forked from facebookresearch/faiss
-
Notifications
You must be signed in to change notification settings - Fork 1
/
IndexPQ.cpp
1074 lines (832 loc) · 27.7 KB
/
IndexPQ.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/**
* Copyright (c) 2015-present, Facebook, Inc.
* All rights reserved.
*
* This source code is licensed under the BSD+Patents license found in the
* LICENSE file in the root directory of this source tree.
*/
/* Copyright 2004-present Facebook. All Rights Reserved.
Index based on product quantiztion.
*/
#include "IndexPQ.h"
#include <cstddef>
#include <cstring>
#include <cstdio>
#include <cmath>
#include <algorithm>
#include "FaissAssert.h"
#include "hamming.h"
namespace faiss {
/*********************************************************
* IndexPQ implementation
********************************************************/
IndexPQ::IndexPQ (int d, size_t M, size_t nbits, MetricType metric):
Index(d, metric), pq(d, M, nbits)
{
is_trained = false;
do_polysemous_training = false;
polysemous_ht = nbits * M + 1;
search_type = ST_PQ;
encode_signs = false;
}
IndexPQ::IndexPQ ()
{
metric_type = METRIC_L2;
is_trained = false;
do_polysemous_training = false;
polysemous_ht = pq.nbits * pq.M + 1;
search_type = ST_PQ;
encode_signs = false;
}
void IndexPQ::train (idx_t n, const float *x)
{
if (!do_polysemous_training) { // standard training
pq.train(n, x);
} else {
idx_t ntrain_perm = polysemous_training.ntrain_permutation;
if (ntrain_perm > n / 4)
ntrain_perm = n / 4;
if (verbose) {
printf ("PQ training on %ld points, remains %ld points: "
"training polysemous on %s\n",
n - ntrain_perm, ntrain_perm,
ntrain_perm == 0 ? "centroids" : "these");
}
pq.train(n - ntrain_perm, x);
polysemous_training.optimize_pq_for_hamming (
pq, ntrain_perm, x + (n - ntrain_perm) * d);
}
is_trained = true;
}
void IndexPQ::add (idx_t n, const float *x)
{
FAISS_THROW_IF_NOT (is_trained);
codes.resize ((n + ntotal) * pq.code_size);
pq.compute_codes (x, &codes[ntotal * pq.code_size], n);
ntotal += n;
}
void IndexPQ::reset()
{
codes.clear();
ntotal = 0;
}
void IndexPQ::reconstruct_n (idx_t i0, idx_t ni, float *recons) const
{
FAISS_THROW_IF_NOT (ni == 0 || (i0 >= 0 && i0 + ni <= ntotal));
for (idx_t i = 0; i < ni; i++) {
const uint8_t * code = &codes[(i0 + i) * pq.code_size];
pq.decode (code, recons + i * d);
}
}
void IndexPQ::reconstruct (idx_t key, float * recons) const
{
FAISS_THROW_IF_NOT (key >= 0 && key < ntotal);
pq.decode (&codes[key * pq.code_size], recons);
}
/*****************************************
* IndexPQ polysemous search routines
******************************************/
void IndexPQ::search (idx_t n, const float *x, idx_t k,
float *distances, idx_t *labels) const
{
FAISS_THROW_IF_NOT (is_trained);
if (search_type == ST_PQ) { // Simple PQ search
if (metric_type == METRIC_L2) {
float_maxheap_array_t res = {
size_t(n), size_t(k), labels, distances };
pq.search (x, n, codes.data(), ntotal, &res, true);
} else {
float_minheap_array_t res = {
size_t(n), size_t(k), labels, distances };
pq.search_ip (x, n, codes.data(), ntotal, &res, true);
}
indexPQ_stats.nq += n;
indexPQ_stats.ncode += n * ntotal;
} else if (search_type == ST_polysemous ||
search_type == ST_polysemous_generalize) {
FAISS_THROW_IF_NOT (metric_type == METRIC_L2);
search_core_polysemous (n, x, k, distances, labels);
} else { // code-to-code distances
uint8_t * q_codes = new uint8_t [n * pq.code_size];
ScopeDeleter<uint8_t> del (q_codes);
if (!encode_signs) {
pq.compute_codes (x, q_codes, n);
} else {
FAISS_THROW_IF_NOT (d == pq.nbits * pq.M);
