-
Notifications
You must be signed in to change notification settings - Fork 12
/
diskann2.cpp
2131 lines (1916 loc) · 91.8 KB
/
diskann2.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
#include "diskann.h"
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>
#include <algorithm>
#include <cstdio>
#include <util/io_perf.h>
template<typename DATAT>
void train_cluster(const std::string& raw_data_bin_file,
const std::string& output_path,
const int32_t K1,
float** centroids,
double& avg_len) {
TimeRecorder rc("train cluster");
std::cout << "train_cluster parameters:" << std::endl;
std::cout << " raw_data_bin_file: " << raw_data_bin_file
<< " output path: " << output_path
<< " K1: " << K1
<< " centroids: " << *centroids
<< std::endl;
assert((*centroids) == nullptr);
DATAT* sample_data = nullptr;
uint32_t nb, dim;
get_bin_metadata(raw_data_bin_file, nb, dim);
int64_t sample_num = nb * K1_SAMPLE_RATE;
std::cout << "nb = " << nb << ", dim = " << dim << ", sample_num 4 K1: " << sample_num << std::endl;
*centroids = new float[K1 * dim];
sample_data = new DATAT[sample_num * dim];
reservoir_sampling(raw_data_bin_file, sample_num, sample_data);
rc.RecordSection("reservoir sample with sample rate: " + std::to_string(K1_SAMPLE_RATE) + " done");
double mxl, mnl;
int64_t stat_n = std::min(static_cast<int64_t>(1000000), sample_num);
stat_length<DATAT>(sample_data, stat_n, dim, mxl, mnl, avg_len);
rc.RecordSection("calculate " + std::to_string(stat_n) + " vectors from sample_data done");
std::cout << "max len: " << mxl << ", min len: " << mnl << ", average len: " << avg_len << std::endl;
kmeans<DATAT>(sample_num, sample_data, dim, K1, *centroids, false, avg_len);
rc.RecordSection("kmeans done");
assert((*centroids) != nullptr);
delete[] sample_data;
rc.ElapseFromBegin("train cluster done.");
}
template<typename DATAT, typename DISTT, typename HEAPT>
void divide_raw_data(const std::string& raw_data_bin_file,
const std::string& output_path,
const float* centroids,
const int32_t K1) {
TimeRecorder rc("divide raw data");
std::cout << "divide_raw_data parameters:" << std::endl;
std::cout << " raw_data_bin_file: " << raw_data_bin_file
<< " output_path: " << output_path
<< " centroids: " << centroids
<< " K1: " << K1
<< std::endl;
IOReader reader(raw_data_bin_file);
uint32_t nb, dim;
reader.read((char*)&nb, sizeof(uint32_t));
reader.read((char*)&dim, sizeof(uint32_t));
uint32_t placeholder = 0, const_one = 1;
std::vector<uint32_t> cluster_size(K1, 0);
std::vector<std::ofstream> cluster_dat_writer(K1);
std::vector<std::ofstream> cluster_ids_writer(K1);
for (int i = 0; i < K1; i ++) {
std::string cluster_raw_data_file_name = output_path + CLUSTER + std::to_string(i) + RAWDATA + BIN;
std::string cluster_ids_data_file_name = output_path + CLUSTER + std::to_string(i) + GLOBAL_IDS + BIN;
cluster_dat_writer[i] = std::ofstream(cluster_raw_data_file_name, std::ios::binary);
cluster_ids_writer[i] = std::ofstream(cluster_ids_data_file_name, std::ios::binary);
cluster_dat_writer[i].write((char*)&placeholder, sizeof(uint32_t));
cluster_dat_writer[i].write((char*)&dim, sizeof(uint32_t));
cluster_ids_writer[i].write((char*)&placeholder, sizeof(uint32_t));
cluster_ids_writer[i].write((char*)&const_one, sizeof(uint32_t));
}
int64_t block_size = 1000000;
assert(nb > 0);
int64_t block_num = (nb - 1) / block_size + 1;
std::vector<int64_t> cluster_id(block_size);
std::vector<DISTT> dists(block_size);
DATAT* block_buf = new DATAT[block_size * dim];
for (int64_t i = 0; i < block_num; i ++) {
TimeRecorder rci("block-" + std::to_string(i));
int64_t sp = i * block_size;
int64_t ep = std::min((int64_t)nb, sp + block_size);
std::cout << "split the " << i << "th block, start position = " << sp << ", end position = " << ep << std::endl;
reader.read((char*)block_buf, (ep - sp) * dim * sizeof(DATAT));
rci.RecordSection("read block data done");
