1 | /* |
2 | * Copyright (c) 2016-present, 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 | * Constructs a dictionary using a heuristic based on the following paper: |
13 | * |
14 | * Liao, Petri, Moffat, Wirth |
15 | * Effective Construction of Relative Lempel-Ziv Dictionaries |
16 | * Published in WWW 2016. |
17 | * |
18 | * Adapted from code originally written by @ot (Giuseppe Ottaviano). |
19 | ******************************************************************************/ |
20 | |
21 | /*-************************************* |
22 | * Dependencies |
23 | ***************************************/ |
24 | #include <stdio.h> /* fprintf */ |
25 | #include <stdlib.h> /* malloc, free, qsort */ |
26 | #include <string.h> /* memset */ |
27 | #include <time.h> /* clock */ |
28 | |
29 | #include "mem.h" /* read */ |
30 | #include "pool.h" |
31 | #include "threading.h" |
32 | #include "zstd_internal.h" /* includes zstd.h */ |
33 | #ifndef ZDICT_STATIC_LINKING_ONLY |
34 | #define ZDICT_STATIC_LINKING_ONLY |
35 | #endif |
36 | #include "zdict.h" |
37 | |
38 | /*-************************************* |
39 | * Constants |
40 | ***************************************/ |
41 | #define COVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((U32)-1) : ((U32)1 GB)) |
42 | |
43 | /*-************************************* |
44 | * Console display |
45 | ***************************************/ |
46 | static int g_displayLevel = 2; |
47 | #define DISPLAY(...) \ |
48 | { \ |
49 | fprintf(stderr, __VA_ARGS__); \ |
50 | fflush(stderr); \ |
51 | } |
52 | #define LOCALDISPLAYLEVEL(displayLevel, l, ...) \ |
53 | if (displayLevel >= l) { \ |
54 | DISPLAY(__VA_ARGS__); \ |
55 | } /* 0 : no display; 1: errors; 2: default; 3: details; 4: debug */ |
56 | #define DISPLAYLEVEL(l, ...) LOCALDISPLAYLEVEL(g_displayLevel, l, __VA_ARGS__) |
57 | |
58 | #define LOCALDISPLAYUPDATE(displayLevel, l, ...) \ |
59 | if (displayLevel >= l) { \ |
60 | if ((clock() - g_time > refreshRate) || (displayLevel >= 4)) { \ |
61 | g_time = clock(); \ |
62 | DISPLAY(__VA_ARGS__); \ |
63 | } \ |
64 | } |
65 | #define DISPLAYUPDATE(l, ...) LOCALDISPLAYUPDATE(g_displayLevel, l, __VA_ARGS__) |
66 | static const clock_t refreshRate = CLOCKS_PER_SEC * 15 / 100; |
67 | static clock_t g_time = 0; |
68 | |
69 | /*-************************************* |
70 | * Hash table |
71 | *************************************** |
72 | * A small specialized hash map for storing activeDmers. |
73 | * The map does not resize, so if it becomes full it will loop forever. |
74 | * Thus, the map must be large enough to store every value. |
75 | * The map implements linear probing and keeps its load less than 0.5. |
76 | */ |
77 | |
78 | #define MAP_EMPTY_VALUE ((U32)-1) |
79 | typedef struct COVER_map_pair_t_s { |
80 | U32 key; |
81 | U32 value; |
82 | } COVER_map_pair_t; |
83 | |
84 | typedef struct COVER_map_s { |
85 | COVER_map_pair_t *data; |
86 | U32 sizeLog; |
87 | U32 size; |
88 | U32 sizeMask; |
89 | } COVER_map_t; |
90 | |
91 | /** |
92 | * Clear the map. |
93 | */ |
94 | static void COVER_map_clear(COVER_map_t *map) { |
95 | memset(map->data, MAP_EMPTY_VALUE, map->size * sizeof(COVER_map_pair_t)); |
96 | } |
97 | |
98 | /** |
99 | * Initializes a map of the given size. |
100 | * Returns 1 on success and 0 on failure. |
101 | * The map must be destroyed with COVER_map_destroy(). |
102 | * The map is only guaranteed to be large enough to hold size elements. |
103 | */ |
104 | static int COVER_map_init(COVER_map_t *map, U32 size) { |
105 | map->sizeLog = ZSTD_highbit32(size) + 2; |
106 | map->size = (U32)1 << map->sizeLog; |
107 | map->sizeMask = map->size - 1; |
108 | map->data = (COVER_map_pair_t *)malloc(map->size * sizeof(COVER_map_pair_t)); |
109 | if (!map->data) { |
110 | map->sizeLog = 0; |
111 | map->size = 0; |
112 | return 0; |
113 | } |
114 | COVER_map_clear(map); |
115 | return 1; |
116 | } |
117 | |
118 | /** |
119 | * Internal hash function |
120 | */ |
121 | static const U32 prime4bytes = 2654435761U; |
122 | static U32 COVER_map_hash(COVER_map_t *map, U32 key) { |
123 | return (key * prime4bytes) >> (32 - map->sizeLog); |
124 | } |
125 | |
126 | /** |
127 | * Helper function that returns the index that a key should be placed into. |
128 | */ |
129 | static U32 COVER_map_index(COVER_map_t *map, U32 key) { |
130 | const U32 hash = COVER_map_hash(map, key); |
131 | U32 i; |
132 | for (i = hash;; i = (i + 1) & map->sizeMask) { |
133 | COVER_map_pair_t *pos = &map->data[i]; |
134 | if (pos->value == MAP_EMPTY_VALUE) { |
135 | return i; |
136 | } |
137 | if (pos->key == key) { |
138 | return i; |
139 | } |
140 | } |
141 | } |
142 | |
143 | /** |
144 | * Returns the pointer to the value for key. |
