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