elkan_L2_assign<const DATAT, const float, DISTT>(block_buf, centroids, dim, ep -sp, K1, cluster_id.data(), dists.data());
//knn_1<HEAPT, DATAT, float> (
// block_buf, centroids, ep - sp, K1, dim, 1,
// dists.data(), cluster_id.data(), L2sqr<const DATAT, const float, DISTT>);
rci.RecordSection("select file done");
for (int64_t j = 0; j < ep - sp; j ++) {
int64_t cid = cluster_id[j];
uint32_t uid = (uint32_t)(j + sp);
// for debug
/*
if (0 == uid) {
std::cout << "vector0 was split into cluster " << cid << std::endl;
std::cout << " show vector0:" << std::endl;
for (auto si = 0; si < dim; si ++) {
std::cout << *(block_buf + j * dim + si) << " ";
}
std::cout << std::endl;
}
*/
cluster_dat_writer[cid].write((char*)(block_buf + j * dim), sizeof(DATAT) * dim);
cluster_ids_writer[cid].write((char*)&uid, sizeof(uint32_t));
cluster_size[cid] ++;
}
rci.RecordSection("write done");
rci.ElapseFromBegin("split block " + std::to_string(i) + " done");
}
rc.RecordSection("split done");
size_t sump = 0;
std::cout << "split_raw_data done in ... seconds, show statistics:" << std::endl;
for (int i = 0; i < K1; i ++) {
uint32_t cis = cluster_size[i];
cluster_dat_writer[i].seekp(0);
cluster_dat_writer[i].write((char*)&cis, sizeof(uint32_t));
cluster_dat_writer[i].close();
cluster_ids_writer[i].seekp(0);
cluster_ids_writer[i].write((char*)&cis, sizeof(uint32_t));
cluster_ids_writer[i].close();
std::cout << "cluster-" << i << " has " << cis << " points." << std::endl;
sump += cis;
}
rc.RecordSection("rewrite header done");
std::cout << "total points num: " << sump << std::endl;
delete[] block_buf;
block_buf = nullptr;
rc.ElapseFromBegin("split_raw_data totally done");
}
template<typename DATAT, typename DISTT, typename HEAPT>
void conquer_clusters(const std::string& output_path,
const int K1, const double avg_len, const int threshold) {
TimeRecorder rc("conquer clusters");
std::cout << "conquer clusters parameters:" << std::endl;
std::cout << " output_path: " << output_path
<< " vector avg length: " << avg_len
<< std::endl;
std::vector<int64_t> cluster_id;
std::vector<DISTT> dists;
uint32_t placeholder = 1;
std::string bucket_centroids_file = output_path + BUCKET + CENTROIDS + BIN;
std::string bucket_ids_file = output_path + CLUSTER + COMBINE_IDS + BIN;
std::ofstream bucket_ids_writer(bucket_ids_file, std::ios::binary);
std::ofstream bucket_ctd_writer(bucket_centroids_file, std::ios::binary);
bucket_ctd_writer.write((char*)&placeholder, sizeof(uint32_t));
bucket_ctd_writer.write((char*)&placeholder, sizeof(uint32_t));
bucket_ids_writer.write((char*)&placeholder, sizeof(uint32_t));
bucket_ids_writer.write((char*)&placeholder, sizeof(uint32_t));
uint32_t graph_nb = 0, graph_dim;
for (int i = 0; i < K1; i ++) {
TimeRecorder rci("train-cluster-" + std::to_string(i));
std::string data_file = output_path + CLUSTER + std::to_string(i) + RAWDATA + BIN;
std::string ids_file = output_path + CLUSTER + std::to_string(i) + GLOBAL_IDS + BIN;
std::string meta_file = output_path + CLUSTER + std::to_string(i) + META + BIN;
uint32_t cluster_size, cluster_dim, ids_size, ids_dim;
IOReader data_reader(data_file);
IOReader ids_reader(ids_file);
data_reader.read((char*)&cluster_size, sizeof(uint32_t));
data_reader.read((char*)&cluster_dim, sizeof(uint32_t));
ids_reader.read((char*)&ids_size, sizeof(uint32_t));
ids_reader.read((char*)&ids_dim, sizeof(uint32_t));
assert(cluster_size == ids_size);
assert(ids_dim == 1);
DATAT* datai = new DATAT[(uint64_t)cluster_size * cluster_dim];
data_reader.read((char*)datai, (uint64_t)cluster_size * cluster_dim * sizeof(DATAT));
uint32_t* idsi = new uint32_t[(uint64_t)ids_size * ids_dim];
ids_reader.read((char*)idsi, (uint64_t)ids_size * ids_dim * sizeof(uint32_t));
int64_t K2 = (cluster_size - 1) / threshold + 1;
std::cout << "cluster-" << i << " will split into " << K2 << " buckets." << std::endl;
