xref: /freebsd/sys/contrib/zstd/lib/dictBuilder/cover.c (revision a25896ca1270e25b657ceaa8d47d5699515f5c25)
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 "cover.h"
33 #include "zstd_internal.h" /* includes zstd.h */
34 #ifndef ZDICT_STATIC_LINKING_ONLY
35 #define ZDICT_STATIC_LINKING_ONLY
36 #endif
37 #include "zdict.h"
38 
39 /*-*************************************
40 *  Constants
41 ***************************************/
42 #define COVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((U32)-1) : ((U32)1 GB))
43 #define DEFAULT_SPLITPOINT 1.0
44 
45 /*-*************************************
46 *  Console display
47 ***************************************/
48 static int g_displayLevel = 2;
49 #define DISPLAY(...)                                                           \
50   {                                                                            \
51     fprintf(stderr, __VA_ARGS__);                                              \
52     fflush(stderr);                                                            \
53   }
54 #define LOCALDISPLAYLEVEL(displayLevel, l, ...)                                \
55   if (displayLevel >= l) {                                                     \
56     DISPLAY(__VA_ARGS__);                                                      \
57   } /* 0 : no display;   1: errors;   2: default;  3: details;  4: debug */
58 #define DISPLAYLEVEL(l, ...) LOCALDISPLAYLEVEL(g_displayLevel, l, __VA_ARGS__)
59 
60 #define LOCALDISPLAYUPDATE(displayLevel, l, ...)                               \
61   if (displayLevel >= l) {                                                     \
62     if ((clock() - g_time > refreshRate) || (displayLevel >= 4)) {             \
63       g_time = clock();                                                        \
64       DISPLAY(__VA_ARGS__);                                                    \
65     }                                                                          \
66   }
67 #define DISPLAYUPDATE(l, ...) LOCALDISPLAYUPDATE(g_displayLevel, l, __VA_ARGS__)
68 static const clock_t refreshRate = CLOCKS_PER_SEC * 15 / 100;
69 static clock_t g_time = 0;
70 
71 /*-*************************************
72 * Hash table
73 ***************************************
74 * A small specialized hash map for storing activeDmers.
75 * The map does not resize, so if it becomes full it will loop forever.
76 * Thus, the map must be large enough to store every value.
77 * The map implements linear probing and keeps its load less than 0.5.
78 */
79 
80 #define MAP_EMPTY_VALUE ((U32)-1)
81 typedef struct COVER_map_pair_t_s {
82   U32 key;
83   U32 value;
84 } COVER_map_pair_t;
85 
86 typedef struct COVER_map_s {
87   COVER_map_pair_t *data;
88   U32 sizeLog;
89   U32 size;
90   U32 sizeMask;
91 } COVER_map_t;
92 
93 /**
94  * Clear the map.
95  */
96 static void COVER_map_clear(COVER_map_t *map) {
97   memset(map->data, MAP_EMPTY_VALUE, map->size * sizeof(COVER_map_pair_t));
98 }
99 
100 /**
101  * Initializes a map of the given size.
102  * Returns 1 on success and 0 on failure.
103  * The map must be destroyed with COVER_map_destroy().
104  * The map is only guaranteed to be large enough to hold size elements.
105  */
106 static int COVER_map_init(COVER_map_t *map, U32 size) {
107   map->sizeLog = ZSTD_highbit32(size) + 2;
108   map->size = (U32)1 << map->sizeLog;
109   map->sizeMask = map->size - 1;
110   map->data = (COVER_map_pair_t *)malloc(map->size * sizeof(COVER_map_pair_t));
111   if (!map->data) {
112     map->sizeLog = 0;
113     map->size = 0;
114     return 0;
115   }
116   COVER_map_clear(map);
117   return 1;
118 }
119 
120 /**
121  * Internal hash function
122  */
123 static const U32 prime4bytes = 2654435761U;
124 static U32 COVER_map_hash(COVER_map_t *map, U32 key) {
125   return (key * prime4bytes) >> (32 - map->sizeLog);
126 }
127 
128 /**
129  * Helper function that returns the index that a key should be placed into.
130  */
131 static U32 COVER_map_index(COVER_map_t *map, U32 key) {
132   const U32 hash = COVER_map_hash(map, key);
133   U32 i;
134   for (i = hash;; i = (i + 1) & map->sizeMask) {
135     COVER_map_pair_t *pos = &map->data[i];
136     if (pos->value == MAP_EMPTY_VALUE) {
137       return i;
138     }
139     if (pos->key == key) {
140       return i;
141     }
142   }
143 }
144 
145 /**
146  * Returns the pointer to the value for key.
147  * If key is not in the map, it is inserted and the value is set to 0.
148  * The map must not be full.
149  */
150 static U32 *COVER_map_at(COVER_map_t *map, U32 key) {
151   COVER_map_pair_t *pos = &map->data[COVER_map_index(map, key)];
152   if (pos->value == MAP_EMPTY_VALUE) {
153     pos->key = key;
154     pos->value = 0;
155   }
156   return &pos->value;
157 }
158 
159 /**
160  * Deletes key from the map if present.
