/* ******************************************************************1* FSE : Finite State Entropy codec2* Public Prototypes declaration3* Copyright (c) Meta Platforms, Inc. and affiliates.4*5* You can contact the author at :6* - Source repository : https://github.com/Cyan4973/FiniteStateEntropy7*8* This source code is licensed under both the BSD-style license (found in the9* LICENSE file in the root directory of this source tree) and the GPLv2 (found10* in the COPYING file in the root directory of this source tree).11* You may select, at your option, one of the above-listed licenses.12****************************************************************** */1314#if defined (__cplusplus)15extern "C" {16#endif1718#ifndef FSE_H19#define FSE_H202122/*-*****************************************23* Dependencies24******************************************/25#include "zstd_deps.h" /* size_t, ptrdiff_t */262728/*-*****************************************29* FSE_PUBLIC_API : control library symbols visibility30******************************************/31#if defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) && defined(__GNUC__) && (__GNUC__ >= 4)32# define FSE_PUBLIC_API __attribute__ ((visibility ("default")))33#elif defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) /* Visual expected */34# define FSE_PUBLIC_API __declspec(dllexport)35#elif defined(FSE_DLL_IMPORT) && (FSE_DLL_IMPORT==1)36# 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.*/37#else38# define FSE_PUBLIC_API39#endif4041/*------ Version ------*/42#define FSE_VERSION_MAJOR 043#define FSE_VERSION_MINOR 944#define FSE_VERSION_RELEASE 04546#define FSE_LIB_VERSION FSE_VERSION_MAJOR.FSE_VERSION_MINOR.FSE_VERSION_RELEASE47#define FSE_QUOTE(str) #str48#define FSE_EXPAND_AND_QUOTE(str) FSE_QUOTE(str)49#define FSE_VERSION_STRING FSE_EXPAND_AND_QUOTE(FSE_LIB_VERSION)5051#define FSE_VERSION_NUMBER (FSE_VERSION_MAJOR *100*100 + FSE_VERSION_MINOR *100 + FSE_VERSION_RELEASE)52FSE_PUBLIC_API unsigned FSE_versionNumber(void); /**< library version number; to be used when checking dll version */535455/*-*****************************************56* Tool functions57******************************************/58FSE_PUBLIC_API size_t FSE_compressBound(size_t size); /* maximum compressed size */5960/* Error Management */61FSE_PUBLIC_API unsigned FSE_isError(size_t code); /* tells if a return value is an error code */62FSE_PUBLIC_API const char* FSE_getErrorName(size_t code); /* provides error code string (useful for debugging) */636465/*-*****************************************66* FSE detailed API67******************************************/68/*!69FSE_compress() does the following:701. count symbol occurrence from source[] into table count[] (see hist.h)712. normalize counters so that sum(count[]) == Power_of_2 (2^tableLog)723. save normalized counters to memory buffer using writeNCount()734. build encoding table 'CTable' from normalized counters745. encode the data stream using encoding table 'CTable'7576FSE_decompress() does the following:771. read normalized counters with readNCount()782. build decoding table 'DTable' from normalized counters793. decode the data stream using decoding table 'DTable'8081The following API allows targeting specific sub-functions for advanced tasks.82For example, it's possible to compress several blocks using the same 'CTable',83or to save and provide normalized distribution using external method.84*/8586/* *** COMPRESSION *** */8788/*! FSE_optimalTableLog():89dynamically downsize 'tableLog' when conditions are met.90It saves CPU time, by using smaller tables, while preserving or even improving compression ratio.91@return : recommended tableLog (necessarily <= 'maxTableLog') */92FSE_PUBLIC_API unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue);9394/*! FSE_normalizeCount():95normalize counts so that sum(count[]) == Power_of_2 (2^tableLog)96'normalizedCounter' is a table of short, of minimum size (maxSymbolValue+1).97useLowProbCount is a boolean parameter which trades off compressed size for98faster header decoding. When it is set to 1, the compressed data will