float* centroids_i = new float[K2 * cluster_dim];
kmeans<DATAT>(cluster_size, datai, (int32_t)cluster_dim, K2, centroids_i, false, avg_len);
rci.RecordSection("kmeans done");
cluster_id.resize(cluster_size);
dists.resize(cluster_size);
elkan_L2_assign<>(datai, centroids_i, cluster_dim, cluster_size, K2, cluster_id.data(), dists.data());
// knn_1<HEAPT, DATAT, float> (
// datai, centroids_i, cluster_size, K2, cluster_dim, 1,
// dists.data(), cluster_id.data(), L2sqr<const DATAT, const float, DISTT>);
rci.RecordSection("assign done");
std::vector<uint32_t> buckets_size(K2 + 1, 0);
std::vector<std::pair<uint32_t, uint32_t>> cluster_off;
cluster_off.resize(cluster_size);
for (int j = 0; j < cluster_size; j ++) {
buckets_size[cluster_id[j] + 1] ++;
}
{// validate bucket size
std::vector<int> empty_bkid;
for (int j = 1; j <= K2; j ++) {
assert(buckets_size[j] >= 0);
if (buckets_size[j] == 0)
empty_bkid.push_back(j - 1);
}
std::cout << "cluster-" << i << " has " << empty_bkid.size() << " empty buckets:" << std::endl;
for (int j = 0; j < empty_bkid.size(); j ++)
std::cout << empty_bkid[j] << " ";
std::cout << std::endl;
}
// write meta file
write_bin_file<uint32_t>(meta_file, &buckets_size[1], K2, 1);
rci.RecordSection("save meta into file: " + meta_file + " done");
for (int j = 1; j <= K2; j ++) {
buckets_size[j] += buckets_size[j - 1];
}
// write buckets's centroids and combine ids
// write_bin_file<float>(bucket_centroids_file, centroids_i, K2, cluster_dim);
bucket_ctd_writer.write((char*)centroids_i, K2 * cluster_dim * sizeof(float));
rci.RecordSection("append centroids_i into bucket_centroids_file");
for (int j = 0; j < K2; j ++) {
assert(buckets_size[j] <= cluster_size);
uint64_t gid = gen_id(i, j, buckets_size[j]);
bucket_ids_writer.write((char*)&gid, sizeof(uint64_t));
}
rci.RecordSection("append combine ids into bucket_ids_file");
for (int j = 0; j < cluster_size; j ++) {
cluster_off[j].first = buckets_size[cluster_id[j]] ++;
cluster_off[j].second = j;
}
std::sort(cluster_off.begin(), cluster_off.end(), [](const auto &l, const auto &r) {
return l.first < r.first;
});
rci.RecordSection("sort done");
// rewrite raw_data and global ids by bucket order
{ // data_writer will close and auto flush in de-constructor
IOWriter data_writer(data_file, MEGABYTE * 100);
IOWriter ids_writer(ids_file, MEGABYTE * 10);
data_writer.write((char*)&cluster_size, sizeof(uint32_t));
data_writer.write((char*)&cluster_dim, sizeof(uint32_t));
ids_writer.write((char*)&ids_size, sizeof(uint32_t));
ids_writer.write((char*)&ids_dim, sizeof(uint32_t));
for (int j = 0; j < cluster_size; j ++) {
uint64_t ori_pos = cluster_off[j].second;
data_writer.write((char*)(datai + ori_pos * cluster_dim), sizeof(DATAT) * cluster_dim);
ids_writer.write((char*)(idsi + ori_pos * ids_dim), sizeof(uint32_t) * ids_dim);
// for debug
/*
if (*(idsi + ori_pos * ids_dim) == 0) {
std::cout << "vector0 is arranged to new pos: " << j << " in cluster " << i << std::endl;
std::cout << " show vector0:" << std::endl;
for (auto si = 0; si < cluster_dim; si ++) {
std::cout << *(datai + ori_pos * cluster_dim + si) << " ";
}
std::cout << std::endl;
v0cid = i;
v0pos = j;
}
*/
}
}
rci.RecordSection("rearrange raw_data and global ids done");
// load buckets_size from meta_file again
// uint32_t meta_numi, meta_dimi;
// read_bin_file<uint32_t>(meta_file, buckets_size.data(), meta_numi, meta_dimi);
// assert(meta_numi == K2);
// assert(meta_dimi == 1);
graph_nb += K2;
graph_dim = cluster_dim;
delete[] datai;
delete[] idsi;
delete[] centroids_i;
rci.ElapseFromBegin("done");
rc.RecordSection("conquer the " + std::to_string(i) + "th cluster done");
}
bucket_ids_writer.close();
bucket_ctd_writer.close();
std::cout << "total bucket num = " << graph_nb << std::endl;
set_bin_metadata(bucket_centroids_file, graph_nb, graph_dim);
assert(1 == placeholder);
