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 * Dependencies 13 ***************************************/ 14 #include "zstd_compress_sequences.h" 15 16 /* 17 * -log2(x / 256) lookup table for x in [0, 256). 18 * If x == 0: Return 0 19 * Else: Return floor(-log2(x / 256) * 256) 20 */ 21 static unsigned const kInverseProbabilityLog256[256] = { 22 0, 2048, 1792, 1642, 1536, 1453, 1386, 1329, 1280, 1236, 1197, 1162, 23 1130, 1100, 1073, 1047, 1024, 1001, 980, 960, 941, 923, 906, 889, 24 874, 859, 844, 830, 817, 804, 791, 779, 768, 756, 745, 734, 25 724, 714, 704, 694, 685, 676, 667, 658, 650, 642, 633, 626, 26 618, 610, 603, 595, 588, 581, 574, 567, 561, 554, 548, 542, 27 535, 529, 523, 517, 512, 506, 500, 495, 489, 484, 478, 473, 28 468, 463, 458, 453, 448, 443, 438, 434, 429, 424, 420, 415, 29 411, 407, 402, 398, 394, 390, 386, 382, 377, 373, 370, 366, 30 362, 358, 354, 350, 347, 343, 339, 336, 332, 329, 325, 322, 31 318, 315, 311, 308, 305, 302, 298, 295, 292, 289, 286, 282, 32 279, 276, 273, 270, 267, 264, 261, 258, 256, 253, 250, 247, 33 244, 241, 239, 236, 233, 230, 228, 225, 222, 220, 217, 215, 34 212, 209, 207, 204, 202, 199, 197, 194, 192, 190, 187, 185, 35 182, 180, 178, 175, 173, 171, 168, 166, 164, 162, 159, 157, 36 155, 153, 151, 149, 146, 144, 142, 140, 138, 136, 134, 132, 37 130, 128, 126, 123, 121, 119, 117, 115, 114, 112, 110, 108, 38 106, 104, 102, 100, 98, 96, 94, 93, 91, 89, 87, 85, 39 83, 82, 80, 78, 76, 74, 73, 71, 69, 67, 66, 64, 40 62, 61, 59, 57, 55, 54, 52, 50, 49, 47, 46, 44, 41 42, 41, 39, 37, 36, 34, 33, 31, 30, 28, 26, 25, 42 23, 22, 20, 19, 17, 16, 14, 13, 11, 10, 8, 7, 43 5, 4, 2, 1, 44 }; 45 46 static unsigned ZSTD_getFSEMaxSymbolValue(FSE_CTable const* ctable) { 47 void const* ptr = ctable; 48 U16 const* u16ptr = (U16 const*)ptr; 49 U32 const maxSymbolValue = MEM_read16(u16ptr + 1); 50 return maxSymbolValue; 51 } 52 53 /* 54 * Returns true if we should use ncount=-1 else we should 55 * use ncount=1 for low probability symbols instead. 56 */ 57 static unsigned ZSTD_useLowProbCount(size_t const nbSeq) 58 { 59 /* Heuristic: This should cover most blocks <= 16K and 60 * start to fade out after 16K to about 32K depending on 61 * comprssibility. 62 */ 63 return nbSeq >= 2048; 64 } 65 66 /* 67 * Returns the cost in bytes of encoding the normalized count header. 68 * Returns an error if any of the helper functions return an error. 69 */ 70 static size_t ZSTD_NCountCost(unsigned const* count, unsigned const max, 71 size_t const nbSeq, unsigned const FSELog) 72 { 73 BYTE wksp[FSE_NCOUNTBOUND]; 74 S16 norm[MaxSeq + 1]; 75 const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max); 76 FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq, max, ZSTD_useLowProbCount(nbSeq)), ""); 77 return FSE_writeNCount(wksp, sizeof(wksp), norm, max, tableLog); 78 } 79 80 /* 81 * Returns the cost in bits of encoding the distribution described by count 82 * using the entropy bound. 83 */ 84 static size_t ZSTD_entropyCost(unsigned const* count, unsigned const max, size_t const total) 85 { 86 unsigned cost = 0; 87 unsigned s; 88 for (s = 0; s <= max; ++s) { 89 unsigned norm = (unsigned)((256 * count[s]) / total); 90 if (count[s] != 0 && norm == 0) 91 norm = 1; 92 assert(count[s] < total); 93 cost += count[s] * kInverseProbabilityLog256[norm]; 94 } 95 return cost >> 8; 96 } 97 98 /* 99 * Returns the cost in bits of encoding the distribution in count using ctable. 