SphinxBase  5prealpha
lm_trie.c
1 /* -*- c-basic-offset: 4; indent-tabs-mode: nil -*- */
2 /* ====================================================================
3  * Copyright (c) 2015 Carnegie Mellon University. All rights
4  * reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  *
10  * 1. Redistributions of source code must retain the above copyright
11  * notice, this list of conditions and the following disclaimer.
12  *
13  * 2. Redistributions in binary form must reproduce the above copyright
14  * notice, this list of conditions and the following disclaimer in
15  * the documentation and/or other materials provided with the
16  * distribution.
17  *
18  * This work was supported in part by funding from the Defense Advanced
19  * Research Projects Agency and the National Science Foundation of the
20  * United States of America, and the CMU Sphinx Speech Consortium.
21  *
22  * THIS SOFTWARE IS PROVIDED BY CARNEGIE MELLON UNIVERSITY ``AS IS'' AND
23  * ANY EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
24  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
25  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY
26  * NOR ITS EMPLOYEES BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
27  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
28  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
32  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * ====================================================================
35  *
36  */
37 
38 #include <string.h>
39 #include <stdio.h>
40 #include <assert.h>
41 
42 #include <sphinxbase/prim_type.h>
43 #include <sphinxbase/ckd_alloc.h>
44 #include <sphinxbase/err.h>
45 #include <sphinxbase/priority_queue.h>
46 
47 #include "lm_trie.h"
48 #include "lm_trie_quant.h"
49 
50 static void lm_trie_alloc_ngram(lm_trie_t * trie, uint32 * counts, int order);
51 
52 static uint32
53 base_size(uint32 entries, uint32 max_vocab, uint8 remaining_bits)
54 {
55  uint8 total_bits = bitarr_required_bits(max_vocab) + remaining_bits;
56  /* Extra entry for next pointer at the end.
57  * +7 then / 8 to round up bits and convert to bytes
58  * +sizeof(uint64) so that ReadInt57 etc don't go segfault.
59  * Note that this waste is O(order), not O(number of ngrams).*/
60  return ((1 + entries) * total_bits + 7) / 8 + sizeof(uint64);
61 }
62 
63 uint32
64 middle_size(uint8 quant_bits, uint32 entries, uint32 max_vocab,
65  uint32 max_ptr)
66 {
67  return base_size(entries, max_vocab,
68  quant_bits + bitarr_required_bits(max_ptr));
69 }
70 
71 uint32
72 longest_size(uint8 quant_bits, uint32 entries, uint32 max_vocab)
73 {
74  return base_size(entries, max_vocab, quant_bits);
75 }
76 
77 static void
78 base_init(base_t * base, void *base_mem, uint32 max_vocab,
79  uint8 remaining_bits)
80 {
81  base->word_bits = bitarr_required_bits(max_vocab);
82  base->word_mask = (1U << base->word_bits) - 1U;
83  if (base->word_bits > 25)
84  E_ERROR
85  ("Sorry, word indices more than %d are not implemented. Edit util/bit_packing.hh and fix the bit packing functions\n",
86  (1U << 25));
87  base->total_bits = base->word_bits + remaining_bits;
88 
89  base->base = (uint8 *) base_mem;
90  base->insert_index = 0;
91  base->max_vocab = max_vocab;
92 }
93 
94 void
95 middle_init(middle_t * middle, void *base_mem, uint8 quant_bits,
96  uint32 entries, uint32 max_vocab, uint32 max_next,
97  void *next_source)
98 {
99  middle->quant_bits = quant_bits;
100  bitarr_mask_from_max(&middle->next_mask, max_next);
