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 * Tuning parameters
14 ****************************************/
15 #define MINRATIO 4 /* minimum nb of apparition to be selected in dictionary */
16 #define ZDICT_MAX_SAMPLES_SIZE (2000U << 20)
17 #define ZDICT_MIN_SAMPLES_SIZE (ZDICT_CONTENTSIZE_MIN * MINRATIO)
18
19
20 /*-**************************************
21 * Compiler Options
22 ****************************************/
23 /* Unix Large Files support (>4GB) */
24 #define _FILE_OFFSET_BITS 64
25 #if (defined(__sun__) && (!defined(__LP64__))) /* Sun Solaris 32-bits requires specific definitions */
26 # ifndef _LARGEFILE_SOURCE
27 # define _LARGEFILE_SOURCE
28 # endif
29 #elif ! defined(__LP64__) /* No point defining Large file for 64 bit */
30 # ifndef _LARGEFILE64_SOURCE
31 # define _LARGEFILE64_SOURCE
32 # endif
33 #endif
34
35
36 /*-*************************************
37 * Dependencies
38 ***************************************/
39 #include <stdlib.h> /* malloc, free */
40 #include <string.h> /* memset */
41 #include <stdio.h> /* fprintf, fopen, ftello64 */
42 #include <time.h> /* clock */
43
44 #ifndef ZDICT_STATIC_LINKING_ONLY
45 # define ZDICT_STATIC_LINKING_ONLY
46 #endif
47 #define HUF_STATIC_LINKING_ONLY
48
49 #include "../common/mem.h" /* read */
50 #include "../common/fse.h" /* FSE_normalizeCount, FSE_writeNCount */
51 #include "../common/huf.h" /* HUF_buildCTable, HUF_writeCTable */
52 #include "../common/zstd_internal.h" /* includes zstd.h */
53 #include "../common/xxhash.h" /* XXH64 */
54 #include "../compress/zstd_compress_internal.h" /* ZSTD_loadCEntropy() */
55 #include "../zdict.h"
56 #include "divsufsort.h"
57
58
59 /*-*************************************
60 * Constants
61 ***************************************/
62 #define KB *(1 <<10)
63 #define MB *(1 <<20)
64 #define GB *(1U<<30)
65
66 #define DICTLISTSIZE_DEFAULT 10000
67
68 #define NOISELENGTH 32
69
70 static const U32 g_selectivity_default = 9;
71
72
73 /*-*************************************
74 * Console display
75 ***************************************/
76 #undef DISPLAY
77 #define DISPLAY(...) { fprintf(stderr, __VA_ARGS__); fflush( stderr ); }
78 #undef DISPLAYLEVEL
79 #define DISPLAYLEVEL(l, ...) if (notificationLevel>=l) { DISPLAY(__VA_ARGS__); } /* 0 : no display; 1: errors; 2: default; 3: details; 4: debug */
80
ZDICT_clockSpan(clock_t nPrevious)81 static clock_t ZDICT_clockSpan(clock_t nPrevious) { return clock() - nPrevious; }
82
ZDICT_printHex(const void * ptr,size_t length)83 static void ZDICT_printHex(const void* ptr, size_t length)
84 {
85 const BYTE* const b = (const BYTE*)ptr;
86 size_t u;
87 for (u=0; u<length; u++) {
88 BYTE c = b[u];
89 if (c<32 || c>126) c = '.'; /* non-printable char */
90 DISPLAY("%c", c);
91 }
92 }
93
94
95 /*-********************************************************
96 * Helper functions
97 **********************************************************/
ZDICT_isError(size_t errorCode)98 unsigned ZDICT_isError(size_t errorCode) { return ERR_isError(errorCode); }
99
ZDICT_getErrorName(size_t errorCode)100 const char* ZDICT_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCode); }
101
ZDICT_getDictID(const void * dictBuffer,size_t dictSize)102 unsigned ZDICT_getDictID(const void* dictBuffer, size_t dictSize)
103 {
104 if (dictSize < 8) return 0;
105 if (MEM_readLE32(dictBuffer) != ZSTD_MAGIC_DICTIONARY) return 0;
106 return MEM_readLE32((const char*)dictBuffer + 4);
107 }
108
ZDICT_getDictHeaderSize(const void * dictBuffer,size_t dictSize)109 size_t ZDICT_getDictHeaderSize(const void* dictBuffer, size_t dictSize)
110 {
111 size_t headerSize;
112 if (dictSize <= 8 || MEM_readLE32(dictBuffer) != ZSTD_MAGIC_DICTIONARY) return ERROR(dictionary_corrupted);
113
114 { ZSTD_compressedBlockState_t* bs = (ZSTD_compressedBlockState_t*)malloc(sizeof(ZSTD_compressedBlockState_t));
115 U32* wksp = (U32*)malloc(HUF_WORKSPACE_SIZE);
116 if (!bs || !wksp) {
117 headerSize = ERROR(memory_allocation);
118 } else {
119 ZSTD_reset_compressedBlockState(bs);
120 headerSize = ZSTD_loadCEntropy(bs, wksp, dictBuffer, dictSize);
121 }
122
123 free(bs);
124 free(wksp);
125 }
126
127 return headerSize;
128 }
129
130 /*-********************************************************
131 * Dictionary training functions
132 **********************************************************/
ZDICT_NbCommonBytes(size_t val)133 static unsigned ZDICT_NbCommonBytes (size_t val)
134 {
135 if (MEM_isLittleEndian()) {
136 if (MEM_64bits()) {
137 # if defined(_MSC_VER) && defined(_WIN64)
138 unsigned long r = 0;
139 _BitScanForward64( &r, (U64)val );
140 return (unsigned)(r>>3);
141 # elif defined(__GNUC__) && (__GNUC__ >= 3)
142 return (__builtin_ctzll((U64)val) >> 3);
143 # else
144 static const int DeBruijnBytePos[64] = { 0, 0, 0, 0, 0, 1, 1, 2, 0, 3, 1, 3, 1, 4, 2, 7, 0, 2, 3, 6, 1, 5, 3, 5, 1, 3, 4, 4, 2, 5, 6, 7, 7, 0, 1, 2, 3, 3, 4, 6, 2, 6, 5, 5, 3, 4, 5, 6, 7, 1, 2, 4, 6, 4, 4, 5, 7, 2, 6, 5, 7, 6, 7, 7 };
145 return DeBruijnBytePos[((U64)((val & -(long long)val) * 0x0218A392CDABBD3FULL)) >> 58];
146 # endif
147 } else { /* 32 bits */
148 # if defined(_MSC_VER)
149 unsigned long r=0;
150 _BitScanForward( &r, (U32)val );
151 return (unsigned)(r>>3);
152 # elif defined(__GNUC__) && (__GNUC__ >= 3)
153 return (__builtin_ctz((U32)val) >> 3);
154 # else
155 static const int DeBruijnBytePos[32] = { 0, 0, 3, 0, 3, 1, 3, 0, 3, 2, 2, 1, 3, 2, 0, 1, 3, 3, 1, 2, 2, 2, 2, 0, 3, 1, 2, 0, 1, 0, 1, 1 };
156 return DeBruijnBytePos[((U32)((val & -(S32)val) * 0x077CB531U)) >> 27];
157 # endif
158 }
159 } else { /* Big Endian CPU */
160 if (MEM_64bits()) {
161 # if defined(_MSC_VER) && defined(_WIN64)
162 unsigned long r = 0;
163 _BitScanReverse64( &r, val );
164 return (unsigned)(r>>3);
165 # elif defined(__GNUC__) && (__GNUC__ >= 3)
166 return (__builtin_clzll(val) >> 3);
167 # else
168 unsigned r;
169 const unsigned n32 = sizeof(size_t)*4; /* calculate this way due to compiler complaining in 32-bits mode */
170 if (!(val>>n32)) { r=4; } else { r=0; val>>=n32; }
171 if (!(val>>16)) { r+=2; val>>=8; } else { val>>=24; }
172 r += (!val);
173 return r;
174 # endif
175 } else { /* 32 bits */
176 # if defined(_MSC_VER)
177 unsigned long r = 0;
178 _BitScanReverse( &r, (unsigned long)val );
179 return (unsigned)(r>>3);
180 # elif defined(__GNUC__) && (__GNUC__ >= 3)
181 return (__builtin_clz((U32)val) >> 3);
182 # else
183 unsigned r;
184 if (!(val>>16)) { r=2; val>>=8; } else { r=0; val>>=24; }
185 r += (!val);
186 return r;
187 # endif
188 } }
189 }
190
191
192 /*! ZDICT_count() :
