/src/zstd/lib/compress/fse_compress.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* ****************************************************************** |
2 | | * FSE : Finite State Entropy encoder |
3 | | * Copyright (c) Meta Platforms, Inc. and affiliates. |
4 | | * |
5 | | * You can contact the author at : |
6 | | * - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy |
7 | | * - Public forum : https://groups.google.com/forum/#!forum/lz4c |
8 | | * |
9 | | * This source code is licensed under both the BSD-style license (found in the |
10 | | * LICENSE file in the root directory of this source tree) and the GPLv2 (found |
11 | | * in the COPYING file in the root directory of this source tree). |
12 | | * You may select, at your option, one of the above-listed licenses. |
13 | | ****************************************************************** */ |
14 | | |
15 | | /* ************************************************************** |
16 | | * Includes |
17 | | ****************************************************************/ |
18 | | #include "../common/compiler.h" |
19 | | #include "../common/mem.h" /* U32, U16, etc. */ |
20 | | #include "../common/debug.h" /* assert, DEBUGLOG */ |
21 | | #include "hist.h" /* HIST_count_wksp */ |
22 | | #include "../common/bitstream.h" |
23 | | #define FSE_STATIC_LINKING_ONLY |
24 | | #include "../common/fse.h" |
25 | | #include "../common/error_private.h" |
26 | | #define ZSTD_DEPS_NEED_MALLOC |
27 | | #define ZSTD_DEPS_NEED_MATH64 |
28 | | #include "../common/zstd_deps.h" /* ZSTD_memset */ |
29 | | #include "../common/bits.h" /* ZSTD_highbit32 */ |
30 | | |
31 | | |
32 | | /* ************************************************************** |
33 | | * Error Management |
34 | | ****************************************************************/ |
35 | 4.92M | #define FSE_isError ERR_isError |
36 | | |
37 | | |
38 | | /* ************************************************************** |
39 | | * Templates |
40 | | ****************************************************************/ |
41 | | /* |
42 | | designed to be included |
43 | | for type-specific functions (template emulation in C) |
44 | | Objective is to write these functions only once, for improved maintenance |
45 | | */ |
46 | | |
47 | | /* safety checks */ |
48 | | #ifndef FSE_FUNCTION_EXTENSION |
49 | | # error "FSE_FUNCTION_EXTENSION must be defined" |
50 | | #endif |
51 | | #ifndef FSE_FUNCTION_TYPE |
52 | | # error "FSE_FUNCTION_TYPE must be defined" |
53 | | #endif |
54 | | |
55 | | /* Function names */ |
56 | | #define FSE_CAT(X,Y) X##Y |
57 | | #define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y) |
58 | | #define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y) |
59 | | |
60 | | |
61 | | /* Function templates */ |
62 | | |
63 | | /* FSE_buildCTable_wksp() : |
64 | | * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`). |
65 | | * wkspSize should be sized to handle worst case situation, which is `1<<max_tableLog * sizeof(FSE_FUNCTION_TYPE)` |
66 | | * workSpace must also be properly aligned with FSE_FUNCTION_TYPE requirements |
67 | | */ |
68 | | size_t FSE_buildCTable_wksp(FSE_CTable* ct, |
69 | | const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, |
70 | | void* workSpace, size_t wkspSize) |
71 | 20.8M | { |
72 | 20.8M | U32 const tableSize = 1 << tableLog; |
73 | 20.8M | U32 const tableMask = tableSize - 1; |
74 | 20.8M | void* const ptr = ct; |