101  middle->next_source = next_source;
102  if (entries + 1 >= (1U << 25) || (max_next >= (1U << 25)))
103  E_ERROR
104  ("Sorry, this does not support more than %d n-grams of a particular order. Edit util/bit_packing.hh and fix the bit packing functions\n",
105  (1U << 25));
106  base_init(&middle->base, base_mem, max_vocab,
107  quant_bits + middle->next_mask.bits);
108 }
109 
110 void
111 longest_init(longest_t * longest, void *base_mem, uint8 quant_bits,
112  uint32 max_vocab)
113 {
114  base_init(&longest->base, base_mem, max_vocab, quant_bits);
115 }
116 
117 static bitarr_address_t
118 middle_insert(middle_t * middle, uint32 word, int order, int max_order)
119 {
120  uint32 at_pointer;
121  uint32 next;
122  bitarr_address_t address;
123  assert(word <= middle->base.word_mask);
124  address.base = middle->base.base;
125  address.offset = middle->base.insert_index * middle->base.total_bits;
126  bitarr_write_int25(address, middle->base.word_bits, word);
127  address.offset += middle->base.word_bits;
128  at_pointer = address.offset;
129  address.offset += middle->quant_bits;
130  if (order == max_order - 1) {
131  next = ((longest_t *) middle->next_source)->base.insert_index;
132  }
133  else {
134  next = ((middle_t *) middle->next_source)->base.insert_index;
135  }
136 
137  bitarr_write_int25(address, middle->next_mask.bits, next);
138  middle->base.insert_index++;
139  address.offset = at_pointer;
140  return address;
141 }
142 
143 static bitarr_address_t
144 longest_insert(longest_t * longest, uint32 index)
145 {
146  bitarr_address_t address;
147  assert(index <= longest->base.word_mask);
148  address.base = longest->base.base;
149  address.offset = longest->base.insert_index * longest->base.total_bits;
150  bitarr_write_int25(address, longest->base.word_bits, index);
151  address.offset += longest->base.word_bits;
152  longest->base.insert_index++;
153  return address;
154 }
155 
156 static void
157 middle_finish_loading(middle_t * middle, uint32 next_end)
158 {
159  bitarr_address_t address;
160  address.base = middle->base.base;
161  address.offset =
162  (middle->base.insert_index + 1) * middle->base.total_bits -
163  middle->next_mask.bits;
164  bitarr_write_int25(address, middle->next_mask.bits, next_end);
165 }
166 
167 static uint32
168 unigram_next(lm_trie_t * trie, int order)
169 {
170  return order ==
171  2 ? trie->longest->base.insert_index : trie->middle_begin->base.
172  insert_index;
173 }
174 
175 void
176 lm_trie_fix_counts(ngram_raw_t ** raw_ngrams, uint32 * counts,
177  uint32 * fixed_counts, int order)
178 {
179  priority_queue_t *ngrams =
180  priority_queue_create(order - 1, &ngram_ord_comparator);
181  uint32 raw_ngram_ptrs[NGRAM_MAX_ORDER - 1];
182  uint32 words[NGRAM_MAX_ORDER];
183  int i;
184 
185  memset(words, -1, sizeof(words));
186  memcpy(fixed_counts, counts, order * sizeof(*fixed_counts));
187  for (i = 2; i <= order; i++) {
188  ngram_raw_t *tmp_ngram;
189 
190  if (counts[i - 1] <= 0)
191  continue;
192 
193  raw_ngram_ptrs[i - 2] = 0;
194 
195  tmp_ngram =
196  (ngram_raw_t *) ckd_calloc(1, sizeof(*tmp_ngram));
197  *tmp_ngram = raw_ngrams[i - 2][0];
198  tmp_ngram->order = i;
199  priority_queue_add(ngrams, tmp_ngram);
200  }
201 
202  for (;;) {
203  int32 to_increment = TRUE;
204  ngram_raw_t *top;
205  if (priority_queue_size(ngrams) == 0) {
206  break;
207  }
208  top = (ngram_raw_t *) priority_queue_poll(ngrams);
209  if (top->order == 2) {