set_bin_metadata(bucket_ids_file, graph_nb, placeholder);
rc.ElapseFromBegin("conquer cluster totally done");
}
void build_graph(const std::string& index_path,
const int hnswM, const int hnswefC,
MetricType metric_type) {
TimeRecorder rc("create_graph_index");
std::cout << "build hnsw parameters:" << std::endl;
std::cout << " index_path: " << index_path
<< " hnsw.M: " << hnswM
<< " hnsw.efConstruction: " << hnswefC
<< " metric_type: " << (int)metric_type
<< std::endl;
float* pdata = nullptr;
uint64_t* pids = nullptr;
uint32_t npts, ndim, nids, nidsdim;
read_bin_file<float>(index_path + BUCKET + CENTROIDS + BIN, pdata, npts, ndim);
rc.RecordSection("load centroids of buckets done");
std::cout << "there are " << npts << " of dimension " << ndim << " points of hnsw" << std::endl;
assert(pdata != nullptr);
hnswlib::SpaceInterface<float>* space;
if (MetricType::L2 == metric_type) {
space = new hnswlib::L2Space(ndim);
} else if (MetricType::IP == metric_type) {
space = new hnswlib::InnerProductSpace(ndim);
} else {
std::cout << "invalid metric_type = " << (int)metric_type << std::endl;
return;
}
read_bin_file<uint64_t>(index_path + CLUSTER + COMBINE_IDS + BIN, pids, nids, nidsdim);
rc.RecordSection("load combine ids of buckets done");
std::cout << "there are " << nids << " of dimension " << nidsdim << " combine ids of hnsw" << std::endl;
assert(pids != nullptr);
assert(npts == nids);
assert(nidsdim == 1);
auto index_hnsw = std::make_shared<hnswlib::HierarchicalNSW<float>>(space, npts, hnswM, hnswefC);
index_hnsw->addPoint(pdata, pids[0]);
#pragma omp parallel for
for (int64_t i = 1; i < npts; i ++) {
index_hnsw->addPoint(pdata + i * ndim, pids[i]);
}
std::cout << "hnsw totally add " << npts << " points" << std::endl;
rc.RecordSection("create index hnsw done");
index_hnsw->saveIndex(index_path + HNSW + INDEX + BIN);
rc.RecordSection("hnsw save index done");
delete[] pdata;
pdata = nullptr;
delete[] pids;
pids = nullptr;
rc.ElapseFromBegin("create index hnsw totally done");
}
template<typename DATAT, typename DISTT, typename HEAPT>
void train_pq_quantizer(
const std::string& raw_data_bin_file,
const std::string& output_path,
const int K1,
const int PQM,
const int PQnbits,
MetricType metric_type) {
TimeRecorder rc("train pq quantizer");
std::cout << "train quantizer parameters:" << std::endl;
std::cout << " raw_data_bin_file: " << raw_data_bin_file
<< " output_path: " << output_path
<< " PQM: " << PQM
<< " PQnbits: " << PQnbits
<< std::endl;
uint32_t nb, dim;
get_bin_metadata(raw_data_bin_file, nb, dim);
int64_t pq_sample_num = 100000;
DATAT* pq_sample_data = new DATAT[pq_sample_num * dim];
reservoir_sampling(raw_data_bin_file, pq_sample_num, pq_sample_data);
rc.RecordSection("reservoir_sampling 4 pq train set done");
ProductQuantizer<HEAPT, DATAT, uint8_t> pq_quantizer(dim, PQM, PQnbits, metric_type);
pq_quantizer.train(pq_sample_num, pq_sample_data);
rc.RecordSection("pq quantizer train done");
pq_quantizer.save_centroids(output_path + PQ_CENTROIDS + BIN);
rc.RecordSection("pq quantizer save centroids done");
float* precomputer_table = nullptr;
for (int i = 0; i < K1; i ++) {
std::string data_file = output_path + CLUSTER + std::to_string(i) + RAWDATA + BIN;
std::string pq_codebook_file = output_path + CLUSTER + std::to_string(i) + PQ + CODEBOOK + BIN;
uint32_t cluster_size, cluster_dim;
IOReader data_reader(data_file);
data_reader.read((char*)&cluster_size, sizeof(uint32_t));
data_reader.read((char*)&cluster_dim, sizeof(uint32_t));
DATAT* datai = new DATAT[(uint64_t)cluster_size * cluster_dim];
data_reader.read((char*)datai, (uint64_t)cluster_size * cluster_dim * sizeof(DATAT));
pq_quantizer.encode_vectors_and_save(precomputer_table, cluster_size, datai, pq_codebook_file);
delete[] datai;
rc.RecordSection("the " + std::to_string(i) + "th cluster encode and save codebook done.");
}
delete[] pq_sample_data;