193 Count the nb of common bytes between 2 pointers.
194 Note : this function presumes end of buffer followed by noisy guard band.
195 */
ZDICT_count(const void * pIn,const void * pMatch)196 static size_t ZDICT_count(const void* pIn, const void* pMatch)
197 {
198 const char* const pStart = (const char*)pIn;
199 for (;;) {
200 size_t const diff = MEM_readST(pMatch) ^ MEM_readST(pIn);
201 if (!diff) {
202 pIn = (const char*)pIn+sizeof(size_t);
203 pMatch = (const char*)pMatch+sizeof(size_t);
204 continue;
205 }
206 pIn = (const char*)pIn+ZDICT_NbCommonBytes(diff);
207 return (size_t)((const char*)pIn - pStart);
208 }
209 }
210
211
212 typedef struct {
213 U32 pos;
214 U32 length;
215 U32 savings;
216 } dictItem;
217
ZDICT_initDictItem(dictItem * d)218 static void ZDICT_initDictItem(dictItem* d)
219 {
220 d->pos = 1;
221 d->length = 0;
222 d->savings = (U32)(-1);
223 }
224
225
226 #define LLIMIT 64 /* heuristic determined experimentally */
227 #define MINMATCHLENGTH 7 /* heuristic determined experimentally */
ZDICT_analyzePos(BYTE * doneMarks,const int * suffix,U32 start,const void * buffer,U32 minRatio,U32 notificationLevel)228 static dictItem ZDICT_analyzePos(
229 BYTE* doneMarks,
230 const int* suffix, U32 start,
231 const void* buffer, U32 minRatio, U32 notificationLevel)
232 {
233 U32 lengthList[LLIMIT] = {0};
234 U32 cumulLength[LLIMIT] = {0};
235 U32 savings[LLIMIT] = {0};
236 const BYTE* b = (const BYTE*)buffer;
237 size_t maxLength = LLIMIT;
238 size_t pos = suffix[start];
239 U32 end = start;
240 dictItem solution;
241
242 /* init */
243 memset(&solution, 0, sizeof(solution));
244 doneMarks[pos] = 1;
245
246 /* trivial repetition cases */
247 if ( (MEM_read16(b+pos+0) == MEM_read16(b+pos+2))
248 ||(MEM_read16(b+pos+1) == MEM_read16(b+pos+3))
249 ||(MEM_read16(b+pos+2) == MEM_read16(b+pos+4)) ) {
250 /* skip and mark segment */
251 U16 const pattern16 = MEM_read16(b+pos+4);
252 U32 u, patternEnd = 6;
253 while (MEM_read16(b+pos+patternEnd) == pattern16) patternEnd+=2 ;
254 if (b[pos+patternEnd] == b[pos+patternEnd-1]) patternEnd++;
255 for (u=1; u<patternEnd; u++)
256 doneMarks[pos+u] = 1;
257 return solution;
258 }
259
260 /* look forward */
261 { size_t length;
262 do {
263 end++;
264 length = ZDICT_count(b + pos, b + suffix[end]);
265 } while (length >= MINMATCHLENGTH);
266 }
267
268 /* look backward */
269 { size_t length;
270 do {
271 length = ZDICT_count(b + pos, b + *(suffix+start-1));
272 if (length >=MINMATCHLENGTH) start--;
273 } while(length >= MINMATCHLENGTH);
274 }
275
276 /* exit if not found a minimum nb of repetitions */
277 if (end-start < minRatio) {
278 U32 idx;
279 for(idx=start; idx<end; idx++)
280 doneMarks[suffix[idx]] = 1;
281 return solution;
282 }
283
284 { int i;
285 U32 mml;
286 U32 refinedStart = start;
287 U32 refinedEnd = end;
288
289 DISPLAYLEVEL(4, "\n");
290 DISPLAYLEVEL(4, "found %3u matches of length >= %i at pos %7u ", (unsigned)(end-start), MINMATCHLENGTH, (unsigned)pos);
291 DISPLAYLEVEL(4, "\n");
292
293 for (mml = MINMATCHLENGTH ; ; mml++) {
294 BYTE currentChar = 0;
295 U32 currentCount = 0;
296 U32 currentID = refinedStart;
297 U32 id;
298 U32 selectedCount = 0;
299 U32 selectedID = currentID;
300 for (id =refinedStart; id < refinedEnd; id++) {
301 if (b[suffix[id] + mml] != currentChar) {
302 if (currentCount > selectedCount) {
303 selectedCount = currentCount;
304 selectedID = currentID;
305 }
306 currentID = id;
307 currentChar = b[ suffix[id] + mml];
308 currentCount = 0;
309 }
310 currentCount ++;
311 }
312 if (currentCount > selectedCount) { /* for last */
313 selectedCount = currentCount;
314 selectedID = currentID;
315 }
316
317 if (selectedCount < minRatio)
318 break;
319 refinedStart = selectedID;
320 refinedEnd = refinedStart + selectedCount;
321 }
322
323 /* evaluate gain based on new dict */
324 start = refinedStart;
325 pos = suffix[refinedStart];
326 end = start;
327 memset(lengthList, 0, sizeof(lengthList));
328
329 /* look forward */
330 { size_t length;
331 do {
332 end++;
333 length = ZDICT_count(b + pos, b + suffix[end]);
334 if (length >= LLIMIT) length = LLIMIT-1;
335 lengthList[length]++;
336 } while (length >=MINMATCHLENGTH);
337 }
338
339 /* look backward */
340 { size_t length = MINMATCHLENGTH;
341 while ((length >= MINMATCHLENGTH) & (start > 0)) {
342 length = ZDICT_count(b + pos, b + suffix[start - 1]);
343 if (length >= LLIMIT) length = LLIMIT - 1;
344 lengthList[length]++;
345 if (length >= MINMATCHLENGTH) start--;
346 }
347 }
348
349 /* largest useful length */
350 memset(cumulLength, 0, sizeof(cumulLength));
351 cumulLength[maxLength-1] = lengthList[maxLength-1];
352 for (i=(int)(maxLength-2); i>=0; i--)
353 cumulLength[i] = cumulLength[i+1] + lengthList[i];
354
355 for (i=LLIMIT-1; i>=MINMATCHLENGTH; i--) if (cumulLength[i]>=minRatio) break;
356 maxLength = i;
357
358 /* reduce maxLength in case of final into repetitive data */
359 { U32 l = (U32)maxLength;
360 BYTE const c = b[pos + maxLength-1];
361 while (b[pos+l-2]==c) l--;
362 maxLength = l;
363 }
364 if (maxLength < MINMATCHLENGTH) return solution; /* skip : no long-enough solution */
365
366 /* calculate savings */
367 savings[5] = 0;
368 for (i=MINMATCHLENGTH; i<=(int)maxLength; i++)
