1 /* 2 * Copyright (c) 2016-present, Yann Collet, Facebook, Inc. 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 <time.h> /* clock_t, clock, CLOCKS_PER_SEC */ 30 #include <errno.h> /* errno */ 31 32 #include "mem.h" /* read */ 33 #include "error_private.h" 34 #include "dibio.h" 35 36 37 /*-************************************* 38 * Constants 39 ***************************************/ 40 #define KB *(1 <<10) 41 #define MB *(1 <<20) 42 #define GB *(1U<<30) 43 44 #define SAMPLESIZE_MAX (128 KB) 45 #define MEMMULT 11 /* rough estimation : memory cost to analyze 1 byte of sample */ 46 #define COVER_MEMMULT 9 /* rough estimation : memory cost to analyze 1 byte of sample */ 47 static const size_t g_maxMemory = (sizeof(size_t) == 4) ? (2 GB - 64 MB) : ((size_t)(512 MB) << sizeof(size_t)); 48 49 #define NOISELENGTH 32 50 51 52 /*-************************************* 53 * Console display 54 ***************************************/ 55 #define DISPLAY(...) fprintf(stderr, __VA_ARGS__) 56 #define DISPLAYLEVEL(l, ...) if (displayLevel>=l) { DISPLAY(__VA_ARGS__); } 57 58 #define DISPLAYUPDATE(l, ...) if (displayLevel>=l) { \ 59 if ((DIB_clockSpan(g_time) > refreshRate) || (displayLevel>=4)) \ 60 { g_time = clock(); DISPLAY(__VA_ARGS__); \ 61 if (displayLevel>=4) fflush(stderr); } } 62 static const clock_t refreshRate = CLOCKS_PER_SEC * 2 / 10; 63 static clock_t g_time = 0; 64 65 static clock_t DIB_clockSpan(clock_t nPrevious) { return clock() - nPrevious; } 66 67 68 /*-************************************* 69 * Exceptions 70 ***************************************/ 71 #ifndef DEBUG 72 # define DEBUG 0 73 #endif 74 #define DEBUGOUTPUT(...) if (DEBUG) DISPLAY(__VA_ARGS__); 75 #define EXM_THROW(error, ...) \ 76 { \ 77 DEBUGOUTPUT("Error defined at %s, line %i : \n", __FILE__, __LINE__); \ 78 DISPLAY("Error %i : ", error); \ 79 DISPLAY(__VA_ARGS__); \ 80 DISPLAY("\n"); \ 81 exit(error); \ 82 } 83 84 85 /* ******************************************************** 86 * Helper functions 87 **********************************************************/ 88 unsigned DiB_isError(size_t errorCode) { return ERR_isError(errorCode); } 89 90 const char* DiB_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCode); } 91 92 #undef MIN 93 #define MIN(a,b) ((a) < (b) ? (a) : (b)) 94 95 96 /* ******************************************************** 97 * File related operations 98 **********************************************************/ 99 /** DiB_loadFiles() : 100 * load samples from files listed in fileNamesTable into buffer. 101 * works even if buffer is too small to load all samples. 102 * Also provides the size of each sample into sampleSizes table 103 * which must be sized correctly, using DiB_fileStats(). 104 * @return : nb of samples effectively loaded into `buffer` 105 * *bufferSizePtr is modified, it provides the amount data loaded within buffer. 106 * sampleSizes is filled with the size of each sample. 