delete[] precomputer_table;
rc.ElapseFromBegin("train quantizer totally done.");
}
template<typename DATAT, typename DISTT, typename HEAPT>
void train_pq_residual_quantizer(
const std::string& raw_data_bin_file,
const std::string& output_path,
const int K1,
const int PQM,
const int PQnbits,
MetricType metric_type) {
TimeRecorder rc("train pq quantizer with residual");
std::cout << "train quantizer parameters:" << std::endl;
std::cout << " raw_data_bin_file: " << raw_data_bin_file
<< " output_path: " << output_path
<< " PQM: " << PQM
<< " PQnbits: " << PQnbits
<< std::endl;
uint32_t nb, dim;
get_bin_metadata(raw_data_bin_file, nb, dim);
std::vector<std::vector<uint32_t> > metas(K1);
load_meta_impl(output_path, metas, K1);
assert(metas.size() == K1);
int64_t pq_sample_num = std::min(
static_cast<int64_t>(65536),
std::min(
static_cast<int64_t>(nb * PQ_SAMPLE_RATE),
static_cast<int64_t>(std::accumulate(metas[0].begin(), metas[0].end(), 0))));
DATAT* sample_data = new DATAT[pq_sample_num * dim];
assert(sample_data != nullptr);
float* sample_ivf_cen = new float[pq_sample_num * dim];
assert(sample_ivf_cen != nullptr);
// read ivf centroids
float *ivf_cen = nullptr;
uint32_t ivf_n, ivf_dim;
read_bin_file<float>(output_path + BUCKET + CENTROIDS + BIN, ivf_cen, ivf_n, ivf_dim);
assert(ivf_dim == dim);
assert(ivf_cen != nullptr);
reservoir_sampling_residual<DATAT>(output_path, metas, ivf_cen, dim, pq_sample_num, sample_data, sample_ivf_cen, K1);
rc.RecordSection("reservoir_sampling_residual for pq training set done");
PQResidualQuantizer<HEAPT, DATAT, uint8_t> quantizer(dim, PQM, PQnbits, metric_type);
quantizer.train(pq_sample_num, sample_data, sample_ivf_cen);
rc.RecordSection("pq residual quantizer train done");
delete[] sample_data;
delete[] sample_ivf_cen;
quantizer.save_centroids(output_path + PQ_CENTROIDS + BIN);
rc.RecordSection("pq residual quantizer save centroids done");
float* precompute_table = nullptr;
uint64_t bucket_cnt = 0;
for (int i = 0; i < K1; ++i) {
std::string data_file = output_path + CLUSTER + std::to_string(i) + RAWDATA + BIN;
std::string pq_codebook_file = output_path + CLUSTER + std::to_string(i) + PQ + CODEBOOK + BIN;
uint32_t cluster_size, cluster_dim;
IOReader data_reader(data_file);
data_reader.read((char*)&cluster_size, sizeof(uint32_t));
data_reader.read((char*)&cluster_dim, sizeof(uint32_t));
DATAT* datai = new DATAT[cluster_size * cluster_dim];
data_reader.read((char*)datai, cluster_size * cluster_dim * sizeof(DATAT));
quantizer.encode_vectors_and_save(precompute_table, cluster_size, datai,
ivf_cen + bucket_cnt * cluster_dim, metas[i],
pq_codebook_file);
bucket_cnt += metas[i].size();
delete[] datai;
rc.RecordSection("the " + std::to_string(i) + "th cluster encode and save codebook done.");
}
delete[] precompute_table;
delete[] ivf_cen;
rc.ElapseFromBegin("train pq residual quantizer totally done.");
}
// page alignment 4 raw data and uid in each cluster 4 better refine performance, uid follow after rawdata
template<typename DATAT>
void page_align(const std::string& raw_data_bin_file, const std::string& output_path, const int K1) {
TimeRecorder rc("page align");
uint32_t nb, dim;
get_bin_metadata(raw_data_bin_file, nb, dim);
uint32_t page_size = PAGESIZE;
uint32_t vector_size = dim * sizeof(DATAT);
uint32_t id_size = sizeof(uint32_t);
uint32_t node_size = vector_size + id_size;
// number of nodes per page
uint32_t nnpp = page_size / node_size;
assert(nnpp > 0);
std::cout << "page size = " << page_size << ", number of nodes(vector + id) in per page: " << nnpp
<< std::endl;
char* data_page_buf = new char[page_size];
for (uint32_t i = 0; i < K1; i ++) {
std::string cluster_raw_data_file_name = output_path + CLUSTER + std::to_string(i) + RAWDATA + BIN;
std::string cluster_ids_data_file_name = output_path + CLUSTER + std::to_string(i) + GLOBAL_IDS + BIN;