369 savings[i] = savings[i-1] + (lengthList[i] * (i-3));
370
371 DISPLAYLEVEL(4, "Selected dict at position %u, of length %u : saves %u (ratio: %.2f) \n",
372 (unsigned)pos, (unsigned)maxLength, (unsigned)savings[maxLength], (double)savings[maxLength] / maxLength);
373
374 solution.pos = (U32)pos;
375 solution.length = (U32)maxLength;
376 solution.savings = savings[maxLength];
377
378 /* mark positions done */
379 { U32 id;
380 for (id=start; id<end; id++) {
381 U32 p, pEnd, length;
382 U32 const testedPos = suffix[id];
383 if (testedPos == pos)
384 length = solution.length;
385 else {
386 length = (U32)ZDICT_count(b+pos, b+testedPos);
387 if (length > solution.length) length = solution.length;
388 }
389 pEnd = (U32)(testedPos + length);
390 for (p=testedPos; p<pEnd; p++)
391 doneMarks[p] = 1;
392 } } }
393
394 return solution;
395 }
396
397
isIncluded(const void * in,const void * container,size_t length)398 static int isIncluded(const void* in, const void* container, size_t length)
399 {
400 const char* const ip = (const char*) in;
401 const char* const into = (const char*) container;
402 size_t u;
403
404 for (u=0; u<length; u++) { /* works because end of buffer is a noisy guard band */
405 if (ip[u] != into[u]) break;
406 }
407
408 return u==length;
409 }
410
411 /*! ZDICT_tryMerge() :
412 check if dictItem can be merged, do it if possible
413 @return : id of destination elt, 0 if not merged
414 */
ZDICT_tryMerge(dictItem * table,dictItem elt,U32 eltNbToSkip,const void * buffer)415 static U32 ZDICT_tryMerge(dictItem* table, dictItem elt, U32 eltNbToSkip, const void* buffer)
416 {
417 const U32 tableSize = table->pos;
418 const U32 eltEnd = elt.pos + elt.length;
419 const char* const buf = (const char*) buffer;
420
421 /* tail overlap */
422 U32 u; for (u=1; u<tableSize; u++) {
423 if (u==eltNbToSkip) continue;
424 if ((table[u].pos > elt.pos) && (table[u].pos <= eltEnd)) { /* overlap, existing > new */
425 /* append */
426 U32 const addedLength = table[u].pos - elt.pos;
427 table[u].length += addedLength;
428 table[u].pos = elt.pos;
429 table[u].savings += elt.savings * addedLength / elt.length; /* rough approx */
430 table[u].savings += elt.length / 8; /* rough approx bonus */
431 elt = table[u];
432 /* sort : improve rank */
433 while ((u>1) && (table[u-1].savings < elt.savings))
434 table[u] = table[u-1], u--;
435 table[u] = elt;
436 return u;
437 } }
438
439 /* front overlap */
440 for (u=1; u<tableSize; u++) {
441 if (u==eltNbToSkip) continue;
442
443 if ((table[u].pos + table[u].length >= elt.pos) && (table[u].pos < elt.pos)) { /* overlap, existing < new */
444 /* append */
445 int const addedLength = (int)eltEnd - (table[u].pos + table[u].length);
446 table[u].savings += elt.length / 8; /* rough approx bonus */
447 if (addedLength > 0) { /* otherwise, elt fully included into existing */
448 table[u].length += addedLength;
449 table[u].savings += elt.savings * addedLength / elt.length; /* rough approx */
450 }
451 /* sort : improve rank */
452 elt = table[u];
453 while ((u>1) && (table[u-1].savings < elt.savings))
454 table[u] = table[u-1], u--;
455 table[u] = elt;
456 return u;
457 }
458
459 if (MEM_read64(buf + table[u].pos) == MEM_read64(buf + elt.pos + 1)) {
460 if (isIncluded(buf + table[u].pos, buf + elt.pos + 1, table[u].length)) {
461 size_t const addedLength = MAX( (int)elt.length - (int)table[u].length , 1 );
462 table[u].pos = elt.pos;
463 table[u].savings += (U32)(elt.savings * addedLength / elt.length);
464 table[u].length = MIN(elt.length, table[u].length + 1);
465 return u;
466 }
467 }
468 }
469
470 return 0;
471 }
472
473
ZDICT_removeDictItem(dictItem * table,U32 id)474 static void ZDICT_removeDictItem(dictItem* table, U32 id)
475 {
476 /* convention : table[0].pos stores nb of elts */
477 U32 const max = table[0].pos;
478 U32 u;
479 if (!id) return; /* protection, should never happen */
480 for (u=id; u<max-1; u++)
481 table[u] = table[u+1];
482 table->pos--;
483 }
484
485
ZDICT_insertDictItem(dictItem * table,U32 maxSize,dictItem elt,const void * buffer)486 static void ZDICT_insertDictItem(dictItem* table, U32 maxSize, dictItem elt, const void* buffer)
487 {
488 /* merge if possible */
489 U32 mergeId = ZDICT_tryMerge(table, elt, 0, buffer);
490 if (mergeId) {
491 U32 newMerge = 1;
492 while (newMerge) {
493 newMerge = ZDICT_tryMerge(table, table[mergeId], mergeId, buffer);
494 if (newMerge) ZDICT_removeDictItem(table, mergeId);
495 mergeId = newMerge;
496 }
497 return;
498 }
499
500 /* insert */
501 { U32 current;
502 U32 nextElt = table->pos;
503 if (nextElt >= maxSize) nextElt = maxSize-1;
504 current = nextElt-1;
505 while (table[current].savings < elt.savings) {
506 table[current+1] = table[current];
507 current--;
508 }
509 table[current+1] = elt;
510 table->pos = nextElt+1;
511 }
512 }
513
514
ZDICT_dictSize(const dictItem * dictList)515 static U32 ZDICT_dictSize(const dictItem* dictList)
516 {
517 U32 u, dictSize = 0;
518 for (u=1; u<dictList[0].pos; u++)
519 dictSize += dictList[u].length;
520 return dictSize;
521 }
522
523
ZDICT_trainBuffer_legacy(dictItem * dictList,U32 dictListSize,const void * const buffer,size_t bufferSize,const size_t * fileSizes,unsigned nbFiles,unsigned minRatio,U32 notificationLevel)524 static size_t ZDICT_trainBuffer_legacy(dictItem* dictList, U32 dictListSize,
