xref: /dflybsd-src/contrib/zstd/lib/compress/fse_compress.c (revision a28cd43d19e8b720a6c852a4bbc5ae147a26165a)
1a28cd43dSSascha Wildner /* ******************************************************************
2a28cd43dSSascha Wildner  * FSE : Finite State Entropy encoder
3a28cd43dSSascha Wildner  * Copyright (c) 2013-2020, Yann Collet, Facebook, Inc.
4a28cd43dSSascha Wildner  *
5a28cd43dSSascha Wildner  *  You can contact the author at :
6a28cd43dSSascha Wildner  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
7a28cd43dSSascha Wildner  *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
8a28cd43dSSascha Wildner  *
9a28cd43dSSascha Wildner  * This source code is licensed under both the BSD-style license (found in the
10a28cd43dSSascha Wildner  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
11a28cd43dSSascha Wildner  * in the COPYING file in the root directory of this source tree).
12a28cd43dSSascha Wildner  * You may select, at your option, one of the above-listed licenses.
13a28cd43dSSascha Wildner ****************************************************************** */
14a28cd43dSSascha Wildner 
15a28cd43dSSascha Wildner /* **************************************************************
16a28cd43dSSascha Wildner *  Includes
17a28cd43dSSascha Wildner ****************************************************************/
18a28cd43dSSascha Wildner #include "../common/compiler.h"
19a28cd43dSSascha Wildner #include "../common/mem.h"        /* U32, U16, etc. */
20a28cd43dSSascha Wildner #include "../common/debug.h"      /* assert, DEBUGLOG */
21a28cd43dSSascha Wildner #include "hist.h"       /* HIST_count_wksp */
22a28cd43dSSascha Wildner #include "../common/bitstream.h"
23a28cd43dSSascha Wildner #define FSE_STATIC_LINKING_ONLY
24a28cd43dSSascha Wildner #include "../common/fse.h"
25a28cd43dSSascha Wildner #include "../common/error_private.h"
26a28cd43dSSascha Wildner #define ZSTD_DEPS_NEED_MALLOC
27a28cd43dSSascha Wildner #define ZSTD_DEPS_NEED_MATH64
28a28cd43dSSascha Wildner #include "../common/zstd_deps.h"  /* ZSTD_malloc, ZSTD_free, ZSTD_memcpy, ZSTD_memset */
29a28cd43dSSascha Wildner 
30a28cd43dSSascha Wildner 
31a28cd43dSSascha Wildner /* **************************************************************
32a28cd43dSSascha Wildner *  Error Management
33a28cd43dSSascha Wildner ****************************************************************/
34a28cd43dSSascha Wildner #define FSE_isError ERR_isError
35a28cd43dSSascha Wildner 
36a28cd43dSSascha Wildner 
37a28cd43dSSascha Wildner /* **************************************************************
38a28cd43dSSascha Wildner *  Templates
39a28cd43dSSascha Wildner ****************************************************************/
40a28cd43dSSascha Wildner /*
41a28cd43dSSascha Wildner   designed to be included
42a28cd43dSSascha Wildner   for type-specific functions (template emulation in C)
43a28cd43dSSascha Wildner   Objective is to write these functions only once, for improved maintenance
44a28cd43dSSascha Wildner */
45a28cd43dSSascha Wildner 
46a28cd43dSSascha Wildner /* safety checks */
47a28cd43dSSascha Wildner #ifndef FSE_FUNCTION_EXTENSION
48a28cd43dSSascha Wildner #  error "FSE_FUNCTION_EXTENSION must be defined"
49a28cd43dSSascha Wildner #endif
50a28cd43dSSascha Wildner #ifndef FSE_FUNCTION_TYPE
51a28cd43dSSascha Wildner #  error "FSE_FUNCTION_TYPE must be defined"
52a28cd43dSSascha Wildner #endif
53a28cd43dSSascha Wildner 
54a28cd43dSSascha Wildner /* Function names */
55a28cd43dSSascha Wildner #define FSE_CAT(X,Y) X##Y
56a28cd43dSSascha Wildner #define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
57a28cd43dSSascha Wildner #define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
58a28cd43dSSascha Wildner 
59a28cd43dSSascha Wildner 
60a28cd43dSSascha Wildner /* Function templates */
61a28cd43dSSascha Wildner 
62a28cd43dSSascha Wildner /* FSE_buildCTable_wksp() :
63a28cd43dSSascha Wildner  * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
64a28cd43dSSascha Wildner  * wkspSize should be sized to handle worst case situation, which is `1<<max_tableLog * sizeof(FSE_FUNCTION_TYPE)`
65a28cd43dSSascha Wildner  * workSpace must also be properly aligned with FSE_FUNCTION_TYPE requirements
66a28cd43dSSascha Wildner  */
FSE_buildCTable_wksp(FSE_CTable * ct,const short * normalizedCounter,unsigned maxSymbolValue,unsigned tableLog,void * workSpace,size_t wkspSize)67a28cd43dSSascha Wildner size_t FSE_buildCTable_wksp(FSE_CTable* ct,
68a28cd43dSSascha Wildner                       const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog,
69a28cd43dSSascha Wildner                             void* workSpace, size_t wkspSize)
70a28cd43dSSascha Wildner {
71a28cd43dSSascha Wildner     U32 const tableSize = 1 << tableLog;
72a28cd43dSSascha Wildner     U32 const tableMask = tableSize - 1;
73a28cd43dSSascha Wildner     void* const ptr = ct;
74a28cd43dSSascha Wildner     U16* const tableU16 = ( (U16*) ptr) + 2;
75a28cd43dSSascha Wildner     void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableLog ? tableSize>>1 : 1) ;
76a28cd43dSSascha Wildner     FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
77a28cd43dSSascha Wildner     U32 const step = FSE_TABLESTEP(tableSize);
78a28cd43dSSascha Wildner 
79a28cd43dSSascha Wildner     U32* cumul = (U32*)workSpace;
80a28cd43dSSascha Wildner     FSE_FUNCTION_TYPE* tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSymbolValue + 2));
81a28cd43dSSascha Wildner 
82a28cd43dSSascha Wildner     U32 highThreshold = tableSize-1;
83a28cd43dSSascha Wildner 
84a28cd43dSSascha Wildner     if ((size_t)workSpace & 3) return ERROR(GENERIC); /* Must be 4 byte aligned */
85a28cd43dSSascha Wildner     if (FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) > wkspSize) return ERROR(tableLog_tooLarge);
86a28cd43dSSascha Wildner     /* CTable header */
87a28cd43dSSascha Wildner     tableU16[-2] = (U16) tableLog;
