1/*
2 * Copyright (c) Yann Collet, Facebook, Inc.
3 * All rights reserved.
4 *
5 * This source code is licensed under both the BSD-style license (found in the
6 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
7 * in the COPYING file in the root directory of this source tree).
8 * You may select, at your option, one of the above-listed licenses.
9 */
10
11 /*-*************************************
12 * Dependencies
13 ***************************************/
14#include "zstd_compress_sequences.h"
15
16/*
17 * -log2(x / 256) lookup table for x in [0, 256).
18 * If x == 0: Return 0
19 * Else: Return floor(-log2(x / 256) * 256)
20 */
21static unsigned const kInverseProbabilityLog256[256] = {
22 0, 2048, 1792, 1642, 1536, 1453, 1386, 1329, 1280, 1236, 1197, 1162,
23 1130, 1100, 1073, 1047, 1024, 1001, 980, 960, 941, 923, 906, 889,
24 874, 859, 844, 830, 817, 804, 791, 779, 768, 756, 745, 734,
25 724, 714, 704, 694, 685, 676, 667, 658, 650, 642, 633, 626,
26 618, 610, 603, 595, 588, 581, 574, 567, 561, 554, 548, 542,
27 535, 529, 523, 517, 512, 506, 500, 495, 489, 484, 478, 473,
28 468, 463, 458, 453, 448, 443, 438, 434, 429, 424, 420, 415,
29 411, 407, 402, 398, 394, 390, 386, 382, 377, 373, 370, 366,
30 362, 358, 354, 350, 347, 343, 339, 336, 332, 329, 325, 322,
31 318, 315, 311, 308, 305, 302, 298, 295, 292, 289, 286, 282,
32 279, 276, 273, 270, 267, 264, 261, 258, 256, 253, 250, 247,
33 244, 241, 239, 236, 233, 230, 228, 225, 222, 220, 217, 215,
34 212, 209, 207, 204, 202, 199, 197, 194, 192, 190, 187, 185,
35 182, 180, 178, 175, 173, 171, 168, 166, 164, 162, 159, 157,
36 155, 153, 151, 149, 146, 144, 142, 140, 138, 136, 134, 132,
37 130, 128, 126, 123, 121, 119, 117, 115, 114, 112, 110, 108,
38 106, 104, 102, 100, 98, 96, 94, 93, 91, 89, 87, 85,
39 83, 82, 80, 78, 76, 74, 73, 71, 69, 67, 66, 64,
40 62, 61, 59, 57, 55, 54, 52, 50, 49, 47, 46, 44,
41 42, 41, 39, 37, 36, 34, 33, 31, 30, 28, 26, 25,
42 23, 22, 20, 19, 17, 16, 14, 13, 11, 10, 8, 7,
43 5, 4, 2, 1,
44};
45
46static unsigned ZSTD_getFSEMaxSymbolValue(FSE_CTable const* ctable) {
47 void const* ptr = ctable;
48 U16 const* u16ptr = (U16 const*)ptr;
49 U32 const maxSymbolValue = MEM_read16(memPtr: u16ptr + 1);
50 return maxSymbolValue;
51}
52
53/*
54 * Returns true if we should use ncount=-1 else we should
55 * use ncount=1 for low probability symbols instead.
56 */
57static unsigned ZSTD_useLowProbCount(size_t const nbSeq)
58{
59 /* Heuristic: This should cover most blocks <= 16K and
60 * start to fade out after 16K to about 32K depending on
61 * comprssibility.
62 */
63 return nbSeq >= 2048;
64}
65
66/*
67 * Returns the cost in bytes of encoding the normalized count header.
68 * Returns an error if any of the helper functions return an error.
