xref: /freebsd-src/sys/contrib/openzfs/module/zstd/lib/common/fse.h (revision c03c5b1c80914ec656fbee84539355d1fad68bf9)
1*c03c5b1cSMartin Matuska /* ******************************************************************
2*c03c5b1cSMartin Matuska  * FSE : Finite State Entropy codec
3*c03c5b1cSMartin Matuska  * Public Prototypes declaration
4*c03c5b1cSMartin Matuska  * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
5*c03c5b1cSMartin Matuska  *
6*c03c5b1cSMartin Matuska  * You can contact the author at :
7*c03c5b1cSMartin Matuska  * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
8*c03c5b1cSMartin Matuska  *
9*c03c5b1cSMartin Matuska  * This source code is licensed under both the BSD-style license (found in the
10*c03c5b1cSMartin Matuska  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
11*c03c5b1cSMartin Matuska  * in the COPYING file in the root directory of this source tree).
12*c03c5b1cSMartin Matuska  * You may select, at your option, one of the above-listed licenses.
13*c03c5b1cSMartin Matuska ****************************************************************** */
14*c03c5b1cSMartin Matuska 
15*c03c5b1cSMartin Matuska #if defined (__cplusplus)
16*c03c5b1cSMartin Matuska extern "C" {
17*c03c5b1cSMartin Matuska #endif
18*c03c5b1cSMartin Matuska 
19*c03c5b1cSMartin Matuska #ifndef FSE_H
20*c03c5b1cSMartin Matuska #define FSE_H
21*c03c5b1cSMartin Matuska 
22*c03c5b1cSMartin Matuska 
23*c03c5b1cSMartin Matuska /*-*****************************************
24*c03c5b1cSMartin Matuska *  Dependencies
25*c03c5b1cSMartin Matuska ******************************************/
26*c03c5b1cSMartin Matuska #include <stddef.h>    /* size_t, ptrdiff_t */
27*c03c5b1cSMartin Matuska 
28*c03c5b1cSMartin Matuska 
29*c03c5b1cSMartin Matuska /*-*****************************************
30*c03c5b1cSMartin Matuska *  FSE_PUBLIC_API : control library symbols visibility
31*c03c5b1cSMartin Matuska ******************************************/
32*c03c5b1cSMartin Matuska #if defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) && defined(__GNUC__) && (__GNUC__ >= 4)
33*c03c5b1cSMartin Matuska #  define FSE_PUBLIC_API __attribute__ ((visibility ("default")))
34*c03c5b1cSMartin Matuska #elif defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1)   /* Visual expected */
35*c03c5b1cSMartin Matuska #  define FSE_PUBLIC_API __declspec(dllexport)
36*c03c5b1cSMartin Matuska #elif defined(FSE_DLL_IMPORT) && (FSE_DLL_IMPORT==1)
37*c03c5b1cSMartin Matuska #  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.*/
38*c03c5b1cSMartin Matuska #else
39*c03c5b1cSMartin Matuska #  define FSE_PUBLIC_API
40*c03c5b1cSMartin Matuska #endif
41*c03c5b1cSMartin Matuska 
42*c03c5b1cSMartin Matuska /*------   Version   ------*/
43*c03c5b1cSMartin Matuska #define FSE_VERSION_MAJOR    0
44*c03c5b1cSMartin Matuska #define FSE_VERSION_MINOR    9
45*c03c5b1cSMartin Matuska #define FSE_VERSION_RELEASE  0
46*c03c5b1cSMartin Matuska 
47*c03c5b1cSMartin Matuska #define FSE_LIB_VERSION FSE_VERSION_MAJOR.FSE_VERSION_MINOR.FSE_VERSION_RELEASE
48*c03c5b1cSMartin Matuska #define FSE_QUOTE(str) #str
49*c03c5b1cSMartin Matuska #define FSE_EXPAND_AND_QUOTE(str) FSE_QUOTE(str)
50*c03c5b1cSMartin Matuska #define FSE_VERSION_STRING FSE_EXPAND_AND_QUOTE(FSE_LIB_VERSION)
51*c03c5b1cSMartin Matuska 
52*c03c5b1cSMartin Matuska #define FSE_VERSION_NUMBER  (FSE_VERSION_MAJOR *100*100 + FSE_VERSION_MINOR *100 + FSE_VERSION_RELEASE)
53*c03c5b1cSMartin Matuska FSE_PUBLIC_API unsigned FSE_versionNumber(void);   /**< library version number; to be used when checking dll version */
54*c03c5b1cSMartin Matuska 
55*c03c5b1cSMartin Matuska 
56*c03c5b1cSMartin Matuska /*-****************************************
57*c03c5b1cSMartin Matuska *  FSE simple functions
58*c03c5b1cSMartin Matuska ******************************************/
59*c03c5b1cSMartin Matuska /*! FSE_compress() :
60*c03c5b1cSMartin Matuska     Compress content of buffer 'src', of size 'srcSize', into destination buffer 'dst'.
61*c03c5b1cSMartin Matuska     'dst' buffer must be already allocated. Compression runs faster is dstCapacity >= FSE_compressBound(srcSize).
62*c03c5b1cSMartin Matuska     @return : size of compressed data (<= dstCapacity).
63*c03c5b1cSMartin Matuska     Special values : if return == 0, srcData is not compressible => Nothing is stored within dst !!!
64*c03c5b1cSMartin Matuska                      if return == 1, srcData is a single byte symbol * srcSize times. Use RLE compression instead.
65*c03c5b1cSMartin Matuska                      if FSE_isError(return), compression failed (more details using FSE_getErrorName())
66*c03c5b1cSMartin Matuska */
67*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_compress(void* dst, size_t dstCapacity,
68*c03c5b1cSMartin Matuska                              const void* src, size_t srcSize);
69*c03c5b1cSMartin Matuska 
70*c03c5b1cSMartin Matuska /*! FSE_decompress():
71*c03c5b1cSMartin Matuska     Decompress FSE data from buffer 'cSrc', of size 'cSrcSize',
72*c03c5b1cSMartin Matuska     into already allocated destination buffer 'dst', of size 'dstCapacity'.
73*c03c5b1cSMartin Matuska     @return : size of regenerated data (<= maxDstSize),
74*c03c5b1cSMartin Matuska               or an error code, which can be tested using FSE_isError() .
75*c03c5b1cSMartin Matuska 
76*c03c5b1cSMartin Matuska     ** Important ** : FSE_decompress() does not decompress non-compressible nor RLE data !!!
77*c03c5b1cSMartin Matuska     Why ? : making this distinction requires a header.
78*c03c5b1cSMartin Matuska     Header management is intentionally delegated to the user layer, which can better manage special cases.
