1 /* SPDX-License-Identifier: GPL-2.0+ OR BSD-3-Clause */ 2 /* ****************************************************************** 3 * FSE : Finite State Entropy codec 4 * Public Prototypes declaration 5 * Copyright (c) Meta Platforms, Inc. and affiliates. 6 * 7 * You can contact the author at : 8 * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy 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 #ifndef FSE_H 16 #define FSE_H 17 18 19 /*-***************************************** 20 * Dependencies 21 ******************************************/ 22 #include "zstd_deps.h" /* size_t, ptrdiff_t */ 23 24 /*-***************************************** 25 * FSE_PUBLIC_API : control library symbols visibility 26 ******************************************/ 27 #if defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) && defined(__GNUC__) && (__GNUC__ >= 4) 28 # define FSE_PUBLIC_API __attribute__ ((visibility ("default"))) 29 #elif defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) /* Visual expected */ 30 # define FSE_PUBLIC_API __declspec(dllexport) 31 #elif defined(FSE_DLL_IMPORT) && (FSE_DLL_IMPORT==1) 32 # define FSE_PUBLIC_API __declspec(dllimport) /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/ 33 #else 34 # define FSE_PUBLIC_API 35 #endif 36 37 /*------ Version ------*/ 38 #define FSE_VERSION_MAJOR 0 39 #define FSE_VERSION_MINOR 9 40 #define FSE_VERSION_RELEASE 0 41 42 #define FSE_LIB_VERSION FSE_VERSION_MAJOR.FSE_VERSION_MINOR.FSE_VERSION_RELEASE 43 #define FSE_QUOTE(str) #str 44 #define FSE_EXPAND_AND_QUOTE(str) FSE_QUOTE(str) 45 #define FSE_VERSION_STRING FSE_EXPAND_AND_QUOTE(FSE_LIB_VERSION) 46 47 #define FSE_VERSION_NUMBER (FSE_VERSION_MAJOR *100*100 + FSE_VERSION_MINOR *100 + FSE_VERSION_RELEASE) 48 FSE_PUBLIC_API unsigned FSE_versionNumber(void); /*< library version number; to be used when checking dll version */ 49 50 51 /*-***************************************** 52 * Tool functions 53 ******************************************/ 54 FSE_PUBLIC_API size_t FSE_compressBound(size_t size); /* maximum compressed size */ 55 56 /* Error Management */ 57 FSE_PUBLIC_API unsigned FSE_isError(size_t code); /* tells if a return value is an error code */ 58 FSE_PUBLIC_API const char* FSE_getErrorName(size_t code); /* provides error code string (useful for debugging) */ 59 60 61 /*-***************************************** 62 * FSE detailed API 63 ******************************************/ 64 /*! 65 FSE_compress() does the following: 66 1. count symbol occurrence from source[] into table count[] (see hist.h) 67 2. normalize counters so that sum(count[]) == Power_of_2 (2^tableLog) 68 3. save normalized counters to memory buffer using writeNCount() 69 4. build encoding table 'CTable' from normalized counters 70 5. encode the data stream using encoding table 'CTable' 71 72 FSE_decompress() does the following: 73 1. read normalized counters with readNCount() 74 2. build decoding table 'DTable' from normalized counters 75 3. decode the data stream using decoding table 'DTable' 76 77 The following API allows targeting specific sub-functions for advanced tasks. 78 For example, it's possible to compress several blocks using the same 'CTable', 79 or to save and provide normalized distribution using external method. 80 */ 81 82 /* *** COMPRESSION *** */ 83 84 /*! FSE_optimalTableLog(): 85 dynamically downsize 'tableLog' when conditions are met. 86 It saves CPU time, by using smaller tables, while preserving or even improving compression ratio. 87 @return : recommended tableLog (necessarily <= 'maxTableLog') */ 88 FSE_PUBLIC_API unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue); 89 90 /*! FSE_normalizeCount(): 91 normalize counts so that sum(count[]) == Power_of_2 (2^tableLog) 92 'normalizedCounter' is a table of short, of minimum size (maxSymbolValue+1). 93 useLowProbCount is a boolean parameter which trades off compressed size for 94 faster header decoding. When it is set to 1, the compressed data will be slightly 95 smaller. And when it is set to 0, FSE_readNCount() and FSE_buildDTable() will be 96 faster. If you are compressing a small amount of data (< 2 KB) then useLowProbCount=0 97 is a good default, since header deserialization makes a big speed difference. 