69 */
70static size_t ZSTD_NCountCost(unsigned const* count, unsigned const max,
71 size_t const nbSeq, unsigned const FSELog)
72{
73 BYTE wksp[FSE_NCOUNTBOUND];
74 S16 norm[MaxSeq + 1];
75 const U32 tableLog = FSE_optimalTableLog(maxTableLog: FSELog, srcSize: nbSeq, maxSymbolValue: max);
76 FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq, max, ZSTD_useLowProbCount(nbSeq)), "");
77 return FSE_writeNCount(buffer: wksp, bufferSize: sizeof(wksp), normalizedCounter: norm, maxSymbolValue: max, tableLog);
78}
79
80/*
81 * Returns the cost in bits of encoding the distribution described by count
82 * using the entropy bound.
83 */
84static size_t ZSTD_entropyCost(unsigned const* count, unsigned const max, size_t const total)
85{
86 unsigned cost = 0;
87 unsigned s;
88
89 assert(total > 0);
90 for (s = 0; s <= max; ++s) {
91 unsigned norm = (unsigned)((256 * count[s]) / total);
92 if (count[s] != 0 && norm == 0)
93 norm = 1;
94 assert(count[s] < total);
95 cost += count[s] * kInverseProbabilityLog256[norm];
96 }
97 return cost >> 8;
98}
99
100/*
101 * Returns the cost in bits of encoding the distribution in count using ctable.
102 * Returns an error if ctable cannot represent all the symbols in count.
103 */
104size_t ZSTD_fseBitCost(
105 FSE_CTable const* ctable,
106 unsigned const* count,
107 unsigned const max)
108{
109 unsigned const kAccuracyLog = 8;
110 size_t cost = 0;
111 unsigned s;
112 FSE_CState_t cstate;
113 FSE_initCState(statePtr: &cstate, ct: ctable);
114 if (ZSTD_getFSEMaxSymbolValue(ctable) < max) {
115 DEBUGLOG(5, "Repeat FSE_CTable has maxSymbolValue %u < %u",
116 ZSTD_getFSEMaxSymbolValue(ctable), max);
117 return ERROR(GENERIC);
118 }
119 for (s = 0; s <= max; ++s) {
120 unsigned const tableLog = cstate.stateLog;
121 unsigned const badCost = (tableLog + 1) << kAccuracyLog;
122 unsigned const bitCost = FSE_bitCost(symbolTTPtr: cstate.symbolTT, tableLog, symbolValue: s, accuracyLog: kAccuracyLog);
123 if (count[s] == 0)
124 continue;
125 if (bitCost >= badCost) {
126 DEBUGLOG(5, "Repeat FSE_CTable has Prob[%u] == 0", s);
127 return ERROR(GENERIC);
128 }
129 cost += (size_t)count[s] * bitCost;
130 }
131 return cost >> kAccuracyLog;
132}
133
134/*
135 * Returns the cost in bits of encoding the distribution in count using the
136 * table described by norm. The max symbol support by norm is assumed >= max.
137 * norm must be valid for every symbol with non-zero probability in count.
138 */
139size_t ZSTD_crossEntropyCost(short const* norm, unsigned accuracyLog,
140 unsigned const* count, unsigned const max)
141{
142 unsigned const shift = 8 - accuracyLog;
143 size_t cost = 0;
144 unsigned s;
145 assert(accuracyLog <= 8);
146 for (s = 0; s <= max; ++s) {
147 unsigned const normAcc = (norm[s] != -1) ? (unsigned)norm[s] : 1;
148 unsigned const norm256 = normAcc << shift;
149 assert(norm256 > 0);
150 assert(norm256 < 256);
151 cost += count[s] * kInverseProbabilityLog256[norm256];
152 }
153 return cost >> 8;
154}
155
156symbolEncodingType_e
157ZSTD_selectEncodingType(
158 FSE_repeat* repeatMode, unsigned const* count, unsigned const max,
159 size_t const mostFrequent, size_t nbSeq, unsigned const FSELog,
160 FSE_CTable const* prevCTable,
161 short const* defaultNorm, U32 defaultNormLog,
162 ZSTD_defaultPolicy_e const isDefaultAllowed,
163 ZSTD_strategy const strategy)
164{
165 ZSTD_STATIC_ASSERT(ZSTD_defaultDisallowed == 0 && ZSTD_defaultAllowed != 0);
166 if (mostFrequent == nbSeq) {
167 *repeatMode = FSE_repeat_none;
168 if (isDefaultAllowed && nbSeq <= 2) {
169 /* Prefer set_basic over set_rle when there are 2 or less symbols,
170 * since RLE uses 1 byte, but set_basic uses 5-6 bits per symbol.
