ce188d4d |
1 | /* LzFindMt.c -- multithreaded Match finder for LZ algorithms\r |
2 | 2015-10-15 : Igor Pavlov : Public domain */\r |
3 | \r |
4 | #include "Precomp.h"\r |
5 | \r |
6 | #include "LzHash.h"\r |
7 | \r |
8 | #include "LzFindMt.h"\r |
9 | \r |
10 | static void MtSync_Construct(CMtSync *p)\r |
11 | {\r |
12 | p->wasCreated = False;\r |
13 | p->csWasInitialized = False;\r |
14 | p->csWasEntered = False;\r |
15 | Thread_Construct(&p->thread);\r |
16 | Event_Construct(&p->canStart);\r |
17 | Event_Construct(&p->wasStarted);\r |
18 | Event_Construct(&p->wasStopped);\r |
19 | Semaphore_Construct(&p->freeSemaphore);\r |
20 | Semaphore_Construct(&p->filledSemaphore);\r |
21 | }\r |
22 | \r |
23 | static void MtSync_GetNextBlock(CMtSync *p)\r |
24 | {\r |
25 | if (p->needStart)\r |
26 | {\r |
27 | p->numProcessedBlocks = 1;\r |
28 | p->needStart = False;\r |
29 | p->stopWriting = False;\r |
30 | p->exit = False;\r |
31 | Event_Reset(&p->wasStarted);\r |
32 | Event_Reset(&p->wasStopped);\r |
33 | \r |
34 | Event_Set(&p->canStart);\r |
35 | Event_Wait(&p->wasStarted);\r |
36 | }\r |
37 | else\r |
38 | {\r |
39 | CriticalSection_Leave(&p->cs);\r |
40 | p->csWasEntered = False;\r |
41 | p->numProcessedBlocks++;\r |
42 | Semaphore_Release1(&p->freeSemaphore);\r |
43 | }\r |
44 | Semaphore_Wait(&p->filledSemaphore);\r |
45 | CriticalSection_Enter(&p->cs);\r |
46 | p->csWasEntered = True;\r |
47 | }\r |
48 | \r |
49 | /* MtSync_StopWriting must be called if Writing was started */\r |
50 | \r |
51 | static void MtSync_StopWriting(CMtSync *p)\r |
52 | {\r |
53 | UInt32 myNumBlocks = p->numProcessedBlocks;\r |
54 | if (!Thread_WasCreated(&p->thread) || p->needStart)\r |
55 | return;\r |
56 | p->stopWriting = True;\r |
57 | if (p->csWasEntered)\r |
58 | {\r |
59 | CriticalSection_Leave(&p->cs);\r |
60 | p->csWasEntered = False;\r |
61 | }\r |
62 | Semaphore_Release1(&p->freeSemaphore);\r |
63 | \r |
64 | Event_Wait(&p->wasStopped);\r |
65 | \r |
66 | while (myNumBlocks++ != p->numProcessedBlocks)\r |
67 | {\r |
68 | Semaphore_Wait(&p->filledSemaphore);\r |
69 | Semaphore_Release1(&p->freeSemaphore);\r |
70 | }\r |
71 | p->needStart = True;\r |
72 | }\r |
73 | \r |
74 | static void MtSync_Destruct(CMtSync *p)\r |
75 | {\r |
76 | if (Thread_WasCreated(&p->thread))\r |
77 | {\r |
78 | MtSync_StopWriting(p);\r |
79 | p->exit = True;\r |
80 | if (p->needStart)\r |
81 | Event_Set(&p->canStart);\r |
82 | Thread_Wait(&p->thread);\r |
83 | Thread_Close(&p->thread);\r |
84 | }\r |
85 | if (p->csWasInitialized)\r |
86 | {\r |
87 | CriticalSection_Delete(&p->cs);\r |
88 | p->csWasInitialized = False;\r |
89 | }\r |
90 | \r |
91 | Event_Close(&p->canStart);\r |
92 | Event_Close(&p->wasStarted);\r |
93 | Event_Close(&p->wasStopped);\r |
94 | Semaphore_Close(&p->freeSemaphore);\r |
95 | Semaphore_Close(&p->filledSemaphore);\r |
96 | \r |
97 | p->wasCreated = False;\r |
98 | }\r |
99 | \r |
100 | #define RINOK_THREAD(x) { if ((x) != 0) return SZ_ERROR_THREAD; }\r |
101 | \r |
102 | static SRes MtSync_Create2(CMtSync *p, THREAD_FUNC_TYPE startAddress, void *obj, UInt32 numBlocks)\r |
103 | {\r |
104 | if (p->wasCreated)\r |
105 | return SZ_OK;\r |
106 | \r |
107 | RINOK_THREAD(CriticalSection_Init(&p->cs));\r |
108 | p->csWasInitialized = True;\r |
109 | \r |
110 | RINOK_THREAD(AutoResetEvent_CreateNotSignaled(&p->canStart));\r |
111 | RINOK_THREAD(AutoResetEvent_CreateNotSignaled(&p->wasStarted));\r |
