1 /*
2 * Copyright (c) 2016-2020, 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 <errno.h> /* errno */
30 #include <assert.h>
31
32 #include "timefn.h" /* UTIL_time_t, UTIL_clockSpanMicro, UTIL_getTime */
33 #include "../lib/common/mem.h" /* read */
34 #include "../lib/common/error_private.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
53
54 /*-*************************************
55 * Console display
56 ***************************************/
57 #define DISPLAY(...) fprintf(stderr, __VA_ARGS__)
58 #define DISPLAYLEVEL(l, ...) if (displayLevel>=l) { DISPLAY(__VA_ARGS__); }
59
60 static const U64 g_refreshRate = SEC_TO_MICRO / 6;
61 static UTIL_time_t g_displayClock = UTIL_TIME_INITIALIZER;
62
63 #define DISPLAYUPDATE(l, ...) { if (displayLevel>=l) { \
64 if ((UTIL_clockSpanMicro(g_displayClock) > g_refreshRate) || (displayLevel>=4)) \
65 { g_displayClock = UTIL_getTime(); DISPLAY(__VA_ARGS__); \
66 if (displayLevel>=4) fflush(stderr); } } }
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 #undef MIN
89 #define MIN(a,b) ((a) < (b) ? (a) : (b))
90
91
92 /* ********************************************************
93 * File related operations
94 **********************************************************/
95 /** DiB_loadFiles() :
96 * load samples from files listed in fileNamesTable into buffer.
97 * works even if buffer is too small to load all samples.
98 * Also provides the size of each sample into sampleSizes table
99 * which must be sized correctly, using DiB_fileStats().
100 * @return : nb of samples effectively loaded into `buffer`
101 * *bufferSizePtr is modified, it provides the amount data loaded within buffer.
102 * sampleSizes is filled with the size of each sample.
103 */
DiB_loadFiles(void * buffer,size_t * bufferSizePtr,size_t * sampleSizes,unsigned sstSize,const char ** fileNamesTable,unsigned nbFiles,size_t targetChunkSize,unsigned displayLevel)104 static unsigned DiB_loadFiles(void* buffer, size_t* bufferSizePtr,
105 size_t* sampleSizes, unsigned sstSize,
106 const char** fileNamesTable, unsigned nbFiles, size_t targetChunkSize,
107 unsigned displayLevel)
108 {
109 char* const buff = (char*)buffer;
110 size_t pos = 0;
111 unsigned nbLoadedChunks = 0, fileIndex;
112
113 for (fileIndex=0; fileIndex<nbFiles; fileIndex++) {
114 const char* const fileName = fileNamesTable[fileIndex];
115 unsigned long long const fs64 = UTIL_getFileSize(fileName);
116 unsigned long long remainingToLoad = (fs64 == UTIL_FILESIZE_UNKNOWN) ? 0 : fs64;
117 U32 const nbChunks = targetChunkSize ? (U32)((fs64 + (targetChunkSize-1)) / targetChunkSize) : 1;
118 U64 const chunkSize = targetChunkSize ? MIN(targetChunkSize, fs64) : fs64;
119 size_t const maxChunkSize = (size_t)MIN(chunkSize, SAMPLESIZE_MAX);
120 U32 cnb;
121 FILE* const f = fopen(fileName, "rb");
122 if (f==NULL) EXM_THROW(10, "zstd: dictBuilder: %s %s ", fileName, strerror(errno));
123 DISPLAYUPDATE(2, "Loading %s... \r", fileName);
124 for (cnb=0; cnb<nbChunks; cnb++) {
125 size_t const toLoad = (size_t)MIN(maxChunkSize, remainingToLoad);
126 if (toLoad > *bufferSizePtr-pos) break;
127 { size_t const readSize = fread(buff+pos, 1, toLoad, f);
128 if (readSize != toLoad) EXM_THROW(11, "Pb reading %s", fileName);
129 pos += readSize;
130 sampleSizes[nbLoadedChunks++] = toLoad;
131 remainingToLoad -= targetChunkSize;
132 if (nbLoadedChunks == sstSize) { /* no more space left in sampleSizes table */
