/src/zstd/lib/dictBuilder/fastcover.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * Copyright (c) Meta Platforms, Inc. and affiliates. |
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 <stdio.h> /* fprintf */ |
15 | | #include <stdlib.h> /* malloc, free, qsort */ |
16 | | #include <string.h> /* memset */ |
17 | | #include <time.h> /* clock */ |
18 | | |
19 | | #ifndef ZDICT_STATIC_LINKING_ONLY |
20 | | # define ZDICT_STATIC_LINKING_ONLY |
21 | | #endif |
22 | | |
23 | | #include "../common/mem.h" /* read */ |
24 | | #include "../common/pool.h" |
25 | | #include "../common/threading.h" |
26 | | #include "../common/zstd_internal.h" /* includes zstd.h */ |
27 | | #include "../compress/zstd_compress_internal.h" /* ZSTD_hash*() */ |
28 | | #include "../zdict.h" |
29 | | #include "cover.h" |
30 | | |
31 | | |
32 | | /*-************************************* |
33 | | * Constants |
34 | | ***************************************/ |
35 | | /** |
36 | | * There are 32bit indexes used to ref samples, so limit samples size to 4GB |
37 | | * on 64bit builds. |
38 | | * For 32bit builds we choose 1 GB. |
39 | | * Most 32bit platforms have 2GB user-mode addressable space and we allocate a large |
40 | | * contiguous buffer, so 1GB is already a high limit. |
41 | | */ |
42 | 0 | #define FASTCOVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((unsigned)-1) : ((unsigned)1 GB)) |
43 | 0 | #define FASTCOVER_MAX_F 31 |
44 | 0 | #define FASTCOVER_MAX_ACCEL 10 |
45 | 0 | #define FASTCOVER_DEFAULT_SPLITPOINT 0.75 |
46 | 0 | #define DEFAULT_F 20 |
47 | 0 | #define DEFAULT_ACCEL 1 |
48 | | |
49 | | |
50 | | /*-************************************* |
51 | | * Console display |
52 | | * |
53 | | * Captures the `displayLevel` variable in the local scope. |
54 | | ***************************************/ |
55 | | #undef DISPLAY |
56 | | #define DISPLAY(...) \ |
57 | 0 | { \ |
58 | 0 | fprintf(stderr, __VA_ARGS__); \ |
59 | 0 | fflush(stderr); \ |
60 | 0 | } |
61 | | #undef DISPLAYLEVEL |
62 | | #define DISPLAYLEVEL(l, ...) \ |
63 | 0 | if (displayLevel >= l) { \ |
64 | 0 | DISPLAY(__VA_ARGS__); \ |
65 | 0 | } /* 0 : no display; 1: errors; 2: default; 3: details; 4: debug */ |
66 | | |
67 | | #undef DISPLAYUPDATE |
68 | | #define DISPLAYUPDATE(lastUpdateTime, l, ...) \ |
69 | 0 | if (displayLevel >= l) { \ |
70 | 0 | const clock_t refreshRate = CLOCKS_PER_SEC * 15 / 100; \ |
71 | 0 | if ((clock() - lastUpdateTime > refreshRate) || (displayLevel >= 4)) { \ |
72 | 0 | lastUpdateTime = clock(); \ |
73 | 0 | DISPLAY(__VA_ARGS__); \ |
74 | 0 | } \ |
75 | 0 | } |
76 | | |
77 | | |
78 | | /*-************************************* |
79 | | * Hash Functions |
80 | | ***************************************/ |
81 | | /** |
82 | | * Hash the d-byte value pointed to by p and mod 2^f into the frequency vector |
83 | | */ |
84 | 0 | static size_t FASTCOVER_hashPtrToIndex(const void* p, U32 f, unsigned d) { |
85 | 0 | if (d == 6) { |
86 | 0 | return ZSTD_hash6Ptr(p, f); |
87 | 0 | } |
88 | 0 | return ZSTD_hash8Ptr(p, f); |
89 | 0 | } |
90 | | |
91 | | |
92 | | /*-************************************* |
93 | | * Acceleration |
94 | | ***************************************/ |
95 | | typedef struct { |
96 | | unsigned finalize; /* Percentage of training samples used for ZDICT_finalizeDictionary */ |
97 | | unsigned skip; /* Number of dmer skipped between each dmer counted in computeFrequency */ |
98 | | } FASTCOVER_accel_t; |
99 | | |
100 | | |
101 | | static const FASTCOVER_accel_t FASTCOVER_defaultAccelParameters[FASTCOVER_MAX_ACCEL+1] = { |
102 | | { 100, 0 }, /* accel = 0, should not happen because accel = 0 defaults to accel = 1 */ |
103 | | { 100, 0 }, /* accel = 1 */ |
104 | | { 50, 1 }, /* accel = 2 */ |
105 | | { 34, 2 }, /* accel = 3 */ |
106 | | { 25, 3 }, /* accel = 4 */ |
107 | | { 20, 4 }, /* accel = 5 */ |
108 | | { 17, 5 }, /* accel = 6 */ |
109 | | { 14, 6 }, /* accel = 7 */ |
110 | | { 13, 7 }, /* accel = 8 */ |
111 | | { 11, 8 }, /* accel = 9 */ |
