1 /* 2 * Copyright (c) 2016-2020, 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 #ifndef DICTBUILDER_H_001 12 #define DICTBUILDER_H_001 13 14 #if defined (__cplusplus) 15 extern "C" { 16 #endif 17 18 19 /*====== Dependencies ======*/ 20 #include <stddef.h> /* size_t */ 21 22 23 /* ===== ZDICTLIB_API : control library symbols visibility ===== */ 24 #ifndef ZDICTLIB_VISIBILITY 25 # if defined(__GNUC__) && (__GNUC__ >= 4) 26 # define ZDICTLIB_VISIBILITY __attribute__ ((visibility ("default"))) 27 # else 28 # define ZDICTLIB_VISIBILITY 29 # endif 30 #endif 31 #if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1) 32 # define ZDICTLIB_API __declspec(dllexport) ZDICTLIB_VISIBILITY 33 #elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1) 34 # define ZDICTLIB_API __declspec(dllimport) ZDICTLIB_VISIBILITY /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/ 35 #else 36 # define ZDICTLIB_API ZDICTLIB_VISIBILITY 37 #endif 38 39 40 /*! ZDICT_trainFromBuffer(): 41 * Train a dictionary from an array of samples. 42 * Redirect towards ZDICT_optimizeTrainFromBuffer_fastCover() single-threaded, with d=8, steps=4, 43 * f=20, and accel=1. 44 * Samples must be stored concatenated in a single flat buffer `samplesBuffer`, 45 * supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order. 46 * The resulting dictionary will be saved into `dictBuffer`. 47 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`) 48 * or an error code, which can be tested with ZDICT_isError(). 49 * Note: Dictionary training will fail if there are not enough samples to construct a 50 * dictionary, or if most of the samples are too small (< 8 bytes being the lower limit). 51 * If dictionary training fails, you should use zstd without a dictionary, as the dictionary 52 * would've been ineffective anyways. If you believe your samples would benefit from a dictionary 53 * please open an issue with details, and we can look into it. 54 * Note: ZDICT_trainFromBuffer()'s memory usage is about 6 MB. 55 * Tips: In general, a reasonable dictionary has a size of ~ 100 KB. 56 * It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`. 57 * In general, it's recommended to provide a few thousands samples, though this can vary a lot. 58 * It's recommended that total size of all samples be about ~x100 times the target size of dictionary. 59 */ 60 ZDICTLIB_API size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCapacity, 61 const void* samplesBuffer, 62 const size_t* samplesSizes, unsigned nbSamples); 63 64 typedef struct { 65 int compressionLevel; /*< optimize for a specific zstd compression level; 0 means default */ 66 unsigned notificationLevel; /*< Write log to stderr; 0 = none (default); 1 = errors; 2 = progression; 3 = details; 4 = debug; */ 67 unsigned dictID; /*< force dictID value; 0 means auto mode (32-bits random value) */ 68 } ZDICT_params_t; 69 70 /*! ZDICT_finalizeDictionary(): 71 * Given a custom content as a basis for dictionary, and a set of samples, 72 * finalize dictionary by adding headers and statistics according to the zstd 73 * dictionary format. 74 * 75 * Samples must be stored concatenated in a flat buffer `samplesBuffer`, 76 * supplied with an array of sizes `samplesSizes`, providing the size of each 77 * sample in order. The samples are used to construct the statistics, so they 78 * should be representative of what you will compress with this dictionary. 79 * 80 * The compression level can be set in `parameters`. You should pass the 81 * compression level you expect to use in production. The statistics for each 82 * compression level differ, so tuning the dictionary for the compression level 83 * can help quite a bit. 84 * 85 * You can set an explicit dictionary ID in `parameters`, or allow us to pick 86 * a random dictionary ID for you, but we can't guarantee no collisions. 87 * 88 * The dstDictBuffer and the dictContent may overlap, and the content will be 89 * appended to the end of the header. If the header + the content doesn't fit in 90 * maxDictSize the beginning of the content is truncated to make room, since it 91 * is presumed that the most profitable content is at the end of the dictionary, 92 * since that is the cheapest to reference. 93 * 94 * `dictContentSize` must be >= ZDICT_CONTENTSIZE_MIN bytes. 95 * `maxDictSize` must be >= max(dictContentSize, ZSTD_DICTSIZE_MIN). 