98 Otherwise, useLowProbCount=1 is a good default, since the speed difference is small. 99 @return : tableLog, 100 or an errorCode, which can be tested using FSE_isError() */ 101 FSE_PUBLIC_API size_t FSE_normalizeCount(short* normalizedCounter, unsigned tableLog, 102 const unsigned* count, size_t srcSize, unsigned maxSymbolValue, unsigned useLowProbCount); 103 104 /*! FSE_NCountWriteBound(): 105 Provides the maximum possible size of an FSE normalized table, given 'maxSymbolValue' and 'tableLog'. 106 Typically useful for allocation purpose. */ 107 FSE_PUBLIC_API size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog); 108 109 /*! FSE_writeNCount(): 110 Compactly save 'normalizedCounter' into 'buffer'. 111 @return : size of the compressed table, 112 or an errorCode, which can be tested using FSE_isError(). */ 113 FSE_PUBLIC_API size_t FSE_writeNCount (void* buffer, size_t bufferSize, 114 const short* normalizedCounter, 115 unsigned maxSymbolValue, unsigned tableLog); 116 117 /*! Constructor and Destructor of FSE_CTable. 118 Note that FSE_CTable size depends on 'tableLog' and 'maxSymbolValue' */ 119 typedef unsigned FSE_CTable; /* don't allocate that. It's only meant to be more restrictive than void* */ 120 121 /*! FSE_buildCTable(): 122 Builds `ct`, which must be already allocated, using FSE_createCTable(). 123 @return : 0, or an errorCode, which can be tested using FSE_isError() */ 124 FSE_PUBLIC_API size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog); 125 126 /*! FSE_compress_usingCTable(): 127 Compress `src` using `ct` into `dst` which must be already allocated. 128 @return : size of compressed data (<= `dstCapacity`), 129 or 0 if compressed data could not fit into `dst`, 130 or an errorCode, which can be tested using FSE_isError() */ 131 FSE_PUBLIC_API size_t FSE_compress_usingCTable (void* dst, size_t dstCapacity, const void* src, size_t srcSize, const FSE_CTable* ct); 132 133 /*! 134 Tutorial : 135 ---------- 136 The first step is to count all symbols. FSE_count() does this job very fast. 137 Result will be saved into 'count', a table of unsigned int, which must be already allocated, and have 'maxSymbolValuePtr[0]+1' cells. 138 'src' is a table of bytes of size 'srcSize'. All values within 'src' MUST be <= maxSymbolValuePtr[0] 139 maxSymbolValuePtr[0] will be updated, with its real value (necessarily <= original value) 140 FSE_count() will return the number of occurrence of the most frequent symbol. 141 This can be used to know if there is a single symbol within 'src', and to quickly evaluate its compressibility. 142 If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()). 143 144 The next step is to normalize the frequencies. 145 FSE_normalizeCount() will ensure that sum of frequencies is == 2 ^'tableLog'. 146 It also guarantees a minimum of 1 to any Symbol with frequency >= 1. 147 You can use 'tableLog'==0 to mean "use default tableLog value". 148 If you are unsure of which tableLog value to use, you can ask FSE_optimalTableLog(), 149 which will provide the optimal valid tableLog given sourceSize, maxSymbolValue, and a user-defined maximum (0 means "default"). 150 151 The result of FSE_normalizeCount() will be saved into a table, 152 called 'normalizedCounter', which is a table of signed short. 153 'normalizedCounter' must be already allocated, and have at least 'maxSymbolValue+1' cells. 154 The return value is tableLog if everything proceeded as expected. 155 It is 0 if there is a single symbol within distribution. 156 If there is an error (ex: invalid tableLog value), the function will return an ErrorCode (which can be tested using FSE_isError()). 157 158 'normalizedCounter' can be saved in a compact manner to a memory area using FSE_writeNCount(). 159 'buffer' must be already allocated. 160 For guaranteed success, buffer size must be at least FSE_headerBound(). 161 The result of the function is the number of bytes written into 'buffer'. 162 If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError(); ex : buffer size too small). 163 164 'normalizedCounter' can then be used to create the compression table 'CTable'. 165 The space required by 'CTable' must be already allocated, using FSE_createCTable(). 166 You can then use FSE_buildCTable() to fill 'CTable'. 