| 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 | |
| 13 | /* ************************************** |
| 14 | * Compiler Warnings |
| 15 | ****************************************/ |
| 16 | #ifdef _MSC_VER |
| 17 | # pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */ |
| 18 | #endif |
| 19 | |
| 20 | |
| 21 | /*-************************************* |
| 22 | * Includes |
| 23 | ***************************************/ |
| 24 | #include "platform.h" /* Large Files support */ |
| 25 | #include "util.h" /* UTIL_getFileSize, UTIL_getTotalFileSize */ |
| 26 | #include <stdlib.h> /* malloc, free */ |
| 27 | #include <string.h> /* memset */ |
| 28 | #include <stdio.h> /* fprintf, fopen, ftello64 */ |
| 29 | #include <errno.h> /* errno */ |
| 30 | |
| 31 | #include "timefn.h" /* UTIL_time_t, UTIL_clockSpanMicro, UTIL_getTime */ |
| 32 | #include "../lib/common/debug.h" /* assert */ |
| 33 | #include "../lib/common/mem.h" /* read */ |
| 34 | #include "../lib/zstd_errors.h" |
| 35 | #include "dibio.h" |
| 36 | |
| 37 | |
| 38 | /*-************************************* |
| 39 | * Constants |
| 40 | ***************************************/ |
| 41 | #define KB *(1 <<10) |
| 42 | #define MB *(1 <<20) |
| 43 | #define GB *(1U<<30) |
| 44 | |
| 45 | #define SAMPLESIZE_MAX (128 KB) |
| 46 | #define MEMMULT 11 /* rough estimation : memory cost to analyze 1 byte of sample */ |
| 47 | #define COVER_MEMMULT 9 /* rough estimation : memory cost to analyze 1 byte of sample */ |
| 48 | #define FASTCOVER_MEMMULT 1 /* rough estimation : memory cost to analyze 1 byte of sample */ |
| 49 | static const size_t g_maxMemory = (sizeof(size_t) == 4) ? (2 GB - 64 MB) : ((size_t)(512 MB) << sizeof(size_t)); |
| 50 | |
| 51 | #define NOISELENGTH 32 |
| 52 | #define MAX_SAMPLES_SIZE (2 GB) /* training dataset limited to 2GB */ |
| 53 | |
| 54 | |
| 55 | /*-************************************* |
| 56 | * Console display |
| 57 | ***************************************/ |
| 58 | #define DISPLAY(...) fprintf(stderr, __VA_ARGS__) |
| 59 | #define DISPLAYLEVEL(l, ...) if (displayLevel>=l) { DISPLAY(__VA_ARGS__); } |
| 60 | |
| 61 | static const U64 g_refreshRate = SEC_TO_MICRO / 6; |
| 62 | static UTIL_time_t g_displayClock = UTIL_TIME_INITIALIZER; |
| 63 | |
| 64 | #define DISPLAYUPDATE(l, ...) { if (displayLevel>=l) { \ |
| 65 | if ((UTIL_clockSpanMicro(g_displayClock) > g_refreshRate) || (displayLevel>=4)) \ |
| 66 | { g_displayClock = UTIL_getTime(); DISPLAY(__VA_ARGS__); \ |
| 67 | if (displayLevel>=4) fflush(stderr); } } } |
| 68 | |
| 69 | /*-************************************* |
| 70 | * Exceptions |
| 71 | ***************************************/ |
| 72 | #ifndef DEBUG |
| 73 | # define DEBUG 0 |
| 74 | #endif |
| 75 | #define DEBUGOUTPUT(...) if (DEBUG) DISPLAY(__VA_ARGS__); |
| 76 | #define EXM_THROW(error, ...) \ |
| 77 | { \ |
| 78 | DEBUGOUTPUT("Error defined at %s, line %i : \n", __FILE__, __LINE__); \ |
| 79 | DISPLAY("Error %i : ", error); \ |
| 80 | DISPLAY(__VA_ARGS__); \ |
| 81 | DISPLAY("\n"); \ |
| 82 | exit(error); \ |
| 83 | } |
| 84 | |
| 85 | |
| 86 | /* ******************************************************** |