79*c03c5b1cSMartin Matuska */
80*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_decompress(void* dst,  size_t dstCapacity,
81*c03c5b1cSMartin Matuska                                const void* cSrc, size_t cSrcSize);
82*c03c5b1cSMartin Matuska 
83*c03c5b1cSMartin Matuska 
84*c03c5b1cSMartin Matuska /*-*****************************************
85*c03c5b1cSMartin Matuska *  Tool functions
86*c03c5b1cSMartin Matuska ******************************************/
87*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_compressBound(size_t size);       /* maximum compressed size */
88*c03c5b1cSMartin Matuska 
89*c03c5b1cSMartin Matuska /* Error Management */
90*c03c5b1cSMartin Matuska FSE_PUBLIC_API unsigned    FSE_isError(size_t code);        /* tells if a return value is an error code */
91*c03c5b1cSMartin Matuska FSE_PUBLIC_API const char* FSE_getErrorName(size_t code);   /* provides error code string (useful for debugging) */
92*c03c5b1cSMartin Matuska 
93*c03c5b1cSMartin Matuska 
94*c03c5b1cSMartin Matuska /*-*****************************************
95*c03c5b1cSMartin Matuska *  FSE advanced functions
96*c03c5b1cSMartin Matuska ******************************************/
97*c03c5b1cSMartin Matuska /*! FSE_compress2() :
98*c03c5b1cSMartin Matuska     Same as FSE_compress(), but allows the selection of 'maxSymbolValue' and 'tableLog'
99*c03c5b1cSMartin Matuska     Both parameters can be defined as '0' to mean : use default value
100*c03c5b1cSMartin Matuska     @return : size of compressed data
101*c03c5b1cSMartin Matuska     Special values : if return == 0, srcData is not compressible => Nothing is stored within cSrc !!!
102*c03c5b1cSMartin Matuska                      if return == 1, srcData is a single byte symbol * srcSize times. Use RLE compression.
103*c03c5b1cSMartin Matuska                      if FSE_isError(return), it's an error code.
104*c03c5b1cSMartin Matuska */
105*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_compress2 (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog);
106*c03c5b1cSMartin Matuska 
107*c03c5b1cSMartin Matuska 
108*c03c5b1cSMartin Matuska /*-*****************************************
109*c03c5b1cSMartin Matuska *  FSE detailed API
110*c03c5b1cSMartin Matuska ******************************************/
111*c03c5b1cSMartin Matuska /*!
112*c03c5b1cSMartin Matuska FSE_compress() does the following:
113*c03c5b1cSMartin Matuska 1. count symbol occurrence from source[] into table count[] (see hist.h)
114*c03c5b1cSMartin Matuska 2. normalize counters so that sum(count[]) == Power_of_2 (2^tableLog)
115*c03c5b1cSMartin Matuska 3. save normalized counters to memory buffer using writeNCount()
116*c03c5b1cSMartin Matuska 4. build encoding table 'CTable' from normalized counters
117*c03c5b1cSMartin Matuska 5. encode the data stream using encoding table 'CTable'
118*c03c5b1cSMartin Matuska 
119*c03c5b1cSMartin Matuska FSE_decompress() does the following:
120*c03c5b1cSMartin Matuska 1. read normalized counters with readNCount()
121*c03c5b1cSMartin Matuska 2. build decoding table 'DTable' from normalized counters
122*c03c5b1cSMartin Matuska 3. decode the data stream using decoding table 'DTable'
123*c03c5b1cSMartin Matuska 
124*c03c5b1cSMartin Matuska The following API allows targeting specific sub-functions for advanced tasks.
125*c03c5b1cSMartin Matuska For example, it's possible to compress several blocks using the same 'CTable',
126*c03c5b1cSMartin Matuska or to save and provide normalized distribution using external method.
127*c03c5b1cSMartin Matuska */
128*c03c5b1cSMartin Matuska 
129*c03c5b1cSMartin Matuska /* *** COMPRESSION *** */
130*c03c5b1cSMartin Matuska 
131*c03c5b1cSMartin Matuska /*! FSE_optimalTableLog():
132*c03c5b1cSMartin Matuska     dynamically downsize 'tableLog' when conditions are met.
133*c03c5b1cSMartin Matuska     It saves CPU time, by using smaller tables, while preserving or even improving compression ratio.
134*c03c5b1cSMartin Matuska     @return : recommended tableLog (necessarily <= 'maxTableLog') */
135*c03c5b1cSMartin Matuska FSE_PUBLIC_API unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue);
136*c03c5b1cSMartin Matuska 
137*c03c5b1cSMartin Matuska /*! FSE_normalizeCount():
138*c03c5b1cSMartin Matuska     normalize counts so that sum(count[]) == Power_of_2 (2^tableLog)
139*c03c5b1cSMartin Matuska     'normalizedCounter' is a table of short, of minimum size (maxSymbolValue+1).
140*c03c5b1cSMartin Matuska     @return : tableLog,
141*c03c5b1cSMartin Matuska               or an errorCode, which can be tested using FSE_isError() */
142*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_normalizeCount(short* normalizedCounter, unsigned tableLog,
143*c03c5b1cSMartin Matuska                     const unsigned* count, size_t srcSize, unsigned maxSymbolValue);
144*c03c5b1cSMartin Matuska 
145*c03c5b1cSMartin Matuska /*! FSE_NCountWriteBound():
146*c03c5b1cSMartin Matuska     Provides the maximum possible size of an FSE normalized table, given 'maxSymbolValue' and 'tableLog'.
147*c03c5b1cSMartin Matuska     Typically useful for allocation purpose. */
148*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog);
149*c03c5b1cSMartin Matuska 
150*c03c5b1cSMartin Matuska /*! FSE_writeNCount():
151*c03c5b1cSMartin Matuska     Compactly save 'normalizedCounter' into 'buffer'.
152*c03c5b1cSMartin Matuska     @return : size of the compressed table,
153*c03c5b1cSMartin Matuska               or an errorCode, which can be tested using FSE_isError(). */
154*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_writeNCount (void* buffer, size_t bufferSize,
155*c03c5b1cSMartin Matuska                                  const short* normalizedCounter,
156*c03c5b1cSMartin Matuska                                  unsigned maxSymbolValue, unsigned tableLog);
157*c03c5b1cSMartin Matuska 
158*c03c5b1cSMartin Matuska /*! Constructor and Destructor of FSE_CTable.
159*c03c5b1cSMartin Matuska     Note that FSE_CTable size depends on 'tableLog' and 'maxSymbolValue' */
160*c03c5b1cSMartin Matuska typedef unsigned FSE_CTable;   /* don't allocate that. It's only meant to be more restrictive than void* */
161*c03c5b1cSMartin Matuska FSE_PUBLIC_API FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog);
162*c03c5b1cSMartin Matuska FSE_PUBLIC_API void        FSE_freeCTable (FSE_CTable* ct);
163*c03c5b1cSMartin Matuska 
164*c03c5b1cSMartin Matuska /*! FSE_buildCTable():
165*c03c5b1cSMartin Matuska     Builds `ct`, which must be already allocated, using FSE_createCTable().
166*c03c5b1cSMartin Matuska     @return : 0, or an errorCode, which can be tested using FSE_isError() */
167*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
168*c03c5b1cSMartin Matuska 
169*c03c5b1cSMartin Matuska /*! FSE_compress_usingCTable():
170*c03c5b1cSMartin Matuska     Compress `src` using `ct` into `dst` which must be already allocated.
171*c03c5b1cSMartin Matuska     @return : size of compressed data (<= `dstCapacity`),
172*c03c5b1cSMartin Matuska               or 0 if compressed data could not fit into `dst`,
173*c03c5b1cSMartin Matuska               or an errorCode, which can be tested using FSE_isError() */
174*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_compress_usingCTable (void* dst, size_t dstCapacity, const void* src, size_t srcSize, const FSE_CTable* ct);
175*c03c5b1cSMartin Matuska 
176*c03c5b1cSMartin Matuska /*!
177*c03c5b1cSMartin Matuska Tutorial :
178*c03c5b1cSMartin Matuska ----------
179*c03c5b1cSMartin Matuska The first step is to count all symbols. FSE_count() does this job very fast.
