xref: /freebsd/sys/contrib/zstd/programs/dibio.c (revision 1719886f6d08408b834d270c59ffcfd821c8f63a)
1 /*
2  * Copyright (c) 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 "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 #define FASTCOVER_MEMMULT 1    /* rough estimation : memory cost to analyze 1 byte of sample */
48 static const size_t g_maxMemory = (sizeof(size_t) == 4) ? (2 GB - 64 MB) : ((size_t)(512 MB) << sizeof(size_t));
49 
50 #define NOISELENGTH 32
51 #define MAX_SAMPLES_SIZE (2 GB) /* training dataset limited to 2GB */
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   Returns the size of a file.
93   If error returns -1.
94 */
95 static S64 DiB_getFileSize (const char * fileName)
96 {
97     U64 const fileSize = UTIL_getFileSize(fileName);
98     return (fileSize == UTIL_FILESIZE_UNKNOWN) ? -1 : (S64)fileSize;
99 }
100 
101 /* ********************************************************
102 *  File related operations
103 **********************************************************/
104 /** DiB_loadFiles() :
105  *  load samples from files listed in fileNamesTable into buffer.
106  *  works even if buffer is too small to load all samples.
107  *  Also provides the size of each sample into sampleSizes table
108  *  which must be sized correctly, using DiB_fileStats().
109  * @return : nb of samples effectively loaded into `buffer`
110  * *bufferSizePtr is modified, it provides the amount data loaded within buffer.
111  *  sampleSizes is filled with the size of each sample.
112  */
113 static int DiB_loadFiles(
114     void* buffer, size_t* bufferSizePtr,
115     size_t* sampleSizes, int sstSize,
116     const char** fileNamesTable, int nbFiles,
117     size_t targetChunkSize, int displayLevel )
118 {
119     char* const buff = (char*)buffer;
120     size_t totalDataLoaded = 0;
121     int nbSamplesLoaded = 0;
122     int fileIndex = 0;
123     FILE * f = NULL;
124 
125     assert(targetChunkSize <= SAMPLESIZE_MAX);
126 
127     while ( nbSamplesLoaded < sstSize && fileIndex < nbFiles ) {
128         size_t fileDataLoaded;
129         S64 const fileSize = DiB_getFileSize(fileNamesTable[fileIndex]);
130         if (fileSize <= 0) /* skip if zero-size or file error */
131             continue;
132 
133         f = fopen( fileNamesTable[fileIndex], "rb");
134         if (f == NULL)
135             EXM_THROW(10, "zstd: dictBuilder: %s %s ", fileNamesTable[fileIndex], strerror(errno));
136         DISPLAYUPDATE(2, "Loading %s...       \r", fileNamesTable[fileIndex]);
137 
138         /* Load the first chunk of data from the file */
139         fileDataLoaded = targetChunkSize > 0 ?
140                             (size_t)MIN(fileSize, (S64)targetChunkSize) :
141                             (size_t)MIN(fileSize, SAMPLESIZE_MAX );
142         if (totalDataLoaded + fileDataLoaded > *bufferSizePtr)
143             break;
144         if (fread( buff+totalDataLoaded, 1, fileDataLoaded, f ) != fileDataLoaded)
145             EXM_THROW(11, "Pb reading %s", fileNamesTable[fileIndex]);
146         sampleSizes[nbSamplesLoaded++] = fileDataLoaded;
147         totalDataLoaded += fileDataLoaded;
148 
149         /* If file-chunking is enabled, load the rest of the file as more samples */
150         if (targetChunkSize > 0) {
151             while( (S64)fileDataLoaded < fileSize && nbSamplesLoaded < sstSize ) {
152                 size_t const chunkSize = MIN((size_t)(fileSize-fileDataLoaded), targetChunkSize);
153                 if (totalDataLoaded + chunkSize > *bufferSizePtr) /* buffer is full */
154                     break;
155 
156                 if (fread( buff+totalDataLoaded, 1, chunkSize, f ) != chunkSize)
157                     EXM_THROW(11, "Pb reading %s", fileNamesTable[fileIndex]);
158                 sampleSizes[nbSamplesLoaded++] = chunkSize;
159                 totalDataLoaded += chunkSize;
160                 fileDataLoaded += chunkSize;
161             }
