xref: /netbsd-src/external/bsd/zstd/dist/lib/compress/fse_compress.c (revision 3117ece4fc4a4ca4489ba793710b60b0d26bab6c)
1*3117ece4Schristos /* ******************************************************************
2*3117ece4Schristos  * FSE : Finite State Entropy encoder
3*3117ece4Schristos  * Copyright (c) Meta Platforms, Inc. and affiliates.
4*3117ece4Schristos  *
5*3117ece4Schristos  *  You can contact the author at :
6*3117ece4Schristos  *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
7*3117ece4Schristos  *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
8*3117ece4Schristos  *
9*3117ece4Schristos  * This source code is licensed under both the BSD-style license (found in the
10*3117ece4Schristos  * LICENSE file in the root directory of this source tree) and the GPLv2 (found
11*3117ece4Schristos  * in the COPYING file in the root directory of this source tree).
12*3117ece4Schristos  * You may select, at your option, one of the above-listed licenses.
13*3117ece4Schristos ****************************************************************** */
14*3117ece4Schristos 
15*3117ece4Schristos /* **************************************************************
16*3117ece4Schristos *  Includes
17*3117ece4Schristos ****************************************************************/
18*3117ece4Schristos #include "../common/compiler.h"
19*3117ece4Schristos #include "../common/mem.h"        /* U32, U16, etc. */
20*3117ece4Schristos #include "../common/debug.h"      /* assert, DEBUGLOG */
21*3117ece4Schristos #include "hist.h"       /* HIST_count_wksp */
22*3117ece4Schristos #include "../common/bitstream.h"
23*3117ece4Schristos #define FSE_STATIC_LINKING_ONLY
24*3117ece4Schristos #include "../common/fse.h"
25*3117ece4Schristos #include "../common/error_private.h"
26*3117ece4Schristos #define ZSTD_DEPS_NEED_MALLOC
27*3117ece4Schristos #define ZSTD_DEPS_NEED_MATH64
28*3117ece4Schristos #include "../common/zstd_deps.h"  /* ZSTD_memset */
29*3117ece4Schristos #include "../common/bits.h" /* ZSTD_highbit32 */
30*3117ece4Schristos 
31*3117ece4Schristos 
32*3117ece4Schristos /* **************************************************************
33*3117ece4Schristos *  Error Management
34*3117ece4Schristos ****************************************************************/
35*3117ece4Schristos #define FSE_isError ERR_isError
36*3117ece4Schristos 
37*3117ece4Schristos 
38*3117ece4Schristos /* **************************************************************
39*3117ece4Schristos *  Templates
40*3117ece4Schristos ****************************************************************/
41*3117ece4Schristos /*
42*3117ece4Schristos   designed to be included
43*3117ece4Schristos   for type-specific functions (template emulation in C)
44*3117ece4Schristos   Objective is to write these functions only once, for improved maintenance
45*3117ece4Schristos */
46*3117ece4Schristos 
47*3117ece4Schristos /* safety checks */
48*3117ece4Schristos #ifndef FSE_FUNCTION_EXTENSION
49*3117ece4Schristos #  error "FSE_FUNCTION_EXTENSION must be defined"
50*3117ece4Schristos #endif
51*3117ece4Schristos #ifndef FSE_FUNCTION_TYPE
52*3117ece4Schristos #  error "FSE_FUNCTION_TYPE must be defined"
53*3117ece4Schristos #endif
54*3117ece4Schristos 
55*3117ece4Schristos /* Function names */
56*3117ece4Schristos #define FSE_CAT(X,Y) X##Y
57*3117ece4Schristos #define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
58*3117ece4Schristos #define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)
59*3117ece4Schristos 
60*3117ece4Schristos 
61*3117ece4Schristos /* Function templates */
62*3117ece4Schristos 
63*3117ece4Schristos /* FSE_buildCTable_wksp() :
64*3117ece4Schristos  * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
65*3117ece4Schristos  * wkspSize should be sized to handle worst case situation, which is `1<<max_tableLog * sizeof(FSE_FUNCTION_TYPE)`
66*3117ece4Schristos  * workSpace must also be properly aligned with FSE_FUNCTION_TYPE requirements
67*3117ece4Schristos  */
68*3117ece4Schristos size_t FSE_buildCTable_wksp(FSE_CTable* ct,
69*3117ece4Schristos                       const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog,
70*3117ece4Schristos                             void* workSpace, size_t wkspSize)
71*3117ece4Schristos {