be slightly99smaller. And when it is set to 0, FSE_readNCount() and FSE_buildDTable() will be100faster. If you are compressing a small amount of data (< 2 KB) then useLowProbCount=0101is a good default, since header deserialization makes a big speed difference.102Otherwise, useLowProbCount=1 is a good default, since the speed difference is small.103@return : tableLog,104or an errorCode, which can be tested using FSE_isError() */105FSE_PUBLIC_API size_t FSE_normalizeCount(short* normalizedCounter, unsigned tableLog,106const unsigned* count, size_t srcSize, unsigned maxSymbolValue, unsigned useLowProbCount);107108/*! FSE_NCountWriteBound():109Provides the maximum possible size of an FSE normalized table, given 'maxSymbolValue' and 'tableLog'.110Typically useful for allocation purpose. */111FSE_PUBLIC_API size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog);112113/*! FSE_writeNCount():114Compactly save 'normalizedCounter' into 'buffer'.115@return : size of the compressed table,116or an errorCode, which can be tested using FSE_isError(). */117FSE_PUBLIC_API size_t FSE_writeNCount (void* buffer, size_t bufferSize,118const short* normalizedCounter,119unsigned maxSymbolValue, unsigned tableLog);120121/*! Constructor and Destructor of FSE_CTable.122Note that FSE_CTable size depends on 'tableLog' and 'maxSymbolValue' */123typedef unsigned FSE_CTable; /* don't allocate that. It's only meant to be more restrictive than void* */124125/*! FSE_buildCTable():126Builds `ct`, which must be already allocated, using FSE_createCTable().127@return : 0, or an errorCode, which can be tested using FSE_isError() */128FSE_PUBLIC_API size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);129130/*! FSE_compress_usingCTable():131Compress `src` using `ct` into `dst` which must be already allocated.132@return : size of compressed data (<= `dstCapacity`),133or 0 if compressed data could not fit into `dst`,134or an errorCode, which can be tested using FSE_isError() */135FSE_PUBLIC_API size_t FSE_compress_usingCTable (void* dst, size_t dstCapacity, const void* src, size_t srcSize, const FSE_CTable* ct);136137/*!138Tutorial :139----------140The first step is to count all symbols. FSE_count() does this job very fast.141Result will be saved into 'count', a table of unsigned int, which must be already allocated, and have 'maxSymbolValuePtr[0]+1' cells.142'src' is a table of bytes of size 'srcSize'. All values within 'src' MUST be <= maxSymbolValuePtr[0]143maxSymbolValuePtr[0] will be updated, with its real value (necessarily <= original value)144FSE_count() will return the number of occurrence of the most frequent symbol.145This can be used to know if there is a single symbol within 'src', and to quickly evaluate its compressibility.146If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()).147148The next step is to normalize the frequencies.149FSE_normalizeCount() will ensure that sum of frequencies is == 2 ^'tableLog'.150It also guarantees a minimum of 1 to any Symbol with frequency >= 1.151You can use 'tableLog'==0 to mean "use default tableLog value".152If you are unsure of which tableLog value to use, you can ask FSE_optimalTableLog(),153which will provide the optimal valid tableLog given sourceSize, maxSymbolValue, and a user-defined maximum (0 means "default").154155The result of FSE_normalizeCount() will be saved into a table,156called 'normalizedCounter', which is a table of signed short.157'normalizedCounter' must be already allocated, and have at least 'maxSymbolValue+1' cells.158The return value is tableLog if everything proceeded as expected.159It is 0 if there is a single symbol within distribution.160If there is an error (ex: invalid tableLog value), the function will return an ErrorCode (which can be tested using FSE_isError()).161162'normalizedCounter' can be saved in a compact manner to a memory area using FSE_writeNCount().163'buffer' must be already allocated.164For guaranteed success, buffer size must be at least FSE_headerBound().165The result of the function is the number of bytes written into 'buffer'.166If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError(); ex : buffer size too small).167168'normalizedCounter' can then be used to create the compression table 'CTable'.169The space required by 'CTable' must be already allocated, using FSE_createCTable().170You can then use FSE_buildCTable() to fill 'CTable'.171If there is an error, both functions will return an ErrorCode (which can be tested using FSE_isError()).172173'CTable' can then be used to compress 'src', with FSE_compress_usingCTable().174Similar to FSE_count(), the convention is that 'src' is assumed to be a table of char of size 'srcSize'175The function returns the size of compressed data (without header), necessarily <= `dstCapacity`.176If it returns '0', compressed data could not fit into 'dst'.177If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()).178*/179180181/* *** DECOMPRESSION *** */182183/*! FSE_readNCount():184Read compactly saved 'normalizedCounter' from 'rBuffer'.185@return : size read from 'rBuffer',186or an errorCode, which can be tested using FSE_isError().187maxSymbolValuePtr[0] and tableLogPtr[0] will also be updated with their respective values */188FSE_PUBLIC_API size_t FSE_readNCount (short* normalizedCounter,189unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,190const void* rBuffer, size_t rBuffSize);191192/*! FSE_readNCount_bmi2():193* Same as FSE_readNCount() but pass bmi2=1 when your CPU supports BMI2 and 0 otherwise.194*/195FSE_PUBLIC_API size_t FSE_readNCount_bmi2(short* normalizedCounter,196unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,197const void* rBuffer, size_t rBuffSize, int bmi2);198199typedef unsigned FSE_DTable; /* don't allocate that. It's just a way to be more restrictive than void* */200201/*!202Tutorial :203----------204(Note : these functions only decompress FSE-compressed blocks.205If block is uncompressed, use memcpy() instead206If block is a single repeated byte, use memset() instead )207208The first step is to obtain the normalized frequencies of symbols.209This can be performed by FSE_readNCount() if it was saved using FSE_writeNCount().210'normalizedCounter' must be already allocated, and have at least 'maxSymbolValuePtr[0]+1' cells of signed short.211In practice, that means it's necessary to know 'maxSymbolValue' beforehand,212or size the table to handle worst case situations (typically 256).213FSE_readNCount() will provide 'tableLog' and 'maxSymbolValue'.214The result of FSE_readNCount() is the number of bytes read from 'rBuffer'.215Note that 'rBufferSize' must be at least 4 bytes, even if useful information is less than that.216If there is an error, the function will return an error code, which can be tested using FSE_isError().217218The next step is to build the decompression tables 'FSE_DTable' from 'normalizedCounter'.219This is performed by the function FSE_buildDTable().220The space required by 'FSE_DTable' must be already allocated using FSE_createDTable().221If there is an error, the function will return an error code, which can be tested using FSE_isError().222223`FSE_DTable` can then be used to decompress `cSrc`, with FSE_decompress_usingDTable().224`cSrcSize` must be strictly correct, otherwise decompression will fail.225FSE_decompress_usingDTable() result will tell how many bytes were regenerated (<=`dstCapacity`).226If there is an error, the function will return an error code, which can be tested using FSE_isError(). (ex: dst buffer too small)227*/228229#endif /* FSE_H */230231#if defined(FSE_STATIC_LINKING_ONLY) && !defined(FSE_H_FSE_STATIC_LINKING_ONLY)232#define FSE_H_FSE_STATIC_LINKING_ONLY233234/* *** Dependency *** */235#include "bitstream.h"236237238/* *****************************************239* Static allocation240*******************************************/241/* FSE buffer bounds */242#define FSE_NCOUNTBOUND 512243#define FSE_BLOCKBOUND(size) ((size) + ((size)>>7) + 4 /* fse states */ + sizeof(size_t) /* bitContainer */)244#define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size)) /* Macro version, useful for static allocation */245246/* It is possible to statically allocate FSE CTable/DTable as a table of