525 const void* const buffer, size_t bufferSize, /* buffer must end with noisy guard band */
526 const size_t* fileSizes, unsigned nbFiles,
527 unsigned minRatio, U32 notificationLevel)
528 {
529 int* const suffix0 = (int*)malloc((bufferSize+2)*sizeof(*suffix0));
530 int* const suffix = suffix0+1;
531 U32* reverseSuffix = (U32*)malloc((bufferSize)*sizeof(*reverseSuffix));
532 BYTE* doneMarks = (BYTE*)malloc((bufferSize+16)*sizeof(*doneMarks)); /* +16 for overflow security */
533 U32* filePos = (U32*)malloc(nbFiles * sizeof(*filePos));
534 size_t result = 0;
535 clock_t displayClock = 0;
536 clock_t const refreshRate = CLOCKS_PER_SEC * 3 / 10;
537
538 # undef DISPLAYUPDATE
539 # define DISPLAYUPDATE(l, ...) if (notificationLevel>=l) { \
540 if (ZDICT_clockSpan(displayClock) > refreshRate) \
541 { displayClock = clock(); DISPLAY(__VA_ARGS__); \
542 if (notificationLevel>=4) fflush(stderr); } }
543
544 /* init */
545 DISPLAYLEVEL(2, "\r%70s\r", ""); /* clean display line */
546 if (!suffix0 || !reverseSuffix || !doneMarks || !filePos) {
547 result = ERROR(memory_allocation);
548 goto _cleanup;
549 }
550 if (minRatio < MINRATIO) minRatio = MINRATIO;
551 memset(doneMarks, 0, bufferSize+16);
552
553 /* limit sample set size (divsufsort limitation)*/
554 if (bufferSize > ZDICT_MAX_SAMPLES_SIZE) DISPLAYLEVEL(3, "sample set too large : reduced to %u MB ...\n", (unsigned)(ZDICT_MAX_SAMPLES_SIZE>>20));
555 while (bufferSize > ZDICT_MAX_SAMPLES_SIZE) bufferSize -= fileSizes[--nbFiles];
556
557 /* sort */
558 DISPLAYLEVEL(2, "sorting %u files of total size %u MB ...\n", nbFiles, (unsigned)(bufferSize>>20));
559 { int const divSuftSortResult = divsufsort((const unsigned char*)buffer, suffix, (int)bufferSize, 0);
560 if (divSuftSortResult != 0) { result = ERROR(GENERIC); goto _cleanup; }
561 }
562 suffix[bufferSize] = (int)bufferSize; /* leads into noise */
563 suffix0[0] = (int)bufferSize; /* leads into noise */
564 /* build reverse suffix sort */
565 { size_t pos;
566 for (pos=0; pos < bufferSize; pos++)
567 reverseSuffix[suffix[pos]] = (U32)pos;
568 /* note filePos tracks borders between samples.
569 It's not used at this stage, but planned to become useful in a later update */
570 filePos[0] = 0;
571 for (pos=1; pos<nbFiles; pos++)
572 filePos[pos] = (U32)(filePos[pos-1] + fileSizes[pos-1]);
573 }
574
575 DISPLAYLEVEL(2, "finding patterns ... \n");
576 DISPLAYLEVEL(3, "minimum ratio : %u \n", minRatio);
577
578 { U32 cursor; for (cursor=0; cursor < bufferSize; ) {
579 dictItem solution;
580 if (doneMarks[cursor]) { cursor++; continue; }
581 solution = ZDICT_analyzePos(doneMarks, suffix, reverseSuffix[cursor], buffer, minRatio, notificationLevel);
582 if (solution.length==0) { cursor++; continue; }
583 ZDICT_insertDictItem(dictList, dictListSize, solution, buffer);
584 cursor += solution.length;
585 DISPLAYUPDATE(2, "\r%4.2f %% \r", (double)cursor / bufferSize * 100);
586 } }
587
588 _cleanup:
589 free(suffix0);
590 free(reverseSuffix);
591 free(doneMarks);
592 free(filePos);
593 return result;
594 }
595
596
ZDICT_fillNoise(void * buffer,size_t length)597 static void ZDICT_fillNoise(void* buffer, size_t length)
598 {
599 unsigned const prime1 = 2654435761U;
600 unsigned const prime2 = 2246822519U;
601 unsigned acc = prime1;
602 size_t p=0;
603 for (p=0; p<length; p++) {
604 acc *= prime2;
605 ((unsigned char*)buffer)[p] = (unsigned char)(acc >> 21);
606 }
607 }
608
609
610 typedef struct
611 {
612 ZSTD_CDict* dict; /* dictionary */
613 ZSTD_CCtx* zc; /* working context */
614 void* workPlace; /* must be ZSTD_BLOCKSIZE_MAX allocated */
615 } EStats_ress_t;
616
617 #define MAXREPOFFSET 1024
618
ZDICT_countEStats(EStats_ress_t esr,const ZSTD_parameters * params,unsigned * countLit,unsigned * offsetcodeCount,unsigned * matchlengthCount,unsigned * litlengthCount,U32 * repOffsets,const void * src,size_t srcSize,U32 notificationLevel)619 static void ZDICT_countEStats(EStats_ress_t esr, const ZSTD_parameters* params,
620 unsigned* countLit, unsigned* offsetcodeCount, unsigned* matchlengthCount, unsigned* litlengthCount, U32* repOffsets,
621 const void* src, size_t srcSize,
622 U32 notificationLevel)
623 {
624 size_t const blockSizeMax = MIN (ZSTD_BLOCKSIZE_MAX, 1 << params->cParams.windowLog);
625 size_t cSize;
626
627 if (srcSize > blockSizeMax) srcSize = blockSizeMax; /* protection vs large samples */
628 { size_t const errorCode = ZSTD_compressBegin_usingCDict(esr.zc, esr.dict);
629 if (ZSTD_isError(errorCode)) { DISPLAYLEVEL(1, "warning : ZSTD_compressBegin_usingCDict failed \n"); return; }
630
631 }
632 cSize = ZSTD_compressBlock(esr.zc, esr.workPlace, ZSTD_BLOCKSIZE_MAX, src, srcSize);
633 if (ZSTD_isError(cSize)) { DISPLAYLEVEL(3, "warning : could not compress sample size %u \n", (unsigned)srcSize); return; }
634
635 if (cSize) { /* if == 0; block is not compressible */
636 const seqStore_t* const seqStorePtr = ZSTD_getSeqStore(esr.zc);
637
638 /* literals stats */
639 { const BYTE* bytePtr;
640 for(bytePtr = seqStorePtr->litStart; bytePtr < seqStorePtr->lit; bytePtr++)
641 countLit[*bytePtr]++;
642 }
643
644 /* seqStats */
645 { U32 const nbSeq = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