| 87 | * Helper functions |
| 88 | **********************************************************/ |
| 89 | #undef MIN |
| 90 | #define MIN(a,b) ((a) < (b) ? (a) : (b)) |
| 91 | |
| 92 | /** |
| 93 | Returns the size of a file. |
| 94 | If error returns -1. |
| 95 | */ |
| 96 | static S64 DiB_getFileSize (const char * fileName) |
| 97 | { |
| 98 | U64 const fileSize = UTIL_getFileSize(fileName); |
| 99 | return (fileSize == UTIL_FILESIZE_UNKNOWN) ? -1 : (S64)fileSize; |
| 100 | } |
| 101 | |
| 102 | /* ******************************************************** |
| 103 | * File related operations |
| 104 | **********************************************************/ |
| 105 | /** DiB_loadFiles() : |
| 106 | * load samples from files listed in fileNamesTable into buffer. |
| 107 | * works even if buffer is too small to load all samples. |
| 108 | * Also provides the size of each sample into sampleSizes table |
| 109 | * which must be sized correctly, using DiB_fileStats(). |
| 110 | * @return : nb of samples effectively loaded into `buffer` |
| 111 | * *bufferSizePtr is modified, it provides the amount data loaded within buffer. |
| 112 | * sampleSizes is filled with the size of each sample. |
| 113 | */ |
| 114 | static int DiB_loadFiles( |
| 115 | void* buffer, size_t* bufferSizePtr, |
| 116 | size_t* sampleSizes, int sstSize, |
| 117 | const char** fileNamesTable, int nbFiles, |
| 118 | size_t targetChunkSize, int displayLevel ) |
| 119 | { |
| 120 | char* const buff = (char*)buffer; |
| 121 | size_t totalDataLoaded = 0; |
| 122 | int nbSamplesLoaded = 0; |
| 123 | int fileIndex = 0; |
| 124 | FILE * f = NULL; |
| 125 | |
| 126 | assert(targetChunkSize <= SAMPLESIZE_MAX); |
| 127 | |
| 128 | while ( nbSamplesLoaded < sstSize && fileIndex < nbFiles ) { |
| 129 | size_t fileDataLoaded; |
| 130 | S64 const fileSize = DiB_getFileSize(fileNamesTable[fileIndex]); |
| 131 | if (fileSize <= 0) { |
| 132 | /* skip if zero-size or file error */ |
| 133 | ++fileIndex; |
| 134 | continue; |
| 135 | } |
| 136 | |
| 137 | f = fopen( fileNamesTable[fileIndex], "rb"); |
| 138 | if (f == NULL) |
| 139 | EXM_THROW(10, "zstd: dictBuilder: %s %s ", fileNamesTable[fileIndex], strerror(errno)); |
| 140 | DISPLAYUPDATE(2, "Loading %s... \r", fileNamesTable[fileIndex]); |
| 141 | |
| 142 | /* Load the first chunk of data from the file */ |
| 143 | fileDataLoaded = targetChunkSize > 0 ? |
| 144 | (size_t)MIN(fileSize, (S64)targetChunkSize) : |
| 145 | (size_t)MIN(fileSize, SAMPLESIZE_MAX ); |
| 146 | if (totalDataLoaded + fileDataLoaded > *bufferSizePtr) |
| 147 | break; |
| 148 | if (fread( buff+totalDataLoaded, 1, fileDataLoaded, f ) != fileDataLoaded) |
| 149 | EXM_THROW(11, "Pb reading %s", fileNamesTable[fileIndex]); |
| 150 | sampleSizes[nbSamplesLoaded++] = fileDataLoaded; |
| 151 | totalDataLoaded += fileDataLoaded; |
| 152 | |
| 153 | /* If file-chunking is enabled, load the rest of the file as more samples */ |
| 154 | if (targetChunkSize > 0) { |
| 155 | while( (S64)fileDataLoaded < fileSize && nbSamplesLoaded < sstSize ) { |
| 156 | size_t const chunkSize = MIN((size_t)(fileSize-fileDataLoaded), targetChunkSize); |