180*c03c5b1cSMartin Matuska Result will be saved into 'count', a table of unsigned int, which must be already allocated, and have 'maxSymbolValuePtr[0]+1' cells.
181*c03c5b1cSMartin Matuska 'src' is a table of bytes of size 'srcSize'. All values within 'src' MUST be <= maxSymbolValuePtr[0]
182*c03c5b1cSMartin Matuska maxSymbolValuePtr[0] will be updated, with its real value (necessarily <= original value)
183*c03c5b1cSMartin Matuska FSE_count() will return the number of occurrence of the most frequent symbol.
184*c03c5b1cSMartin Matuska This can be used to know if there is a single symbol within 'src', and to quickly evaluate its compressibility.
185*c03c5b1cSMartin Matuska If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()).
186*c03c5b1cSMartin Matuska 
187*c03c5b1cSMartin Matuska The next step is to normalize the frequencies.
188*c03c5b1cSMartin Matuska FSE_normalizeCount() will ensure that sum of frequencies is == 2 ^'tableLog'.
189*c03c5b1cSMartin Matuska It also guarantees a minimum of 1 to any Symbol with frequency >= 1.
190*c03c5b1cSMartin Matuska You can use 'tableLog'==0 to mean "use default tableLog value".
191*c03c5b1cSMartin Matuska If you are unsure of which tableLog value to use, you can ask FSE_optimalTableLog(),
192*c03c5b1cSMartin Matuska which will provide the optimal valid tableLog given sourceSize, maxSymbolValue, and a user-defined maximum (0 means "default").
193*c03c5b1cSMartin Matuska 
194*c03c5b1cSMartin Matuska The result of FSE_normalizeCount() will be saved into a table,
195*c03c5b1cSMartin Matuska called 'normalizedCounter', which is a table of signed short.
196*c03c5b1cSMartin Matuska 'normalizedCounter' must be already allocated, and have at least 'maxSymbolValue+1' cells.
197*c03c5b1cSMartin Matuska The return value is tableLog if everything proceeded as expected.
198*c03c5b1cSMartin Matuska It is 0 if there is a single symbol within distribution.
199*c03c5b1cSMartin Matuska If there is an error (ex: invalid tableLog value), the function will return an ErrorCode (which can be tested using FSE_isError()).
200*c03c5b1cSMartin Matuska 
201*c03c5b1cSMartin Matuska 'normalizedCounter' can be saved in a compact manner to a memory area using FSE_writeNCount().
202*c03c5b1cSMartin Matuska 'buffer' must be already allocated.
203*c03c5b1cSMartin Matuska For guaranteed success, buffer size must be at least FSE_headerBound().
204*c03c5b1cSMartin Matuska The result of the function is the number of bytes written into 'buffer'.
205*c03c5b1cSMartin Matuska If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError(); ex : buffer size too small).
206*c03c5b1cSMartin Matuska 
207*c03c5b1cSMartin Matuska 'normalizedCounter' can then be used to create the compression table 'CTable'.
208*c03c5b1cSMartin Matuska The space required by 'CTable' must be already allocated, using FSE_createCTable().
209*c03c5b1cSMartin Matuska You can then use FSE_buildCTable() to fill 'CTable'.
210*c03c5b1cSMartin Matuska If there is an error, both functions will return an ErrorCode (which can be tested using FSE_isError()).
211*c03c5b1cSMartin Matuska 
212*c03c5b1cSMartin Matuska 'CTable' can then be used to compress 'src', with FSE_compress_usingCTable().
213*c03c5b1cSMartin Matuska Similar to FSE_count(), the convention is that 'src' is assumed to be a table of char of size 'srcSize'
214*c03c5b1cSMartin Matuska The function returns the size of compressed data (without header), necessarily <= `dstCapacity`.
215*c03c5b1cSMartin Matuska If it returns '0', compressed data could not fit into 'dst'.
216*c03c5b1cSMartin Matuska If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()).
217*c03c5b1cSMartin Matuska */
218*c03c5b1cSMartin Matuska 
219*c03c5b1cSMartin Matuska 
220*c03c5b1cSMartin Matuska /* *** DECOMPRESSION *** */
221*c03c5b1cSMartin Matuska 
222*c03c5b1cSMartin Matuska /*! FSE_readNCount():
223*c03c5b1cSMartin Matuska     Read compactly saved 'normalizedCounter' from 'rBuffer'.
224*c03c5b1cSMartin Matuska     @return : size read from 'rBuffer',
225*c03c5b1cSMartin Matuska               or an errorCode, which can be tested using FSE_isError().
226*c03c5b1cSMartin Matuska               maxSymbolValuePtr[0] and tableLogPtr[0] will also be updated with their respective values */
227*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_readNCount (short* normalizedCounter,
228*c03c5b1cSMartin Matuska                            unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
229*c03c5b1cSMartin Matuska                            const void* rBuffer, size_t rBuffSize);
230*c03c5b1cSMartin Matuska 
231*c03c5b1cSMartin Matuska /*! Constructor and Destructor of FSE_DTable.
232*c03c5b1cSMartin Matuska     Note that its size depends on 'tableLog' */
233*c03c5b1cSMartin Matuska typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */
234*c03c5b1cSMartin Matuska FSE_PUBLIC_API FSE_DTable* FSE_createDTable(unsigned tableLog);
235*c03c5b1cSMartin Matuska FSE_PUBLIC_API void        FSE_freeDTable(FSE_DTable* dt);
236*c03c5b1cSMartin Matuska 
237*c03c5b1cSMartin Matuska /*! FSE_buildDTable():
238*c03c5b1cSMartin Matuska     Builds 'dt', which must be already allocated, using FSE_createDTable().
239*c03c5b1cSMartin Matuska     return : 0, or an errorCode, which can be tested using FSE_isError() */
240*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_buildDTable (FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);
241*c03c5b1cSMartin Matuska 
242*c03c5b1cSMartin Matuska /*! FSE_decompress_usingDTable():
243*c03c5b1cSMartin Matuska     Decompress compressed source `cSrc` of size `cSrcSize` using `dt`
244*c03c5b1cSMartin Matuska     into `dst` which must be already allocated.
245*c03c5b1cSMartin Matuska     @return : size of regenerated data (necessarily <= `dstCapacity`),
246*c03c5b1cSMartin Matuska               or an errorCode, which can be tested using FSE_isError() */
247*c03c5b1cSMartin Matuska FSE_PUBLIC_API size_t FSE_decompress_usingDTable(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, const FSE_DTable* dt);
248*c03c5b1cSMartin Matuska 
249*c03c5b1cSMartin Matuska /*!
250*c03c5b1cSMartin Matuska Tutorial :
251*c03c5b1cSMartin Matuska ----------
252*c03c5b1cSMartin Matuska (Note : these functions only decompress FSE-compressed blocks.
253*c03c5b1cSMartin Matuska  If block is uncompressed, use memcpy() instead
254*c03c5b1cSMartin Matuska  If block is a single repeated byte, use memset() instead )
255*c03c5b1cSMartin Matuska 
256*c03c5b1cSMartin Matuska The first step is to obtain the normalized frequencies of symbols.
257*c03c5b1cSMartin Matuska This can be performed by FSE_readNCount() if it was saved using FSE_writeNCount().
258*c03c5b1cSMartin Matuska 'normalizedCounter' must be already allocated, and have at least 'maxSymbolValuePtr[0]+1' cells of signed short.