112 | | { 10, 9 }, /* accel = 10 */ |
113 | | }; |
114 | | |
115 | | |
116 | | /*-************************************* |
117 | | * Context |
118 | | ***************************************/ |
119 | | typedef struct { |
120 | | const BYTE *samples; |
121 | | size_t *offsets; |
122 | | const size_t *samplesSizes; |
123 | | size_t nbSamples; |
124 | | size_t nbTrainSamples; |
125 | | size_t nbTestSamples; |
126 | | size_t nbDmers; |
127 | | U32 *freqs; |
128 | | unsigned d; |
129 | | unsigned f; |
130 | | FASTCOVER_accel_t accelParams; |
131 | | int displayLevel; |
132 | | } FASTCOVER_ctx_t; |
133 | | |
134 | | |
135 | | /*-************************************* |
136 | | * Helper functions |
137 | | ***************************************/ |
138 | | /** |
139 | | * Selects the best segment in an epoch. |
140 | | * Segments of are scored according to the function: |
141 | | * |
142 | | * Let F(d) be the frequency of all dmers with hash value d. |
143 | | * Let S_i be hash value of the dmer at position i of segment S which has length k. |
144 | | * |
145 | | * Score(S) = F(S_1) + F(S_2) + ... + F(S_{k-d+1}) |
146 | | * |
147 | | * Once the dmer with hash value d is in the dictionary we set F(d) = 0. |
148 | | */ |
149 | | static COVER_segment_t FASTCOVER_selectSegment(const FASTCOVER_ctx_t *ctx, |
150 | | U32 *freqs, U32 begin, U32 end, |
151 | | ZDICT_cover_params_t parameters, |
152 | 0 | U16* segmentFreqs) { |
153 | | /* Constants */ |
154 | 0 | const U32 k = parameters.k; |
155 | 0 | const U32 d = parameters.d; |
156 | 0 | const U32 f = ctx->f; |
157 | 0 | const U32 dmersInK = k - d + 1; |
158 | | |
159 | | /* Try each segment (activeSegment) and save the best (bestSegment) */ |
160 | 0 | COVER_segment_t bestSegment = {0, 0, 0}; |
161 | 0 | COVER_segment_t activeSegment; |
162 | | |
163 | | /* Reset the activeDmers in the segment */ |
164 | | /* The activeSegment starts at the beginning of the epoch. */ |
165 | 0 | activeSegment.begin = begin; |
166 | 0 | activeSegment.end = begin; |
167 | 0 | activeSegment.score = 0; |
168 | | |
169 | | /* Slide the activeSegment through the whole epoch. |
170 | | * Save the best segment in bestSegment. |
171 | | */ |
172 | 0 | while (activeSegment.end < end) { |
173 | | /* Get hash value of current dmer */ |
174 | 0 | const size_t idx = FASTCOVER_hashPtrToIndex(ctx->samples + activeSegment.end, f, d); |
175 | | |
176 | | /* Add frequency of this index to score if this is the first occurrence of index in active segment */ |
177 | 0 | if (segmentFreqs[idx] == 0) { |
178 | 0 | activeSegment.score += freqs[idx]; |
179 | 0 | } |
180 | | /* Increment end of segment and segmentFreqs*/ |
181 | 0 | activeSegment.end += 1; |
182 | 0 | segmentFreqs[idx] += 1; |
183 | | /* If the window is now too large, drop the first position */ |
184 | 0 | if (activeSegment.end - activeSegment.begin == dmersInK + 1) { |
185 | | /* Get hash value of the dmer to be eliminated from active segment */ |
186 | 0 | const size_t delIndex = FASTCOVER_hashPtrToIndex(ctx->samples + activeSegment.begin, f, d); |
187 | 0 | segmentFreqs[delIndex] -= 1; |
188 | | /* Subtract frequency of this index from score if this is the last occurrence of this index in active segment */ |
189 | 0 | if (segmentFreqs[delIndex] == 0) { |
190 | 0 | activeSegment.score -= freqs[delIndex]; |
191 | 0 | } |
192 | | /* Increment start of segment */ |
193 | 0 | activeSegment.begin += 1; |
194 | 0 | } |
195 | | |
196 | | /* If this segment is the best so far save it */ |
197 | 0 | if (activeSegment.score > bestSegment.score) { |
198 | 0 | bestSegment = activeSegment; |
199 | 0 | } |
200 | 0 | } |
201 | | |
202 | | /* Zero out rest of segmentFreqs array */ |
203 | 0 | while (activeSegment.begin < end) { |
204 | 0 | const size_t delIndex = FASTCOVER_hashPtrToIndex(ctx->samples + activeSegment.begin, f, d); |
205 | 0 | segmentFreqs[delIndex] -= 1; |
206 | 0 | activeSegment.begin += 1; |
207 | 0 | } |
208 | |
|
209 | 0 | { |
210 | | /* Zero the frequency of hash value of each dmer covered by the chosen segment. */ |