210  memcpy(words, top->words, 2 * sizeof(*words));
211  }
212  else {
213  for (i = 0; i < top->order - 1; i++) {
214  if (words[i] != top->words[i]) {
215  int num;
216  num = (i == 0) ? 1 : i;
217  memcpy(words, top->words,
218  (num + 1) * sizeof(*words));
219  fixed_counts[num]++;
220  to_increment = FALSE;
221  break;
222  }
223  }
224  words[top->order - 1] = top->words[top->order - 1];
225  }
226  if (to_increment) {
227  raw_ngram_ptrs[top->order - 2]++;
228  }
229  if (raw_ngram_ptrs[top->order - 2] < counts[top->order - 1]) {
230  *top = raw_ngrams[top->order - 2][raw_ngram_ptrs[top->order - 2]];
231  priority_queue_add(ngrams, top);
232  }
233  else {
234  ckd_free(top);
235  }
236  }
237 
238  assert(priority_queue_size(ngrams) == 0);
239  priority_queue_free(ngrams, NULL);
240 }
241 
242 
243 static void
244 recursive_insert(lm_trie_t * trie, ngram_raw_t ** raw_ngrams,
245  uint32 * counts, int order)
246 {
247  uint32 unigram_idx = 0;
248  uint32 *words;
249  float *probs;
250  const uint32 unigram_count = (uint32) counts[0];
251  priority_queue_t *ngrams =
252  priority_queue_create(order, &ngram_ord_comparator);
253  ngram_raw_t *ngram;
254  uint32 *raw_ngrams_ptr;
255  int i;
256 
257  words = (uint32 *) ckd_calloc(order, sizeof(*words));
258  probs = (float *) ckd_calloc(order - 1, sizeof(*probs));
259  ngram = (ngram_raw_t *) ckd_calloc(1, sizeof(*ngram));
260  ngram->order = 1;
261  ngram->words = &unigram_idx;
262  priority_queue_add(ngrams, ngram);
263  raw_ngrams_ptr =
264  (uint32 *) ckd_calloc(order - 1, sizeof(*raw_ngrams_ptr));
265  for (i = 2; i <= order; ++i) {
266  ngram_raw_t *tmp_ngram;
267 
268  if (counts[i - 1] <= 0)
269  continue;
270 
271  raw_ngrams_ptr[i - 2] = 0;
272  tmp_ngram =
273  (ngram_raw_t *) ckd_calloc(1, sizeof(*tmp_ngram));
274  *tmp_ngram = raw_ngrams[i - 2][0];
275  tmp_ngram->order = i;
276 
277  priority_queue_add(ngrams, tmp_ngram);
278  }
279 
280  for (;;) {
281  ngram_raw_t *top =
282  (ngram_raw_t *) priority_queue_poll(ngrams);
283 
284  if (top->order == 1) {
285  trie->unigrams[unigram_idx].next = unigram_next(trie, order);
286  words[0] = unigram_idx;
287  probs[0] = trie->unigrams[unigram_idx].prob;
288  if (++unigram_idx == unigram_count + 1) {
289  ckd_free(top);
290  break;
291  }
292  priority_queue_add(ngrams, top);
293  }
294  else {
295  for (i = 0; i < top->order - 1; i++) {
296  if (words[i] != top->words[i]) {
297  /* need to insert dummy suffixes to make ngram of higher order reachable */
298  int j;
299  assert(i > 0); /* unigrams are not pruned without removing ngrams that contains them */
300  for (j = i; j < top->order - 1; j++) {
301  middle_t *middle = &trie->middle_begin[j - 1];
302  bitarr_address_t address =
303  middle_insert(middle, top->words[j],
304  j + 1, order);
305  /* calculate prob for blank */
306  float calc_prob =
307  probs[j - 1] +
308  trie->unigrams[top->words[j]].bo;
309  probs[j] = calc_prob;
310  lm_trie_quant_mwrite(trie->quant, address, j - 1,
311  calc_prob, 0.0f);
312  }
313  }
314  }
315  memcpy(words, top->words,
316  top->order * sizeof(*words));
317  if (top->order == order) {
318  bitarr_address_t address =
319  longest_insert(trie->longest,
320  top->words[top->order - 1]);
321  lm_trie_quant_lwrite(trie->quant, address, top->prob);
322  }
323  else {
324  middle_t *middle = &trie->middle_begin[top->order - 2];
325  bitarr_address_t address =
326  middle_insert(middle,
327  top->words[top->order - 1],