259*c03c5b1cSMartin Matuska In practice, that means it's necessary to know 'maxSymbolValue' beforehand,
260*c03c5b1cSMartin Matuska or size the table to handle worst case situations (typically 256).
261*c03c5b1cSMartin Matuska FSE_readNCount() will provide 'tableLog' and 'maxSymbolValue'.
262*c03c5b1cSMartin Matuska The result of FSE_readNCount() is the number of bytes read from 'rBuffer'.
263*c03c5b1cSMartin Matuska Note that 'rBufferSize' must be at least 4 bytes, even if useful information is less than that.
264*c03c5b1cSMartin Matuska If there is an error, the function will return an error code, which can be tested using FSE_isError().
265*c03c5b1cSMartin Matuska 
266*c03c5b1cSMartin Matuska The next step is to build the decompression tables 'FSE_DTable' from 'normalizedCounter'.
267*c03c5b1cSMartin Matuska This is performed by the function FSE_buildDTable().
268*c03c5b1cSMartin Matuska The space required by 'FSE_DTable' must be already allocated using FSE_createDTable().
269*c03c5b1cSMartin Matuska If there is an error, the function will return an error code, which can be tested using FSE_isError().
270*c03c5b1cSMartin Matuska 
271*c03c5b1cSMartin Matuska `FSE_DTable` can then be used to decompress `cSrc`, with FSE_decompress_usingDTable().
272*c03c5b1cSMartin Matuska `cSrcSize` must be strictly correct, otherwise decompression will fail.
273*c03c5b1cSMartin Matuska FSE_decompress_usingDTable() result will tell how many bytes were regenerated (<=`dstCapacity`).
274*c03c5b1cSMartin Matuska If there is an error, the function will return an error code, which can be tested using FSE_isError(). (ex: dst buffer too small)
275*c03c5b1cSMartin Matuska */
276*c03c5b1cSMartin Matuska 
277*c03c5b1cSMartin Matuska #endif  /* FSE_H */
278*c03c5b1cSMartin Matuska 
279*c03c5b1cSMartin Matuska #if defined(FSE_STATIC_LINKING_ONLY) && !defined(FSE_H_FSE_STATIC_LINKING_ONLY)
280*c03c5b1cSMartin Matuska #define FSE_H_FSE_STATIC_LINKING_ONLY
281*c03c5b1cSMartin Matuska 
282*c03c5b1cSMartin Matuska /* *** Dependency *** */
283*c03c5b1cSMartin Matuska #include "bitstream.h"
284*c03c5b1cSMartin Matuska 
285*c03c5b1cSMartin Matuska 
286*c03c5b1cSMartin Matuska /* *****************************************
287*c03c5b1cSMartin Matuska *  Static allocation
288*c03c5b1cSMartin Matuska *******************************************/
289*c03c5b1cSMartin Matuska /* FSE buffer bounds */
290*c03c5b1cSMartin Matuska #define FSE_NCOUNTBOUND 512
291*c03c5b1cSMartin Matuska #define FSE_BLOCKBOUND(size) (size + (size>>7) + 4 /* fse states */ + sizeof(size_t) /* bitContainer */)
292*c03c5b1cSMartin Matuska #define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size))   /* Macro version, useful for static allocation */
293*c03c5b1cSMartin Matuska 
294*c03c5b1cSMartin Matuska /* It is possible to statically allocate FSE CTable/DTable as a table of FSE_CTable/FSE_DTable using below macros */
295*c03c5b1cSMartin Matuska #define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue)   (1 + (1<<(maxTableLog-1)) + ((maxSymbolValue+1)*2))
296*c03c5b1cSMartin Matuska #define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<maxTableLog))
297*c03c5b1cSMartin Matuska 
298*c03c5b1cSMartin Matuska /* or use the size to malloc() space directly. Pay attention to alignment restrictions though */
299*c03c5b1cSMartin Matuska #define FSE_CTABLE_SIZE(maxTableLog, maxSymbolValue)   (FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(FSE_CTable))
300*c03c5b1cSMartin Matuska #define FSE_DTABLE_SIZE(maxTableLog)                   (FSE_DTABLE_SIZE_U32(maxTableLog) * sizeof(FSE_DTable))
301*c03c5b1cSMartin Matuska 
302*c03c5b1cSMartin Matuska 
303*c03c5b1cSMartin Matuska /* *****************************************
304*c03c5b1cSMartin Matuska  *  FSE advanced API
305*c03c5b1cSMartin Matuska  ***************************************** */
306*c03c5b1cSMartin Matuska 
307*c03c5b1cSMartin Matuska unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus);
308*c03c5b1cSMartin Matuska /**< same as FSE_optimalTableLog(), which used `minus==2` */
309*c03c5b1cSMartin Matuska 
310*c03c5b1cSMartin Matuska /* FSE_compress_wksp() :
311*c03c5b1cSMartin Matuska  * Same as FSE_compress2(), but using an externally allocated scratch buffer (`workSpace`).
312*c03c5b1cSMartin Matuska  * FSE_WKSP_SIZE_U32() provides the minimum size required for `workSpace` as a table of FSE_CTable.
313*c03c5b1cSMartin Matuska  */
314*c03c5b1cSMartin Matuska #define FSE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue)   ( FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) + ((maxTableLog > 12) ? (1 << (maxTableLog - 2)) : 1024) )
315*c03c5b1cSMartin Matuska size_t FSE_compress_wksp (void* dst, size_t dstSize, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
316*c03c5b1cSMartin Matuska 
317*c03c5b1cSMartin Matuska size_t FSE_buildCTable_raw (FSE_CTable* ct, unsigned nbBits);
318*c03c5b1cSMartin Matuska /**< build a fake FSE_CTable, designed for a flat distribution, where each symbol uses nbBits */
319*c03c5b1cSMartin Matuska 
320*c03c5b1cSMartin Matuska size_t FSE_buildCTable_rle (FSE_CTable* ct, unsigned char symbolValue);
321*c03c5b1cSMartin Matuska /**< build a fake FSE_CTable, designed to compress always the same symbolValue */
322*c03c5b1cSMartin Matuska 
323*c03c5b1cSMartin Matuska /* FSE_buildCTable_wksp() :
324*c03c5b1cSMartin Matuska  * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
325*c03c5b1cSMartin Matuska  * `wkspSize` must be >= `(1<<tableLog)`.
326*c03c5b1cSMartin Matuska  */
327*c03c5b1cSMartin Matuska size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
328*c03c5b1cSMartin Matuska 
329*c03c5b1cSMartin Matuska size_t FSE_buildDTable_raw (FSE_DTable* dt, unsigned nbBits);
330*c03c5b1cSMartin Matuska /**< build a fake FSE_DTable, designed to read a flat distribution where each symbol uses nbBits */
331*c03c5b1cSMartin Matuska 
332*c03c5b1cSMartin Matuska size_t FSE_buildDTable_rle (FSE_DTable* dt, unsigned char symbolValue);
333*c03c5b1cSMartin Matuska /**< build a fake FSE_DTable, designed to always generate the same symbolValue */
334*c03c5b1cSMartin Matuska 
335*c03c5b1cSMartin Matuska size_t FSE_decompress_wksp(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, FSE_DTable* workSpace, unsigned maxLog);
336*c03c5b1cSMartin Matuska /**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DTABLE_SIZE_U32(maxLog)` */
337*c03c5b1cSMartin Matuska 
338*c03c5b1cSMartin Matuska typedef enum {
339*c03c5b1cSMartin Matuska    FSE_repeat_none,  /**< Cannot use the previous table */
340*c03c5b1cSMartin Matuska    FSE_repeat_check, /**< Can use the previous table but it must be checked */
341*c03c5b1cSMartin Matuska    FSE_repeat_valid  /**< Can use the previous table and it is assumed to be valid */
342*c03c5b1cSMartin Matuska  } FSE_repeat;
343*c03c5b1cSMartin Matuska 
344*c03c5b1cSMartin Matuska /* *****************************************
345*c03c5b1cSMartin Matuska *  FSE symbol compression API
346*c03c5b1cSMartin Matuska *******************************************/
347*c03c5b1cSMartin Matuska /*!
