| 1 | |
| 2 | |
| 3 | * FSE : Finite State Entropy encoder
|
| 4 | * Copyright (c) Meta Platforms, Inc. and affiliates.
|
| 5 | *
|
| 6 | * You can contact the author at :
|
| 7 | * - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
|
| 8 | * - Public forum : https://groups.google.com/forum/#!forum/lz4c
|
| 9 | *
|
| 10 | * This source code is licensed under both the BSD-style license (found in the
|
| 11 | * LICENSE file in the root directory of this source tree) and the GPLv2 (found
|
| 12 | * in the COPYING file in the root directory of this source tree).
|
| 13 | * You may select, at your option, one of the above-listed licenses.
|
| 14 | ******************************************************************
|
| 15 |
|
| 16 | |
| 17 | * Includes
|
| 18 | ***************************************************************
|
| 19 | #include "../common/compiler.h"
|
| 20 | #include "../common/mem.h"
|
| 21 | #include "../common/debug.h"
|
| 22 | #include "hist.h"
|
| 23 | #include "../common/bitstream.h"
|
| 24 | #define FSE_STATIC_LINKING_ONLY
|
| 25 | #include "../common/fse.h"
|
| 26 | #include "../common/error_private.h"
|
| 27 | #define ZSTD_DEPS_NEED_MALLOC
|
| 28 | #define ZSTD_DEPS_NEED_MATH64
|
| 29 | #include "../common/zstd_deps.h"
|
| 30 | #include "../common/bits.h"
|
| 31 |
|
| 32 |
|
| 33 | |
| 34 | * Error Management
|
| 35 | ***************************************************************
|
| 36 | #define FSE_isError ERR_isError
|
| 37 |
|
| 38 |
|
| 39 | |
| 40 | * Templates
|
| 41 | ***************************************************************
|
| 42 | |
| 43 | designed to be included
|
| 44 | for type-specific functions (template emulation in C)
|
| 45 | Objective is to write these functions only once, for improved maintenance
|
| 46 |
|
| 47 |
|
| 48 |
|
| 49 | #ifndef FSE_FUNCTION_EXTENSION
|
| 50 | # error "FSE_FUNCTION_EXTENSION must be defined"
|
| 51 | #endif
|
| 52 | #ifndef FSE_FUNCTION_TYPE
|
| 53 | # error "FSE_FUNCTION_TYPE must be defined"
|
| 54 | #endif
|
| 55 |
|
| 56 |
|
| 57 | #define FSE_CAT(X,Y) X##Y
|
| 58 | #define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
|
| 59 | #define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
|
| 60 |
|
| 61 |
|
| 62 |
|
| 63 |
|
| 64 | |
| 65 | * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
|
| 66 | * wkspSize should be sized to handle worst case situation, which is `1<<max_tableLog * sizeof(FSE_FUNCTION_TYPE)`
|
| 67 | * workSpace must also be properly aligned with FSE_FUNCTION_TYPE requirements
|
| 68 |
|
| 69 | size_t FSE_buildCTable_wksp(FSE_CTable* ct,
|
| 70 | const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog,
|
| 71 | void* workSpace, size_t wkspSize)
|
| 72 | {
|
| 73 | U32 const tableSize = 1 << tableLog;
|
| 74 | U32 const tableMask = tableSize - 1;
|
| 75 | void* const ptr = ct;
|
| 76 | U16* const tableU16 = ( (U16*) ptr) + 2;
|
| 77 | void* const FSCT = ((U32*)ptr) + 1 + (tableLog ? tableSize>>1 : 1) ;
|
| 78 | FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
|
| 79 | U32 const step = FSE_TABLESTEP(tableSize);
|
| 80 | U32 const maxSV1 = maxSymbolValue+1;
|
| 81 |
|
| 82 | U16* cumul = (U16*)workSpace;
|
| 83 | FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSV1+1));
|
| 84 |
|
| 85 | U32 highThreshold = tableSize-1;
|
| 86 |
|
| 87 | assert(((size_t)workSpace & 1) == 0);
|