167 If there is an error, both functions will return an ErrorCode (which can be tested using FSE_isError()). 168 169 'CTable' can then be used to compress 'src', with FSE_compress_usingCTable(). 170 Similar to FSE_count(), the convention is that 'src' is assumed to be a table of char of size 'srcSize' 171 The function returns the size of compressed data (without header), necessarily <= `dstCapacity`. 172 If it returns '0', compressed data could not fit into 'dst'. 173 If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()). 174 */ 175 176 177 /* *** DECOMPRESSION *** */ 178 179 /*! FSE_readNCount(): 180 Read compactly saved 'normalizedCounter' from 'rBuffer'. 181 @return : size read from 'rBuffer', 182 or an errorCode, which can be tested using FSE_isError(). 183 maxSymbolValuePtr[0] and tableLogPtr[0] will also be updated with their respective values */ 184 FSE_PUBLIC_API size_t FSE_readNCount (short* normalizedCounter, 185 unsigned* maxSymbolValuePtr, unsigned* tableLogPtr, 186 const void* rBuffer, size_t rBuffSize); 187 188 /*! FSE_readNCount_bmi2(): 189 * Same as FSE_readNCount() but pass bmi2=1 when your CPU supports BMI2 and 0 otherwise. 190 */ 191 FSE_PUBLIC_API size_t FSE_readNCount_bmi2(short* normalizedCounter, 192 unsigned* maxSymbolValuePtr, unsigned* tableLogPtr, 193 const void* rBuffer, size_t rBuffSize, int bmi2); 194 195 typedef unsigned FSE_DTable; /* don't allocate that. It's just a way to be more restrictive than void* */ 196 197 /*! 198 Tutorial : 199 ---------- 200 (Note : these functions only decompress FSE-compressed blocks. 201 If block is uncompressed, use memcpy() instead 202 If block is a single repeated byte, use memset() instead ) 203 204 The first step is to obtain the normalized frequencies of symbols. 205 This can be performed by FSE_readNCount() if it was saved using FSE_writeNCount(). 206 'normalizedCounter' must be already allocated, and have at least 'maxSymbolValuePtr[0]+1' cells of signed short. 207 In practice, that means it's necessary to know 'maxSymbolValue' beforehand, 208 or size the table to handle worst case situations (typically 256). 209 FSE_readNCount() will provide 'tableLog' and 'maxSymbolValue'. 210 The result of FSE_readNCount() is the number of bytes read from 'rBuffer'. 211 Note that 'rBufferSize' must be at least 4 bytes, even if useful information is less than that. 212 If there is an error, the function will return an error code, which can be tested using FSE_isError(). 213 214 The next step is to build the decompression tables 'FSE_DTable' from 'normalizedCounter'. 215 This is performed by the function FSE_buildDTable(). 216 The space required by 'FSE_DTable' must be already allocated using FSE_createDTable(). 217 If there is an error, the function will return an error code, which can be tested using FSE_isError(). 218 219 `FSE_DTable` can then be used to decompress `cSrc`, with FSE_decompress_usingDTable(). 220 `cSrcSize` must be strictly correct, otherwise decompression will fail. 221 FSE_decompress_usingDTable() result will tell how many bytes were regenerated (<=`dstCapacity`). 222 If there is an error, the function will return an error code, which can be tested using FSE_isError(). (ex: dst buffer too small) 223 */ 224 225 #endif /* FSE_H */ 226 227 228 #if !defined(FSE_H_FSE_STATIC_LINKING_ONLY) 229 #define FSE_H_FSE_STATIC_LINKING_ONLY 230 #include "bitstream.h" 231 232 /* ***************************************** 233 * Static allocation 234 *******************************************/ 235 /* FSE buffer bounds */ 236 #define FSE_NCOUNTBOUND 512 237 #define FSE_BLOCKBOUND(size) ((size) + ((size)>>7) + 4 /* fse states */ + sizeof(size_t) /* bitContainer */) 238 #define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size)) /* Macro version, useful for static allocation */ 239 240 /* It is possible to statically allocate FSE CTable/DTable as a table of FSE_CTable/FSE_DTable using below macros */ 241 #define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) (1 + (1<<((maxTableLog)-1)) + (((maxSymbolValue)+1)*2)) 242 #define FSE_DTABLE_SIZE_U32(maxTableLog) (1 + (1<<(maxTableLog))) 243 244 /* or use the size to malloc() space directly. Pay attention to alignment restrictions though */ 245 #define