88a28cd43dSSascha Wildner     tableU16[-1] = (U16) maxSymbolValue;
89a28cd43dSSascha Wildner     assert(tableLog < 16);   /* required for threshold strategy to work */
90a28cd43dSSascha Wildner 
91a28cd43dSSascha Wildner     /* For explanations on how to distribute symbol values over the table :
92a28cd43dSSascha Wildner      * http://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */
93a28cd43dSSascha Wildner 
94a28cd43dSSascha Wildner      #ifdef __clang_analyzer__
95a28cd43dSSascha Wildner      ZSTD_memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize);   /* useless initialization, just to keep scan-build happy */
96a28cd43dSSascha Wildner      #endif
97a28cd43dSSascha Wildner 
98a28cd43dSSascha Wildner     /* symbol start positions */
99a28cd43dSSascha Wildner     {   U32 u;
100a28cd43dSSascha Wildner         cumul[0] = 0;
101a28cd43dSSascha Wildner         for (u=1; u <= maxSymbolValue+1; u++) {
102a28cd43dSSascha Wildner             if (normalizedCounter[u-1]==-1) {  /* Low proba symbol */
103a28cd43dSSascha Wildner                 cumul[u] = cumul[u-1] + 1;
104a28cd43dSSascha Wildner                 tableSymbol[highThreshold--] = (FSE_FUNCTION_TYPE)(u-1);
105a28cd43dSSascha Wildner             } else {
106a28cd43dSSascha Wildner                 cumul[u] = cumul[u-1] + normalizedCounter[u-1];
107a28cd43dSSascha Wildner         }   }
108a28cd43dSSascha Wildner         cumul[maxSymbolValue+1] = tableSize+1;
109a28cd43dSSascha Wildner     }
110a28cd43dSSascha Wildner 
111a28cd43dSSascha Wildner     /* Spread symbols */
112a28cd43dSSascha Wildner     {   U32 position = 0;
113a28cd43dSSascha Wildner         U32 symbol;
114a28cd43dSSascha Wildner         for (symbol=0; symbol<=maxSymbolValue; symbol++) {
115a28cd43dSSascha Wildner             int nbOccurrences;
116a28cd43dSSascha Wildner             int const freq = normalizedCounter[symbol];
117a28cd43dSSascha Wildner             for (nbOccurrences=0; nbOccurrences<freq; nbOccurrences++) {
118a28cd43dSSascha Wildner                 tableSymbol[position] = (FSE_FUNCTION_TYPE)symbol;
119a28cd43dSSascha Wildner                 position = (position + step) & tableMask;
120a28cd43dSSascha Wildner                 while (position > highThreshold)
121a28cd43dSSascha Wildner                     position = (position + step) & tableMask;   /* Low proba area */
122a28cd43dSSascha Wildner         }   }
123a28cd43dSSascha Wildner 
124a28cd43dSSascha Wildner         assert(position==0);  /* Must have initialized all positions */
125a28cd43dSSascha Wildner     }
126a28cd43dSSascha Wildner 
127a28cd43dSSascha Wildner     /* Build table */
128a28cd43dSSascha Wildner     {   U32 u; for (u=0; u<tableSize; u++) {
129a28cd43dSSascha Wildner         FSE_FUNCTION_TYPE s = tableSymbol[u];   /* note : static analyzer may not understand tableSymbol is properly initialized */
130a28cd43dSSascha Wildner         tableU16[cumul[s]++] = (U16) (tableSize+u);   /* TableU16 : sorted by symbol order; gives next state value */
131a28cd43dSSascha Wildner     }   }
132a28cd43dSSascha Wildner 
133a28cd43dSSascha Wildner     /* Build Symbol Transformation Table */
134a28cd43dSSascha Wildner     {   unsigned total = 0;
135a28cd43dSSascha Wildner         unsigned s;
136a28cd43dSSascha Wildner         for (s=0; s<=maxSymbolValue; s++) {
137a28cd43dSSascha Wildner             switch (normalizedCounter[s])
138a28cd43dSSascha Wildner             {
139a28cd43dSSascha Wildner             case  0:
140a28cd43dSSascha Wildner                 /* filling nonetheless, for compatibility with FSE_getMaxNbBits() */
141a28cd43dSSascha Wildner                 symbolTT[s].deltaNbBits = ((tableLog+1) << 16) - (1<<tableLog);
142a28cd43dSSascha Wildner                 break;
143a28cd43dSSascha Wildner 
144a28cd43dSSascha Wildner             case -1:
145a28cd43dSSascha Wildner             case  1:
146a28cd43dSSascha Wildner                 symbolTT[s].deltaNbBits = (tableLog << 16) - (1<<tableLog);
147a28cd43dSSascha Wildner                 symbolTT[s].deltaFindState = total - 1;
148a28cd43dSSascha Wildner                 total ++;
149a28cd43dSSascha Wildner                 break;
150a28cd43dSSascha Wildner             default :
151a28cd43dSSascha Wildner                 {
152a28cd43dSSascha Wildner                     U32 const maxBitsOut = tableLog - BIT_highbit32 (normalizedCounter[s]-1);
153a28cd43dSSascha Wildner                     U32 const minStatePlus = normalizedCounter[s] << maxBitsOut;
154a28cd43dSSascha Wildner                     symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus;
155a28cd43dSSascha Wildner                     symbolTT[s].deltaFindState = total - normalizedCounter[s];
156a28cd43dSSascha Wildner                     total +=  normalizedCounter[s];
157a28cd43dSSascha Wildner     }   }   }   }
158a28cd43dSSascha Wildner 
159a28cd43dSSascha Wildner #if 0  /* debug : symbol costs */
160a28cd43dSSascha Wildner     DEBUGLOG(5, "\n --- table statistics : ");
161a28cd43dSSascha Wildner     {   U32 symbol;
162a28cd43dSSascha Wildner         for (symbol=0; symbol<=maxSymbolValue; symbol++) {
163a28cd43dSSascha Wildner             DEBUGLOG(5, "%3u: w=%3i,   maxBits=%u, fracBits=%.2f",
164a28cd43dSSascha Wildner                 symbol, normalizedCounter[symbol],
165a28cd43dSSascha Wildner                 FSE_getMaxNbBits(symbolTT, symbol),
166a28cd43dSSascha Wildner                 (double)FSE_bitCost(symbolTT, tableLog, symbol, 8) / 256);
167a28cd43dSSascha Wildner         }
168a28cd43dSSascha Wildner     }
169a28cd43dSSascha Wildner #endif
170a28cd43dSSascha Wildner 
171a28cd43dSSascha Wildner     return 0;
172a28cd43dSSascha Wildner }
173a28cd43dSSascha Wildner 
174a28cd43dSSascha Wildner #ifndef ZSTD_NO_UNUSED_FUNCTIONS
FSE_buildCTable(FSE_CTable * ct,const short * normalizedCounter,unsigned maxSymbolValue,unsigned tableLog)175a28cd43dSSascha Wildner size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