75 | 20.8M | U16* const tableU16 = ( (U16*) ptr) + 2; |
76 | 18.4E | void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableLog ? tableSize>>1 : 1) ; |
77 | 20.8M | FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT); |
78 | 20.8M | U32 const step = FSE_TABLESTEP(tableSize); |
79 | 20.8M | U32 const maxSV1 = maxSymbolValue+1; |
80 | | |
81 | 20.8M | U16* cumul = (U16*)workSpace; /* size = maxSV1 */ |
82 | 20.8M | FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSV1+1)); /* size = tableSize */ |
83 | | |
84 | 20.8M | U32 highThreshold = tableSize-1; |
85 | | |
86 | 20.8M | assert(((size_t)workSpace & 1) == 0); /* Must be 2 bytes-aligned */ |
87 | 20.8M | if (FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) > wkspSize) return ERROR(tableLog_tooLarge); |
88 | | /* CTable header */ |
89 | 20.8M | tableU16[-2] = (U16) tableLog; |
90 | 20.8M | tableU16[-1] = (U16) maxSymbolValue; |
91 | 20.8M | assert(tableLog < 16); /* required for threshold strategy to work */ |
92 | | |
93 | | /* For explanations on how to distribute symbol values over the table : |
94 | | * https://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */ |
95 | | |
96 | | #ifdef __clang_analyzer__ |
97 | | ZSTD_memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize); /* useless initialization, just to keep scan-build happy */ |
98 | | #endif |
99 | | |
100 | | /* symbol start positions */ |
101 | 20.8M | { U32 u; |
102 | 20.8M | cumul[0] = 0; |
103 | 603M | for (u=1; u <= maxSV1; u++) { |
104 | 583M | if (normalizedCounter[u-1]==-1) { /* Low proba symbol */ |
105 | 64.1M | cumul[u] = cumul[u-1] + 1; |
106 | 64.1M | tableSymbol[highThreshold--] = (FSE_FUNCTION_TYPE)(u-1); |
107 | 518M | } else { |
108 | 518M | assert(normalizedCounter[u-1] >= 0); |
109 | 518M | cumul[u] = cumul[u-1] + (U16)normalizedCounter[u-1]; |
110 | 518M | assert(cumul[u] >= cumul[u-1]); /* no overflow */ |
111 | 518M | } } |
112 | 20.8M | cumul[maxSV1] = (U16)(tableSize+1); |
113 | 20.8M | } |
114 | | |
115 | | /* Spread symbols */ |
116 | 20.8M | if (highThreshold == tableSize - 1) { |
117 | | /* Case for no low prob count symbols. Lay down 8 bytes at a time |
118 | | * to reduce branch misses since we are operating on a small block |
119 | | */ |
120 | 9.20M | BYTE* const spread = tableSymbol + tableSize; /* size = tableSize + 8 (may write beyond tableSize) */ |
121 | 9.20M | { U64 const add = 0x0101010101010101ull; |
122 | 9.20M | size_t pos = 0; |
123 | 9.20M | U64 sv = 0; |
124 | 9.20M | U32 s; |
125 | 121M | for (s=0; s<maxSV1; ++s, sv += add) { |
126 | 112M | int i; |
127 | 112M | int const n = normalizedCounter[s]; |
128 | 112M | MEM_write64(spread + pos, sv); |
129 | 149M | for (i = 8; i < n; i += 8) { |
130 | 37.0M | MEM_write64(spread + pos + i, sv); |
131 | 37.0M | } |
132 | 112M | assert(n>=0); |
133 | 112M | pos += (size_t)n; |
134 | 112M | } |
135 | 9.20M | } |
136 | | /* Spread symbols across the table. Lack of lowprob symbols means that |
137 | | * we don't need variable sized inner loop, so we can unroll the loop and |
138 | | * reduce branch misses. |
139 | | */ |
140 | 9.20M | { size_t position = 0; |
141 | 9.20M | size_t s; |
142 | 9.20M | size_t const unroll = 2; /* Experimentally determined optimal unroll */ |
143 | 9.20M | assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */ |
144 | 259M | for (s = 0; s < (size_t)tableSize; s += unroll) { |
145 | 249M | size_t u; |
146 | 749M | for (u = 0; u < unroll; ++u) { |