FSE_CTable/FSE_DTable using below macros */247#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) (1 + (1<<((maxTableLog)-1)) + (((maxSymbolValue)+1)*2))248#define FSE_DTABLE_SIZE_U32(maxTableLog) (1 + (1<<(maxTableLog)))249250/* or use the size to malloc() space directly. Pay attention to alignment restrictions though */251#define FSE_CTABLE_SIZE(maxTableLog, maxSymbolValue) (FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(FSE_CTable))252#define FSE_DTABLE_SIZE(maxTableLog) (FSE_DTABLE_SIZE_U32(maxTableLog) * sizeof(FSE_DTable))253254255/* *****************************************256* FSE advanced API257***************************************** */258259unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus);260/**< same as FSE_optimalTableLog(), which used `minus==2` */261262size_t FSE_buildCTable_rle (FSE_CTable* ct, unsigned char symbolValue);263/**< build a fake FSE_CTable, designed to compress always the same symbolValue */264265/* FSE_buildCTable_wksp() :266* Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).267* `wkspSize` must be >= `FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog)` of `unsigned`.268* See FSE_buildCTable_wksp() for breakdown of workspace usage.269*/270#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 */)271#define FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) (sizeof(unsigned) * FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog))272size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);273274#define FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) (sizeof(short) * (maxSymbolValue + 1) + (1ULL << maxTableLog) + 8)275#define FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) ((FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) + sizeof(unsigned) - 1) / sizeof(unsigned))276FSE_PUBLIC_API size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);277/**< Same as FSE_buildDTable(), using an externally allocated `workspace` produced with `FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxSymbolValue)` */278279#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)280#define FSE_DECOMPRESS_WKSP_SIZE(maxTableLog, maxSymbolValue) (FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(unsigned))281size_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);282/**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DECOMPRESS_WKSP_SIZE_U32(maxLog, maxSymbolValue)`.283* Set bmi2 to 1 if your CPU supports BMI2 or 0 if it doesn't */284285typedef enum {286FSE_repeat_none, /**< Cannot use the previous table */287FSE_repeat_check, /**< Can use the previous table but it must be checked */288FSE_repeat_valid /**< Can use the previous table and it is assumed to be valid */289} FSE_repeat;290291/* *****************************************292* FSE symbol compression API293*******************************************/294/*!295This API consists of small unitary functions, which highly benefit from being inlined.296Hence their body are included in next section.297*/298typedef struct {299ptrdiff_t value;300const void* stateTable;301const void* symbolTT;302unsigned stateLog;303} FSE_CState_t;304305static void FSE_initCState(FSE_CState_t* CStatePtr, const FSE_CTable* ct);306307static void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* CStatePtr, unsigned symbol);308309static void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* CStatePtr);310311/**<312These functions are inner components of FSE_compress_usingCTable().313They allow the creation of custom streams, mixing multiple tables and bit sources.314315A key property to keep in mind is that encoding and decoding are done **in reverse direction**.316So the first symbol you will encode is the last you will decode, like a LIFO stack.317318You will need a few variables to track your CStream. They are :319320FSE_CTable ct; // Provided by FSE_buildCTable()321BIT_CStream_t bitStream; // bitStream tracking structure322FSE_CState_t state; // State tracking structure (can have several)323324325The first thing to do is to init bitStream and state.326size_t errorCode = BIT_initCStream(&bitStream, dstBuffer, maxDstSize);327FSE_initCState(&state, ct);328329Note