161  */
162 static void COVER_map_remove(COVER_map_t *map, U32 key) {
163   U32 i = COVER_map_index(map, key);
164   COVER_map_pair_t *del = &map->data[i];
165   U32 shift = 1;
166   if (del->value == MAP_EMPTY_VALUE) {
167     return;
168   }
169   for (i = (i + 1) & map->sizeMask;; i = (i + 1) & map->sizeMask) {
170     COVER_map_pair_t *const pos = &map->data[i];
171     /* If the position is empty we are done */
172     if (pos->value == MAP_EMPTY_VALUE) {
173       del->value = MAP_EMPTY_VALUE;
174       return;
175     }
176     /* If pos can be moved to del do so */
177     if (((i - COVER_map_hash(map, pos->key)) & map->sizeMask) >= shift) {
178       del->key = pos->key;
179       del->value = pos->value;
180       del = pos;
181       shift = 1;
182     } else {
183       ++shift;
184     }
185   }
186 }
187 
188 /**
189  * Destroys a map that is inited with COVER_map_init().
190  */
191 static void COVER_map_destroy(COVER_map_t *map) {
192   if (map->data) {
193     free(map->data);
194   }
195   map->data = NULL;
196   map->size = 0;
197 }
198 
199 /*-*************************************
200 * Context
201 ***************************************/
202 
203 typedef struct {
204   const BYTE *samples;
205   size_t *offsets;
206   const size_t *samplesSizes;
207   size_t nbSamples;
208   size_t nbTrainSamples;
209   size_t nbTestSamples;
210   U32 *suffix;
211   size_t suffixSize;
212   U32 *freqs;
213   U32 *dmerAt;
214   unsigned d;
215 } COVER_ctx_t;
216 
217 /* We need a global context for qsort... */
218 static COVER_ctx_t *g_ctx = NULL;
219 
220 /*-*************************************
221 *  Helper functions
222 ***************************************/
223 
224 /**
225  * Returns the sum of the sample sizes.
226  */
227 size_t COVER_sum(const size_t *samplesSizes, unsigned nbSamples) {
228   size_t sum = 0;
229   unsigned i;
230   for (i = 0; i < nbSamples; ++i) {
231     sum += samplesSizes[i];
232   }
233   return sum;
234 }
235 
236 /**
237  * Returns -1 if the dmer at lp is less than the dmer at rp.
238  * Return 0 if the dmers at lp and rp are equal.
239  * Returns 1 if the dmer at lp is greater than the dmer at rp.
240  */
241 static int COVER_cmp(COVER_ctx_t *ctx, const void *lp, const void *rp) {
242   U32 const lhs = *(U32 const *)lp;
243   U32 const rhs = *(U32 const *)rp;
244   return memcmp(ctx->samples + lhs, ctx->samples + rhs, ctx->d);
245 }
246 /**
247  * Faster version for d <= 8.
248  */
249 static int COVER_cmp8(COVER_ctx_t *ctx, const void *lp, const void *rp) {
250   U64 const mask = (ctx->d == 8) ? (U64)-1 : (((U64)1 << (8 * ctx->d)) - 1);
251   U64 const lhs = MEM_readLE64(ctx->samples + *(U32 const *)lp) & mask;
252   U64 const rhs = MEM_readLE64(ctx->samples + *(U32 const *)rp) & mask;
253   if (lhs < rhs) {
254     return -1;
255   }
256   return (lhs > rhs);
257 }
258 
259 /**
260  * Same as COVER_cmp() except ties are broken by pointer value
261  * NOTE: g_ctx must be set to call this function.  A global is required because
262  * qsort doesn't take an opaque pointer.
263  */
264 static int COVER_strict_cmp(const void *lp, const void *rp) {
265   int result = COVER_cmp(g_ctx, lp, rp);
266   if (result == 0) {
267     result = lp < rp ? -1 : 1;
268   }
269   return result;
270 }
271 /**
272  * Faster version for d <= 8.
273  */
274 static int COVER_strict_cmp8(const void *lp, const void *rp) {
275   int result = COVER_cmp8(g_ctx, lp, rp);
276   if (result == 0) {
277     result = lp < rp ? -1 : 1;
278   }
279   return result;
280 }
281 
282 /**
283  * Returns the first pointer in [first, last) whose element does not compare
284  * less than value.  If no such element exists it returns last.
285  */
286 static const size_t *COVER_lower_bound(const size_t *first, const size_t *last,
287                                        size_t value) {
288   size_t count = last - first;
289   while (count != 0) {
290     size_t step = count / 2;
291     const size_t *ptr = first;
292     ptr += step;
293     if (*ptr < value) {
294       first = ++ptr;
295       count -= step + 1;
296     } else {
297       count = step;
298     }
299   }
300   return first;
301 }
302 
303 /**
304  * Generic groupBy function.
305  * Groups an array sorted by cmp into groups with equivalent values.
306  * Calls grp for each group.