328  top->order, order);
329  /* write prob and backoff */
330  probs[top->order - 1] = top->prob;
331  lm_trie_quant_mwrite(trie->quant, address, top->order - 2,
332  top->prob, top->backoff);
333  }
334  raw_ngrams_ptr[top->order - 2]++;
335  if (raw_ngrams_ptr[top->order - 2] < counts[top->order - 1]) {
336  *top = raw_ngrams[top->order -
337  2][raw_ngrams_ptr[top->order - 2]];
338 
339  priority_queue_add(ngrams, top);
340  }
341  else {
342  ckd_free(top);
343  }
344  }
345  }
346  assert(priority_queue_size(ngrams) == 0);
347  priority_queue_free(ngrams, NULL);
348  ckd_free(raw_ngrams_ptr);
349  ckd_free(words);
350  ckd_free(probs);
351 }
352 
353 static lm_trie_t *
354 lm_trie_init(uint32 unigram_count)
355 {
356  lm_trie_t *trie;
357 
358  trie = (lm_trie_t *) ckd_calloc(1, sizeof(*trie));
359  memset(trie->hist_cache, -1, sizeof(trie->hist_cache)); /* prepare request history */
360  memset(trie->backoff_cache, 0, sizeof(trie->backoff_cache));
361  trie->unigrams =
362  (unigram_t *) ckd_calloc((unigram_count + 1),
363  sizeof(*trie->unigrams));
364  trie->ngram_mem = NULL;
365  return trie;
366 }
367 
368 lm_trie_t *
369 lm_trie_create(uint32 unigram_count, int order)
370 {
371  lm_trie_t *trie = lm_trie_init(unigram_count);
372  trie->quant =
373  (order > 1) ? lm_trie_quant_create(order) : 0;
374  return trie;
375 }
376 
377 lm_trie_t *
378 lm_trie_read_bin(uint32 * counts, int order, FILE * fp)
379 {
380  lm_trie_t *trie = lm_trie_init(counts[0]);
381  trie->quant = (order > 1) ? lm_trie_quant_read_bin(fp, order) : NULL;
382  fread(trie->unigrams, sizeof(*trie->unigrams), (counts[0] + 1), fp);
383  if (order > 1) {
384  lm_trie_alloc_ngram(trie, counts, order);
385  fread(trie->ngram_mem, 1, trie->ngram_mem_size, fp);
386  }
387  return trie;
388 }
389 
390 void
391 lm_trie_write_bin(lm_trie_t * trie, uint32 unigram_count, FILE * fp)
392 {
393 
394  if (trie->quant)
395  lm_trie_quant_write_bin(trie->quant, fp);
396  fwrite(trie->unigrams, sizeof(*trie->unigrams), (unigram_count + 1),
397  fp);
398  if (trie->ngram_mem)
399  fwrite(trie->ngram_mem, 1, trie->ngram_mem_size, fp);
400 }
401 
402 void
403 lm_trie_free(lm_trie_t * trie)
404 {
405  if (trie->ngram_mem) {
406  ckd_free(trie->ngram_mem);
407  ckd_free(trie->middle_begin);
408  ckd_free(trie->longest);
409  }
410  if (trie->quant)
411  lm_trie_quant_free(trie->quant);
412  ckd_free(trie->unigrams);
413  ckd_free(trie);
414 }
415 
416 static void
417 lm_trie_alloc_ngram(lm_trie_t * trie, uint32 * counts, int order)
418 {
419  int i;
420  uint8 *mem_ptr;
421  uint8 **middle_starts;
422 
423  trie->ngram_mem_size = 0;
424  for (i = 1; i < order - 1; i++) {
425  trie->ngram_mem_size +=
426  middle_size(lm_trie_quant_msize(trie->quant), counts[i],
427  counts[0], counts[i + 1]);
428  }
429  trie->ngram_mem_size +=
430  longest_size(lm_trie_quant_lsize(trie->quant), counts[order - 1],
431  counts[0]);
432  trie->ngram_mem =
433  (uint8 *) ckd_calloc(trie->ngram_mem_size,
434  sizeof(*trie->ngram_mem));
435  mem_ptr = trie->ngram_mem;
436  trie->middle_begin =
437  (middle_t *) ckd_calloc(order - 2, sizeof(*trie->middle_begin));
438  trie->middle_end = trie->middle_begin + (order - 2);
439  middle_starts =
440  (uint8 **) ckd_calloc(order - 2, sizeof(*middle_starts));
441  for (i = 2; i < order; i++) {
442  middle_starts[i - 2] = mem_ptr;
443  mem_ptr +=
444  middle_size(lm_trie_quant_msize(trie->quant), counts[i - 1],
445  counts[0], counts[i]);