646 ZSTD_seqToCodes(seqStorePtr);
647
648 { const BYTE* codePtr = seqStorePtr->ofCode;
649 U32 u;
650 for (u=0; u<nbSeq; u++) offsetcodeCount[codePtr[u]]++;
651 }
652
653 { const BYTE* codePtr = seqStorePtr->mlCode;
654 U32 u;
655 for (u=0; u<nbSeq; u++) matchlengthCount[codePtr[u]]++;
656 }
657
658 { const BYTE* codePtr = seqStorePtr->llCode;
659 U32 u;
660 for (u=0; u<nbSeq; u++) litlengthCount[codePtr[u]]++;
661 }
662
663 if (nbSeq >= 2) { /* rep offsets */
664 const seqDef* const seq = seqStorePtr->sequencesStart;
665 U32 offset1 = seq[0].offset - 3;
666 U32 offset2 = seq[1].offset - 3;
667 if (offset1 >= MAXREPOFFSET) offset1 = 0;
668 if (offset2 >= MAXREPOFFSET) offset2 = 0;
669 repOffsets[offset1] += 3;
670 repOffsets[offset2] += 1;
671 } } }
672 }
673
ZDICT_totalSampleSize(const size_t * fileSizes,unsigned nbFiles)674 static size_t ZDICT_totalSampleSize(const size_t* fileSizes, unsigned nbFiles)
675 {
676 size_t total=0;
677 unsigned u;
678 for (u=0; u<nbFiles; u++) total += fileSizes[u];
679 return total;
680 }
681
682 typedef struct { U32 offset; U32 count; } offsetCount_t;
683
ZDICT_insertSortCount(offsetCount_t table[ZSTD_REP_NUM+1],U32 val,U32 count)684 static void ZDICT_insertSortCount(offsetCount_t table[ZSTD_REP_NUM+1], U32 val, U32 count)
685 {
686 U32 u;
687 table[ZSTD_REP_NUM].offset = val;
688 table[ZSTD_REP_NUM].count = count;
689 for (u=ZSTD_REP_NUM; u>0; u--) {
690 offsetCount_t tmp;
691 if (table[u-1].count >= table[u].count) break;
692 tmp = table[u-1];
693 table[u-1] = table[u];
694 table[u] = tmp;
695 }
696 }
697
698 /* ZDICT_flatLit() :
699 * rewrite `countLit` to contain a mostly flat but still compressible distribution of literals.
700 * necessary to avoid generating a non-compressible distribution that HUF_writeCTable() cannot encode.
701 */
ZDICT_flatLit(unsigned * countLit)702 static void ZDICT_flatLit(unsigned* countLit)
703 {
704 int u;
705 for (u=1; u<256; u++) countLit[u] = 2;
706 countLit[0] = 4;
707 countLit[253] = 1;
708 countLit[254] = 1;
709 }
710
711 #define OFFCODE_MAX 30 /* only applicable to first block */
ZDICT_analyzeEntropy(void * dstBuffer,size_t maxDstSize,int compressionLevel,const void * srcBuffer,const size_t * fileSizes,unsigned nbFiles,const void * dictBuffer,size_t dictBufferSize,unsigned notificationLevel)712 static size_t ZDICT_analyzeEntropy(void* dstBuffer, size_t maxDstSize,
713 int compressionLevel,
714 const void* srcBuffer, const size_t* fileSizes, unsigned nbFiles,
715 const void* dictBuffer, size_t dictBufferSize,
716 unsigned notificationLevel)
717 {
718 unsigned countLit[256];
719 HUF_CREATE_STATIC_CTABLE(hufTable, 255);
720 unsigned offcodeCount[OFFCODE_MAX+1];
721 short offcodeNCount[OFFCODE_MAX+1];
722 U32 offcodeMax = ZSTD_highbit32((U32)(dictBufferSize + 128 KB));
723 unsigned matchLengthCount[MaxML+1];
724 short matchLengthNCount[MaxML+1];
725 unsigned litLengthCount[MaxLL+1];
726 short litLengthNCount[MaxLL+1];
727 U32 repOffset[MAXREPOFFSET];
728 offsetCount_t bestRepOffset[ZSTD_REP_NUM+1];
729 EStats_ress_t esr = { NULL, NULL, NULL };
730 ZSTD_parameters params;
731 U32 u, huffLog = 11, Offlog = OffFSELog, mlLog = MLFSELog, llLog = LLFSELog, total;
732 size_t pos = 0, errorCode;
733 size_t eSize = 0;
734 size_t const totalSrcSize = ZDICT_totalSampleSize(fileSizes, nbFiles);
735 size_t const averageSampleSize = totalSrcSize / (nbFiles + !nbFiles);
736 BYTE* dstPtr = (BYTE*)dstBuffer;
737
738 /* init */
739 DEBUGLOG(4, "ZDICT_analyzeEntropy");
740 if (offcodeMax>OFFCODE_MAX) { eSize = ERROR(dictionaryCreation_failed); goto _cleanup; } /* too large dictionary */
741 for (u=0; u<256; u++) countLit[u] = 1; /* any character must be described */
742 for (u=0; u<=offcodeMax; u++) offcodeCount[u] = 1;
743 for (u=0; u<=MaxML; u++) matchLengthCount[u] = 1;
744 for (u=0; u<=MaxLL; u++) litLengthCount[u] = 1;
745 memset(repOffset, 0, sizeof(repOffset));
746 repOffset[1] = repOffset[4] = repOffset[8] = 1;
747 memset(bestRepOffset, 0, sizeof(bestRepOffset));
748 if (compressionLevel==0) compressionLevel = ZSTD_CLEVEL_DEFAULT;
749 params = ZSTD_getParams(compressionLevel, averageSampleSize, dictBufferSize);
750
751 esr.dict = ZSTD_createCDict_advanced(dictBuffer, dictBufferSize, ZSTD_dlm_byRef, ZSTD_dct_rawContent, params.cParams, ZSTD_defaultCMem);
752 esr.zc = ZSTD_createCCtx();
753 esr.workPlace = malloc(ZSTD_BLOCKSIZE_MAX);
754 if (!esr.dict || !esr.zc || !esr.workPlace) {
755 eSize = ERROR(memory_allocation);
756 DISPLAYLEVEL(1, "Not enough memory \n");
757 goto _cleanup;
758 }
759
760 /* collect stats on all samples */
761 for (u=0; u<nbFiles; u++) {
762 ZDICT_countEStats(esr, ¶ms,
763 countLit, offcodeCount, matchLengthCount, litLengthCount, repOffset,
764 (const char*)srcBuffer + pos, fileSizes[u],
765 notificationLevel);
766 pos += fileSizes[u];
767 }
768
769 /* analyze, build stats, starting with literals */
770 { size_t maxNbBits = HUF_buildCTable (hufTable, countLit, 255, huffLog);
771 if (HUF_isError(maxNbBits)) {
772 eSize = maxNbBits;
773 DISPLAYLEVEL(1, " HUF_buildCTable error \n");
774 goto _cleanup;
775 }
776 if (maxNbBits==8) { /* not compressible : will fail on HUF_writeCTable() */
777 DISPLAYLEVEL(2, "warning : pathological dataset : literals are not compressible : samples are noisy or too regular \n");
