maxDictSize; |
| 651 | if (ratio >= 10) { |
| 652 | return; |
| 653 | } |
| 654 | LOCALDISPLAYLEVEL(displayLevel, 1, |
| 655 | "WARNING: The maximum dictionary size %u is too large " |
| 656 | "compared to the source size %u! " |
| 657 | "size(source)/size(dictionary) = %f, but it should be >= " |
| 658 | "10! This may lead to a subpar dictionary! We recommend " |
| 659 | "training on sources at least 10x, and preferably 100x " |
| 660 | "the size of the dictionary! \n", (U32)maxDictSize, |
| 661 | (U32)nbDmers, ratio); |
| 662 | } |
| 663 | |
| 664 | COVER_epoch_info_t COVER_computeEpochs(U32 maxDictSize, |
| 665 | U32 nbDmers, U32 k, U32 passes) |
| 666 | { |
| 667 | const U32 minEpochSize = k * 10; |
| 668 | COVER_epoch_info_t epochs; |
| 669 | epochs.num = MAX(1, maxDictSize / k / passes); |
| 670 | epochs.size = nbDmers / epochs.num; |
| 671 | if (epochs.size >= minEpochSize) { |
| 672 | assert(epochs.size * epochs.num <= nbDmers); |
| 673 | return epochs; |
| 674 | } |
| 675 | epochs.size = MIN(minEpochSize, nbDmers); |
| 676 | epochs.num = nbDmers / epochs.size; |
| 677 | assert(epochs.size * epochs.num <= nbDmers); |
| 678 | return epochs; |
| 679 | } |
| 680 | |
| 681 | /** |
| 682 | * Given the prepared context build the dictionary. |
| 683 | */ |
| 684 | static size_t COVER_buildDictionary(const COVER_ctx_t *ctx, U32 *freqs, |
| 685 | COVER_map_t *activeDmers, void *dictBuffer, |
| 686 | size_t dictBufferCapacity, |
| 687 | ZDICT_cover_params_t parameters) { |
| 688 | BYTE *const dict = (BYTE *)dictBuffer; |
| 689 | size_t tail = dictBufferCapacity; |
| 690 | /* Divide the data into epochs. We will select one segment from each epoch. */ |
| 691 | const COVER_epoch_info_t epochs = COVER_computeEpochs( |
| 692 | (U32)dictBufferCapacity, (U32)ctx->suffixSize, parameters.k, 4); |
| 693 | const size_t maxZeroScoreRun = MAX(10, MIN(100, epochs.num >> 3)); |
| 694 | size_t zeroScoreRun = 0; |
| 695 | size_t epoch; |
| 696 | DISPLAYLEVEL(2, "Breaking content into %u epochs of size %u\n", |
| 697 | (U32)epochs.num, (U32)epochs.size); |
| 698 | /* Loop through the epochs until there are no more segments or the dictionary |
| 699 | * is full. |
| 700 | */ |
| 701 | for (epoch = 0; tail > 0; epoch = (epoch + 1) % epochs.num) { |
| 702 | const U32 epochBegin = (U32)(epoch * epochs.size); |
| 703 | const U32 epochEnd = epochBegin + epochs.size; |
| 704 | size_t segmentSize; |
| 705 | /* Select a segment */ |
| 706 | COVER_segment_t segment = COVER_selectSegment( |
| 707 | ctx, freqs, activeDmers, epochBegin, epochEnd, parameters); |
| 708 | /* If the segment covers no dmers, then we are out of content. |
| 709 | * There may be new content in other epochs, for continue for some time. |
| 710 | */ |
| 711 | if (segment.score == 0) { |
| 712 | if (++zeroScoreRun >= maxZeroScoreRun) { |
| 713 | break; |
| 714 | } |
| 715 | continue; |
| 716 | } |
| 717 | zeroScoreRun = 0; |
| 718 | /* Trim the segment if necessary and if it is too small then we are done */ |
| 719 | segmentSize = MIN(segment.end - segment.begin + parameters.d - 1, tail); |
| 720 | if (segmentSize < parameters.d) { |