446  }
447  trie->longest = (longest_t *) ckd_calloc(1, sizeof(*trie->longest));
448  /* Crazy backwards thing so we initialize using pointers to ones that have already been initialized */
449  for (i = order - 1; i >= 2; --i) {
450  middle_t *middle_ptr = &trie->middle_begin[i - 2];
451  middle_init(middle_ptr, middle_starts[i - 2],
452  lm_trie_quant_msize(trie->quant), counts[i - 1],
453  counts[0], counts[i],
454  (i ==
455  order -
456  1) ? (void *) trie->longest : (void *) &trie->
457  middle_begin[i - 1]);
458  }
459  ckd_free(middle_starts);
460  longest_init(trie->longest, mem_ptr, lm_trie_quant_lsize(trie->quant),
461  counts[0]);
462 }
463 
464 void
465 lm_trie_build(lm_trie_t * trie, ngram_raw_t ** raw_ngrams, uint32 * counts, uint32 *out_counts,
466  int order)
467 {
468  int i;
469 
470  lm_trie_fix_counts(raw_ngrams, counts, out_counts, order);
471  lm_trie_alloc_ngram(trie, out_counts, order);
472 
473  if (order > 1)
474  E_INFO("Training quantizer\n");
475  for (i = 2; i < order; i++) {
476  lm_trie_quant_train(trie->quant, i, counts[i - 1],
477  raw_ngrams[i - 2]);
478  }
479  lm_trie_quant_train_prob(trie->quant, order, counts[order - 1],
480  raw_ngrams[order - 2]);
481 
482  E_INFO("Building LM trie\n");
483  recursive_insert(trie, raw_ngrams, counts, order);
484  /* Set ending offsets so the last entry will be sized properly */
485  /* Last entry for unigrams was already set. */
486  if (trie->middle_begin != trie->middle_end) {
487  middle_t *middle_ptr;
488  for (middle_ptr = trie->middle_begin;
489  middle_ptr != trie->middle_end - 1; ++middle_ptr) {
490  middle_t *next_middle_ptr = middle_ptr + 1;
491  middle_finish_loading(middle_ptr,
492  next_middle_ptr->base.insert_index);
493  }
494  middle_ptr = trie->middle_end - 1;
495  middle_finish_loading(middle_ptr,
496  trie->longest->base.insert_index);
497  }
498 }
499 
500 unigram_t *
501 unigram_find(unigram_t * u, uint32 word, node_range_t * next)
502 {
503  unigram_t *ptr = &u[word];
504  next->begin = ptr->next;
505  next->end = (ptr + 1)->next;
506  return ptr;
507 }
508 
509 static size_t
510 calc_pivot(uint32 off, uint32 range, uint32 width)
511 {
512  return (size_t) ((off * width) / (range + 1));
513 }
514 
515 static uint8
516 uniform_find(void *base, uint8 total_bits, uint8 key_bits, uint32 key_mask,
517  uint32 before_it, uint32 before_v,
518  uint32 after_it, uint32 after_v, uint32 key, uint32 * out)
519 {
520  bitarr_address_t address;
521  address.base = base;
522 
523  /* If we look for unigram added later */
524  if (key > after_v)
525  return FALSE;
526 
527  while (after_it - before_it > 1) {
528  uint32 mid;
529  uint32 pivot =
530  before_it + (1 +
531  calc_pivot(key - before_v, after_v - before_v,
532  after_it - before_it - 1));
533  /* access by pivot */
534  address.offset = pivot * (uint32) total_bits;
535  mid = bitarr_read_int25(address, key_bits, key_mask);
536  if (mid < key) {
537  before_it = pivot;
538  before_v = mid;
539  }
540  else if (mid > key) {
541  after_it = pivot;
542  after_v = mid;
543  }
544  else {
545  *out = pivot;
546  return TRUE;
547  }
548  }
549  return FALSE;
550 }
551 
552 static bitarr_address_t
553 middle_find(middle_t * middle, uint32 word, node_range_t * range)
554 {
555  uint32 at_pointer;
556  bitarr_address_t address;
557 
558  /* finding BitPacked with uniform find */
559  if (!uniform_find
560  ((void *) middle->base.base, middle->base.total_bits,