171 * If basic encoding isn't possible, always choose RLE.
172 */
173 DEBUGLOG(5, "Selected set_basic");
174 return set_basic;
175 }
176 DEBUGLOG(5, "Selected set_rle");
177 return set_rle;
178 }
179 if (strategy < ZSTD_lazy) {
180 if (isDefaultAllowed) {
181 size_t const staticFse_nbSeq_max = 1000;
182 size_t const mult = 10 - strategy;
183 size_t const baseLog = 3;
184 size_t const dynamicFse_nbSeq_min = (((size_t)1 << defaultNormLog) * mult) >> baseLog; /* 28-36 for offset, 56-72 for lengths */
185 assert(defaultNormLog >= 5 && defaultNormLog <= 6); /* xx_DEFAULTNORMLOG */
186 assert(mult <= 9 && mult >= 7);
187 if ( (*repeatMode == FSE_repeat_valid)
188 && (nbSeq < staticFse_nbSeq_max) ) {
189 DEBUGLOG(5, "Selected set_repeat");
190 return set_repeat;
191 }
192 if ( (nbSeq < dynamicFse_nbSeq_min)
193 || (mostFrequent < (nbSeq >> (defaultNormLog-1))) ) {
194 DEBUGLOG(5, "Selected set_basic");
195 /* The format allows default tables to be repeated, but it isn't useful.
196 * When using simple heuristics to select encoding type, we don't want
197 * to confuse these tables with dictionaries. When running more careful
198 * analysis, we don't need to waste time checking both repeating tables
199 * and default tables.
200 */
201 *repeatMode = FSE_repeat_none;
202 return set_basic;
203 }
204 }
205 } else {
206 size_t const basicCost = isDefaultAllowed ? ZSTD_crossEntropyCost(norm: defaultNorm, accuracyLog: defaultNormLog, count, max) : ERROR(GENERIC);
207 size_t const repeatCost = *repeatMode != FSE_repeat_none ? ZSTD_fseBitCost(ctable: prevCTable, count, max) : ERROR(GENERIC);
208 size_t const NCountCost = ZSTD_NCountCost(count, max, nbSeq, FSELog);
209 size_t const compressedCost = (NCountCost << 3) + ZSTD_entropyCost(count, max, total: nbSeq);
210
211 if (isDefaultAllowed) {
212 assert(!ZSTD_isError(basicCost));
213 assert(!(*repeatMode == FSE_repeat_valid && ZSTD_isError(repeatCost)));
214 }
215 assert(!ZSTD_isError(NCountCost));
216 assert(compressedCost < ERROR(maxCode));
217 DEBUGLOG(5, "Estimated bit costs: basic=%u\trepeat=%u\tcompressed=%u",
218 (unsigned)basicCost, (unsigned)repeatCost, (unsigned)compressedCost);
219 if (basicCost <= repeatCost && basicCost <= compressedCost) {
220 DEBUGLOG(5, "Selected set_basic");
221 assert(isDefaultAllowed);
222 *repeatMode = FSE_repeat_none;
223 return set_basic;
224 }
225 if (repeatCost <= compressedCost) {
226 DEBUGLOG(5, "Selected set_repeat");
227 assert(!ZSTD_isError(repeatCost));
228 return set_repeat;
229 }
230 assert(compressedCost < basicCost && compressedCost < repeatCost);
231 }
232 DEBUGLOG(5, "Selected set_compressed");
233 *repeatMode = FSE_repeat_check;
234 return set_compressed;
235}
236
237typedef struct {
238 S16 norm[MaxSeq + 1];
239 U32 wksp[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(MaxSeq, MaxFSELog)];
240} ZSTD_BuildCTableWksp;
241
242size_t
243ZSTD_buildCTable(void* dst, size_t dstCapacity,
244 FSE_CTable* nextCTable, U32 FSELog, symbolEncodingType_e type,