| 157 | if (totalDataLoaded + chunkSize > *bufferSizePtr) /* buffer is full */ |
| 158 | break; |
| 159 | |
| 160 | if (fread( buff+totalDataLoaded, 1, chunkSize, f ) != chunkSize) |
| 161 | EXM_THROW(11, "Pb reading %s", fileNamesTable[fileIndex]); |
| 162 | sampleSizes[nbSamplesLoaded++] = chunkSize; |
| 163 | totalDataLoaded += chunkSize; |
| 164 | fileDataLoaded += chunkSize; |
| 165 | } |
| 166 | } |
| 167 | fileIndex += 1; |
| 168 | fclose(f); f = NULL; |
| 169 | } |
| 170 | if (f != NULL) |
| 171 | fclose(f); |
| 172 | |
| 173 | DISPLAYLEVEL(2, "\r%79s\r", ""); |
| 174 | DISPLAYLEVEL(4, "Loaded %d KB total training data, %d nb samples \n", |
| 175 | (int)(totalDataLoaded / (1 KB)), nbSamplesLoaded ); |
| 176 | *bufferSizePtr = totalDataLoaded; |
| 177 | return nbSamplesLoaded; |
| 178 | } |
| 179 | |
| 180 | #define DiB_rotl32(x,r) ((x << r) | (x >> (32 - r))) |
| 181 | static U32 DiB_rand(U32* src) |
| 182 | { |
| 183 | static const U32 prime1 = 2654435761U; |
| 184 | static const U32 prime2 = 2246822519U; |
| 185 | U32 rand32 = *src; |
| 186 | rand32 *= prime1; |
| 187 | rand32 ^= prime2; |
| 188 | rand32 = DiB_rotl32(rand32, 13); |
| 189 | *src = rand32; |
| 190 | return rand32 >> 5; |
| 191 | } |
| 192 | |
| 193 | /* DiB_shuffle() : |
| 194 | * shuffle a table of file names in a semi-random way |
| 195 | * It improves dictionary quality by reducing "locality" impact, so if sample set is very large, |
| 196 | * it will load random elements from it, instead of just the first ones. */ |
| 197 | static void DiB_shuffle(const char** fileNamesTable, unsigned nbFiles) { |
| 198 | U32 seed = 0xFD2FB528; |
| 199 | unsigned i; |
| 200 | if (nbFiles == 0) |
| 201 | return; |
| 202 | for (i = nbFiles - 1; i > 0; --i) { |
| 203 | unsigned const j = DiB_rand(&seed) % (i + 1); |
| 204 | const char* const tmp = fileNamesTable[j]; |
| 205 | fileNamesTable[j] = fileNamesTable[i]; |
| 206 | fileNamesTable[i] = tmp; |
| 207 | } |
| 208 | } |
| 209 | |
| 210 | |
| 211 | /*-******************************************************** |
| 212 | * Dictionary training functions |
| 213 | **********************************************************/ |
| 214 | static size_t DiB_findMaxMem(unsigned long long requiredMem) |
| 215 | { |
| 216 | size_t const step = 8 MB; |
| 217 | void* testmem = NULL; |
| 218 | |
| 219 | requiredMem = (((requiredMem >> 23) + 1) << 23); |
| 220 | requiredMem += step; |
| 221 | if (requiredMem > g_maxMemory) requiredMem = g_maxMemory; |
| 222 | |
| 223 | while (!testmem) { |
| 224 | testmem = malloc((size_t)requiredMem); |
| 225 | requiredMem -= step; |
| 226 | } |
| 227 | |
| 228 | free(testmem); |
| 229 | return (size_t)requiredMem; |
| 230 | } |
| 231 | |
| 232 | |
| 233 | static void DiB_fillNoise(void* buffer, size_t length) |
| 234 | { |
| 235 | unsigned const prime1 = 2654435761U; |
| 236 | unsigned const prime2 = 2246822519U; |
| 237 | unsigned acc = prime1; |
| 238 | size_t p=0; |
| 239 | |
| 240 | for (p=0; p<length; p++) { |
| 241 | acc *= prime2; |
| 242 | ((unsigned char*)buffer)[p] = (unsigned char)(acc >> 21); |
| 243 | } |
| 244 | } |