std::string cluster_raw_data_file_name2 = output_path + CLUSTER + std::to_string(i) + "-" + RAWDATA + BIN;
uint32_t cluster_size, cluster_dim, ids_size, ids_dim;
IOReader data_reader(cluster_raw_data_file_name, (uint64_t)(2) * GIGABYTE);
IOReader ids_reader(cluster_ids_data_file_name);
IOWriter data_writer(cluster_raw_data_file_name2, (uint64_t)(2) * GIGABYTE);
data_reader.read((char*)&cluster_size, sizeof(uint32_t));
data_reader.read((char*)&cluster_dim, sizeof(uint32_t));
ids_reader.read((char*)&ids_size, sizeof(uint32_t));
ids_reader.read((char*)&ids_dim, sizeof(uint32_t));
// number of pages 4 node, the first page is meta page
uint32_t npn = (cluster_size - 1) / nnpp + 2;
std::cout << "there are " << npn << " pages 4 nodes in cluster " << std::to_string(i) << std::endl;
memset(data_page_buf, 0, page_size);
*(uint32_t*)(data_page_buf + 0 * sizeof(uint32_t)) = cluster_size;
*(uint32_t*)(data_page_buf + 1 * sizeof(uint32_t)) = cluster_dim;
*(uint32_t*)(data_page_buf + 2 * sizeof(uint32_t)) = page_size;
*(uint32_t*)(data_page_buf + 3 * sizeof(uint32_t)) = nnpp;
*(uint32_t*)(data_page_buf + 4 * sizeof(uint32_t)) = npn;
*(uint32_t*)(data_page_buf + 5 * sizeof(uint32_t)) = i;
data_writer.write(data_page_buf, page_size);
uint32_t write_cnt = 0;
for (auto j = 1; j < npn; j ++) {
memset(data_page_buf, 0, sizeof(page_size));
for (auto k = 0; k < nnpp && write_cnt < cluster_size; k ++) {
data_reader.read(data_page_buf + k * node_size, vector_size);
ids_reader.read(data_page_buf + k * node_size + vector_size, id_size);
write_cnt ++;
}
data_writer.write(data_page_buf, page_size);
}
auto rm_dat_ret = std::remove(cluster_raw_data_file_name.c_str());
if (rm_dat_ret == 0) {
std::cout << "remove old raw data file: " << cluster_raw_data_file_name << " success." << std::endl;
} else {
std::cout << "remove old raw data file: " << cluster_raw_data_file_name << " failed." << std::endl;
}
auto rm_ids_ret = std::remove(cluster_ids_data_file_name.c_str());
if (rm_ids_ret == 0) {
std::cout << "remove old ids data file: " << cluster_ids_data_file_name << " success." << std::endl;
} else {
std::cout << "remove old ids data file: " << cluster_ids_data_file_name << " failed." << std::endl;
}
rc.RecordSection("cluster " + std::to_string(i) + " page aligned done.");
}
delete[] data_page_buf;
rc.ElapseFromBegin("page align totally done.");
}
template<typename DATAT, typename DISTT, typename HEAPT>
void build_bigann(const std::string& raw_data_bin_file,
const std::string& output_path,
const int hnswM, const int hnswefC,
const int PQM, const int PQnbits,
const int K1, const int threshold,
MetricType metric_type,
QuantizerType quantizer_type) {
TimeRecorder rc("build bigann");
std::cout << "build bigann parameters:" << std::endl;
std::cout << " raw_data_bin_file: " << raw_data_bin_file
<< " output_path: " << output_path
<< " hnsw.M: " << hnswM
<< " hnsw.efConstruction: " << hnswefC
<< " PQ.M: " << PQM
<< " PQ.nbits: " << PQnbits
<< " K1: " << K1
<< " bucket split threshold: " << threshold
<< std::endl;
float* centroids = nullptr;
double avg_len;
// sampling and do K1-means to get the first round centroids
train_cluster<DATAT>(raw_data_bin_file, output_path, K1, ¢roids, avg_len);
assert(centroids != nullptr);
rc.RecordSection("train cluster to get " + std::to_string(K1) + " centroids done.");
divide_raw_data<DATAT, DISTT, HEAPT>(raw_data_bin_file, output_path, centroids, K1);
rc.RecordSection("divide raw data into " + std::to_string(K1) + " clusters done");
conquer_clusters<DATAT, DISTT, HEAPT>(output_path, K1, avg_len, threshold);
rc.RecordSection("conquer each cluster into buckets done");
build_graph(output_path, hnswM, hnswefC, metric_type);
rc.RecordSection("build hnsw done.");
if (QuantizerType::PQ == quantizer_type) {
train_pq_quantizer<DATAT, DISTT, HEAPT>(raw_data_bin_file, output_path, K1, PQM, PQnbits, metric_type);