memset (q_codes, 0, n * pq.code_size);
for (size_t i = 0; i < n; i++) {
const float *xi = x + i * d;
uint8_t *code = q_codes + i * pq.code_size;
for (int j = 0; j < d; j++)
if (xi[j] > 0) code [j>>3] |= 1 << (j & 7);
}
}
if (search_type == ST_SDC) {
float_maxheap_array_t res = {
size_t(n), size_t(k), labels, distances};
pq.search_sdc (q_codes, n, codes.data(), ntotal, &res, true);
} else {
int * idistances = new int [n * k];
ScopeDeleter<int> del (idistances);
int_maxheap_array_t res = {
size_t (n), size_t (k), labels, idistances};
if (search_type == ST_HE) {
hammings_knn (&res, q_codes, codes.data(),
ntotal, pq.code_size, true);
} else if (search_type == ST_generalized_HE) {
generalized_hammings_knn (&res, q_codes, codes.data(),
ntotal, pq.code_size, true);
}
// convert distances to floats
for (int i = 0; i < k * n; i++)
distances[i] = idistances[i];
}
indexPQ_stats.nq += n;
indexPQ_stats.ncode += n * ntotal;
}
}
void IndexPQStats::reset()
{
nq = ncode = n_hamming_pass = 0;
}
IndexPQStats indexPQ_stats;
template <class HammingComputer>
static size_t polysemous_inner_loop (
const IndexPQ & index,
const float *dis_table_qi, const uint8_t *q_code,
size_t k, float *heap_dis, long *heap_ids)
{
int M = index.pq.M;
int code_size = index.pq.code_size;
int ksub = index.pq.ksub;
size_t ntotal = index.ntotal;
int ht = index.polysemous_ht;
const uint8_t *b_code = index.codes.data();
size_t n_pass_i = 0;
HammingComputer hc (q_code, code_size);
for (long bi = 0; bi < ntotal; bi++) {
int hd = hc.hamming (b_code);
if (hd < ht) {
n_pass_i ++;
float dis = 0;
const float * dis_table = dis_table_qi;
for (int m = 0; m < M; m++) {
dis += dis_table [b_code[m]];
dis_table += ksub;
}
if (dis < heap_dis[0]) {
maxheap_pop (k, heap_dis, heap_ids);
maxheap_push (k, heap_dis, heap_ids, dis, bi);
}
}
b_code += code_size;
}
return n_pass_i;
}
void IndexPQ::search_core_polysemous (idx_t n, const float *x, idx_t k,
float *distances, idx_t *labels) const
{
FAISS_THROW_IF_NOT (pq.byte_per_idx == 1);
// PQ distance tables
float * dis_tables = new float [n * pq.ksub * pq.M];
ScopeDeleter<float> del (dis_tables);
pq.compute_distance_tables (n, x, dis_tables);
// Hamming embedding queries
uint8_t * q_codes = new uint8_t [n * pq.code_size];
ScopeDeleter<uint8_t> del2 (q_codes);
if (false) {
pq.compute_codes (x, q_codes, n);
} else {
#pragma omp parallel for
for (idx_t qi = 0; qi < n; qi++) {
pq.compute_code_from_distance_table
(dis_tables + qi * pq.M * pq.ksub,
q_codes + qi * pq.code_size);
}
}
size_t n_pass = 0;
#pragma omp parallel for reduction (+: n_pass)
for (idx_t qi = 0; qi < n; qi++) {
const uint8_t * q_code = q_codes + qi * pq.code_size;
const float * dis_table_qi = dis_tables + qi * pq.M * pq.ksub;
long * heap_ids = labels + qi * k;
float *heap_dis = distances + qi * k;
maxheap_heapify (k, heap_dis, heap_ids);
if (search_type == ST_polysemous) {
switch (pq.code_size) {
case 4:
n_pass += polysemous_inner_loop<HammingComputer4>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
break;
case 8:
n_pass += polysemous_inner_loop<HammingComputer8>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
break;
case 16:
n_pass += polysemous_inner_loop<HammingComputer16>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
break;
case 32:
n_pass += polysemous_inner_loop<HammingComputer32>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
break;
case 20:
n_pass += polysemous_inner_loop<HammingComputer20>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
break;
default:
if (pq.code_size % 8 == 0) {