| 88 | if (FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) > wkspSize) return ERROR(tableLog_tooLarge);
|
| 89 |
|
| 90 | tableU16[-2] = (U16) tableLog;
|
| 91 | tableU16[-1] = (U16) maxSymbolValue;
|
| 92 | assert(tableLog < 16);
|
| 93 |
|
| 94 | |
| 95 | * https://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html
|
| 96 |
|
| 97 | #ifdef __clang_analyzer__
|
| 98 | ZSTD_memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize);
|
| 99 | #endif
|
| 100 |
|
| 101 |
|
| 102 | { U32 u;
|
| 103 | cumul[0] = 0;
|
| 104 | for (u=1; u <= maxSV1; u++) {
|
| 105 | if (normalizedCounter[u-1]==-1) {
|
| 106 | cumul[u] = cumul[u-1] + 1;
|
| 107 | tableSymbol[highThreshold--] = (FSE_FUNCTION_TYPE)(u-1);
|
| 108 | } else {
|
| 109 | assert(normalizedCounter[u-1] >= 0);
|
| 110 | cumul[u] = cumul[u-1] + (U16)normalizedCounter[u-1];
|
| 111 | assert(cumul[u] >= cumul[u-1]);
|
| 112 | } }
|
| 113 | cumul[maxSV1] = (U16)(tableSize+1);
|
| 114 | }
|
| 115 |
|
| 116 |
|
| 117 | if (highThreshold == tableSize - 1) {
|
| 118 | |
| 119 | * to reduce branch misses since we are operating on a small block
|
| 120 |
|
| 121 | BYTE* const spread = tableSymbol + tableSize;
|
| 122 | { U64 const add = 0x0101010101010101ull;
|
| 123 | size_t pos = 0;
|
| 124 | U64 sv = 0;
|
| 125 | U32 s;
|
| 126 | for (s=0; s<maxSV1; ++s, sv += add) {
|
| 127 | int i;
|
| 128 | int const n = normalizedCounter[s];
|
| 129 | MEM_write64(spread + pos, sv);
|
| 130 | for (i = 8; i < n; i += 8) {
|
| 131 | MEM_write64(spread + pos + i, sv);
|
| 132 | }
|
| 133 | assert(n>=0);
|
| 134 | pos += (size_t)n;
|
| 135 | }
|
| 136 | }
|
| 137 | |
| 138 | * we don't need variable sized inner loop, so we can unroll the loop and
|
| 139 | * reduce branch misses.
|
| 140 |
|
| 141 | { size_t position = 0;
|
| 142 | size_t s;
|
| 143 | size_t const unroll = 2;
|
| 144 | assert(tableSize % unroll == 0);
|
| 145 | for (s = 0; s < (size_t)tableSize; s += unroll) {
|
| 146 | size_t u;
|
| 147 | for (u = 0; u < unroll; ++u) {
|
| 148 | size_t const uPosition = (position + (u * step)) & tableMask;
|
| 149 | tableSymbol[uPosition] = spread[s + u];
|
| 150 | }
|
| 151 | position = (position + (unroll * step)) & tableMask;
|
| 152 | }
|
| 153 | assert(position == 0);
|
| 154 | }
|
| 155 | } else {
|
| 156 | U32 position = 0;
|
| 157 | U32 symbol;
|
| 158 | for (symbol=0; symbol<maxSV1; symbol++) {
|
| 159 | int nbOccurrences;
|
| 160 | int const freq = normalizedCounter[symbol];
|
| 161 | for (nbOccurrences=0; nbOccurrences<freq; nbOccurrences++) {
|
| 162 | tableSymbol[position] = (FSE_FUNCTION_TYPE)symbol;
|
| 163 | position = (position + step) & tableMask;
|
| 164 | while (position > highThreshold)
|
| 165 | position = (position + step) & tableMask;
|
| 166 | } }
|
| 167 | assert(position==0);
|
| 168 | }
|
| 169 |
|
| 170 |
|
| 171 | { U32 u; for (u=0; u<tableSize; u++) {
|
| 172 | FSE_FUNCTION_TYPE s = tableSymbol[u];
|
| 173 | tableU16[cumul[s]++] = (U16) (tableSize+u);
|
| 174 | } }
|
| 175 |
|
| 176 |
|
| 177 | { unsigned total = 0;
|
| 178 | unsigned s;
|
| 179 | for (s=0; s<=maxSymbolValue; s++) {
|
| 180 | switch (normalizedCounter[s])
|
| 181 | {
|
| 182 | case 0:
|
| 183 |
|
| 184 | symbolTT[s].deltaNbBits = ((tableLog+1) << 16) - (1<<tableLog);
|
| 185 | break;
|
| 186 |
|
| 187 | case -1:
|
| 188 | case 1:
|
| 189 | symbolTT[s].deltaNbBits = (tableLog << 16) - (1<<tableLog);