176a28cd43dSSascha Wildner {
177a28cd43dSSascha Wildner     FSE_FUNCTION_TYPE tableSymbol[FSE_MAX_TABLESIZE];   /* memset() is not necessary, even if static analyzer complain about it */
178a28cd43dSSascha Wildner     return FSE_buildCTable_wksp(ct, normalizedCounter, maxSymbolValue, tableLog, tableSymbol, sizeof(tableSymbol));
179a28cd43dSSascha Wildner }
180a28cd43dSSascha Wildner #endif
181a28cd43dSSascha Wildner 
182a28cd43dSSascha Wildner 
183a28cd43dSSascha Wildner 
184a28cd43dSSascha Wildner #ifndef FSE_COMMONDEFS_ONLY
185a28cd43dSSascha Wildner 
186a28cd43dSSascha Wildner 
187a28cd43dSSascha Wildner /*-**************************************************************
188a28cd43dSSascha Wildner *  FSE NCount encoding
189a28cd43dSSascha Wildner ****************************************************************/
FSE_NCountWriteBound(unsigned maxSymbolValue,unsigned tableLog)190a28cd43dSSascha Wildner size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog)
191a28cd43dSSascha Wildner {
192a28cd43dSSascha Wildner     size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog) >> 3) + 3;
193a28cd43dSSascha Wildner     return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND;  /* maxSymbolValue==0 ? use default */
194a28cd43dSSascha Wildner }
195a28cd43dSSascha Wildner 
196a28cd43dSSascha Wildner static size_t
FSE_writeNCount_generic(void * header,size_t headerBufferSize,const short * normalizedCounter,unsigned maxSymbolValue,unsigned tableLog,unsigned writeIsSafe)197a28cd43dSSascha Wildner FSE_writeNCount_generic (void* header, size_t headerBufferSize,
198a28cd43dSSascha Wildner                    const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog,
199a28cd43dSSascha Wildner                          unsigned writeIsSafe)
200a28cd43dSSascha Wildner {
201a28cd43dSSascha Wildner     BYTE* const ostart = (BYTE*) header;
202a28cd43dSSascha Wildner     BYTE* out = ostart;
203a28cd43dSSascha Wildner     BYTE* const oend = ostart + headerBufferSize;
204a28cd43dSSascha Wildner     int nbBits;
205a28cd43dSSascha Wildner     const int tableSize = 1 << tableLog;
206a28cd43dSSascha Wildner     int remaining;
207a28cd43dSSascha Wildner     int threshold;
208a28cd43dSSascha Wildner     U32 bitStream = 0;
209a28cd43dSSascha Wildner     int bitCount = 0;
210a28cd43dSSascha Wildner     unsigned symbol = 0;
211a28cd43dSSascha Wildner     unsigned const alphabetSize = maxSymbolValue + 1;
212a28cd43dSSascha Wildner     int previousIs0 = 0;
213a28cd43dSSascha Wildner 
214a28cd43dSSascha Wildner     /* Table Size */
215a28cd43dSSascha Wildner     bitStream += (tableLog-FSE_MIN_TABLELOG) << bitCount;
216a28cd43dSSascha Wildner     bitCount  += 4;
217a28cd43dSSascha Wildner 
218a28cd43dSSascha Wildner     /* Init */
219a28cd43dSSascha Wildner     remaining = tableSize+1;   /* +1 for extra accuracy */
220a28cd43dSSascha Wildner     threshold = tableSize;
221a28cd43dSSascha Wildner     nbBits = tableLog+1;
222a28cd43dSSascha Wildner 
223a28cd43dSSascha Wildner     while ((symbol < alphabetSize) && (remaining>1)) {  /* stops at 1 */
224a28cd43dSSascha Wildner         if (previousIs0) {
225a28cd43dSSascha Wildner             unsigned start = symbol;
226a28cd43dSSascha Wildner             while ((symbol < alphabetSize) && !normalizedCounter[symbol]) symbol++;
227a28cd43dSSascha Wildner             if (symbol == alphabetSize) break;   /* incorrect distribution */
228a28cd43dSSascha Wildner             while (symbol >= start+24) {
229a28cd43dSSascha Wildner                 start+=24;
230a28cd43dSSascha Wildner                 bitStream += 0xFFFFU << bitCount;
231a28cd43dSSascha Wildner                 if ((!writeIsSafe) && (out > oend-2))
232a28cd43dSSascha Wildner                     return ERROR(dstSize_tooSmall);   /* Buffer overflow */
233a28cd43dSSascha Wildner                 out[0] = (BYTE) bitStream;
234a28cd43dSSascha Wildner                 out[1] = (BYTE)(bitStream>>8);
235a28cd43dSSascha Wildner                 out+=2;
236a28cd43dSSascha Wildner                 bitStream>>=16;
237a28cd43dSSascha Wildner             }
238a28cd43dSSascha Wildner             while (symbol >= start+3) {
239a28cd43dSSascha Wildner                 start+=3;
240a28cd43dSSascha Wildner                 bitStream += 3 << bitCount;
241a28cd43dSSascha Wildner                 bitCount += 2;
242a28cd43dSSascha Wildner             }
243a28cd43dSSascha Wildner             bitStream += (symbol-start) << bitCount;
244a28cd43dSSascha Wildner             bitCount += 2;
245a28cd43dSSascha Wildner             if (bitCount>16) {
246a28cd43dSSascha Wildner                 if ((!writeIsSafe) && (out > oend - 2))
247a28cd43dSSascha Wildner                     return ERROR(dstSize_tooSmall);   /* Buffer overflow */
248a28cd43dSSascha Wildner                 out[0] = (BYTE)bitStream;
249a28cd43dSSascha Wildner                 out[1] = (BYTE)(bitStream>>8);
250a28cd43dSSascha Wildner                 out += 2;
251a28cd43dSSascha Wildner                 bitStream >>= 16;
252a28cd43dSSascha Wildner                 bitCount -= 16;
253a28cd43dSSascha Wildner         }   }
254a28cd43dSSascha Wildner         {   int count = normalizedCounter[symbol++];
255a28cd43dSSascha Wildner             int const max = (2*threshold-1) - remaining;
256a28cd43dSSascha Wildner             remaining -= count < 0 ? -count : count;
257a28cd43dSSascha Wildner             count++;   /* +1 for extra accuracy */
258a28cd43dSSascha Wildner             if (count>=threshold)
259a28cd43dSSascha Wildner                 count += max;   /* [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[ */
260a28cd43dSSascha Wildner             bitStream += count << bitCount;
261a28cd43dSSascha Wildner             bitCount  += nbBits;
262a28cd43dSSascha Wildner             bitCount  -= (count<max);
263a28cd43dSSascha Wildner             previousIs0  = (count==1);