778 ZDICT_flatLit(countLit); /* replace distribution by a fake "mostly flat but still compressible" distribution, that HUF_writeCTable() can encode */
779 maxNbBits = HUF_buildCTable (hufTable, countLit, 255, huffLog);
780 assert(maxNbBits==9);
781 }
782 huffLog = (U32)maxNbBits;
783 }
784
785 /* looking for most common first offsets */
786 { U32 offset;
787 for (offset=1; offset<MAXREPOFFSET; offset++)
788 ZDICT_insertSortCount(bestRepOffset, offset, repOffset[offset]);
789 }
790 /* note : the result of this phase should be used to better appreciate the impact on statistics */
791
792 total=0; for (u=0; u<=offcodeMax; u++) total+=offcodeCount[u];
793 errorCode = FSE_normalizeCount(offcodeNCount, Offlog, offcodeCount, total, offcodeMax, /* useLowProbCount */ 1);
794 if (FSE_isError(errorCode)) {
795 eSize = errorCode;
796 DISPLAYLEVEL(1, "FSE_normalizeCount error with offcodeCount \n");
797 goto _cleanup;
798 }
799 Offlog = (U32)errorCode;
800
801 total=0; for (u=0; u<=MaxML; u++) total+=matchLengthCount[u];
802 errorCode = FSE_normalizeCount(matchLengthNCount, mlLog, matchLengthCount, total, MaxML, /* useLowProbCount */ 1);
803 if (FSE_isError(errorCode)) {
804 eSize = errorCode;
805 DISPLAYLEVEL(1, "FSE_normalizeCount error with matchLengthCount \n");
806 goto _cleanup;
807 }
808 mlLog = (U32)errorCode;
809
810 total=0; for (u=0; u<=MaxLL; u++) total+=litLengthCount[u];
811 errorCode = FSE_normalizeCount(litLengthNCount, llLog, litLengthCount, total, MaxLL, /* useLowProbCount */ 1);
812 if (FSE_isError(errorCode)) {
813 eSize = errorCode;
814 DISPLAYLEVEL(1, "FSE_normalizeCount error with litLengthCount \n");
815 goto _cleanup;
816 }
817 llLog = (U32)errorCode;
818
819 /* write result to buffer */
820 { size_t const hhSize = HUF_writeCTable(dstPtr, maxDstSize, hufTable, 255, huffLog);
821 if (HUF_isError(hhSize)) {
822 eSize = hhSize;
823 DISPLAYLEVEL(1, "HUF_writeCTable error \n");
824 goto _cleanup;
825 }
826 dstPtr += hhSize;
827 maxDstSize -= hhSize;
828 eSize += hhSize;
829 }
830
831 { size_t const ohSize = FSE_writeNCount(dstPtr, maxDstSize, offcodeNCount, OFFCODE_MAX, Offlog);
832 if (FSE_isError(ohSize)) {
833 eSize = ohSize;
834 DISPLAYLEVEL(1, "FSE_writeNCount error with offcodeNCount \n");
835 goto _cleanup;
836 }
837 dstPtr += ohSize;
838 maxDstSize -= ohSize;
839 eSize += ohSize;
840 }
841
842 { size_t const mhSize = FSE_writeNCount(dstPtr, maxDstSize, matchLengthNCount, MaxML, mlLog);
843 if (FSE_isError(mhSize)) {
844 eSize = mhSize;
845 DISPLAYLEVEL(1, "FSE_writeNCount error with matchLengthNCount \n");
846 goto _cleanup;
847 }
848 dstPtr += mhSize;
849 maxDstSize -= mhSize;
850 eSize += mhSize;
851 }
852
853 { size_t const lhSize = FSE_writeNCount(dstPtr, maxDstSize, litLengthNCount, MaxLL, llLog);
854 if (FSE_isError(lhSize)) {
855 eSize = lhSize;
856 DISPLAYLEVEL(1, "FSE_writeNCount error with litlengthNCount \n");
857 goto _cleanup;
858 }
859 dstPtr += lhSize;
860 maxDstSize -= lhSize;
861 eSize += lhSize;
862 }
863
864 if (maxDstSize<12) {
865 eSize = ERROR(dstSize_tooSmall);
866 DISPLAYLEVEL(1, "not enough space to write RepOffsets \n");
867 goto _cleanup;
868 }
869 # if 0
870 MEM_writeLE32(dstPtr+0, bestRepOffset[0].offset);
871 MEM_writeLE32(dstPtr+4, bestRepOffset[1].offset);
872 MEM_writeLE32(dstPtr+8, bestRepOffset[2].offset);
873 #else
874 /* at this stage, we don't use the result of "most common first offset",
875 as the impact of statistics is not properly evaluated */
876 MEM_writeLE32(dstPtr+0, repStartValue[0]);
877 MEM_writeLE32(dstPtr+4, repStartValue[1]);
878 MEM_writeLE32(dstPtr+8, repStartValue[2]);
879 #endif
880 eSize += 12;
881
882 _cleanup:
883 ZSTD_freeCDict(esr.dict);
884 ZSTD_freeCCtx(esr.zc);
885 free(esr.workPlace);
886
887 return eSize;
888 }
889
890
891
ZDICT_finalizeDictionary(void * dictBuffer,size_t dictBufferCapacity,const void * customDictContent,size_t dictContentSize,const void * samplesBuffer,const size_t * samplesSizes,unsigned nbSamples,ZDICT_params_t params)892 size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
893 const void* customDictContent, size_t dictContentSize,
894 const void* samplesBuffer, const size_t* samplesSizes,
895 unsigned nbSamples, ZDICT_params_t params)
896 {
897 size_t hSize;
898 #define HBUFFSIZE 256 /* should prove large enough for all entropy headers */
899 BYTE header[HBUFFSIZE];
900 int const compressionLevel = (params.compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : params.compressionLevel;
901 U32 const notificationLevel = params.notificationLevel;
902
903 /* check conditions */
904 DEBUGLOG(4, "ZDICT_finalizeDictionary");
905 if (dictBufferCapacity < dictContentSize) return ERROR(dstSize_tooSmall);
906 if (dictContentSize < ZDICT_CONTENTSIZE_MIN) return ERROR(srcSize_wrong);
907 if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) return ERROR(dstSize_tooSmall);
908
909 /* dictionary header */
910 MEM_writeLE32(header, ZSTD_MAGIC_DICTIONARY);
911 { U64 const randomID = XXH64(customDictContent, dictContentSize, 0);
912 U32 const compliantID = (randomID % ((1U<<31)-32768)) + 32768;
913 U32 const dictID = params.dictID ? params.dictID : compliantID;
914 MEM_writeLE32(header+4, dictID);
915 }
916 hSize = 8;
917
918 /* entropy tables */