|
| 190 | assert(total <= INT_MAX);
|
| 191 | symbolTT[s].deltaFindState = (int)(total - 1);
|
| 192 | total ++;
|
| 193 | break;
|
| 194 | default :
|
| 195 | assert(normalizedCounter[s] > 1);
|
| 196 | { U32 const maxBitsOut = tableLog - ZSTD_highbit32 ((U32)normalizedCounter[s]-1);
|
| 197 | U32 const minStatePlus = (U32)normalizedCounter[s] << maxBitsOut;
|
| 198 | symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus;
|
| 199 | symbolTT[s].deltaFindState = (int)(total - (unsigned)normalizedCounter[s]);
|
| 200 | total += (unsigned)normalizedCounter[s];
|
| 201 | } } } }
|
| 202 |
|
| 203 | #if 0 |
| 204 | DEBUGLOG(5, "\n --- table statistics : ");
|
| 205 | { U32 symbol;
|
| 206 | for (symbol=0; symbol<=maxSymbolValue; symbol++) {
|
| 207 | DEBUGLOG(5, "%3u: w=%3i, maxBits=%u, fracBits=%.2f",
|
| 208 | symbol, normalizedCounter[symbol],
|
| 209 | FSE_getMaxNbBits(symbolTT, symbol),
|
| 210 | (double)FSE_bitCost(symbolTT, tableLog, symbol, 8) / 256);
|
| 211 | } }
|
| 212 | #endif
|
| 213 |
|
| 214 | return 0;
|
| 215 | }
|
| 216 |
|
| 217 |
|
| 218 |
|
| 219 | #ifndef FSE_COMMONDEFS_ONLY
|
| 220 |
|
| 221 | |
| 222 | * FSE NCount encoding
|
| 223 | ***************************************************************
|
| 224 | size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog)
|
| 225 | {
|
| 226 | size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog
|
| 227 | + 4
|
| 228 | + 2 ) / 8)
|
| 229 | + 1
|
| 230 | + 2 ;
|
| 231 | return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND;
|
| 232 | }
|
| 233 |
|
| 234 | static size_t
|
| 235 | FSE_writeNCount_generic (void* header, size_t headerBufferSize,
|
| 236 | const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog,
|
| 237 | unsigned writeIsSafe)
|
| 238 | {
|
| 239 | BYTE* const ostart = (BYTE*) header;
|
| 240 | BYTE* out = ostart;
|
| 241 | BYTE* const oend = ostart + headerBufferSize;
|
| 242 | int nbBits;
|
| 243 | const int tableSize = 1 << tableLog;
|
| 244 | int remaining;
|
| 245 | int threshold;
|
| 246 | U32 bitStream = 0;
|
| 247 | int bitCount = 0;
|
| 248 | unsigned symbol = 0;
|
| 249 | unsigned const alphabetSize = maxSymbolValue + 1;
|
| 250 | int previousIs0 = 0;
|
| 251 |
|
| 252 |
|
| 253 | bitStream += (tableLog-FSE_MIN_TABLELOG) << bitCount;
|
| 254 | bitCount += 4;
|
| 255 |
|
| 256 |
|
| 257 | remaining = tableSize+1;
|
| 258 | threshold = tableSize;
|
| 259 | nbBits = (int)tableLog+1;
|
| 260 |
|
| 261 | while ((symbol < alphabetSize) && (remaining>1)) {
|
| 262 | if (previousIs0) {
|
| 263 | unsigned start = symbol;
|
| 264 | while ((symbol < alphabetSize) && !normalizedCounter[symbol]) symbol++;
|
| 265 | if (symbol == alphabetSize) break;
|
| 266 | while (symbol >= start+24) {
|
| 267 | start+=24;
|
| 268 | bitStream += 0xFFFFU << bitCount;
|
| 269 | if ((!writeIsSafe) && (out > oend-2))
|
| 270 | return ERROR(dstSize_tooSmall);
|
| 271 | out[0] = (BYTE) bitStream;
|
| 272 | out[1] = (BYTE)(bitStream>>8);
|
| 273 | out+=2;
|
| 274 | bitStream>>=16;
|
| 275 | }
|
| 276 | while (symbol >= start+3) {
|
| 277 | start+=3;
|
| 278 | bitStream += 3U << bitCount;
|
| 279 | bitCount += 2;
|
| 280 | }
|
| 281 | bitStream += (symbol-start) << bitCount;
|
| 282 | bitCount += 2;
|
| 283 | if (bitCount>16) {
|
| 284 | if ((!writeIsSafe) && (out > oend - 2))
|
| 285 | return ERROR(dstSize_tooSmall);
|