307  */
308 static void
309 COVER_groupBy(const void *data, size_t count, size_t size, COVER_ctx_t *ctx,
310               int (*cmp)(COVER_ctx_t *, const void *, const void *),
311               void (*grp)(COVER_ctx_t *, const void *, const void *)) {
312   const BYTE *ptr = (const BYTE *)data;
313   size_t num = 0;
314   while (num < count) {
315     const BYTE *grpEnd = ptr + size;
316     ++num;
317     while (num < count && cmp(ctx, ptr, grpEnd) == 0) {
318       grpEnd += size;
319       ++num;
320     }
321     grp(ctx, ptr, grpEnd);
322     ptr = grpEnd;
323   }
324 }
325 
326 /*-*************************************
327 *  Cover functions
328 ***************************************/
329 
330 /**
331  * Called on each group of positions with the same dmer.
332  * Counts the frequency of each dmer and saves it in the suffix array.
333  * Fills `ctx->dmerAt`.
334  */
335 static void COVER_group(COVER_ctx_t *ctx, const void *group,
336                         const void *groupEnd) {
337   /* The group consists of all the positions with the same first d bytes. */
338   const U32 *grpPtr = (const U32 *)group;
339   const U32 *grpEnd = (const U32 *)groupEnd;
340   /* The dmerId is how we will reference this dmer.
341    * This allows us to map the whole dmer space to a much smaller space, the
342    * size of the suffix array.
343    */
344   const U32 dmerId = (U32)(grpPtr - ctx->suffix);
345   /* Count the number of samples this dmer shows up in */
346   U32 freq = 0;
347   /* Details */
348   const size_t *curOffsetPtr = ctx->offsets;
349   const size_t *offsetsEnd = ctx->offsets + ctx->nbSamples;
350   /* Once *grpPtr >= curSampleEnd this occurrence of the dmer is in a
351    * different sample than the last.
352    */
353   size_t curSampleEnd = ctx->offsets[0];
354   for (; grpPtr != grpEnd; ++grpPtr) {
355     /* Save the dmerId for this position so we can get back to it. */
356     ctx->dmerAt[*grpPtr] = dmerId;
357     /* Dictionaries only help for the first reference to the dmer.
358      * After that zstd can reference the match from the previous reference.
359      * So only count each dmer once for each sample it is in.
360      */
361     if (*grpPtr < curSampleEnd) {
362       continue;
363     }
364     freq += 1;
365     /* Binary search to find the end of the sample *grpPtr is in.
366      * In the common case that grpPtr + 1 == grpEnd we can skip the binary
367      * search because the loop is over.
368      */
369     if (grpPtr + 1 != grpEnd) {
370       const size_t *sampleEndPtr =
371           COVER_lower_bound(curOffsetPtr, offsetsEnd, *grpPtr);
372       curSampleEnd = *sampleEndPtr;
373       curOffsetPtr = sampleEndPtr + 1;
374     }
375   }
376   /* At this point we are never going to look at this segment of the suffix
377    * array again.  We take advantage of this fact to save memory.
378    * We store the frequency of the dmer in the first position of the group,
379    * which is dmerId.
380    */
381   ctx->suffix[dmerId] = freq;
382 }
383 
384 
385 /**
386  * Selects the best segment in an epoch.
387  * Segments of are scored according to the function:
388  *
389  * Let F(d) be the frequency of dmer d.
390  * Let S_i be the dmer at position i of segment S which has length k.
391  *
392  *     Score(S) = F(S_1) + F(S_2) + ... + F(S_{k-d+1})
393  *
394  * Once the dmer d is in the dictionay we set F(d) = 0.
395  */
396 static COVER_segment_t COVER_selectSegment(const COVER_ctx_t *ctx, U32 *freqs,
397                                            COVER_map_t *activeDmers, U32 begin,
398                                            U32 end,
399                                            ZDICT_cover_params_t parameters) {
400   /* Constants */
401   const U32 k = parameters.k;
402   const U32 d = parameters.d;
403   const U32 dmersInK = k - d + 1;
404   /* Try each segment (activeSegment) and save the best (bestSegment) */
405   COVER_segment_t bestSegment = {0, 0, 0};
406   COVER_segment_t activeSegment;
407   /* Reset the activeDmers in the segment */
408   COVER_map_clear(activeDmers);
409   /* The activeSegment starts at the beginning of the epoch. */
410   activeSegment.begin = begin;
411   activeSegment.end = begin;
412   activeSegment.score = 0;
413   /* Slide the activeSegment through the whole epoch.
414    * Save the best segment in bestSegment.
415    */
416   while (activeSegment.end < end) {
417     /* The dmerId for the dmer at the next position */
418     U32 newDmer = ctx->dmerAt[activeSegment.end];
419     /* The entry in activeDmers for this dmerId */
420     U32 *newDmerOcc = COVER_map_at(activeDmers, newDmer);
421     /* If the dmer isn't already present in the segment add its score. */
422     if (*newDmerOcc == 0) {
423       /* The paper suggest using the L-0.5 norm, but experiments show that it
424        * doesn't help.