n_pass += polysemous_inner_loop<HammingComputerM8>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
} else if (pq.code_size % 4 == 0) {
n_pass += polysemous_inner_loop<HammingComputerM4>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
} else {
FAISS_THROW_FMT(
"code size %zd not supported for polysemous",
pq.code_size);
}
break;
}
} else {
switch (pq.code_size) {
case 8:
n_pass += polysemous_inner_loop<GenHammingComputer8>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
break;
case 16:
n_pass += polysemous_inner_loop<GenHammingComputer16>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
break;
case 32:
n_pass += polysemous_inner_loop<GenHammingComputer32>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
break;
default:
if (pq.code_size % 8 == 0) {
n_pass += polysemous_inner_loop<GenHammingComputerM8>
(*this, dis_table_qi, q_code, k, heap_dis, heap_ids);
} else {
FAISS_THROW_FMT(
"code size %zd not supported for polysemous",
pq.code_size);
}
break;
}
}
maxheap_reorder (k, heap_dis, heap_ids);
}
indexPQ_stats.nq += n;
indexPQ_stats.ncode += n * ntotal;
indexPQ_stats.n_hamming_pass += n_pass;
}
/*****************************************
* Stats of IndexPQ codes
******************************************/
void IndexPQ::hamming_distance_table (idx_t n, const float *x,
int32_t *dis) const
{
uint8_t * q_codes = new uint8_t [n * pq.code_size];
ScopeDeleter<uint8_t> del (q_codes);
pq.compute_codes (x, q_codes, n);
hammings (q_codes, codes.data(), n, ntotal, pq.code_size, dis);
}
void IndexPQ::hamming_distance_histogram (idx_t n, const float *x,
idx_t nb, const float *xb,
long *hist)
{
FAISS_THROW_IF_NOT (metric_type == METRIC_L2);
FAISS_THROW_IF_NOT (pq.code_size % 8 == 0);
FAISS_THROW_IF_NOT (pq.byte_per_idx == 1);
// Hamming embedding queries
uint8_t * q_codes = new uint8_t [n * pq.code_size];
ScopeDeleter <uint8_t> del (q_codes);
pq.compute_codes (x, q_codes, n);
uint8_t * b_codes ;
ScopeDeleter <uint8_t> del_b_codes;
if (xb) {
b_codes = new uint8_t [nb * pq.code_size];
del_b_codes.set (b_codes);
pq.compute_codes (xb, b_codes, nb);
} else {
nb = ntotal;
b_codes = codes.data();
}
int nbits = pq.M * pq.nbits;
memset (hist, 0, sizeof(*hist) * (nbits + 1));
size_t bs = 256;
#pragma omp parallel
{
std::vector<long> histi (nbits + 1);
hamdis_t *distances = new hamdis_t [nb * bs];
ScopeDeleter<hamdis_t> del (distances);
#pragma omp for
for (size_t q0 = 0; q0 < n; q0 += bs) {
// printf ("dis stats: %ld/%ld\n", q0, n);
size_t q1 = q0 + bs;
if (q1 > n) q1 = n;
hammings (q_codes + q0 * pq.code_size, b_codes,
q1 - q0, nb,
pq.code_size, distances);
for (size_t i = 0; i < nb * (q1 - q0); i++)
histi [distances [i]]++;
}
#pragma omp critical
{
for (int i = 0; i <= nbits; i++)
hist[i] += histi[i];
}
}
}
/*****************************************
* MultiIndexQuantizer
******************************************/
namespace {
template <typename T>
struct PreSortedArray {
const T * x;
int N;
explicit PreSortedArray (int N): N(N) {
}
void init (const T*x) {
this->x = x;
}
// get smallest value
T get_0 () {
return x[0];
}
// get delta between n-smallest and n-1 -smallest
T get_diff (int n) {
return x[n] - x[n - 1];
}
// remap orders counted from smallest to indices in array
int get_ord (int n) {
return n;
}
};
template <typename T>
struct ArgSort {
const T * x;
bool operator() (size_t i, size_t j) {
return x[i] < x[j];
}
};
/** Array that maintains a permutation of its elements so that the
* array's elements are sorted
*/
template <typename T>
struct SortedArray {
const T * x;
int N;
std::vector<int> perm;
explicit SortedArray (int N) {