561  middle->base.word_bits, middle->base.word_mask, range->begin - 1,
562  0, range->end, middle->base.max_vocab, word, &at_pointer)) {
563  address.base = NULL;
564  address.offset = 0;
565  return address;
566  }
567 
568  address.base = middle->base.base;
569  at_pointer *= middle->base.total_bits;
570  at_pointer += middle->base.word_bits;
571  address.offset = at_pointer + middle->quant_bits;
572  range->begin =
573  bitarr_read_int25(address, middle->next_mask.bits,
574  middle->next_mask.mask);
575  address.offset += middle->base.total_bits;
576  range->end =
577  bitarr_read_int25(address, middle->next_mask.bits,
578  middle->next_mask.mask);
579  address.offset = at_pointer;
580 
581  return address;
582 }
583 
584 static bitarr_address_t
585 longest_find(longest_t * longest, uint32 word, node_range_t * range)
586 {
587  uint32 at_pointer;
588  bitarr_address_t address;
589 
590  /* finding BitPacked with uniform find */
591  if (!uniform_find
592  ((void *) longest->base.base, longest->base.total_bits,
593  longest->base.word_bits, longest->base.word_mask,
594  range->begin - 1, 0, range->end, longest->base.max_vocab, word,
595  &at_pointer)) {
596  address.base = NULL;
597  address.offset = 0;
598  return address;
599  }
600  address.base = longest->base.base;
601  address.offset =
602  at_pointer * longest->base.total_bits + longest->base.word_bits;
603  return address;
604 }
605 
606 static float
607 get_available_prob(lm_trie_t * trie, int32 wid, int32 * hist,
608  int max_order, int32 n_hist, int32 * n_used)
609 {
610  float prob;
611  node_range_t node;
612  bitarr_address_t address;
613  int order_minus_2;
614  uint8 independent_left;
615  int32 *hist_iter, *hist_end;
616 
617  *n_used = 1;
618  prob = unigram_find(trie->unigrams, wid, &node)->prob;
619  if (n_hist == 0) {
620  return prob;
621  }
622 
623  /* find ngrams of higher order if any */
624  order_minus_2 = 0;
625  independent_left = (node.begin == node.end);
626  hist_iter = hist;
627  hist_end = hist + n_hist;
628  for (;; order_minus_2++, hist_iter++) {
629  if (hist_iter == hist_end)
630  return prob;
631  if (independent_left)
632  return prob;
633  if (order_minus_2 == max_order - 2)
634  break;
635 
636  address =
637  middle_find(&trie->middle_begin[order_minus_2], *hist_iter,
638  &node);
639  independent_left = (address.base == NULL)
640  || (node.begin == node.end);
641 
642  /* didn't find entry */
643  if (address.base == NULL)
644  return prob;
645  prob = lm_trie_quant_mpread(trie->quant, address, order_minus_2);
646  *n_used = order_minus_2 + 2;
647  }
648 
649  address = longest_find(trie->longest, *hist_iter, &node);
650  if (address.base != NULL) {
651  prob = lm_trie_quant_lpread(trie->quant, address);
652  *n_used = max_order;
653  }
654  return prob;
655 }
656 
657 static float
658 get_available_backoff(lm_trie_t * trie, int32 start, int32 * hist,
659  int32 n_hist)
660 {
661  float backoff = 0.0f;
662  int order_minus_2;
663  int32 *hist_iter;
664  node_range_t node;
665  unigram_t *first_hist = unigram_find(trie->unigrams, hist[0], &node);
666  if (start <= 1) {
667  backoff += first_hist->bo;
668  start = 2;
669  }
670  order_minus_2 = start - 2;
671  for (hist_iter = hist + start - 1; hist_iter < hist + n_hist;
672  hist_iter++, order_minus_2++) {
673  bitarr_address_t address =
674  middle_find(&trie->middle_begin[order_minus_2], *hist_iter,
675  &node);
676  if (address.base == NULL)
677  break;
678  backoff +=