| 245 | |
| 246 | |
| 247 | static void DiB_saveDict(const char* dictFileName, |
| 248 | const void* buff, size_t buffSize) |
| 249 | { |
| 250 | FILE* const f = fopen(dictFileName, "wb"); |
| 251 | if (f==NULL) EXM_THROW(3, "cannot open %s ", dictFileName); |
| 252 | |
| 253 | { size_t const n = fwrite(buff, 1, buffSize, f); |
| 254 | if (n!=buffSize) EXM_THROW(4, "%s : write error", dictFileName) } |
| 255 | |
| 256 | { size_t const n = (size_t)fclose(f); |
| 257 | if (n!=0) EXM_THROW(5, "%s : flush error", dictFileName) } |
| 258 | } |
| 259 | |
| 260 | typedef struct { |
| 261 | S64 totalSizeToLoad; |
| 262 | int nbSamples; |
| 263 | int oneSampleTooLarge; |
| 264 | } fileStats; |
| 265 | |
| 266 | /*! DiB_fileStats() : |
| 267 | * Given a list of files, and a chunkSize (0 == no chunk, whole files) |
| 268 | * provides the amount of data to be loaded and the resulting nb of samples. |
| 269 | * This is useful primarily for allocation purpose => sample buffer, and sample sizes table. |
| 270 | */ |
| 271 | static fileStats DiB_fileStats(const char** fileNamesTable, int nbFiles, size_t chunkSize, int displayLevel) |
| 272 | { |
| 273 | fileStats fs; |
| 274 | int n; |
| 275 | memset(&fs, 0, sizeof(fs)); |
| 276 | |
| 277 | /* We assume that if chunking is requested, the chunk size is < SAMPLESIZE_MAX */ |
| 278 | assert( chunkSize <= SAMPLESIZE_MAX ); |
| 279 | |
| 280 | for (n=0; n<nbFiles; n++) { |
| 281 | S64 const fileSize = DiB_getFileSize(fileNamesTable[n]); |
| 282 | /* TODO: is there a minimum sample size? What if the file is 1-byte? */ |
| 283 | if (fileSize == 0) { |
| 284 | DISPLAYLEVEL(3, "Sample file '%s' has zero size, skipping...\n", fileNamesTable[n]); |
| 285 | continue; |
| 286 | } |
| 287 | |
| 288 | /* the case where we are breaking up files in sample chunks */ |
| 289 | if (chunkSize > 0) { |
| 290 | /* TODO: is there a minimum sample size? Can we have a 1-byte sample? */ |
| 291 | fs.nbSamples += (int)((fileSize + chunkSize-1) / chunkSize); |
| 292 | fs.totalSizeToLoad += fileSize; |
| 293 | } |
| 294 | else { |
| 295 | /* the case where one file is one sample */ |
| 296 | if (fileSize > SAMPLESIZE_MAX) { |
| 297 | /* flag excessively large sample files */ |
| 298 | fs.oneSampleTooLarge |= (fileSize > 2*SAMPLESIZE_MAX); |
| 299 | |
| 300 | /* Limit to the first SAMPLESIZE_MAX (128kB) of the file */ |
| 301 | DISPLAYLEVEL(3, "Sample file '%s' is too large, limiting to %d KB", |
| 302 | fileNamesTable[n], SAMPLESIZE_MAX / (1 KB)); |
| 303 | } |
| 304 | fs.nbSamples += 1; |
| 305 | fs.totalSizeToLoad += MIN(fileSize, SAMPLESIZE_MAX); |
| 306 | } |
| 307 | } |
| 308 | DISPLAYLEVEL(4, "Found training data %d files, %d KB, %d samples\n", nbFiles, (int)(fs.totalSizeToLoad / (1 KB)), fs.nbSamples); |
| 309 | return fs; |
| 310 | } |
| 311 | |
| 312 | int DiB_trainFromFiles(const char* dictFileName, size_t maxDictSize, |
| 313 | const char** fileNamesTable, int nbFiles, size_t chunkSize, |
| 314 | ZDICT_legacy_params_t* params, ZDICT_cover_params_t* coverParams, |
| 315 | ZDICT_fastCover_params_t* fastCoverParams, int optimize, unsigned memLimit) |
| 316 | { |
| 317 | fileStats fs; |