147 | 499M | size_t const uPosition = (position + (u * step)) & tableMask; |
148 | 499M | tableSymbol[uPosition] = spread[s + u]; |
149 | 499M | } |
150 | 249M | position = (position + (unroll * step)) & tableMask; |
151 | 249M | } |
152 | 9.19M | assert(position == 0); /* Must have initialized all positions */ |
153 | 9.19M | } |
154 | 11.6M | } else { |
155 | 11.6M | U32 position = 0; |
156 | 11.6M | U32 symbol; |
157 | 482M | for (symbol=0; symbol<maxSV1; symbol++) { |
158 | 470M | int nbOccurrences; |
159 | 470M | int const freq = normalizedCounter[symbol]; |
160 | 1.08G | for (nbOccurrences=0; nbOccurrences<freq; nbOccurrences++) { |
161 | 612M | tableSymbol[position] = (FSE_FUNCTION_TYPE)symbol; |
162 | 612M | position = (position + step) & tableMask; |
163 | 677M | while (position > highThreshold) |
164 | 64.1M | position = (position + step) & tableMask; /* Low proba area */ |
165 | 612M | } } |
166 | 11.6M | assert(position==0); /* Must have initialized all positions */ |
167 | 11.6M | } |
168 | | |
169 | | /* Build table */ |
170 | 1.19G | { U32 u; for (u=0; u<tableSize; u++) { |
171 | 1.17G | FSE_FUNCTION_TYPE s = tableSymbol[u]; /* note : static analyzer may not understand tableSymbol is properly initialized */ |
172 | 1.17G | tableU16[cumul[s]++] = (U16) (tableSize+u); /* TableU16 : sorted by symbol order; gives next state value */ |
173 | 1.17G | } } |
174 | | |
175 | | /* Build Symbol Transformation Table */ |
176 | 20.8M | { unsigned total = 0; |
177 | 20.8M | unsigned s; |
178 | 603M | for (s=0; s<=maxSymbolValue; s++) { |
179 | 583M | switch (normalizedCounter[s]) |
180 | 583M | { |
181 | 36.6M | case 0: |
182 | | /* filling nonetheless, for compatibility with FSE_getMaxNbBits() */ |
183 | 36.6M | symbolTT[s].deltaNbBits = ((tableLog+1) << 16) - (1<<tableLog); |
184 | 36.6M | break; |
185 | | |
186 | 64.1M | case -1: |
187 | 367M | case 1: |
188 | 367M | symbolTT[s].deltaNbBits = (tableLog << 16) - (1<<tableLog); |
189 | 367M | assert(total <= INT_MAX); |
190 | 367M | symbolTT[s].deltaFindState = (int)(total - 1); |
191 | 367M | total ++; |
192 | 367M | break; |
193 | 179M | default : |
194 | 179M | assert(normalizedCounter[s] > 1); |
195 | 179M | { U32 const maxBitsOut = tableLog - ZSTD_highbit32 ((U32)normalizedCounter[s]-1); |
196 | 179M | U32 const minStatePlus = (U32)normalizedCounter[s] << maxBitsOut; |
197 | 179M | symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus; |
198 | 179M | symbolTT[s].deltaFindState = (int)(total - (unsigned)normalizedCounter[s]); |
199 | 179M | total += (unsigned)normalizedCounter[s]; |
200 | 179M | } } } } |
201 | | |
202 | | #if 0 /* debug : symbol costs */ |
203 | | DEBUGLOG(5, "\n --- table statistics : "); |
204 | | { U32 symbol; |
205 | | for (symbol=0; symbol<=maxSymbolValue; symbol++) { |
206 | | DEBUGLOG(5, "%3u: w=%3i, maxBits=%u, fracBits=%.2f", |
207 | | symbol, normalizedCounter[symbol], |
208 | | FSE_getMaxNbBits(symbolTT, symbol), |
209 | | (double)FSE_bitCost(symbolTT, tableLog, symbol, 8) / 256); |
210 | | } } |
211 | | #endif |
212 | | |
213 | 20.8M | return 0; |
214 | 20.8M | } |
215 | | |
216 | | |
217 | | |
218 | | #ifndef FSE_COMMONDEFS_ONLY |
219 | | |
220 | | /*-************************************************************** |
221 | | * FSE NCount encoding |
222 | | ****************************************************************/ |
223 | | size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog) |