425        */
426       activeSegment.score += freqs[newDmer];
427     }
428     /* Add the dmer to the segment */
429     activeSegment.end += 1;
430     *newDmerOcc += 1;
431 
432     /* If the window is now too large, drop the first position */
433     if (activeSegment.end - activeSegment.begin == dmersInK + 1) {
434       U32 delDmer = ctx->dmerAt[activeSegment.begin];
435       U32 *delDmerOcc = COVER_map_at(activeDmers, delDmer);
436       activeSegment.begin += 1;
437       *delDmerOcc -= 1;
438       /* If this is the last occurence of the dmer, subtract its score */
439       if (*delDmerOcc == 0) {
440         COVER_map_remove(activeDmers, delDmer);
441         activeSegment.score -= freqs[delDmer];
442       }
443     }
444 
445     /* If this segment is the best so far save it */
446     if (activeSegment.score > bestSegment.score) {
447       bestSegment = activeSegment;
448     }
449   }
450   {
451     /* Trim off the zero frequency head and tail from the segment. */
452     U32 newBegin = bestSegment.end;
453     U32 newEnd = bestSegment.begin;
454     U32 pos;
455     for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) {
456       U32 freq = freqs[ctx->dmerAt[pos]];
457       if (freq != 0) {
458         newBegin = MIN(newBegin, pos);
459         newEnd = pos + 1;
460       }
461     }
462     bestSegment.begin = newBegin;
463     bestSegment.end = newEnd;
464   }
465   {
466     /* Zero out the frequency of each dmer covered by the chosen segment. */
467     U32 pos;
468     for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) {
469       freqs[ctx->dmerAt[pos]] = 0;
470     }
471   }
472   return bestSegment;
473 }
474 
475 /**
476  * Check the validity of the parameters.
477  * Returns non-zero if the parameters are valid and 0 otherwise.
478  */
479 static int COVER_checkParameters(ZDICT_cover_params_t parameters,
480                                  size_t maxDictSize) {
481   /* k and d are required parameters */
482   if (parameters.d == 0 || parameters.k == 0) {
483     return 0;
484   }
485   /* k <= maxDictSize */
486   if (parameters.k > maxDictSize) {
487     return 0;
488   }
489   /* d <= k */
490   if (parameters.d > parameters.k) {
491     return 0;
492   }
493   /* 0 < splitPoint <= 1 */
494   if (parameters.splitPoint <= 0 || parameters.splitPoint > 1){
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, double splitPoint) {
535   const BYTE *const samples = (const BYTE *)samplesBuffer;
536   const size_t totalSamplesSize = COVER_sum(samplesSizes, nbSamples);
537   /* Split samples into testing and training sets */
538   const unsigned nbTrainSamples = splitPoint < 1.0 ? (unsigned)((double)nbSamples * splitPoint) : nbSamples;
539   const unsigned nbTestSamples = splitPoint < 1.0 ? nbSamples - nbTrainSamples : nbSamples;
540   const size_t trainingSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes, nbTrainSamples) : totalSamplesSize;
541   const size_t testSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes + nbTrainSamples, nbTestSamples) : totalSamplesSize;
542   /* Checks */
543   if (totalSamplesSize < MAX(d, sizeof(U64)) ||
544       totalSamplesSize >= (size_t)COVER_MAX_SAMPLES_SIZE) {
545     DISPLAYLEVEL(1, "Total samples size is too large (%u MB), maximum size is %u MB\n",
546                  (U32)(totalSamplesSize>>20), (COVER_MAX_SAMPLES_SIZE >> 20));
547     return 0;
548   }
549   /* Check if there are at least 5 training samples */
550   if (nbTrainSamples < 5) {
551     DISPLAYLEVEL(1, "Total number of training samples is %u and is invalid.", nbTrainSamples);
552     return 0;
553   }
554   /* Check if there's testing sample */
555   if (nbTestSamples < 1) {
556     DISPLAYLEVEL(1, "Total number of testing samples is %u and is invalid.", nbTestSamples);
557     return 0;
558   }
559   /* Zero the context */
560   memset(ctx, 0, sizeof(*ctx));
561   DISPLAYLEVEL(2, "Training on %u samples of total size %u\n", nbTrainSamples,
562                (U32)trainingSamplesSize);
563   DISPLAYLEVEL(2, "Testing on %u samples of total size %u\n", nbTestSamples,
564                (U32)testSamplesSize);
565   ctx->samples = samples;
566   ctx->samplesSizes = samplesSizes;
567   ctx->nbSamples = nbSamples;
568   ctx->nbTrainSamples = nbTrainSamples;
569   ctx->nbTestSamples = nbTestSamples;
570   /* Partial suffix array */
571   ctx->suffixSize = trainingSamplesSize - MAX(d, sizeof(U64)) + 1;
572   ctx->suffix = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
573   /* Maps index to the dmerID */
574   ctx->dmerAt = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
575   /* The offsets of each file */
576   ctx->offsets = (size_t *)malloc((nbSamples + 1) * sizeof(size_t));
577   if (!ctx->suffix || !ctx->dmerAt || !ctx->offsets) {
578     DISPLAYLEVEL(1, "Failed to allocate scratch buffers\n");
579     COVER_ctx_destroy(ctx);
580     return 0;
581   }
582   ctx->freqs = NULL;
583   ctx->d = d;
584 
585   /* Fill offsets from the samplesSizes */
586   {
587     U32 i;
588     ctx->offsets[0] = 0;
589     for (i = 1; i <= nbSamples; ++i) {
590       ctx->offsets[i] = ctx->offsets[i - 1] + samplesSizes[i - 1];
591     }
592   }
593   DISPLAYLEVEL(2, "Constructing partial suffix array\n");
594   {
595     /* suffix is a partial suffix array.