that BIT_initCStream() can produce an error code, so its result should be tested, using FSE_isError();330You can then encode your input data, byte after byte.331FSE_encodeSymbol() outputs a maximum of 'tableLog' bits at a time.332Remember decoding will be done in reverse direction.333FSE_encodeByte(&bitStream, &state, symbol);334335At any time, you can also add any bit sequence.336Note : maximum allowed nbBits is 25, for compatibility with 32-bits decoders337BIT_addBits(&bitStream, bitField, nbBits);338339The above methods don't commit data to memory, they just store it into local register, for speed.340Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (size_t).341Writing data to memory is a manual operation, performed by the flushBits function.342BIT_flushBits(&bitStream);343344Your last FSE encoding operation shall be to flush your last state value(s).345FSE_flushState(&bitStream, &state);346347Finally, you must close the bitStream.348The function returns the size of CStream in bytes.349If data couldn't fit into dstBuffer, it will return a 0 ( == not compressible)350If there is an error, it returns an errorCode (which can be tested using FSE_isError()).351size_t size = BIT_closeCStream(&bitStream);352*/353354355/* *****************************************356* FSE symbol decompression API357*******************************************/358typedef struct {359size_t state;360const void* table; /* precise table may vary, depending on U16 */361} FSE_DState_t;362363364static void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt);365366static unsigned char FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);367368static unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr);369370/**<371Let's now decompose FSE_decompress_usingDTable() into its unitary components.372You will decode FSE-encoded symbols from the bitStream,373and also any other bitFields you put in, **in reverse order**.374375You will need a few variables to track your bitStream. They are :376377BIT_DStream_t DStream; // Stream context378FSE_DState_t DState; // State context. Multiple ones are possible379FSE_DTable* DTablePtr; // Decoding table, provided by FSE_buildDTable()380381The first thing to do is to init the bitStream.382errorCode = BIT_initDStream(&DStream, srcBuffer, srcSize);383384You should then retrieve your initial state(s)385(in reverse flushing order if you have several ones) :386errorCode = FSE_initDState(&DState, &DStream, DTablePtr);387388You can then decode your data, symbol after symbol.389For information the maximum number of bits read by FSE_decodeSymbol() is 'tableLog'.390Keep in mind that symbols are decoded in reverse order, like a LIFO stack (last in, first out).391unsigned char symbol = FSE_decodeSymbol(&DState, &DStream);392393You can retrieve any bitfield you eventually stored into the bitStream (in reverse order)394Note : maximum allowed nbBits is 25, for 32-bits compatibility395size_t bitField = BIT_readBits(&DStream, nbBits);396397All above operations only read from local register (which size depends on size_t).398Refueling the register from memory is manually performed by the reload method.399endSignal = FSE_reloadDStream(&DStream);400401BIT_reloadDStream() result tells if there is still some more data to read from DStream.402BIT_DStream_unfinished : there is still some data left into the DStream.403BIT_DStream_endOfBuffer : Dstream reached end of buffer. Its container may no longer be completely filled.404BIT_DStream_completed : Dstream reached its exact end, corresponding in general to decompression completed.405BIT_DStream_tooFar : Dstream went too far. Decompression result is corrupted.406407When reaching end of buffer (BIT_DStream_endOfBuffer), progress slowly, notably if you decode multiple symbols per loop,408to properly detect the exact end of stream.409After each decoded symbol, check if DStream is fully consumed using this simple test :410BIT_reloadDStream(&DStream) >= BIT_DStream_completed411412When it's done, verify decompression is fully completed, by checking both DStream and the relevant