this->N = N;
perm.resize (N);
}
void init (const T*x) {
this->x = x;
for (int n = 0; n < N; n++)
perm[n] = n;
ArgSort<T> cmp = {x };
std::sort (perm.begin(), perm.end(), cmp);
}
// get smallest value
T get_0 () {
return x[perm[0]];
}
// get delta between n-smallest and n-1 -smallest
T get_diff (int n) {
return x[perm[n]] - x[perm[n - 1]];
}
// remap orders counted from smallest to indices in array
int get_ord (int n) {
return perm[n];
}
};
/** Array has n values. Sort the k first ones and copy the other ones
* into elements k..n-1
*/
template <class C>
void partial_sort (int k, int n,
const typename C::T * vals, typename C::TI * perm) {
// insert first k elts in heap
for (int i = 1; i < k; i++) {
indirect_heap_push<C> (i + 1, vals, perm, perm[i]);
}
// insert next n - k elts in heap
for (int i = k; i < n; i++) {
typename C::TI id = perm[i];
typename C::TI top = perm[0];
if (C::cmp(vals[top], vals[id])) {
indirect_heap_pop<C> (k, vals, perm);
indirect_heap_push<C> (k, vals, perm, id);
perm[i] = top;
} else {
// nothing, elt at i is good where it is.
}
}
// order the k first elements in heap
for (int i = k - 1; i > 0; i--) {
typename C::TI top = perm[0];
indirect_heap_pop<C> (i + 1, vals, perm);
perm[i] = top;
}
}
/** same as SortedArray, but only the k first elements are sorted */
template <typename T>
struct SemiSortedArray {
const T * x;
int N;
// type of the heap: CMax = sort ascending
typedef CMax<T, int> HC;
std::vector<int> perm;
int k; // k elements are sorted
int initial_k, k_factor;
explicit SemiSortedArray (int N) {
this->N = N;
perm.resize (N);
perm.resize (N);
initial_k = 3;
k_factor = 4;
}
void init (const T*x) {
this->x = x;
for (int n = 0; n < N; n++)
perm[n] = n;
k = 0;
grow (initial_k);
}
/// grow the sorted part of the array to size next_k
void grow (int next_k) {
if (next_k < N) {
partial_sort<HC> (next_k - k, N - k, x, &perm[k]);
k = next_k;
} else { // full sort of remainder of array
ArgSort<T> cmp = {x };
std::sort (perm.begin() + k, perm.end(), cmp);
k = N;
}
}
// get smallest value
T get_0 () {
return x[perm[0]];
}
// get delta between n-smallest and n-1 -smallest
T get_diff (int n) {
if (n >= k) {
// want to keep powers of 2 - 1
int next_k = (k + 1) * k_factor - 1;
grow (next_k);
}
return x[perm[n]] - x[perm[n - 1]];
}
// remap orders counted from smallest to indices in array
int get_ord (int n) {
assert (n < k);
return perm[n];
}
};
/*****************************************
* Find the k smallest sums of M terms, where each term is taken in a
* table x of n values.
*
* A combination of terms is encoded as a scalar 0 <= t < n^M. The
* combination t0 ... t(M-1) that correspond to the sum
*
* sum = x[0, t0] + x[1, t1] + .... + x[M-1, t(M-1)]
*
* is encoded as
*
* t = t0 + t1 * n + t2 * n^2 + ... + t(M-1) * n^(M-1)
*
* MinSumK is an object rather than a function, so that storage can be
* re-used over several computations with the same sizes. use_seen is
* good when there may be ties in the x array and it is a concern if
* occasionally several t's are returned.
*
* @param x size M * n, values to add up
* @parms k nb of results to retrieve
* @param M nb of terms
* @param n nb of distinct values
* @param sums output, size k, sorted
* @prarm terms output, size k, with encoding as above
*
******************************************/
template <typename T, class SSA, bool use_seen>
struct MinSumK {
int K; ///< nb of sums to return
int M; ///< nb of elements to sum up
int nbit; ///< nb of bits to encode one entry
int N; ///< nb of possible elements for each of the M terms
/** the heap.