FSE_CTABLE_SIZE(maxTableLog, maxSymbolValue) (FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(FSE_CTable)) 246 #define FSE_DTABLE_SIZE(maxTableLog) (FSE_DTABLE_SIZE_U32(maxTableLog) * sizeof(FSE_DTable)) 247 248 249 /* ***************************************** 250 * FSE advanced API 251 ***************************************** */ 252 253 unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus); 254 /*< same as FSE_optimalTableLog(), which used `minus==2` */ 255 256 size_t FSE_buildCTable_rle (FSE_CTable* ct, unsigned char symbolValue); 257 /*< build a fake FSE_CTable, designed to compress always the same symbolValue */ 258 259 /* FSE_buildCTable_wksp() : 260 * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`). 261 * `wkspSize` must be >= `FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog)` of `unsigned`. 262 * See FSE_buildCTable_wksp() for breakdown of workspace usage. 263 */ 264 #define FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog) (((maxSymbolValue + 2) + (1ull << (tableLog)))/2 + sizeof(U64)/sizeof(U32) /* additional 8 bytes for potential table overwrite */) 265 #define FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) (sizeof(unsigned) * FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog)) 266 size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize); 267 268 #define FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) (sizeof(short) * (maxSymbolValue + 1) + (1ULL << maxTableLog) + 8) 269 #define FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) ((FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) + sizeof(unsigned) - 1) / sizeof(unsigned)) 270 FSE_PUBLIC_API size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize); 271 /*< Same as FSE_buildDTable(), using an externally allocated `workspace` produced with `FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxSymbolValue)` */ 272 273 #define FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) (FSE_DTABLE_SIZE_U32(maxTableLog) + 1 + FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) + (FSE_MAX_SYMBOL_VALUE + 1) / 2 + 1) 274 #define FSE_DECOMPRESS_WKSP_SIZE(maxTableLog, maxSymbolValue) (FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(unsigned)) 275 size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2); 276 /*< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DECOMPRESS_WKSP_SIZE_U32(maxLog, maxSymbolValue)`. 277 * Set bmi2 to 1 if your CPU supports BMI2 or 0 if it doesn't */ 278 279 typedef enum { 280 FSE_repeat_none, /*< Cannot use the previous table */ 281 FSE_repeat_check, /*< Can use the previous table but it must be checked */ 282 FSE_repeat_valid /*< Can use the previous table and it is assumed to be valid */ 283 } FSE_repeat; 284 285 /* ***************************************** 286 * FSE symbol compression API 287 *******************************************/ 288 /*! 289 This API consists of small unitary functions, which highly benefit from being inlined. 290 Hence their body are included in next section. 291 */ 292 typedef struct { 293 ptrdiff_t value; 294 const void* stateTable; 295 const void* symbolTT; 296 unsigned stateLog; 297 } FSE_CState_t; 298 299 static void FSE_initCState(FSE_CState_t* CStatePtr, const FSE_CTable* ct); 300 301 static void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* CStatePtr, unsigned symbol); 302 303 static void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* CStatePtr); 304 305 /*< 306 These functions are inner components of FSE_compress_usingCTable(). 307 They allow the creation of custom streams, mixing multiple tables and bit sources. 308 309 A key property to keep in mind is that encoding and decoding are done **in reverse direction**. 310 So the first symbol you will encode is the last you will decode, like a LIFO stack. 311 312 You will need a few variables to track your CStream. They are : 313 314 FSE_CTable ct; // Provided by FSE_buildCTable() 315 BIT_CStream_t bitStream; // bitStream tracking structure 316 FSE_CState_t state; // State tracking structure (can have several) 317 318 319 The first thing to do is to init bitStream and state. 