145 | * If key is not in the map, it is inserted and the value is set to 0. |
146 | * The map must not be full. |
147 | */ |
148 | static U32 *COVER_map_at(COVER_map_t *map, U32 key) { |
149 | COVER_map_pair_t *pos = &map->data[COVER_map_index(map, key)]; |
150 | if (pos->value == MAP_EMPTY_VALUE) { |
151 | pos->key = key; |
152 | pos->value = 0; |
153 | } |
154 | return &pos->value; |
155 | } |
156 | |
157 | /** |
158 | * Deletes key from the map if present. |
159 | */ |
160 | static void COVER_map_remove(COVER_map_t *map, U32 key) { |
161 | U32 i = COVER_map_index(map, key); |
162 | COVER_map_pair_t *del = &map->data[i]; |
163 | U32 shift = 1; |
164 | if (del->value == MAP_EMPTY_VALUE) { |
165 | return; |
166 | } |
167 | for (i = (i + 1) & map->sizeMask;; i = (i + 1) & map->sizeMask) { |
168 | COVER_map_pair_t *const pos = &map->data[i]; |
169 | /* If the position is empty we are done */ |
170 | if (pos->value == MAP_EMPTY_VALUE) { |
171 | del->value = MAP_EMPTY_VALUE; |
172 | return; |
173 | } |
174 | /* If pos can be moved to del do so */ |
175 | if (((i - COVER_map_hash(map, pos->key)) & map->sizeMask) >= shift) { |
176 | del->key = pos->key; |
177 | del->value = pos->value; |
178 | del = pos; |
179 | shift = 1; |
180 | } else { |
181 | ++shift; |
182 | } |
183 | } |
184 | } |
185 | |
186 | /** |
187 | * Destroyes a map that is inited with COVER_map_init(). |
188 | */ |
189 | static void COVER_map_destroy(COVER_map_t *map) { |
190 | if (map->data) { |
191 | free(map->data); |
192 | } |
193 | map->data = NULL; |
194 | map->size = 0; |
195 | } |
196 | |
197 | /*-************************************* |
198 | * Context |
199 | ***************************************/ |
200 | |
201 | typedef struct { |
202 | const BYTE *samples; |
203 | size_t *offsets; |
204 | const size_t *samplesSizes; |
205 | size_t nbSamples; |
206 | U32 *suffix; |
207 | size_t suffixSize; |
208 | U32 *freqs; |
209 | U32 *dmerAt; |
210 | unsigned d; |
211 | } COVER_ctx_t; |
212 | |
213 | /* We need a global context for qsort... */ |
214 | static COVER_ctx_t *g_ctx = NULL; |
215 | |
216 | /*-************************************* |
217 | * Helper functions |
218 | ***************************************/ |
219 | |
220 | /** |
221 | * Returns the sum of the sample sizes. |
222 | */ |
223 | static size_t COVER_sum(const size_t *samplesSizes, unsigned nbSamples) { |
224 | size_t sum = 0; |
225 | size_t i; |
226 | for (i = 0; i < nbSamples; ++i) { |
227 | sum += samplesSizes[i]; |
228 | } |
229 | return sum; |
230 | } |
231 | |
232 | /** |
233 | * Returns -1 if the dmer at lp is less than the dmer at rp. |
234 | * Return 0 if the dmers at lp and rp are equal. |
235 | * Returns 1 if the dmer at lp is greater than the dmer at rp. |
236 | */ |
237 | static int COVER_cmp(COVER_ctx_t *ctx, const void *lp, const void *rp) { |
238 | U32 const lhs = *(U32 const *)lp; |
239 | U32 const rhs = *(U32 const *)rp; |
240 | return memcmp(ctx->samples + lhs, ctx->samples + rhs, ctx->d); |
241 | } |
242 | /** |
243 | * Faster version for d <= 8. |
244 | */ |
245 | static int COVER_cmp8(COVER_ctx_t *ctx, const void *lp, const void *rp) { |
246 | U64 const mask = (ctx->d == 8) ? (U64)-1 : (((U64)1 << (8 * ctx->d)) - 1); |
247 | U64 const lhs = MEM_readLE64(ctx->samples + *(U32 const *)lp) & mask; |
248 | U64 const rhs = MEM_readLE64(ctx->samples + *(U32 const *)rp) & mask; |
249 | if (lhs < rhs) { |
250 | return -1; |
251 | } |
252 | return (lhs > rhs); |
253 | } |
254 | |
255 | /** |
256 | * Same as COVER_cmp() except ties are broken by pointer value |
257 | * NOTE: g_ctx must be set to call this function. A global is required because |
258 | * qsort doesn't take an opaque pointer. |
259 | */ |
260 | static int COVER_strict_cmp(const void *lp, const void *rp) { |
261 | int result = COVER_cmp(g_ctx, lp, rp); |
262 | if (result == 0) { |
263 | result = lp < rp ? -1 : 1; |
264 | } |
265 | return result; |
266 | } |
267 | /** |
268 | * Faster version for d <= 8. |
269 | */ |
270 | static int COVER_strict_cmp8(const void *lp, const void *rp) { |
271 | int result = COVER_cmp8(g_ctx, lp, rp); |
272 | if (result == 0) { |
273 | result = lp < rp ? -1 : 1; |
274 | } |
275 | return result; |
276 | } |
277 | |
278 | /** |
279 | * Returns the first pointer in [first, last) whose element does not compare |
280 | * less than value. If no such element exists it returns last. |
281 | */ |
282 | static const size_t *COVER_lower_bound(const size_t *first, const size_t *last, |
283 | size_t value) { |
284 | size_t count = last - first; |
285 | while (count != 0) { |
286 | size_t step = count / 2; |
287 | const size_t *ptr = first; |
288 | ptr += step; |
289 | if (*ptr < value) { |
290 | first = ++ptr; |