states.413Checking if DStream has reached its end is performed by :414BIT_endOfDStream(&DStream);415Check also the states. There might be some symbols left there, if some high probability ones (>50%) are possible.416FSE_endOfDState(&DState);417*/418419420/* *****************************************421* FSE unsafe API422*******************************************/423static unsigned char FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);424/* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */425426427/* *****************************************428* Implementation of inlined functions429*******************************************/430typedef struct {431int deltaFindState;432U32 deltaNbBits;433} FSE_symbolCompressionTransform; /* total 8 bytes */434435MEM_STATIC void FSE_initCState(FSE_CState_t* statePtr, const FSE_CTable* ct)436{437const void* ptr = ct;438const U16* u16ptr = (const U16*) ptr;439const U32 tableLog = MEM_read16(ptr);440statePtr->value = (ptrdiff_t)1<<tableLog;441statePtr->stateTable = u16ptr+2;442statePtr->symbolTT = ct + 1 + (tableLog ? (1<<(tableLog-1)) : 1);443statePtr->stateLog = tableLog;444}445446447/*! FSE_initCState2() :448* Same as FSE_initCState(), but the first symbol to include (which will be the last to be read)449* uses the smallest state value possible, saving the cost of this symbol */450MEM_STATIC void FSE_initCState2(FSE_CState_t* statePtr, const FSE_CTable* ct, U32 symbol)451{452FSE_initCState(statePtr, ct);453{ const FSE_symbolCompressionTransform symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];454const U16* stateTable = (const U16*)(statePtr->stateTable);455U32 nbBitsOut = (U32)((symbolTT.deltaNbBits + (1<<15)) >> 16);456statePtr->value = (nbBitsOut << 16) - symbolTT.deltaNbBits;457statePtr->value = stateTable[(statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];458}459}460461MEM_STATIC void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* statePtr, unsigned symbol)462{463FSE_symbolCompressionTransform const symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];464const U16* const stateTable = (const U16*)(statePtr->stateTable);465U32 const nbBitsOut = (U32)((statePtr->value + symbolTT.deltaNbBits) >> 16);466BIT_addBits(bitC, statePtr->value, nbBitsOut);467statePtr->value = stateTable[ (statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];468}469470MEM_STATIC void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* statePtr)471{472BIT_addBits(bitC, statePtr->value, statePtr->stateLog);473BIT_flushBits(bitC);474}475476477/* FSE_getMaxNbBits() :478* Approximate maximum cost of a symbol, in bits.479* Fractional get rounded up (i.e. a symbol with a normalized frequency of 3 gives the same result as a frequency of 2)480* note 1 : assume symbolValue is valid (<= maxSymbolValue)481* note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */482MEM_STATIC U32 FSE_getMaxNbBits(const void* symbolTTPtr, U32 symbolValue)483{484const FSE_symbolCompressionTransform* symbolTT = (const FSE_symbolCompressionTransform*) symbolTTPtr;485return (symbolTT[symbolValue].deltaNbBits + ((1<<16)-1)) >> 16;486}487488/* FSE_bitCost() :489* Approximate symbol cost, as fractional value, using fixed-point format (accuracyLog fractional bits)490* note 1 : assume symbolValue is valid (<= maxSymbolValue)491* note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */492MEM_STATIC U32 FSE_bitCost(const void* symbolTTPtr, U32 tableLog, U32 symbolValue, U32 accuracyLog)493{494const FSE_symbolCompressionTransform* symbolTT = (const FSE_symbolCompressionTransform*) symbolTTPtr;495U32 const minNbBits = symbolTT[symbolValue].deltaNbBits >> 16;496U32 const threshold = (minNbBits+1) << 16;497assert(tableLog < 16);498assert(accuracyLog < 31-tableLog); /* ensure enough room for renormalization double shift */499{ U32 const tableSize = 1 << tableLog;500U32 const deltaFromThreshold = threshold - (symbolTT[symbolValue].deltaNbBits + tableSize);501U32 const normalizedDeltaFromThreshold = (deltaFromThreshold << accuracyLog) >> tableLog; /* linear interpolation (very approximate) */502U32 const bitMultiplier = 1 << accuracyLog;503assert(symbolTT[symbolValue].deltaNbBits + tableSize <= threshold);504assert(normalizedDeltaFromThreshold <= bitMultiplier);505return (minNbBits+1)*bitMultiplier - normalizedDeltaFromThreshold;506}507}508509510/* ====== Decompression ====== */511512typedef struct {513U16 tableLog;514U16 fastMode;515} FSE_DTableHeader; /* sizeof U32 */516517typedef struct518{519unsigned short newState;520unsigned char symbol;521unsigned char nbBits;522} FSE_decode_t; /* size == U32 */523524MEM_STATIC void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt)525{526const void* ptr = dt;527const FSE_DTableHeader* const DTableH = (const FSE_DTableHeader*)ptr;528DStatePtr->state = BIT_readBits(bitD, DTableH->tableLog);529BIT_reloadDStream(bitD);530DStatePtr->table = dt + 1;531}532533MEM_STATIC BYTE FSE_peekSymbol(const FSE_DState_t* DStatePtr)534{535FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];536return DInfo.symbol;537}538539MEM_STATIC void FSE_updateState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)540{541FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];542U32 const nbBits = DInfo.nbBits;543size_t const lowBits = BIT_readBits(bitD, nbBits);544DStatePtr->state = DInfo.newState + lowBits;545}546547MEM_STATIC BYTE FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)548{549FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];550U32 const nbBits = DInfo.nbBits;551BYTE const symbol = DInfo.symbol;552size_t const lowBits = BIT_readBits(bitD, nbBits);553554DStatePtr->state = DInfo.newState + lowBits;555return symbol;556}557558/*! FSE_decodeSymbolFast() :559unsafe, only works if no symbol has a probability > 50% */560MEM_STATIC BYTE FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)561{562FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];563U32 const nbBits = DInfo.nbBits;564BYTE const symbol = DInfo.symbol;565size_t const lowBits = BIT_readBitsFast(bitD, nbBits);566567DStatePtr->state = DInfo.newState + lowBits;568return symbol;569}570571MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)572{573return DStatePtr->state == 0;574}575576577578#ifndef FSE_COMMONDEFS_ONLY579580/* **************************************************************581* Tuning parameters582****************************************************************/583/*!MEMORY_USAGE :584* Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)585* Increasing memory usage improves compression ratio586* Reduced memory usage can improve speed, due to cache effect587* Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */588#ifndef FSE_MAX_MEMORY_USAGE589# define FSE_MAX_MEMORY_USAGE 14590#endif591#ifndef FSE_DEFAULT_MEMORY_USAGE592# define FSE_DEFAULT_MEMORY_USAGE 13593#endif594#if (FSE_DEFAULT_MEMORY_USAGE > FSE_MAX_MEMORY_USAGE)595# error "FSE_DEFAULT_MEMORY_USAGE must be <= FSE_MAX_MEMORY_USAGE"596#endif597598/*!FSE_MAX_SYMBOL_VALUE :599* Maximum symbol value authorized.600* Required for proper stack allocation */601#ifndef FSE_MAX_SYMBOL_VALUE602# define FSE_MAX_SYMBOL_VALUE 255603#endif604605/* **************************************************************606* template functions type & suffix607****************************************************************/608#define FSE_FUNCTION_TYPE BYTE609#define FSE_FUNCTION_EXTENSION610#define FSE_DECODE_TYPE FSE_decode_t611612613#endif /* !FSE_COMMONDEFS_ONLY */614615616/* ***************************************************************617* Constants618*****************************************************************/619#define FSE_MAX_TABLELOG (FSE_MAX_MEMORY_USAGE-2)620#define FSE_MAX_TABLESIZE (1U<<FSE_MAX_TABLELOG)621#define FSE_MAXTABLESIZE_MASK (FSE_MAX_TABLESIZE-1)622#define FSE_DEFAULT_TABLELOG (FSE_DEFAULT_MEMORY_USAGE-2)623#define FSE_MIN_TABLELOG 5624625#define FSE_TABLELOG_ABSOLUTE_MAX 15626#if FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX627# error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"628#endif629630#define FSE_TABLESTEP(tableSize) (((tableSize)>>1) + ((tableSize)>>3) + 3)631632633#endif /* FSE_STATIC_LINKING_ONLY */634635636#if defined (__cplusplus)637}638#endif639640641