679  lm_trie_quant_mboread(trie->quant, address, order_minus_2);
680  }
681  return backoff;
682 }
683 
684 static float
685 lm_trie_nobo_score(lm_trie_t * trie, int32 wid, int32 * hist,
686  int max_order, int32 n_hist, int32 * n_used)
687 {
688  float prob =
689  get_available_prob(trie, wid, hist, max_order, n_hist, n_used);
690  if (n_hist < *n_used)
691  return prob;
692  return prob + get_available_backoff(trie, *n_used, hist, n_hist);
693 }
694 
695 static float
696 lm_trie_hist_score(lm_trie_t * trie, int32 wid, int32 * hist, int32 n_hist,
697  int32 * n_used)
698 {
699  float prob;
700  int i, j;
701  node_range_t node;
702  bitarr_address_t address;
703 
704  *n_used = 1;
705  prob = unigram_find(trie->unigrams, wid, &node)->prob;
706  if (n_hist == 0)
707  return prob;
708  for (i = 0; i < n_hist - 1; i++) {
709  address = middle_find(&trie->middle_begin[i], hist[i], &node);
710  if (address.base == NULL) {
711  for (j = i; j < n_hist; j++) {
712  prob += trie->backoff_cache[j];
713  }
714  return prob;
715  }
716  else {
717  (*n_used)++;
718  prob = lm_trie_quant_mpread(trie->quant, address, i);
719  }
720  }
721  address = longest_find(trie->longest, hist[n_hist - 1], &node);
722  if (address.base == NULL) {
723  return prob + trie->backoff_cache[n_hist - 1];
724  }
725  else {
726  (*n_used)++;
727  return lm_trie_quant_lpread(trie->quant, address);
728  }
729 }
730 
731 static uint8
732 history_matches(int32 * hist, int32 * prev_hist, int32 n_hist)
733 {
734  int i;
735  for (i = 0; i < n_hist; i++) {
736  if (hist[i] != prev_hist[i]) {
737  return FALSE;
738  }
739  }
740  return TRUE;
741 }
742 
743 static void
744 update_backoff(lm_trie_t * trie, int32 * hist, int32 n_hist)
745 {
746  int i;
747  node_range_t node;
748  bitarr_address_t address;
749 
750  memset(trie->backoff_cache, 0, sizeof(trie->backoff_cache));
751  trie->backoff_cache[0] = unigram_find(trie->unigrams, hist[0], &node)->bo;
752  for (i = 1; i < n_hist; i++) {
753  address = middle_find(&trie->middle_begin[i - 1], hist[i], &node);
754  if (address.base == NULL) {
755  break;
756  }
757  trie->backoff_cache[i] =
758  lm_trie_quant_mboread(trie->quant, address, i - 1);
759  }
760  memcpy(trie->hist_cache, hist, n_hist * sizeof(*hist));
761 }
762 
763 float
764 lm_trie_score(lm_trie_t * trie, int order, int32 wid, int32 * hist,
765  int32 n_hist, int32 * n_used)
766 {
767  if (n_hist < order - 1) {
768  return lm_trie_nobo_score(trie, wid, hist, order, n_hist, n_used);
769  }
770  else {
771  assert(n_hist == order - 1);
772  if (!history_matches(hist, (int32 *) trie->hist_cache, n_hist)) {
773  update_backoff(trie, hist, n_hist);
774  }
775  return lm_trie_hist_score(trie, wid, hist, n_hist, n_used);
776  }
777 }
778 
779 void
780 lm_trie_fill_raw_ngram(lm_trie_t * trie,
781  ngram_raw_t * raw_ngrams, uint32 * raw_ngram_idx,
782  uint32 * counts, node_range_t range, uint32 * hist,
783  int n_hist, int order, int max_order)
784 {
785  if (n_hist > 0 && range.begin == range.end) {
786  return;
787  }
788  if (n_hist == 0) {
789  uint32 i;
790  for (i = 0; i < counts[0]; i++) {
791  node_range_t node;
792  unigram_find(trie->unigrams, i, &node);
793  hist[0] = i;
794  lm_trie_fill_raw_ngram(trie, raw_ngrams, raw_ngram_idx, counts,
795  node, hist, 1, order, max_order);
796  }
797  }
798  else if (n_hist < order - 1) {
799  uint32 ptr;
800  node_range_t node;
801  bitarr_address_t address;
802  uint32 new_word;
803  middle_t *middle = &trie->middle_begin[n_hist - 1];