| 721 | break; |
| 722 | } |
| 723 | /* We fill the dictionary from the back to allow the best segments to be |
| 724 | * referenced with the smallest offsets. |
| 725 | */ |
| 726 | tail -= segmentSize; |
| 727 | memcpy(dict + tail, ctx->samples + segment.begin, segmentSize); |
| 728 | DISPLAYUPDATE( |
| 729 | 2, "\r%u%% ", |
| 730 | (unsigned)(((dictBufferCapacity - tail) * 100) / dictBufferCapacity)); |
| 731 | } |
| 732 | DISPLAYLEVEL(2, "\r%79s\r", ""); |
| 733 | return tail; |
| 734 | } |
| 735 | |
| 736 | ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover( |
| 737 | void *dictBuffer, size_t dictBufferCapacity, |
| 738 | const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples, |
| 739 | ZDICT_cover_params_t parameters) |
| 740 | { |
| 741 | BYTE* const dict = (BYTE*)dictBuffer; |
| 742 | COVER_ctx_t ctx; |
| 743 | COVER_map_t activeDmers; |
| 744 | parameters.splitPoint = 1.0; |
| 745 | /* Initialize global data */ |
| 746 | g_displayLevel = (int)parameters.zParams.notificationLevel; |
| 747 | /* Checks */ |
| 748 | if (!COVER_checkParameters(parameters, dictBufferCapacity)) { |
| 749 | DISPLAYLEVEL(1, "Cover parameters incorrect\n"); |
| 750 | return ERROR(parameter_outOfBound); |
| 751 | } |
| 752 | if (nbSamples == 0) { |
| 753 | DISPLAYLEVEL(1, "Cover must have at least one input file\n"); |
| 754 | return ERROR(srcSize_wrong); |
| 755 | } |
| 756 | if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) { |
| 757 | DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n", |
| 758 | ZDICT_DICTSIZE_MIN); |
| 759 | return ERROR(dstSize_tooSmall); |
| 760 | } |
| 761 | /* Initialize context and activeDmers */ |
| 762 | { |
| 763 | size_t const initVal = COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, |
| 764 | parameters.d, parameters.splitPoint); |
| 765 | if (ZSTD_isError(initVal)) { |
| 766 | return initVal; |
| 767 | } |
| 768 | } |
| 769 | COVER_warnOnSmallCorpus(dictBufferCapacity, ctx.suffixSize, g_displayLevel); |
| 770 | if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) { |
| 771 | DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n"); |
| 772 | COVER_ctx_destroy(&ctx); |
| 773 | return ERROR(memory_allocation); |
| 774 | } |
| 775 | |
| 776 | DISPLAYLEVEL(2, "Building dictionary\n"); |
| 777 | { |
| 778 | const size_t tail = |
| 779 | COVER_buildDictionary(&ctx, ctx.freqs, &activeDmers, dictBuffer, |
| 780 | dictBufferCapacity, parameters); |
| 781 | const size_t dictionarySize = ZDICT_finalizeDictionary( |
| 782 | dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail, |
| 783 | samplesBuffer, samplesSizes, nbSamples, parameters.zParams); |
| 784 | if (!ZSTD_isError(dictionarySize)) { |
| 785 | DISPLAYLEVEL(2, "Constructed dictionary of size %u\n", |
| 786 | (unsigned)dictionarySize); |
| 787 | } |
| 788 | COVER_ctx_destroy(&ctx); |
| 789 | COVER_map_destroy(&activeDmers); |
| 790 | return dictionarySize; |
| 791 | } |
| 792 | } |
| 793 | |
| 794 | |
| 795 | |
| 796 | size_t COVER_checkTotalCompressedSize(const ZDICT_cover_params_t parameters, |
| 797 | const size_t *samplesSizes, const BYTE *samples, |
| 798 | size_t *offsets, |