224 | 16.6M | { |
225 | 16.6M | size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog |
226 | 16.6M | + 4 /* bitCount initialized at 4 */ |
227 | 16.6M | + 2 /* first two symbols may use one additional bit each */) / 8) |
228 | 16.6M | + 1 /* round up to whole nb bytes */ |
229 | 16.6M | + 2 /* additional two bytes for bitstream flush */; |
230 | 18.4E | return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND; /* maxSymbolValue==0 ? use default */ |
231 | 16.6M | } |
232 | | |
233 | | static size_t |
234 | | FSE_writeNCount_generic (void* header, size_t headerBufferSize, |
235 | | const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, |
236 | | unsigned writeIsSafe) |
237 | 16.6M | { |
238 | 16.6M | BYTE* const ostart = (BYTE*) header; |
239 | 16.6M | BYTE* out = ostart; |
240 | 16.6M | BYTE* const oend = ostart + headerBufferSize; |
241 | 16.6M | int nbBits; |
242 | 16.6M | const int tableSize = 1 << tableLog; |
243 | 16.6M | int remaining; |
244 | 16.6M | int threshold; |
245 | 16.6M | U32 bitStream = 0; |
246 | 16.6M | int bitCount = 0; |
247 | 16.6M | unsigned symbol = 0; |
248 | 16.6M | unsigned const alphabetSize = maxSymbolValue + 1; |
249 | 16.6M | int previousIs0 = 0; |
250 | | |
251 | | /* Table Size */ |
252 | 16.6M | bitStream += (tableLog-FSE_MIN_TABLELOG) << bitCount; |
253 | 16.6M | bitCount += 4; |
254 | | |
255 | | /* Init */ |
256 | 16.6M | remaining = tableSize+1; /* +1 for extra accuracy */ |
257 | 16.6M | threshold = tableSize; |
258 | 16.6M | nbBits = (int)tableLog+1; |
259 | | |
260 | 198M | while ((symbol < alphabetSize) && (remaining>1)) { /* stops at 1 */ |
261 | 182M | if (previousIs0) { |
262 | 32.3M | unsigned start = symbol; |
263 | 106M | while ((symbol < alphabetSize) && !normalizedCounter[symbol]) symbol++; |
264 | 32.3M | if (symbol == alphabetSize) break; /* incorrect distribution */ |
265 | 32.8M | while (symbol >= start+24) { |
266 | 505k | start+=24; |
267 | 505k | bitStream += 0xFFFFU << bitCount; |
268 | 505k | if ((!writeIsSafe) && (out > oend-2)) |
269 | 3 | return ERROR(dstSize_tooSmall); /* Buffer overflow */ |
270 | 505k | out[0] = (BYTE) bitStream; |
271 | 505k | out[1] = (BYTE)(bitStream>>8); |
272 | 505k | out+=2; |
273 | 505k | bitStream>>=16; |
274 | 505k | } |
275 | 47.1M | while (symbol >= start+3) { |
276 | 14.7M | start+=3; |
277 | 14.7M | bitStream += 3U << bitCount; |
278 | 14.7M | bitCount += 2; |
279 | 14.7M | } |
280 | 32.3M | bitStream += (symbol-start) << bitCount; |
281 | 32.3M | bitCount += 2; |
282 | 32.3M | if (bitCount>16) { |
283 | 5.44M | if ((!writeIsSafe) && (out > oend - 2)) |
284 | 6 | return ERROR(dstSize_tooSmall); /* Buffer overflow */ |
285 | 5.44M | out[0] = (BYTE)bitStream; |
286 | 5.44M | out[1] = (BYTE)(bitStream>>8); |
287 | 5.44M | out += 2; |
288 | 5.44M | bitStream >>= 16; |
289 | 5.44M | bitCount -= 16; |
290 | 5.44M | } } |
291 | 182M | { int count = normalizedCounter[symbol++]; |
292 | 182M | int const max = (2*threshold-1) - remaining; |
293 | 182M | remaining -= count < 0 ? -count : count; |
294 | 182M | count++; /* +1 for extra accuracy */ |
295 | 182M | if (count>=threshold) |
296 | 30.2M | count += max; /* [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[ */ |
297 | 182M | bitStream += (U32)count << bitCount; |
298 | 182M | bitCount += nbBits; |
299 | 182M | bitCount -= (count<max); |
300 | 182M | previousIs0 = (count==1); |
301 | 182M | if (remaining<1) return ERROR(GENERIC); |