} else if (QuantizerType::PQRES == quantizer_type) {
train_pq_residual_quantizer<DATAT, DISTT, HEAPT>(raw_data_bin_file, output_path, K1, PQM, PQnbits, metric_type);
}
rc.RecordSection("train quantizer done.");
page_align<DATAT>(raw_data_bin_file, output_path, K1);
rc.RecordSection("page align done.");
delete[] centroids;
rc.ElapseFromBegin("build bigann totally done.");
}
void load_pq_codebook(const std::string& index_path,
std::vector<std::vector<uint8_t>>& pq_codebook,
const int K1) {
TimeRecorder rc("load pq codebook");
std::cout << "load pq codebook parameters:" << std::endl;
std::cout << " index path: " << index_path
<< " K1: " << K1
<< std::endl;
for (int i = 0; i < K1; i ++) {
std::ifstream reader(index_path + CLUSTER + std::to_string(i) + PQ + CODEBOOK + BIN, std::ios::binary);
uint32_t sizei, dimi;
reader.read((char*)&sizei, sizeof(uint32_t));
reader.read((char*)&dimi, sizeof(uint32_t));
pq_codebook[i].resize((uint64_t)sizei * dimi);
reader.read((char*)pq_codebook[i].data(), (uint64_t)sizei * dimi * sizeof(uint8_t));
reader.close();
}
rc.ElapseFromBegin("load pq codebook done.");
}
void load_meta(const std::string& index_path,
std::vector<std::vector<uint32_t>>& meta,
const int K1) {
TimeRecorder rc("load meta");
std::cout << "load meta parameters:" << std::endl;
std::cout << " index path: " << index_path
<< " K1: " << K1
<< std::endl;
load_meta_impl(index_path, meta, K1);
rc.ElapseFromBegin("load meta done.");
}
template<typename DATAT>
void search_graph(std::shared_ptr<hnswlib::HierarchicalNSW<float>> index_hnsw,
const int nq,
const int dq,
const int nprobe,
const int refine_nprobe,
const DATAT* pquery,
uint64_t* buckets_label) {
TimeRecorder rc("search graph");
std::cout << "search graph parameters:" << std::endl;
std::cout << " index_hnsw: " << index_hnsw
<< " nq: " << nq
<< " dq: " << dq
<< " nprobe: " << nprobe
<< " refine_nprobe: " << refine_nprobe
<< " pquery: " << static_cast<const void *>(pquery)
<< " buckets_label: " << static_cast<void *>(buckets_label)
<< std::endl;
index_hnsw->setEf(refine_nprobe);
#pragma omp parallel for
for (int64_t i = 0; i < nq; i ++) {
// auto queryi = pquery + i * dq;
// todo: hnsw need to support query data is not float
float* queryi = new float[dq];
for (int j = 0; j < dq;j ++)
queryi[j] = (float)(*(pquery + i * dq + j));
auto reti = index_hnsw->searchKnn(queryi, nprobe);
auto p_labeli = buckets_label + i * nprobe;
while (!reti.empty()) {
*p_labeli++ = reti.top().second;
reti.pop();
}
delete[] queryi;
}
rc.ElapseFromBegin("search graph done.");
}
template<typename DATAT, typename DISTT, typename HEAPTT>
void search_pq_quantizer(ProductQuantizer<HEAPTT, DATAT, uint8_t>& pq_quantizer,
const uint32_t nq,
const uint32_t dq,
uint64_t* buckets_label,
const int nprobe,
const int refine_topk,
const int K1,
const DATAT* pquery,
std::vector<std::vector<uint8_t>>& pq_codebook,
std::vector<std::vector<uint32_t>>& meta,
DISTT*& pq_distance,
uint64_t*& pq_offsets) {
TimeRecorder rc("search quantizer");
std::cout << "search quantizer parameters:" << std::endl;
std::cout << " pq_quantizer:" << &pq_quantizer
<< " nq: " << nq
<< " dq: " << dq
<< " buckets_label: " << buckets_label
<< " nprobe: " << nprobe
<< " refine_topk: " << refine_topk
<< " K1: " << K1
<< " pquery: " << static_cast<const void *>(pquery)
<< " pq_distance: " << static_cast<void *>(pq_distance)
<< " pq_offsets: " << static_cast<void *>(pq_offsets)
<< std::endl;
auto pqm = pq_quantizer.getM();
rc.RecordSection("pqm = " + std::to_string(pqm));
#pragma omp parallel
{
float* precompute_table = nullptr;
#pragma omp for
for (int i = 0; i < nq; i ++) {
pq_quantizer.calc_precompute_table(precompute_table, pquery + i * dq);
auto p_labeli = buckets_label + i * nprobe;
auto pq_offseti = pq_offsets + i * refine_topk;
auto pq_distancei = pq_distance + i * refine_topk;
uint32_t cid, bid, off;
for (int j = 0; j < nprobe; j ++) {