348*c03c5b1cSMartin Matuska    This API consists of small unitary functions, which highly benefit from being inlined.
349*c03c5b1cSMartin Matuska    Hence their body are included in next section.
350*c03c5b1cSMartin Matuska */
351*c03c5b1cSMartin Matuska typedef struct {
352*c03c5b1cSMartin Matuska     ptrdiff_t   value;
353*c03c5b1cSMartin Matuska     const void* stateTable;
354*c03c5b1cSMartin Matuska     const void* symbolTT;
355*c03c5b1cSMartin Matuska     unsigned    stateLog;
356*c03c5b1cSMartin Matuska } FSE_CState_t;
357*c03c5b1cSMartin Matuska 
358*c03c5b1cSMartin Matuska static void FSE_initCState(FSE_CState_t* CStatePtr, const FSE_CTable* ct);
359*c03c5b1cSMartin Matuska 
360*c03c5b1cSMartin Matuska static void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* CStatePtr, unsigned symbol);
361*c03c5b1cSMartin Matuska 
362*c03c5b1cSMartin Matuska static void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* CStatePtr);
363*c03c5b1cSMartin Matuska 
364*c03c5b1cSMartin Matuska /**<
365*c03c5b1cSMartin Matuska These functions are inner components of FSE_compress_usingCTable().
366*c03c5b1cSMartin Matuska They allow the creation of custom streams, mixing multiple tables and bit sources.
367*c03c5b1cSMartin Matuska 
368*c03c5b1cSMartin Matuska A key property to keep in mind is that encoding and decoding are done **in reverse direction**.
369*c03c5b1cSMartin Matuska So the first symbol you will encode is the last you will decode, like a LIFO stack.
370*c03c5b1cSMartin Matuska 
371*c03c5b1cSMartin Matuska You will need a few variables to track your CStream. They are :
372*c03c5b1cSMartin Matuska 
373*c03c5b1cSMartin Matuska FSE_CTable    ct;         // Provided by FSE_buildCTable()
374*c03c5b1cSMartin Matuska BIT_CStream_t bitStream;  // bitStream tracking structure
375*c03c5b1cSMartin Matuska FSE_CState_t  state;      // State tracking structure (can have several)
376*c03c5b1cSMartin Matuska 
377*c03c5b1cSMartin Matuska 
378*c03c5b1cSMartin Matuska The first thing to do is to init bitStream and state.
379*c03c5b1cSMartin Matuska     size_t errorCode = BIT_initCStream(&bitStream, dstBuffer, maxDstSize);
380*c03c5b1cSMartin Matuska     FSE_initCState(&state, ct);
381*c03c5b1cSMartin Matuska 
382*c03c5b1cSMartin Matuska Note that BIT_initCStream() can produce an error code, so its result should be tested, using FSE_isError();
383*c03c5b1cSMartin Matuska You can then encode your input data, byte after byte.
384*c03c5b1cSMartin Matuska FSE_encodeSymbol() outputs a maximum of 'tableLog' bits at a time.
385*c03c5b1cSMartin Matuska Remember decoding will be done in reverse direction.
386*c03c5b1cSMartin Matuska     FSE_encodeByte(&bitStream, &state, symbol);
387*c03c5b1cSMartin Matuska 
388*c03c5b1cSMartin Matuska At any time, you can also add any bit sequence.
389*c03c5b1cSMartin Matuska Note : maximum allowed nbBits is 25, for compatibility with 32-bits decoders
390*c03c5b1cSMartin Matuska     BIT_addBits(&bitStream, bitField, nbBits);
391*c03c5b1cSMartin Matuska 
392*c03c5b1cSMartin Matuska The above methods don't commit data to memory, they just store it into local register, for speed.
393*c03c5b1cSMartin Matuska Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (size_t).
394*c03c5b1cSMartin Matuska Writing data to memory is a manual operation, performed by the flushBits function.
395*c03c5b1cSMartin Matuska     BIT_flushBits(&bitStream);
396*c03c5b1cSMartin Matuska 
397*c03c5b1cSMartin Matuska Your last FSE encoding operation shall be to flush your last state value(s).
398*c03c5b1cSMartin Matuska     FSE_flushState(&bitStream, &state);
399*c03c5b1cSMartin Matuska 
400*c03c5b1cSMartin Matuska Finally, you must close the bitStream.
401*c03c5b1cSMartin Matuska The function returns the size of CStream in bytes.
402*c03c5b1cSMartin Matuska If data couldn't fit into dstBuffer, it will return a 0 ( == not compressible)
403*c03c5b1cSMartin Matuska If there is an error, it returns an errorCode (which can be tested using FSE_isError()).
404*c03c5b1cSMartin Matuska     size_t size = BIT_closeCStream(&bitStream);
405*c03c5b1cSMartin Matuska */
406*c03c5b1cSMartin Matuska 
407*c03c5b1cSMartin Matuska 
408*c03c5b1cSMartin Matuska /* *****************************************
409*c03c5b1cSMartin Matuska *  FSE symbol decompression API
410*c03c5b1cSMartin Matuska *******************************************/
411*c03c5b1cSMartin Matuska typedef struct {
412*c03c5b1cSMartin Matuska     size_t      state;
413*c03c5b1cSMartin Matuska     const void* table;   /* precise table may vary, depending on U16 */
414*c03c5b1cSMartin Matuska } FSE_DState_t;
415*c03c5b1cSMartin Matuska 
416*c03c5b1cSMartin Matuska 
417*c03c5b1cSMartin Matuska static void     FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt);
418*c03c5b1cSMartin Matuska 
419*c03c5b1cSMartin Matuska static unsigned char FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
420*c03c5b1cSMartin Matuska 
421*c03c5b1cSMartin Matuska static unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr);
422*c03c5b1cSMartin Matuska 
423*c03c5b1cSMartin Matuska /**<
424*c03c5b1cSMartin Matuska Let's now decompose FSE_decompress_usingDTable() into its unitary components.
425*c03c5b1cSMartin Matuska You will decode FSE-encoded symbols from the bitStream,
426*c03c5b1cSMartin Matuska and also any other bitFields you put in, **in reverse order**.
427*c03c5b1cSMartin Matuska 
428*c03c5b1cSMartin Matuska You will need a few variables to track your bitStream. They are :
429*c03c5b1cSMartin Matuska 
430*c03c5b1cSMartin Matuska BIT_DStream_t DStream;    // Stream context
431*c03c5b1cSMartin Matuska FSE_DState_t  DState;     // State context. Multiple ones are possible
432*c03c5b1cSMartin Matuska FSE_DTable*   DTablePtr;  // Decoding table, provided by FSE_buildDTable()
433*c03c5b1cSMartin Matuska 
434*c03c5b1cSMartin Matuska The first thing to do is to init the bitStream.