302 | 276M | while (remaining<threshold) { nbBits--; threshold>>=1; } |
303 | 182M | } |
304 | 182M | if (bitCount>16) { |
305 | 40.8M | if ((!writeIsSafe) && (out > oend - 2)) |
306 | 30 | return ERROR(dstSize_tooSmall); /* Buffer overflow */ |
307 | 40.8M | out[0] = (BYTE)bitStream; |
308 | 40.8M | out[1] = (BYTE)(bitStream>>8); |
309 | 40.8M | out += 2; |
310 | 40.8M | bitStream >>= 16; |
311 | 40.8M | bitCount -= 16; |
312 | 40.8M | } } |
313 | | |
314 | 16.6M | if (remaining != 1) |
315 | 0 | return ERROR(GENERIC); /* incorrect normalized distribution */ |
316 | 16.6M | assert(symbol <= alphabetSize); |
317 | | |
318 | | /* flush remaining bitStream */ |
319 | 16.6M | if ((!writeIsSafe) && (out > oend - 2)) |
320 | 38 | return ERROR(dstSize_tooSmall); /* Buffer overflow */ |
321 | 16.6M | out[0] = (BYTE)bitStream; |
322 | 16.6M | out[1] = (BYTE)(bitStream>>8); |
323 | 16.6M | out+= (bitCount+7) /8; |
324 | | |
325 | 16.6M | assert(out >= ostart); |
326 | 16.6M | return (size_t)(out-ostart); |
327 | 16.6M | } |
328 | | |
329 | | |
330 | | size_t FSE_writeNCount (void* buffer, size_t bufferSize, |
331 | | const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog) |
332 | 16.6M | { |
333 | 16.6M | if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge); /* Unsupported */ |
334 | 16.6M | if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC); /* Unsupported */ |
335 | | |
336 | 16.6M | if (bufferSize < FSE_NCountWriteBound(maxSymbolValue, tableLog)) |
337 | 294 | return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 0); |
338 | | |
339 | 16.6M | return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 1 /* write in buffer is safe */); |
340 | 16.6M | } |
341 | | |
342 | | |
343 | | /*-************************************************************** |
344 | | * FSE Compression Code |
345 | | ****************************************************************/ |
346 | | |
347 | | /* provides the minimum logSize to safely represent a distribution */ |
348 | | static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue) |
349 | 34.9M | { |
350 | 34.9M | U32 minBitsSrc = ZSTD_highbit32((U32)(srcSize)) + 1; |
351 | 34.9M | U32 minBitsSymbols = ZSTD_highbit32(maxSymbolValue) + 2; |
352 | 34.9M | U32 minBits = minBitsSrc < minBitsSymbols ? minBitsSrc : minBitsSymbols; |
353 | 34.9M | assert(srcSize > 1); /* Not supported, RLE should be used instead */ |
354 | 34.9M | return minBits; |
355 | 34.9M | } |
356 | | |
357 | | unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus) |
358 | 18.3M | { |
359 | 18.3M | U32 maxBitsSrc = ZSTD_highbit32((U32)(srcSize - 1)) - minus; |
360 | 18.3M | U32 tableLog = maxTableLog; |
361 | 18.3M | U32 minBits = FSE_minTableLog(srcSize, maxSymbolValue); |
362 | 18.3M | assert(srcSize > 1); /* Not supported, RLE should be used instead */ |
363 | 18.3M | if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG; |
364 | 18.3M | if (maxBitsSrc < tableLog) tableLog = maxBitsSrc; /* Accuracy can be reduced */ |
365 | 18.3M | if (minBits > tableLog) tableLog = minBits; /* Need a minimum to safely represent all symbol values */ |
366 | 18.3M | if (tableLog < FSE_MIN_TABLELOG) tableLog = FSE_MIN_TABLELOG; |
367 | 18.3M | if (tableLog > FSE_MAX_TABLELOG) tableLog = FSE_MAX_TABLELOG; |
368 | 18.3M | return tableLog; |
369 | 18.3M | } |
370 | | |
371 | | unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue) |
372 | 16.4M | { |