72*3117ece4Schristos     U32 const tableSize = 1 << tableLog;
73*3117ece4Schristos     U32 const tableMask = tableSize - 1;
74*3117ece4Schristos     void* const ptr = ct;
75*3117ece4Schristos     U16* const tableU16 = ( (U16*) ptr) + 2;
76*3117ece4Schristos     void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableLog ? tableSize>>1 : 1) ;
77*3117ece4Schristos     FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
78*3117ece4Schristos     U32 const step = FSE_TABLESTEP(tableSize);
79*3117ece4Schristos     U32 const maxSV1 = maxSymbolValue+1;
80*3117ece4Schristos 
81*3117ece4Schristos     U16* cumul = (U16*)workSpace;   /* size = maxSV1 */
82*3117ece4Schristos     FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSV1+1));  /* size = tableSize */
83*3117ece4Schristos 
84*3117ece4Schristos     U32 highThreshold = tableSize-1;
85*3117ece4Schristos 
86*3117ece4Schristos     assert(((size_t)workSpace & 1) == 0);  /* Must be 2 bytes-aligned */
87*3117ece4Schristos     if (FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) > wkspSize) return ERROR(tableLog_tooLarge);
88*3117ece4Schristos     /* CTable header */
89*3117ece4Schristos     tableU16[-2] = (U16) tableLog;
90*3117ece4Schristos     tableU16[-1] = (U16) maxSymbolValue;
91*3117ece4Schristos     assert(tableLog < 16);   /* required for threshold strategy to work */
92*3117ece4Schristos 
93*3117ece4Schristos     /* For explanations on how to distribute symbol values over the table :
94*3117ece4Schristos      * https://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */
95*3117ece4Schristos 
96*3117ece4Schristos      #ifdef __clang_analyzer__
97*3117ece4Schristos      ZSTD_memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize);   /* useless initialization, just to keep scan-build happy */
98*3117ece4Schristos      #endif
99*3117ece4Schristos 
100*3117ece4Schristos     /* symbol start positions */
101*3117ece4Schristos     {   U32 u;
102*3117ece4Schristos         cumul[0] = 0;
103*3117ece4Schristos         for (u=1; u <= maxSV1; u++) {
104*3117ece4Schristos             if (normalizedCounter[u-1]==-1) {  /* Low proba symbol */
105*3117ece4Schristos                 cumul[u] = cumul[u-1] + 1;
106*3117ece4Schristos                 tableSymbol[highThreshold--] = (FSE_FUNCTION_TYPE)(u-1);
107*3117ece4Schristos             } else {
108*3117ece4Schristos                 assert(normalizedCounter[u-1] >= 0);
109*3117ece4Schristos                 cumul[u] = cumul[u-1] + (U16)normalizedCounter[u-1];
110*3117ece4Schristos                 assert(cumul[u] >= cumul[u-1]);  /* no overflow */
111*3117ece4Schristos         }   }
112*3117ece4Schristos         cumul[maxSV1] = (U16)(tableSize+1);
113*3117ece4Schristos     }
114*3117ece4Schristos 
115*3117ece4Schristos     /* Spread symbols */
116*3117ece4Schristos     if (highThreshold == tableSize - 1) {
117*3117ece4Schristos         /* Case for no low prob count symbols. Lay down 8 bytes at a time
118*3117ece4Schristos          * to reduce branch misses since we are operating on a small block
119*3117ece4Schristos          */
120*3117ece4Schristos         BYTE* const spread = tableSymbol + tableSize; /* size = tableSize + 8 (may write beyond tableSize) */
121*3117ece4Schristos         {   U64 const add = 0x0101010101010101ull;
122*3117ece4Schristos             size_t pos = 0;
123*3117ece4Schristos             U64 sv = 0;
124*3117ece4Schristos             U32 s;
125*3117ece4Schristos             for (s=0; s<maxSV1; ++s, sv += add) {
126*3117ece4Schristos                 int i;
127*3117ece4Schristos                 int const n = normalizedCounter[s];
128*3117ece4Schristos                 MEM_write64(spread + pos, sv);
129*3117ece4Schristos                 for (i = 8; i < n; i += 8) {
130*3117ece4Schristos                     MEM_write64(spread + pos + i, sv);
131*3117ece4Schristos                 }
132*3117ece4Schristos                 assert(n>=0);
133*3117ece4Schristos                 pos += (size_t)n;
134*3117ece4Schristos             }
135*3117ece4Schristos         }
136*3117ece4Schristos         /* Spread symbols across the table. Lack of lowprob symbols means that
137*3117ece4Schristos          * we don't need variable sized inner loop, so we can unroll the loop and
138*3117ece4Schristos          * reduce branch misses.