100 * Returns an error if ctable cannot represent all the symbols in count. 101 */ 102 size_t ZSTD_fseBitCost( 103 FSE_CTable const* ctable, 104 unsigned const* count, 105 unsigned const max) 106 { 107 unsigned const kAccuracyLog = 8; 108 size_t cost = 0; 109 unsigned s; 110 FSE_CState_t cstate; 111 FSE_initCState(&cstate, ctable); 112 if (ZSTD_getFSEMaxSymbolValue(ctable) < max) { 113 DEBUGLOG(5, "Repeat FSE_CTable has maxSymbolValue %u < %u", 114 ZSTD_getFSEMaxSymbolValue(ctable), max); 115 return ERROR(GENERIC); 116 } 117 for (s = 0; s <= max; ++s) { 118 unsigned const tableLog = cstate.stateLog; 119 unsigned const badCost = (tableLog + 1) << kAccuracyLog; 120 unsigned const bitCost = FSE_bitCost(cstate.symbolTT, tableLog, s, kAccuracyLog); 121 if (count[s] == 0) 122 continue; 123 if (bitCost >= badCost) { 124 DEBUGLOG(5, "Repeat FSE_CTable has Prob[%u] == 0", s); 125 return ERROR(GENERIC); 126 } 127 cost += (size_t)count[s] * bitCost; 128 } 129 return cost >> kAccuracyLog; 130 } 131 132 /* 133 * Returns the cost in bits of encoding the distribution in count using the 134 * table described by norm. The max symbol support by norm is assumed >= max. 135 * norm must be valid for every symbol with non-zero probability in count. 136 */ 137 size_t ZSTD_crossEntropyCost(short const* norm, unsigned accuracyLog, 138 unsigned const* count, unsigned const max) 139 { 140 unsigned const shift = 8 - accuracyLog; 141 size_t cost = 0; 142 unsigned s; 143 assert(accuracyLog <= 8); 144 for (s = 0; s <= max; ++s) { 145 unsigned const normAcc = (norm[s] != -1) ? (unsigned)norm[s] : 1; 146 unsigned const norm256 = normAcc << shift; 147 assert(norm256 > 0); 148 assert(norm256 < 256); 149 cost += count[s] * kInverseProbabilityLog256[norm256]; 150 } 151 return cost >> 8; 152 } 153 154 symbolEncodingType_e 155 ZSTD_selectEncodingType( 156 FSE_repeat* repeatMode, unsigned const* count, unsigned const max, 157 size_t const mostFrequent, size_t nbSeq, unsigned const FSELog, 158 FSE_CTable const* prevCTable, 159 short const* defaultNorm, U32 defaultNormLog, 160 ZSTD_defaultPolicy_e const isDefaultAllowed, 161 ZSTD_strategy const strategy) 162 { 163 ZSTD_STATIC_ASSERT(ZSTD_defaultDisallowed == 0 && ZSTD_defaultAllowed != 0); 164 if (mostFrequent == nbSeq) { 165 *repeatMode = FSE_repeat_none; 166 if (isDefaultAllowed && nbSeq <= 2) { 167 /* Prefer set_basic over set_rle when there are 2 or less symbols, 168 * since RLE uses 1 byte, but set_basic uses 5-6 bits per symbol. 169 * If basic encoding isn't possible, always choose RLE. 170 */ 171 DEBUGLOG(5, "Selected set_basic"); 172 return set_basic; 173 } 174 DEBUGLOG(5, "Selected set_rle"); 175 return set_rle; 176 } 177 if (strategy < ZSTD_lazy) { 178 if (isDefaultAllowed) { 179 size_t const staticFse_nbSeq_max = 1000; 180 size_t const mult = 10 - strategy; 181 size_t const baseLog = 3; 182 size_t const dynamicFse_nbSeq_min = (((size_t)1 << defaultNormLog) * mult) >> baseLog; /* 28-36 for offset, 56-72 for lengths */ 183 assert(defaultNormLog >= 5 && defaultNormLog <= 6); /* xx_DEFAULTNORMLOG */ 184 assert(mult <= 9 && mult >= 7); 185 if ( (*repeatMode == FSE_repeat_valid) 186 && (nbSeq < staticFse_nbSeq_max) ) { 187 DEBUGLOG(5, "Selected set_repeat"); 188 return set_repeat; 189 } 190 if ( (nbSeq < dynamicFse_nbSeq_min) 191 || (mostFrequent < (nbSeq >> (defaultNormLog-1))) ) { 192 DEBUGLOG(5, "Selected set_basic"); 193 /* The format allows default tables to be repeated, but it isn't useful. 