291 | count -= step + 1; |
292 | } else { |
293 | count = step; |
294 | } |
295 | } |
296 | return first; |
297 | } |
298 | |
299 | /** |
300 | * Generic groupBy function. |
301 | * Groups an array sorted by cmp into groups with equivalent values. |
302 | * Calls grp for each group. |
303 | */ |
304 | static void |
305 | COVER_groupBy(const void *data, size_t count, size_t size, COVER_ctx_t *ctx, |
306 | int (*cmp)(COVER_ctx_t *, const void *, const void *), |
307 | void (*grp)(COVER_ctx_t *, const void *, const void *)) { |
308 | const BYTE *ptr = (const BYTE *)data; |
309 | size_t num = 0; |
310 | while (num < count) { |
311 | const BYTE *grpEnd = ptr + size; |
312 | ++num; |
313 | while (num < count && cmp(ctx, ptr, grpEnd) == 0) { |
314 | grpEnd += size; |
315 | ++num; |
316 | } |
317 | grp(ctx, ptr, grpEnd); |
318 | ptr = grpEnd; |
319 | } |
320 | } |
321 | |
322 | /*-************************************* |
323 | * Cover functions |
324 | ***************************************/ |
325 | |
326 | /** |
327 | * Called on each group of positions with the same dmer. |
328 | * Counts the frequency of each dmer and saves it in the suffix array. |
329 | * Fills `ctx->dmerAt`. |
330 | */ |
331 | static void COVER_group(COVER_ctx_t *ctx, const void *group, |
332 | const void *groupEnd) { |
333 | /* The group consists of all the positions with the same first d bytes. */ |
334 | const U32 *grpPtr = (const U32 *)group; |
335 | const U32 *grpEnd = (const U32 *)groupEnd; |
336 | /* The dmerId is how we will reference this dmer. |
337 | * This allows us to map the whole dmer space to a much smaller space, the |
338 | * size of the suffix array. |
339 | */ |
340 | const U32 dmerId = (U32)(grpPtr - ctx->suffix); |
341 | /* Count the number of samples this dmer shows up in */ |
342 | U32 freq = 0; |
343 | /* Details */ |
344 | const size_t *curOffsetPtr = ctx->offsets; |
345 | const size_t *offsetsEnd = ctx->offsets + ctx->nbSamples; |
346 | /* Once *grpPtr >= curSampleEnd this occurrence of the dmer is in a |
347 | * different sample than the last. |
348 | */ |
349 | size_t curSampleEnd = ctx->offsets[0]; |
350 | for (; grpPtr != grpEnd; ++grpPtr) { |
351 | /* Save the dmerId for this position so we can get back to it. */ |
352 | ctx->dmerAt[*grpPtr] = dmerId; |
353 | /* Dictionaries only help for the first reference to the dmer. |
354 | * After that zstd can reference the match from the previous reference. |
355 | * So only count each dmer once for each sample it is in. |
356 | */ |
357 | if (*grpPtr < curSampleEnd) { |
358 | continue; |
359 | } |
360 | freq += 1; |
361 | /* Binary search to find the end of the sample *grpPtr is in. |
362 | * In the common case that grpPtr + 1 == grpEnd we can skip the binary |
363 | * search because the loop is over. |
364 | */ |
365 | if (grpPtr + 1 != grpEnd) { |
366 | const size_t *sampleEndPtr = |
367 | COVER_lower_bound(curOffsetPtr, offsetsEnd, *grpPtr); |
368 | curSampleEnd = *sampleEndPtr; |
369 | curOffsetPtr = sampleEndPtr + 1; |
370 | } |
371 | } |
372 | /* At this point we are never going to look at this segment of the suffix |
373 | * array again. We take advantage of this fact to save memory. |
374 | * We store the frequency of the dmer in the first position of the group, |
375 | * which is dmerId. |
376 | */ |
377 | ctx->suffix[dmerId] = freq; |
378 | } |
379 | |
380 | /** |
381 | * A segment is a range in the source as well as the score of the segment. |
382 | */ |
383 | typedef struct { |
384 | U32 begin; |
385 | U32 end; |
386 | U32 score; |
387 | } COVER_segment_t; |
388 | |
389 | /** |
390 | * Selects the best segment in an epoch. |
391 | * Segments of are scored according to the function: |
392 | * |
393 | * Let F(d) be the frequency of dmer d. |
394 | * Let S_i be the dmer at position i of segment S which has length k. |
395 | * |
396 | * Score(S) = F(S_1) + F(S_2) + ... + F(S_{k-d+1}) |
397 | * |
398 | * Once the dmer d is in the dictionay we set F(d) = 0. |
399 | */ |
400 | static COVER_segment_t COVER_selectSegment(const COVER_ctx_t *ctx, U32 *freqs, |
401 | COVER_map_t *activeDmers, U32 begin, |
402 | U32 end, |
403 | ZDICT_cover_params_t parameters) { |
404 | /* Constants */ |
405 | const U32 k = parameters.k; |
406 | const U32 d = parameters.d; |
407 | const U32 dmersInK = k - d + 1; |
408 | /* Try each segment (activeSegment) and save the best (bestSegment) */ |
409 | COVER_segment_t bestSegment = {0, 0, 0}; |
410 | COVER_segment_t activeSegment; |
411 | /* Reset the activeDmers in the segment */ |
412 | COVER_map_clear(activeDmers); |
413 | /* The activeSegment starts at the beginning of the epoch. */ |
414 | activeSegment.begin = begin; |
415 | activeSegment.end = begin; |