| 318 | size_t* sampleSizes; /* vector of sample sizes. Each sample can be up to SAMPLESIZE_MAX */ |
| 319 | int nbSamplesLoaded; /* nb of samples effectively loaded in srcBuffer */ |
| 320 | size_t loadedSize; /* total data loaded in srcBuffer for all samples */ |
| 321 | void* srcBuffer /* contiguous buffer with training data/samples */; |
| 322 | void* const dictBuffer = malloc(maxDictSize); |
| 323 | int result = 0; |
| 324 | |
| 325 | int const displayLevel = params ? params->zParams.notificationLevel : |
| 326 | coverParams ? coverParams->zParams.notificationLevel : |
| 327 | fastCoverParams ? fastCoverParams->zParams.notificationLevel : 0; |
| 328 | |
| 329 | /* Shuffle input files before we start assessing how much sample datA to load. |
| 330 | The purpose of the shuffle is to pick random samples when the sample |
| 331 | set is larger than what we can load in memory. */ |
| 332 | DISPLAYLEVEL(3, "Shuffling input files\n"); |
| 333 | DiB_shuffle(fileNamesTable, nbFiles); |
| 334 | |
| 335 | /* Figure out how much sample data to load with how many samples */ |
| 336 | fs = DiB_fileStats(fileNamesTable, nbFiles, chunkSize, displayLevel); |
| 337 | |
| 338 | { |
| 339 | int const memMult = params ? MEMMULT : |
| 340 | coverParams ? COVER_MEMMULT: |
| 341 | FASTCOVER_MEMMULT; |
| 342 | size_t const maxMem = DiB_findMaxMem(fs.totalSizeToLoad * memMult) / memMult; |
| 343 | /* Limit the size of the training data to the free memory */ |
| 344 | /* Limit the size of the training data to 2GB */ |
| 345 | /* TODO: there is opportunity to stop DiB_fileStats() early when the data limit is reached */ |
| 346 | loadedSize = (size_t)MIN( MIN((S64)maxMem, fs.totalSizeToLoad), MAX_SAMPLES_SIZE ); |
| 347 | if (memLimit != 0) { |
| 348 | DISPLAYLEVEL(2, "! Warning : setting manual memory limit for dictionary training data at %u MB \n", |
| 349 | (unsigned)(memLimit / (1 MB))); |
| 350 | loadedSize = (size_t)MIN(loadedSize, memLimit); |
| 351 | } |
| 352 | srcBuffer = malloc(loadedSize+NOISELENGTH); |
| 353 | sampleSizes = (size_t*)malloc(fs.nbSamples * sizeof(size_t)); |
| 354 | } |
| 355 | |
| 356 | /* Checks */ |
| 357 | if ((fs.nbSamples && !sampleSizes) || (!srcBuffer) || (!dictBuffer)) |
| 358 | EXM_THROW(12, "not enough memory for DiB_trainFiles"); /* should not happen */ |
| 359 | if (fs.oneSampleTooLarge) { |
| 360 | DISPLAYLEVEL(2, "! Warning : some sample(s) are very large \n"); |
| 361 | DISPLAYLEVEL(2, "! Note that dictionary is only useful for small samples. \n"); |
| 362 | DISPLAYLEVEL(2, "! As a consequence, only the first %u bytes of each sample are loaded \n", SAMPLESIZE_MAX); |
| 363 | } |
| 364 | if (fs.nbSamples < 5) { |
| 365 | DISPLAYLEVEL(2, "! Warning : nb of samples too low for proper processing ! \n"); |
| 366 | DISPLAYLEVEL(2, "! Please provide _one file per sample_. \n"); |
| 367 | DISPLAYLEVEL(2, "! Alternatively, split files into fixed-size blocks representative of samples, with -B# \n"); |
| 368 | EXM_THROW(14, "nb of samples too low"); /* we now clearly forbid this case */ |
| 369 | } |
| 370 | if (fs.totalSizeToLoad < (S64)maxDictSize * 8) { |