373 | 16.4M | return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 2); |
374 | 16.4M | } |
375 | | |
376 | | /* Secondary normalization method. |
377 | | To be used when primary method fails. */ |
378 | | |
379 | | static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue, short lowProbCount) |
380 | 180k | { |
381 | 180k | short const NOT_YET_ASSIGNED = -2; |
382 | 180k | U32 s; |
383 | 180k | U32 distributed = 0; |
384 | 180k | U32 ToDistribute; |
385 | | |
386 | | /* Init */ |
387 | 180k | U32 const lowThreshold = (U32)(total >> tableLog); |
388 | 180k | U32 lowOne = (U32)((total * 3) >> (tableLog + 1)); |
389 | | |
390 | 6.16M | for (s=0; s<=maxSymbolValue; s++) { |
391 | 5.98M | if (count[s] == 0) { |
392 | 3.05M | norm[s]=0; |
393 | 3.05M | continue; |
394 | 3.05M | } |
395 | 2.92M | if (count[s] <= lowThreshold) { |
396 | 299k | norm[s] = lowProbCount; |
397 | 299k | distributed++; |
398 | 299k | total -= count[s]; |
399 | 299k | continue; |
400 | 299k | } |
401 | 2.62M | if (count[s] <= lowOne) { |
402 | 384k | norm[s] = 1; |
403 | 384k | distributed++; |
404 | 384k | total -= count[s]; |
405 | 384k | continue; |
406 | 384k | } |
407 | | |
408 | 2.23M | norm[s]=NOT_YET_ASSIGNED; |
409 | 2.23M | } |
410 | 180k | ToDistribute = (1 << tableLog) - distributed; |
411 | | |
412 | 180k | if (ToDistribute == 0) |
413 | 0 | return 0; |
414 | | |
415 | 180k | if ((total / ToDistribute) > lowOne) { |
416 | | /* risk of rounding to zero */ |
417 | 108 | lowOne = (U32)((total * 3) / (ToDistribute * 2)); |
418 | 2.54k | for (s=0; s<=maxSymbolValue; s++) { |
419 | 2.43k | if ((norm[s] == NOT_YET_ASSIGNED) && (count[s] <= lowOne)) { |
420 | 94 | norm[s] = 1; |
421 | 94 | distributed++; |
422 | 94 | total -= count[s]; |
423 | 94 | continue; |
424 | 94 | } } |
425 | 108 | ToDistribute = (1 << tableLog) - distributed; |
426 | 108 | } |
427 | | |
428 | 180k | if (distributed == maxSymbolValue+1) { |
429 | | /* all values are pretty poor; |
430 | | probably incompressible data (should have already been detected); |
431 | | find max, then give all remaining points to max */ |
432 | 0 | U32 maxV = 0, maxC = 0; |
433 | 0 | for (s=0; s<=maxSymbolValue; s++) |
434 | 0 | if (count[s] > maxC) { maxV=s; maxC=count[s]; } |
435 | 0 | norm[maxV] += (short)ToDistribute; |
436 | 0 | return 0; |
437 | 0 | } |
438 | | |
439 | 180k | if (total == 0) { |
440 | | /* all of the symbols were low enough for the lowOne or lowThreshold */ |
441 | 1.27k | for (s=0; ToDistribute > 0; s = (s+1)%(maxSymbolValue+1)) |
442 | 1.20k | if (norm[s] > 0) { ToDistribute--; norm[s]++; } |
443 | 71 | return 0; |
444 | 71 | } |
445 | | |
446 | 180k | { U64 const vStepLog = 62 - tableLog; |
447 | 180k | U64 const mid = (1ULL << (vStepLog-1)) - 1; |
448 | 180k | U64 const rStep = ZSTD_div64((((U64)1<<vStepLog) * ToDistribute) + mid, (U32)total); /* scale on remaining */ |
449 | 180k | U64 tmpTotal = mid; |
450 | 6.16M | for (s=0; s<=maxSymbolValue; s++) { |
451 | 5.97M | if (norm[s]==NOT_YET_ASSIGNED) { |
452 | 2.23M | U64 const end = tmpTotal + (count[s] * rStep); |
453 | 2.23M | U32 const sStart = (U32)(tmpTotal >> vStepLog); |
454 | 2.23M | U32 const sEnd = (U32)(end >> vStepLog); |
455 | 2.23M | U32 const weight = sEnd - sStart; |
456 | 2.23M | if (weight < 1) |
457 | 0 | return ERROR(GENERIC); |
458 | 2.23M | norm[s] = (short)weight; |
459 | 2.23M | tmpTotal = end; |
460 | 2.23M | } } } |
461 | | |
462 | 180k | return 0; |
463 | 180k | } |
464 | | |