* We use a heap to maintain a queue of sums, with the associated
* terms involved in the sum.
*/
typedef CMin<T, long> HC;
size_t heap_capacity, heap_size;
T *bh_val;
long *bh_ids;
std::vector <SSA> ssx;
// all results get pushed several times. When there are ties, they
// are popped interleaved with others, so it is not easy to
// identify them. Therefore, this bit array just marks elements
// that were seen before.
std::vector <uint8_t> seen;
MinSumK (int K, int M, int nbit, int N):
K(K), M(M), nbit(nbit), N(N) {
heap_capacity = K * M;
assert (N <= (1 << nbit));
// we'll do k steps, each step pushes at most M vals
bh_val = new T[heap_capacity];
bh_ids = new long[heap_capacity];
if (use_seen) {
long n_ids = weight(M);
seen.resize ((n_ids + 7) / 8);
}
for (int m = 0; m < M; m++)
ssx.push_back (SSA(N));
}
long weight (int i) {
return 1 << (i * nbit);
}
bool is_seen (long i) {
return (seen[i >> 3] >> (i & 7)) & 1;
}
void mark_seen (long i) {
if (use_seen)
seen [i >> 3] |= 1 << (i & 7);
}
void run (const T *x, long ldx,
T * sums, long * terms) {
heap_size = 0;
for (int m = 0; m < M; m++) {
ssx[m].init(x);
x += ldx;
}
{ // intial result: take min for all elements
T sum = 0;
terms[0] = 0;
mark_seen (0);
for (int m = 0; m < M; m++) {
sum += ssx[m].get_0();
}
sums[0] = sum;
for (int m = 0; m < M; m++) {
heap_push<HC> (++heap_size, bh_val, bh_ids,
sum + ssx[m].get_diff(1),
weight(m));
}
}
for (int k = 1; k < K; k++) {
// pop smallest value from heap
if (use_seen) {// skip already seen elements
while (is_seen (bh_ids[0])) {
assert (heap_size > 0);
heap_pop<HC> (heap_size--, bh_val, bh_ids);
}
}
assert (heap_size > 0);
T sum = sums[k] = bh_val[0];
long ti = terms[k] = bh_ids[0];
if (use_seen) {
mark_seen (ti);
heap_pop<HC> (heap_size--, bh_val, bh_ids);
} else {
do {
heap_pop<HC> (heap_size--, bh_val, bh_ids);
} while (heap_size > 0 && bh_ids[0] == ti);
}
// enqueue followers
long ii = ti;
for (int m = 0; m < M; m++) {
long n = ii & ((1 << nbit) - 1);
ii >>= nbit;
if (n + 1 >= N) continue;
enqueue_follower (ti, m, n, sum);
}
}
/*
for (int k = 0; k < K; k++)
for (int l = k + 1; l < K; l++)
assert (terms[k] != terms[l]);
*/
// convert indices by applying permutation
for (int k = 0; k < K; k++) {
long ii = terms[k];
if (use_seen) {
// clear seen for reuse at next loop
seen[ii >> 3] = 0;
}
long ti = 0;
for (int m = 0; m < M; m++) {
long n = ii & ((1 << nbit) - 1);
ti += ssx[m].get_ord(n) << (nbit * m);
ii >>= nbit;
}
terms[k] = ti;
}
}
void enqueue_follower (long ti, int m, int n, T sum) {
T next_sum = sum + ssx[m].get_diff(n + 1);
long next_ti = ti + weight(m);
heap_push<HC> (++heap_size, bh_val, bh_ids, next_sum, next_ti);
}
~MinSumK () {
delete [] bh_ids;
delete [] bh_val;
}
};
} // anonymous namespace
MultiIndexQuantizer::MultiIndexQuantizer (int d,
size_t M,
size_t nbits):
Index(d, METRIC_L2), pq(d, M, nbits)
{
is_trained = false;
pq.verbose = verbose;
}
void MultiIndexQuantizer::train(idx_t n, const float *x)
{
pq.verbose = verbose;
pq.train (n, x);
is_trained = true;
// count virtual elements in index