435*c03c5b1cSMartin Matuska     errorCode = BIT_initDStream(&DStream, srcBuffer, srcSize);
436*c03c5b1cSMartin Matuska 
437*c03c5b1cSMartin Matuska You should then retrieve your initial state(s)
438*c03c5b1cSMartin Matuska (in reverse flushing order if you have several ones) :
439*c03c5b1cSMartin Matuska     errorCode = FSE_initDState(&DState, &DStream, DTablePtr);
440*c03c5b1cSMartin Matuska 
441*c03c5b1cSMartin Matuska You can then decode your data, symbol after symbol.
442*c03c5b1cSMartin Matuska For information the maximum number of bits read by FSE_decodeSymbol() is 'tableLog'.
443*c03c5b1cSMartin Matuska Keep in mind that symbols are decoded in reverse order, like a LIFO stack (last in, first out).
444*c03c5b1cSMartin Matuska     unsigned char symbol = FSE_decodeSymbol(&DState, &DStream);
445*c03c5b1cSMartin Matuska 
446*c03c5b1cSMartin Matuska You can retrieve any bitfield you eventually stored into the bitStream (in reverse order)
447*c03c5b1cSMartin Matuska Note : maximum allowed nbBits is 25, for 32-bits compatibility
448*c03c5b1cSMartin Matuska     size_t bitField = BIT_readBits(&DStream, nbBits);
449*c03c5b1cSMartin Matuska 
450*c03c5b1cSMartin Matuska All above operations only read from local register (which size depends on size_t).
451*c03c5b1cSMartin Matuska Refueling the register from memory is manually performed by the reload method.
452*c03c5b1cSMartin Matuska     endSignal = FSE_reloadDStream(&DStream);
453*c03c5b1cSMartin Matuska 
454*c03c5b1cSMartin Matuska BIT_reloadDStream() result tells if there is still some more data to read from DStream.
455*c03c5b1cSMartin Matuska BIT_DStream_unfinished : there is still some data left into the DStream.
456*c03c5b1cSMartin Matuska BIT_DStream_endOfBuffer : Dstream reached end of buffer. Its container may no longer be completely filled.
457*c03c5b1cSMartin Matuska BIT_DStream_completed : Dstream reached its exact end, corresponding in general to decompression completed.
458*c03c5b1cSMartin Matuska BIT_DStream_tooFar : Dstream went too far. Decompression result is corrupted.
459*c03c5b1cSMartin Matuska 
460*c03c5b1cSMartin Matuska When reaching end of buffer (BIT_DStream_endOfBuffer), progress slowly, notably if you decode multiple symbols per loop,
461*c03c5b1cSMartin Matuska to properly detect the exact end of stream.
462*c03c5b1cSMartin Matuska After each decoded symbol, check if DStream is fully consumed using this simple test :
463*c03c5b1cSMartin Matuska     BIT_reloadDStream(&DStream) >= BIT_DStream_completed
464*c03c5b1cSMartin Matuska 
465*c03c5b1cSMartin Matuska When it's done, verify decompression is fully completed, by checking both DStream and the relevant states.
466*c03c5b1cSMartin Matuska Checking if DStream has reached its end is performed by :
467*c03c5b1cSMartin Matuska     BIT_endOfDStream(&DStream);
468*c03c5b1cSMartin Matuska Check also the states. There might be some symbols left there, if some high probability ones (>50%) are possible.
469*c03c5b1cSMartin Matuska     FSE_endOfDState(&DState);
470*c03c5b1cSMartin Matuska */
471*c03c5b1cSMartin Matuska 
472*c03c5b1cSMartin Matuska 
473*c03c5b1cSMartin Matuska /* *****************************************
474*c03c5b1cSMartin Matuska *  FSE unsafe API
475*c03c5b1cSMartin Matuska *******************************************/
476*c03c5b1cSMartin Matuska static unsigned char FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
477*c03c5b1cSMartin Matuska /* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */
478*c03c5b1cSMartin Matuska 
479*c03c5b1cSMartin Matuska 
480*c03c5b1cSMartin Matuska /* *****************************************
481*c03c5b1cSMartin Matuska *  Implementation of inlined functions
482*c03c5b1cSMartin Matuska *******************************************/
483*c03c5b1cSMartin Matuska typedef struct {
484*c03c5b1cSMartin Matuska     int deltaFindState;
485*c03c5b1cSMartin Matuska     U32 deltaNbBits;
486*c03c5b1cSMartin Matuska } FSE_symbolCompressionTransform; /* total 8 bytes */
487*c03c5b1cSMartin Matuska 
FSE_initCState(FSE_CState_t * statePtr,const FSE_CTable * ct)488*c03c5b1cSMartin Matuska MEM_STATIC void FSE_initCState(FSE_CState_t* statePtr, const FSE_CTable* ct)
489*c03c5b1cSMartin Matuska {
490*c03c5b1cSMartin Matuska     const void* ptr = ct;
491*c03c5b1cSMartin Matuska     const U16* u16ptr = (const U16*) ptr;
492*c03c5b1cSMartin Matuska     const U32 tableLog = MEM_read16(ptr);
493*c03c5b1cSMartin Matuska     statePtr->value = (ptrdiff_t)1<<tableLog;
494*c03c5b1cSMartin Matuska     statePtr->stateTable = u16ptr+2;
495*c03c5b1cSMartin Matuska     statePtr->symbolTT = ct + 1 + (tableLog ? (1<<(tableLog-1)) : 1);
496*c03c5b1cSMartin Matuska     statePtr->stateLog = tableLog;
497*c03c5b1cSMartin Matuska }
498*c03c5b1cSMartin Matuska 
499*c03c5b1cSMartin Matuska 
500*c03c5b1cSMartin Matuska /*! FSE_initCState2() :
501*c03c5b1cSMartin Matuska *   Same as FSE_initCState(), but the first symbol to include (which will be the last to be read)
502*c03c5b1cSMartin Matuska *   uses the smallest state value possible, saving the cost of this symbol */
FSE_initCState2(FSE_CState_t * statePtr,const FSE_CTable * ct,U32 symbol)503*c03c5b1cSMartin Matuska MEM_STATIC void FSE_initCState2(FSE_CState_t* statePtr, const FSE_CTable* ct, U32 symbol)
504*c03c5b1cSMartin Matuska {
505*c03c5b1cSMartin Matuska     FSE_initCState(statePtr, ct);
506*c03c5b1cSMartin Matuska     {   const FSE_symbolCompressionTransform symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];
507*c03c5b1cSMartin Matuska         const U16* stateTable = (const U16*)(statePtr->stateTable);
508*c03c5b1cSMartin Matuska         U32 nbBitsOut  = (U32)((symbolTT.deltaNbBits + (1<<15)) >> 16);
509*c03c5b1cSMartin Matuska         statePtr->value = (nbBitsOut << 16) - symbolTT.deltaNbBits;
510*c03c5b1cSMartin Matuska         statePtr->value = stateTable[(statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];
511*c03c5b1cSMartin Matuska     }
512*c03c5b1cSMartin Matuska }
513*c03c5b1cSMartin Matuska 
FSE_encodeSymbol(BIT_CStream_t * bitC,FSE_CState_t * statePtr,unsigned symbol)514*c03c5b1cSMartin Matuska MEM_STATIC void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* statePtr, unsigned symbol)
515*c03c5b1cSMartin Matuska {
516*c03c5b1cSMartin Matuska     FSE_symbolCompressionTransform const symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];
517*c03c5b1cSMartin Matuska     const U16* const stateTable = (const U16*)(statePtr->stateTable);
518*c03c5b1cSMartin Matuska     U32 const nbBitsOut  = (U32)((statePtr->value + symbolTT.deltaNbBits) >> 16);
519*c03c5b1cSMartin Matuska     BIT_addBits(bitC, statePtr->value, nbBitsOut);
520*c03c5b1cSMartin Matuska     statePtr->value = stateTable[ (statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];
521*c03c5b1cSMartin Matuska }
522*c03c5b1cSMartin Matuska 
FSE_flushCState(BIT_CStream_t * bitC,const FSE_CState_t * statePtr)523*c03c5b1cSMartin Matuska MEM_STATIC void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* statePtr)
524*c03c5b1cSMartin Matuska {
525*c03c5b1cSMartin Matuska     BIT_addBits(bitC, statePtr->value, statePtr->stateLog);
526*c03c5b1cSMartin Matuska     BIT_flushBits(bitC);
527*c03c5b1cSMartin Matuska }
528*c03c5b1cSMartin Matuska 
529*c03c5b1cSMartin Matuska 
530*c03c5b1cSMartin Matuska /* FSE_getMaxNbBits() :
531*c03c5b1cSMartin Matuska  * Approximate maximum cost of a symbol, in bits.