416 | activeSegment.score = 0; |
417 | /* Slide the activeSegment through the whole epoch. |
418 | * Save the best segment in bestSegment. |
419 | */ |
420 | while (activeSegment.end < end) { |
421 | /* The dmerId for the dmer at the next position */ |
422 | U32 newDmer = ctx->dmerAt[activeSegment.end]; |
423 | /* The entry in activeDmers for this dmerId */ |
424 | U32 *newDmerOcc = COVER_map_at(activeDmers, newDmer); |
425 | /* If the dmer isn't already present in the segment add its score. */ |
426 | if (*newDmerOcc == 0) { |
427 | /* The paper suggest using the L-0.5 norm, but experiments show that it |
428 | * doesn't help. |
429 | */ |
430 | activeSegment.score += freqs[newDmer]; |
431 | } |
432 | /* Add the dmer to the segment */ |
433 | activeSegment.end += 1; |
434 | *newDmerOcc += 1; |
435 | |
436 | /* If the window is now too large, drop the first position */ |
437 | if (activeSegment.end - activeSegment.begin == dmersInK + 1) { |
438 | U32 delDmer = ctx->dmerAt[activeSegment.begin]; |
439 | U32 *delDmerOcc = COVER_map_at(activeDmers, delDmer); |
440 | activeSegment.begin += 1; |
441 | *delDmerOcc -= 1; |
442 | /* If this is the last occurence of the dmer, subtract its score */ |
443 | if (*delDmerOcc == 0) { |
444 | COVER_map_remove(activeDmers, delDmer); |
445 | activeSegment.score -= freqs[delDmer]; |
446 | } |
447 | } |
448 | |
449 | /* If this segment is the best so far save it */ |
450 | if (activeSegment.score > bestSegment.score) { |
451 | bestSegment = activeSegment; |
452 | } |
453 | } |
454 | { |
455 | /* Trim off the zero frequency head and tail from the segment. */ |
456 | U32 newBegin = bestSegment.end; |
457 | U32 newEnd = bestSegment.begin; |
458 | U32 pos; |
459 | for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) { |
460 | U32 freq = freqs[ctx->dmerAt[pos]]; |
461 | if (freq != 0) { |
462 | newBegin = MIN(newBegin, pos); |
463 | newEnd = pos + 1; |
464 | } |
465 | } |
466 | bestSegment.begin = newBegin; |
467 | bestSegment.end = newEnd; |
468 | } |
469 | { |
470 | /* Zero out the frequency of each dmer covered by the chosen segment. */ |
471 | U32 pos; |
472 | for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) { |
473 | freqs[ctx->dmerAt[pos]] = 0; |
474 | } |
475 | } |
476 | return bestSegment; |
477 | } |
478 | |
479 | /** |
480 | * Check the validity of the parameters. |
481 | * Returns non-zero if the parameters are valid and 0 otherwise. |
482 | */ |
483 | static int COVER_checkParameters(ZDICT_cover_params_t parameters, |
484 | size_t maxDictSize) { |
485 | /* k and d are required parameters */ |
486 | if (parameters.d == 0 || parameters.k == 0) { |
487 | return 0; |
488 | } |
489 | /* k <= maxDictSize */ |
490 | if (parameters.k > maxDictSize) { |
491 | return 0; |
492 | } |
493 | /* d <= k */ |
494 | if (parameters.d > parameters.k) { |
495 | return 0; |
496 | } |
497 | return 1; |
498 | } |
499 | |
500 | /** |
501 | * Clean up a context initialized with `COVER_ctx_init()`. |
502 | */ |
503 | static void COVER_ctx_destroy(COVER_ctx_t *ctx) { |
504 | if (!ctx) { |
505 | return; |
506 | } |
507 | if (ctx->suffix) { |
508 | free(ctx->suffix); |
509 | ctx->suffix = NULL; |
510 | } |
511 | if (ctx->freqs) { |
512 | free(ctx->freqs); |
513 | ctx->freqs = NULL; |
514 | } |
515 | if (ctx->dmerAt) { |
516 | free(ctx->dmerAt); |
517 | ctx->dmerAt = NULL; |
518 | } |
519 | if (ctx->offsets) { |
520 | free(ctx->offsets); |
521 | ctx->offsets = NULL; |
522 | } |
523 | } |
524 | |
525 | /** |
526 | * Prepare a context for dictionary building. |
527 | * The context is only dependent on the parameter `d` and can used multiple |
528 | * times. |
529 | * Returns 1 on success or zero on error. |
530 | * The context must be destroyed with `COVER_ctx_destroy()`. |
531 | */ |
532 | static int COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer, |
533 | const size_t *samplesSizes, unsigned nbSamples, |
534 | unsigned d) { |
535 | const BYTE *const samples = (const BYTE *)samplesBuffer; |
536 | const size_t totalSamplesSize = COVER_sum(samplesSizes, nbSamples); |
537 | /* Checks */ |
538 | if (totalSamplesSize < MAX(d, sizeof(U64)) || |
539 | totalSamplesSize >= (size_t)COVER_MAX_SAMPLES_SIZE) { |
540 | DISPLAYLEVEL(1, "Total samples size is too large (%u MB), maximum size is %u MB\n" , |
541 | (U32)(totalSamplesSize>>20), (COVER_MAX_SAMPLES_SIZE >> 20)); |
542 | return 0; |
543 | } |
544 | /* Zero the context */ |
545 | memset(ctx, 0, sizeof(*ctx)); |
546 | DISPLAYLEVEL(2, "Training on %u samples of total size %u\n" , nbSamples, |
547 | (U32)totalSamplesSize); |
548 | ctx->samples = samples; |
549 | ctx->samplesSizes = samplesSizes; |
550 | ctx->nbSamples = nbSamples; |
551 | /* Partial suffix array */ |