139*3117ece4Schristos          */
140*3117ece4Schristos         {   size_t position = 0;
141*3117ece4Schristos             size_t s;
142*3117ece4Schristos             size_t const unroll = 2; /* Experimentally determined optimal unroll */
143*3117ece4Schristos             assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
144*3117ece4Schristos             for (s = 0; s < (size_t)tableSize; s += unroll) {
145*3117ece4Schristos                 size_t u;
146*3117ece4Schristos                 for (u = 0; u < unroll; ++u) {
147*3117ece4Schristos                     size_t const uPosition = (position + (u * step)) & tableMask;
148*3117ece4Schristos                     tableSymbol[uPosition] = spread[s + u];
149*3117ece4Schristos                 }
150*3117ece4Schristos                 position = (position + (unroll * step)) & tableMask;
151*3117ece4Schristos             }
152*3117ece4Schristos             assert(position == 0);   /* Must have initialized all positions */
153*3117ece4Schristos         }
154*3117ece4Schristos     } else {
155*3117ece4Schristos         U32 position = 0;
156*3117ece4Schristos         U32 symbol;
157*3117ece4Schristos         for (symbol=0; symbol<maxSV1; symbol++) {
158*3117ece4Schristos             int nbOccurrences;
159*3117ece4Schristos             int const freq = normalizedCounter[symbol];
160*3117ece4Schristos             for (nbOccurrences=0; nbOccurrences<freq; nbOccurrences++) {
161*3117ece4Schristos                 tableSymbol[position] = (FSE_FUNCTION_TYPE)symbol;
162*3117ece4Schristos                 position = (position + step) & tableMask;
163*3117ece4Schristos                 while (position > highThreshold)
164*3117ece4Schristos                     position = (position + step) & tableMask;   /* Low proba area */
165*3117ece4Schristos         }   }
166*3117ece4Schristos         assert(position==0);  /* Must have initialized all positions */
167*3117ece4Schristos     }
168*3117ece4Schristos 
169*3117ece4Schristos     /* Build table */
170*3117ece4Schristos     {   U32 u; for (u=0; u<tableSize; u++) {
171*3117ece4Schristos         FSE_FUNCTION_TYPE s = tableSymbol[u];   /* note : static analyzer may not understand tableSymbol is properly initialized */
172*3117ece4Schristos         tableU16[cumul[s]++] = (U16) (tableSize+u);   /* TableU16 : sorted by symbol order; gives next state value */
173*3117ece4Schristos     }   }
174*3117ece4Schristos 
175*3117ece4Schristos     /* Build Symbol Transformation Table */
176*3117ece4Schristos     {   unsigned total = 0;
177*3117ece4Schristos         unsigned s;
178*3117ece4Schristos         for (s=0; s<=maxSymbolValue; s++) {
179*3117ece4Schristos             switch (normalizedCounter[s])
180*3117ece4Schristos             {
181*3117ece4Schristos             case  0:
182*3117ece4Schristos                 /* filling nonetheless, for compatibility with FSE_getMaxNbBits() */
183*3117ece4Schristos                 symbolTT[s].deltaNbBits = ((tableLog+1) << 16) - (1<<tableLog);
184*3117ece4Schristos                 break;
185*3117ece4Schristos 
186*3117ece4Schristos             case -1:
187*3117ece4Schristos             case  1:
188*3117ece4Schristos                 symbolTT[s].deltaNbBits = (tableLog << 16) - (1<<tableLog);
189*3117ece4Schristos                 assert(total <= INT_MAX);
190*3117ece4Schristos                 symbolTT[s].deltaFindState = (int)(total - 1);
191*3117ece4Schristos                 total ++;
192*3117ece4Schristos                 break;
193*3117ece4Schristos             default :
194*3117ece4Schristos                 assert(normalizedCounter[s] > 1);
195*3117ece4Schristos                 {   U32 const maxBitsOut = tableLog - ZSTD_highbit32 ((U32)normalizedCounter[s]-1);
196*3117ece4Schristos                     U32 const minStatePlus = (U32)normalizedCounter[s] << maxBitsOut;
197*3117ece4Schristos                     symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus;
198*3117ece4Schristos                     symbolTT[s].deltaFindState = (int)(total - (unsigned)normalizedCounter[s]);
199*3117ece4Schristos                     total +=  (unsigned)normalizedCounter[s];
200*3117ece4Schristos     }   }   }   }
201*3117ece4Schristos 
202*3117ece4Schristos #if 0  /* debug : symbol costs */
203*3117ece4Schristos     DEBUGLOG(5, "\n --- table statistics : ");
204*3117ece4Schristos     {   U32 symbol;
205*3117ece4Schristos         for (symbol=0; symbol<=maxSymbolValue; symbol++) {
206*3117ece4Schristos             DEBUGLOG(5, "%3u: w=%3i,   maxBits=%u, fracBits=%.2f",
207*3117ece4Schristos                 symbol, normalizedCounter[symbol],
208*3117ece4Schristos                 FSE_getMaxNbBits(symbolTT, symbol),
209*3117ece4Schristos                 (double)FSE_bitCost(symbolTT, tableLog, symbol, 8) / 256);
210*3117ece4Schristos     }   }
211*3117ece4Schristos #endif
212*3117ece4Schristos 
213*3117ece4Schristos     return 0;