112 | RINOK_THREAD(AutoResetEvent_CreateNotSignaled(&p->wasStopped));\r |
113 | \r |
114 | RINOK_THREAD(Semaphore_Create(&p->freeSemaphore, numBlocks, numBlocks));\r |
115 | RINOK_THREAD(Semaphore_Create(&p->filledSemaphore, 0, numBlocks));\r |
116 | \r |
117 | p->needStart = True;\r |
118 | \r |
119 | RINOK_THREAD(Thread_Create(&p->thread, startAddress, obj));\r |
120 | p->wasCreated = True;\r |
121 | return SZ_OK;\r |
122 | }\r |
123 | \r |
124 | static SRes MtSync_Create(CMtSync *p, THREAD_FUNC_TYPE startAddress, void *obj, UInt32 numBlocks)\r |
125 | {\r |
126 | SRes res = MtSync_Create2(p, startAddress, obj, numBlocks);\r |
127 | if (res != SZ_OK)\r |
128 | MtSync_Destruct(p);\r |
129 | return res;\r |
130 | }\r |
131 | \r |
132 | void MtSync_Init(CMtSync *p) { p->needStart = True; }\r |
133 | \r |
134 | #define kMtMaxValForNormalize 0xFFFFFFFF\r |
135 | \r |
136 | #define DEF_GetHeads2(name, v, action) \\r |
137 | static void GetHeads ## name(const Byte *p, UInt32 pos, \\r |
138 | UInt32 *hash, UInt32 hashMask, UInt32 *heads, UInt32 numHeads, const UInt32 *crc) \\r |
139 | { action; for (; numHeads != 0; numHeads--) { \\r |
140 | const UInt32 value = (v); p++; *heads++ = pos - hash[value]; hash[value] = pos++; } }\r |
141 | \r |
142 | #define DEF_GetHeads(name, v) DEF_GetHeads2(name, v, ;)\r |
143 | \r |
144 | DEF_GetHeads2(2, (p[0] | ((UInt32)p[1] << 8)), UNUSED_VAR(hashMask); UNUSED_VAR(crc); )\r |
145 | DEF_GetHeads(3, (crc[p[0]] ^ p[1] ^ ((UInt32)p[2] << 8)) & hashMask)\r |
146 | DEF_GetHeads(4, (crc[p[0]] ^ p[1] ^ ((UInt32)p[2] << 8) ^ (crc[p[3]] << 5)) & hashMask)\r |
147 | DEF_GetHeads(4b, (crc[p[0]] ^ p[1] ^ ((UInt32)p[2] << 8) ^ ((UInt32)p[3] << 16)) & hashMask)\r |
148 | /* DEF_GetHeads(5, (crc[p[0]] ^ p[1] ^ ((UInt32)p[2] << 8) ^ (crc[p[3]] << 5) ^ (crc[p[4]] << 3)) & hashMask) */\r |
149 | \r |
150 | static void HashThreadFunc(CMatchFinderMt *mt)\r |
151 | {\r |
152 | CMtSync *p = &mt->hashSync;\r |
153 | for (;;)\r |
154 | {\r |
155 | UInt32 numProcessedBlocks = 0;\r |
156 | Event_Wait(&p->canStart);\r |
157 | Event_Set(&p->wasStarted);\r |
158 | for (;;)\r |
159 | {\r |
160 | if (p->exit)\r |
161 | return;\r |
162 | if (p->stopWriting)\r |
163 | {\r |
164 | p->numProcessedBlocks = numProcessedBlocks;\r |
165 | Event_Set(&p->wasStopped);\r |
166 | break;\r |
167 | }\r |
168 | \r |
169 | {\r |
170 | CMatchFinder *mf = mt->MatchFinder;\r |
171 | if (MatchFinder_NeedMove(mf))\r |
172 | {\r |
173 | CriticalSection_Enter(&mt->btSync.cs);\r |
174 | CriticalSection_Enter(&mt->hashSync.cs);\r |
175 | {\r |
176 | const Byte *beforePtr = Inline_MatchFinder_GetPointerToCurrentPos(mf);\r |
177 | ptrdiff_t offset;\r |
178 | MatchFinder_MoveBlock(mf);\r |
179 | offset = beforePtr - Inline_MatchFinder_GetPointerToCurrentPos(mf);\r |
180 | mt->pointerToCurPos -= offset;\r |
181 | mt->buffer -= offset;\r |
182 | }\r |
183 | CriticalSection_Leave(&mt->btSync.cs);\r |
184 | CriticalSection_Leave(&mt->hashSync.cs);\r |
185 | continue;\r |
186 | }\r |
187 | \r |
188 | Semaphore_Wait(&p->freeSemaphore);\r |
189 | \r |
190 | MatchFinder_ReadIfRequired(mf);\r |
191 | if (mf->pos > (kMtMaxValForNormalize - kMtHashBlockSize))\r |
192 | {\r |
193 | UInt32 subValue = (mf->pos - mf->historySize - 1);\r |
194 | MatchFinder_ReduceOffsets(mf, subValue);\r |
195 | MatchFinder_Normalize3(subValue, mf->hash + mf->fixedHashSize, (size_t)mf->hashMask + 1);\r |
196 | }\r |