162         }
163         fileIndex += 1;
164         fclose(f); f = NULL;
165     }
166     if (f != NULL)
167         fclose(f);
168 
169     DISPLAYLEVEL(2, "\r%79s\r", "");
170     DISPLAYLEVEL(4, "Loaded %d KB total training data, %d nb samples \n",
171         (int)(totalDataLoaded / (1 KB)), nbSamplesLoaded );
172     *bufferSizePtr = totalDataLoaded;
173     return nbSamplesLoaded;
174 }
175 
176 #define DiB_rotl32(x,r) ((x << r) | (x >> (32 - r)))
177 static U32 DiB_rand(U32* src)
178 {
179     static const U32 prime1 = 2654435761U;
180     static const U32 prime2 = 2246822519U;
181     U32 rand32 = *src;
182     rand32 *= prime1;
183     rand32 ^= prime2;
184     rand32  = DiB_rotl32(rand32, 13);
185     *src = rand32;
186     return rand32 >> 5;
187 }
188 
189 /* DiB_shuffle() :
190  * shuffle a table of file names in a semi-random way
191  * It improves dictionary quality by reducing "locality" impact, so if sample set is very large,
192  * it will load random elements from it, instead of just the first ones. */
193 static void DiB_shuffle(const char** fileNamesTable, unsigned nbFiles) {
194     U32 seed = 0xFD2FB528;
195     unsigned i;
196     assert(nbFiles >= 1);
197     for (i = nbFiles - 1; i > 0; --i) {
198         unsigned const j = DiB_rand(&seed) % (i + 1);
199         const char* const tmp = fileNamesTable[j];
200         fileNamesTable[j] = fileNamesTable[i];
201         fileNamesTable[i] = tmp;
202     }
203 }
204 
205 
206 /*-********************************************************
207 *  Dictionary training functions
208 **********************************************************/
209 static size_t DiB_findMaxMem(unsigned long long requiredMem)
210 {
211     size_t const step = 8 MB;
212     void* testmem = NULL;
213 
214     requiredMem = (((requiredMem >> 23) + 1) << 23);
215     requiredMem += step;
216     if (requiredMem > g_maxMemory) requiredMem = g_maxMemory;
217 
218     while (!testmem) {
219         testmem = malloc((size_t)requiredMem);
220         requiredMem -= step;
221     }
222 
223     free(testmem);
224     return (size_t)requiredMem;
225 }
226 
227 
228 static void DiB_fillNoise(void* buffer, size_t length)
229 {
230     unsigned const prime1 = 2654435761U;
231     unsigned const prime2 = 2246822519U;
232     unsigned acc = prime1;
233     size_t p=0;
234 
235     for (p=0; p<length; p++) {
236         acc *= prime2;
237         ((unsigned char*)buffer)[p] = (unsigned char)(acc >> 21);
238     }
239 }
240 
241 
242 static void DiB_saveDict(const char* dictFileName,
243                          const void* buff, size_t buffSize)
244 {
245     FILE* const f = fopen(dictFileName, "wb");
246     if (f==NULL) EXM_THROW(3, "cannot open %s ", dictFileName);
247 
248     { size_t const n = fwrite(buff, 1, buffSize, f);
249       if (n!=buffSize) EXM_THROW(4, "%s : write error", dictFileName) }
250 
251     { size_t const n = (size_t)fclose(f);
252       if (n!=0) EXM_THROW(5, "%s : flush error", dictFileName) }
253 }
254 
255 typedef struct {
256     S64 totalSizeToLoad;
257     int nbSamples;
258     int oneSampleTooLarge;
259 } fileStats;
260 
261 /*! DiB_fileStats() :
262  *  Given a list of files, and a chunkSize (0 == no chunk, whole files)
263  *  provides the amount of data to be loaded and the resulting nb of samples.
264  *  This is useful primarily for allocation purpose => sample buffer, and sample sizes table.
265  */
266 static fileStats DiB_fileStats(const char** fileNamesTable, int nbFiles, size_t chunkSize, int displayLevel)
267 {
268     fileStats fs;
269     int n;
270     memset(&fs, 0, sizeof(fs));
271 
272     // We assume that if chunking is requested, the chunk size is < SAMPLESIZE_MAX
273     assert( chunkSize <= SAMPLESIZE_MAX );
274 
275     for (n=0; n<nbFiles; n++) {
276       S64 const fileSize = DiB_getFileSize(fileNamesTable[n]);
277       // TODO: is there a minimum sample size? What if the file is 1-byte?