194 * When using simple heuristics to select encoding type, we don't want 195 * to confuse these tables with dictionaries. When running more careful 196 * analysis, we don't need to waste time checking both repeating tables 197 * and default tables. 198 */ 199 *repeatMode = FSE_repeat_none; 200 return set_basic; 201 } 202 } 203 } else { 204 size_t const basicCost = isDefaultAllowed ? ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, count, max) : ERROR(GENERIC); 205 size_t const repeatCost = *repeatMode != FSE_repeat_none ? ZSTD_fseBitCost(prevCTable, count, max) : ERROR(GENERIC); 206 size_t const NCountCost = ZSTD_NCountCost(count, max, nbSeq, FSELog); 207 size_t const compressedCost = (NCountCost << 3) + ZSTD_entropyCost(count, max, nbSeq); 208 209 if (isDefaultAllowed) { 210 assert(!ZSTD_isError(basicCost)); 211 assert(!(*repeatMode == FSE_repeat_valid && ZSTD_isError(repeatCost))); 212 } 213 assert(!ZSTD_isError(NCountCost)); 214 assert(compressedCost < ERROR(maxCode)); 215 DEBUGLOG(5, "Estimated bit costs: basic=%u\trepeat=%u\tcompressed=%u", 216 (unsigned)basicCost, (unsigned)repeatCost, (unsigned)compressedCost); 217 if (basicCost <= repeatCost && basicCost <= compressedCost) { 218 DEBUGLOG(5, "Selected set_basic"); 219 assert(isDefaultAllowed); 220 *repeatMode = FSE_repeat_none; 221 return set_basic; 222 } 223 if (repeatCost <= compressedCost) { 224 DEBUGLOG(5, "Selected set_repeat"); 225 assert(!ZSTD_isError(repeatCost)); 226 return set_repeat; 227 } 228 assert(compressedCost < basicCost && compressedCost < repeatCost); 229 } 230 DEBUGLOG(5, "Selected set_compressed"); 231 *repeatMode = FSE_repeat_check; 232 return set_compressed; 233 } 234 235 typedef struct { 236 S16 norm[MaxSeq + 1]; 237 U32 wksp[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(MaxSeq, MaxFSELog)]; 238 } ZSTD_BuildCTableWksp; 239 240 size_t 241 ZSTD_buildCTable(void* dst, size_t dstCapacity, 242 FSE_CTable* nextCTable, U32 FSELog, symbolEncodingType_e type, 243 unsigned* count, U32 max, 244 const BYTE* codeTable, size_t nbSeq, 245 const S16* defaultNorm, U32 defaultNormLog, U32 defaultMax, 246 const FSE_CTable* prevCTable, size_t prevCTableSize, 247 void* entropyWorkspace, size_t entropyWorkspaceSize) 248 { 249 BYTE* op = (BYTE*)dst; 250 const BYTE* const oend = op + dstCapacity; 251 DEBUGLOG(6, "ZSTD_buildCTable (dstCapacity=%u)", (unsigned)dstCapacity); 252 253 switch (type) { 254 case set_rle: 255 FORWARD_IF_ERROR(FSE_buildCTable_rle(nextCTable, (BYTE)max), ""); 256 RETURN_ERROR_IF(dstCapacity==0, dstSize_tooSmall, "not enough space"); 257 *op = codeTable[0]; 258 return 1; 259 case set_repeat: 260 ZSTD_memcpy(nextCTable, prevCTable, prevCTableSize); 261 return 0; 262 case set_basic: 263 FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, defaultNorm, defaultMax, defaultNormLog, entropyWorkspace, entropyWorkspaceSize), ""); /* note : could be pre-calculated */ 264 return 0; 265 case set_compressed: { 266 ZSTD_BuildCTableWksp* wksp = (ZSTD_BuildCTableWksp*)entropyWorkspace; 267 size_t nbSeq_1 = nbSeq; 268 const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max); 269 if (count[codeTable[nbSeq-1]] > 1) { 270 count[codeTable[nbSeq-1]]--; 271 nbSeq_1--; 272 } 273 assert(nbSeq_1 > 1); 274 assert(entropyWorkspaceSize >= sizeof(ZSTD_BuildCTableWksp)); 275 (void)entropyWorkspaceSize; 276 