264a28cd43dSSascha Wildner             if (remaining<1) return ERROR(GENERIC);
265a28cd43dSSascha Wildner             while (remaining<threshold) { nbBits--; threshold>>=1; }
266a28cd43dSSascha Wildner         }
267a28cd43dSSascha Wildner         if (bitCount>16) {
268a28cd43dSSascha Wildner             if ((!writeIsSafe) && (out > oend - 2))
269a28cd43dSSascha Wildner                 return ERROR(dstSize_tooSmall);   /* Buffer overflow */
270a28cd43dSSascha Wildner             out[0] = (BYTE)bitStream;
271a28cd43dSSascha Wildner             out[1] = (BYTE)(bitStream>>8);
272a28cd43dSSascha Wildner             out += 2;
273a28cd43dSSascha Wildner             bitStream >>= 16;
274a28cd43dSSascha Wildner             bitCount -= 16;
275a28cd43dSSascha Wildner     }   }
276a28cd43dSSascha Wildner 
277a28cd43dSSascha Wildner     if (remaining != 1)
278a28cd43dSSascha Wildner         return ERROR(GENERIC);  /* incorrect normalized distribution */
279a28cd43dSSascha Wildner     assert(symbol <= alphabetSize);
280a28cd43dSSascha Wildner 
281a28cd43dSSascha Wildner     /* flush remaining bitStream */
282a28cd43dSSascha Wildner     if ((!writeIsSafe) && (out > oend - 2))
283a28cd43dSSascha Wildner         return ERROR(dstSize_tooSmall);   /* Buffer overflow */
284a28cd43dSSascha Wildner     out[0] = (BYTE)bitStream;
285a28cd43dSSascha Wildner     out[1] = (BYTE)(bitStream>>8);
286a28cd43dSSascha Wildner     out+= (bitCount+7) /8;
287a28cd43dSSascha Wildner 
288a28cd43dSSascha Wildner     return (out-ostart);
289a28cd43dSSascha Wildner }
290a28cd43dSSascha Wildner 
291a28cd43dSSascha Wildner 
FSE_writeNCount(void * buffer,size_t bufferSize,const short * normalizedCounter,unsigned maxSymbolValue,unsigned tableLog)292a28cd43dSSascha Wildner size_t FSE_writeNCount (void* buffer, size_t bufferSize,
293a28cd43dSSascha Wildner                   const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
294a28cd43dSSascha Wildner {
295a28cd43dSSascha Wildner     if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);   /* Unsupported */
296a28cd43dSSascha Wildner     if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC);   /* Unsupported */
297a28cd43dSSascha Wildner 
298a28cd43dSSascha Wildner     if (bufferSize < FSE_NCountWriteBound(maxSymbolValue, tableLog))
299a28cd43dSSascha Wildner         return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 0);
300a28cd43dSSascha Wildner 
301a28cd43dSSascha Wildner     return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 1 /* write in buffer is safe */);
302a28cd43dSSascha Wildner }
303a28cd43dSSascha Wildner 
304a28cd43dSSascha Wildner 
305a28cd43dSSascha Wildner /*-**************************************************************
306a28cd43dSSascha Wildner *  FSE Compression Code
307a28cd43dSSascha Wildner ****************************************************************/
308a28cd43dSSascha Wildner 
FSE_createCTable(unsigned maxSymbolValue,unsigned tableLog)309a28cd43dSSascha Wildner FSE_CTable* FSE_createCTable (unsigned maxSymbolValue, unsigned tableLog)
310a28cd43dSSascha Wildner {
311a28cd43dSSascha Wildner     size_t size;
312a28cd43dSSascha Wildner     if (tableLog > FSE_TABLELOG_ABSOLUTE_MAX) tableLog = FSE_TABLELOG_ABSOLUTE_MAX;
313a28cd43dSSascha Wildner     size = FSE_CTABLE_SIZE_U32 (tableLog, maxSymbolValue) * sizeof(U32);
314a28cd43dSSascha Wildner     return (FSE_CTable*)ZSTD_malloc(size);
315a28cd43dSSascha Wildner }
316a28cd43dSSascha Wildner 
FSE_freeCTable(FSE_CTable * ct)317a28cd43dSSascha Wildner void FSE_freeCTable (FSE_CTable* ct) { ZSTD_free(ct); }
318a28cd43dSSascha Wildner 
319a28cd43dSSascha Wildner /* provides the minimum logSize to safely represent a distribution */
FSE_minTableLog(size_t srcSize,unsigned maxSymbolValue)320a28cd43dSSascha Wildner static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue)
321a28cd43dSSascha Wildner {
322a28cd43dSSascha Wildner     U32 minBitsSrc = BIT_highbit32((U32)(srcSize)) + 1;
323a28cd43dSSascha Wildner     U32 minBitsSymbols = BIT_highbit32(maxSymbolValue) + 2;
324a28cd43dSSascha Wildner     U32 minBits = minBitsSrc < minBitsSymbols ? minBitsSrc : minBitsSymbols;
325a28cd43dSSascha Wildner     assert(srcSize > 1); /* Not supported, RLE should be used instead */
326a28cd43dSSascha Wildner     return minBits;
327a28cd43dSSascha Wildner }
328a28cd43dSSascha Wildner 
FSE_optimalTableLog_internal(unsigned maxTableLog,size_t srcSize,unsigned maxSymbolValue,unsigned minus)329a28cd43dSSascha Wildner unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus)
330a28cd43dSSascha Wildner {
331a28cd43dSSascha Wildner     U32 maxBitsSrc = BIT_highbit32((U32)(srcSize - 1)) - minus;
332a28cd43dSSascha Wildner     U32 tableLog = maxTableLog;
333a28cd43dSSascha Wildner     U32 minBits = FSE_minTableLog(srcSize, maxSymbolValue);
334a28cd43dSSascha Wildner     assert(srcSize > 1); /* Not supported, RLE should be used instead */
335a28cd43dSSascha Wildner     if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
336a28cd43dSSascha Wildner     if (maxBitsSrc < tableLog) tableLog = maxBitsSrc;   /* Accuracy can be reduced */
337a28cd43dSSascha Wildner     if (minBits > tableLog) tableLog = minBits;   /* Need a minimum to safely represent all symbol values */
338a28cd43dSSascha Wildner     if (tableLog < FSE_MIN_TABLELOG) tableLog = FSE_MIN_TABLELOG;
339a28cd43dSSascha Wildner     if (tableLog > FSE_MAX_TABLELOG) tableLog = FSE_MAX_TABLELOG;
340a28cd43dSSascha Wildner     return tableLog;
341a28cd43dSSascha Wildner }
342a28cd43dSSascha Wildner 
FSE_optimalTableLog(unsigned maxTableLog,size_t srcSize,unsigned maxSymbolValue)343a28cd43dSSascha Wildner unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
344a28cd43dSSascha Wildner {
345a28cd43dSSascha Wildner     return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 2);
346a28cd43dSSascha Wildner }
347a28cd43dSSascha Wildner 
348a28cd43dSSascha Wildner /* Secondary normalization method.