133 fileIndex = nbFiles; /* stop there */
134 break;
135 }
136 if (toLoad < targetChunkSize) {
137 fseek(f, (long)(targetChunkSize - toLoad), SEEK_CUR);
138 } } }
139 fclose(f);
140 }
141 DISPLAYLEVEL(2, "\r%79s\r", "");
142 *bufferSizePtr = pos;
143 DISPLAYLEVEL(4, "loaded : %u KB \n", (unsigned)(pos >> 10))
144 return nbLoadedChunks;
145 }
146
147 #define DiB_rotl32(x,r) ((x << r) | (x >> (32 - r)))
DiB_rand(U32 * src)148 static U32 DiB_rand(U32* src)
149 {
150 static const U32 prime1 = 2654435761U;
151 static const U32 prime2 = 2246822519U;
152 U32 rand32 = *src;
153 rand32 *= prime1;
154 rand32 ^= prime2;
155 rand32 = DiB_rotl32(rand32, 13);
156 *src = rand32;
157 return rand32 >> 5;
158 }
159
160 /* DiB_shuffle() :
161 * shuffle a table of file names in a semi-random way
162 * It improves dictionary quality by reducing "locality" impact, so if sample set is very large,
163 * it will load random elements from it, instead of just the first ones. */
DiB_shuffle(const char ** fileNamesTable,unsigned nbFiles)164 static void DiB_shuffle(const char** fileNamesTable, unsigned nbFiles) {
165 U32 seed = 0xFD2FB528;
166 unsigned i;
167 assert(nbFiles >= 1);
168 for (i = nbFiles - 1; i > 0; --i) {
169 unsigned const j = DiB_rand(&seed) % (i + 1);
170 const char* const tmp = fileNamesTable[j];
171 fileNamesTable[j] = fileNamesTable[i];
172 fileNamesTable[i] = tmp;
173 }
174 }
175
176
177 /*-********************************************************
178 * Dictionary training functions
179 **********************************************************/
DiB_findMaxMem(unsigned long long requiredMem)180 static size_t DiB_findMaxMem(unsigned long long requiredMem)
181 {
182 size_t const step = 8 MB;
183 void* testmem = NULL;
184
185 requiredMem = (((requiredMem >> 23) + 1) << 23);
186 requiredMem += step;
187 if (requiredMem > g_maxMemory) requiredMem = g_maxMemory;
188
189 while (!testmem) {
190 testmem = malloc((size_t)requiredMem);
191 requiredMem -= step;
192 }
193
194 free(testmem);
195 return (size_t)requiredMem;
196 }
197
198
DiB_fillNoise(void * buffer,size_t length)199 static void DiB_fillNoise(void* buffer, size_t length)
200 {
201 unsigned const prime1 = 2654435761U;
202 unsigned const prime2 = 2246822519U;
203 unsigned acc = prime1;
204 size_t p=0;
205
206 for (p=0; p<length; p++) {
207 acc *= prime2;
208 ((unsigned char*)buffer)[p] = (unsigned char)(acc >> 21);
209 }
210 }
211
212
DiB_saveDict(const char * dictFileName,const void * buff,size_t buffSize)213 static void DiB_saveDict(const char* dictFileName,
214 const void* buff, size_t buffSize)
215 {
216 FILE* const f = fopen(dictFileName, "wb");
217 if (f==NULL) EXM_THROW(3, "cannot open %s ", dictFileName);
218
219 { size_t const n = fwrite(buff, 1, buffSize, f);
220 if (n!=buffSize) EXM_THROW(4, "%s : write error", dictFileName) }
221
222 { size_t const n = (size_t)fclose(f);
223 if (n!=0) EXM_THROW(5, "%s : flush error", dictFileName) }
224 }
225
226
227 typedef struct {
228 U64 totalSizeToLoad;
229 unsigned oneSampleTooLarge;
230 unsigned nbSamples;
231 } fileStats;
232
233 /*! DiB_fileStats() :
234 * Given a list of files, and a chunkSize (0 == no chunk, whole files)
235 * provides the amount of data to be loaded and the resulting nb of samples.
236 * This is useful primarily for allocation purpose => sample buffer, and sample sizes table.