465 | | size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog, |
466 | | const unsigned* count, size_t total, |
467 | | unsigned maxSymbolValue, unsigned useLowProbCount) |
468 | 16.6M | { |
469 | | /* Sanity checks */ |
470 | 16.6M | if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG; |
471 | 16.6M | if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC); /* Unsupported size */ |
472 | 16.6M | if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge); /* Unsupported size */ |
473 | 16.6M | if (tableLog < FSE_minTableLog(total, maxSymbolValue)) return ERROR(GENERIC); /* Too small tableLog, compression potentially impossible */ |
474 | | |
475 | 16.6M | { static U32 const rtbTable[] = { 0, 473195, 504333, 520860, 550000, 700000, 750000, 830000 }; |
476 | 16.6M | short const lowProbCount = useLowProbCount ? -1 : 1; |
477 | 16.6M | U64 const scale = 62 - tableLog; |
478 | 16.6M | U64 const step = ZSTD_div64((U64)1<<62, (U32)total); /* <== here, one division ! */ |
479 | 16.6M | U64 const vStep = 1ULL<<(scale-20); |
480 | 16.6M | int stillToDistribute = 1<<tableLog; |
481 | 16.6M | unsigned s; |
482 | 16.6M | unsigned largest=0; |
483 | 16.6M | short largestP=0; |
484 | 16.6M | U32 lowThreshold = (U32)(total >> tableLog); |
485 | | |
486 | 272M | for (s=0; s<=maxSymbolValue; s++) { |
487 | 256M | if (count[s] == total) return 0; /* rle special case */ |
488 | 256M | if (count[s] == 0) { normalizedCounter[s]=0; continue; } |
489 | 149M | if (count[s] <= lowThreshold) { |
490 | 40.0M | normalizedCounter[s] = lowProbCount; |
491 | 40.0M | stillToDistribute--; |
492 | 109M | } else { |
493 | 109M | short proba = (short)((count[s]*step) >> scale); |
494 | 109M | if (proba<8) { |
495 | 80.8M | U64 restToBeat = vStep * rtbTable[proba]; |
496 | 80.8M | proba += (count[s]*step) - ((U64)proba<<scale) > restToBeat; |
497 | 80.8M | } |
498 | 109M | if (proba > largestP) { largestP=proba; largest=s; } |
499 | 109M | normalizedCounter[s] = proba; |
500 | 109M | stillToDistribute -= proba; |
501 | 109M | } } |
502 | 16.6M | if (-stillToDistribute >= (normalizedCounter[largest] >> 1)) { |
503 | | /* corner case, need another normalization method */ |
504 | 180k | size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue, lowProbCount); |
505 | 180k | if (FSE_isError(errorCode)) return errorCode; |
506 | 180k | } |
507 | 16.4M | else normalizedCounter[largest] += (short)stillToDistribute; |
508 | 16.6M | } |
509 | | |
510 | | #if 0 |
511 | | { /* Print Table (debug) */ |
512 | | U32 s; |
513 | | U32 nTotal = 0; |
514 | | for (s=0; s<=maxSymbolValue; s++) |
515 | | RAWLOG(2, "%3i: %4i \n", s, normalizedCounter[s]); |
516 | | for (s=0; s<=maxSymbolValue; s++) |
517 | | nTotal += abs(normalizedCounter[s]); |
518 | | if (nTotal != (1U<<tableLog)) |
519 | | RAWLOG(2, "Warning !!! Total == %u != %u !!!", nTotal, 1U<<tableLog); |
520 | | getchar(); |
521 | | } |
522 | | #endif |
523 | | |
524 | 16.6M | return tableLog; |
525 | 16.6M | } |
526 | | |
527 | | /* fake FSE_CTable, for rle input (always same symbol) */ |
528 | | size_t FSE_buildCTable_rle (FSE_CTable* ct, BYTE symbolValue) |
529 | 959k | { |
530 | 959k | void* ptr = ct; |
531 | 959k | U16* tableU16 = ( (U16*) ptr) + 2; |
532 | 959k | void* FSCTptr = (U32*)ptr + 2; |
533 | 959k | FSE_symbolCompressionTransform* symbolTT = (FSE_symbolCompressionTransform*) FSCTptr; |
534 | | |
535 | | /* header */ |
536 | 959k | tableU16[-2] = (U16) 0; |