211 | 0 | U32 pos; |
212 | 0 | for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) { |
213 | 0 | const size_t i = FASTCOVER_hashPtrToIndex(ctx->samples + pos, f, d); |
214 | 0 | freqs[i] = 0; |
215 | 0 | } |
216 | 0 | } |
217 | |
|
218 | 0 | return bestSegment; |
219 | 0 | } |
220 | | |
221 | | |
222 | | static int FASTCOVER_checkParameters(ZDICT_cover_params_t parameters, |
223 | | size_t maxDictSize, unsigned f, |
224 | 0 | unsigned accel) { |
225 | | /* k, d, and f are required parameters */ |
226 | 0 | if (parameters.d == 0 || parameters.k == 0) { |
227 | 0 | return 0; |
228 | 0 | } |
229 | | /* d has to be 6 or 8 */ |
230 | 0 | if (parameters.d != 6 && parameters.d != 8) { |
231 | 0 | return 0; |
232 | 0 | } |
233 | | /* k <= maxDictSize */ |
234 | 0 | if (parameters.k > maxDictSize) { |
235 | 0 | return 0; |
236 | 0 | } |
237 | | /* d <= k */ |
238 | 0 | if (parameters.d > parameters.k) { |
239 | 0 | return 0; |
240 | 0 | } |
241 | | /* 0 < f <= FASTCOVER_MAX_F*/ |
242 | 0 | if (f > FASTCOVER_MAX_F || f == 0) { |
243 | 0 | return 0; |
244 | 0 | } |
245 | | /* 0 < splitPoint <= 1 */ |
246 | 0 | if (parameters.splitPoint <= 0 || parameters.splitPoint > 1) { |
247 | 0 | return 0; |
248 | 0 | } |
249 | | /* 0 < accel <= 10 */ |
250 | 0 | if (accel > 10 || accel == 0) { |
251 | 0 | return 0; |
252 | 0 | } |
253 | 0 | return 1; |
254 | 0 | } |
255 | | |
256 | | |
257 | | /** |
258 | | * Clean up a context initialized with `FASTCOVER_ctx_init()`. |
259 | | */ |
260 | | static void |
261 | | FASTCOVER_ctx_destroy(FASTCOVER_ctx_t* ctx) |
262 | 0 | { |
263 | 0 | if (!ctx) return; |
264 | | |
265 | 0 | free(ctx->freqs); |
266 | 0 | ctx->freqs = NULL; |
267 | |
|
268 | 0 | free(ctx->offsets); |
269 | 0 | ctx->offsets = NULL; |
270 | 0 | } |
271 | | |
272 | | |
273 | | /** |
274 | | * Calculate for frequency of hash value of each dmer in ctx->samples |
275 | | */ |
276 | | static void |
277 | | FASTCOVER_computeFrequency(U32* freqs, const FASTCOVER_ctx_t* ctx) |
278 | 0 | { |
279 | 0 | const unsigned f = ctx->f; |
280 | 0 | const unsigned d = ctx->d; |
281 | 0 | const unsigned skip = ctx->accelParams.skip; |
282 | 0 | const unsigned readLength = MAX(d, 8); |
283 | 0 | size_t i; |
284 | 0 | assert(ctx->nbTrainSamples >= 5); |
285 | 0 | assert(ctx->nbTrainSamples <= ctx->nbSamples); |
286 | 0 | for (i = 0; i < ctx->nbTrainSamples; i++) { |
287 | 0 | size_t start = ctx->offsets[i]; /* start of current dmer */ |
288 | 0 | size_t const currSampleEnd = ctx->offsets[i+1]; |
289 | 0 | while (start + readLength <= currSampleEnd) { |
290 | 0 | const size_t dmerIndex = FASTCOVER_hashPtrToIndex(ctx->samples + start, f, d); |
291 | 0 | freqs[dmerIndex]++; |
292 | 0 | start = start + skip + 1; |
293 | 0 | } |
294 | 0 | } |
295 | 0 | } |
296 | | |
297 | | |
298 | | /** |
299 | | * Prepare a context for dictionary building. |
300 | | * The context is only dependent on the parameter `d` and can be used multiple |
301 | | * times. |
302 | | * Returns 0 on success or error code on error. |
303 | | * The context must be destroyed with `FASTCOVER_ctx_destroy()`. |
304 | | */ |
305 | | static size_t |
306 | | FASTCOVER_ctx_init(FASTCOVER_ctx_t* ctx, |
307 | | const void* samplesBuffer, |
308 | | const size_t* samplesSizes, unsigned nbSamples, |
309 | | unsigned d, double splitPoint, unsigned f, |
310 | | FASTCOVER_accel_t accelParams, |
311 | | int displayLevel) |
312 | 0 | { |
313 | 0 | const BYTE* const samples = (const BYTE*)samplesBuffer; |
314 | 0 | const size_t totalSamplesSize = COVER_sum(samplesSizes, nbSamples); |
315 | | /* Split samples into testing and training sets */ |
316 | 0 | const unsigned nbTrainSamples = splitPoint < 1.0 ? (unsigned)((double)nbSamples * splitPoint) : nbSamples; |
317 | 0 | const unsigned nbTestSamples = splitPoint < 1.0 ? nbSamples - nbTrainSamples : nbSamples; |
318 | 0 | const size_t trainingSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes, nbTrainSamples) : totalSamplesSize; |