237 */
DiB_fileStats(const char ** fileNamesTable,unsigned nbFiles,size_t chunkSize,unsigned displayLevel)238 static fileStats DiB_fileStats(const char** fileNamesTable, unsigned nbFiles, size_t chunkSize, unsigned displayLevel)
239 {
240 fileStats fs;
241 unsigned n;
242 memset(&fs, 0, sizeof(fs));
243 for (n=0; n<nbFiles; n++) {
244 U64 const fileSize = UTIL_getFileSize(fileNamesTable[n]);
245 U64 const srcSize = (fileSize == UTIL_FILESIZE_UNKNOWN) ? 0 : fileSize;
246 U32 const nbSamples = (U32)(chunkSize ? (srcSize + (chunkSize-1)) / chunkSize : 1);
247 U64 const chunkToLoad = chunkSize ? MIN(chunkSize, srcSize) : srcSize;
248 size_t const cappedChunkSize = (size_t)MIN(chunkToLoad, SAMPLESIZE_MAX);
249 fs.totalSizeToLoad += cappedChunkSize * nbSamples;
250 fs.oneSampleTooLarge |= (chunkSize > 2*SAMPLESIZE_MAX);
251 fs.nbSamples += nbSamples;
252 }
253 DISPLAYLEVEL(4, "Preparing to load : %u KB \n", (unsigned)(fs.totalSizeToLoad >> 10));
254 return fs;
255 }
256
257
258 /*! ZDICT_trainFromBuffer_unsafe_legacy() :
259 Strictly Internal use only !!
260 Same as ZDICT_trainFromBuffer_legacy(), but does not control `samplesBuffer`.
261 `samplesBuffer` must be followed by noisy guard band to avoid out-of-buffer reads.
262 @return : size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
263 or an error code.
264 */
265 size_t ZDICT_trainFromBuffer_unsafe_legacy(void* dictBuffer, size_t dictBufferCapacity,
266 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
267 ZDICT_legacy_params_t parameters);
268
269
DiB_trainFromFiles(const char * dictFileName,unsigned maxDictSize,const char ** fileNamesTable,unsigned nbFiles,size_t chunkSize,ZDICT_legacy_params_t * params,ZDICT_cover_params_t * coverParams,ZDICT_fastCover_params_t * fastCoverParams,int optimize)270 int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize,
271 const char** fileNamesTable, unsigned nbFiles, size_t chunkSize,
272 ZDICT_legacy_params_t* params, ZDICT_cover_params_t* coverParams,
273 ZDICT_fastCover_params_t* fastCoverParams, int optimize)
274 {
275 unsigned const displayLevel = params ? params->zParams.notificationLevel :
276 coverParams ? coverParams->zParams.notificationLevel :
277 fastCoverParams ? fastCoverParams->zParams.notificationLevel :
278 0; /* should never happen */
279 void* const dictBuffer = malloc(maxDictSize);
280 fileStats const fs = DiB_fileStats(fileNamesTable, nbFiles, chunkSize, displayLevel);
281 size_t* const sampleSizes = (size_t*)malloc(fs.nbSamples * sizeof(size_t));
282 size_t const memMult = params ? MEMMULT :
283 coverParams ? COVER_MEMMULT:
284 FASTCOVER_MEMMULT;
285 size_t const maxMem = DiB_findMaxMem(fs.totalSizeToLoad * memMult) / memMult;
286 size_t loadedSize = (size_t) MIN ((unsigned long long)maxMem, fs.totalSizeToLoad);
287 void* const srcBuffer = malloc(loadedSize+NOISELENGTH);
288 int result = 0;
289
290 /* Checks */
291 if ((!sampleSizes) || (!srcBuffer) || (!dictBuffer))
292 EXM_THROW(12, "not enough memory for DiB_trainFiles"); /* should not happen */
293 if (fs.oneSampleTooLarge) {
294 DISPLAYLEVEL(2, "! Warning : some sample(s) are very large \n");
295 DISPLAYLEVEL(2, "! Note that dictionary is only useful for small samples. \n");