552 | ctx->suffixSize = totalSamplesSize - MAX(d, sizeof(U64)) + 1; |
553 | ctx->suffix = (U32 *)malloc(ctx->suffixSize * sizeof(U32)); |
554 | /* Maps index to the dmerID */ |
555 | ctx->dmerAt = (U32 *)malloc(ctx->suffixSize * sizeof(U32)); |
556 | /* The offsets of each file */ |
557 | ctx->offsets = (size_t *)malloc((nbSamples + 1) * sizeof(size_t)); |
558 | if (!ctx->suffix || !ctx->dmerAt || !ctx->offsets) { |
559 | DISPLAYLEVEL(1, "Failed to allocate scratch buffers\n" ); |
560 | COVER_ctx_destroy(ctx); |
561 | return 0; |
562 | } |
563 | ctx->freqs = NULL; |
564 | ctx->d = d; |
565 | |
566 | /* Fill offsets from the samlesSizes */ |
567 | { |
568 | U32 i; |
569 | ctx->offsets[0] = 0; |
570 | for (i = 1; i <= nbSamples; ++i) { |
571 | ctx->offsets[i] = ctx->offsets[i - 1] + samplesSizes[i - 1]; |
572 | } |
573 | } |
574 | DISPLAYLEVEL(2, "Constructing partial suffix array\n" ); |
575 | { |
576 | /* suffix is a partial suffix array. |
577 | * It only sorts suffixes by their first parameters.d bytes. |
578 | * The sort is stable, so each dmer group is sorted by position in input. |
579 | */ |
580 | U32 i; |
581 | for (i = 0; i < ctx->suffixSize; ++i) { |
582 | ctx->suffix[i] = i; |
583 | } |
584 | /* qsort doesn't take an opaque pointer, so pass as a global */ |
585 | g_ctx = ctx; |
586 | qsort(ctx->suffix, ctx->suffixSize, sizeof(U32), |
587 | (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp)); |
588 | } |
589 | DISPLAYLEVEL(2, "Computing frequencies\n" ); |
590 | /* For each dmer group (group of positions with the same first d bytes): |
591 | * 1. For each position we set dmerAt[position] = dmerID. The dmerID is |
592 | * (groupBeginPtr - suffix). This allows us to go from position to |
593 | * dmerID so we can look up values in freq. |
594 | * 2. We calculate how many samples the dmer occurs in and save it in |
595 | * freqs[dmerId]. |
596 | */ |
597 | COVER_groupBy(ctx->suffix, ctx->suffixSize, sizeof(U32), ctx, |
598 | (ctx->d <= 8 ? &COVER_cmp8 : &COVER_cmp), &COVER_group); |
599 | ctx->freqs = ctx->suffix; |
600 | ctx->suffix = NULL; |
601 | return 1; |
602 | } |
603 | |
604 | /** |
605 | * Given the prepared context build the dictionary. |
606 | */ |
607 | static size_t COVER_buildDictionary(const COVER_ctx_t *ctx, U32 *freqs, |
608 | COVER_map_t *activeDmers, void *dictBuffer, |
609 | size_t dictBufferCapacity, |
610 | ZDICT_cover_params_t parameters) { |
611 | BYTE *const dict = (BYTE *)dictBuffer; |
612 | size_t tail = dictBufferCapacity; |
613 | /* Divide the data up into epochs of equal size. |
614 | * We will select at least one segment from each epoch. |
615 | */ |
616 | const U32 epochs = (U32)(dictBufferCapacity / parameters.k); |
617 | const U32 epochSize = (U32)(ctx->suffixSize / epochs); |
618 | size_t epoch; |
619 | DISPLAYLEVEL(2, "Breaking content into %u epochs of size %u\n" , epochs, |
620 | epochSize); |
621 | /* Loop through the epochs until there are no more segments or the dictionary |
622 | * is full. |
623 | */ |
624 | for (epoch = 0; tail > 0; epoch = (epoch + 1) % epochs) { |
625 | const U32 epochBegin = (U32)(epoch * epochSize); |
626 | const U32 epochEnd = epochBegin + epochSize; |
627 | size_t segmentSize; |
628 | /* Select a segment */ |
629 | COVER_segment_t segment = COVER_selectSegment( |
630 | ctx, freqs, activeDmers, epochBegin, epochEnd, parameters); |
631 | /* If the segment covers no dmers, then we are out of content */ |
632 | if (segment.score == 0) { |
633 | break; |
634 | } |
635 | /* Trim the segment if necessary and if it is too small then we are done */ |
636 | segmentSize = MIN(segment.end - segment.begin + parameters.d - 1, tail); |
637 | if (segmentSize < parameters.d) { |
638 | break; |
639 | } |
640 | /* We fill the dictionary from the back to allow the best segments to be |
641 | * referenced with the smallest offsets. |
642 | */ |
643 | tail -= segmentSize; |
644 | memcpy(dict + tail, ctx->samples + segment.begin, segmentSize); |
645 | DISPLAYUPDATE( |
646 | 2, "\r%u%% " , |
647 | (U32)(((dictBufferCapacity - tail) * 100) / dictBufferCapacity)); |
648 | } |
649 | DISPLAYLEVEL(2, "\r%79s\r" , "" ); |
650 | return tail; |
651 | } |
652 | |
653 | ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover( |
654 | void *dictBuffer, size_t dictBufferCapacity, |
655 | const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples, |
656 | ZDICT_cover_params_t parameters) |
657 | { |
658 | BYTE* const dict = (BYTE*)dictBuffer; |
659 | COVER_ctx_t ctx; |
660 | COVER_map_t activeDmers; |
661 | |
662 | /* Initialize global data */ |
663 | g_displayLevel = parameters.zParams.notificationLevel; |
664 | /* Checks */ |
665 | if (!COVER_checkParameters(parameters, dictBufferCapacity)) { |