596      * It only sorts suffixes by their first parameters.d bytes.
597      * The sort is stable, so each dmer group is sorted by position in input.
598      */
599     U32 i;
600     for (i = 0; i < ctx->suffixSize; ++i) {
601       ctx->suffix[i] = i;
602     }
603     /* qsort doesn't take an opaque pointer, so pass as a global.
604      * On OpenBSD qsort() is not guaranteed to be stable, their mergesort() is.
605      */
606     g_ctx = ctx;
607 #if defined(__OpenBSD__)
608     mergesort(ctx->suffix, ctx->suffixSize, sizeof(U32),
609           (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp));
610 #else
611     qsort(ctx->suffix, ctx->suffixSize, sizeof(U32),
612           (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp));
613 #endif
614   }
615   DISPLAYLEVEL(2, "Computing frequencies\n");
616   /* For each dmer group (group of positions with the same first d bytes):
617    * 1. For each position we set dmerAt[position] = dmerID.  The dmerID is
618    *    (groupBeginPtr - suffix).  This allows us to go from position to
619    *    dmerID so we can look up values in freq.
620    * 2. We calculate how many samples the dmer occurs in and save it in
621    *    freqs[dmerId].
622    */
623   COVER_groupBy(ctx->suffix, ctx->suffixSize, sizeof(U32), ctx,
624                 (ctx->d <= 8 ? &COVER_cmp8 : &COVER_cmp), &COVER_group);
625   ctx->freqs = ctx->suffix;
626   ctx->suffix = NULL;
627   return 1;
628 }
629 
630 /**
631  * Given the prepared context build the dictionary.
632  */
633 static size_t COVER_buildDictionary(const COVER_ctx_t *ctx, U32 *freqs,
634                                     COVER_map_t *activeDmers, void *dictBuffer,
635                                     size_t dictBufferCapacity,
636                                     ZDICT_cover_params_t parameters) {
637   BYTE *const dict = (BYTE *)dictBuffer;
638   size_t tail = dictBufferCapacity;
639   /* Divide the data up into epochs of equal size.
640    * We will select at least one segment from each epoch.
641    */
642   const U32 epochs = MAX(1, (U32)(dictBufferCapacity / parameters.k / 4));
643   const U32 epochSize = (U32)(ctx->suffixSize / epochs);
644   size_t epoch;
645   DISPLAYLEVEL(2, "Breaking content into %u epochs of size %u\n", epochs,
646                epochSize);
647   /* Loop through the epochs until there are no more segments or the dictionary
648    * is full.
649    */
650   for (epoch = 0; tail > 0; epoch = (epoch + 1) % epochs) {
651     const U32 epochBegin = (U32)(epoch * epochSize);
652     const U32 epochEnd = epochBegin + epochSize;
653     size_t segmentSize;
654     /* Select a segment */
655     COVER_segment_t segment = COVER_selectSegment(
656         ctx, freqs, activeDmers, epochBegin, epochEnd, parameters);
657     /* If the segment covers no dmers, then we are out of content */
658     if (segment.score == 0) {
659       break;
660     }
661     /* Trim the segment if necessary and if it is too small then we are done */
662     segmentSize = MIN(segment.end - segment.begin + parameters.d - 1, tail);
663     if (segmentSize < parameters.d) {
664       break;
665     }
666     /* We fill the dictionary from the back to allow the best segments to be
667      * referenced with the smallest offsets.