197 | {\r |
198 | UInt32 *heads = mt->hashBuf + ((numProcessedBlocks++) & kMtHashNumBlocksMask) * kMtHashBlockSize;\r |
199 | UInt32 num = mf->streamPos - mf->pos;\r |
200 | heads[0] = 2;\r |
201 | heads[1] = num;\r |
202 | if (num >= mf->numHashBytes)\r |
203 | {\r |
204 | num = num - mf->numHashBytes + 1;\r |
205 | if (num > kMtHashBlockSize - 2)\r |
206 | num = kMtHashBlockSize - 2;\r |
207 | mt->GetHeadsFunc(mf->buffer, mf->pos, mf->hash + mf->fixedHashSize, mf->hashMask, heads + 2, num, mf->crc);\r |
208 | heads[0] += num;\r |
209 | }\r |
210 | mf->pos += num;\r |
211 | mf->buffer += num;\r |
212 | }\r |
213 | }\r |
214 | \r |
215 | Semaphore_Release1(&p->filledSemaphore);\r |
216 | }\r |
217 | }\r |
218 | }\r |
219 | \r |
220 | static void MatchFinderMt_GetNextBlock_Hash(CMatchFinderMt *p)\r |
221 | {\r |
222 | MtSync_GetNextBlock(&p->hashSync);\r |
223 | p->hashBufPosLimit = p->hashBufPos = ((p->hashSync.numProcessedBlocks - 1) & kMtHashNumBlocksMask) * kMtHashBlockSize;\r |
224 | p->hashBufPosLimit += p->hashBuf[p->hashBufPos++];\r |
225 | p->hashNumAvail = p->hashBuf[p->hashBufPos++];\r |
226 | }\r |
227 | \r |
228 | #define kEmptyHashValue 0\r |
229 | \r |
230 | /* #define MFMT_GM_INLINE */\r |
231 | \r |
232 | #ifdef MFMT_GM_INLINE\r |
233 | \r |
234 | #define NO_INLINE MY_FAST_CALL\r |
235 | \r |
236 | static Int32 NO_INLINE GetMatchesSpecN(UInt32 lenLimit, UInt32 pos, const Byte *cur, CLzRef *son,\r |
237 | UInt32 _cyclicBufferPos, UInt32 _cyclicBufferSize, UInt32 _cutValue,\r |
238 | UInt32 *_distances, UInt32 _maxLen, const UInt32 *hash, Int32 limit, UInt32 size, UInt32 *posRes)\r |
239 | {\r |
240 | do\r |
241 | {\r |
242 | UInt32 *distances = _distances + 1;\r |
243 | UInt32 curMatch = pos - *hash++;\r |
244 | \r |
245 | CLzRef *ptr0 = son + (_cyclicBufferPos << 1) + 1;\r |
246 | CLzRef *ptr1 = son + (_cyclicBufferPos << 1);\r |
247 | UInt32 len0 = 0, len1 = 0;\r |
248 | UInt32 cutValue = _cutValue;\r |
249 | UInt32 maxLen = _maxLen;\r |
250 | for (;;)\r |
251 | {\r |
252 | UInt32 delta = pos - curMatch;\r |
253 | if (cutValue-- == 0 || delta >= _cyclicBufferSize)\r |
254 | {\r |
255 | *ptr0 = *ptr1 = kEmptyHashValue;\r |
256 | break;\r |
257 | }\r |
258 | {\r |
259 | CLzRef *pair = son + ((_cyclicBufferPos - delta + ((delta > _cyclicBufferPos) ? _cyclicBufferSize : 0)) << 1);\r |
260 | const Byte *pb = cur - delta;\r |
261 | UInt32 len = (len0 < len1 ? len0 : len1);\r |
262 | if (pb[len] == cur[len])\r |
263 | {\r |
264 | if (++len != lenLimit && pb[len] == cur[len])\r |
265 | while (++len != lenLimit)\r |
266 | if (pb[len] != cur[len])\r |
267 | break;\r |
268 | if (maxLen < len)\r |
269 | {\r |
270 | *distances++ = maxLen = len;\r |
271 | *distances++ = delta - 1;\r |
272 | if (len == lenLimit)\r |
273 | {\r |
274 | *ptr1 = pair[0];\r |
275 | *ptr0 = pair[1];\r |
276 | break;\r |
277 | }\r |
278 | }\r |
279 | }\r |
280 | if (pb[len] < cur[len])\r |
281 | {\r |
282 | *ptr1 = curMatch;\r |
283 | ptr1 = pair + 1;\r |
284 | curMatch = *ptr1;\r |
285 | len1 = len;\r |
286 | }\r |
287 | else\r |
288 | {\r |
289 | *ptr0 = curMatch;\r |
290 | ptr0 = pair;\r |
291 | curMatch = *ptr0;\r |
292 | len0 = len;\r |
293 | }\r |
294 | }\r |
295 | }\r |
296 | pos++;\r |
297 | _cyclicBufferPos++;\r |
298 | cur++;\r |
299 | {\r |
300 | UInt32 num = (UInt32)(distances - _distances);\r |
301 | *_distances = num - 1;\r |
302 | _distances += num;\r |
303 | limit -= num;\r |