278       if (fileSize == 0) {
279         DISPLAYLEVEL(3, "Sample file '%s' has zero size, skipping...\n", fileNamesTable[n]);
280         continue;
281       }
282 
283       /* the case where we are breaking up files in sample chunks */
284       if (chunkSize > 0)
285       {
286         // TODO: is there a minimum sample size? Can we have a 1-byte sample?
287         fs.nbSamples += (int)((fileSize + chunkSize-1) / chunkSize);
288         fs.totalSizeToLoad += fileSize;
289       }
290       else {
291       /* the case where one file is one sample */
292         if (fileSize > SAMPLESIZE_MAX) {
293           /* flag excessively large sample files */
294           fs.oneSampleTooLarge |= (fileSize > 2*SAMPLESIZE_MAX);
295 
296           /* Limit to the first SAMPLESIZE_MAX (128kB) of the file */
297           DISPLAYLEVEL(3, "Sample file '%s' is too large, limiting to %d KB",
298               fileNamesTable[n], SAMPLESIZE_MAX / (1 KB));
299         }
300         fs.nbSamples += 1;
301         fs.totalSizeToLoad += MIN(fileSize, SAMPLESIZE_MAX);
302       }
303     }
304     DISPLAYLEVEL(4, "Found training data %d files, %d KB, %d samples\n", nbFiles, (int)(fs.totalSizeToLoad / (1 KB)), fs.nbSamples);
305     return fs;
306 }
307 
308 int DiB_trainFromFiles(const char* dictFileName, size_t maxDictSize,
309                        const char** fileNamesTable, int nbFiles, size_t chunkSize,
310                        ZDICT_legacy_params_t* params, ZDICT_cover_params_t* coverParams,
311                        ZDICT_fastCover_params_t* fastCoverParams, int optimize, unsigned memLimit)
312 {
313     fileStats fs;
314     size_t* sampleSizes; /* vector of sample sizes. Each sample can be up to SAMPLESIZE_MAX */
315     int nbSamplesLoaded; /* nb of samples effectively loaded in srcBuffer */
316     size_t loadedSize; /* total data loaded in srcBuffer for all samples */
317     void* srcBuffer /* contiguous buffer with training data/samples */;
318     void* const dictBuffer = malloc(maxDictSize);
319     int result = 0;
320 
321     int const displayLevel = params ? params->zParams.notificationLevel :
322         coverParams ? coverParams->zParams.notificationLevel :
323         fastCoverParams ? fastCoverParams->zParams.notificationLevel : 0;
324 
325     /* Shuffle input files before we start assessing how much sample datA to load.
326        The purpose of the shuffle is to pick random samples when the sample
327        set is larger than what we can load in memory. */
328     DISPLAYLEVEL(3, "Shuffling input files\n");
329     DiB_shuffle(fileNamesTable, nbFiles);
330 
331     /* Figure out how much sample data to load with how many samples */
332     fs = DiB_fileStats(fileNamesTable, nbFiles, chunkSize, displayLevel);
333 
334     {
335         int const memMult = params ? MEMMULT :
336                             coverParams ? COVER_MEMMULT:
337                             FASTCOVER_MEMMULT;
338         size_t const maxMem =  DiB_findMaxMem(fs.totalSizeToLoad * memMult) / memMult;
339         /* Limit the size of the training data to the free memory */
340         /* Limit the size of the training data to 2GB */
341         /* TODO: there is opportunity to stop DiB_fileStats() early when the data limit is reached */
342         loadedSize = (size_t)MIN( MIN((S64)maxMem, fs.totalSizeToLoad), MAX_SAMPLES_SIZE );
343         if (memLimit != 0) {
344             DISPLAYLEVEL(2, "!  Warning : setting manual memory limit for dictionary training data at %u MB \n",
345                 (unsigned)(memLimit / (1 MB)));
346             loadedSize = (size_t)MIN(loadedSize, memLimit);
347         }
348         srcBuffer = malloc(loadedSize+NOISELENGTH);
349         sampleSizes = (size_t*)malloc(fs.nbSamples * sizeof(size_t));
350     }
351 
352     /* Checks */
353     if ((!sampleSizes) || (!srcBuffer) || (!dictBuffer))
354         EXM_THROW(12, "not enough memory for DiB_trainFiles");   /* should not happen */
355     if (fs.oneSampleTooLarge) {
356         DISPLAYLEVEL(2, "!  Warning : some sample(s) are very large \n");
357         DISPLAYLEVEL(2, "!  Note that dictionary is only useful for small samples. \n");
358         DISPLAYLEVEL(2, "!  As a consequence, only the first %u bytes of each sample are loaded \n", SAMPLESIZE_MAX);