| 286 | out[0] = (BYTE)bitStream;
|
| 287 | out[1] = (BYTE)(bitStream>>8);
|
| 288 | out += 2;
|
| 289 | bitStream >>= 16;
|
| 290 | bitCount -= 16;
|
| 291 | } }
|
| 292 | { int count = normalizedCounter[symbol++];
|
| 293 | int const max = (2*threshold-1) - remaining;
|
| 294 | remaining -= count < 0 ? -count : count;
|
| 295 | count++;
|
| 296 | if (count>=threshold)
|
| 297 | count += max;
|
| 298 | bitStream += (U32)count << bitCount;
|
| 299 | bitCount += nbBits;
|
| 300 | bitCount -= (count<max);
|
| 301 | previousIs0 = (count==1);
|
| 302 | if (remaining<1) return ERROR(GENERIC);
|
| 303 | while (remaining<threshold) { nbBits--; threshold>>=1; }
|
| 304 | }
|
| 305 | if (bitCount>16) {
|
| 306 | if ((!writeIsSafe) && (out > oend - 2))
|
| 307 | return ERROR(dstSize_tooSmall);
|
| 308 | out[0] = (BYTE)bitStream;
|
| 309 | out[1] = (BYTE)(bitStream>>8);
|
| 310 | out += 2;
|
| 311 | bitStream >>= 16;
|
| 312 | bitCount -= 16;
|
| 313 | } }
|
| 314 |
|
| 315 | if (remaining != 1)
|
| 316 | return ERROR(GENERIC);
|
| 317 | assert(symbol <= alphabetSize);
|
| 318 |
|
| 319 |
|
| 320 | if ((!writeIsSafe) && (out > oend - 2))
|
| 321 | return ERROR(dstSize_tooSmall);
|
| 322 | out[0] = (BYTE)bitStream;
|
| 323 | out[1] = (BYTE)(bitStream>>8);
|
| 324 | out+= (bitCount+7) /8;
|
| 325 |
|
| 326 | assert(out >= ostart);
|
| 327 | return (size_t)(out-ostart);
|
| 328 | }
|
| 329 |
|
| 330 |
|
| 331 | size_t FSE_writeNCount (void* buffer, size_t bufferSize,
|
| 332 | const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
|
| 333 | {
|
| 334 | if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
|
| 335 | if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC);
|
| 336 |
|
| 337 | if (bufferSize < FSE_NCountWriteBound(maxSymbolValue, tableLog))
|
| 338 | return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 0);
|
| 339 |
|
| 340 | return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 1 );
|
| 341 | }
|
| 342 |
|
| 343 |
|
| 344 | |
| 345 | * FSE Compression Code
|
| 346 | ***************************************************************
|
| 347 |
|
| 348 |
|
| 349 | static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue)
|
| 350 | {
|
| 351 | U32 minBitsSrc = ZSTD_highbit32((U32)(srcSize)) + 1;
|
| 352 | U32 minBitsSymbols = ZSTD_highbit32(maxSymbolValue) + 2;
|
| 353 | U32 minBits = minBitsSrc < minBitsSymbols ? minBitsSrc : minBitsSymbols;
|
| 354 | assert(srcSize > 1);
|
| 355 | return minBits;
|
| 356 | }
|
| 357 |
|
| 358 | unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus)
|
| 359 | {
|
| 360 | U32 maxBitsSrc = ZSTD_highbit32((U32)(srcSize - 1)) - minus;
|
| 361 | U32 tableLog = maxTableLog;
|
| 362 | U32 minBits = FSE_minTableLog(srcSize, maxSymbolValue);
|
| 363 | assert(srcSize > 1);
|
| 364 | if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
|
| 365 | if (maxBitsSrc < tableLog) tableLog = maxBitsSrc;
|
| 366 | if (minBits > tableLog) tableLog = minBits;
|
| 367 | if (tableLog < FSE_MIN_TABLELOG) tableLog = FSE_MIN_TABLELOG;
|
| 368 | if (tableLog > FSE_MAX_TABLELOG) tableLog = FSE_MAX_TABLELOG;
|
| 369 | return tableLog;
|
| 370 | }
|
| 371 |
|
| 372 | unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
|
| 373 | {
|
| 374 | return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 2);
|
| 375 | }
|
| 376 |
|
| 377 | |
| 378 | To be used when primary method fails.