214*3117ece4Schristos }
215*3117ece4Schristos 
216*3117ece4Schristos 
217*3117ece4Schristos 
218*3117ece4Schristos #ifndef FSE_COMMONDEFS_ONLY
219*3117ece4Schristos 
220*3117ece4Schristos /*-**************************************************************
221*3117ece4Schristos *  FSE NCount encoding
222*3117ece4Schristos ****************************************************************/
223*3117ece4Schristos size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog)
224*3117ece4Schristos {
225*3117ece4Schristos     size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog
226*3117ece4Schristos                                    + 4 /* bitCount initialized at 4 */
227*3117ece4Schristos                                    + 2 /* first two symbols may use one additional bit each */) / 8)
228*3117ece4Schristos                                    + 1 /* round up to whole nb bytes */
229*3117ece4Schristos                                    + 2 /* additional two bytes for bitstream flush */;
230*3117ece4Schristos     return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND;  /* maxSymbolValue==0 ? use default */
231*3117ece4Schristos }
232*3117ece4Schristos 
233*3117ece4Schristos static size_t
234*3117ece4Schristos FSE_writeNCount_generic (void* header, size_t headerBufferSize,
235*3117ece4Schristos                    const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog,
236*3117ece4Schristos                          unsigned writeIsSafe)
237*3117ece4Schristos {
238*3117ece4Schristos     BYTE* const ostart = (BYTE*) header;
239*3117ece4Schristos     BYTE* out = ostart;
240*3117ece4Schristos     BYTE* const oend = ostart + headerBufferSize;
241*3117ece4Schristos     int nbBits;
242*3117ece4Schristos     const int tableSize = 1 << tableLog;
243*3117ece4Schristos     int remaining;
244*3117ece4Schristos     int threshold;
245*3117ece4Schristos     U32 bitStream = 0;
246*3117ece4Schristos     int bitCount = 0;
247*3117ece4Schristos     unsigned symbol = 0;
248*3117ece4Schristos     unsigned const alphabetSize = maxSymbolValue + 1;
249*3117ece4Schristos     int previousIs0 = 0;
250*3117ece4Schristos 
251*3117ece4Schristos     /* Table Size */
252*3117ece4Schristos     bitStream += (tableLog-FSE_MIN_TABLELOG) << bitCount;
253*3117ece4Schristos     bitCount  += 4;
254*3117ece4Schristos 
255*3117ece4Schristos     /* Init */
256*3117ece4Schristos     remaining = tableSize+1;   /* +1 for extra accuracy */
257*3117ece4Schristos     threshold = tableSize;
258*3117ece4Schristos     nbBits = (int)tableLog+1;
259*3117ece4Schristos 
260*3117ece4Schristos     while ((symbol < alphabetSize) && (remaining>1)) {  /* stops at 1 */
261*3117ece4Schristos         if (previousIs0) {
262*3117ece4Schristos             unsigned start = symbol;
263*3117ece4Schristos             while ((symbol < alphabetSize) && !normalizedCounter[symbol]) symbol++;
264*3117ece4Schristos             if (symbol == alphabetSize) break;   /* incorrect distribution */
265*3117ece4Schristos             while (symbol >= start+24) {
266*3117ece4Schristos                 start+=24;
267*3117ece4Schristos                 bitStream += 0xFFFFU << bitCount;
268*3117ece4Schristos                 if ((!writeIsSafe) && (out > oend-2))
269*3117ece4Schristos                     return ERROR(dstSize_tooSmall);   /* Buffer overflow */
270*3117ece4Schristos                 out[0] = (BYTE) bitStream;
271*3117ece4Schristos                 out[1] = (BYTE)(bitStream>>8);
272*3117ece4Schristos                 out+=2;
273*3117ece4Schristos                 bitStream>>=16;
274*3117ece4Schristos             }
275*3117ece4Schristos             while (symbol >= start+3) {
276*3117ece4Schristos                 start+=3;
277*3117ece4Schristos                 bitStream += 3U << bitCount;
278*3117ece4Schristos                 bitCount += 2;
279*3117ece4Schristos             }
280*3117ece4Schristos             bitStream += (symbol-start) << bitCount;
281*3117ece4Schristos             bitCount += 2;
282*3117ece4Schristos             if (bitCount>16) {
283*3117ece4Schristos                 if ((!writeIsSafe) && (out > oend - 2))
284*3117ece4Schristos                     return ERROR(dstSize_tooSmall);   /* Buffer overflow */
285*3117ece4Schristos                 out[0] = (BYTE)bitStream;
286*3117ece4Schristos                 out[1] = (BYTE)(bitStream>>8);
287*3117ece4Schristos                 out += 2;
288*3117ece4Schristos                 bitStream >>= 16;
289*3117ece4Schristos                 bitCount -= 16;
290*3117ece4Schristos         }   }
291*3117ece4Schristos         {   int count = normalizedCounter[symbol++];
292*3117ece4Schristos             int const max = (2*threshold-1) - remaining;
293*3117ece4Schristos             remaining -= count < 0 ? -count : count;
294*3117ece4Schristos             count++;   /* +1 for extra accuracy */
295*3117ece4Schristos             if (count>=threshold)