| 799 | size_t nbTrainSamples, size_t nbSamples, |
| 800 | BYTE *const dict, size_t dictBufferCapacity) { |
| 801 | size_t totalCompressedSize = ERROR(GENERIC); |
| 802 | /* Pointers */ |
| 803 | ZSTD_CCtx *cctx; |
| 804 | ZSTD_CDict *cdict; |
| 805 | void *dst; |
| 806 | /* Local variables */ |
| 807 | size_t dstCapacity; |
| 808 | size_t i; |
| 809 | /* Allocate dst with enough space to compress the maximum sized sample */ |
| 810 | { |
| 811 | size_t maxSampleSize = 0; |
| 812 | i = parameters.splitPoint < 1.0 ? nbTrainSamples : 0; |
| 813 | for (; i < nbSamples; ++i) { |
| 814 | maxSampleSize = MAX(samplesSizes[i], maxSampleSize); |
| 815 | } |
| 816 | dstCapacity = ZSTD_compressBound(maxSampleSize); |
| 817 | dst = malloc(dstCapacity); |
| 818 | } |
| 819 | /* Create the cctx and cdict */ |
| 820 | cctx = ZSTD_createCCtx(); |
| 821 | cdict = ZSTD_createCDict(dict, dictBufferCapacity, |
| 822 | parameters.zParams.compressionLevel); |
| 823 | if (!dst || !cctx || !cdict) { |
| 824 | goto _compressCleanup; |
| 825 | } |
| 826 | /* Compress each sample and sum their sizes (or error) */ |
| 827 | totalCompressedSize = dictBufferCapacity; |
| 828 | i = parameters.splitPoint < 1.0 ? nbTrainSamples : 0; |
| 829 | for (; i < nbSamples; ++i) { |
| 830 | const size_t size = ZSTD_compress_usingCDict( |
| 831 | cctx, dst, dstCapacity, samples + offsets[i], |
| 832 | samplesSizes[i], cdict); |
| 833 | if (ZSTD_isError(size)) { |
| 834 | totalCompressedSize = size; |
| 835 | goto _compressCleanup; |
| 836 | } |
| 837 | totalCompressedSize += size; |
| 838 | } |
| 839 | _compressCleanup: |
| 840 | ZSTD_freeCCtx(cctx); |
| 841 | ZSTD_freeCDict(cdict); |
| 842 | if (dst) { |
| 843 | free(dst); |
| 844 | } |
| 845 | return totalCompressedSize; |
| 846 | } |
| 847 | |
| 848 | |
| 849 | /** |
| 850 | * Initialize the `COVER_best_t`. |
| 851 | */ |
| 852 | void COVER_best_init(COVER_best_t *best) { |
| 853 | if (best==NULL) return; /* compatible with init on NULL */ |
| 854 | (void)ZSTD_pthread_mutex_init(&best->mutex, NULL); |
| 855 | (void)ZSTD_pthread_cond_init(&best->cond, NULL); |
| 856 | best->liveJobs = 0; |
| 857 | best->dict = NULL; |
| 858 | best->dictSize = 0; |
| 859 | best->compressedSize = (size_t)-1; |
| 860 | memset(&best->parameters, 0, sizeof(best->parameters)); |
| 861 | } |
| 862 | |
| 863 | /** |
| 864 | * Wait until liveJobs == 0. |
| 865 | */ |
| 866 | void COVER_best_wait(COVER_best_t *best) { |
| 867 | if (!best) { |
| 868 | return; |
| 869 | } |
| 870 | ZSTD_pthread_mutex_lock(&best->mutex); |
| 871 | while (best->liveJobs != 0) { |
| 872 | ZSTD_pthread_cond_wait(&best->cond, &best->mutex); |
| 873 | } |
| 874 | ZSTD_pthread_mutex_unlock(&best->mutex); |
| 875 | } |
| 876 | |
| 877 | /** |
| 878 | * Call COVER_best_wait() and then destroy the COVER_best_t. |
| 879 | */ |
| 880 | void COVER_best_destroy(COVER_best_t *best) { |
| 881 | if (!best) { |
| 882 | return; |
| 883 | } |
| 884 | COVER_best_wait(best); |
| 885 | if (best->dict) { |
| 886 | free(best->dict); |
| 887 | } |
| 888 | ZSTD_pthread_mutex_destroy(&best->mutex); |
| 889 | ZSTD_pthread_cond_destroy(&best->cond); |