668      */
669     tail -= segmentSize;
670     memcpy(dict + tail, ctx->samples + segment.begin, segmentSize);
671     DISPLAYUPDATE(
672         2, "\r%u%%       ",
673         (U32)(((dictBufferCapacity - tail) * 100) / dictBufferCapacity));
674   }
675   DISPLAYLEVEL(2, "\r%79s\r", "");
676   return tail;
677 }
678 
679 ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
680     void *dictBuffer, size_t dictBufferCapacity,
681     const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples,
682     ZDICT_cover_params_t parameters)
683 {
684   BYTE* const dict = (BYTE*)dictBuffer;
685   COVER_ctx_t ctx;
686   COVER_map_t activeDmers;
687   parameters.splitPoint = 1.0;
688   /* Initialize global data */
689   g_displayLevel = parameters.zParams.notificationLevel;
690   /* Checks */
691   if (!COVER_checkParameters(parameters, dictBufferCapacity)) {
692     DISPLAYLEVEL(1, "Cover parameters incorrect\n");
693     return ERROR(GENERIC);
694   }
695   if (nbSamples == 0) {
696     DISPLAYLEVEL(1, "Cover must have at least one input file\n");
697     return ERROR(GENERIC);
698   }
699   if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) {
700     DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n",
701                  ZDICT_DICTSIZE_MIN);
702     return ERROR(dstSize_tooSmall);
703   }
704   /* Initialize context and activeDmers */
705   if (!COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples,
706                       parameters.d, parameters.splitPoint)) {
707     return ERROR(GENERIC);
708   }
709   if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) {
710     DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n");
711     COVER_ctx_destroy(&ctx);
712     return ERROR(GENERIC);
713   }
714 
715   DISPLAYLEVEL(2, "Building dictionary\n");
716   {
717     const size_t tail =
718         COVER_buildDictionary(&ctx, ctx.freqs, &activeDmers, dictBuffer,
719                               dictBufferCapacity, parameters);
720     const size_t dictionarySize = ZDICT_finalizeDictionary(
721         dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail,
722         samplesBuffer, samplesSizes, nbSamples, parameters.zParams);
723     if (!ZSTD_isError(dictionarySize)) {
724       DISPLAYLEVEL(2, "Constructed dictionary of size %u\n",
725                    (U32)dictionarySize);
726     }
727     COVER_ctx_destroy(&ctx);
728     COVER_map_destroy(&activeDmers);
729     return dictionarySize;
730   }
731 }
732 
733 
734 
735 size_t COVER_checkTotalCompressedSize(const ZDICT_cover_params_t parameters,
736                                     const size_t *samplesSizes, const BYTE *samples,
737                                     size_t *offsets,
738                                     size_t nbTrainSamples, size_t nbSamples,
739                                     BYTE *const dict, size_t dictBufferCapacity) {
740   size_t totalCompressedSize = ERROR(GENERIC);
741   /* Pointers */
742   ZSTD_CCtx *cctx;
743   ZSTD_CDict *cdict;
744   void *dst;
745   /* Local variables */
746   size_t dstCapacity;
747   size_t i;
748   /* Allocate dst with enough space to compress the maximum sized sample */
749   {
750     size_t maxSampleSize = 0;
751     i = parameters.splitPoint < 1.0 ? nbTrainSamples : 0;
752     for (; i < nbSamples; ++i) {
753       maxSampleSize = MAX(samplesSizes[i], maxSampleSize);
754     }
755     dstCapacity = ZSTD_compressBound(maxSampleSize);
756     dst = malloc(dstCapacity);
757   }
758   /* Create the cctx and cdict */
759   cctx = ZSTD_createCCtx();
760   cdict = ZSTD_createCDict(dict, dictBufferCapacity,
761                            parameters.zParams.compressionLevel);
762   if (!dst || !cctx || !cdict) {
763     goto _compressCleanup;
764   }
765   /* Compress each sample and sum their sizes (or error) */
766   totalCompressedSize = dictBufferCapacity;
767   i = parameters.splitPoint < 1.0 ? nbTrainSamples : 0;
768   for (; i < nbSamples; ++i) {
769     const size_t size = ZSTD_compress_usingCDict(
770         cctx, dst, dstCapacity, samples + offsets[i],
771         samplesSizes[i], cdict);
772     if (ZSTD_isError(size)) {
773       totalCompressedSize = ERROR(GENERIC);
774       goto _compressCleanup;
775     }
776     totalCompressedSize += size;
777   }
778 _compressCleanup:
779   ZSTD_freeCCtx(cctx);
780   ZSTD_freeCDict(cdict);
781   if (dst) {
782     free(dst);
783   }
784   return totalCompressedSize;
785 }
786 
787 
788 /**
789  * Initialize the `COVER_best_t`.
790  */
791 void COVER_best_init(COVER_best_t *best) {
792   if (best==NULL) return; /* compatible with init on NULL */
793   (void)ZSTD_pthread_mutex_init(&best->mutex, NULL);
794   (void)ZSTD_pthread_cond_init(&best->cond, NULL);
795   best->liveJobs = 0;
796   best->dict = NULL;
797   best->dictSize = 0;
798   best->compressedSize = (size_t)-1;
799   memset(&best->parameters, 0, sizeof(best->parameters));
800 }
801 
802 /**
803  * Wait until liveJobs == 0.
804  */
805 void COVER_best_wait(COVER_best_t *best) {
806   if (!best) {
807     return;
808   }
809   ZSTD_pthread_mutex_lock(&best->mutex);
810   while (best->liveJobs != 0) {
811     ZSTD_pthread_cond_wait(&best->cond, &best->mutex);
812   }
813   ZSTD_pthread_mutex_unlock(&best->mutex);
814 }
815 
816 /**
817  * Call COVER_best_wait() and then destroy the COVER_best_t.
818  */
819 void COVER_best_destroy(COVER_best_t *best) {
820   if (!best) {
821     return;
822   }
823   COVER_best_wait(best);
824   if (best->dict) {
825     free(best->dict);
826   }
827   ZSTD_pthread_mutex_destroy(&best->mutex);
828   ZSTD_pthread_cond_destroy(&best->cond);
829 }
830 
831 /**
832  * Called when a thread is about to be launched.