349a28cd43dSSascha Wildner    To be used when primary method fails. */
350a28cd43dSSascha Wildner 
FSE_normalizeM2(short * norm,U32 tableLog,const unsigned * count,size_t total,U32 maxSymbolValue,short lowProbCount)351a28cd43dSSascha Wildner static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue, short lowProbCount)
352a28cd43dSSascha Wildner {
353a28cd43dSSascha Wildner     short const NOT_YET_ASSIGNED = -2;
354a28cd43dSSascha Wildner     U32 s;
355a28cd43dSSascha Wildner     U32 distributed = 0;
356a28cd43dSSascha Wildner     U32 ToDistribute;
357a28cd43dSSascha Wildner 
358a28cd43dSSascha Wildner     /* Init */
359a28cd43dSSascha Wildner     U32 const lowThreshold = (U32)(total >> tableLog);
360a28cd43dSSascha Wildner     U32 lowOne = (U32)((total * 3) >> (tableLog + 1));
361a28cd43dSSascha Wildner 
362a28cd43dSSascha Wildner     for (s=0; s<=maxSymbolValue; s++) {
363a28cd43dSSascha Wildner         if (count[s] == 0) {
364a28cd43dSSascha Wildner             norm[s]=0;
365a28cd43dSSascha Wildner             continue;
366a28cd43dSSascha Wildner         }
367a28cd43dSSascha Wildner         if (count[s] <= lowThreshold) {
368a28cd43dSSascha Wildner             norm[s] = lowProbCount;
369a28cd43dSSascha Wildner             distributed++;
370a28cd43dSSascha Wildner             total -= count[s];
371a28cd43dSSascha Wildner             continue;
372a28cd43dSSascha Wildner         }
373a28cd43dSSascha Wildner         if (count[s] <= lowOne) {
374a28cd43dSSascha Wildner             norm[s] = 1;
375a28cd43dSSascha Wildner             distributed++;
376a28cd43dSSascha Wildner             total -= count[s];
377a28cd43dSSascha Wildner             continue;
378a28cd43dSSascha Wildner         }
379a28cd43dSSascha Wildner 
380a28cd43dSSascha Wildner         norm[s]=NOT_YET_ASSIGNED;
381a28cd43dSSascha Wildner     }
382a28cd43dSSascha Wildner     ToDistribute = (1 << tableLog) - distributed;
383a28cd43dSSascha Wildner 
384a28cd43dSSascha Wildner     if (ToDistribute == 0)
385a28cd43dSSascha Wildner         return 0;
386a28cd43dSSascha Wildner 
387a28cd43dSSascha Wildner     if ((total / ToDistribute) > lowOne) {
388a28cd43dSSascha Wildner         /* risk of rounding to zero */
389a28cd43dSSascha Wildner         lowOne = (U32)((total * 3) / (ToDistribute * 2));
390a28cd43dSSascha Wildner         for (s=0; s<=maxSymbolValue; s++) {
391a28cd43dSSascha Wildner             if ((norm[s] == NOT_YET_ASSIGNED) && (count[s] <= lowOne)) {
392a28cd43dSSascha Wildner                 norm[s] = 1;
393a28cd43dSSascha Wildner                 distributed++;
394a28cd43dSSascha Wildner                 total -= count[s];
395a28cd43dSSascha Wildner                 continue;
396a28cd43dSSascha Wildner         }   }
397a28cd43dSSascha Wildner         ToDistribute = (1 << tableLog) - distributed;
398a28cd43dSSascha Wildner     }
399a28cd43dSSascha Wildner 
400a28cd43dSSascha Wildner     if (distributed == maxSymbolValue+1) {
401a28cd43dSSascha Wildner         /* all values are pretty poor;
402a28cd43dSSascha Wildner            probably incompressible data (should have already been detected);
403a28cd43dSSascha Wildner            find max, then give all remaining points to max */
404a28cd43dSSascha Wildner         U32 maxV = 0, maxC = 0;
405a28cd43dSSascha Wildner         for (s=0; s<=maxSymbolValue; s++)
406a28cd43dSSascha Wildner             if (count[s] > maxC) { maxV=s; maxC=count[s]; }
407a28cd43dSSascha Wildner         norm[maxV] += (short)ToDistribute;
408a28cd43dSSascha Wildner         return 0;
409a28cd43dSSascha Wildner     }
410a28cd43dSSascha Wildner 
411a28cd43dSSascha Wildner     if (total == 0) {
412a28cd43dSSascha Wildner         /* all of the symbols were low enough for the lowOne or lowThreshold */
413a28cd43dSSascha Wildner         for (s=0; ToDistribute > 0; s = (s+1)%(maxSymbolValue+1))
414a28cd43dSSascha Wildner             if (norm[s] > 0) { ToDistribute--; norm[s]++; }
415a28cd43dSSascha Wildner         return 0;
416a28cd43dSSascha Wildner     }
417a28cd43dSSascha Wildner 
418a28cd43dSSascha Wildner     {   U64 const vStepLog = 62 - tableLog;
419a28cd43dSSascha Wildner         U64 const mid = (1ULL << (vStepLog-1)) - 1;
420a28cd43dSSascha Wildner         U64 const rStep = ZSTD_div64((((U64)1<<vStepLog) * ToDistribute) + mid, (U32)total);   /* scale on remaining */
421a28cd43dSSascha Wildner         U64 tmpTotal = mid;
422a28cd43dSSascha Wildner         for (s=0; s<=maxSymbolValue; s++) {
423a28cd43dSSascha Wildner             if (norm[s]==NOT_YET_ASSIGNED) {
424a28cd43dSSascha Wildner                 U64 const end = tmpTotal + (count[s] * rStep);
425a28cd43dSSascha Wildner                 U32 const sStart = (U32)(tmpTotal >> vStepLog);
426a28cd43dSSascha Wildner                 U32 const sEnd = (U32)(end >> vStepLog);
427a28cd43dSSascha Wildner                 U32 const weight = sEnd - sStart;
428a28cd43dSSascha Wildner                 if (weight < 1)
429a28cd43dSSascha Wildner                     return ERROR(GENERIC);
430a28cd43dSSascha Wildner                 norm[s] = (short)weight;
431a28cd43dSSascha Wildner                 tmpTotal = end;
432a28cd43dSSascha Wildner     }   }   }
433a28cd43dSSascha Wildner 
434a28cd43dSSascha Wildner     return 0;
435a28cd43dSSascha Wildner }
436a28cd43dSSascha Wildner 
FSE_normalizeCount(short * normalizedCounter,unsigned tableLog,const unsigned * count,size_t total,unsigned maxSymbolValue,unsigned useLowProbCount)437a28cd43dSSascha Wildner size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
438a28cd43dSSascha Wildner                            const unsigned* count, size_t total,
439a28cd43dSSascha Wildner                            unsigned maxSymbolValue, unsigned useLowProbCount)
440a28cd43dSSascha Wildner {
441a28cd43dSSascha Wildner     /* Sanity checks */
442a28cd43dSSascha Wildner     if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
443a28cd43dSSascha Wildner     if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC);   /* Unsupported size */