245 unsigned* count, U32 max,
246 const BYTE* codeTable, size_t nbSeq,
247 const S16* defaultNorm, U32 defaultNormLog, U32 defaultMax,
248 const FSE_CTable* prevCTable, size_t prevCTableSize,
249 void* entropyWorkspace, size_t entropyWorkspaceSize)
250{
251 BYTE* op = (BYTE*)dst;
252 const BYTE* const oend = op + dstCapacity;
253 DEBUGLOG(6, "ZSTD_buildCTable (dstCapacity=%u)", (unsigned)dstCapacity);
254
255 switch (type) {
256 case set_rle:
257 FORWARD_IF_ERROR(FSE_buildCTable_rle(nextCTable, (BYTE)max), "");
258 RETURN_ERROR_IF(dstCapacity==0, dstSize_tooSmall, "not enough space");
259 *op = codeTable[0];
260 return 1;
261 case set_repeat:
262 ZSTD_memcpy(nextCTable, prevCTable, prevCTableSize);
263 return 0;
264 case set_basic:
265 FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, defaultNorm, defaultMax, defaultNormLog, entropyWorkspace, entropyWorkspaceSize), ""); /* note : could be pre-calculated */
266 return 0;
267 case set_compressed: {
268 ZSTD_BuildCTableWksp* wksp = (ZSTD_BuildCTableWksp*)entropyWorkspace;
269 size_t nbSeq_1 = nbSeq;
270 const U32 tableLog = FSE_optimalTableLog(maxTableLog: FSELog, srcSize: nbSeq, maxSymbolValue: max);
271 if (count[codeTable[nbSeq-1]] > 1) {
272 count[codeTable[nbSeq-1]]--;
273 nbSeq_1--;
274 }
275 assert(nbSeq_1 > 1);
276 assert(entropyWorkspaceSize >= sizeof(ZSTD_BuildCTableWksp));
277 (void)entropyWorkspaceSize;
278 FORWARD_IF_ERROR(FSE_normalizeCount(wksp->norm, tableLog, count, nbSeq_1, max, ZSTD_useLowProbCount(nbSeq_1)), "FSE_normalizeCount failed");
279 assert(oend >= op);
280 { size_t const NCountSize = FSE_writeNCount(buffer: op, bufferSize: (size_t)(oend - op), normalizedCounter: wksp->norm, maxSymbolValue: max, tableLog); /* overflow protected */
281 FORWARD_IF_ERROR(NCountSize, "FSE_writeNCount failed");
282 FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, wksp->norm, max, tableLog, wksp->wksp, sizeof(wksp->wksp)), "FSE_buildCTable_wksp failed");
283 return NCountSize;
284 }
285 }
286 default: assert(0); RETURN_ERROR(GENERIC, "impossible to reach");
287 }
288}
289
290FORCE_INLINE_TEMPLATE size_t
291ZSTD_encodeSequences_body(
292 void* dst, size_t dstCapacity,
293 FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
294 FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
295 FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
296 seqDef const* sequences, size_t nbSeq, int longOffsets)
297{
298 BIT_CStream_t blockStream;
299 FSE_CState_t stateMatchLength;
300 FSE_CState_t stateOffsetBits;
301 FSE_CState_t stateLitLength;
302
303 RETURN_ERROR_IF(
304 ERR_isError(BIT_initCStream(&blockStream, dst, dstCapacity)),
305 dstSize_tooSmall, "not enough space remaining");
306 DEBUGLOG(6, "available space for bitstream : %i (dstCapacity=%u)",
307 (int)(blockStream.endPtr - blockStream.startPtr),
308 (unsigned)dstCapacity);
309
310 /* first symbols */
311 FSE_initCState2(statePtr: &stateMatchLength, ct: CTable_MatchLength, symbol: mlCodeTable[nbSeq-1]);