| 371 | DISPLAYLEVEL(2, "! Warning : data size of samples too small for target dictionary size \n"); |
| 372 | DISPLAYLEVEL(2, "! Samples should be about 100x larger than target dictionary size \n"); |
| 373 | } |
| 374 | |
| 375 | /* init */ |
| 376 | if ((S64)loadedSize < fs.totalSizeToLoad) |
| 377 | DISPLAYLEVEL(1, "Training samples set too large (%u MB); training on %u MB only...\n", |
| 378 | (unsigned)(fs.totalSizeToLoad / (1 MB)), |
| 379 | (unsigned)(loadedSize / (1 MB))); |
| 380 | |
| 381 | /* Load input buffer */ |
| 382 | nbSamplesLoaded = DiB_loadFiles( |
| 383 | srcBuffer, &loadedSize, sampleSizes, fs.nbSamples, fileNamesTable, |
| 384 | nbFiles, chunkSize, displayLevel); |
| 385 | |
| 386 | { size_t dictSize = ZSTD_error_GENERIC; |
| 387 | if (params) { |
| 388 | DiB_fillNoise((char*)srcBuffer + loadedSize, NOISELENGTH); /* guard band, for end of buffer condition */ |
| 389 | dictSize = ZDICT_trainFromBuffer_legacy(dictBuffer, maxDictSize, |
| 390 | srcBuffer, sampleSizes, nbSamplesLoaded, |
| 391 | *params); |
| 392 | } else if (coverParams) { |
| 393 | if (optimize) { |
| 394 | dictSize = ZDICT_optimizeTrainFromBuffer_cover(dictBuffer, maxDictSize, |
| 395 | srcBuffer, sampleSizes, nbSamplesLoaded, |
| 396 | coverParams); |
| 397 | if (!ZDICT_isError(dictSize)) { |
| 398 | unsigned splitPercentage = (unsigned)(coverParams->splitPoint * 100); |
| 399 | DISPLAYLEVEL(2, "k=%u\nd=%u\nsteps=%u\nsplit=%u\n", coverParams->k, coverParams->d, |
| 400 | coverParams->steps, splitPercentage); |
| 401 | } |
| 402 | } else { |
| 403 | dictSize = ZDICT_trainFromBuffer_cover(dictBuffer, maxDictSize, srcBuffer, |
| 404 | sampleSizes, nbSamplesLoaded, *coverParams); |
| 405 | } |
| 406 | } else if (fastCoverParams != NULL) { |
| 407 | if (optimize) { |
| 408 | dictSize = ZDICT_optimizeTrainFromBuffer_fastCover(dictBuffer, maxDictSize, |
| 409 | srcBuffer, sampleSizes, nbSamplesLoaded, |
| 410 | fastCoverParams); |
| 411 | if (!ZDICT_isError(dictSize)) { |
| 412 | unsigned splitPercentage = (unsigned)(fastCoverParams->splitPoint * 100); |
| 413 | DISPLAYLEVEL(2, "k=%u\nd=%u\nf=%u\nsteps=%u\nsplit=%u\naccel=%u\n", fastCoverParams->k, |
| 414 | fastCoverParams->d, fastCoverParams->f, fastCoverParams->steps, splitPercentage, |
| 415 | fastCoverParams->accel); |
| 416 | } |
| 417 | } else { |
| 418 | dictSize = ZDICT_trainFromBuffer_fastCover(dictBuffer, maxDictSize, srcBuffer, |
| 419 | sampleSizes, nbSamplesLoaded, *fastCoverParams); |
| 420 | } |
| 421 | } else { |
| 422 | assert(0 /* Impossible */); |
| 423 | } |
| 424 | if (ZDICT_isError(dictSize)) { |
| 425 | DISPLAYLEVEL(1, "dictionary training failed : %s \n", ZDICT_getErrorName(dictSize)); /* should not happen */ |
| 426 | result = 1; |
| 427 | goto _cleanup; |
| 428 | } |
| 429 | /* save dict */ |
| 430 | DISPLAYLEVEL(2, "Save dictionary of size %u into file %s \n", (unsigned)dictSize, dictFileName); |
| 431 | DiB_saveDict(dictFileName, dictBuffer, dictSize); |
| 432 | } |
| 433 | |
| 434 | /* clean up */ |
| 435 | _cleanup: |
| 436 | free(srcBuffer); |
| 437 | free(sampleSizes); |
| 438 | free(dictBuffer); |
| 439 | return result; |
| 440 | } |