666 | DISPLAYLEVEL(1, "Cover parameters incorrect\n" ); |
667 | return ERROR(GENERIC); |
668 | } |
669 | if (nbSamples == 0) { |
670 | DISPLAYLEVEL(1, "Cover must have at least one input file\n" ); |
671 | return ERROR(GENERIC); |
672 | } |
673 | if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) { |
674 | DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n" , |
675 | ZDICT_DICTSIZE_MIN); |
676 | return ERROR(dstSize_tooSmall); |
677 | } |
678 | /* Initialize context and activeDmers */ |
679 | if (!COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, |
680 | parameters.d)) { |
681 | return ERROR(GENERIC); |
682 | } |
683 | if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) { |
684 | DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n" ); |
685 | COVER_ctx_destroy(&ctx); |
686 | return ERROR(GENERIC); |
687 | } |
688 | |
689 | DISPLAYLEVEL(2, "Building dictionary\n" ); |
690 | { |
691 | const size_t tail = |
692 | COVER_buildDictionary(&ctx, ctx.freqs, &activeDmers, dictBuffer, |
693 | dictBufferCapacity, parameters); |
694 | const size_t dictionarySize = ZDICT_finalizeDictionary( |
695 | dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail, |
696 | samplesBuffer, samplesSizes, nbSamples, parameters.zParams); |
697 | if (!ZSTD_isError(dictionarySize)) { |
698 | DISPLAYLEVEL(2, "Constructed dictionary of size %u\n" , |
699 | (U32)dictionarySize); |
700 | } |
701 | COVER_ctx_destroy(&ctx); |
702 | COVER_map_destroy(&activeDmers); |
703 | return dictionarySize; |
704 | } |
705 | } |
706 | |
707 | /** |
708 | * COVER_best_t is used for two purposes: |
709 | * 1. Synchronizing threads. |
710 | * 2. Saving the best parameters and dictionary. |
711 | * |
712 | * All of the methods except COVER_best_init() are thread safe if zstd is |
713 | * compiled with multithreaded support. |
714 | */ |
715 | typedef struct COVER_best_s { |
716 | ZSTD_pthread_mutex_t mutex; |
717 | ZSTD_pthread_cond_t cond; |
718 | size_t liveJobs; |
719 | void *dict; |
720 | size_t dictSize; |
721 | ZDICT_cover_params_t parameters; |
722 | size_t compressedSize; |
723 | } COVER_best_t; |
724 | |
725 | /** |
726 | * Initialize the `COVER_best_t`. |
727 | */ |
728 | static void COVER_best_init(COVER_best_t *best) { |
729 | if (best==NULL) return; /* compatible with init on NULL */ |
730 | (void)ZSTD_pthread_mutex_init(&best->mutex, NULL); |
731 | (void)ZSTD_pthread_cond_init(&best->cond, NULL); |
732 | best->liveJobs = 0; |
733 | best->dict = NULL; |
734 | best->dictSize = 0; |
735 | best->compressedSize = (size_t)-1; |
736 | memset(&best->parameters, 0, sizeof(best->parameters)); |
737 | } |
738 | |
739 | /** |
740 | * Wait until liveJobs == 0. |
741 | */ |
742 | static void COVER_best_wait(COVER_best_t *best) { |
743 | if (!best) { |
744 | return; |
745 | } |
746 | ZSTD_pthread_mutex_lock(&best->mutex); |
747 | while (best->liveJobs != 0) { |
748 | ZSTD_pthread_cond_wait(&best->cond, &best->mutex); |
749 | } |
750 | ZSTD_pthread_mutex_unlock(&best->mutex); |
751 | } |
752 | |
753 | /** |
754 | * Call COVER_best_wait() and then destroy the COVER_best_t. |
755 | */ |
756 | static void COVER_best_destroy(COVER_best_t *best) { |
757 | if (!best) { |
758 | return; |
759 | } |
760 | COVER_best_wait(best); |
761 | if (best->dict) { |
762 | free(best->dict); |
763 | } |
764 | ZSTD_pthread_mutex_destroy(&best->mutex); |
765 | ZSTD_pthread_cond_destroy(&best->cond); |
766 | } |
767 | |
768 | /** |
769 | * Called when a thread is about to be launched. |
770 | * Increments liveJobs. |
771 | */ |
772 | static void COVER_best_start(COVER_best_t *best) { |
773 | if (!best) { |
774 | return; |
775 | } |
776 | ZSTD_pthread_mutex_lock(&best->mutex); |
777 | ++best->liveJobs; |
778 | ZSTD_pthread_mutex_unlock(&best->mutex); |
779 | } |
780 | |
781 | /** |
782 | * Called when a thread finishes executing, both on error or success. |
783 | * Decrements liveJobs and signals any waiting threads if liveJobs == 0. |
784 | * If this dictionary is the best so far save it and its parameters. |
785 | */ |
786 | static void COVER_best_finish(COVER_best_t *best, size_t compressedSize, |
787 | ZDICT_cover_params_t parameters, void *dict, |
788 | size_t dictSize) { |
789 | if (!best) { |
790 | return; |
791 | } |
792 | { |
793 | size_t liveJobs; |
794 | ZSTD_pthread_mutex_lock(&best->mutex); |
795 | --best->liveJobs; |
796 | liveJobs = best->liveJobs; |
797 | /* If the new dictionary is better */ |
798 | if (compressedSize < best->compressedSize) { |
799 | /* Allocate space if necessary */ |
800 | if (!best->dict || best->dictSize < dictSize) { |
801 | if (best->dict) { |
802 | free(best->dict); |
803 | } |
804 | best->dict = malloc(dictSize); |