444a28cd43dSSascha Wildner     if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);   /* Unsupported size */
445a28cd43dSSascha Wildner     if (tableLog < FSE_minTableLog(total, maxSymbolValue)) return ERROR(GENERIC);   /* Too small tableLog, compression potentially impossible */
446a28cd43dSSascha Wildner 
447a28cd43dSSascha Wildner     {   static U32 const rtbTable[] = {     0, 473195, 504333, 520860, 550000, 700000, 750000, 830000 };
448a28cd43dSSascha Wildner         short const lowProbCount = useLowProbCount ? -1 : 1;
449a28cd43dSSascha Wildner         U64 const scale = 62 - tableLog;
450a28cd43dSSascha Wildner         U64 const step = ZSTD_div64((U64)1<<62, (U32)total);   /* <== here, one division ! */
451a28cd43dSSascha Wildner         U64 const vStep = 1ULL<<(scale-20);
452a28cd43dSSascha Wildner         int stillToDistribute = 1<<tableLog;
453a28cd43dSSascha Wildner         unsigned s;
454a28cd43dSSascha Wildner         unsigned largest=0;
455a28cd43dSSascha Wildner         short largestP=0;
456a28cd43dSSascha Wildner         U32 lowThreshold = (U32)(total >> tableLog);
457a28cd43dSSascha Wildner 
458a28cd43dSSascha Wildner         for (s=0; s<=maxSymbolValue; s++) {
459a28cd43dSSascha Wildner             if (count[s] == total) return 0;   /* rle special case */
460a28cd43dSSascha Wildner             if (count[s] == 0) { normalizedCounter[s]=0; continue; }
461a28cd43dSSascha Wildner             if (count[s] <= lowThreshold) {
462a28cd43dSSascha Wildner                 normalizedCounter[s] = lowProbCount;
463a28cd43dSSascha Wildner                 stillToDistribute--;
464a28cd43dSSascha Wildner             } else {
465a28cd43dSSascha Wildner                 short proba = (short)((count[s]*step) >> scale);
466a28cd43dSSascha Wildner                 if (proba<8) {
467a28cd43dSSascha Wildner                     U64 restToBeat = vStep * rtbTable[proba];
468a28cd43dSSascha Wildner                     proba += (count[s]*step) - ((U64)proba<<scale) > restToBeat;
469a28cd43dSSascha Wildner                 }
470a28cd43dSSascha Wildner                 if (proba > largestP) { largestP=proba; largest=s; }
471a28cd43dSSascha Wildner                 normalizedCounter[s] = proba;
472a28cd43dSSascha Wildner                 stillToDistribute -= proba;
473a28cd43dSSascha Wildner         }   }
474a28cd43dSSascha Wildner         if (-stillToDistribute >= (normalizedCounter[largest] >> 1)) {
475a28cd43dSSascha Wildner             /* corner case, need another normalization method */
476a28cd43dSSascha Wildner             size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue, lowProbCount);
477a28cd43dSSascha Wildner             if (FSE_isError(errorCode)) return errorCode;
478a28cd43dSSascha Wildner         }
479a28cd43dSSascha Wildner         else normalizedCounter[largest] += (short)stillToDistribute;
480a28cd43dSSascha Wildner     }
481a28cd43dSSascha Wildner 
482a28cd43dSSascha Wildner #if 0
483a28cd43dSSascha Wildner     {   /* Print Table (debug) */
484a28cd43dSSascha Wildner         U32 s;
485a28cd43dSSascha Wildner         U32 nTotal = 0;
486a28cd43dSSascha Wildner         for (s=0; s<=maxSymbolValue; s++)
487a28cd43dSSascha Wildner             RAWLOG(2, "%3i: %4i \n", s, normalizedCounter[s]);
488a28cd43dSSascha Wildner         for (s=0; s<=maxSymbolValue; s++)
489a28cd43dSSascha Wildner             nTotal += abs(normalizedCounter[s]);
490a28cd43dSSascha Wildner         if (nTotal != (1U<<tableLog))
491a28cd43dSSascha Wildner             RAWLOG(2, "Warning !!! Total == %u != %u !!!", nTotal, 1U<<tableLog);
492a28cd43dSSascha Wildner         getchar();
493a28cd43dSSascha Wildner     }
494a28cd43dSSascha Wildner #endif
495a28cd43dSSascha Wildner 
496a28cd43dSSascha Wildner     return tableLog;
497a28cd43dSSascha Wildner }
498a28cd43dSSascha Wildner 
499a28cd43dSSascha Wildner 
500a28cd43dSSascha Wildner /* fake FSE_CTable, for raw (uncompressed) input */
FSE_buildCTable_raw(FSE_CTable * ct,unsigned nbBits)501a28cd43dSSascha Wildner size_t FSE_buildCTable_raw (FSE_CTable* ct, unsigned nbBits)
502a28cd43dSSascha Wildner {
503a28cd43dSSascha Wildner     const unsigned tableSize = 1 << nbBits;
504a28cd43dSSascha Wildner     const unsigned tableMask = tableSize - 1;
505a28cd43dSSascha Wildner     const unsigned maxSymbolValue = tableMask;
506a28cd43dSSascha Wildner     void* const ptr = ct;
507a28cd43dSSascha Wildner     U16* const tableU16 = ( (U16*) ptr) + 2;
508a28cd43dSSascha Wildner     void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableSize>>1);   /* assumption : tableLog >= 1 */
509a28cd43dSSascha Wildner     FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
510a28cd43dSSascha Wildner     unsigned s;
511a28cd43dSSascha Wildner 
512a28cd43dSSascha Wildner     /* Sanity checks */
513a28cd43dSSascha Wildner     if (nbBits < 1) return ERROR(GENERIC);             /* min size */
514a28cd43dSSascha Wildner 
515a28cd43dSSascha Wildner     /* header */
516a28cd43dSSascha Wildner     tableU16[-2] = (U16) nbBits;
517a28cd43dSSascha Wildner     tableU16[-1] = (U16) maxSymbolValue;
518a28cd43dSSascha Wildner 
519a28cd43dSSascha Wildner     /* Build table */
520a28cd43dSSascha Wildner     for (s=0; s<tableSize; s++)
521a28cd43dSSascha Wildner         tableU16[s] = (U16)(tableSize + s);
522a28cd43dSSascha Wildner 
523a28cd43dSSascha Wildner     /* Build Symbol Transformation Table */
524a28cd43dSSascha Wildner     {   const U32 deltaNbBits = (nbBits << 16) - (1 << nbBits);
525a28cd43dSSascha Wildner         for (s=0; s<=maxSymbolValue; s++) {
526a28cd43dSSascha Wildner             symbolTT[s].deltaNbBits = deltaNbBits;
527a28cd43dSSascha Wildner             symbolTT[s].deltaFindState = s-1;
528a28cd43dSSascha Wildner     }   }
529a28cd43dSSascha Wildner 
530a28cd43dSSascha Wildner     return 0;
531a28cd43dSSascha Wildner }
532a28cd43dSSascha Wildner 
533a28cd43dSSascha Wildner /* fake FSE_CTable, for rle input (always same symbol) */
FSE_buildCTable_rle(FSE_CTable * ct,BYTE symbolValue)534a28cd43dSSascha Wildner size_t FSE_buildCTable_rle (FSE_CTable* ct, BYTE symbolValue)