|
| 379 |
|
| 380 | static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue, short lowProbCount)
|
| 381 | {
|
| 382 | short const NOT_YET_ASSIGNED = -2;
|
| 383 | U32 s;
|
| 384 | U32 distributed = 0;
|
| 385 | U32 ToDistribute;
|
| 386 |
|
| 387 |
|
| 388 | U32 const lowThreshold = (U32)(total >> tableLog);
|
| 389 | U32 lowOne = (U32)((total * 3) >> (tableLog + 1));
|
| 390 |
|
| 391 | for (s=0; s<=maxSymbolValue; s++) {
|
| 392 | if (count[s] == 0) {
|
| 393 | norm[s]=0;
|
| 394 | continue;
|
| 395 | }
|
| 396 | if (count[s] <= lowThreshold) {
|
| 397 | norm[s] = lowProbCount;
|
| 398 | distributed++;
|
| 399 | total -= count[s];
|
| 400 | continue;
|
| 401 | }
|
| 402 | if (count[s] <= lowOne) {
|
| 403 | norm[s] = 1;
|
| 404 | distributed++;
|
| 405 | total -= count[s];
|
| 406 | continue;
|
| 407 | }
|
| 408 |
|
| 409 | norm[s]=NOT_YET_ASSIGNED;
|
| 410 | }
|
| 411 | ToDistribute = (1 << tableLog) - distributed;
|
| 412 |
|
| 413 | if (ToDistribute == 0)
|
| 414 | return 0;
|
| 415 |
|
| 416 | if ((total / ToDistribute) > lowOne) {
|
| 417 |
|
| 418 | lowOne = (U32)((total * 3) / (ToDistribute * 2));
|
| 419 | for (s=0; s<=maxSymbolValue; s++) {
|
| 420 | if ((norm[s] == NOT_YET_ASSIGNED) && (count[s] <= lowOne)) {
|
| 421 | norm[s] = 1;
|
| 422 | distributed++;
|
| 423 | total -= count[s];
|
| 424 | continue;
|
| 425 | } }
|
| 426 | ToDistribute = (1 << tableLog) - distributed;
|
| 427 | }
|
| 428 |
|
| 429 | if (distributed == maxSymbolValue+1) {
|
| 430 | |
| 431 | probably incompressible data (should have already been detected);
|
| 432 | find max, then give all remaining points to max
|
| 433 | U32 maxV = 0, maxC = 0;
|
| 434 | for (s=0; s<=maxSymbolValue; s++)
|
| 435 | if (count[s] > maxC) { maxV=s; maxC=count[s]; }
|
| 436 | norm[maxV] += (short)ToDistribute;
|
| 437 | return 0;
|
| 438 | }
|
| 439 |
|
| 440 | if (total == 0) {
|
| 441 |
|
| 442 | for (s=0; ToDistribute > 0; s = (s+1)%(maxSymbolValue+1))
|
| 443 | if (norm[s] > 0) { ToDistribute--; norm[s]++; }
|
| 444 | return 0;
|
| 445 | }
|
| 446 |
|
| 447 | { U64 const vStepLog = 62 - tableLog;
|
| 448 | U64 const mid = (1ULL << (vStepLog-1)) - 1;
|
| 449 | U64 const rStep = ZSTD_div64((((U64)1<<vStepLog) * ToDistribute) + mid, (U32)total);
|
| 450 | U64 tmpTotal = mid;
|
| 451 | for (s=0; s<=maxSymbolValue; s++) {
|
| 452 | if (norm[s]==NOT_YET_ASSIGNED) {
|
| 453 | U64 const end = tmpTotal + (count[s] * rStep);
|
| 454 | U32 const sStart = (U32)(tmpTotal >> vStepLog);
|
| 455 | U32 const sEnd = (U32)(end >> vStepLog);
|
| 456 | U32 const weight = sEnd - sStart;
|
| 457 | if (weight < 1)
|
| 458 | return ERROR(GENERIC);
|
| 459 | norm[s] = (short)weight;
|
| 460 | tmpTotal = end;
|
| 461 | } } }
|
| 462 |
|
| 463 | return 0;
|
| 464 | }
|
| 465 |
|
| 466 | size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
|