FORWARD_IF_ERROR(FSE_normalizeCount(wksp->norm, tableLog, count, nbSeq_1, max, ZSTD_useLowProbCount(nbSeq_1)), ""); 277 { size_t const NCountSize = FSE_writeNCount(op, oend - op, wksp->norm, max, tableLog); /* overflow protected */ 278 FORWARD_IF_ERROR(NCountSize, "FSE_writeNCount failed"); 279 FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, wksp->norm, max, tableLog, wksp->wksp, sizeof(wksp->wksp)), ""); 280 return NCountSize; 281 } 282 } 283 default: assert(0); RETURN_ERROR(GENERIC, "impossible to reach"); 284 } 285 } 286 287 FORCE_INLINE_TEMPLATE size_t 288 ZSTD_encodeSequences_body( 289 void* dst, size_t dstCapacity, 290 FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable, 291 FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable, 292 FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable, 293 seqDef const* sequences, size_t nbSeq, int longOffsets) 294 { 295 BIT_CStream_t blockStream; 296 FSE_CState_t stateMatchLength; 297 FSE_CState_t stateOffsetBits; 298 FSE_CState_t stateLitLength; 299 300 RETURN_ERROR_IF( 301 ERR_isError(BIT_initCStream(&blockStream, dst, dstCapacity)), 302 dstSize_tooSmall, "not enough space remaining"); 303 DEBUGLOG(6, "available space for bitstream : %i (dstCapacity=%u)", 304 (int)(blockStream.endPtr - blockStream.startPtr), 305 (unsigned)dstCapacity); 306 307 /* first symbols */ 308 FSE_initCState2(&stateMatchLength, CTable_MatchLength, mlCodeTable[nbSeq-1]); 309 FSE_initCState2(&stateOffsetBits, CTable_OffsetBits, ofCodeTable[nbSeq-1]); 310 FSE_initCState2(&stateLitLength, CTable_LitLength, llCodeTable[nbSeq-1]); 311 BIT_addBits(&blockStream, sequences[nbSeq-1].litLength, LL_bits[llCodeTable[nbSeq-1]]); 312 if (MEM_32bits()) BIT_flushBits(&blockStream); 313 BIT_addBits(&blockStream, sequences[nbSeq-1].matchLength, ML_bits[mlCodeTable[nbSeq-1]]); 314 if (MEM_32bits()) BIT_flushBits(&blockStream); 315 if (longOffsets) { 316 U32 const ofBits = ofCodeTable[nbSeq-1]; 317 unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1); 318 if (extraBits) { 319 BIT_addBits(&blockStream, sequences[nbSeq-1].offset, extraBits); 320 BIT_flushBits(&blockStream); 321 } 322 BIT_addBits(&blockStream, sequences[nbSeq-1].offset >> extraBits, 323 ofBits - extraBits); 324 } else { 325 BIT_addBits(&blockStream, sequences[nbSeq-1].offset, ofCodeTable[nbSeq-1]); 326 } 327 BIT_flushBits(&blockStream); 328 329 { size_t n; 330 for (n=nbSeq-2 ; n<nbSeq ; n--) { /* intentional underflow */ 331 BYTE const llCode = llCodeTable[n]; 332 BYTE const ofCode = ofCodeTable[n]; 333 BYTE const mlCode = mlCodeTable[n]; 334 U32 const llBits = LL_bits[llCode]; 335 U32 const ofBits = ofCode; 336 U32 const mlBits = ML_bits[mlCode]; 337 DEBUGLOG(6, "encoding: litlen:%2u - matchlen:%2u - offCode:%7u", 338 (unsigned)sequences[n].litLength, 339 (unsigned)sequences[n].matchLength + MINMATCH, 340 (unsigned)sequences[n].offset); 341 /* 32b*/ /* 64b*/ 342 /* (7)*/ /* (7)*/ 343 FSE_encodeSymbol(&blockStream, &stateOffsetBits, ofCode); /* 15 */ /* 15 */ 344 FSE_encodeSymbol(&blockStream, &stateMatchLength, mlCode); /* 24 */ /* 24 */ 345 if (MEM_32bits()) BIT_flushBits(&blockStream); /* (7)*/ 346 FSE_encodeSymbol(&blockStream, &stateLitLength, llCode); /* 16 */ /* 33 */ 347 if (MEM_32bits() || (ofBits+mlBits+llBits >= 64-7-(LLFSELog+MLFSELog+OffFSELog))) 348 BIT_flushBits(&blockStream); /* (7)*/ 349 BIT_addBits(&blockStream, sequences[n].litLength, llBits); 350 if (MEM_32bits() && ((llBits+mlBits)>24)) BIT_flushBits(&blockStream); 