535a28cd43dSSascha Wildner {
536a28cd43dSSascha Wildner     void* ptr = ct;
537a28cd43dSSascha Wildner     U16* tableU16 = ( (U16*) ptr) + 2;
538a28cd43dSSascha Wildner     void* FSCTptr = (U32*)ptr + 2;
539a28cd43dSSascha Wildner     FSE_symbolCompressionTransform* symbolTT = (FSE_symbolCompressionTransform*) FSCTptr;
540a28cd43dSSascha Wildner 
541a28cd43dSSascha Wildner     /* header */
542a28cd43dSSascha Wildner     tableU16[-2] = (U16) 0;
543a28cd43dSSascha Wildner     tableU16[-1] = (U16) symbolValue;
544a28cd43dSSascha Wildner 
545a28cd43dSSascha Wildner     /* Build table */
546a28cd43dSSascha Wildner     tableU16[0] = 0;
547a28cd43dSSascha Wildner     tableU16[1] = 0;   /* just in case */
548a28cd43dSSascha Wildner 
549a28cd43dSSascha Wildner     /* Build Symbol Transformation Table */
550a28cd43dSSascha Wildner     symbolTT[symbolValue].deltaNbBits = 0;
551a28cd43dSSascha Wildner     symbolTT[symbolValue].deltaFindState = 0;
552a28cd43dSSascha Wildner 
553a28cd43dSSascha Wildner     return 0;
554a28cd43dSSascha Wildner }
555a28cd43dSSascha Wildner 
556a28cd43dSSascha Wildner 
FSE_compress_usingCTable_generic(void * dst,size_t dstSize,const void * src,size_t srcSize,const FSE_CTable * ct,const unsigned fast)557a28cd43dSSascha Wildner static size_t FSE_compress_usingCTable_generic (void* dst, size_t dstSize,
558a28cd43dSSascha Wildner                            const void* src, size_t srcSize,
559a28cd43dSSascha Wildner                            const FSE_CTable* ct, const unsigned fast)
560a28cd43dSSascha Wildner {
561a28cd43dSSascha Wildner     const BYTE* const istart = (const BYTE*) src;
562a28cd43dSSascha Wildner     const BYTE* const iend = istart + srcSize;
563a28cd43dSSascha Wildner     const BYTE* ip=iend;
564a28cd43dSSascha Wildner 
565a28cd43dSSascha Wildner     BIT_CStream_t bitC;
566a28cd43dSSascha Wildner     FSE_CState_t CState1, CState2;
567a28cd43dSSascha Wildner 
568a28cd43dSSascha Wildner     /* init */
569a28cd43dSSascha Wildner     if (srcSize <= 2) return 0;
570a28cd43dSSascha Wildner     { size_t const initError = BIT_initCStream(&bitC, dst, dstSize);
571a28cd43dSSascha Wildner       if (FSE_isError(initError)) return 0; /* not enough space available to write a bitstream */ }
572a28cd43dSSascha Wildner 
573a28cd43dSSascha Wildner #define FSE_FLUSHBITS(s)  (fast ? BIT_flushBitsFast(s) : BIT_flushBits(s))
574a28cd43dSSascha Wildner 
575a28cd43dSSascha Wildner     if (srcSize & 1) {
576a28cd43dSSascha Wildner         FSE_initCState2(&CState1, ct, *--ip);
577a28cd43dSSascha Wildner         FSE_initCState2(&CState2, ct, *--ip);
578a28cd43dSSascha Wildner         FSE_encodeSymbol(&bitC, &CState1, *--ip);
579a28cd43dSSascha Wildner         FSE_FLUSHBITS(&bitC);
580a28cd43dSSascha Wildner     } else {
581a28cd43dSSascha Wildner         FSE_initCState2(&CState2, ct, *--ip);
582a28cd43dSSascha Wildner         FSE_initCState2(&CState1, ct, *--ip);
583a28cd43dSSascha Wildner     }
584a28cd43dSSascha Wildner 
585a28cd43dSSascha Wildner     /* join to mod 4 */
586a28cd43dSSascha Wildner     srcSize -= 2;
587a28cd43dSSascha Wildner     if ((sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) && (srcSize & 2)) {  /* test bit 2 */
588a28cd43dSSascha Wildner         FSE_encodeSymbol(&bitC, &CState2, *--ip);
589a28cd43dSSascha Wildner         FSE_encodeSymbol(&bitC, &CState1, *--ip);
590a28cd43dSSascha Wildner         FSE_FLUSHBITS(&bitC);
591a28cd43dSSascha Wildner     }
592a28cd43dSSascha Wildner 
593a28cd43dSSascha Wildner     /* 2 or 4 encoding per loop */
594a28cd43dSSascha Wildner     while ( ip>istart ) {
595a28cd43dSSascha Wildner 
596a28cd43dSSascha Wildner         FSE_encodeSymbol(&bitC, &CState2, *--ip);
597a28cd43dSSascha Wildner 
598a28cd43dSSascha Wildner         if (sizeof(bitC.bitContainer)*8 < FSE_MAX_TABLELOG*2+7 )   /* this test must be static */
599a28cd43dSSascha Wildner             FSE_FLUSHBITS(&bitC);
600a28cd43dSSascha Wildner 
601a28cd43dSSascha Wildner         FSE_encodeSymbol(&bitC, &CState1, *--ip);
602a28cd43dSSascha Wildner 
603a28cd43dSSascha Wildner         if (sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) {  /* this test must be static */
604a28cd43dSSascha Wildner             FSE_encodeSymbol(&bitC, &CState2, *--ip);
605a28cd43dSSascha Wildner             FSE_encodeSymbol(&bitC, &CState1, *--ip);
606a28cd43dSSascha Wildner         }
607a28cd43dSSascha Wildner 
608a28cd43dSSascha Wildner         FSE_FLUSHBITS(&bitC);
609a28cd43dSSascha Wildner     }
610a28cd43dSSascha Wildner 
611a28cd43dSSascha Wildner     FSE_flushCState(&bitC, &CState2);
612a28cd43dSSascha Wildner     FSE_flushCState(&bitC, &CState1);
613a28cd43dSSascha Wildner     return BIT_closeCStream(&bitC);
614a28cd43dSSascha Wildner }
615a28cd43dSSascha Wildner 
FSE_compress_usingCTable(void * dst,size_t dstSize,const void * src,size_t srcSize,const FSE_CTable * ct)616a28cd43dSSascha Wildner size_t FSE_compress_usingCTable (void* dst, size_t dstSize,
617a28cd43dSSascha Wildner                            const void* src, size_t srcSize,
618a28cd43dSSascha Wildner                            const FSE_CTable* ct)
619a28cd43dSSascha Wildner {
620a28cd43dSSascha Wildner     unsigned const fast = (dstSize >= FSE_BLOCKBOUND(srcSize));
621a28cd43dSSascha Wildner 
622a28cd43dSSascha Wildner     if (fast)
623a28cd43dSSascha Wildner         return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 1);
624a28cd43dSSascha Wildner     else
625a28cd43dSSascha Wildner         return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 0);
626a28cd43dSSascha Wildner }
627a28cd43dSSascha Wildner 
628a28cd43dSSascha Wildner 
FSE_compressBound(size_t size)629a28cd43dSSascha Wildner size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); }
630a28cd43dSSascha Wildner 
631a28cd43dSSascha Wildner #ifndef ZSTD_NO_UNUSED_FUNCTIONS
632a28cd43dSSascha Wildner /* FSE_compress_wksp() :
633a28cd43dSSascha Wildner  * Same as FSE_compress2(), but using an externally allocated scratch buffer (`workSpace`).