320 size_t errorCode = BIT_initCStream(&bitStream, dstBuffer, maxDstSize); 321 FSE_initCState(&state, ct); 322 323 Note that BIT_initCStream() can produce an error code, so its result should be tested, using FSE_isError(); 324 You can then encode your input data, byte after byte. 325 FSE_encodeSymbol() outputs a maximum of 'tableLog' bits at a time. 326 Remember decoding will be done in reverse direction. 327 FSE_encodeByte(&bitStream, &state, symbol); 328 329 At any time, you can also add any bit sequence. 330 Note : maximum allowed nbBits is 25, for compatibility with 32-bits decoders 331 BIT_addBits(&bitStream, bitField, nbBits); 332 333 The above methods don't commit data to memory, they just store it into local register, for speed. 334 Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (size_t). 335 Writing data to memory is a manual operation, performed by the flushBits function. 336 BIT_flushBits(&bitStream); 337 338 Your last FSE encoding operation shall be to flush your last state value(s). 339 FSE_flushState(&bitStream, &state); 340 341 Finally, you must close the bitStream. 342 The function returns the size of CStream in bytes. 343 If data couldn't fit into dstBuffer, it will return a 0 ( == not compressible) 344 If there is an error, it returns an errorCode (which can be tested using FSE_isError()). 345 size_t size = BIT_closeCStream(&bitStream); 346 */ 347 348 349 /* ***************************************** 350 * FSE symbol decompression API 351 *******************************************/ 352 typedef struct { 353 size_t state; 354 const void* table; /* precise table may vary, depending on U16 */ 355 } FSE_DState_t; 356 357 358 static void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt); 359 360 static unsigned char FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD); 361 362 static unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr); 363 364 /*< 365 Let's now decompose FSE_decompress_usingDTable() into its unitary components. 366 You will decode FSE-encoded symbols from the bitStream, 367 and also any other bitFields you put in, **in reverse order**. 368 369 You will need a few variables to track your bitStream. They are : 370 371 BIT_DStream_t DStream; // Stream context 372 FSE_DState_t DState; // State context. Multiple ones are possible 373 FSE_DTable* DTablePtr; // Decoding table, provided by FSE_buildDTable() 374 375 The first thing to do is to init the bitStream. 376 errorCode = BIT_initDStream(&DStream, srcBuffer, srcSize); 377 378 You should then retrieve your initial state(s) 379 (in reverse flushing order if you have several ones) : 380 errorCode = FSE_initDState(&DState, &DStream, DTablePtr); 381 382 You can then decode your data, symbol after symbol. 383 For information the maximum number of bits read by FSE_decodeSymbol() is 'tableLog'. 384 Keep in mind that symbols are decoded in reverse order, like a LIFO stack (last in, first out). 385 unsigned char symbol = FSE_decodeSymbol(&DState, &DStream); 386 387 You can retrieve any bitfield you eventually stored into the bitStream (in reverse order) 388 Note : maximum allowed nbBits is 25, for 32-bits compatibility 389 size_t bitField = BIT_readBits(&DStream, nbBits); 390 391 All above operations only read from local register (which size depends on size_t). 392 Refueling the register from memory is manually performed by the reload method. 393 endSignal = FSE_reloadDStream(&DStream); 394 395 BIT_reloadDStream() result tells if there is still some more data to read from DStream. 396 BIT_DStream_unfinished : there is still some data left into the DStream. 397 BIT_DStream_endOfBuffer : Dstream reached end of buffer. Its container may no longer be completely filled. 398 BIT_DStream_completed : Dstream reached its exact end, corresponding in general to decompression completed. 399 BIT_DStream_tooFar : Dstream went too far. Decompression result is corrupted. 400 401 When reaching end of buffer (BIT_DStream_endOfBuffer), progress slowly, notably if you decode multiple symbols per loop, 402 to properly detect the exact end of stream. 403 After each decoded symbol, check if DStream is fully consumed using this simple test : 404 BIT_reloadDStream(&DStream) >= BIT_DStream_completed 405 406 When it's done, verify decompression is fully completed, by checking both DStream and the relevant states. 