96 * 97 * @return: size of dictionary stored into `dstDictBuffer` (<= `maxDictSize`), 98 * or an error code, which can be tested by ZDICT_isError(). 99 * Note: ZDICT_finalizeDictionary() will push notifications into stderr if 100 * instructed to, using notificationLevel>0. 101 * NOTE: This function currently may fail in several edge cases including: 102 * * Not enough samples 103 * * Samples are uncompressible 104 * * Samples are all exactly the same 105 */ 106 ZDICTLIB_API size_t ZDICT_finalizeDictionary(void* dstDictBuffer, size_t maxDictSize, 107 const void* dictContent, size_t dictContentSize, 108 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples, 109 ZDICT_params_t parameters); 110 111 112 /*====== Helper functions ======*/ 113 ZDICTLIB_API unsigned ZDICT_getDictID(const void* dictBuffer, size_t dictSize); /**< extracts dictID; @return zero if error (not a valid dictionary) */ 114 ZDICTLIB_API size_t ZDICT_getDictHeaderSize(const void* dictBuffer, size_t dictSize); /* returns dict header size; returns a ZSTD error code on failure */ 115 ZDICTLIB_API unsigned ZDICT_isError(size_t errorCode); 116 ZDICTLIB_API const char* ZDICT_getErrorName(size_t errorCode); 117 118 119 120 #ifdef ZDICT_STATIC_LINKING_ONLY 121 122 /* ==================================================================================== 123 * The definitions in this section are considered experimental. 124 * They should never be used with a dynamic library, as they may change in the future. 125 * They are provided for advanced usages. 126 * Use them only in association with static linking. 127 * ==================================================================================== */ 128 129 #define ZDICT_CONTENTSIZE_MIN 128 130 #define ZDICT_DICTSIZE_MIN 256 131 132 /*! ZDICT_cover_params_t: 133 * k and d are the only required parameters. 134 * For others, value 0 means default. 135 */ 136 typedef struct { 137 unsigned k; /* Segment size : constraint: 0 < k : Reasonable range [16, 2048+] */ 138 unsigned d; /* dmer size : constraint: 0 < d <= k : Reasonable range [6, 16] */ 139 unsigned steps; /* Number of steps : Only used for optimization : 0 means default (40) : Higher means more parameters checked */ 140 unsigned nbThreads; /* Number of threads : constraint: 0 < nbThreads : 1 means single-threaded : Only used for optimization : Ignored if ZSTD_MULTITHREAD is not defined */ 141 double splitPoint; /* Percentage of samples used for training: Only used for optimization : the first nbSamples * splitPoint samples will be used to training, the last nbSamples * (1 - splitPoint) samples will be used for testing, 0 means default (1.0), 1.0 when all samples are used for both training and testing */ 142 unsigned shrinkDict; /* Train dictionaries to shrink in size starting from the minimum size and selects the smallest dictionary that is shrinkDictMaxRegression% worse than the largest dictionary. 0 means no shrinking and 1 means shrinking */ 143 unsigned shrinkDictMaxRegression; /* Sets shrinkDictMaxRegression so that a smaller dictionary can be at worse shrinkDictMaxRegression% worse than the max dict size dictionary. */ 144 ZDICT_params_t zParams; 145 } ZDICT_cover_params_t; 146 147 typedef struct { 148 unsigned k; /* Segment size : constraint: 0 < k : Reasonable range [16, 2048+] */ 149 unsigned d; /* dmer size : constraint: 0 < d <= k : Reasonable range [6, 16] */ 150 unsigned f; /* log of size of frequency array : constraint: 0 < f <= 31 : 1 means default(20)*/ 151 unsigned steps; /* Number of steps : Only used for optimization : 0 means default (40) : Higher means more parameters checked */ 152 unsigned nbThreads; /* Number of threads : constraint: 0 < nbThreads : 1 means single-threaded : Only used for optimization : Ignored if ZSTD_MULTITHREAD is not defined */ 153 double splitPoint; /* Percentage of samples used for training: Only used for optimization : the first nbSamples * splitPoint samples will be used to training, the last nbSamples * (1 - splitPoint) samples will be used for testing, 0 means default (0.75), 1.0 when all samples are used for both training and testing */ 154 unsigned accel; /* Acceleration level: constraint: 0 < accel <= 10, higher means faster and less accurate, 0 means default(1) */ 155 unsigned shrinkDict; /* Train dictionaries to shrink in size starting from the minimum size and selects the smallest dictionary that is shrinkDictMaxRegression% worse than the largest dictionary. 