296*3117ece4Schristos                 count += max;   /* [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[ */
297*3117ece4Schristos             bitStream += (U32)count << bitCount;
298*3117ece4Schristos             bitCount  += nbBits;
299*3117ece4Schristos             bitCount  -= (count<max);
300*3117ece4Schristos             previousIs0  = (count==1);
301*3117ece4Schristos             if (remaining<1) return ERROR(GENERIC);
302*3117ece4Schristos             while (remaining<threshold) { nbBits--; threshold>>=1; }
303*3117ece4Schristos         }
304*3117ece4Schristos         if (bitCount>16) {
305*3117ece4Schristos             if ((!writeIsSafe) && (out > oend - 2))
306*3117ece4Schristos                 return ERROR(dstSize_tooSmall);   /* Buffer overflow */
307*3117ece4Schristos             out[0] = (BYTE)bitStream;
308*3117ece4Schristos             out[1] = (BYTE)(bitStream>>8);
309*3117ece4Schristos             out += 2;
310*3117ece4Schristos             bitStream >>= 16;
311*3117ece4Schristos             bitCount -= 16;
312*3117ece4Schristos     }   }
313*3117ece4Schristos 
314*3117ece4Schristos     if (remaining != 1)
315*3117ece4Schristos         return ERROR(GENERIC);  /* incorrect normalized distribution */
316*3117ece4Schristos     assert(symbol <= alphabetSize);
317*3117ece4Schristos 
318*3117ece4Schristos     /* flush remaining bitStream */
319*3117ece4Schristos     if ((!writeIsSafe) && (out > oend - 2))
320*3117ece4Schristos         return ERROR(dstSize_tooSmall);   /* Buffer overflow */
321*3117ece4Schristos     out[0] = (BYTE)bitStream;
322*3117ece4Schristos     out[1] = (BYTE)(bitStream>>8);
323*3117ece4Schristos     out+= (bitCount+7) /8;
324*3117ece4Schristos 
325*3117ece4Schristos     assert(out >= ostart);
326*3117ece4Schristos     return (size_t)(out-ostart);
327*3117ece4Schristos }
328*3117ece4Schristos 
329*3117ece4Schristos 
330*3117ece4Schristos size_t FSE_writeNCount (void* buffer, size_t bufferSize,
331*3117ece4Schristos                   const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
332*3117ece4Schristos {
333*3117ece4Schristos     if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);   /* Unsupported */
334*3117ece4Schristos     if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC);   /* Unsupported */
335*3117ece4Schristos 
336*3117ece4Schristos     if (bufferSize < FSE_NCountWriteBound(maxSymbolValue, tableLog))
337*3117ece4Schristos         return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 0);
338*3117ece4Schristos 
339*3117ece4Schristos     return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 1 /* write in buffer is safe */);
340*3117ece4Schristos }
341*3117ece4Schristos 
342*3117ece4Schristos 
343*3117ece4Schristos /*-**************************************************************
344*3117ece4Schristos *  FSE Compression Code
345*3117ece4Schristos ****************************************************************/
346*3117ece4Schristos 
347*3117ece4Schristos /* provides the minimum logSize to safely represent a distribution */
348*3117ece4Schristos static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue)
349*3117ece4Schristos {
350*3117ece4Schristos     U32 minBitsSrc = ZSTD_highbit32((U32)(srcSize)) + 1;
351*3117ece4Schristos     U32 minBitsSymbols = ZSTD_highbit32(maxSymbolValue) + 2;
352*3117ece4Schristos     U32 minBits = minBitsSrc < minBitsSymbols ? minBitsSrc : minBitsSymbols;
353*3117ece4Schristos     assert(srcSize > 1); /* Not supported, RLE should be used instead */
354*3117ece4Schristos     return minBits;
355*3117ece4Schristos }
356*3117ece4Schristos 
357*3117ece4Schristos unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus)
358*3117ece4Schristos {
359*3117ece4Schristos     U32 maxBitsSrc = ZSTD_highbit32((U32)(srcSize - 1)) - minus;
360*3117ece4Schristos     U32 tableLog = maxTableLog;
361*3117ece4Schristos     U32 minBits = FSE_minTableLog(srcSize, maxSymbolValue);
362*3117ece4Schristos     assert(srcSize > 1); /* Not supported, RLE should be used instead */
363*3117ece4Schristos     if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
364*3117ece4Schristos     if (maxBitsSrc < tableLog) tableLog = maxBitsSrc;   /* Accuracy can be reduced */
365*3117ece4Schristos     if (minBits > tableLog) tableLog = minBits;   /* Need a minimum to safely represent all symbol values */
366*3117ece4Schristos     if (tableLog < FSE_MIN_TABLELOG) tableLog = FSE_MIN_TABLELOG;
367*3117ece4Schristos     if (tableLog > FSE_MAX_TABLELOG) tableLog = FSE_MAX_TABLELOG;
368*3117ece4Schristos     return tableLog;
369*3117ece4Schristos }
370*3117ece4Schristos 
371*3117ece4Schristos unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
372*3117ece4Schristos {
373*3117ece4Schristos     return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 2);
374*3117ece4Schristos }
375*3117ece4Schristos 
376*3117ece4Schristos /* Secondary normalization method.