407 Checking if DStream has reached its end is performed by : 408 BIT_endOfDStream(&DStream); 409 Check also the states. There might be some symbols left there, if some high probability ones (>50%) are possible. 410 FSE_endOfDState(&DState); 411 */ 412 413 414 /* ***************************************** 415 * FSE unsafe API 416 *******************************************/ 417 static unsigned char FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD); 418 /* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */ 419 420 421 /* ***************************************** 422 * Implementation of inlined functions 423 *******************************************/ 424 typedef struct { 425 int deltaFindState; 426 U32 deltaNbBits; 427 } FSE_symbolCompressionTransform; /* total 8 bytes */ 428 429 MEM_STATIC void FSE_initCState(FSE_CState_t* statePtr, const FSE_CTable* ct) 430 { 431 const void* ptr = ct; 432 const U16* u16ptr = (const U16*) ptr; 433 const U32 tableLog = MEM_read16(ptr); 434 statePtr->value = (ptrdiff_t)1<<tableLog; 435 statePtr->stateTable = u16ptr+2; 436 statePtr->symbolTT = ct + 1 + (tableLog ? (1<<(tableLog-1)) : 1); 437 statePtr->stateLog = tableLog; 438 } 439 440 441 /*! FSE_initCState2() : 442 * Same as FSE_initCState(), but the first symbol to include (which will be the last to be read) 443 * uses the smallest state value possible, saving the cost of this symbol */ 444 MEM_STATIC void FSE_initCState2(FSE_CState_t* statePtr, const FSE_CTable* ct, U32 symbol) 445 { 446 FSE_initCState(statePtr, ct); 447 { const FSE_symbolCompressionTransform symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol]; 448 const U16* stateTable = (const U16*)(statePtr->stateTable); 449 U32 nbBitsOut = (U32)((symbolTT.deltaNbBits + (1<<15)) >> 16); 450 statePtr->value = (nbBitsOut << 16) - symbolTT.deltaNbBits; 451 statePtr->value = stateTable[(statePtr->value >> nbBitsOut) + symbolTT.deltaFindState]; 452 } 453 } 454 455 MEM_STATIC void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* statePtr, unsigned symbol) 456 { 457 FSE_symbolCompressionTransform const symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol]; 458 const U16* const stateTable = (const U16*)(statePtr->stateTable); 459 U32 const nbBitsOut = (U32)((statePtr->value + symbolTT.deltaNbBits) >> 16); 460 BIT_addBits(bitC, (BitContainerType)statePtr->value, nbBitsOut); 461 statePtr->value = stateTable[ (statePtr->value >> nbBitsOut) + symbolTT.deltaFindState]; 462 } 463 464 MEM_STATIC void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* statePtr) 465 { 466 BIT_addBits(bitC, (BitContainerType)statePtr->value, statePtr->stateLog); 467 BIT_flushBits(bitC); 468 } 469 470 471 /* FSE_getMaxNbBits() : 472 * Approximate maximum cost of a symbol, in bits. 473 * Fractional get rounded up (i.e. a symbol with a normalized frequency of 3 gives the same result as a frequency of 2) 474 * note 1 : assume symbolValue is valid (<= maxSymbolValue) 475 * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */ 476 MEM_STATIC U32 FSE_getMaxNbBits(const void* symbolTTPtr, U32 symbolValue) 477 { 478 const FSE_symbolCompressionTransform* symbolTT = (const FSE_symbolCompressionTransform*) symbolTTPtr; 479 return (symbolTT[symbolValue].deltaNbBits + ((1<<16)-1)) >> 16; 480 } 481 482 /* FSE_bitCost() : 483 * Approximate symbol cost, as fractional value, using fixed-point format (accuracyLog fractional bits) 484 * note 1 : assume symbolValue is valid (<= maxSymbolValue) 485 * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */ 486 MEM_STATIC U32 FSE_bitCost(const void* symbolTTPtr, U32 tableLog, U32 symbolValue, U32 accuracyLog) 487 { 488 const FSE_symbolCompressionTransform* symbolTT = (const FSE_symbolCompressionTransform*) symbolTTPtr; 489 U32 const minNbBits = symbolTT[symbolValue].deltaNbBits >> 16; 490 U32 const threshold = (minNbBits+1) << 16; 491 assert(tableLog < 16); 492 assert(accuracyLog < 31-tableLog); /* ensure enough room for renormalization double shift */ 493 { U32 const tableSize = 1 << tableLog; 494 U32 const deltaFromThreshold = threshold - (symbolTT[symbolValue].deltaNbBits + tableSize); 495 U32 const normalizedDeltaFromThreshold = (deltaFromThreshold << accuracyLog) >> tableLog; /* linear interpolation (very approximate) */ 496 U32 const bitMultiplier = 1 << accuracyLog; 497 assert(symbolTT[symbolValue].deltaNbBits + tableSize <= threshold); 498 assert(normalizedDeltaFromThreshold <= bitMultiplier); 499 return (minNbBits+1)*bitMultiplier - normalizedDeltaFromThreshold; 500 } 501 } 502 503 504 /* ====== Decompression ====== */ 505 506 typedef struct { 507 U16 tableLog; 508 U16 fastMode; 509 } FSE_DTableHeader; /* sizeof U32 */ 510 511 typedef struct 512 { 513 unsigned short newState; 514 unsigned char symbol; 515 unsigned char nbBits; 516 } FSE_decode_t; /* size == U32 */ 517 518 MEM_STATIC void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt) 519 { 520 const void* ptr = dt; 521 const FSE_DTableHeader* const DTableH = (const FSE_DTableHeader*)ptr; 522 DStatePtr->state = BIT_readBits(bitD, DTableH->tableLog); 523 BIT_reloadDStream(bitD); 524 DStatePtr->table = dt + 1; 525 } 526 527 MEM_STATIC BYTE FSE_peekSymbol(const FSE_DState_t* DStatePtr) 528 { 529 FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state]; 530 return DInfo.symbol; 531 } 532 533 MEM_STATIC void FSE_updateState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD) 534 { 535 FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state]; 536 U32 const nbBits = DInfo.nbBits; 537 size_t const lowBits = BIT_readBits(bitD, nbBits); 538 DStatePtr->state = DInfo.newState + lowBits; 539 } 540 541 MEM_STATIC BYTE FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD) 542 { 543 FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state]; 544 U32 const nbBits = DInfo.nbBits; 545 BYTE const symbol = DInfo.symbol; 546 size_t const lowBits = BIT_readBits(bitD, nbBits); 547 548 DStatePtr->state = DInfo.newState + lowBits; 549 return symbol; 550 } 551 552 /*! FSE_decodeSymbolFast() : 553 unsafe, only works if no symbol has a probability > 50% */ 554 MEM_STATIC BYTE FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD) 555 { 556 FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state]; 557 U32 const nbBits = DInfo.nbBits; 558 BYTE const symbol = DInfo.symbol; 559 size_t const lowBits = BIT_readBitsFast(bitD, nbBits); 560 561 DStatePtr->state = DInfo.newState + lowBits; 562 return symbol; 563 } 564 565 MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr) 566 { 567 return DStatePtr->state == 0; 568 } 569 570 571 572 #ifndef FSE_COMMONDEFS_ONLY 573 574 /* ************************************************************** 575 * Tuning parameters 576 ****************************************************************/ 577 /*!MEMORY_USAGE : 578 * Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.) 579 * Increasing memory usage improves compression ratio 580 * Reduced memory usage can improve speed, due to cache effect 581 * Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */ 582 #ifndef FSE_MAX_MEMORY_USAGE 583 # define FSE_MAX_MEMORY_USAGE 14 584 #endif 585 #ifndef FSE_DEFAULT_MEMORY_USAGE 586 # define FSE_DEFAULT_MEMORY_USAGE 13 587 #endif 588 #if (FSE_DEFAULT_MEMORY_USAGE > FSE_MAX_MEMORY_USAGE) 589 # error "FSE_DEFAULT_MEMORY_USAGE must be <= FSE_MAX_MEMORY_USAGE" 590 #endif 591 592 /*!FSE_MAX_SYMBOL_VALUE : 593 * Maximum symbol value authorized. 594 * Required for proper stack allocation */ 595 #ifndef FSE_MAX_SYMBOL_VALUE 596 # define FSE_MAX_SYMBOL_VALUE 255 597 #endif 598 599 /* ************************************************************** 600 * template functions type & suffix 601 ****************************************************************/ 602 #define FSE_FUNCTION_TYPE BYTE 603 #define FSE_FUNCTION_EXTENSION 604 #define FSE_DECODE_TYPE FSE_decode_t 605 606 607 #endif /* !FSE_COMMONDEFS_ONLY */ 608 609 610 /* *************************************************************** 611 * Constants 612 *****************************************************************/ 613 #define FSE_MAX_TABLELOG (FSE_MAX_MEMORY_USAGE-2) 614 #define FSE_MAX_TABLESIZE (1U<<FSE_MAX_TABLELOG) 615 #define FSE_MAXTABLESIZE_MASK (FSE_MAX_TABLESIZE-1) 616 #define FSE_DEFAULT_TABLELOG (FSE_DEFAULT_MEMORY_USAGE-2) 617 #define FSE_MIN_TABLELOG 5 618 619 #define FSE_TABLELOG_ABSOLUTE_MAX 15 620 #if FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX 621 # error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported" 622 #endif 623 624 #define FSE_TABLESTEP(tableSize) (((tableSize)>>1) + ((tableSize)>>3) + 3) 625 626 #endif /* FSE_STATIC_LINKING_ONLY */ 627