805 | if (!best->dict) { |
806 | best->compressedSize = ERROR(GENERIC); |
807 | best->dictSize = 0; |
808 | return; |
809 | } |
810 | } |
811 | /* Save the dictionary, parameters, and size */ |
812 | memcpy(best->dict, dict, dictSize); |
813 | best->dictSize = dictSize; |
814 | best->parameters = parameters; |
815 | best->compressedSize = compressedSize; |
816 | } |
817 | ZSTD_pthread_mutex_unlock(&best->mutex); |
818 | if (liveJobs == 0) { |
819 | ZSTD_pthread_cond_broadcast(&best->cond); |
820 | } |
821 | } |
822 | } |
823 | |
824 | /** |
825 | * Parameters for COVER_tryParameters(). |
826 | */ |
827 | typedef struct COVER_tryParameters_data_s { |
828 | const COVER_ctx_t *ctx; |
829 | COVER_best_t *best; |
830 | size_t dictBufferCapacity; |
831 | ZDICT_cover_params_t parameters; |
832 | } COVER_tryParameters_data_t; |
833 | |
834 | /** |
835 | * Tries a set of parameters and upates the COVER_best_t with the results. |
836 | * This function is thread safe if zstd is compiled with multithreaded support. |
837 | * It takes its parameters as an *OWNING* opaque pointer to support threading. |
838 | */ |
839 | static void COVER_tryParameters(void *opaque) { |
840 | /* Save parameters as local variables */ |
841 | COVER_tryParameters_data_t *const data = (COVER_tryParameters_data_t *)opaque; |
842 | const COVER_ctx_t *const ctx = data->ctx; |
843 | const ZDICT_cover_params_t parameters = data->parameters; |
844 | size_t dictBufferCapacity = data->dictBufferCapacity; |
845 | size_t totalCompressedSize = ERROR(GENERIC); |
846 | /* Allocate space for hash table, dict, and freqs */ |
847 | COVER_map_t activeDmers; |
848 | BYTE *const dict = (BYTE * const)malloc(dictBufferCapacity); |
849 | U32 *freqs = (U32 *)malloc(ctx->suffixSize * sizeof(U32)); |
850 | if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) { |
851 | DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n" ); |
852 | goto _cleanup; |
853 | } |
854 | if (!dict || !freqs) { |
855 | DISPLAYLEVEL(1, "Failed to allocate buffers: out of memory\n" ); |
856 | goto _cleanup; |
857 | } |
858 | /* Copy the frequencies because we need to modify them */ |
859 | memcpy(freqs, ctx->freqs, ctx->suffixSize * sizeof(U32)); |
860 | /* Build the dictionary */ |
861 | { |
862 | const size_t tail = COVER_buildDictionary(ctx, freqs, &activeDmers, dict, |
863 | dictBufferCapacity, parameters); |
864 | dictBufferCapacity = ZDICT_finalizeDictionary( |
865 | dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail, |
866 | ctx->samples, ctx->samplesSizes, (unsigned)ctx->nbSamples, |
867 | parameters.zParams); |
868 | if (ZDICT_isError(dictBufferCapacity)) { |
869 | DISPLAYLEVEL(1, "Failed to finalize dictionary\n" ); |
870 | goto _cleanup; |
871 | } |
872 | } |
873 | /* Check total compressed size */ |
874 | { |
875 | /* Pointers */ |
876 | ZSTD_CCtx *cctx; |
877 | ZSTD_CDict *cdict; |
878 | void *dst; |
879 | /* Local variables */ |
880 | size_t dstCapacity; |
881 | size_t i; |
882 | /* Allocate dst with enough space to compress the maximum sized sample */ |
883 | { |
884 | size_t maxSampleSize = 0; |
885 | for (i = 0; i < ctx->nbSamples; ++i) { |
886 | maxSampleSize = MAX(ctx->samplesSizes[i], maxSampleSize); |
887 | } |
888 | dstCapacity = ZSTD_compressBound(maxSampleSize); |
889 | dst = malloc(dstCapacity); |
890 | } |
891 | /* Create the cctx and cdict */ |
892 | cctx = ZSTD_createCCtx(); |
893 | cdict = ZSTD_createCDict(dict, dictBufferCapacity, |
894 | parameters.zParams.compressionLevel); |
895 | if (!dst || !cctx || !cdict) { |
896 | goto _compressCleanup; |
897 | } |
898 | /* Compress each sample and sum their sizes (or error) */ |
899 | totalCompressedSize = dictBufferCapacity; |
900 | for (i = 0; i < ctx->nbSamples; ++i) { |
901 | const size_t size = ZSTD_compress_usingCDict( |
902 | cctx, dst, dstCapacity, ctx->samples + ctx->offsets[i], |
903 | ctx->samplesSizes[i], cdict); |
904 | if (ZSTD_isError(size)) { |
905 | totalCompressedSize = ERROR(GENERIC); |
906 | goto _compressCleanup; |
907 | } |
908 | totalCompressedSize += size; |
909 | } |
910 | _compressCleanup: |
911 | ZSTD_freeCCtx(cctx); |
912 | ZSTD_freeCDict(cdict); |
913 | if (dst) { |
914 | free(dst); |
915 | } |
916 | } |
917 | |
918 | _cleanup: |
919 | COVER_best_finish(data->best, totalCompressedSize, parameters, dict, |
920 | dictBufferCapacity); |
921 | free(data); |
922 | COVER_map_destroy(&activeDmers); |
923 | if (dict) { |
924 | free(dict); |
925 | } |
926 | if (freqs) { |
927 | free(freqs); |
928 | } |
929 | } |
930 | |
931 | ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover( |
932 | void *dictBuffer, size_t dictBufferCapacity, const void *samplesBuffer, |
933 | const size_t *samplesSizes, unsigned nbSamples, |