304 | }\r |
305 | }\r |
306 | while (limit > 0 && --size != 0);\r |
307 | *posRes = pos;\r |
308 | return limit;\r |
309 | }\r |
310 | \r |
311 | #endif\r |
312 | \r |
313 | static void BtGetMatches(CMatchFinderMt *p, UInt32 *distances)\r |
314 | {\r |
315 | UInt32 numProcessed = 0;\r |
316 | UInt32 curPos = 2;\r |
317 | UInt32 limit = kMtBtBlockSize - (p->matchMaxLen * 2);\r |
318 | \r |
319 | distances[1] = p->hashNumAvail;\r |
320 | \r |
321 | while (curPos < limit)\r |
322 | {\r |
323 | if (p->hashBufPos == p->hashBufPosLimit)\r |
324 | {\r |
325 | MatchFinderMt_GetNextBlock_Hash(p);\r |
326 | distances[1] = numProcessed + p->hashNumAvail;\r |
327 | if (p->hashNumAvail >= p->numHashBytes)\r |
328 | continue;\r |
329 | distances[0] = curPos + p->hashNumAvail;\r |
330 | distances += curPos;\r |
331 | for (; p->hashNumAvail != 0; p->hashNumAvail--)\r |
332 | *distances++ = 0;\r |
333 | return;\r |
334 | }\r |
335 | {\r |
336 | UInt32 size = p->hashBufPosLimit - p->hashBufPos;\r |
337 | UInt32 lenLimit = p->matchMaxLen;\r |
338 | UInt32 pos = p->pos;\r |
339 | UInt32 cyclicBufferPos = p->cyclicBufferPos;\r |
340 | if (lenLimit >= p->hashNumAvail)\r |
341 | lenLimit = p->hashNumAvail;\r |
342 | {\r |
343 | UInt32 size2 = p->hashNumAvail - lenLimit + 1;\r |
344 | if (size2 < size)\r |
345 | size = size2;\r |
346 | size2 = p->cyclicBufferSize - cyclicBufferPos;\r |
347 | if (size2 < size)\r |
348 | size = size2;\r |
349 | }\r |
350 | \r |
351 | #ifndef MFMT_GM_INLINE\r |
352 | while (curPos < limit && size-- != 0)\r |
353 | {\r |
354 | UInt32 *startDistances = distances + curPos;\r |
355 | UInt32 num = (UInt32)(GetMatchesSpec1(lenLimit, pos - p->hashBuf[p->hashBufPos++],\r |
356 | pos, p->buffer, p->son, cyclicBufferPos, p->cyclicBufferSize, p->cutValue,\r |
357 | startDistances + 1, p->numHashBytes - 1) - startDistances);\r |
358 | *startDistances = num - 1;\r |
359 | curPos += num;\r |
360 | cyclicBufferPos++;\r |
361 | pos++;\r |
362 | p->buffer++;\r |
363 | }\r |
364 | #else\r |
365 | {\r |
366 | UInt32 posRes;\r |
367 | curPos = limit - GetMatchesSpecN(lenLimit, pos, p->buffer, p->son, cyclicBufferPos, p->cyclicBufferSize, p->cutValue,\r |
368 | distances + curPos, p->numHashBytes - 1, p->hashBuf + p->hashBufPos, (Int32)(limit - curPos), size, &posRes);\r |
369 | p->hashBufPos += posRes - pos;\r |
370 | cyclicBufferPos += posRes - pos;\r |
371 | p->buffer += posRes - pos;\r |
372 | pos = posRes;\r |
373 | }\r |
374 | #endif\r |
375 | \r |
376 | numProcessed += pos - p->pos;\r |
377 | p->hashNumAvail -= pos - p->pos;\r |
378 | p->pos = pos;\r |
379 | if (cyclicBufferPos == p->cyclicBufferSize)\r |
380 | cyclicBufferPos = 0;\r |
381 | p->cyclicBufferPos = cyclicBufferPos;\r |
382 | }\r |
383 | }\r |
384 | \r |
385 | distances[0] = curPos;\r |
386 | }\r |
387 | \r |
388 | static void BtFillBlock(CMatchFinderMt *p, UInt32 globalBlockIndex)\r |
389 | {\r |
390 | CMtSync *sync = &p->hashSync;\r |
391 | if (!sync->needStart)\r |
392 | {\r |
393 | CriticalSection_Enter(&sync->cs);\r |
394 | sync->csWasEntered = True;\r |
395 | }\r |
396 | \r |
397 | BtGetMatches(p, p->btBuf + (globalBlockIndex & kMtBtNumBlocksMask) * kMtBtBlockSize);\r |
398 | \r |
399 | if (p->pos > kMtMaxValForNormalize - kMtBtBlockSize)\r |
400 | {\r |
401 | UInt32 subValue = p->pos - p->cyclicBufferSize;\r |
402 | MatchFinder_Normalize3(subValue, p->son, (size_t)p->cyclicBufferSize * 2);\r |
403 | p->pos -= subValue;\r |