359     }
360     if (fs.nbSamples < 5) {
361         DISPLAYLEVEL(2, "!  Warning : nb of samples too low for proper processing ! \n");
362         DISPLAYLEVEL(2, "!  Please provide _one file per sample_. \n");
363         DISPLAYLEVEL(2, "!  Alternatively, split files into fixed-size blocks representative of samples, with -B# \n");
364         EXM_THROW(14, "nb of samples too low");   /* we now clearly forbid this case */
365     }
366     if (fs.totalSizeToLoad < (S64)maxDictSize * 8) {
367         DISPLAYLEVEL(2, "!  Warning : data size of samples too small for target dictionary size \n");
368         DISPLAYLEVEL(2, "!  Samples should be about 100x larger than target dictionary size \n");
369     }
370 
371     /* init */
372     if ((S64)loadedSize < fs.totalSizeToLoad)
373         DISPLAYLEVEL(1, "Training samples set too large (%u MB); training on %u MB only...\n",
374             (unsigned)(fs.totalSizeToLoad / (1 MB)),
375             (unsigned)(loadedSize / (1 MB)));
376 
377     /* Load input buffer */
378     nbSamplesLoaded = DiB_loadFiles(
379         srcBuffer, &loadedSize, sampleSizes, fs.nbSamples, fileNamesTable,
380         nbFiles, chunkSize, displayLevel);
381 
382     {   size_t dictSize;
383         if (params) {
384             DiB_fillNoise((char*)srcBuffer + loadedSize, NOISELENGTH);   /* guard band, for end of buffer condition */
385             dictSize = ZDICT_trainFromBuffer_legacy(dictBuffer, maxDictSize,
386                                                     srcBuffer, sampleSizes, nbSamplesLoaded,
387                                                     *params);
388         } else if (coverParams) {
389             if (optimize) {
390               dictSize = ZDICT_optimizeTrainFromBuffer_cover(dictBuffer, maxDictSize,
391                                                              srcBuffer, sampleSizes, nbSamplesLoaded,
392                                                              coverParams);
393               if (!ZDICT_isError(dictSize)) {
394                   unsigned splitPercentage = (unsigned)(coverParams->splitPoint * 100);
395                   DISPLAYLEVEL(2, "k=%u\nd=%u\nsteps=%u\nsplit=%u\n", coverParams->k, coverParams->d,
396                               coverParams->steps, splitPercentage);
397               }
398             } else {
399               dictSize = ZDICT_trainFromBuffer_cover(dictBuffer, maxDictSize, srcBuffer,
400                                                      sampleSizes, nbSamplesLoaded, *coverParams);
401             }
402         } else {
403             assert(fastCoverParams != NULL);
404             if (optimize) {
405               dictSize = ZDICT_optimizeTrainFromBuffer_fastCover(dictBuffer, maxDictSize,
406                                                               srcBuffer, sampleSizes, nbSamplesLoaded,
407                                                               fastCoverParams);
408               if (!ZDICT_isError(dictSize)) {
409                 unsigned splitPercentage = (unsigned)(fastCoverParams->splitPoint * 100);
410                 DISPLAYLEVEL(2, "k=%u\nd=%u\nf=%u\nsteps=%u\nsplit=%u\naccel=%u\n", fastCoverParams->k,
411                             fastCoverParams->d, fastCoverParams->f, fastCoverParams->steps, splitPercentage,
412                             fastCoverParams->accel);
413               }
414             } else {
415               dictSize = ZDICT_trainFromBuffer_fastCover(dictBuffer, maxDictSize, srcBuffer,
416                                                         sampleSizes, nbSamplesLoaded, *fastCoverParams);
417             }
418         }
419         if (ZDICT_isError(dictSize)) {
420             DISPLAYLEVEL(1, "dictionary training failed : %s \n", ZDICT_getErrorName(dictSize));   /* should not happen */
421             result = 1;
422             goto _cleanup;
423         }
424         /* save dict */
425         DISPLAYLEVEL(2, "Save dictionary of size %u into file %s \n", (unsigned)dictSize, dictFileName);
426         DiB_saveDict(dictFileName, dictBuffer, dictSize);
427     }
428 
429     /* clean up */
430 _cleanup:
431     free(srcBuffer);
432     free(sampleSizes);
433     free(dictBuffer);
434     return result;
435 }
436