319 | 0 | const size_t testSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes + nbTrainSamples, nbTestSamples) : totalSamplesSize; |
320 | 0 | ctx->displayLevel = displayLevel; |
321 | | |
322 | | /* Checks */ |
323 | 0 | if (totalSamplesSize < MAX(d, sizeof(U64)) || |
324 | 0 | totalSamplesSize >= (size_t)FASTCOVER_MAX_SAMPLES_SIZE) { |
325 | 0 | DISPLAYLEVEL(1, "Total samples size is too large (%u MB), maximum size is %u MB\n", |
326 | 0 | (unsigned)(totalSamplesSize >> 20), (FASTCOVER_MAX_SAMPLES_SIZE >> 20)); |
327 | 0 | return ERROR(srcSize_wrong); |
328 | 0 | } |
329 | | |
330 | | /* Check if there are at least 5 training samples */ |
331 | 0 | if (nbTrainSamples < 5) { |
332 | 0 | DISPLAYLEVEL(1, "Total number of training samples is %u and is invalid\n", nbTrainSamples); |
333 | 0 | return ERROR(srcSize_wrong); |
334 | 0 | } |
335 | | |
336 | | /* Check if there's testing sample */ |
337 | 0 | if (nbTestSamples < 1) { |
338 | 0 | DISPLAYLEVEL(1, "Total number of testing samples is %u and is invalid.\n", nbTestSamples); |
339 | 0 | return ERROR(srcSize_wrong); |
340 | 0 | } |
341 | | |
342 | | /* Zero the context */ |
343 | 0 | memset(ctx, 0, sizeof(*ctx)); |
344 | 0 | DISPLAYLEVEL(2, "Training on %u samples of total size %u\n", nbTrainSamples, |
345 | 0 | (unsigned)trainingSamplesSize); |
346 | 0 | DISPLAYLEVEL(2, "Testing on %u samples of total size %u\n", nbTestSamples, |
347 | 0 | (unsigned)testSamplesSize); |
348 | |
|
349 | 0 | ctx->samples = samples; |
350 | 0 | ctx->samplesSizes = samplesSizes; |
351 | 0 | ctx->nbSamples = nbSamples; |
352 | 0 | ctx->nbTrainSamples = nbTrainSamples; |
353 | 0 | ctx->nbTestSamples = nbTestSamples; |
354 | 0 | ctx->nbDmers = trainingSamplesSize - MAX(d, sizeof(U64)) + 1; |
355 | 0 | ctx->d = d; |
356 | 0 | ctx->f = f; |
357 | 0 | ctx->accelParams = accelParams; |
358 | | |
359 | | /* The offsets of each file */ |
360 | 0 | ctx->offsets = (size_t*)calloc((nbSamples + 1), sizeof(size_t)); |
361 | 0 | if (ctx->offsets == NULL) { |
362 | 0 | DISPLAYLEVEL(1, "Failed to allocate scratch buffers \n"); |
363 | 0 | FASTCOVER_ctx_destroy(ctx); |
364 | 0 | return ERROR(memory_allocation); |
365 | 0 | } |
366 | | |
367 | | /* Fill offsets from the samplesSizes */ |
368 | 0 | { U32 i; |
369 | 0 | ctx->offsets[0] = 0; |
370 | 0 | assert(nbSamples >= 5); |
371 | 0 | for (i = 1; i <= nbSamples; ++i) { |
372 | 0 | ctx->offsets[i] = ctx->offsets[i - 1] + samplesSizes[i - 1]; |
373 | 0 | } |
374 | 0 | } |
375 | | |
376 | | /* Initialize frequency array of size 2^f */ |
377 | 0 | ctx->freqs = (U32*)calloc(((U64)1 << f), sizeof(U32)); |
378 | 0 | if (ctx->freqs == NULL) { |
379 | 0 | DISPLAYLEVEL(1, "Failed to allocate frequency table \n"); |
380 | 0 | FASTCOVER_ctx_destroy(ctx); |
381 | 0 | return ERROR(memory_allocation); |
382 | 0 | } |
383 | | |
384 | 0 | DISPLAYLEVEL(2, "Computing frequencies\n"); |
385 | 0 | FASTCOVER_computeFrequency(ctx->freqs, ctx); |
386 | |
|
387 | 0 | return 0; |
388 | 0 | } |
389 | | |
390 | | |
391 | | /** |
392 | | * Given the prepared context build the dictionary. |
393 | | */ |
394 | | static size_t |
395 | | FASTCOVER_buildDictionary(const FASTCOVER_ctx_t* ctx, |
396 | | U32* freqs, |
397 | | void* dictBuffer, size_t dictBufferCapacity, |
398 | | ZDICT_cover_params_t parameters, |
399 | | U16* segmentFreqs) |
400 | 0 | { |
401 | 0 | BYTE *const dict = (BYTE *)dictBuffer; |
402 | 0 | size_t tail = dictBufferCapacity; |
403 | | /* Divide the data into epochs. We will select one segment from each epoch. */ |
404 | 0 | const COVER_epoch_info_t epochs = COVER_computeEpochs( |
405 | 0 | (U32)dictBufferCapacity, (U32)ctx->nbDmers, parameters.k, 1); |
406 | 0 | const size_t maxZeroScoreRun = 10; |
407 | 0 | const int displayLevel = ctx->displayLevel; |
408 | 0 | size_t zeroScoreRun = 0; |
409 | 0 | clock_t lastUpdateTime = 0; |
410 | 0 | size_t epoch; |
411 | 0 | DISPLAYLEVEL(2, "Breaking content into %u epochs of size %u\n", |
412 | 0 | (U32)epochs.num, (U32)epochs.size); |