934 | ZDICT_cover_params_t *parameters) { |
935 | /* constants */ |
936 | const unsigned nbThreads = parameters->nbThreads; |
937 | const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d; |
938 | const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d; |
939 | const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k; |
940 | const unsigned kMaxK = parameters->k == 0 ? 2000 : parameters->k; |
941 | const unsigned kSteps = parameters->steps == 0 ? 40 : parameters->steps; |
942 | const unsigned kStepSize = MAX((kMaxK - kMinK) / kSteps, 1); |
943 | const unsigned kIterations = |
944 | (1 + (kMaxD - kMinD) / 2) * (1 + (kMaxK - kMinK) / kStepSize); |
945 | /* Local variables */ |
946 | const int displayLevel = parameters->zParams.notificationLevel; |
947 | unsigned iteration = 1; |
948 | unsigned d; |
949 | unsigned k; |
950 | COVER_best_t best; |
951 | POOL_ctx *pool = NULL; |
952 | |
953 | /* Checks */ |
954 | if (kMinK < kMaxD || kMaxK < kMinK) { |
955 | LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect parameters\n" ); |
956 | return ERROR(GENERIC); |
957 | } |
958 | if (nbSamples == 0) { |
959 | DISPLAYLEVEL(1, "Cover must have at least one input file\n" ); |
960 | return ERROR(GENERIC); |
961 | } |
962 | if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) { |
963 | DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n" , |
964 | ZDICT_DICTSIZE_MIN); |
965 | return ERROR(dstSize_tooSmall); |
966 | } |
967 | if (nbThreads > 1) { |
968 | pool = POOL_create(nbThreads, 1); |
969 | if (!pool) { |
970 | return ERROR(memory_allocation); |
971 | } |
972 | } |
973 | /* Initialization */ |
974 | COVER_best_init(&best); |
975 | /* Turn down global display level to clean up display at level 2 and below */ |
976 | g_displayLevel = displayLevel == 0 ? 0 : displayLevel - 1; |
977 | /* Loop through d first because each new value needs a new context */ |
978 | LOCALDISPLAYLEVEL(displayLevel, 2, "Trying %u different sets of parameters\n" , |
979 | kIterations); |
980 | for (d = kMinD; d <= kMaxD; d += 2) { |
981 | /* Initialize the context for this value of d */ |
982 | COVER_ctx_t ctx; |
983 | LOCALDISPLAYLEVEL(displayLevel, 3, "d=%u\n" , d); |
984 | if (!COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, d)) { |
985 | LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to initialize context\n" ); |
986 | COVER_best_destroy(&best); |
987 | POOL_free(pool); |
988 | return ERROR(GENERIC); |
989 | } |
990 | /* Loop through k reusing the same context */ |
991 | for (k = kMinK; k <= kMaxK; k += kStepSize) { |
992 | /* Prepare the arguments */ |
993 | COVER_tryParameters_data_t *data = (COVER_tryParameters_data_t *)malloc( |
994 | sizeof(COVER_tryParameters_data_t)); |
995 | LOCALDISPLAYLEVEL(displayLevel, 3, "k=%u\n" , k); |
996 | if (!data) { |
997 | LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to allocate parameters\n" ); |
998 | COVER_best_destroy(&best); |
999 | COVER_ctx_destroy(&ctx); |
1000 | POOL_free(pool); |
1001 | return ERROR(GENERIC); |
1002 | } |
1003 | data->ctx = &ctx; |
1004 | data->best = &best; |
1005 | data->dictBufferCapacity = dictBufferCapacity; |
1006 | data->parameters = *parameters; |
1007 | data->parameters.k = k; |
1008 | data->parameters.d = d; |
1009 | data->parameters.steps = kSteps; |
1010 | data->parameters.zParams.notificationLevel = g_displayLevel; |
1011 | /* Check the parameters */ |
1012 | if (!COVER_checkParameters(data->parameters, dictBufferCapacity)) { |
1013 | DISPLAYLEVEL(1, "Cover parameters incorrect\n" ); |
1014 | free(data); |
1015 | continue; |
1016 | } |
1017 | /* Call the function and pass ownership of data to it */ |
1018 | COVER_best_start(&best); |
1019 | if (pool) { |
1020 | POOL_add(pool, &COVER_tryParameters, data); |
1021 | } else { |
1022 | COVER_tryParameters(data); |
1023 | } |
1024 | /* Print status */ |
1025 | LOCALDISPLAYUPDATE(displayLevel, 2, "\r%u%% " , |
1026 | (U32)((iteration * 100) / kIterations)); |
1027 | ++iteration; |
1028 | } |
1029 | COVER_best_wait(&best); |
1030 | COVER_ctx_destroy(&ctx); |
1031 | } |
1032 | LOCALDISPLAYLEVEL(displayLevel, 2, "\r%79s\r" , "" ); |
1033 | /* Fill the output buffer and parameters with output of the best parameters */ |
1034 | { |
1035 | const size_t dictSize = best.dictSize; |
1036 | if (ZSTD_isError(best.compressedSize)) { |
1037 | const size_t compressedSize = best.compressedSize; |
1038 | COVER_best_destroy(&best); |
1039 | POOL_free(pool); |
1040 | return compressedSize; |
1041 | } |
1042 | *parameters = best.parameters; |
1043 | memcpy(dictBuffer, best.dict, dictSize); |
1044 | COVER_best_destroy(&best); |
1045 | POOL_free(pool); |
1046 | return dictSize; |
1047 | } |
1048 | } |
1049 | |