| 467 | const unsigned* count, size_t total,
|
| 468 | unsigned maxSymbolValue, unsigned useLowProbCount)
|
| 469 | {
|
| 470 |
|
| 471 | if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
|
| 472 | if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC);
|
| 473 | if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
|
| 474 | if (tableLog < FSE_minTableLog(total, maxSymbolValue)) return ERROR(GENERIC);
|
| 475 |
|
| 476 | { static U32 const rtbTable[] = { 0, 473195, 504333, 520860, 550000, 700000, 750000, 830000 };
|
| 477 | short const lowProbCount = useLowProbCount ? -1 : 1;
|
| 478 | U64 const scale = 62 - tableLog;
|
| 479 | U64 const step = ZSTD_div64((U64)1<<62, (U32)total);
|
| 480 | U64 const vStep = 1ULL<<(scale-20);
|
| 481 | int stillToDistribute = 1<<tableLog;
|
| 482 | unsigned s;
|
| 483 | unsigned largest=0;
|
| 484 | short largestP=0;
|
| 485 | U32 lowThreshold = (U32)(total >> tableLog);
|
| 486 |
|
| 487 | for (s=0; s<=maxSymbolValue; s++) {
|
| 488 | if (count[s] == total) return 0;
|
| 489 | if (count[s] == 0) { normalizedCounter[s]=0; continue; }
|
| 490 | if (count[s] <= lowThreshold) {
|
| 491 | normalizedCounter[s] = lowProbCount;
|
| 492 | stillToDistribute--;
|
| 493 | } else {
|
| 494 | short proba = (short)((count[s]*step) >> scale);
|
| 495 | if (proba<8) {
|
| 496 | U64 restToBeat = vStep * rtbTable[proba];
|
| 497 | proba += (count[s]*step) - ((U64)proba<<scale) > restToBeat;
|
| 498 | }
|
| 499 | if (proba > largestP) { largestP=proba; largest=s; }
|
| 500 | normalizedCounter[s] = proba;
|
| 501 | stillToDistribute -= proba;
|
| 502 | } }
|
| 503 | if (-stillToDistribute >= (normalizedCounter[largest] >> 1)) {
|
| 504 |
|
| 505 | size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue, lowProbCount);
|
| 506 | if (FSE_isError(errorCode)) return errorCode;
|
| 507 | }
|
| 508 | else normalizedCounter[largest] += (short)stillToDistribute;
|
| 509 | }
|
| 510 |
|
| 511 | #if 0 |
| 512 | { /* Print Table (debug) */
|
| 513 | U32 s;
|
| 514 | U32 nTotal = 0;
|
| 515 | for (s=0; s<=maxSymbolValue; s++)
|
| 516 | RAWLOG(2, "%3i: %4i \n", s, normalizedCounter[s]);
|
| 517 | for (s=0; s<=maxSymbolValue; s++)
|
| 518 | nTotal += abs(normalizedCounter[s]);
|
| 519 | if (nTotal != (1U<<tableLog))
|
| 520 | RAWLOG(2, "Warning !!! Total == %u != %u !!!", nTotal, 1U<<tableLog);
|
| 521 | getchar();
|
| 522 | }
|
| 523 | #endif
|
| 524 |
|
| 525 | return tableLog;
|
| 526 | }
|
| 527 |
|
| 528 |
|
| 529 | size_t FSE_buildCTable_rle (FSE_CTable* ct, BYTE symbolValue)
|
| 530 | {
|
| 531 | void* ptr = ct;
|
| 532 | U16* tableU16 = ( (U16*) ptr) + 2;
|
| 533 | void* FSCTptr = (U32*)ptr + 2;
|
| 534 | FSE_symbolCompressionTransform* symbolTT = (FSE_symbolCompressionTransform*) FSCTptr;
|
| 535 |
|
| 536 |
|
| 537 | tableU16[-2] = (U16) 0;
|
| 538 | tableU16[-1] = (U16) symbolValue;