| 890 | } |
| 891 | |
| 892 | /** |
| 893 | * Called when a thread is about to be launched. |
| 894 | * Increments liveJobs. |
| 895 | */ |
| 896 | void COVER_best_start(COVER_best_t *best) { |
| 897 | if (!best) { |
| 898 | return; |
| 899 | } |
| 900 | ZSTD_pthread_mutex_lock(&best->mutex); |
| 901 | ++best->liveJobs; |
| 902 | ZSTD_pthread_mutex_unlock(&best->mutex); |
| 903 | } |
| 904 | |
| 905 | /** |
| 906 | * Called when a thread finishes executing, both on error or success. |
| 907 | * Decrements liveJobs and signals any waiting threads if liveJobs == 0. |
| 908 | * If this dictionary is the best so far save it and its parameters. |
| 909 | */ |
| 910 | void COVER_best_finish(COVER_best_t *best, ZDICT_cover_params_t parameters, |
| 911 | COVER_dictSelection_t selection) { |
| 912 | void* dict = selection.dictContent; |
| 913 | size_t compressedSize = selection.totalCompressedSize; |
| 914 | size_t dictSize = selection.dictSize; |
| 915 | if (!best) { |
| 916 | return; |
| 917 | } |
| 918 | { |
| 919 | size_t liveJobs; |
| 920 | ZSTD_pthread_mutex_lock(&best->mutex); |
| 921 | --best->liveJobs; |
| 922 | liveJobs = best->liveJobs; |
| 923 | /* If the new dictionary is better */ |
| 924 | if (compressedSize < best->compressedSize) { |
| 925 | /* Allocate space if necessary */ |
| 926 | if (!best->dict || best->dictSize < dictSize) { |
| 927 | if (best->dict) { |
| 928 | free(best->dict); |
| 929 | } |
| 930 | best->dict = malloc(dictSize); |
| 931 | if (!best->dict) { |
| 932 | best->compressedSize = ERROR(GENERIC); |
| 933 | best->dictSize = 0; |
| 934 | ZSTD_pthread_cond_signal(&best->cond); |
| 935 | ZSTD_pthread_mutex_unlock(&best->mutex); |
| 936 | return; |
| 937 | } |
| 938 | } |
| 939 | /* Save the dictionary, parameters, and size */ |
| 940 | if (dict) { |
| 941 | memcpy(best->dict, dict, dictSize); |
| 942 | best->dictSize = dictSize; |
| 943 | best->parameters = parameters; |
| 944 | best->compressedSize = compressedSize; |
| 945 | } |
| 946 | } |
| 947 | if (liveJobs == 0) { |
| 948 | ZSTD_pthread_cond_broadcast(&best->cond); |
| 949 | } |
| 950 | ZSTD_pthread_mutex_unlock(&best->mutex); |
| 951 | } |
| 952 | } |
| 953 | |
| 954 | static COVER_dictSelection_t setDictSelection(BYTE* buf, size_t s, size_t csz) |
| 955 | { |
| 956 | COVER_dictSelection_t ds; |
| 957 | ds.dictContent = buf; |
| 958 | ds.dictSize = s; |
| 959 | ds.totalCompressedSize = csz; |
| 960 | return ds; |
| 961 | } |
| 962 | |
| 963 | COVER_dictSelection_t COVER_dictSelectionError(size_t error) { |
| 964 | return setDictSelection(NULL, 0, error); |
| 965 | } |
| 966 | |
| 967 | unsigned COVER_dictSelectionIsError(COVER_dictSelection_t selection) { |
| 968 | return (ZSTD_isError(selection.totalCompressedSize) || !selection.dictContent); |
| 969 | } |
| 970 | |
| 971 | void COVER_dictSelectionFree(COVER_dictSelection_t selection){ |
| 972 | free(selection.dictContent); |
| 973 | } |
| 974 | |
| 975 | COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent, size_t dictBufferCapacity, |
| 976 | size_t dictContentSize, const BYTE* samplesBuffer, const size_t* samplesSizes, unsigned nbFinalizeSamples, |