919 DISPLAYLEVEL(2, "\r%70s\r", ""); /* clean display line */
920 DISPLAYLEVEL(2, "statistics ... \n");
921 { size_t const eSize = ZDICT_analyzeEntropy(header+hSize, HBUFFSIZE-hSize,
922 compressionLevel,
923 samplesBuffer, samplesSizes, nbSamples,
924 customDictContent, dictContentSize,
925 notificationLevel);
926 if (ZDICT_isError(eSize)) return eSize;
927 hSize += eSize;
928 }
929
930 /* copy elements in final buffer ; note : src and dst buffer can overlap */
931 if (hSize + dictContentSize > dictBufferCapacity) dictContentSize = dictBufferCapacity - hSize;
932 { size_t const dictSize = hSize + dictContentSize;
933 char* dictEnd = (char*)dictBuffer + dictSize;
934 memmove(dictEnd - dictContentSize, customDictContent, dictContentSize);
935 memcpy(dictBuffer, header, hSize);
936 return dictSize;
937 }
938 }
939
940
ZDICT_addEntropyTablesFromBuffer_advanced(void * dictBuffer,size_t dictContentSize,size_t dictBufferCapacity,const void * samplesBuffer,const size_t * samplesSizes,unsigned nbSamples,ZDICT_params_t params)941 static size_t ZDICT_addEntropyTablesFromBuffer_advanced(
942 void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity,
943 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
944 ZDICT_params_t params)
945 {
946 int const compressionLevel = (params.compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : params.compressionLevel;
947 U32 const notificationLevel = params.notificationLevel;
948 size_t hSize = 8;
949
950 /* calculate entropy tables */
951 DISPLAYLEVEL(2, "\r%70s\r", ""); /* clean display line */
952 DISPLAYLEVEL(2, "statistics ... \n");
953 { size_t const eSize = ZDICT_analyzeEntropy((char*)dictBuffer+hSize, dictBufferCapacity-hSize,
954 compressionLevel,
955 samplesBuffer, samplesSizes, nbSamples,
956 (char*)dictBuffer + dictBufferCapacity - dictContentSize, dictContentSize,
957 notificationLevel);
958 if (ZDICT_isError(eSize)) return eSize;
959 hSize += eSize;
960 }
961
962 /* add dictionary header (after entropy tables) */
963 MEM_writeLE32(dictBuffer, ZSTD_MAGIC_DICTIONARY);
964 { U64 const randomID = XXH64((char*)dictBuffer + dictBufferCapacity - dictContentSize, dictContentSize, 0);
965 U32 const compliantID = (randomID % ((1U<<31)-32768)) + 32768;
966 U32 const dictID = params.dictID ? params.dictID : compliantID;
967 MEM_writeLE32((char*)dictBuffer+4, dictID);
968 }
969
970 if (hSize + dictContentSize < dictBufferCapacity)
971 memmove((char*)dictBuffer + hSize, (char*)dictBuffer + dictBufferCapacity - dictContentSize, dictContentSize);
972 return MIN(dictBufferCapacity, hSize+dictContentSize);
973 }
974
975 /*! ZDICT_trainFromBuffer_unsafe_legacy() :
976 * Warning : `samplesBuffer` must be followed by noisy guard band !!!
977 * @return : size of dictionary, or an error code which can be tested with ZDICT_isError()
978 */
ZDICT_trainFromBuffer_unsafe_legacy(void * dictBuffer,size_t maxDictSize,const void * samplesBuffer,const size_t * samplesSizes,unsigned nbSamples,ZDICT_legacy_params_t params)979 static size_t ZDICT_trainFromBuffer_unsafe_legacy(
980 void* dictBuffer, size_t maxDictSize,
981 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
982 ZDICT_legacy_params_t params)
983 {
984 U32 const dictListSize = MAX(MAX(DICTLISTSIZE_DEFAULT, nbSamples), (U32)(maxDictSize/16));
985 dictItem* const dictList = (dictItem*)malloc(dictListSize * sizeof(*dictList));
986 unsigned const selectivity = params.selectivityLevel == 0 ? g_selectivity_default : params.selectivityLevel;
987 unsigned const minRep = (selectivity > 30) ? MINRATIO : nbSamples >> selectivity;
988 size_t const targetDictSize = maxDictSize;
989 size_t const samplesBuffSize = ZDICT_totalSampleSize(samplesSizes, nbSamples);
990 size_t dictSize = 0;
991 U32 const notificationLevel = params.zParams.notificationLevel;
992
993 /* checks */
994 if (!dictList) return ERROR(memory_allocation);
995 if (maxDictSize < ZDICT_DICTSIZE_MIN) { free(dictList); return ERROR(dstSize_tooSmall); } /* requested dictionary size is too small */
996 if (samplesBuffSize < ZDICT_MIN_SAMPLES_SIZE) { free(dictList); return ERROR(dictionaryCreation_failed); } /* not enough source to create dictionary */
997
998 /* init */
999 ZDICT_initDictItem(dictList);
1000
1001 /* build dictionary */
1002 ZDICT_trainBuffer_legacy(dictList, dictListSize,
1003 samplesBuffer, samplesBuffSize,
1004 samplesSizes, nbSamples,
1005 minRep, notificationLevel);
1006
1007 /* display best matches */
1008 if (params.zParams.notificationLevel>= 3) {
1009 unsigned const nb = MIN(25, dictList[0].pos);
1010 unsigned const dictContentSize = ZDICT_dictSize(dictList);
1011 unsigned u;
1012 DISPLAYLEVEL(3, "\n %u segments found, of total size %u \n", (unsigned)dictList[0].pos-1, dictContentSize);
1013 DISPLAYLEVEL(3, "list %u best segments \n", nb-1);
1014 for (u=1; u<nb; u++) {
1015 unsigned const pos = dictList[u].pos;
1016 unsigned const length = dictList[u].length;
1017 U32 const printedLength = MIN(40, length);
1018 if ((pos > samplesBuffSize) || ((pos + length) > samplesBuffSize)) {
1019 free(dictList);
1020 return ERROR(GENERIC); /* should never happen */
1021 }
1022 DISPLAYLEVEL(3, "%3u:%3u bytes at pos %8u, savings %7u bytes |",
1023 u, length, pos, (unsigned)dictList[u].savings);