634a28cd43dSSascha Wildner  * `wkspSize` size must be `(1<<tableLog)`.
635a28cd43dSSascha Wildner  */
FSE_compress_wksp(void * dst,size_t dstSize,const void * src,size_t srcSize,unsigned maxSymbolValue,unsigned tableLog,void * workSpace,size_t wkspSize)636a28cd43dSSascha Wildner 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)
637a28cd43dSSascha Wildner {
638a28cd43dSSascha Wildner     BYTE* const ostart = (BYTE*) dst;
639a28cd43dSSascha Wildner     BYTE* op = ostart;
640a28cd43dSSascha Wildner     BYTE* const oend = ostart + dstSize;
641a28cd43dSSascha Wildner 
642a28cd43dSSascha Wildner     unsigned count[FSE_MAX_SYMBOL_VALUE+1];
643a28cd43dSSascha Wildner     S16   norm[FSE_MAX_SYMBOL_VALUE+1];
644a28cd43dSSascha Wildner     FSE_CTable* CTable = (FSE_CTable*)workSpace;
645a28cd43dSSascha Wildner     size_t const CTableSize = FSE_CTABLE_SIZE_U32(tableLog, maxSymbolValue);
646a28cd43dSSascha Wildner     void* scratchBuffer = (void*)(CTable + CTableSize);
647a28cd43dSSascha Wildner     size_t const scratchBufferSize = wkspSize - (CTableSize * sizeof(FSE_CTable));
648a28cd43dSSascha Wildner 
649a28cd43dSSascha Wildner     /* init conditions */
650a28cd43dSSascha Wildner     if (wkspSize < FSE_COMPRESS_WKSP_SIZE_U32(tableLog, maxSymbolValue)) return ERROR(tableLog_tooLarge);
651a28cd43dSSascha Wildner     if (srcSize <= 1) return 0;  /* Not compressible */
652a28cd43dSSascha Wildner     if (!maxSymbolValue) maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
653a28cd43dSSascha Wildner     if (!tableLog) tableLog = FSE_DEFAULT_TABLELOG;
654a28cd43dSSascha Wildner 
655a28cd43dSSascha Wildner     /* Scan input and build symbol stats */
656a28cd43dSSascha Wildner     {   CHECK_V_F(maxCount, HIST_count_wksp(count, &maxSymbolValue, src, srcSize, scratchBuffer, scratchBufferSize) );
657a28cd43dSSascha Wildner         if (maxCount == srcSize) return 1;   /* only a single symbol in src : rle */
658a28cd43dSSascha Wildner         if (maxCount == 1) return 0;         /* each symbol present maximum once => not compressible */
659a28cd43dSSascha Wildner         if (maxCount < (srcSize >> 7)) return 0;   /* Heuristic : not compressible enough */
660a28cd43dSSascha Wildner     }
661a28cd43dSSascha Wildner 
662a28cd43dSSascha Wildner     tableLog = FSE_optimalTableLog(tableLog, srcSize, maxSymbolValue);
663a28cd43dSSascha Wildner     CHECK_F( FSE_normalizeCount(norm, tableLog, count, srcSize, maxSymbolValue, /* useLowProbCount */ srcSize >= 2048) );
664a28cd43dSSascha Wildner 
665a28cd43dSSascha Wildner     /* Write table description header */
666a28cd43dSSascha Wildner     {   CHECK_V_F(nc_err, FSE_writeNCount(op, oend-op, norm, maxSymbolValue, tableLog) );
667a28cd43dSSascha Wildner         op += nc_err;
668a28cd43dSSascha Wildner     }
669a28cd43dSSascha Wildner 
670a28cd43dSSascha Wildner     /* Compress */
671a28cd43dSSascha Wildner     CHECK_F( FSE_buildCTable_wksp(CTable, norm, maxSymbolValue, tableLog, scratchBuffer, scratchBufferSize) );
672a28cd43dSSascha Wildner     {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, oend - op, src, srcSize, CTable) );
673a28cd43dSSascha Wildner         if (cSize == 0) return 0;   /* not enough space for compressed data */
674a28cd43dSSascha Wildner         op += cSize;
675a28cd43dSSascha Wildner     }
676a28cd43dSSascha Wildner 
677a28cd43dSSascha Wildner     /* check compressibility */
678a28cd43dSSascha Wildner     if ( (size_t)(op-ostart) >= srcSize-1 ) return 0;
679a28cd43dSSascha Wildner 
680a28cd43dSSascha Wildner     return op-ostart;
681a28cd43dSSascha Wildner }
682a28cd43dSSascha Wildner 
683a28cd43dSSascha Wildner typedef struct {
684a28cd43dSSascha Wildner     FSE_CTable CTable_max[FSE_CTABLE_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE)];
685a28cd43dSSascha Wildner     union {
686a28cd43dSSascha Wildner       U32 hist_wksp[HIST_WKSP_SIZE_U32];
687a28cd43dSSascha Wildner       BYTE scratchBuffer[1 << FSE_MAX_TABLELOG];
688a28cd43dSSascha Wildner     } workspace;
689a28cd43dSSascha Wildner } fseWkspMax_t;
690a28cd43dSSascha Wildner 
FSE_compress2(void * dst,size_t dstCapacity,const void * src,size_t srcSize,unsigned maxSymbolValue,unsigned tableLog)691a28cd43dSSascha Wildner size_t FSE_compress2 (void* dst, size_t dstCapacity, const void* src, size_t srcSize, unsigned maxSymbolValue, unsigned tableLog)
692a28cd43dSSascha Wildner {
693a28cd43dSSascha Wildner     fseWkspMax_t scratchBuffer;
694a28cd43dSSascha Wildner     DEBUG_STATIC_ASSERT(sizeof(scratchBuffer) >= FSE_COMPRESS_WKSP_SIZE_U32(FSE_MAX_TABLELOG, FSE_MAX_SYMBOL_VALUE));   /* compilation failures here means scratchBuffer is not large enough */
695a28cd43dSSascha Wildner     if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);
696a28cd43dSSascha Wildner     return FSE_compress_wksp(dst, dstCapacity, src, srcSize, maxSymbolValue, tableLog, &scratchBuffer, sizeof(scratchBuffer));
697a28cd43dSSascha Wildner }
698a28cd43dSSascha Wildner 
FSE_compress(void * dst,size_t dstCapacity,const void * src,size_t srcSize)699a28cd43dSSascha Wildner size_t FSE_compress (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
700a28cd43dSSascha Wildner {
701a28cd43dSSascha Wildner     return FSE_compress2(dst, dstCapacity, src, srcSize, FSE_MAX_SYMBOL_VALUE, FSE_DEFAULT_TABLELOG);
702a28cd43dSSascha Wildner }
703a28cd43dSSascha Wildner #endif
704a28cd43dSSascha Wildner 
705a28cd43dSSascha Wildner #endif   /* FSE_COMMONDEFS_ONLY */
706