| 977 | size_t nbCheckSamples, size_t nbSamples, ZDICT_cover_params_t params, size_t* offsets, size_t totalCompressedSize) { |
| 978 | |
| 979 | size_t largestDict = 0; |
| 980 | size_t largestCompressed = 0; |
| 981 | BYTE* customDictContentEnd = customDictContent + dictContentSize; |
| 982 | |
| 983 | BYTE * largestDictbuffer = (BYTE *)malloc(dictBufferCapacity); |
| 984 | BYTE * candidateDictBuffer = (BYTE *)malloc(dictBufferCapacity); |
| 985 | double regressionTolerance = ((double)params.shrinkDictMaxRegression / 100.0) + 1.00; |
| 986 | |
| 987 | if (!largestDictbuffer || !candidateDictBuffer) { |
| 988 | free(largestDictbuffer); |
| 989 | free(candidateDictBuffer); |
| 990 | return COVER_dictSelectionError(dictContentSize); |
| 991 | } |
| 992 | |
| 993 | /* Initial dictionary size and compressed size */ |
| 994 | memcpy(largestDictbuffer, customDictContent, dictContentSize); |
| 995 | dictContentSize = ZDICT_finalizeDictionary( |
| 996 | largestDictbuffer, dictBufferCapacity, customDictContent, dictContentSize, |
| 997 | samplesBuffer, samplesSizes, nbFinalizeSamples, params.zParams); |
| 998 | |
| 999 | if (ZDICT_isError(dictContentSize)) { |
| 1000 | free(largestDictbuffer); |
| 1001 | free(candidateDictBuffer); |
| 1002 | return COVER_dictSelectionError(dictContentSize); |
| 1003 | } |
| 1004 | |
| 1005 | totalCompressedSize = COVER_checkTotalCompressedSize(params, samplesSizes, |
| 1006 | samplesBuffer, offsets, |
| 1007 | nbCheckSamples, nbSamples, |
| 1008 | largestDictbuffer, dictContentSize); |
| 1009 | |
| 1010 | if (ZSTD_isError(totalCompressedSize)) { |
| 1011 | free(largestDictbuffer); |
| 1012 | free(candidateDictBuffer); |
| 1013 | return COVER_dictSelectionError(totalCompressedSize); |
| 1014 | } |
| 1015 | |
| 1016 | if (params.shrinkDict == 0) { |
| 1017 | free(candidateDictBuffer); |
| 1018 | return setDictSelection(largestDictbuffer, dictContentSize, totalCompressedSize); |
| 1019 | } |
| 1020 | |
| 1021 | largestDict = dictContentSize; |
| 1022 | largestCompressed = totalCompressedSize; |
| 1023 | dictContentSize = ZDICT_DICTSIZE_MIN; |
| 1024 | |
| 1025 | /* Largest dict is initially at least ZDICT_DICTSIZE_MIN */ |
| 1026 | while (dictContentSize < largestDict) { |
| 1027 | memcpy(candidateDictBuffer, largestDictbuffer, largestDict); |
| 1028 | dictContentSize = ZDICT_finalizeDictionary( |
| 1029 | candidateDictBuffer, dictBufferCapacity, customDictContentEnd - dictContentSize, dictContentSize, |
| 1030 | samplesBuffer, samplesSizes, nbFinalizeSamples, params.zParams); |
| 1031 | |
| 1032 | if (ZDICT_isError(dictContentSize)) { |
| 1033 | free(largestDictbuffer); |
| 1034 | free(candidateDictBuffer); |
| 1035 | return COVER_dictSelectionError(dictContentSize); |
| 1036 | |
| 1037 | } |
| 1038 | |
| 1039 | totalCompressedSize = COVER_checkTotalCompressedSize(params, samplesSizes, |
| 1040 | samplesBuffer, offsets, |
| 1041 | nbCheckSamples, nbSamples, |
| 1042 | candidateDictBuffer, dictContentSize); |
| 1043 | |
| 1044 | if (ZSTD_isError(totalCompressedSize)) { |
| 1045 | free(largestDictbuffer); |
| 1046 | free(candidateDictBuffer); |