0 means no shrinking and 1 means shrinking */ 156 unsigned shrinkDictMaxRegression; /* Sets shrinkDictMaxRegression so that a smaller dictionary can be at worse shrinkDictMaxRegression% worse than the max dict size dictionary. */ 157 158 ZDICT_params_t zParams; 159 } ZDICT_fastCover_params_t; 160 161 /*! ZDICT_trainFromBuffer_cover(): 162 * Train a dictionary from an array of samples using the COVER algorithm. 163 * Samples must be stored concatenated in a single flat buffer `samplesBuffer`, 164 * supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order. 165 * The resulting dictionary will be saved into `dictBuffer`. 166 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`) 167 * or an error code, which can be tested with ZDICT_isError(). 168 * See ZDICT_trainFromBuffer() for details on failure modes. 169 * Note: ZDICT_trainFromBuffer_cover() requires about 9 bytes of memory for each input byte. 170 * Tips: In general, a reasonable dictionary has a size of ~ 100 KB. 171 * It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`. 172 * In general, it's recommended to provide a few thousands samples, though this can vary a lot. 173 * It's recommended that total size of all samples be about ~x100 times the target size of dictionary. 174 */ 175 ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover( 176 void *dictBuffer, size_t dictBufferCapacity, 177 const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples, 178 ZDICT_cover_params_t parameters); 179 180 /*! ZDICT_optimizeTrainFromBuffer_cover(): 181 * The same requirements as above hold for all the parameters except `parameters`. 182 * This function tries many parameter combinations and picks the best parameters. 183 * `*parameters` is filled with the best parameters found, 184 * dictionary constructed with those parameters is stored in `dictBuffer`. 185 * 186 * All of the parameters d, k, steps are optional. 187 * If d is non-zero then we don't check multiple values of d, otherwise we check d = {6, 8}. 188 * if steps is zero it defaults to its default value. 189 * If k is non-zero then we don't check multiple values of k, otherwise we check steps values in [50, 2000]. 190 * 191 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`) 192 * or an error code, which can be tested with ZDICT_isError(). 193 * On success `*parameters` contains the parameters selected. 194 * See ZDICT_trainFromBuffer() for details on failure modes. 195 * Note: ZDICT_optimizeTrainFromBuffer_cover() requires about 8 bytes of memory for each input byte and additionally another 5 bytes of memory for each byte of memory for each thread. 196 */ 197 ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover( 198 void* dictBuffer, size_t dictBufferCapacity, 199 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples, 200 ZDICT_cover_params_t* parameters); 201 202 /*! ZDICT_trainFromBuffer_fastCover(): 203 * Train a dictionary from an array of samples using a modified version of COVER algorithm. 204 * Samples must be stored concatenated in a single flat buffer `samplesBuffer`, 205 * supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order. 206 * d and k are required. 207 * All other parameters are optional, will use default values if not provided 208 * The resulting dictionary will be saved into `dictBuffer`. 209 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`) 210 * or an error code, which can be tested with ZDICT_isError(). 211 * See ZDICT_trainFromBuffer() for details on failure modes. 212 * Note: ZDICT_trainFromBuffer_fastCover() requires 6 * 2^f bytes of memory. 213 * Tips: In general, a reasonable dictionary has a size of ~ 100 KB. 214 * It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`. 215 * In general, it's recommended to provide a few thousands samples, though this can vary a lot. 216 * It's recommended that total size of all samples be about ~x100 times the target size of dictionary. 217 */ 218 ZDICTLIB_API size_t ZDICT_trainFromBuffer_fastCover(void *dictBuffer, 219 size_t dictBufferCapacity, const void *samplesBuffer, 220 const size_t *samplesSizes, unsigned nbSamples, 221 ZDICT_fastCover_params_t parameters); 222 223 /*! ZDICT_optimizeTrainFromBuffer_fastCover(): 224 * The same requirements as above hold for all the parameters except `parameters`. 