1024 ZDICT_printHex((const char*)samplesBuffer+pos, printedLength);
1025 DISPLAYLEVEL(3, "| \n");
1026 } }
1027
1028
1029 /* create dictionary */
1030 { unsigned dictContentSize = ZDICT_dictSize(dictList);
1031 if (dictContentSize < ZDICT_CONTENTSIZE_MIN) { free(dictList); return ERROR(dictionaryCreation_failed); } /* dictionary content too small */
1032 if (dictContentSize < targetDictSize/4) {
1033 DISPLAYLEVEL(2, "! warning : selected content significantly smaller than requested (%u < %u) \n", dictContentSize, (unsigned)maxDictSize);
1034 if (samplesBuffSize < 10 * targetDictSize)
1035 DISPLAYLEVEL(2, "! consider increasing the number of samples (total size : %u MB)\n", (unsigned)(samplesBuffSize>>20));
1036 if (minRep > MINRATIO) {
1037 DISPLAYLEVEL(2, "! consider increasing selectivity to produce larger dictionary (-s%u) \n", selectivity+1);
1038 DISPLAYLEVEL(2, "! note : larger dictionaries are not necessarily better, test its efficiency on samples \n");
1039 }
1040 }
1041
1042 if ((dictContentSize > targetDictSize*3) && (nbSamples > 2*MINRATIO) && (selectivity>1)) {
1043 unsigned proposedSelectivity = selectivity-1;
1044 while ((nbSamples >> proposedSelectivity) <= MINRATIO) { proposedSelectivity--; }
1045 DISPLAYLEVEL(2, "! note : calculated dictionary significantly larger than requested (%u > %u) \n", dictContentSize, (unsigned)maxDictSize);
1046 DISPLAYLEVEL(2, "! consider increasing dictionary size, or produce denser dictionary (-s%u) \n", proposedSelectivity);
1047 DISPLAYLEVEL(2, "! always test dictionary efficiency on real samples \n");
1048 }
1049
1050 /* limit dictionary size */
1051 { U32 const max = dictList->pos; /* convention : nb of useful elts within dictList */
1052 U32 currentSize = 0;
1053 U32 n; for (n=1; n<max; n++) {
1054 currentSize += dictList[n].length;
1055 if (currentSize > targetDictSize) { currentSize -= dictList[n].length; break; }
1056 }
1057 dictList->pos = n;
1058 dictContentSize = currentSize;
1059 }
1060
1061 /* build dict content */
1062 { U32 u;
1063 BYTE* ptr = (BYTE*)dictBuffer + maxDictSize;
1064 for (u=1; u<dictList->pos; u++) {
1065 U32 l = dictList[u].length;
1066 ptr -= l;
1067 if (ptr<(BYTE*)dictBuffer) { free(dictList); return ERROR(GENERIC); } /* should not happen */
1068 memcpy(ptr, (const char*)samplesBuffer+dictList[u].pos, l);
1069 } }
1070
1071 dictSize = ZDICT_addEntropyTablesFromBuffer_advanced(dictBuffer, dictContentSize, maxDictSize,
1072 samplesBuffer, samplesSizes, nbSamples,
1073 params.zParams);
1074 }
1075
1076 /* clean up */
1077 free(dictList);
1078 return dictSize;
1079 }
1080
1081
1082 /* ZDICT_trainFromBuffer_legacy() :
1083 * issue : samplesBuffer need to be followed by a noisy guard band.
1084 * work around : duplicate the buffer, and add the noise */
ZDICT_trainFromBuffer_legacy(void * dictBuffer,size_t dictBufferCapacity,const void * samplesBuffer,const size_t * samplesSizes,unsigned nbSamples,ZDICT_legacy_params_t params)1085 size_t ZDICT_trainFromBuffer_legacy(void* dictBuffer, size_t dictBufferCapacity,
1086 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
1087 ZDICT_legacy_params_t params)
1088 {
1089 size_t result;
1090 void* newBuff;
1091 size_t const sBuffSize = ZDICT_totalSampleSize(samplesSizes, nbSamples);
1092 if (sBuffSize < ZDICT_MIN_SAMPLES_SIZE) return 0; /* not enough content => no dictionary */
1093
1094 newBuff = malloc(sBuffSize + NOISELENGTH);
1095 if (!newBuff) return ERROR(memory_allocation);
1096
1097 memcpy(newBuff, samplesBuffer, sBuffSize);
1098 ZDICT_fillNoise((char*)newBuff + sBuffSize, NOISELENGTH); /* guard band, for end of buffer condition */
1099
1100 result =
1101 ZDICT_trainFromBuffer_unsafe_legacy(dictBuffer, dictBufferCapacity, newBuff,
1102 samplesSizes, nbSamples, params);
1103 free(newBuff);
1104 return result;
1105 }
1106
1107
ZDICT_trainFromBuffer(void * dictBuffer,size_t dictBufferCapacity,const void * samplesBuffer,const size_t * samplesSizes,unsigned nbSamples)1108 size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCapacity,
1109 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples)
1110 {
1111 ZDICT_fastCover_params_t params;
1112 DEBUGLOG(3, "ZDICT_trainFromBuffer");
1113 memset(¶ms, 0, sizeof(params));
1114 params.d = 8;
1115 params.steps = 4;
1116 /* Use default level since no compression level information is available */
1117 params.zParams.compressionLevel = ZSTD_CLEVEL_DEFAULT;
1118 #if defined(DEBUGLEVEL) && (DEBUGLEVEL>=1)
1119 params.zParams.notificationLevel = DEBUGLEVEL;
1120 #endif
1121 return ZDICT_optimizeTrainFromBuffer_fastCover(dictBuffer, dictBufferCapacity,
1122 samplesBuffer, samplesSizes, nbSamples,
1123 ¶ms);
1124 }
1125
ZDICT_addEntropyTablesFromBuffer(void * dictBuffer,size_t dictContentSize,size_t dictBufferCapacity,const void * samplesBuffer,const size_t * samplesSizes,unsigned nbSamples)1126 size_t ZDICT_addEntropyTablesFromBuffer(void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity,
1127 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples)
1128 {
1129 ZDICT_params_t params;
1130 memset(¶ms, 0, sizeof(params));
1131 return ZDICT_addEntropyTablesFromBuffer_advanced(dictBuffer, dictContentSize, dictBufferCapacity,
1132 samplesBuffer, samplesSizes, nbSamples,
1133 params);
1134 }
1135