404 | }\r |
405 | \r |
406 | if (!sync->needStart)\r |
407 | {\r |
408 | CriticalSection_Leave(&sync->cs);\r |
409 | sync->csWasEntered = False;\r |
410 | }\r |
411 | }\r |
412 | \r |
413 | void BtThreadFunc(CMatchFinderMt *mt)\r |
414 | {\r |
415 | CMtSync *p = &mt->btSync;\r |
416 | for (;;)\r |
417 | {\r |
418 | UInt32 blockIndex = 0;\r |
419 | Event_Wait(&p->canStart);\r |
420 | Event_Set(&p->wasStarted);\r |
421 | for (;;)\r |
422 | {\r |
423 | if (p->exit)\r |
424 | return;\r |
425 | if (p->stopWriting)\r |
426 | {\r |
427 | p->numProcessedBlocks = blockIndex;\r |
428 | MtSync_StopWriting(&mt->hashSync);\r |
429 | Event_Set(&p->wasStopped);\r |
430 | break;\r |
431 | }\r |
432 | Semaphore_Wait(&p->freeSemaphore);\r |
433 | BtFillBlock(mt, blockIndex++);\r |
434 | Semaphore_Release1(&p->filledSemaphore);\r |
435 | }\r |
436 | }\r |
437 | }\r |
438 | \r |
439 | void MatchFinderMt_Construct(CMatchFinderMt *p)\r |
440 | {\r |
441 | p->hashBuf = NULL;\r |
442 | MtSync_Construct(&p->hashSync);\r |
443 | MtSync_Construct(&p->btSync);\r |
444 | }\r |
445 | \r |
446 | static void MatchFinderMt_FreeMem(CMatchFinderMt *p, ISzAlloc *alloc)\r |
447 | {\r |
448 | alloc->Free(alloc, p->hashBuf);\r |
449 | p->hashBuf = NULL;\r |
450 | }\r |
451 | \r |
452 | void MatchFinderMt_Destruct(CMatchFinderMt *p, ISzAlloc *alloc)\r |
453 | {\r |
454 | MtSync_Destruct(&p->hashSync);\r |
455 | MtSync_Destruct(&p->btSync);\r |
456 | MatchFinderMt_FreeMem(p, alloc);\r |
457 | }\r |
458 | \r |
459 | #define kHashBufferSize (kMtHashBlockSize * kMtHashNumBlocks)\r |
460 | #define kBtBufferSize (kMtBtBlockSize * kMtBtNumBlocks)\r |
461 | \r |
462 | static THREAD_FUNC_RET_TYPE THREAD_FUNC_CALL_TYPE HashThreadFunc2(void *p) { HashThreadFunc((CMatchFinderMt *)p); return 0; }\r |
463 | static THREAD_FUNC_RET_TYPE THREAD_FUNC_CALL_TYPE BtThreadFunc2(void *p)\r |
464 | {\r |
465 | Byte allocaDummy[0x180];\r |
466 | unsigned i = 0;\r |
467 | for (i = 0; i < 16; i++)\r |
468 | allocaDummy[i] = (Byte)0;\r |
469 | if (allocaDummy[0] == 0)\r |
470 | BtThreadFunc((CMatchFinderMt *)p);\r |
471 | return 0;\r |
472 | }\r |
473 | \r |
474 | SRes MatchFinderMt_Create(CMatchFinderMt *p, UInt32 historySize, UInt32 keepAddBufferBefore,\r |
475 | UInt32 matchMaxLen, UInt32 keepAddBufferAfter, ISzAlloc *alloc)\r |
476 | {\r |
477 | CMatchFinder *mf = p->MatchFinder;\r |
478 | p->historySize = historySize;\r |
479 | if (kMtBtBlockSize <= matchMaxLen * 4)\r |
480 | return SZ_ERROR_PARAM;\r |
481 | if (!p->hashBuf)\r |
482 | {\r |
483 | p->hashBuf = (UInt32 *)alloc->Alloc(alloc, (kHashBufferSize + kBtBufferSize) * sizeof(UInt32));\r |
484 | if (!p->hashBuf)\r |
485 | return SZ_ERROR_MEM;\r |
486 | p->btBuf = p->hashBuf + kHashBufferSize;\r |
487 | }\r |
488 | keepAddBufferBefore += (kHashBufferSize + kBtBufferSize);\r |
489 | keepAddBufferAfter += kMtHashBlockSize;\r |
490 | if (!MatchFinder_Create(mf, historySize, keepAddBufferBefore, matchMaxLen, keepAddBufferAfter, alloc))\r |
491 | return SZ_ERROR_MEM;\r |
492 | \r |
493 | RINOK(MtSync_Create(&p->hashSync, HashThreadFunc2, p, kMtHashNumBlocks));\r |
494 | RINOK(MtSync_Create(&p->btSync, BtThreadFunc2, p, kMtBtNumBlocks));\r |
495 | return SZ_OK;\r |
496 | }\r |
497 | \r |
498 | /* Call it after ReleaseStream / SetStream */\r |
499 | void MatchFinderMt_Init(CMatchFinderMt *p)\r |
500 | {\r |
501 | CMatchFinder *mf = p->MatchFinder;\r |