413 | | /* Loop through the epochs until there are no more segments or the dictionary |
414 | | * is full. |
415 | | */ |
416 | 0 | for (epoch = 0; tail > 0; epoch = (epoch + 1) % epochs.num) { |
417 | 0 | const U32 epochBegin = (U32)(epoch * epochs.size); |
418 | 0 | const U32 epochEnd = epochBegin + epochs.size; |
419 | 0 | size_t segmentSize; |
420 | | /* Select a segment */ |
421 | 0 | COVER_segment_t segment = FASTCOVER_selectSegment( |
422 | 0 | ctx, freqs, epochBegin, epochEnd, parameters, segmentFreqs); |
423 | | |
424 | | /* If the segment covers no dmers, then we are out of content. |
425 | | * There may be new content in other epochs, for continue for some time. |
426 | | */ |
427 | 0 | if (segment.score == 0) { |
428 | 0 | if (++zeroScoreRun >= maxZeroScoreRun) { |
429 | 0 | break; |
430 | 0 | } |
431 | 0 | continue; |
432 | 0 | } |
433 | 0 | zeroScoreRun = 0; |
434 | | |
435 | | /* Trim the segment if necessary and if it is too small then we are done */ |
436 | 0 | segmentSize = MIN(segment.end - segment.begin + parameters.d - 1, tail); |
437 | 0 | if (segmentSize < parameters.d) { |
438 | 0 | break; |
439 | 0 | } |
440 | | |
441 | | /* We fill the dictionary from the back to allow the best segments to be |
442 | | * referenced with the smallest offsets. |
443 | | */ |
444 | 0 | tail -= segmentSize; |
445 | 0 | memcpy(dict + tail, ctx->samples + segment.begin, segmentSize); |
446 | 0 | DISPLAYUPDATE( |
447 | 0 | lastUpdateTime, |
448 | 0 | 2, "\r%u%% ", |
449 | 0 | (unsigned)(((dictBufferCapacity - tail) * 100) / dictBufferCapacity)); |
450 | 0 | } |
451 | 0 | DISPLAYLEVEL(2, "\r%79s\r", ""); |
452 | 0 | return tail; |
453 | 0 | } |
454 | | |
455 | | /** |
456 | | * Parameters for FASTCOVER_tryParameters(). |
457 | | */ |
458 | | typedef struct FASTCOVER_tryParameters_data_s { |
459 | | const FASTCOVER_ctx_t* ctx; |
460 | | COVER_best_t* best; |
461 | | size_t dictBufferCapacity; |
462 | | ZDICT_cover_params_t parameters; |
463 | | } FASTCOVER_tryParameters_data_t; |
464 | | |
465 | | |
466 | | /** |
467 | | * Tries a set of parameters and updates the COVER_best_t with the results. |
468 | | * This function is thread safe if zstd is compiled with multithreaded support. |
469 | | * It takes its parameters as an *OWNING* opaque pointer to support threading. |
470 | | */ |
471 | | static void FASTCOVER_tryParameters(void* opaque) |
472 | 0 | { |
473 | | /* Save parameters as local variables */ |
474 | 0 | FASTCOVER_tryParameters_data_t *const data = (FASTCOVER_tryParameters_data_t*)opaque; |
475 | 0 | const FASTCOVER_ctx_t *const ctx = data->ctx; |
476 | 0 | const ZDICT_cover_params_t parameters = data->parameters; |
477 | 0 | size_t dictBufferCapacity = data->dictBufferCapacity; |
478 | 0 | size_t totalCompressedSize = ERROR(GENERIC); |
479 | | /* Initialize array to keep track of frequency of dmer within activeSegment */ |
480 | 0 | U16* segmentFreqs = (U16*)calloc(((U64)1 << ctx->f), sizeof(U16)); |
481 | | /* Allocate space for hash table, dict, and freqs */ |
482 | 0 | BYTE *const dict = (BYTE*)malloc(dictBufferCapacity); |
483 | 0 | COVER_dictSelection_t selection = COVER_dictSelectionError(ERROR(GENERIC)); |
484 | 0 | U32* freqs = (U32*) malloc(((U64)1 << ctx->f) * sizeof(U32)); |
485 | 0 | const int displayLevel = ctx->displayLevel; |
486 | 0 | if (!segmentFreqs || !dict || !freqs) { |
487 | 0 | DISPLAYLEVEL(1, "Failed to allocate buffers: out of memory\n"); |
488 | 0 | goto _cleanup; |
489 | 0 | } |
490 | | /* Copy the frequencies because we need to modify them */ |
491 | 0 | memcpy(freqs, ctx->freqs, ((U64)1 << ctx->f) * sizeof(U32)); |
492 | | /* Build the dictionary */ |
493 | 0 | { const size_t tail = FASTCOVER_buildDictionary(ctx, freqs, dict, dictBufferCapacity, |
494 | 0 | parameters, segmentFreqs); |
495 | |
|
496 | 0 | const unsigned nbFinalizeSamples = (unsigned)(ctx->nbTrainSamples * ctx->accelParams.finalize / 100); |
497 | 0 | selection = COVER_selectDict(dict + tail, dictBufferCapacity, dictBufferCapacity - tail, |