107 */ 108 static unsigned DiB_loadFiles(void* buffer, size_t* bufferSizePtr, 109 size_t* sampleSizes, unsigned sstSize, 110 const char** fileNamesTable, unsigned nbFiles, size_t targetChunkSize, 111 unsigned displayLevel) 112 { 113 char* const buff = (char*)buffer; 114 size_t pos = 0; 115 unsigned nbLoadedChunks = 0, fileIndex; 116 117 for (fileIndex=0; fileIndex<nbFiles; fileIndex++) { 118 const char* const fileName = fileNamesTable[fileIndex]; 119 unsigned long long const fs64 = UTIL_getFileSize(fileName); 120 unsigned long long remainingToLoad = fs64; 121 U32 const nbChunks = targetChunkSize ? (U32)((fs64 + (targetChunkSize-1)) / targetChunkSize) : 1; 122 U64 const chunkSize = targetChunkSize ? MIN(targetChunkSize, fs64) : fs64; 123 size_t const maxChunkSize = (size_t)MIN(chunkSize, SAMPLESIZE_MAX); 124 U32 cnb; 125 FILE* const f = fopen(fileName, "rb"); 126 if (f==NULL) EXM_THROW(10, "zstd: dictBuilder: %s %s ", fileName, strerror(errno)); 127 DISPLAYUPDATE(2, "Loading %s... \r", fileName); 128 for (cnb=0; cnb<nbChunks; cnb++) { 129 size_t const toLoad = (size_t)MIN(maxChunkSize, remainingToLoad); 130 if (toLoad > *bufferSizePtr-pos) break; 131 { size_t const readSize = fread(buff+pos, 1, toLoad, f); 132 if (readSize != toLoad) EXM_THROW(11, "Pb reading %s", fileName); 133 pos += readSize; 134 sampleSizes[nbLoadedChunks++] = toLoad; 135 remainingToLoad -= targetChunkSize; 136 if (nbLoadedChunks == sstSize) { /* no more space left in sampleSizes table */ 137 fileIndex = nbFiles; /* stop there */ 138 break; 139 } 140 if (toLoad < targetChunkSize) { 141 fseek(f, (long)(targetChunkSize - toLoad), SEEK_CUR); 142 } } } 143 fclose(f); 144 } 145 DISPLAYLEVEL(2, "\r%79s\r", ""); 146 *bufferSizePtr = pos; 147 DISPLAYLEVEL(4, "loaded : %u KB \n", (U32)(pos >> 10)) 148 return nbLoadedChunks; 149 } 150 151 #define DiB_rotl32(x,r) ((x << r) | (x >> (32 - r))) 152 static U32 DiB_rand(U32* src) 153 { 154 static const U32 prime1 = 2654435761U; 155 static const U32 prime2 = 2246822519U; 156 U32 rand32 = *src; 157 rand32 *= prime1; 158 rand32 ^= prime2; 159 rand32 = DiB_rotl32(rand32, 13); 160 *src = rand32; 161 return rand32 >> 5; 162 } 163 164 /* DiB_shuffle() : 165 * shuffle a table of file names in a semi-random way 166 * It improves dictionary quality by reducing "locality" impact, so if sample set is very large, 167 * it will load random elements from it, instead of just the first ones. */ 168 static void DiB_shuffle(const char** fileNamesTable, unsigned nbFiles) { 169 U32 seed = 0xFD2FB528; 170 unsigned i; 171 for (i = nbFiles - 1; i > 0; --i) { 172 unsigned const j = DiB_rand(&seed) % (i + 1); 173 const char* const tmp = fileNamesTable[j]; 174 fileNamesTable[j] = fileNamesTable[i]; 175 fileNamesTable[i] = tmp; 176 } 177 } 178 179 180 /*-******************************************************** 181 * Dictionary training functions 182 **********************************************************/ 183 static size_t DiB_findMaxMem(unsigned long long requiredMem) 184 { 185 size_t const step = 8 MB; 186 void* testmem = NULL; 187 188 requiredMem = (((requiredMem >> 23) + 1) << 23); 189 requiredMem += step; 190 if (requiredMem > g_maxMemory) requiredMem = g_maxMemory; 191 192 while (!testmem) { 193 testmem = malloc((size_t)requiredMem); 194 requiredMem -= step; 195 } 196 197 free(testmem); 198 return (size_t)requiredMem; 199 } 200 201 202 static void DiB_fillNoise(void* buffer, size_t length) 203 { 204 unsigned const prime1 = 2654435761U; 205 unsigned const prime2 = 2246822519U; 206 unsigned acc = prime1; 207 size_t p=0;; 208 209 for (p=0; p<length; p++) { 210 acc *= prime2; 211 ((unsigned char*)buffer)[p] = (unsigned char)(acc >> 21); 212 } 213 } 214 215 216 static void DiB_saveDict(const char* dictFileName, 217 const void* buff, size_t buffSize) 218 { 219 FILE* const f = fopen(dictFileName, "wb"); 220 if (f==NULL) EXM_THROW(3, "cannot open %s ", dictFileName); 221 222 { size_t const n = fwrite(buff, 1, buffSize, f); 223 if (n!=buffSize) EXM_THROW(4, "%s : write error", dictFileName) } 224 225 { size_t const n = (size_t)fclose(f); 226 if (n!=0) EXM_THROW(5, "%s : flush error", dictFileName) } 227 } 228 229 230 typedef struct { 231 U64 totalSizeToLoad; 232 unsigned oneSampleTooLarge; 233 unsigned nbSamples; 234 } fileStats; 235 236 /*! DiB_fileStats() : 237 * Given a list of files, and a chunkSize (0 == no chunk, whole files) 238 * provides the amount of data to be loaded and the resulting nb of samples. 239 * This is useful primarily for allocation purpose => sample buffer, and sample sizes table. 240 */ 241 static fileStats DiB_fileStats(const char** fileNamesTable, unsigned nbFiles, size_t chunkSize, unsigned displayLevel) 242 { 243 fileStats fs; 244 unsigned n; 245 memset(&fs, 0, sizeof(fs)); 246 for (n=0; n<nbFiles; n++) { 247 U64 const fileSize = UTIL_getFileSize(fileNamesTable[n]); 248 U32 const nbSamples = (U32)(chunkSize ? (fileSize + (chunkSize-1)) / chunkSize : 1); 249 U64 const chunkToLoad = chunkSize ? MIN(chunkSize, fileSize) : fileSize; 250 size_t const cappedChunkSize = (size_t)MIN(chunkToLoad, SAMPLESIZE_MAX); 251 fs.totalSizeToLoad += cappedChunkSize * nbSamples; 252 fs.oneSampleTooLarge |= (chunkSize > 2*SAMPLESIZE_MAX); 253 fs.nbSamples += nbSamples; 254 } 255 DISPLAYLEVEL(4, "Preparing to load : %u KB \n", (U32)(fs.totalSizeToLoad >> 10)); 256 return fs; 257 } 258 259 260 /*! ZDICT_trainFromBuffer_unsafe_legacy() : 261 Strictly Internal use only !! 262 Same as ZDICT_trainFromBuffer_legacy(), but does not control `samplesBuffer`. 263 `samplesBuffer` must be followed by noisy guard band to avoid out-of-buffer reads. 264 @return : size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`) 265 or an error code. 266 */ 267 size_t ZDICT_trainFromBuffer_unsafe_legacy(void* dictBuffer, size_t dictBufferCapacity, 268 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples, 269 ZDICT_legacy_params_t parameters); 270 271 272 int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize, 273 const char** fileNamesTable, unsigned nbFiles, size_t chunkSize, 274 ZDICT_legacy_params_t *params, ZDICT_cover_params_t *coverParams, 275 int optimizeCover) 276 { 277 unsigned const displayLevel = params ? params->zParams.notificationLevel : 278 coverParams ? coverParams->zParams.notificationLevel : 279 0; /* should never happen */ 280 void* const dictBuffer = malloc(maxDictSize); 281 fileStats const fs = DiB_fileStats(fileNamesTable, nbFiles, chunkSize, displayLevel); 282 size_t* const