377*3117ece4Schristos    To be used when primary method fails. */
378*3117ece4Schristos 
379*3117ece4Schristos static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue, short lowProbCount)
380*3117ece4Schristos {
381*3117ece4Schristos     short const NOT_YET_ASSIGNED = -2;
382*3117ece4Schristos     U32 s;
383*3117ece4Schristos     U32 distributed = 0;
384*3117ece4Schristos     U32 ToDistribute;
385*3117ece4Schristos 
386*3117ece4Schristos     /* Init */
387*3117ece4Schristos     U32 const lowThreshold = (U32)(total >> tableLog);
388*3117ece4Schristos     U32 lowOne = (U32)((total * 3) >> (tableLog + 1));
389*3117ece4Schristos 
390*3117ece4Schristos     for (s=0; s<=maxSymbolValue; s++) {
391*3117ece4Schristos         if (count[s] == 0) {
392*3117ece4Schristos             norm[s]=0;
393*3117ece4Schristos             continue;
394*3117ece4Schristos         }
395*3117ece4Schristos         if (count[s] <= lowThreshold) {
396*3117ece4Schristos             norm[s] = lowProbCount;
397*3117ece4Schristos             distributed++;
398*3117ece4Schristos             total -= count[s];
399*3117ece4Schristos             continue;
400*3117ece4Schristos         }
401*3117ece4Schristos         if (count[s] <= lowOne) {
402*3117ece4Schristos             norm[s] = 1;
403*3117ece4Schristos             distributed++;
404*3117ece4Schristos             total -= count[s];
405*3117ece4Schristos             continue;
406*3117ece4Schristos         }
407*3117ece4Schristos 
408*3117ece4Schristos         norm[s]=NOT_YET_ASSIGNED;
409*3117ece4Schristos     }
410*3117ece4Schristos     ToDistribute = (1 << tableLog) - distributed;
411*3117ece4Schristos 
412*3117ece4Schristos     if (ToDistribute == 0)
413*3117ece4Schristos         return 0;
414*3117ece4Schristos 
415*3117ece4Schristos     if ((total / ToDistribute) > lowOne) {
416*3117ece4Schristos         /* risk of rounding to zero */
417*3117ece4Schristos         lowOne = (U32)((total * 3) / (ToDistribute * 2));
418*3117ece4Schristos         for (s=0; s<=maxSymbolValue; s++) {
419*3117ece4Schristos             if ((norm[s] == NOT_YET_ASSIGNED) && (count[s] <= lowOne)) {
420*3117ece4Schristos                 norm[s] = 1;
421*3117ece4Schristos                 distributed++;
422*3117ece4Schristos                 total -= count[s];
423*3117ece4Schristos                 continue;
424*3117ece4Schristos         }   }
425*3117ece4Schristos         ToDistribute = (1 << tableLog) - distributed;
426*3117ece4Schristos     }
427*3117ece4Schristos 
428*3117ece4Schristos     if (distributed == maxSymbolValue+1) {
429*3117ece4Schristos         /* all values are pretty poor;
430*3117ece4Schristos            probably incompressible data (should have already been detected);
431*3117ece4Schristos            find max, then give all remaining points to max */
432*3117ece4Schristos         U32 maxV = 0, maxC = 0;
433*3117ece4Schristos         for (s=0; s<=maxSymbolValue; s++)
434*3117ece4Schristos             if (count[s] > maxC) { maxV=s; maxC=count[s]; }
435*3117ece4Schristos         norm[maxV] += (short)ToDistribute;
436*3117ece4Schristos         return 0;
437*3117ece4Schristos     }
438*3117ece4Schristos 
439*3117ece4Schristos     if (total == 0) {
440*3117ece4Schristos         /* all of the symbols were low enough for the lowOne or lowThreshold */
441*3117ece4Schristos         for (s=0; ToDistribute > 0; s = (s+1)%(maxSymbolValue+1))
442*3117ece4Schristos             if (norm[s] > 0) { ToDistribute--; norm[s]++; }
443*3117ece4Schristos         return 0;
444*3117ece4Schristos     }
445*3117ece4Schristos 
446*3117ece4Schristos     {   U64 const vStepLog = 62 - tableLog;
447*3117ece4Schristos         U64 const mid = (1ULL << (vStepLog-1)) - 1;
448*3117ece4Schristos         U64 const rStep = ZSTD_div64((((U64)1<<vStepLog) * ToDistribute) + mid, (U32)total);   /* scale on remaining */
449*3117ece4Schristos         U64 tmpTotal = mid;
450*3117ece4Schristos         for (s=0; s<=maxSymbolValue; s++) {
451*3117ece4Schristos             if (norm[s]==NOT_YET_ASSIGNED) {
452*3117ece4Schristos                 U64 const end = tmpTotal + (count[s] * rStep);
453*3117ece4Schristos                 U32 const sStart = (U32)(tmpTotal >> vStepLog);
454*3117ece4Schristos                 U32 const sEnd = (U32)(end >> vStepLog);
455*3117ece4Schristos                 U32 const weight = sEnd - sStart;
456*3117ece4Schristos                 if (weight < 1)
457*3117ece4Schristos                     return ERROR(GENERIC);
458*3117ece4Schristos                 norm[s] = (short)weight;
459*3117ece4Schristos                 tmpTotal = end;
460*3117ece4Schristos     }   }   }
461*3117ece4Schristos 
462*3117ece4Schristos     return 0;