502 | p->btBufPos = p->btBufPosLimit = 0;\r |
503 | p->hashBufPos = p->hashBufPosLimit = 0;\r |
504 | \r |
505 | /* Init without data reading. We don't want to read data in this thread */\r |
506 | MatchFinder_Init_2(mf, False);\r |
507 | \r |
508 | p->pointerToCurPos = Inline_MatchFinder_GetPointerToCurrentPos(mf);\r |
509 | p->btNumAvailBytes = 0;\r |
510 | p->lzPos = p->historySize + 1;\r |
511 | \r |
512 | p->hash = mf->hash;\r |
513 | p->fixedHashSize = mf->fixedHashSize;\r |
514 | p->crc = mf->crc;\r |
515 | \r |
516 | p->son = mf->son;\r |
517 | p->matchMaxLen = mf->matchMaxLen;\r |
518 | p->numHashBytes = mf->numHashBytes;\r |
519 | p->pos = mf->pos;\r |
520 | p->buffer = mf->buffer;\r |
521 | p->cyclicBufferPos = mf->cyclicBufferPos;\r |
522 | p->cyclicBufferSize = mf->cyclicBufferSize;\r |
523 | p->cutValue = mf->cutValue;\r |
524 | }\r |
525 | \r |
526 | /* ReleaseStream is required to finish multithreading */\r |
527 | void MatchFinderMt_ReleaseStream(CMatchFinderMt *p)\r |
528 | {\r |
529 | MtSync_StopWriting(&p->btSync);\r |
530 | /* p->MatchFinder->ReleaseStream(); */\r |
531 | }\r |
532 | \r |
533 | static void MatchFinderMt_Normalize(CMatchFinderMt *p)\r |
534 | {\r |
535 | MatchFinder_Normalize3(p->lzPos - p->historySize - 1, p->hash, p->fixedHashSize);\r |
536 | p->lzPos = p->historySize + 1;\r |
537 | }\r |
538 | \r |
539 | static void MatchFinderMt_GetNextBlock_Bt(CMatchFinderMt *p)\r |
540 | {\r |
541 | UInt32 blockIndex;\r |
542 | MtSync_GetNextBlock(&p->btSync);\r |
543 | blockIndex = ((p->btSync.numProcessedBlocks - 1) & kMtBtNumBlocksMask);\r |
544 | p->btBufPosLimit = p->btBufPos = blockIndex * kMtBtBlockSize;\r |
545 | p->btBufPosLimit += p->btBuf[p->btBufPos++];\r |
546 | p->btNumAvailBytes = p->btBuf[p->btBufPos++];\r |
547 | if (p->lzPos >= kMtMaxValForNormalize - kMtBtBlockSize)\r |
548 | MatchFinderMt_Normalize(p);\r |
549 | }\r |
550 | \r |
551 | static const Byte * MatchFinderMt_GetPointerToCurrentPos(CMatchFinderMt *p)\r |
552 | {\r |
553 | return p->pointerToCurPos;\r |
554 | }\r |
555 | \r |
556 | #define GET_NEXT_BLOCK_IF_REQUIRED if (p->btBufPos == p->btBufPosLimit) MatchFinderMt_GetNextBlock_Bt(p);\r |
557 | \r |
558 | static UInt32 MatchFinderMt_GetNumAvailableBytes(CMatchFinderMt *p)\r |
559 | {\r |
560 | GET_NEXT_BLOCK_IF_REQUIRED;\r |
561 | return p->btNumAvailBytes;\r |
562 | }\r |
563 | \r |
564 | static UInt32 * MixMatches2(CMatchFinderMt *p, UInt32 matchMinPos, UInt32 *distances)\r |
565 | {\r |
566 | UInt32 h2, curMatch2;\r |
567 | UInt32 *hash = p->hash;\r |
568 | const Byte *cur = p->pointerToCurPos;\r |
569 | UInt32 lzPos = p->lzPos;\r |
570 | MT_HASH2_CALC\r |
571 | \r |
572 | curMatch2 = hash[h2];\r |
573 | hash[h2] = lzPos;\r |
574 | \r |
575 | if (curMatch2 >= matchMinPos)\r |
576 | if (cur[(ptrdiff_t)curMatch2 - lzPos] == cur[0])\r |
577 | {\r |
578 | *distances++ = 2;\r |
579 | *distances++ = lzPos - curMatch2 - 1;\r |
580 | }\r |
581 | \r |
582 | return distances;\r |
583 | }\r |
584 | \r |
585 | static UInt32 * MixMatches3(CMatchFinderMt *p, UInt32 matchMinPos, UInt32 *distances)\r |
586 | {\r |
587 | UInt32 h2, h3, curMatch2, curMatch3;\r |
588 | UInt32 *hash = p->hash;\r |
589 | const Byte *cur = p->pointerToCurPos;\r |
590 | UInt32 lzPos = p->lzPos;\r |
591 | MT_HASH3_CALC\r |
592 | \r |
593 | curMatch2 = hash[ h2];\r |
594 | curMatch3 = hash[kFix3HashSize + h3];\r |
595 | \r |
596 | hash[ h2] = lzPos;\r |