sampleSizes = (size_t*)malloc(fs.nbSamples * sizeof(size_t)); 283 size_t const memMult = params ? MEMMULT : COVER_MEMMULT; 284 size_t const maxMem = DiB_findMaxMem(fs.totalSizeToLoad * memMult) / memMult; 285 size_t loadedSize = (size_t) MIN ((unsigned long long)maxMem, fs.totalSizeToLoad); 286 void* const srcBuffer = malloc(loadedSize+NOISELENGTH); 287 int result = 0; 288 289 /* Checks */ 290 if ((!sampleSizes) || (!srcBuffer) || (!dictBuffer)) 291 EXM_THROW(12, "not enough memory for DiB_trainFiles"); /* should not happen */ 292 if (fs.oneSampleTooLarge) { 293 DISPLAYLEVEL(2, "! Warning : some sample(s) are very large \n"); 294 DISPLAYLEVEL(2, "! Note that dictionary is only useful for small samples. \n"); 295 DISPLAYLEVEL(2, "! As a consequence, only the first %u bytes of each sample are loaded \n", SAMPLESIZE_MAX); 296 } 297 if (fs.nbSamples < 5) { 298 DISPLAYLEVEL(2, "! Warning : nb of samples too low for proper processing ! \n"); 299 DISPLAYLEVEL(2, "! Please provide _one file per sample_. \n"); 300 DISPLAYLEVEL(2, "! Alternatively, split files into fixed-size blocks representative of samples, with -B# \n"); 301 EXM_THROW(14, "nb of samples too low"); /* we now clearly forbid this case */ 302 } 303 if (fs.totalSizeToLoad < (unsigned long long)(8 * maxDictSize)) { 304 DISPLAYLEVEL(2, "! Warning : data size of samples too small for target dictionary size \n"); 305 DISPLAYLEVEL(2, "! Samples should be about 100x larger than target dictionary size \n"); 306 } 307 308 /* init */ 309 if (loadedSize < fs.totalSizeToLoad) 310 DISPLAYLEVEL(1, "Not enough memory; training on %u MB only...\n", (unsigned)(loadedSize >> 20)); 311 312 /* Load input buffer */ 313 DISPLAYLEVEL(3, "Shuffling input files\n"); 314 DiB_shuffle(fileNamesTable, nbFiles); 315 nbFiles = DiB_loadFiles(srcBuffer, &loadedSize, sampleSizes, fs.nbSamples, fileNamesTable, nbFiles, chunkSize, displayLevel); 316 317 { size_t dictSize; 318 if (params) { 319 DiB_fillNoise((char*)srcBuffer + loadedSize, NOISELENGTH); /* guard band, for end of buffer condition */ 320 dictSize = ZDICT_trainFromBuffer_unsafe_legacy(dictBuffer, maxDictSize, 321 srcBuffer, sampleSizes, fs.nbSamples, 322 *params); 323 } else if (optimizeCover) { 324 dictSize = ZDICT_optimizeTrainFromBuffer_cover(dictBuffer, maxDictSize, 325 srcBuffer, sampleSizes, fs.nbSamples, 326 coverParams); 327 if (!ZDICT_isError(dictSize)) { 328 DISPLAYLEVEL(2, "k=%u\nd=%u\nsteps=%u\n", coverParams->k, coverParams->d, coverParams->steps); 329 } 330 } else { 331 dictSize = ZDICT_trainFromBuffer_cover(dictBuffer, maxDictSize, srcBuffer, 332 sampleSizes, fs.nbSamples, *coverParams); 333 } 334 if (ZDICT_isError(dictSize)) { 335 DISPLAYLEVEL(1, "dictionary training failed : %s \n", ZDICT_getErrorName(dictSize)); /* should not happen */ 336 result = 1; 337 goto _cleanup; 338 } 339 /* save dict */ 340 DISPLAYLEVEL(2, "Save dictionary of size %u into file %s \n", (U32)dictSize, dictFileName); 341 DiB_saveDict(dictFileName, dictBuffer, dictSize); 342 } 343 344 /* clean up */ 345 _cleanup: 346 free(srcBuffer); 347 free(sampleSizes); 348 free(dictBuffer); 349 return result; 350 } 351