498 | 0 | ctx->samples, ctx->samplesSizes, nbFinalizeSamples, ctx->nbTrainSamples, ctx->nbSamples, parameters, ctx->offsets, |
499 | 0 | totalCompressedSize); |
500 | |
|
501 | 0 | if (COVER_dictSelectionIsError(selection)) { |
502 | 0 | DISPLAYLEVEL(1, "Failed to select dictionary\n"); |
503 | 0 | goto _cleanup; |
504 | 0 | } |
505 | 0 | } |
506 | 0 | _cleanup: |
507 | 0 | free(dict); |
508 | 0 | COVER_best_finish(data->best, parameters, selection); |
509 | 0 | free(data); |
510 | 0 | free(segmentFreqs); |
511 | 0 | COVER_dictSelectionFree(selection); |
512 | 0 | free(freqs); |
513 | 0 | } |
514 | | |
515 | | |
516 | | static void |
517 | | FASTCOVER_convertToCoverParams(ZDICT_fastCover_params_t fastCoverParams, |
518 | | ZDICT_cover_params_t* coverParams) |
519 | 0 | { |
520 | 0 | coverParams->k = fastCoverParams.k; |
521 | 0 | coverParams->d = fastCoverParams.d; |
522 | 0 | coverParams->steps = fastCoverParams.steps; |
523 | 0 | coverParams->nbThreads = fastCoverParams.nbThreads; |
524 | 0 | coverParams->splitPoint = fastCoverParams.splitPoint; |
525 | 0 | coverParams->zParams = fastCoverParams.zParams; |
526 | 0 | coverParams->shrinkDict = fastCoverParams.shrinkDict; |
527 | 0 | } |
528 | | |
529 | | |
530 | | static void |
531 | | FASTCOVER_convertToFastCoverParams(ZDICT_cover_params_t coverParams, |
532 | | ZDICT_fastCover_params_t* fastCoverParams, |
533 | | unsigned f, unsigned accel) |
534 | 0 | { |
535 | 0 | fastCoverParams->k = coverParams.k; |
536 | 0 | fastCoverParams->d = coverParams.d; |
537 | 0 | fastCoverParams->steps = coverParams.steps; |
538 | 0 | fastCoverParams->nbThreads = coverParams.nbThreads; |
539 | 0 | fastCoverParams->splitPoint = coverParams.splitPoint; |
540 | 0 | fastCoverParams->f = f; |
541 | 0 | fastCoverParams->accel = accel; |
542 | 0 | fastCoverParams->zParams = coverParams.zParams; |
543 | 0 | fastCoverParams->shrinkDict = coverParams.shrinkDict; |
544 | 0 | } |
545 | | |
546 | | |
547 | | ZDICTLIB_STATIC_API size_t |
548 | | ZDICT_trainFromBuffer_fastCover(void* dictBuffer, size_t dictBufferCapacity, |
549 | | const void* samplesBuffer, |
550 | | const size_t* samplesSizes, unsigned nbSamples, |
551 | | ZDICT_fastCover_params_t parameters) |
552 | 0 | { |
553 | 0 | BYTE* const dict = (BYTE*)dictBuffer; |
554 | 0 | FASTCOVER_ctx_t ctx; |
555 | 0 | ZDICT_cover_params_t coverParams; |
556 | 0 | FASTCOVER_accel_t accelParams; |
557 | 0 | const int displayLevel = (int)parameters.zParams.notificationLevel; |
558 | | /* Assign splitPoint and f if not provided */ |
559 | 0 | parameters.splitPoint = 1.0; |
560 | 0 | parameters.f = parameters.f == 0 ? DEFAULT_F : parameters.f; |
561 | 0 | parameters.accel = parameters.accel == 0 ? DEFAULT_ACCEL : parameters.accel; |
562 | | /* Convert to cover parameter */ |
563 | 0 | memset(&coverParams, 0 , sizeof(coverParams)); |
564 | 0 | FASTCOVER_convertToCoverParams(parameters, &coverParams); |
565 | | /* Checks */ |
566 | 0 | if (!FASTCOVER_checkParameters(coverParams, dictBufferCapacity, parameters.f, |
567 | 0 | parameters.accel)) { |
568 | 0 | DISPLAYLEVEL(1, "FASTCOVER parameters incorrect\n"); |
569 | 0 | return ERROR(parameter_outOfBound); |
570 | 0 | } |
571 | 0 | if (nbSamples == 0) { |
572 | 0 | DISPLAYLEVEL(1, "FASTCOVER must have at least one input file\n"); |
573 | 0 | return ERROR(srcSize_wrong); |
574 | 0 | } |
575 | 0 | if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) { |
576 | 0 | DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n", |
577 | 0 | ZDICT_DICTSIZE_MIN); |
578 | 0 | return ERROR(dstSize_tooSmall); |
579 | 0 | } |
580 | | /* Assign corresponding FASTCOVER_accel_t to accelParams*/ |
581 | 0 | accelParams = FASTCOVER_defaultAccelParameters[parameters.accel]; |
582 | | /* Initialize context */ |
583 | 0 | { |
584 | 0 | size_t const initVal = FASTCOVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, |
585 | 0 | coverParams.d, parameters.splitPoint, parameters.f, |
586 | 0 | accelParams, displayLevel); |
587 | 0 | if (ZSTD_isError(initVal)) { |