312 FSE_initCState2(statePtr: &stateOffsetBits, ct: CTable_OffsetBits, symbol: ofCodeTable[nbSeq-1]);
313 FSE_initCState2(statePtr: &stateLitLength, ct: CTable_LitLength, symbol: llCodeTable[nbSeq-1]);
314 BIT_addBits(bitC: &blockStream, value: sequences[nbSeq-1].litLength, nbBits: LL_bits[llCodeTable[nbSeq-1]]);
315 if (MEM_32bits()) BIT_flushBits(bitC: &blockStream);
316 BIT_addBits(bitC: &blockStream, value: sequences[nbSeq-1].mlBase, nbBits: ML_bits[mlCodeTable[nbSeq-1]]);
317 if (MEM_32bits()) BIT_flushBits(bitC: &blockStream);
318 if (longOffsets) {
319 U32 const ofBits = ofCodeTable[nbSeq-1];
320 unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1);
321 if (extraBits) {
322 BIT_addBits(bitC: &blockStream, value: sequences[nbSeq-1].offBase, nbBits: extraBits);
323 BIT_flushBits(bitC: &blockStream);
324 }
325 BIT_addBits(bitC: &blockStream, value: sequences[nbSeq-1].offBase >> extraBits,
326 nbBits: ofBits - extraBits);
327 } else {
328 BIT_addBits(bitC: &blockStream, value: sequences[nbSeq-1].offBase, nbBits: ofCodeTable[nbSeq-1]);
329 }
330 BIT_flushBits(bitC: &blockStream);
331
332 { size_t n;
333 for (n=nbSeq-2 ; n<nbSeq ; n--) { /* intentional underflow */
334 BYTE const llCode = llCodeTable[n];
335 BYTE const ofCode = ofCodeTable[n];
336 BYTE const mlCode = mlCodeTable[n];
337 U32 const llBits = LL_bits[llCode];
338 U32 const ofBits = ofCode;
339 U32 const mlBits = ML_bits[mlCode];
340 DEBUGLOG(6, "encoding: litlen:%2u - matchlen:%2u - offCode:%7u",
341 (unsigned)sequences[n].litLength,
342 (unsigned)sequences[n].mlBase + MINMATCH,
343 (unsigned)sequences[n].offBase);
344 /* 32b*/ /* 64b*/
345 /* (7)*/ /* (7)*/
346 FSE_encodeSymbol(bitC: &blockStream, statePtr: &stateOffsetBits, symbol: ofCode); /* 15 */ /* 15 */
347 FSE_encodeSymbol(bitC: &blockStream, statePtr: &stateMatchLength, symbol: mlCode); /* 24 */ /* 24 */
348 if (MEM_32bits()) BIT_flushBits(bitC: &blockStream); /* (7)*/
349 FSE_encodeSymbol(bitC: &blockStream, statePtr: &stateLitLength, symbol: llCode); /* 16 */ /* 33 */
350 if (MEM_32bits() || (ofBits+mlBits+llBits >= 64-7-(LLFSELog+MLFSELog+OffFSELog)))
351 BIT_flushBits(bitC: &blockStream); /* (7)*/
352 BIT_addBits(bitC: &blockStream, value: sequences[n].litLength, nbBits: llBits);
353 if (MEM_32bits() && ((llBits+mlBits)>24)) BIT_flushBits(bitC: &blockStream);
354 BIT_addBits(bitC: &blockStream, value: sequences[n].mlBase, nbBits: mlBits);
355 if (MEM_32bits() || (ofBits+mlBits+llBits > 56)) BIT_flushBits(bitC: &blockStream);
356 if (longOffsets) {
357 unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1);
358 if (extraBits) {
359 BIT_addBits(bitC: &blockStream, value: sequences[n].offBase, nbBits: extraBits);
360 BIT_flushBits(bitC: &blockStream); /* (7)*/
361 }
362 BIT_addBits(bitC: &blockStream, value: sequences[n].offBase >> extraBits,
363 nbBits: ofBits - extraBits); /* 31 */
364 } else {