|
| 539 |
|
| 540 |
|
| 541 | tableU16[0] = 0;
|
| 542 | tableU16[1] = 0;
|
| 543 |
|
| 544 |
|
| 545 | symbolTT[symbolValue].deltaNbBits = 0;
|
| 546 | symbolTT[symbolValue].deltaFindState = 0;
|
| 547 |
|
| 548 | return 0;
|
| 549 | }
|
| 550 |
|
| 551 |
|
| 552 | static size_t FSE_compress_usingCTable_generic (void* dst, size_t dstSize,
|
| 553 | const void* src, size_t srcSize,
|
| 554 | const FSE_CTable* ct, const unsigned fast)
|
| 555 | {
|
| 556 | const BYTE* const istart = (const BYTE*) src;
|
| 557 | const BYTE* const iend = istart + srcSize;
|
| 558 | const BYTE* ip=iend;
|
| 559 |
|
| 560 | BIT_CStream_t bitC;
|
| 561 | FSE_CState_t CState1, CState2;
|
| 562 |
|
| 563 |
|
| 564 | if (srcSize <= 2) return 0;
|
| 565 | { size_t const initError = BIT_initCStream(&bitC, dst, dstSize);
|
| 566 | if (FSE_isError(initError)) return 0; }
|
| 567 |
|
| 568 | #define FSE_FLUSHBITS(s) (fast ? BIT_flushBitsFast(s) : BIT_flushBits(s))
|
| 569 |
|
| 570 | if (srcSize & 1) {
|
| 571 | FSE_initCState2(&CState1, ct, *--ip);
|
| 572 | FSE_initCState2(&CState2, ct, *--ip);
|
| 573 | FSE_encodeSymbol(&bitC, &CState1, *--ip);
|
| 574 | FSE_FLUSHBITS(&bitC);
|
| 575 | } else {
|
| 576 | FSE_initCState2(&CState2, ct, *--ip);
|
| 577 | FSE_initCState2(&CState1, ct, *--ip);
|
| 578 | }
|
| 579 |
|
| 580 |
|
| 581 | srcSize -= 2;
|
| 582 | if ((sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) && (srcSize & 2)) {
|
| 583 | FSE_encodeSymbol(&bitC, &CState2, *--ip);
|
| 584 | FSE_encodeSymbol(&bitC, &CState1, *--ip);
|
| 585 | FSE_FLUSHBITS(&bitC);
|
| 586 | }
|
| 587 |
|
| 588 |
|
| 589 | while ( ip>istart ) {
|
| 590 |
|
| 591 | FSE_encodeSymbol(&bitC, &CState2, *--ip);
|
| 592 |
|
| 593 | if (sizeof(bitC.bitContainer)*8 < FSE_MAX_TABLELOG*2+7 )
|
| 594 | FSE_FLUSHBITS(&bitC);
|
| 595 |
|
| 596 | FSE_encodeSymbol(&bitC, &CState1, *--ip);
|
| 597 |
|
| 598 | if (sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) {
|
| 599 | FSE_encodeSymbol(&bitC, &CState2, *--ip);
|
| 600 | FSE_encodeSymbol(&bitC, &CState1, *--ip);
|
| 601 | }
|
| 602 |
|
| 603 | FSE_FLUSHBITS(&bitC);
|
| 604 | }
|
| 605 |
|
| 606 | FSE_flushCState(&bitC, &CState2);
|
| 607 | FSE_flushCState(&bitC, &CState1);
|
| 608 | return BIT_closeCStream(&bitC);
|
| 609 | }
|
| 610 |
|
| 611 | size_t FSE_compress_usingCTable (void* dst, size_t dstSize,
|
| 612 | const void* src, size_t srcSize,
|
| 613 | const FSE_CTable* ct)
|
| 614 | {
|
| 615 | unsigned const fast = (dstSize >= FSE_BLOCKBOUND(srcSize));
|
| 616 |
|
| 617 | if (fast)
|
| 618 | return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 1);
|
| 619 | else
|
| 620 | return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 0);
|
| 621 | }
|
| 622 |
|
| 623 |
|
| 624 | size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); }
|
| 625 |
|
| 626 | #endif
|