597 | hash[kFix3HashSize + h3] = lzPos;\r |
598 | \r |
599 | if (curMatch2 >= matchMinPos && cur[(ptrdiff_t)curMatch2 - lzPos] == cur[0])\r |
600 | {\r |
601 | distances[1] = lzPos - curMatch2 - 1;\r |
602 | if (cur[(ptrdiff_t)curMatch2 - lzPos + 2] == cur[2])\r |
603 | {\r |
604 | distances[0] = 3;\r |
605 | return distances + 2;\r |
606 | }\r |
607 | distances[0] = 2;\r |
608 | distances += 2;\r |
609 | }\r |
610 | \r |
611 | if (curMatch3 >= matchMinPos && cur[(ptrdiff_t)curMatch3 - lzPos] == cur[0])\r |
612 | {\r |
613 | *distances++ = 3;\r |
614 | *distances++ = lzPos - curMatch3 - 1;\r |
615 | }\r |
616 | \r |
617 | return distances;\r |
618 | }\r |
619 | \r |
620 | /*\r |
621 | static UInt32 *MixMatches4(CMatchFinderMt *p, UInt32 matchMinPos, UInt32 *distances)\r |
622 | {\r |
623 | UInt32 h2, h3, h4, curMatch2, curMatch3, curMatch4;\r |
624 | UInt32 *hash = p->hash;\r |
625 | const Byte *cur = p->pointerToCurPos;\r |
626 | UInt32 lzPos = p->lzPos;\r |
627 | MT_HASH4_CALC\r |
628 | \r |
629 | curMatch2 = hash[ h2];\r |
630 | curMatch3 = hash[kFix3HashSize + h3];\r |
631 | curMatch4 = hash[kFix4HashSize + h4];\r |
632 | \r |
633 | hash[ h2] = lzPos;\r |
634 | hash[kFix3HashSize + h3] = lzPos;\r |
635 | hash[kFix4HashSize + h4] = lzPos;\r |
636 | \r |
637 | if (curMatch2 >= matchMinPos && cur[(ptrdiff_t)curMatch2 - lzPos] == cur[0])\r |
638 | {\r |
639 | distances[1] = lzPos - curMatch2 - 1;\r |
640 | if (cur[(ptrdiff_t)curMatch2 - lzPos + 2] == cur[2])\r |
641 | {\r |
642 | distances[0] = (cur[(ptrdiff_t)curMatch2 - lzPos + 3] == cur[3]) ? 4 : 3;\r |
643 | return distances + 2;\r |
644 | }\r |
645 | distances[0] = 2;\r |
646 | distances += 2;\r |
647 | }\r |
648 | \r |
649 | if (curMatch3 >= matchMinPos && cur[(ptrdiff_t)curMatch3 - lzPos] == cur[0])\r |
650 | {\r |
651 | distances[1] = lzPos - curMatch3 - 1;\r |
652 | if (cur[(ptrdiff_t)curMatch3 - lzPos + 3] == cur[3])\r |
653 | {\r |
654 | distances[0] = 4;\r |
655 | return distances + 2;\r |
656 | }\r |
657 | distances[0] = 3;\r |
658 | distances += 2;\r |
659 | }\r |
660 | \r |
661 | if (curMatch4 >= matchMinPos)\r |
662 | if (\r |
663 | cur[(ptrdiff_t)curMatch4 - lzPos] == cur[0] &&\r |
664 | cur[(ptrdiff_t)curMatch4 - lzPos + 3] == cur[3]\r |
665 | )\r |
666 | {\r |
667 | *distances++ = 4;\r |
668 | *distances++ = lzPos - curMatch4 - 1;\r |
669 | }\r |
670 | \r |
671 | return distances;\r |
672 | }\r |
673 | */\r |
674 | \r |
675 | #define INCREASE_LZ_POS p->lzPos++; p->pointerToCurPos++;\r |
676 | \r |
677 | static UInt32 MatchFinderMt2_GetMatches(CMatchFinderMt *p, UInt32 *distances)\r |
678 | {\r |
679 | const UInt32 *btBuf = p->btBuf + p->btBufPos;\r |
680 | UInt32 len = *btBuf++;\r |
681 | p->btBufPos += 1 + len;\r |
682 | p->btNumAvailBytes--;\r |
683 | {\r |
684 | UInt32 i;\r |
685 | for (i = 0; i < len; i += 2)\r |
686 | {\r |
687 | *distances++ = *btBuf++;\r |
688 | *distances++ = *btBuf++;\r |
689 | }\r |
690 | }\r |
691 | INCREASE_LZ_POS\r |
692 | return len;\r |
693 | }\r |
694 | \r |
695 | static UInt32 MatchFinderMt_GetMatches(CMatchFinderMt *p, UInt32 *distances)\r |
696 | {\r |
697 | const UInt32 *btBuf = p->btBuf + p->btBufPos;\r |
698 | UInt32 len = *btBuf++;\r |
699 | p->btBufPos += 1 + len;\r |
700 | \r |
701 | if (len == 0)\r |
702 | {\r |
703 | /* change for bt5 ! */\r |
704 | if (p->btNumAvailBytes-- >= 4)\r |
705 | len = (UInt32)(p->MixMatchesFunc(p, p->lzPos - p->historySize, distances) - (distances));\r |