351 BIT_addBits(&blockStream, sequences[n].matchLength, mlBits); 352 if (MEM_32bits() || (ofBits+mlBits+llBits > 56)) BIT_flushBits(&blockStream); 353 if (longOffsets) { 354 unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1); 355 if (extraBits) { 356 BIT_addBits(&blockStream, sequences[n].offset, extraBits); 357 BIT_flushBits(&blockStream); /* (7)*/ 358 } 359 BIT_addBits(&blockStream, sequences[n].offset >> extraBits, 360 ofBits - extraBits); /* 31 */ 361 } else { 362 BIT_addBits(&blockStream, sequences[n].offset, ofBits); /* 31 */ 363 } 364 BIT_flushBits(&blockStream); /* (7)*/ 365 DEBUGLOG(7, "remaining space : %i", (int)(blockStream.endPtr - blockStream.ptr)); 366 } } 367 368 DEBUGLOG(6, "ZSTD_encodeSequences: flushing ML state with %u bits", stateMatchLength.stateLog); 369 FSE_flushCState(&blockStream, &stateMatchLength); 370 DEBUGLOG(6, "ZSTD_encodeSequences: flushing Off state with %u bits", stateOffsetBits.stateLog); 371 FSE_flushCState(&blockStream, &stateOffsetBits); 372 DEBUGLOG(6, "ZSTD_encodeSequences: flushing LL state with %u bits", stateLitLength.stateLog); 373 FSE_flushCState(&blockStream, &stateLitLength); 374 375 { size_t const streamSize = BIT_closeCStream(&blockStream); 376 RETURN_ERROR_IF(streamSize==0, dstSize_tooSmall, "not enough space"); 377 return streamSize; 378 } 379 } 380 381 static size_t 382 ZSTD_encodeSequences_default( 383 void* dst, size_t dstCapacity, 384 FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable, 385 FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable, 386 FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable, 387 seqDef const* sequences, size_t nbSeq, int longOffsets) 388 { 389 return ZSTD_encodeSequences_body(dst, dstCapacity, 390 CTable_MatchLength, mlCodeTable, 391 CTable_OffsetBits, ofCodeTable, 392 CTable_LitLength, llCodeTable, 393 sequences, nbSeq, longOffsets); 394 } 395 396 397 #if DYNAMIC_BMI2 398 399 static TARGET_ATTRIBUTE("bmi2") size_t 400 ZSTD_encodeSequences_bmi2( 401 void* dst, size_t dstCapacity, 402 FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable, 403 FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable, 404 FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable, 405 seqDef const* sequences, size_t nbSeq, int longOffsets) 406 { 407 return ZSTD_encodeSequences_body(dst, dstCapacity, 408 CTable_MatchLength, mlCodeTable, 409 CTable_OffsetBits, ofCodeTable, 410 CTable_LitLength, llCodeTable, 411 sequences, nbSeq, longOffsets); 412 } 413 414 #endif 415 416 size_t ZSTD_encodeSequences( 417 void* dst, size_t dstCapacity, 418 FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable, 419 FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable, 420 FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable, 421 seqDef const* sequences, size_t nbSeq, int longOffsets, int bmi2) 422 { 423 DEBUGLOG(5, "ZSTD_encodeSequences: dstCapacity = %u", (unsigned)dstCapacity); 424 #if DYNAMIC_BMI2 425 if (bmi2) { 426 return ZSTD_encodeSequences_bmi2(dst, dstCapacity, 427 CTable_MatchLength, mlCodeTable, 428 CTable_OffsetBits, ofCodeTable, 429 CTable_LitLength, llCodeTable, 430 sequences, nbSeq, longOffsets); 431 } 432 #endif 433 (void)bmi2; 434 return ZSTD_encodeSequences_default(dst, dstCapacity, 435 CTable_MatchLength, mlCodeTable, 436 CTable_OffsetBits, ofCodeTable, 437 CTable_LitLength, llCodeTable, 438 sequences, nbSeq, longOffsets); 439 } 440