588 | 0 | DISPLAYLEVEL(1, "Failed to initialize context\n"); |
589 | 0 | return initVal; |
590 | 0 | } |
591 | 0 | } |
592 | 0 | COVER_warnOnSmallCorpus(dictBufferCapacity, ctx.nbDmers, displayLevel); |
593 | | /* Build the dictionary */ |
594 | 0 | DISPLAYLEVEL(2, "Building dictionary\n"); |
595 | 0 | { |
596 | | /* Initialize array to keep track of frequency of dmer within activeSegment */ |
597 | 0 | U16* segmentFreqs = (U16 *)calloc(((U64)1 << parameters.f), sizeof(U16)); |
598 | 0 | const size_t tail = FASTCOVER_buildDictionary(&ctx, ctx.freqs, dictBuffer, |
599 | 0 | dictBufferCapacity, coverParams, segmentFreqs); |
600 | 0 | const unsigned nbFinalizeSamples = (unsigned)(ctx.nbTrainSamples * ctx.accelParams.finalize / 100); |
601 | 0 | const size_t dictionarySize = ZDICT_finalizeDictionary( |
602 | 0 | dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail, |
603 | 0 | samplesBuffer, samplesSizes, nbFinalizeSamples, coverParams.zParams); |
604 | 0 | if (!ZSTD_isError(dictionarySize)) { |
605 | 0 | DISPLAYLEVEL(2, "Constructed dictionary of size %u\n", |
606 | 0 | (unsigned)dictionarySize); |
607 | 0 | } |
608 | 0 | FASTCOVER_ctx_destroy(&ctx); |
609 | 0 | free(segmentFreqs); |
610 | 0 | return dictionarySize; |
611 | 0 | } |
612 | 0 | } |
613 | | |
614 | | |
615 | | ZDICTLIB_STATIC_API size_t |
616 | | ZDICT_optimizeTrainFromBuffer_fastCover( |
617 | | void* dictBuffer, size_t dictBufferCapacity, |
618 | | const void* samplesBuffer, |
619 | | const size_t* samplesSizes, unsigned nbSamples, |
620 | | ZDICT_fastCover_params_t* parameters) |
621 | 0 | { |
622 | 0 | ZDICT_cover_params_t coverParams; |
623 | 0 | FASTCOVER_accel_t accelParams; |
624 | | /* constants */ |
625 | 0 | const unsigned nbThreads = parameters->nbThreads; |
626 | 0 | const double splitPoint = |
627 | 0 | parameters->splitPoint <= 0.0 ? FASTCOVER_DEFAULT_SPLITPOINT : parameters->splitPoint; |
628 | 0 | const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d; |
629 | 0 | const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d; |
630 | 0 | const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k; |
631 | 0 | const unsigned kMaxK = parameters->k == 0 ? 2000 : parameters->k; |
632 | 0 | const unsigned kSteps = parameters->steps == 0 ? 40 : parameters->steps; |
633 | 0 | const unsigned kStepSize = MAX((kMaxK - kMinK) / kSteps, 1); |
634 | 0 | const unsigned kIterations = |
635 | 0 | (1 + (kMaxD - kMinD) / 2) * (1 + (kMaxK - kMinK) / kStepSize); |
636 | 0 | const unsigned f = parameters->f == 0 ? DEFAULT_F : parameters->f; |
637 | 0 | const unsigned accel = parameters->accel == 0 ? DEFAULT_ACCEL : parameters->accel; |
638 | 0 | const unsigned shrinkDict = 0; |
639 | | /* Local variables */ |
640 | 0 | const int displayLevel = (int)parameters->zParams.notificationLevel; |
641 | 0 | unsigned iteration = 1; |
642 | 0 | unsigned d; |
643 | 0 | unsigned k; |
644 | 0 | COVER_best_t best; |
645 | 0 | POOL_ctx *pool = NULL; |
646 | 0 | int warned = 0; |
647 | 0 | clock_t lastUpdateTime = 0; |
648 | | /* Checks */ |
649 | 0 | if (splitPoint <= 0 || splitPoint > 1) { |
650 | 0 | DISPLAYLEVEL(1, "Incorrect splitPoint\n"); |
651 | 0 | return ERROR(parameter_outOfBound); |
652 | 0 | } |
653 | 0 | if (accel == 0 || accel > FASTCOVER_MAX_ACCEL) { |
654 | 0 | DISPLAYLEVEL(1, "Incorrect accel\n"); |
655 | 0 | return ERROR(parameter_outOfBound); |
656 | 0 | } |
657 | 0 | if (kMinK < kMaxD || kMaxK < kMinK) { |
658 | 0 | DISPLAYLEVEL(1, "Incorrect k\n"); |
659 | 0 | return ERROR(parameter_outOfBound); |
660 | 0 | } |
661 | 0 | if (nbSamples == 0) { |
662 | 0 | DISPLAYLEVEL(1, "FASTCOVER must have at least one input file\n"); |
663 | 0 | return ERROR(srcSize_wrong); |
664 | 0 | } |
665 | 0 | if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) { |
666 | 0 | DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n", |
667 | 0 | ZDICT_DICTSIZE_MIN); |
668 | 0 | return ERROR(dstSize_tooSmall); |
669 | 0 | } |
670 | 0 | if (nbThreads > 1) { |
671 | 0 | pool = POOL_create(nbThreads, 1); |