706 | }\r |
707 | else\r |
708 | {\r |
709 | /* Condition: there are matches in btBuf with length < p->numHashBytes */\r |
710 | UInt32 *distances2;\r |
711 | p->btNumAvailBytes--;\r |
712 | distances2 = p->MixMatchesFunc(p, p->lzPos - btBuf[1], distances);\r |
713 | do\r |
714 | {\r |
715 | *distances2++ = *btBuf++;\r |
716 | *distances2++ = *btBuf++;\r |
717 | }\r |
718 | while ((len -= 2) != 0);\r |
719 | len = (UInt32)(distances2 - (distances));\r |
720 | }\r |
721 | INCREASE_LZ_POS\r |
722 | return len;\r |
723 | }\r |
724 | \r |
725 | #define SKIP_HEADER2_MT do { GET_NEXT_BLOCK_IF_REQUIRED\r |
726 | #define SKIP_HEADER_MT(n) SKIP_HEADER2_MT if (p->btNumAvailBytes-- >= (n)) { const Byte *cur = p->pointerToCurPos; UInt32 *hash = p->hash;\r |
727 | #define SKIP_FOOTER_MT } INCREASE_LZ_POS p->btBufPos += p->btBuf[p->btBufPos] + 1; } while (--num != 0);\r |
728 | \r |
729 | static void MatchFinderMt0_Skip(CMatchFinderMt *p, UInt32 num)\r |
730 | {\r |
731 | SKIP_HEADER2_MT { p->btNumAvailBytes--;\r |
732 | SKIP_FOOTER_MT\r |
733 | }\r |
734 | \r |
735 | static void MatchFinderMt2_Skip(CMatchFinderMt *p, UInt32 num)\r |
736 | {\r |
737 | SKIP_HEADER_MT(2)\r |
738 | UInt32 h2;\r |
739 | MT_HASH2_CALC\r |
740 | hash[h2] = p->lzPos;\r |
741 | SKIP_FOOTER_MT\r |
742 | }\r |
743 | \r |
744 | static void MatchFinderMt3_Skip(CMatchFinderMt *p, UInt32 num)\r |
745 | {\r |
746 | SKIP_HEADER_MT(3)\r |
747 | UInt32 h2, h3;\r |
748 | MT_HASH3_CALC\r |
749 | hash[kFix3HashSize + h3] =\r |
750 | hash[ h2] =\r |
751 | p->lzPos;\r |
752 | SKIP_FOOTER_MT\r |
753 | }\r |
754 | \r |
755 | /*\r |
756 | static void MatchFinderMt4_Skip(CMatchFinderMt *p, UInt32 num)\r |
757 | {\r |
758 | SKIP_HEADER_MT(4)\r |
759 | UInt32 h2, h3, h4;\r |
760 | MT_HASH4_CALC\r |
761 | hash[kFix4HashSize + h4] =\r |
762 | hash[kFix3HashSize + h3] =\r |
763 | hash[ h2] =\r |
764 | p->lzPos;\r |
765 | SKIP_FOOTER_MT\r |
766 | }\r |
767 | */\r |
768 | \r |
769 | void MatchFinderMt_CreateVTable(CMatchFinderMt *p, IMatchFinder *vTable)\r |
770 | {\r |
771 | vTable->Init = (Mf_Init_Func)MatchFinderMt_Init;\r |
772 | vTable->GetNumAvailableBytes = (Mf_GetNumAvailableBytes_Func)MatchFinderMt_GetNumAvailableBytes;\r |
773 | vTable->GetPointerToCurrentPos = (Mf_GetPointerToCurrentPos_Func)MatchFinderMt_GetPointerToCurrentPos;\r |
774 | vTable->GetMatches = (Mf_GetMatches_Func)MatchFinderMt_GetMatches;\r |
775 | \r |
776 | switch (p->MatchFinder->numHashBytes)\r |
777 | {\r |
778 | case 2:\r |
779 | p->GetHeadsFunc = GetHeads2;\r |
780 | p->MixMatchesFunc = (Mf_Mix_Matches)0;\r |
781 | vTable->Skip = (Mf_Skip_Func)MatchFinderMt0_Skip;\r |
782 | vTable->GetMatches = (Mf_GetMatches_Func)MatchFinderMt2_GetMatches;\r |
783 | break;\r |
784 | case 3:\r |
785 | p->GetHeadsFunc = GetHeads3;\r |
786 | p->MixMatchesFunc = (Mf_Mix_Matches)MixMatches2;\r |
787 | vTable->Skip = (Mf_Skip_Func)MatchFinderMt2_Skip;\r |
788 | break;\r |
789 | default:\r |
790 | /* case 4: */\r |
791 | p->GetHeadsFunc = p->MatchFinder->bigHash ? GetHeads4b : GetHeads4;\r |
792 | p->MixMatchesFunc = (Mf_Mix_Matches)MixMatches3;\r |
793 | vTable->Skip = (Mf_Skip_Func)MatchFinderMt3_Skip;\r |
794 | break;\r |
795 | /*\r |
796 | default:\r |
797 | p->GetHeadsFunc = GetHeads5;\r |
798 | p->MixMatchesFunc = (Mf_Mix_Matches)MixMatches4;\r |
799 | vTable->Skip = (Mf_Skip_Func)MatchFinderMt4_Skip;\r |
800 | break;\r |
801 | */\r |
802 | }\r |
803 | }\r |