804  for (ptr = range.begin; ptr < range.end; ptr++) {
805  address.base = middle->base.base;
806  address.offset = ptr * middle->base.total_bits;
807  new_word =
808  bitarr_read_int25(address, middle->base.word_bits,
809  middle->base.word_mask);
810  hist[n_hist] = new_word;
811  address.offset += middle->base.word_bits + middle->quant_bits;
812  node.begin =
813  bitarr_read_int25(address, middle->next_mask.bits,
814  middle->next_mask.mask);
815  address.offset =
816  (ptr + 1) * middle->base.total_bits +
817  middle->base.word_bits + middle->quant_bits;
818  node.end =
819  bitarr_read_int25(address, middle->next_mask.bits,
820  middle->next_mask.mask);
821  lm_trie_fill_raw_ngram(trie, raw_ngrams, raw_ngram_idx, counts,
822  node, hist, n_hist + 1, order, max_order);
823  }
824  }
825  else {
826  bitarr_address_t address;
827  uint32 ptr;
828  float prob, backoff;
829  int i;
830  assert(n_hist == order - 1);
831  for (ptr = range.begin; ptr < range.end; ptr++) {
832  ngram_raw_t *raw_ngram = &raw_ngrams[*raw_ngram_idx];
833  if (order == max_order) {
834  longest_t *longest = trie->longest;
835  address.base = longest->base.base;
836  address.offset = ptr * longest->base.total_bits;
837  hist[n_hist] =
838  bitarr_read_int25(address, longest->base.word_bits,
839  longest->base.word_mask);
840  address.offset += longest->base.word_bits;
841  prob = lm_trie_quant_lpread(trie->quant, address);
842  }
843  else {
844  middle_t *middle = &trie->middle_begin[n_hist - 1];
845  address.base = middle->base.base;
846  address.offset = ptr * middle->base.total_bits;
847  hist[n_hist] =
848  bitarr_read_int25(address, middle->base.word_bits,
849  middle->base.word_mask);
850  address.offset += middle->base.word_bits;
851  prob =
852  lm_trie_quant_mpread(trie->quant, address, n_hist - 1);
853  backoff =
854  lm_trie_quant_mboread(trie->quant, address,
855  n_hist - 1);
856  raw_ngram->backoff = backoff;
857  }
858  raw_ngram->prob = prob;
859  raw_ngram->words =
860  (uint32 *) ckd_calloc(order, sizeof(*raw_ngram->words));
861  for (i = 0; i <= n_hist; i++) {
862  raw_ngram->words[i] = hist[n_hist - i];
863  }
864  (*raw_ngram_idx)++;
865  }
866  }
867 }
#define E_INFO(...)
Print logging information to standard error stream.
Definition: err.h:114
#define ckd_calloc(n, sz)
Macros to simplify the use of above functions.
Definition: ckd_alloc.h:248
SPHINXBASE_EXPORT void bitarr_mask_from_max(bitarr_mask_t *bit_mask, uint32 max_value)
Fills mask for certain int range according to provided max value.
Definition: bitarr.c:125
#define E_ERROR(...)
Print error message to error log.
Definition: err.h:104
SPHINXBASE_EXPORT uint8 bitarr_required_bits(uint32 max_value)
Computes amount of bits required ti store integers upto value provided.
Definition: bitarr.c:131
Sphinx&#39;s memory allocation/deallocation routines.
Basic type definitions used in Sphinx.
SPHINXBASE_EXPORT void bitarr_write_int25(bitarr_address_t address, uint8 length, uint32 value)
Write specified value into bit array.
Definition: bitarr.c:112
SPHINXBASE_EXPORT void ckd_free(void *ptr)
Test and free a 1-D array.
Definition: ckd_alloc.c:244
Definition: lm_trie.h:58
Structure that stores address of certain value in bit array.
Definition: bitarr.h:73
SPHINXBASE_EXPORT uint32 bitarr_read_int25(bitarr_address_t address, uint8 length, uint32 mask)
Read uint32 value from bit array.
Definition: bitarr.c:100
Implementation of logging routines.
Definition: dtoa.c:178