833  * Increments liveJobs.
834  */
835 void COVER_best_start(COVER_best_t *best) {
836   if (!best) {
837     return;
838   }
839   ZSTD_pthread_mutex_lock(&best->mutex);
840   ++best->liveJobs;
841   ZSTD_pthread_mutex_unlock(&best->mutex);
842 }
843 
844 /**
845  * Called when a thread finishes executing, both on error or success.
846  * Decrements liveJobs and signals any waiting threads if liveJobs == 0.
847  * If this dictionary is the best so far save it and its parameters.
848  */
849 void COVER_best_finish(COVER_best_t *best, size_t compressedSize,
850                               ZDICT_cover_params_t parameters, void *dict,
851                               size_t dictSize) {
852   if (!best) {
853     return;
854   }
855   {
856     size_t liveJobs;
857     ZSTD_pthread_mutex_lock(&best->mutex);
858     --best->liveJobs;
859     liveJobs = best->liveJobs;
860     /* If the new dictionary is better */
861     if (compressedSize < best->compressedSize) {
862       /* Allocate space if necessary */
863       if (!best->dict || best->dictSize < dictSize) {
864         if (best->dict) {
865           free(best->dict);
866         }
867         best->dict = malloc(dictSize);
868         if (!best->dict) {
869           best->compressedSize = ERROR(GENERIC);
870           best->dictSize = 0;
871           return;
872         }
873       }
874       /* Save the dictionary, parameters, and size */
875       memcpy(best->dict, dict, dictSize);
876       best->dictSize = dictSize;
877       best->parameters = parameters;
878       best->compressedSize = compressedSize;
879     }
880     if (liveJobs == 0) {
881       ZSTD_pthread_cond_broadcast(&best->cond);
882     }
883     ZSTD_pthread_mutex_unlock(&best->mutex);
884   }
885 }
886 
887 /**
888  * Parameters for COVER_tryParameters().
889  */
890 typedef struct COVER_tryParameters_data_s {
891   const COVER_ctx_t *ctx;
892   COVER_best_t *best;
893   size_t dictBufferCapacity;
894   ZDICT_cover_params_t parameters;
895 } COVER_tryParameters_data_t;
896 
897 /**
898  * Tries a set of parameters and updates the COVER_best_t with the results.
899  * This function is thread safe if zstd is compiled with multithreaded support.
900  * It takes its parameters as an *OWNING* opaque pointer to support threading.
901  */
902 static void COVER_tryParameters(void *opaque) {
903   /* Save parameters as local variables */
904   COVER_tryParameters_data_t *const data = (COVER_tryParameters_data_t *)opaque;
905   const COVER_ctx_t *const ctx = data->ctx;
906   const ZDICT_cover_params_t parameters = data->parameters;
907   size_t dictBufferCapacity = data->dictBufferCapacity;
908   size_t totalCompressedSize = ERROR(GENERIC);
909   /* Allocate space for hash table, dict, and freqs */
910   COVER_map_t activeDmers;
911   BYTE *const dict = (BYTE * const)malloc(dictBufferCapacity);
912   U32 *freqs = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
913   if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) {
914     DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n");
915     goto _cleanup;
916   }
917   if (!dict || !freqs) {
918     DISPLAYLEVEL(1, "Failed to allocate buffers: out of memory\n");
919     goto _cleanup;
920   }
921   /* Copy the frequencies because we need to modify them */
922   memcpy(freqs, ctx->freqs, ctx->suffixSize * sizeof(U32));
923   /* Build the dictionary */
924   {
925     const size_t tail = COVER_buildDictionary(ctx, freqs, &activeDmers, dict,
926                                               dictBufferCapacity, parameters);
927     dictBufferCapacity = ZDICT_finalizeDictionary(
928         dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail,
929         ctx->samples, ctx->samplesSizes, (unsigned)ctx->nbTrainSamples,
930         parameters.zParams);
931     if (ZDICT_isError(dictBufferCapacity)) {
932       DISPLAYLEVEL(1, "Failed to finalize dictionary\n");
933       goto _cleanup;
934     }
935   }
936   /* Check total compressed size */
937   totalCompressedSize = COVER_checkTotalCompressedSize(parameters, ctx->samplesSizes,
938                                                        ctx->samples, ctx->offsets,
939                                                        ctx->nbTrainSamples, ctx->nbSamples,
940                                                        dict, dictBufferCapacity);
941 
942 _cleanup:
943   COVER_best_finish(data->best, totalCompressedSize, parameters, dict,
944                     dictBufferCapacity);
945   free(data);
946   COVER_map_destroy(&activeDmers);
947   if (dict) {
948     free(dict);
949   }
950   if (freqs) {
951     free(freqs);
952   }
953 }
954 
955 ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
956     void *dictBuffer, size_t dictBufferCapacity, const void *samplesBuffer,
957     const size_t *samplesSizes, unsigned nbSamples,
958     ZDICT_cover_params_t *parameters) {
959   /* constants */
960   const unsigned nbThreads = parameters->nbThreads;
961   const double splitPoint =