296 DISPLAYLEVEL(2, "! As a consequence, only the first %u bytes of each sample are loaded \n", SAMPLESIZE_MAX);
297 }
298 if (fs.nbSamples < 5) {
299 DISPLAYLEVEL(2, "! Warning : nb of samples too low for proper processing ! \n");
300 DISPLAYLEVEL(2, "! Please provide _one file per sample_. \n");
301 DISPLAYLEVEL(2, "! Alternatively, split files into fixed-size blocks representative of samples, with -B# \n");
302 EXM_THROW(14, "nb of samples too low"); /* we now clearly forbid this case */
303 }
304 if (fs.totalSizeToLoad < (unsigned long long)maxDictSize * 8) {
305 DISPLAYLEVEL(2, "! Warning : data size of samples too small for target dictionary size \n");
306 DISPLAYLEVEL(2, "! Samples should be about 100x larger than target dictionary size \n");
307 }
308
309 /* init */
310 if (loadedSize < fs.totalSizeToLoad)
311 DISPLAYLEVEL(1, "Not enough memory; training on %u MB only...\n", (unsigned)(loadedSize >> 20));
312
313 /* Load input buffer */
314 DISPLAYLEVEL(3, "Shuffling input files\n");
315 DiB_shuffle(fileNamesTable, nbFiles);
316
317 DiB_loadFiles(srcBuffer, &loadedSize, sampleSizes, fs.nbSamples, fileNamesTable, nbFiles, chunkSize, displayLevel);
318
319 { size_t dictSize;
320 if (params) {
321 DiB_fillNoise((char*)srcBuffer + loadedSize, NOISELENGTH); /* guard band, for end of buffer condition */
322 dictSize = ZDICT_trainFromBuffer_unsafe_legacy(dictBuffer, maxDictSize,
323 srcBuffer, sampleSizes, fs.nbSamples,
324 *params);
325 } else if (coverParams) {
326 if (optimize) {
327 dictSize = ZDICT_optimizeTrainFromBuffer_cover(dictBuffer, maxDictSize,
328 srcBuffer, sampleSizes, fs.nbSamples,
329 coverParams);
330 if (!ZDICT_isError(dictSize)) {
331 unsigned splitPercentage = (unsigned)(coverParams->splitPoint * 100);
332 DISPLAYLEVEL(2, "k=%u\nd=%u\nsteps=%u\nsplit=%u\n", coverParams->k, coverParams->d,
333 coverParams->steps, splitPercentage);
334 }
335 } else {
336 dictSize = ZDICT_trainFromBuffer_cover(dictBuffer, maxDictSize, srcBuffer,
337 sampleSizes, fs.nbSamples, *coverParams);
338 }
339 } else {
340 assert(fastCoverParams != NULL);
341 if (optimize) {
342 dictSize = ZDICT_optimizeTrainFromBuffer_fastCover(dictBuffer, maxDictSize,
343 srcBuffer, sampleSizes, fs.nbSamples,
344 fastCoverParams);
345 if (!ZDICT_isError(dictSize)) {
346 unsigned splitPercentage = (unsigned)(fastCoverParams->splitPoint * 100);
347 DISPLAYLEVEL(2, "k=%u\nd=%u\nf=%u\nsteps=%u\nsplit=%u\naccel=%u\n", fastCoverParams->k,
348 fastCoverParams->d, fastCoverParams->f, fastCoverParams->steps, splitPercentage,
349 fastCoverParams->accel);
350 }
351 } else {
352 dictSize = ZDICT_trainFromBuffer_fastCover(dictBuffer, maxDictSize, srcBuffer,
353 sampleSizes, fs.nbSamples, *fastCoverParams);
354 }
355 }
356 if (ZDICT_isError(dictSize)) {
357 DISPLAYLEVEL(1, "dictionary training failed : %s \n", ZDICT_getErrorName(dictSize)); /* should not happen */
358 result = 1;
359 goto _cleanup;
360 }
361 /* save dict */
362 DISPLAYLEVEL(2, "Save dictionary of size %u into file %s \n", (unsigned)dictSize, dictFileName);
363 DiB_saveDict(dictFileName, dictBuffer, dictSize);
364 }
365
366 /* clean up */
367 _cleanup:
368 free(srcBuffer);
369 free(sampleSizes);
370 free(dictBuffer);
371 return result;
372 }
373