537 | 959k | tableU16[-1] = (U16) symbolValue; |
538 | | |
539 | | /* Build table */ |
540 | 959k | tableU16[0] = 0; |
541 | 959k | tableU16[1] = 0; /* just in case */ |
542 | | |
543 | | /* Build Symbol Transformation Table */ |
544 | 959k | symbolTT[symbolValue].deltaNbBits = 0; |
545 | 959k | symbolTT[symbolValue].deltaFindState = 0; |
546 | | |
547 | 959k | return 0; |
548 | 959k | } |
549 | | |
550 | | |
551 | | static size_t FSE_compress_usingCTable_generic (void* dst, size_t dstSize, |
552 | | const void* src, size_t srcSize, |
553 | | const FSE_CTable* ct, const unsigned fast) |
554 | 4.74M | { |
555 | 4.74M | const BYTE* const istart = (const BYTE*) src; |
556 | 4.74M | const BYTE* const iend = istart + srcSize; |
557 | 4.74M | const BYTE* ip=iend; |
558 | | |
559 | 4.74M | BIT_CStream_t bitC; |
560 | 4.74M | FSE_CState_t CState1, CState2; |
561 | | |
562 | | /* init */ |
563 | 4.74M | if (srcSize <= 2) return 0; |
564 | 4.74M | { size_t const initError = BIT_initCStream(&bitC, dst, dstSize); |
565 | 4.74M | if (FSE_isError(initError)) return 0; /* not enough space available to write a bitstream */ } |
566 | | |
567 | 251M | #define FSE_FLUSHBITS(s) (fast ? BIT_flushBitsFast(s) : BIT_flushBits(s)) |
568 | | |
569 | 4.74M | if (srcSize & 1) { |
570 | 3.45M | FSE_initCState2(&CState1, ct, *--ip); |
571 | 3.45M | FSE_initCState2(&CState2, ct, *--ip); |
572 | 3.45M | FSE_encodeSymbol(&bitC, &CState1, *--ip); |
573 | 3.45M | FSE_FLUSHBITS(&bitC); |
574 | 3.45M | } else { |
575 | 1.29M | FSE_initCState2(&CState2, ct, *--ip); |
576 | 1.29M | FSE_initCState2(&CState1, ct, *--ip); |
577 | 1.29M | } |
578 | | |
579 | | /* join to mod 4 */ |
580 | 4.74M | srcSize -= 2; |
581 | 4.74M | if ((sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) && (srcSize & 2)) { /* test bit 2 */ |
582 | 771k | FSE_encodeSymbol(&bitC, &CState2, *--ip); |
583 | 771k | FSE_encodeSymbol(&bitC, &CState1, *--ip); |
584 | 771k | FSE_FLUSHBITS(&bitC); |
585 | 771k | } |
586 | | |
587 | | /* 2 or 4 encoding per loop */ |
588 | 252M | while ( ip>istart ) { |
589 | | |
590 | 247M | FSE_encodeSymbol(&bitC, &CState2, *--ip); |
591 | | |
592 | 247M | if (sizeof(bitC.bitContainer)*8 < FSE_MAX_TABLELOG*2+7 ) /* this test must be static */ |
593 | 0 | FSE_FLUSHBITS(&bitC); |
594 | | |
595 | 247M | FSE_encodeSymbol(&bitC, &CState1, *--ip); |
596 | | |
597 | 247M | if (sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) { /* this test must be static */ |
598 | 247M | FSE_encodeSymbol(&bitC, &CState2, *--ip); |
599 | 247M | FSE_encodeSymbol(&bitC, &CState1, *--ip); |
600 | 247M | } |
601 | | |
602 | 247M | FSE_FLUSHBITS(&bitC); |
603 | 247M | } |
604 | | |
605 | 4.74M | FSE_flushCState(&bitC, &CState2); |
606 | 4.74M | FSE_flushCState(&bitC, &CState1); |
607 | 4.74M | return BIT_closeCStream(&bitC); |
608 | 4.74M | } |
609 | | |
610 | | size_t FSE_compress_usingCTable (void* dst, size_t dstSize, |
611 | | const void* src, size_t srcSize, |
612 | | const FSE_CTable* ct) |
613 | 4.74M | { |
614 | 4.74M | unsigned const fast = (dstSize >= FSE_BLOCKBOUND(srcSize)); |
615 | | |
616 | 4.74M | if (fast) |
617 | 3.38M | return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 1); |
618 | 1.36M | else |
619 | 1.36M | return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 0); |
620 | 4.74M | } |
621 | | |
622 | | |
623 | 0 | size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); } |
624 | | |
625 | | #endif /* FSE_COMMONDEFS_ONLY */ |