532*c03c5b1cSMartin Matuska  * Fractional get rounded up (i.e : a symbol with a normalized frequency of 3 gives the same result as a frequency of 2)
533*c03c5b1cSMartin Matuska  * note 1 : assume symbolValue is valid (<= maxSymbolValue)
534*c03c5b1cSMartin Matuska  * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */
FSE_getMaxNbBits(const void * symbolTTPtr,U32 symbolValue)535*c03c5b1cSMartin Matuska MEM_STATIC U32 FSE_getMaxNbBits(const void* symbolTTPtr, U32 symbolValue)
536*c03c5b1cSMartin Matuska {
537*c03c5b1cSMartin Matuska     const FSE_symbolCompressionTransform* symbolTT = (const FSE_symbolCompressionTransform*) symbolTTPtr;
538*c03c5b1cSMartin Matuska     return (symbolTT[symbolValue].deltaNbBits + ((1<<16)-1)) >> 16;
539*c03c5b1cSMartin Matuska }
540*c03c5b1cSMartin Matuska 
541*c03c5b1cSMartin Matuska /* FSE_bitCost() :
542*c03c5b1cSMartin Matuska  * Approximate symbol cost, as fractional value, using fixed-point format (accuracyLog fractional bits)
543*c03c5b1cSMartin Matuska  * note 1 : assume symbolValue is valid (<= maxSymbolValue)
544*c03c5b1cSMartin Matuska  * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */
FSE_bitCost(const void * symbolTTPtr,U32 tableLog,U32 symbolValue,U32 accuracyLog)545*c03c5b1cSMartin Matuska MEM_STATIC U32 FSE_bitCost(const void* symbolTTPtr, U32 tableLog, U32 symbolValue, U32 accuracyLog)
546*c03c5b1cSMartin Matuska {
547*c03c5b1cSMartin Matuska     const FSE_symbolCompressionTransform* symbolTT = (const FSE_symbolCompressionTransform*) symbolTTPtr;
548*c03c5b1cSMartin Matuska     U32 const minNbBits = symbolTT[symbolValue].deltaNbBits >> 16;
549*c03c5b1cSMartin Matuska     U32 const threshold = (minNbBits+1) << 16;
550*c03c5b1cSMartin Matuska     assert(tableLog < 16);
551*c03c5b1cSMartin Matuska     assert(accuracyLog < 31-tableLog);  /* ensure enough room for renormalization double shift */
552*c03c5b1cSMartin Matuska     {   U32 const tableSize = 1 << tableLog;
553*c03c5b1cSMartin Matuska         U32 const deltaFromThreshold = threshold - (symbolTT[symbolValue].deltaNbBits + tableSize);
554*c03c5b1cSMartin Matuska         U32 const normalizedDeltaFromThreshold = (deltaFromThreshold << accuracyLog) >> tableLog;   /* linear interpolation (very approximate) */
555*c03c5b1cSMartin Matuska         U32 const bitMultiplier = 1 << accuracyLog;
556*c03c5b1cSMartin Matuska         assert(symbolTT[symbolValue].deltaNbBits + tableSize <= threshold);
557*c03c5b1cSMartin Matuska         assert(normalizedDeltaFromThreshold <= bitMultiplier);
558*c03c5b1cSMartin Matuska         return (minNbBits+1)*bitMultiplier - normalizedDeltaFromThreshold;
559*c03c5b1cSMartin Matuska     }
560*c03c5b1cSMartin Matuska }
561*c03c5b1cSMartin Matuska 
562*c03c5b1cSMartin Matuska 
563*c03c5b1cSMartin Matuska /* ======    Decompression    ====== */
564*c03c5b1cSMartin Matuska 
565*c03c5b1cSMartin Matuska typedef struct {
566*c03c5b1cSMartin Matuska     U16 tableLog;
567*c03c5b1cSMartin Matuska     U16 fastMode;
568*c03c5b1cSMartin Matuska } FSE_DTableHeader;   /* sizeof U32 */
569*c03c5b1cSMartin Matuska 
570*c03c5b1cSMartin Matuska typedef struct
571*c03c5b1cSMartin Matuska {
572*c03c5b1cSMartin Matuska     unsigned short newState;
573*c03c5b1cSMartin Matuska     unsigned char  symbol;
574*c03c5b1cSMartin Matuska     unsigned char  nbBits;
575*c03c5b1cSMartin Matuska } FSE_decode_t;   /* size == U32 */
576*c03c5b1cSMartin Matuska 
FSE_initDState(FSE_DState_t * DStatePtr,BIT_DStream_t * bitD,const FSE_DTable * dt)577*c03c5b1cSMartin Matuska MEM_STATIC void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt)
578*c03c5b1cSMartin Matuska {
579*c03c5b1cSMartin Matuska     const void* ptr = dt;
580*c03c5b1cSMartin Matuska     const FSE_DTableHeader* const DTableH = (const FSE_DTableHeader*)ptr;
581*c03c5b1cSMartin Matuska     DStatePtr->state = BIT_readBits(bitD, DTableH->tableLog);
582*c03c5b1cSMartin Matuska     BIT_reloadDStream(bitD);
583*c03c5b1cSMartin Matuska     DStatePtr->table = dt + 1;
584*c03c5b1cSMartin Matuska }
585*c03c5b1cSMartin Matuska 
FSE_peekSymbol(const FSE_DState_t * DStatePtr)586*c03c5b1cSMartin Matuska MEM_STATIC BYTE FSE_peekSymbol(const FSE_DState_t* DStatePtr)
587*c03c5b1cSMartin Matuska {
588*c03c5b1cSMartin Matuska     FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
589*c03c5b1cSMartin Matuska     return DInfo.symbol;
590*c03c5b1cSMartin Matuska }
591*c03c5b1cSMartin Matuska 
FSE_updateState(FSE_DState_t * DStatePtr,BIT_DStream_t * bitD)592*c03c5b1cSMartin Matuska MEM_STATIC void FSE_updateState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
593*c03c5b1cSMartin Matuska {
594*c03c5b1cSMartin Matuska     FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
595*c03c5b1cSMartin Matuska     U32 const nbBits = DInfo.nbBits;
596*c03c5b1cSMartin Matuska     size_t const lowBits = BIT_readBits(bitD, nbBits);
597*c03c5b1cSMartin Matuska     DStatePtr->state = DInfo.newState + lowBits;
598*c03c5b1cSMartin Matuska }
599*c03c5b1cSMartin Matuska 
FSE_decodeSymbol(FSE_DState_t * DStatePtr,BIT_DStream_t * bitD)600*c03c5b1cSMartin Matuska MEM_STATIC BYTE FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