463*3117ece4Schristos }
464*3117ece4Schristos 
465*3117ece4Schristos size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
466*3117ece4Schristos                            const unsigned* count, size_t total,
467*3117ece4Schristos                            unsigned maxSymbolValue, unsigned useLowProbCount)
468*3117ece4Schristos {
469*3117ece4Schristos     /* Sanity checks */
470*3117ece4Schristos     if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
471*3117ece4Schristos     if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC);   /* Unsupported size */
472*3117ece4Schristos     if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);   /* Unsupported size */
473*3117ece4Schristos     if (tableLog < FSE_minTableLog(total, maxSymbolValue)) return ERROR(GENERIC);   /* Too small tableLog, compression potentially impossible */
474*3117ece4Schristos 
475*3117ece4Schristos     {   static U32 const rtbTable[] = {     0, 473195, 504333, 520860, 550000, 700000, 750000, 830000 };
476*3117ece4Schristos         short const lowProbCount = useLowProbCount ? -1 : 1;
477*3117ece4Schristos         U64 const scale = 62 - tableLog;
478*3117ece4Schristos         U64 const step = ZSTD_div64((U64)1<<62, (U32)total);   /* <== here, one division ! */
479*3117ece4Schristos         U64 const vStep = 1ULL<<(scale-20);
480*3117ece4Schristos         int stillToDistribute = 1<<tableLog;
481*3117ece4Schristos         unsigned s;
482*3117ece4Schristos         unsigned largest=0;
483*3117ece4Schristos         short largestP=0;
484*3117ece4Schristos         U32 lowThreshold = (U32)(total >> tableLog);
485*3117ece4Schristos 
486*3117ece4Schristos         for (s=0; s<=maxSymbolValue; s++) {
487*3117ece4Schristos             if (count[s] == total) return 0;   /* rle special case */
488*3117ece4Schristos             if (count[s] == 0) { normalizedCounter[s]=0; continue; }
489*3117ece4Schristos             if (count[s] <= lowThreshold) {
490*3117ece4Schristos                 normalizedCounter[s] = lowProbCount;
491*3117ece4Schristos                 stillToDistribute--;
492*3117ece4Schristos             } else {
493*3117ece4Schristos                 short proba = (short)((count[s]*step) >> scale);
494*3117ece4Schristos                 if (proba<8) {
495*3117ece4Schristos                     U64 restToBeat = vStep * rtbTable[proba];
496*3117ece4Schristos                     proba += (count[s]*step) - ((U64)proba<<scale) > restToBeat;
497*3117ece4Schristos                 }
498*3117ece4Schristos                 if (proba > largestP) { largestP=proba; largest=s; }
499*3117ece4Schristos                 normalizedCounter[s] = proba;
500*3117ece4Schristos                 stillToDistribute -= proba;
501*3117ece4Schristos         }   }
502*3117ece4Schristos         if (-stillToDistribute >= (normalizedCounter[largest] >> 1)) {
503*3117ece4Schristos             /* corner case, need another normalization method */
504*3117ece4Schristos             size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue, lowProbCount);
505*3117ece4Schristos             if (FSE_isError(errorCode)) return errorCode;
506*3117ece4Schristos         }
507*3117ece4Schristos         else normalizedCounter[largest] += (short)stillToDistribute;
508*3117ece4Schristos     }
509*3117ece4Schristos 
510*3117ece4Schristos #if 0
511*3117ece4Schristos     {   /* Print Table (debug) */
512*3117ece4Schristos         U32 s;
513*3117ece4Schristos         U32 nTotal = 0;
514*3117ece4Schristos         for (s=0; s<=maxSymbolValue; s++)
515*3117ece4Schristos             RAWLOG(2, "%3i: %4i \n", s, normalizedCounter[s]);
516*3117ece4Schristos         for (s=0; s<=maxSymbolValue; s++)
517*3117ece4Schristos             nTotal += abs(normalizedCounter[s]);
518*3117ece4Schristos         if (nTotal != (1U<<tableLog))
519*3117ece4Schristos             RAWLOG(2, "Warning !!! Total == %u != %u !!!", nTotal, 1U<<tableLog);
520*3117ece4Schristos         getchar();
521*3117ece4Schristos     }
522*3117ece4Schristos #endif
523*3117ece4Schristos 
524*3117ece4Schristos     return tableLog;
525*3117ece4Schristos }
526*3117ece4Schristos 
527*3117ece4Schristos /* fake FSE_CTable, for rle input (always same symbol) */
528*3117ece4Schristos size_t FSE_buildCTable_rle (FSE_CTable* ct, BYTE symbolValue)
529*3117ece4Schristos {
530*3117ece4Schristos     void* ptr = ct;
531*3117ece4Schristos     U16* tableU16 = ( (U16*) ptr) + 2;
532*3117ece4Schristos     void* FSCTptr = (U32*)ptr + 2;
533*3117ece4Schristos     FSE_symbolCompressionTransform* symbolTT = (FSE_symbolCompressionTransform*) FSCTptr;
534*3117ece4Schristos 
535*3117ece4Schristos     /* header */
536*3117ece4Schristos     tableU16[-2] = (U16) 0;