225 * This function tries many parameter combinations (specifically, k and d combinations) 226 * and picks the best parameters. `*parameters` is filled with the best parameters found, 227 * dictionary constructed with those parameters is stored in `dictBuffer`. 228 * All of the parameters d, k, steps, f, and accel are optional. 229 * If d is non-zero then we don't check multiple values of d, otherwise we check d = {6, 8}. 230 * if steps is zero it defaults to its default value. 231 * If k is non-zero then we don't check multiple values of k, otherwise we check steps values in [50, 2000]. 232 * If f is zero, default value of 20 is used. 233 * If accel is zero, default value of 1 is used. 234 * 235 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`) 236 * or an error code, which can be tested with ZDICT_isError(). 237 * On success `*parameters` contains the parameters selected. 238 * See ZDICT_trainFromBuffer() for details on failure modes. 239 * Note: ZDICT_optimizeTrainFromBuffer_fastCover() requires about 6 * 2^f bytes of memory for each thread. 240 */ 241 ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_fastCover(void* dictBuffer, 242 size_t dictBufferCapacity, const void* samplesBuffer, 243 const size_t* samplesSizes, unsigned nbSamples, 244 ZDICT_fastCover_params_t* parameters); 245 246 typedef struct { 247 unsigned selectivityLevel; /* 0 means default; larger => select more => larger dictionary */ 248 ZDICT_params_t zParams; 249 } ZDICT_legacy_params_t; 250 251 /*! ZDICT_trainFromBuffer_legacy(): 252 * Train a dictionary from an array of samples. 253 * Samples must be stored concatenated in a single flat buffer `samplesBuffer`, 254 * supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order. 255 * The resulting dictionary will be saved into `dictBuffer`. 256 * `parameters` is optional and can be provided with values set to 0 to mean "default". 257 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`) 258 * or an error code, which can be tested with ZDICT_isError(). 259 * See ZDICT_trainFromBuffer() for details on failure modes. 260 * Tips: In general, a reasonable dictionary has a size of ~ 100 KB. 261 * It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`. 262 * In general, it's recommended to provide a few thousands samples, though this can vary a lot. 263 * It's recommended that total size of all samples be about ~x100 times the target size of dictionary. 264 * Note: ZDICT_trainFromBuffer_legacy() will send notifications into stderr if instructed to, using notificationLevel>0. 265 */ 266 ZDICTLIB_API size_t ZDICT_trainFromBuffer_legacy( 267 void *dictBuffer, size_t dictBufferCapacity, 268 const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples, 269 ZDICT_legacy_params_t parameters); 270 271 /* Deprecation warnings */ 272 /* It is generally possible to disable deprecation warnings from compiler, 273 for example with -Wno-deprecated-declarations for gcc 274 or _CRT_SECURE_NO_WARNINGS in Visual. 275 Otherwise, it's also possible to manually define ZDICT_DISABLE_DEPRECATE_WARNINGS */ 276 #ifdef ZDICT_DISABLE_DEPRECATE_WARNINGS 277 # define ZDICT_DEPRECATED(message) ZDICTLIB_API /* disable deprecation warnings */ 278 #else 279 # define ZDICT_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__) 280 # if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */ 281 # define ZDICT_DEPRECATED(message) [[deprecated(message)]] ZDICTLIB_API 282 # elif defined(__clang__) || (ZDICT_GCC_VERSION >= 405) 283 # define ZDICT_DEPRECATED(message) ZDICTLIB_API __attribute__((deprecated(message))) 284 # elif (ZDICT_GCC_VERSION >= 301) 285 # define ZDICT_DEPRECATED(message) ZDICTLIB_API __attribute__((deprecated)) 286 # elif defined(_MSC_VER) 287 # define ZDICT_DEPRECATED(message) ZDICTLIB_API __declspec(deprecated(message)) 288 # else 289 # pragma message("WARNING: You need to implement ZDICT_DEPRECATED for this compiler") 290 # define ZDICT_DEPRECATED(message) ZDICTLIB_API 291 # endif 292 #endif /* ZDICT_DISABLE_DEPRECATE_WARNINGS */ 293 294 ZDICT_DEPRECATED("use ZDICT_finalizeDictionary() instead") 295 size_t ZDICT_addEntropyTablesFromBuffer(void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity, 296 const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples); 297 298 299 #endif /* ZDICT_STATIC_LINKING_ONLY */ 300 301 #if defined (__cplusplus) 302 } 303 #endif 304 305 #endif /* DICTBUILDER_H_001 */ 306