365 BIT_addBits(bitC: &blockStream, value: sequences[n].offBase, nbBits: ofBits); /* 31 */
366 }
367 BIT_flushBits(bitC: &blockStream); /* (7)*/
368 DEBUGLOG(7, "remaining space : %i", (int)(blockStream.endPtr - blockStream.ptr));
369 } }
370
371 DEBUGLOG(6, "ZSTD_encodeSequences: flushing ML state with %u bits", stateMatchLength.stateLog);
372 FSE_flushCState(bitC: &blockStream, statePtr: &stateMatchLength);
373 DEBUGLOG(6, "ZSTD_encodeSequences: flushing Off state with %u bits", stateOffsetBits.stateLog);
374 FSE_flushCState(bitC: &blockStream, statePtr: &stateOffsetBits);
375 DEBUGLOG(6, "ZSTD_encodeSequences: flushing LL state with %u bits", stateLitLength.stateLog);
376 FSE_flushCState(bitC: &blockStream, statePtr: &stateLitLength);
377
378 { size_t const streamSize = BIT_closeCStream(bitC: &blockStream);
379 RETURN_ERROR_IF(streamSize==0, dstSize_tooSmall, "not enough space");
380 return streamSize;
381 }
382}
383
384static size_t
385ZSTD_encodeSequences_default(
386 void* dst, size_t dstCapacity,
387 FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
388 FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
389 FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
390 seqDef const* sequences, size_t nbSeq, int longOffsets)
391{
392 return ZSTD_encodeSequences_body(dst, dstCapacity,
393 CTable_MatchLength, mlCodeTable,
394 CTable_OffsetBits, ofCodeTable,
395 CTable_LitLength, llCodeTable,
396 sequences, nbSeq, longOffsets);
397}
398
399
400#if DYNAMIC_BMI2
401
402static BMI2_TARGET_ATTRIBUTE size_t
403ZSTD_encodeSequences_bmi2(
404 void* dst, size_t dstCapacity,
405 FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
406 FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
407 FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
408 seqDef const* sequences, size_t nbSeq, int longOffsets)
409{
410 return ZSTD_encodeSequences_body(dst, dstCapacity,
411 CTable_MatchLength, mlCodeTable,
412 CTable_OffsetBits, ofCodeTable,
413 CTable_LitLength, llCodeTable,
414 sequences, nbSeq, longOffsets);
415}
416
417#endif
418
419size_t ZSTD_encodeSequences(
420 void* dst, size_t dstCapacity,
421 FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
422 FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
423 FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
424 seqDef const* sequences, size_t nbSeq, int longOffsets, int bmi2)
425{
426 DEBUGLOG(5, "ZSTD_encodeSequences: dstCapacity = %u", (unsigned)dstCapacity);
427#if DYNAMIC_BMI2
428 if (bmi2) {
429 return ZSTD_encodeSequences_bmi2(dst, dstCapacity,
430 CTable_MatchLength, mlCodeTable,
431 CTable_OffsetBits, ofCodeTable,
432 CTable_LitLength, llCodeTable,
433 sequences, nbSeq, longOffsets);
434 }
435#endif
436 (void)bmi2;
437 return ZSTD_encodeSequences_default(dst, dstCapacity,
438 CTable_MatchLength, mlCodeTable,
439 CTable_OffsetBits, ofCodeTable,
440 CTable_LitLength, llCodeTable,
441 sequences, nbSeq, longOffsets);
442}
443

source code of linux/lib/zstd/compress/zstd_compress_sequences.c