537*3117ece4Schristos     tableU16[-1] = (U16) symbolValue;
538*3117ece4Schristos 
539*3117ece4Schristos     /* Build table */
540*3117ece4Schristos     tableU16[0] = 0;
541*3117ece4Schristos     tableU16[1] = 0;   /* just in case */
542*3117ece4Schristos 
543*3117ece4Schristos     /* Build Symbol Transformation Table */
544*3117ece4Schristos     symbolTT[symbolValue].deltaNbBits = 0;
545*3117ece4Schristos     symbolTT[symbolValue].deltaFindState = 0;
546*3117ece4Schristos 
547*3117ece4Schristos     return 0;
548*3117ece4Schristos }
549*3117ece4Schristos 
550*3117ece4Schristos 
551*3117ece4Schristos static size_t FSE_compress_usingCTable_generic (void* dst, size_t dstSize,
552*3117ece4Schristos                            const void* src, size_t srcSize,
553*3117ece4Schristos                            const FSE_CTable* ct, const unsigned fast)
554*3117ece4Schristos {
555*3117ece4Schristos     const BYTE* const istart = (const BYTE*) src;
556*3117ece4Schristos     const BYTE* const iend = istart + srcSize;
557*3117ece4Schristos     const BYTE* ip=iend;
558*3117ece4Schristos 
559*3117ece4Schristos     BIT_CStream_t bitC;
560*3117ece4Schristos     FSE_CState_t CState1, CState2;
561*3117ece4Schristos 
562*3117ece4Schristos     /* init */
563*3117ece4Schristos     if (srcSize <= 2) return 0;
564*3117ece4Schristos     { size_t const initError = BIT_initCStream(&bitC, dst, dstSize);
565*3117ece4Schristos       if (FSE_isError(initError)) return 0; /* not enough space available to write a bitstream */ }
566*3117ece4Schristos 
567*3117ece4Schristos #define FSE_FLUSHBITS(s)  (fast ? BIT_flushBitsFast(s) : BIT_flushBits(s))
568*3117ece4Schristos 
569*3117ece4Schristos     if (srcSize & 1) {
570*3117ece4Schristos         FSE_initCState2(&CState1, ct, *--ip);
571*3117ece4Schristos         FSE_initCState2(&CState2, ct, *--ip);
572*3117ece4Schristos         FSE_encodeSymbol(&bitC, &CState1, *--ip);
573*3117ece4Schristos         FSE_FLUSHBITS(&bitC);
574*3117ece4Schristos     } else {
575*3117ece4Schristos         FSE_initCState2(&CState2, ct, *--ip);
576*3117ece4Schristos         FSE_initCState2(&CState1, ct, *--ip);
577*3117ece4Schristos     }
578*3117ece4Schristos 
579*3117ece4Schristos     /* join to mod 4 */
580*3117ece4Schristos     srcSize -= 2;
581*3117ece4Schristos     if ((sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) && (srcSize & 2)) {  /* test bit 2 */
582*3117ece4Schristos         FSE_encodeSymbol(&bitC, &CState2, *--ip);
583*3117ece4Schristos         FSE_encodeSymbol(&bitC, &CState1, *--ip);
584*3117ece4Schristos         FSE_FLUSHBITS(&bitC);
585*3117ece4Schristos     }
586*3117ece4Schristos 
587*3117ece4Schristos     /* 2 or 4 encoding per loop */
588*3117ece4Schristos     while ( ip>istart ) {
589*3117ece4Schristos 
590*3117ece4Schristos         FSE_encodeSymbol(&bitC, &CState2, *--ip);
591*3117ece4Schristos 
592*3117ece4Schristos         if (sizeof(bitC.bitContainer)*8 < FSE_MAX_TABLELOG*2+7 )   /* this test must be static */
593*3117ece4Schristos             FSE_FLUSHBITS(&bitC);
594*3117ece4Schristos 
595*3117ece4Schristos         FSE_encodeSymbol(&bitC, &CState1, *--ip);
596*3117ece4Schristos 
597*3117ece4Schristos         if (sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) {  /* this test must be static */
598*3117ece4Schristos             FSE_encodeSymbol(&bitC, &CState2, *--ip);
599*3117ece4Schristos             FSE_encodeSymbol(&bitC, &CState1, *--ip);
600*3117ece4Schristos         }
601*3117ece4Schristos 
602*3117ece4Schristos         FSE_FLUSHBITS(&bitC);
603*3117ece4Schristos     }
604*3117ece4Schristos 
605*3117ece4Schristos     FSE_flushCState(&bitC, &CState2);
606*3117ece4Schristos     FSE_flushCState(&bitC, &CState1);
607*3117ece4Schristos     return BIT_closeCStream(&bitC);
608*3117ece4Schristos }
609*3117ece4Schristos 
610*3117ece4Schristos size_t FSE_compress_usingCTable (void* dst, size_t dstSize,
611*3117ece4Schristos                            const void* src, size_t srcSize,
612*3117ece4Schristos                            const FSE_CTable* ct)
613*3117ece4Schristos {
614*3117ece4Schristos     unsigned const fast = (dstSize >= FSE_BLOCKBOUND(srcSize));
615*3117ece4Schristos 
616*3117ece4Schristos     if (fast)
617*3117ece4Schristos         return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 1);
618*3117ece4Schristos     else
619*3117ece4Schristos         return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 0);
620*3117ece4Schristos }
621*3117ece4Schristos 
622*3117ece4Schristos 
623*3117ece4Schristos size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); }
624*3117ece4Schristos 
625*3117ece4Schristos #endif   /* FSE_COMMONDEFS_ONLY */
626