ntotal = 1;
for (int m = 0; m < pq.M; m++)
ntotal *= pq.ksub;
}
void MultiIndexQuantizer::search (idx_t n, const float *x, idx_t k,
float *distances, idx_t *labels) const {
if (n == 0) return;
float * dis_tables = new float [n * pq.ksub * pq.M];
ScopeDeleter<float> del (dis_tables);
pq.compute_distance_tables (n, x, dis_tables);
if (k == 1) {
// simple version that just finds the min in each table
#pragma omp parallel for
for (int i = 0; i < n; i++) {
const float * dis_table = dis_tables + i * pq.ksub * pq.M;
float dis = 0;
idx_t label = 0;
for (int s = 0; s < pq.M; s++) {
float vmin = HUGE_VALF;
idx_t lmin = -1;
for (idx_t j = 0; j < pq.ksub; j++) {
if (dis_table[j] < vmin) {
vmin = dis_table[j];
lmin = j;
}
}
dis += vmin;
label |= lmin << (s * pq.nbits);
dis_table += pq.ksub;
}
distances [i] = dis;
labels [i] = label;
}
} else {
#pragma omp parallel if(n > 1)
{
MinSumK <float, SemiSortedArray<float>, false>
msk(k, pq.M, pq.nbits, pq.ksub);
#pragma omp for
for (int i = 0; i < n; i++) {
msk.run (dis_tables + i * pq.ksub * pq.M, pq.ksub,
distances + i * k, labels + i * k);
}
}
}
}
void MultiIndexQuantizer::reconstruct (idx_t key, float * recons) const
{
long jj = key;
for (int m = 0; m < pq.M; m++) {
long n = jj & ((1L << pq.nbits) - 1);
jj >>= pq.nbits;
memcpy(recons, pq.get_centroids(m, n), sizeof(recons[0]) * pq.dsub);
recons += pq.dsub;
}
}
void MultiIndexQuantizer::add(idx_t /*n*/, const float* /*x*/) {
FAISS_THROW_MSG(
"This index has virtual elements, "
"it does not support add");
}
void MultiIndexQuantizer::reset ()
{
FAISS_THROW_MSG ( "This index has virtual elements, "
"it does not support reset");
}
/*****************************************
* MultiIndexQuantizer2
******************************************/
MultiIndexQuantizer2::MultiIndexQuantizer2 (
int d, size_t M, size_t nbits,
Index **indexes):
MultiIndexQuantizer (d, M, nbits)
{
assign_indexes.resize (M);
for (int i = 0; i < M; i++) {
FAISS_THROW_IF_NOT_MSG(
indexes[i]->d == pq.dsub,
"Provided sub-index has incorrect size");
assign_indexes[i] = indexes[i];
}
own_fields = false;
}
MultiIndexQuantizer2::MultiIndexQuantizer2 (
int d, size_t nbits,
Index *assign_index_0,
Index *assign_index_1):
MultiIndexQuantizer (d, 2, nbits)
{
FAISS_THROW_IF_NOT_MSG(
assign_index_0->d == pq.dsub &&
assign_index_1->d == pq.dsub,
"Provided sub-index has incorrect size");
assign_indexes.resize (2);
assign_indexes [0] = assign_index_0;
assign_indexes [1] = assign_index_1;
own_fields = false;
}
void MultiIndexQuantizer2::train(idx_t n, const float* x)
{
MultiIndexQuantizer::train(n, x);
// add centroids to sub-indexes
for (int i = 0; i < pq.M; i++) {
assign_indexes[i]->add(pq.ksub, pq.get_centroids(i, 0));
}
}
void MultiIndexQuantizer2::search(
idx_t n, const float* x, idx_t K,
float* distances, idx_t* labels) const
{
if (n == 0) return;
int k2 = std::min(K, long(pq.ksub));
long M = pq.M;
long dsub = pq.dsub, ksub = pq.ksub;
// size (M, n, k2)
std::vector<idx_t> sub_ids(n * M * k2);
std::vector<float> sub_dis(n * M * k2);
std::vector<float> xsub(n * dsub);