962       parameters->splitPoint <= 0.0 ? DEFAULT_SPLITPOINT : parameters->splitPoint;
963   const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d;
964   const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d;
965   const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k;
966   const unsigned kMaxK = parameters->k == 0 ? 2000 : parameters->k;
967   const unsigned kSteps = parameters->steps == 0 ? 40 : parameters->steps;
968   const unsigned kStepSize = MAX((kMaxK - kMinK) / kSteps, 1);
969   const unsigned kIterations =
970       (1 + (kMaxD - kMinD) / 2) * (1 + (kMaxK - kMinK) / kStepSize);
971   /* Local variables */
972   const int displayLevel = parameters->zParams.notificationLevel;
973   unsigned iteration = 1;
974   unsigned d;
975   unsigned k;
976   COVER_best_t best;
977   POOL_ctx *pool = NULL;
978 
979   /* Checks */
980   if (splitPoint <= 0 || splitPoint > 1) {
981     LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect parameters\n");
982     return ERROR(GENERIC);
983   }
984   if (kMinK < kMaxD || kMaxK < kMinK) {
985     LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect parameters\n");
986     return ERROR(GENERIC);
987   }
988   if (nbSamples == 0) {
989     DISPLAYLEVEL(1, "Cover must have at least one input file\n");
990     return ERROR(GENERIC);
991   }
992   if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) {
993     DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n",
994                  ZDICT_DICTSIZE_MIN);
995     return ERROR(dstSize_tooSmall);
996   }
997   if (nbThreads > 1) {
998     pool = POOL_create(nbThreads, 1);
999     if (!pool) {
1000       return ERROR(memory_allocation);
1001     }
1002   }
1003   /* Initialization */
1004   COVER_best_init(&best);
1005   /* Turn down global display level to clean up display at level 2 and below */
1006   g_displayLevel = displayLevel == 0 ? 0 : displayLevel - 1;
1007   /* Loop through d first because each new value needs a new context */
1008   LOCALDISPLAYLEVEL(displayLevel, 2, "Trying %u different sets of parameters\n",
1009                     kIterations);
1010   for (d = kMinD; d <= kMaxD; d += 2) {
1011     /* Initialize the context for this value of d */
1012     COVER_ctx_t ctx;
1013     LOCALDISPLAYLEVEL(displayLevel, 3, "d=%u\n", d);
1014     if (!COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, d, splitPoint)) {
1015       LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to initialize context\n");
1016       COVER_best_destroy(&best);
1017       POOL_free(pool);
1018       return ERROR(GENERIC);
1019     }
1020     /* Loop through k reusing the same context */
1021     for (k = kMinK; k <= kMaxK; k += kStepSize) {
1022       /* Prepare the arguments */
1023       COVER_tryParameters_data_t *data = (COVER_tryParameters_data_t *)malloc(
1024           sizeof(COVER_tryParameters_data_t));
1025       LOCALDISPLAYLEVEL(displayLevel, 3, "k=%u\n", k);
1026       if (!data) {
1027         LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to allocate parameters\n");
1028         COVER_best_destroy(&best);
1029         COVER_ctx_destroy(&ctx);
1030         POOL_free(pool);
1031         return ERROR(GENERIC);
1032       }
1033       data->ctx = &ctx;
1034       data->best = &best;
1035       data->dictBufferCapacity = dictBufferCapacity;
1036       data->parameters = *parameters;
1037       data->parameters.k = k;
1038       data->parameters.d = d;
1039       data->parameters.splitPoint = splitPoint;
1040       data->parameters.steps = kSteps;
1041       data->parameters.zParams.notificationLevel = g_displayLevel;
1042       /* Check the parameters */
1043       if (!COVER_checkParameters(data->parameters, dictBufferCapacity)) {
1044         DISPLAYLEVEL(1, "Cover parameters incorrect\n");
1045         free(data);
1046         continue;
1047       }
1048       /* Call the function and pass ownership of data to it */
1049       COVER_best_start(&best);
1050       if (pool) {
1051         POOL_add(pool, &COVER_tryParameters, data);
1052       } else {
1053         COVER_tryParameters(data);
1054       }
1055       /* Print status */
1056       LOCALDISPLAYUPDATE(displayLevel, 2, "\r%u%%       ",
1057                          (U32)((iteration * 100) / kIterations));
1058       ++iteration;
1059     }
1060     COVER_best_wait(&best);
1061     COVER_ctx_destroy(&ctx);
1062   }
1063   LOCALDISPLAYLEVEL(displayLevel, 2, "\r%79s\r", "");
1064   /* Fill the output buffer and parameters with output of the best parameters */
1065   {
1066     const size_t dictSize = best.dictSize;
1067     if (ZSTD_isError(best.compressedSize)) {
1068       const size_t compressedSize = best.compressedSize;
1069       COVER_best_destroy(&best);
1070       POOL_free(pool);
1071       return compressedSize;
1072     }
1073     *parameters = best.parameters;
1074     memcpy(dictBuffer, best.dict, dictSize);
1075     COVER_best_destroy(&best);
1076     POOL_free(pool);
1077     return dictSize;
1078   }
1079 }
1080