672 | 0 | if (!pool) { |
673 | 0 | return ERROR(memory_allocation); |
674 | 0 | } |
675 | 0 | } |
676 | | /* Initialization */ |
677 | 0 | COVER_best_init(&best); |
678 | 0 | memset(&coverParams, 0 , sizeof(coverParams)); |
679 | 0 | FASTCOVER_convertToCoverParams(*parameters, &coverParams); |
680 | 0 | accelParams = FASTCOVER_defaultAccelParameters[accel]; |
681 | | /* Loop through d first because each new value needs a new context */ |
682 | 0 | DISPLAYLEVEL(2, "Trying %u different sets of parameters\n", kIterations); |
683 | 0 | for (d = kMinD; d <= kMaxD; d += 2) { |
684 | | /* Initialize the context for this value of d */ |
685 | 0 | FASTCOVER_ctx_t ctx; |
686 | 0 | DISPLAYLEVEL(3, "d=%u\n", d); |
687 | 0 | { |
688 | | /* Turn down global display level to clean up display at level 2 and below */ |
689 | 0 | const int childDisplayLevel = displayLevel == 0 ? 0 : displayLevel - 1; |
690 | 0 | size_t const initVal = FASTCOVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, d, splitPoint, f, accelParams, childDisplayLevel); |
691 | 0 | if (ZSTD_isError(initVal)) { |
692 | 0 | DISPLAYLEVEL(1, "Failed to initialize context\n"); |
693 | 0 | COVER_best_destroy(&best); |
694 | 0 | POOL_free(pool); |
695 | 0 | return initVal; |
696 | 0 | } |
697 | 0 | } |
698 | 0 | if (!warned) { |
699 | 0 | COVER_warnOnSmallCorpus(dictBufferCapacity, ctx.nbDmers, displayLevel); |
700 | 0 | warned = 1; |
701 | 0 | } |
702 | | /* Loop through k reusing the same context */ |
703 | 0 | for (k = kMinK; k <= kMaxK; k += kStepSize) { |
704 | | /* Prepare the arguments */ |
705 | 0 | FASTCOVER_tryParameters_data_t *data = (FASTCOVER_tryParameters_data_t *)malloc( |
706 | 0 | sizeof(FASTCOVER_tryParameters_data_t)); |
707 | 0 | DISPLAYLEVEL(3, "k=%u\n", k); |
708 | 0 | if (!data) { |
709 | 0 | DISPLAYLEVEL(1, "Failed to allocate parameters\n"); |
710 | 0 | COVER_best_destroy(&best); |
711 | 0 | FASTCOVER_ctx_destroy(&ctx); |
712 | 0 | POOL_free(pool); |
713 | 0 | return ERROR(memory_allocation); |
714 | 0 | } |
715 | 0 | data->ctx = &ctx; |
716 | 0 | data->best = &best; |
717 | 0 | data->dictBufferCapacity = dictBufferCapacity; |
718 | 0 | data->parameters = coverParams; |
719 | 0 | data->parameters.k = k; |
720 | 0 | data->parameters.d = d; |
721 | 0 | data->parameters.splitPoint = splitPoint; |
722 | 0 | data->parameters.steps = kSteps; |
723 | 0 | data->parameters.shrinkDict = shrinkDict; |
724 | 0 | data->parameters.zParams.notificationLevel = (unsigned)ctx.displayLevel; |
725 | | /* Check the parameters */ |
726 | 0 | if (!FASTCOVER_checkParameters(data->parameters, dictBufferCapacity, |
727 | 0 | data->ctx->f, accel)) { |
728 | 0 | DISPLAYLEVEL(1, "FASTCOVER parameters incorrect\n"); |
729 | 0 | free(data); |
730 | 0 | continue; |
731 | 0 | } |
732 | | /* Call the function and pass ownership of data to it */ |
733 | 0 | COVER_best_start(&best); |
734 | 0 | if (pool) { |
735 | 0 | POOL_add(pool, &FASTCOVER_tryParameters, data); |
736 | 0 | } else { |
737 | 0 | FASTCOVER_tryParameters(data); |
738 | 0 | } |
739 | | /* Print status */ |
740 | 0 | DISPLAYUPDATE(lastUpdateTime, |
741 | 0 | 2, "\r%u%% ", |
742 | 0 | (unsigned)((iteration * 100) / kIterations)); |
743 | 0 | ++iteration; |
744 | 0 | } |
745 | 0 | COVER_best_wait(&best); |
746 | 0 | FASTCOVER_ctx_destroy(&ctx); |
747 | 0 | } |
748 | 0 | DISPLAYLEVEL(2, "\r%79s\r", ""); |
749 | | /* Fill the output buffer and parameters with output of the best parameters */ |
750 | 0 | { |
751 | 0 | const size_t dictSize = best.dictSize; |
752 | 0 | if (ZSTD_isError(best.compressedSize)) { |
753 | 0 | const size_t compressedSize = best.compressedSize; |
754 | 0 | COVER_best_destroy(&best); |
755 | 0 | POOL_free(pool); |
756 | 0 | return compressedSize; |
757 | 0 | } |
758 | 0 | FASTCOVER_convertToFastCoverParams(best.parameters, parameters, f, accel); |
759 | 0 | memcpy(dictBuffer, best.dict, dictSize); |
760 | 0 | COVER_best_destroy(&best); |
761 | 0 | POOL_free(pool); |
762 | 0 | return dictSize; |
763 | 0 | } |
764 | |
|
765 | 0 | } |