| 1047 | return COVER_dictSelectionError(totalCompressedSize); |
| 1048 | } |
| 1049 | |
| 1050 | if ((double)totalCompressedSize <= (double)largestCompressed * regressionTolerance) { |
| 1051 | free(largestDictbuffer); |
| 1052 | return setDictSelection( candidateDictBuffer, dictContentSize, totalCompressedSize ); |
| 1053 | } |
| 1054 | dictContentSize *= 2; |
| 1055 | } |
| 1056 | dictContentSize = largestDict; |
| 1057 | totalCompressedSize = largestCompressed; |
| 1058 | free(candidateDictBuffer); |
| 1059 | return setDictSelection( largestDictbuffer, dictContentSize, totalCompressedSize ); |
| 1060 | } |
| 1061 | |
| 1062 | /** |
| 1063 | * Parameters for COVER_tryParameters(). |
| 1064 | */ |
| 1065 | typedef struct COVER_tryParameters_data_s { |
| 1066 | const COVER_ctx_t *ctx; |
| 1067 | COVER_best_t *best; |
| 1068 | size_t dictBufferCapacity; |
| 1069 | ZDICT_cover_params_t parameters; |
| 1070 | } COVER_tryParameters_data_t; |
| 1071 | |
| 1072 | /** |
| 1073 | * Tries a set of parameters and updates the COVER_best_t with the results. |
| 1074 | * This function is thread safe if zstd is compiled with multithreaded support. |
| 1075 | * It takes its parameters as an *OWNING* opaque pointer to support threading. |
| 1076 | */ |
| 1077 | static void COVER_tryParameters(void *opaque) |
| 1078 | { |
| 1079 | /* Save parameters as local variables */ |
| 1080 | COVER_tryParameters_data_t *const data = (COVER_tryParameters_data_t*)opaque; |
| 1081 | const COVER_ctx_t *const ctx = data->ctx; |
| 1082 | const ZDICT_cover_params_t parameters = data->parameters; |
| 1083 | size_t dictBufferCapacity = data->dictBufferCapacity; |
| 1084 | size_t totalCompressedSize = ERROR(GENERIC); |
| 1085 | /* Allocate space for hash table, dict, and freqs */ |
| 1086 | COVER_map_t activeDmers; |
| 1087 | BYTE* const dict = (BYTE*)malloc(dictBufferCapacity); |
| 1088 | COVER_dictSelection_t selection = COVER_dictSelectionError(ERROR(GENERIC)); |
| 1089 | U32* const freqs = (U32*)malloc(ctx->suffixSize * sizeof(U32)); |
| 1090 | if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) { |
| 1091 | DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n"); |
| 1092 | goto _cleanup; |
| 1093 | } |
| 1094 | if (!dict || !freqs) { |
| 1095 | DISPLAYLEVEL(1, "Failed to allocate buffers: out of memory\n"); |
| 1096 | goto _cleanup; |
| 1097 | } |
| 1098 | /* Copy the frequencies because we need to modify them */ |
| 1099 | memcpy(freqs, ctx->freqs, ctx->suffixSize * sizeof(U32)); |
| 1100 | /* Build the dictionary */ |
| 1101 | { |
| 1102 | const size_t tail = COVER_buildDictionary(ctx, freqs, &activeDmers, dict, |
| 1103 | dictBufferCapacity, parameters); |
| 1104 | selection = COVER_selectDict(dict + tail, dictBufferCapacity, dictBufferCapacity - tail, |
| 1105 | ctx->samples, ctx->samplesSizes, (unsigned)ctx->nbTrainSamples, ctx->nbTrainSamples, ctx->nbSamples, parameters, ctx->offsets, |
| 1106 | totalCompressedSize); |
| 1107 | |
| 1108 | if (COVER_dictSelectionIsError(selection)) { |
| 1109 | DISPLAYLEVEL(1, "Failed to select dictionary\n"); |
| 1110 | goto _cleanup; |
| 1111 | } |