601*c03c5b1cSMartin Matuska {
602*c03c5b1cSMartin Matuska     FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
603*c03c5b1cSMartin Matuska     U32 const nbBits = DInfo.nbBits;
604*c03c5b1cSMartin Matuska     BYTE const symbol = DInfo.symbol;
605*c03c5b1cSMartin Matuska     size_t const lowBits = BIT_readBits(bitD, nbBits);
606*c03c5b1cSMartin Matuska 
607*c03c5b1cSMartin Matuska     DStatePtr->state = DInfo.newState + lowBits;
608*c03c5b1cSMartin Matuska     return symbol;
609*c03c5b1cSMartin Matuska }
610*c03c5b1cSMartin Matuska 
611*c03c5b1cSMartin Matuska /*! FSE_decodeSymbolFast() :
612*c03c5b1cSMartin Matuska     unsafe, only works if no symbol has a probability > 50% */
FSE_decodeSymbolFast(FSE_DState_t * DStatePtr,BIT_DStream_t * bitD)613*c03c5b1cSMartin Matuska MEM_STATIC BYTE FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
614*c03c5b1cSMartin Matuska {
615*c03c5b1cSMartin Matuska     FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
616*c03c5b1cSMartin Matuska     U32 const nbBits = DInfo.nbBits;
617*c03c5b1cSMartin Matuska     BYTE const symbol = DInfo.symbol;
618*c03c5b1cSMartin Matuska     size_t const lowBits = BIT_readBitsFast(bitD, nbBits);
619*c03c5b1cSMartin Matuska 
620*c03c5b1cSMartin Matuska     DStatePtr->state = DInfo.newState + lowBits;
621*c03c5b1cSMartin Matuska     return symbol;
622*c03c5b1cSMartin Matuska }
623*c03c5b1cSMartin Matuska 
FSE_endOfDState(const FSE_DState_t * DStatePtr)624*c03c5b1cSMartin Matuska MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
625*c03c5b1cSMartin Matuska {
626*c03c5b1cSMartin Matuska     return DStatePtr->state == 0;
627*c03c5b1cSMartin Matuska }
628*c03c5b1cSMartin Matuska 
629*c03c5b1cSMartin Matuska 
630*c03c5b1cSMartin Matuska 
631*c03c5b1cSMartin Matuska #ifndef FSE_COMMONDEFS_ONLY
632*c03c5b1cSMartin Matuska 
633*c03c5b1cSMartin Matuska /* **************************************************************
634*c03c5b1cSMartin Matuska *  Tuning parameters
635*c03c5b1cSMartin Matuska ****************************************************************/
636*c03c5b1cSMartin Matuska /*!MEMORY_USAGE :
637*c03c5b1cSMartin Matuska *  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
638*c03c5b1cSMartin Matuska *  Increasing memory usage improves compression ratio
639*c03c5b1cSMartin Matuska *  Reduced memory usage can improve speed, due to cache effect
640*c03c5b1cSMartin Matuska *  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
641*c03c5b1cSMartin Matuska #ifndef FSE_MAX_MEMORY_USAGE
642*c03c5b1cSMartin Matuska #  define FSE_MAX_MEMORY_USAGE 14
643*c03c5b1cSMartin Matuska #endif
644*c03c5b1cSMartin Matuska #ifndef FSE_DEFAULT_MEMORY_USAGE
645*c03c5b1cSMartin Matuska #  define FSE_DEFAULT_MEMORY_USAGE 13
646*c03c5b1cSMartin Matuska #endif
647*c03c5b1cSMartin Matuska 
648*c03c5b1cSMartin Matuska /*!FSE_MAX_SYMBOL_VALUE :
649*c03c5b1cSMartin Matuska *  Maximum symbol value authorized.
650*c03c5b1cSMartin Matuska *  Required for proper stack allocation */
651*c03c5b1cSMartin Matuska #ifndef FSE_MAX_SYMBOL_VALUE
652*c03c5b1cSMartin Matuska #  define FSE_MAX_SYMBOL_VALUE 255
653*c03c5b1cSMartin Matuska #endif
654*c03c5b1cSMartin Matuska 
655*c03c5b1cSMartin Matuska /* **************************************************************
656*c03c5b1cSMartin Matuska *  template functions type & suffix
657*c03c5b1cSMartin Matuska ****************************************************************/
658*c03c5b1cSMartin Matuska #define FSE_FUNCTION_TYPE BYTE
659*c03c5b1cSMartin Matuska #define FSE_FUNCTION_EXTENSION
660*c03c5b1cSMartin Matuska #define FSE_DECODE_TYPE FSE_decode_t
661*c03c5b1cSMartin Matuska 
662*c03c5b1cSMartin Matuska 
663*c03c5b1cSMartin Matuska #endif   /* !FSE_COMMONDEFS_ONLY */
664*c03c5b1cSMartin Matuska 
665*c03c5b1cSMartin Matuska 
666*c03c5b1cSMartin Matuska /* ***************************************************************
667*c03c5b1cSMartin Matuska *  Constants
668*c03c5b1cSMartin Matuska *****************************************************************/
669*c03c5b1cSMartin Matuska #define FSE_MAX_TABLELOG  (FSE_MAX_MEMORY_USAGE-2)
670*c03c5b1cSMartin Matuska #define FSE_MAX_TABLESIZE (1U<<FSE_MAX_TABLELOG)
671*c03c5b1cSMartin Matuska #define FSE_MAXTABLESIZE_MASK (FSE_MAX_TABLESIZE-1)
672*c03c5b1cSMartin Matuska #define FSE_DEFAULT_TABLELOG (FSE_DEFAULT_MEMORY_USAGE-2)
673*c03c5b1cSMartin Matuska #define FSE_MIN_TABLELOG 5
674*c03c5b1cSMartin Matuska 
675*c03c5b1cSMartin Matuska #define FSE_TABLELOG_ABSOLUTE_MAX 15
676*c03c5b1cSMartin Matuska #if FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX
677*c03c5b1cSMartin Matuska #  error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"
678*c03c5b1cSMartin Matuska #endif
679*c03c5b1cSMartin Matuska 
680*c03c5b1cSMartin Matuska #define FSE_TABLESTEP(tableSize) ((tableSize>>1) + (tableSize>>3) + 3)
681*c03c5b1cSMartin Matuska 
682*c03c5b1cSMartin Matuska 
683*c03c5b1cSMartin Matuska #endif /* FSE_STATIC_LINKING_ONLY */
684*c03c5b1cSMartin Matuska 
685*c03c5b1cSMartin Matuska 
686*c03c5b1cSMartin Matuska #if defined (__cplusplus)
687*c03c5b1cSMartin Matuska }
688*c03c5b1cSMartin Matuska #endif
689