| 1112 | } |
| 1113 | _cleanup: |
| 1114 | free(dict); |
| 1115 | COVER_best_finish(data->best, parameters, selection); |
| 1116 | free(data); |
| 1117 | COVER_map_destroy(&activeDmers); |
| 1118 | COVER_dictSelectionFree(selection); |
| 1119 | free(freqs); |
| 1120 | } |
| 1121 | |
| 1122 | ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover( |
| 1123 | void* dictBuffer, size_t dictBufferCapacity, const void* samplesBuffer, |
| 1124 | const size_t* samplesSizes, unsigned nbSamples, |
| 1125 | ZDICT_cover_params_t* parameters) |
| 1126 | { |
| 1127 | /* constants */ |
| 1128 | const unsigned nbThreads = parameters->nbThreads; |
| 1129 | const double splitPoint = |
| 1130 | parameters->splitPoint <= 0.0 ? COVER_DEFAULT_SPLITPOINT : parameters->splitPoint; |
| 1131 | const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d; |
| 1132 | const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d; |
| 1133 | const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k; |
| 1134 | const unsigned kMaxK = parameters->k == 0 ? 2000 : parameters->k; |
| 1135 | const unsigned kSteps = parameters->steps == 0 ? 40 : parameters->steps; |
| 1136 | const unsigned kStepSize = MAX((kMaxK - kMinK) / kSteps, 1); |
| 1137 | const unsigned kIterations = |
| 1138 | (1 + (kMaxD - kMinD) / 2) * (1 + (kMaxK - kMinK) / kStepSize); |
| 1139 | const unsigned shrinkDict = 0; |
| 1140 | /* Local variables */ |
| 1141 | const int displayLevel = parameters->zParams.notificationLevel; |
| 1142 | unsigned iteration = 1; |
| 1143 | unsigned d; |
| 1144 | unsigned k; |
| 1145 | COVER_best_t best; |
| 1146 | POOL_ctx *pool = NULL; |
| 1147 | int warned = 0; |
| 1148 | |
| 1149 | /* Checks */ |
| 1150 | if (splitPoint <= 0 || splitPoint > 1) { |
| 1151 | LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect parameters\n"); |
| 1152 | return ERROR(parameter_outOfBound); |
| 1153 | } |
| 1154 | if (kMinK < kMaxD || kMaxK < kMinK) { |
| 1155 | LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect parameters\n"); |
| 1156 | return ERROR(parameter_outOfBound); |
| 1157 | } |
| 1158 | if (nbSamples == 0) { |
| 1159 | DISPLAYLEVEL(1, "Cover must have at least one input file\n"); |
| 1160 | return ERROR(srcSize_wrong); |
| 1161 | } |
| 1162 | if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) { |
| 1163 | DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n", |
| 1164 | ZDICT_DICTSIZE_MIN); |
| 1165 | return ERROR(dstSize_tooSmall); |
| 1166 | } |
| 1167 | if (nbThreads > 1) { |
| 1168 | pool = POOL_create(nbThreads, 1); |
| 1169 | if (!pool) { |
| 1170 | return ERROR(memory_allocation); |
| 1171 | } |
| 1172 | } |
| 1173 | /* Initialization */ |
| 1174 | COVER_best_init(&best); |
| 1175 | /* Turn down global display level to clean up display at level 2 and below */ |
| 1176 | g_displayLevel = displayLevel == 0 ? 0 : displayLevel - 1; |
| 1177 | /* Loop through d first because each new value needs a new context */ |
| 1178 | LOCALDISPLAYLEVEL(displayLevel, 2, "Trying %u different sets of parameters\n", |
| 1179 | kIterations); |
| 1180 | for (d = kMinD; d <= kMaxD; d += 2) { |
| 1181 | /* Initialize the context for this value of d */ |
| 1182 | COVER_ctx_t ctx; |
| 1183 | LOCALDISPLAYLEVEL(displayLevel, 3, "d=%u\n", d); |
| 1184 | { |
| 1185 | const size_t initVal = COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, d, splitPoint); |
| 1186 | if (ZSTD_isError(initVal)) { |
| 1187 | LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to initialize context\n"); |
| 1188 | COVER_best_destroy(&best); |
| 1189 | POOL_free(pool); |
| 1190 | return initVal; |
| 1191 | } |
| 1192 | } |
| 1193 | if (!warned) { |
| 1194 | COVER_warnOnSmallCorpus(dictBufferCapacity, ctx.suffixSize, displayLevel); |
| 1195 | warned = 1; |
| 1196 | } |
| 1197 | /* Loop through k reusing the same context */ |
| 1198 | for (k = kMinK; k <= kMaxK; k += kStepSize) { |
| 1199 | /* Prepare the arguments */ |
| 1200 | COVER_tryParameters_data_t *data = (COVER_tryParameters_data_t *)malloc( |
| 1201 | sizeof(COVER_tryParameters_data_t)); |
| 1202 | LOCALDISPLAYLEVEL(displayLevel, 3, "k=%u\n", k); |
| 1203 | if (!data) { |
| 1204 | LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to allocate parameters\n"); |
| 1205 | COVER_best_destroy(&best); |
| 1206 | COVER_ctx_destroy(&ctx); |
| 1207 | POOL_free(pool); |
| 1208 | return ERROR(memory_allocation); |
| 1209 | } |
| 1210 | data->ctx = &ctx; |
| 1211 | data->best = &best; |
| 1212 | data->dictBufferCapacity = dictBufferCapacity; |
| 1213 | data->parameters = *parameters; |
| 1214 | data->parameters.k = k; |
| 1215 | data->parameters.d = d; |
| 1216 | data->parameters.splitPoint = splitPoint; |
| 1217 | data->parameters.steps = kSteps; |
| 1218 | data->parameters.shrinkDict = shrinkDict; |
| 1219 | data->parameters.zParams.notificationLevel = g_displayLevel; |
| 1220 | /* Check the parameters */ |
| 1221 | if (!COVER_checkParameters(data->parameters, dictBufferCapacity)) { |
| 1222 | DISPLAYLEVEL(1, "Cover parameters incorrect\n"); |
| 1223 | free(data); |
| 1224 | continue; |
| 1225 | } |
| 1226 | /* Call the function and pass ownership of data to it */ |
| 1227 | COVER_best_start(&best); |
| 1228 | if (pool) { |
| 1229 | POOL_add(pool, &COVER_tryParameters, data); |
| 1230 | } else { |
| 1231 | COVER_tryParameters(data); |
| 1232 | } |
| 1233 | /* Print status */ |
| 1234 | LOCALDISPLAYUPDATE(displayLevel, 2, "\r%u%% ", |
| 1235 | (unsigned)((iteration * 100) / kIterations)); |
| 1236 | ++iteration; |
| 1237 | } |
| 1238 | COVER_best_wait(&best); |
| 1239 | COVER_ctx_destroy(&ctx); |
| 1240 | } |
| 1241 | LOCALDISPLAYLEVEL(displayLevel, 2, "\r%79s\r", ""); |
| 1242 | /* Fill the output buffer and parameters with output of the best parameters */ |
| 1243 | { |
| 1244 | const size_t dictSize = best.dictSize; |
| 1245 | if (ZSTD_isError(best.compressedSize)) { |
| 1246 | const size_t compressedSize = best.compressedSize; |
| 1247 | COVER_best_destroy(&best); |
| 1248 | POOL_free(pool); |
| 1249 | return compressedSize; |
| 1250 | } |
| 1251 | *parameters = best.parameters; |
| 1252 | memcpy(dictBuffer, best.dict, dictSize); |
| 1253 | COVER_best_destroy(&best); |
| 1254 | POOL_free(pool); |
| 1255 | return dictSize; |
| 1256 | } |
| 1257 | } |