parse_id(p_labeli[j], cid, bid, off);
assert(cid < K1);
pq_quantizer.search(precompute_table, pquery + i * dq,
pq_codebook[cid].data() + (int64_t)off * pqm, meta[cid][bid],
refine_topk, pq_distancei, pq_offseti,
j + 1 == nprobe, j == 0, cid, off, i);
}
}
delete[] precompute_table;
}
/*
#pragma omp parallel for
for (auto i = 0; i < nq; i ++) {
ProductQuantizer<HEAPTT, DATAT, uint8_t> pq_quantizer_copiesi(pq_quantizer);
pq_quantizer_copiesi.cal_precompute_table(pquery + i * dq, pq_cmp);
auto p_labeli = buckets_label + i * nprobe;
auto pq_offseti = pq_offsets + i * refine_topk;
auto pq_distancei = pq_distance + i * refine_topk;
uint32_t cid, bid, off;
for (auto j = 0; j < nprobe; j ++) {
parse_id(p_labeli[j], cid, bid, off);
assert(cid < K1);
pq_quantizer_copiesi.search(pquery + i * dq,
pq_codebook[cid].data() + off * pqm, meta[cid][bid],
refine_topk, pq_distancei, pq_offseti, pq_cmp,
j + 1 == nprobe, j == 0, cid, off, i);
}
pq_quantizer_copiesi.reset();
}
*/
rc.ElapseFromBegin("search quantizer done.");
}
template<typename DATAT, typename DISTT, typename HEAPTT>
void search_pq_residual_quantizer(
PQResidualQuantizer<HEAPTT, DATAT, uint8_t>& quantizer,
const uint32_t nq,
const uint32_t dq,
float* ivf_centroids,
uint64_t* buckets_label,
const int nprobe,
const int refine_topk,
const int K1,
const DATAT* pquery,
std::vector<std::vector<uint8_t>>& pq_codebook,
std::vector<std::vector<uint32_t>>& meta,
DISTT*& pq_distance,
uint64_t*& pq_offsets) {
TimeRecorder rc("search pq residual quantizer");
std::cout << "search quantizer parameters:" << std::endl;
std::cout << " quantizer:" << &quantizer
<< " nq: " << nq
<< " dq: " << dq
<< " buckets_label: " << buckets_label
<< " nprobe: " << nprobe
<< " refine_topk: " << refine_topk
<< " K1: " << K1
<< " pquery: " << static_cast<const void *>(pquery)
<< " pq_distance: " << static_cast<void *>(pq_distance)
<< " pq_offsets: " << static_cast<void *>(pq_offsets)
<< std::endl;
auto code_num = quantizer.getCodeNum();
rc.RecordSection("pq code num = " + std::to_string(code_num));
std::vector<uint32_t> pre(K1);
for (int i = 0; i < K1; ++i) {
pre[i] = (i == 0 ? 0 : pre[i-1] + meta[i-1].size());
}
#pragma omp parallel
{
float* precompute_table = nullptr;
#pragma omp for
for (int64_t i = 0; i < nq; i ++) {
quantizer.calc_precompute_table(precompute_table, pquery + i * dq);
auto p_labeli = buckets_label + i * nprobe;
auto pq_offseti = pq_offsets + i * refine_topk;
auto pq_distancei = pq_distance + i * refine_topk;
uint32_t cid, bid, off;
for (auto j = 0; j < nprobe; j ++) {
parse_id(p_labeli[j], cid, bid, off);
assert(cid < K1);
const float* cen = ivf_centroids + (pre[cid] + bid) * dq;
quantizer.search(
precompute_table,
pquery + i * dq,
cen,
pq_codebook[cid].data() + (uint64_t)off * code_num,
meta[cid][bid],
refine_topk,
pq_distancei,
pq_offseti,
j + 1 == nprobe,
j == 0,
cid,
off,
i);
}
}
delete[] precompute_table;
}
/*
#pragma omp parallel for
for (auto i = 0; i < nq; i ++) {
ProductQuantizer<HEAPTT, DATAT, uint8_t> pq_quantizer_copiesi(pq_quantizer);
pq_quantizer_copiesi.cal_precompute_table(pquery + i * dq, pq_cmp);
auto p_labeli = buckets_label + i * nprobe;
auto pq_offseti = pq_offsets + i * refine_topk;
auto pq_distancei = pq_distance + i * refine_topk;
uint32_t cid, bid, off;
for (auto j = 0; j < nprobe; j ++) {
parse_id(p_labeli[j], cid, bid, off);
assert(cid < K1);
pq_quantizer_copiesi.search(pquery + i * dq,
pq_codebook[cid].data() + off * pqm, meta[cid][bid],
refine_topk, pq_distancei, pq_offseti, pq_cmp,
j + 1 == nprobe, j == 0, cid, off, i);
}
pq_quantizer_copiesi.reset();
}
*/
rc.ElapseFromBegin("search quantizer done.");
}
// cannot work
template<typename DATAT, typename DISTT, typename HEAPT>
void refine(const std::string& index_path,
const int K1,
const uint32_t nq,
const uint32_t dq,
const int topk,
const int refine_topk,
uint64_t* pq_offsets,
const DATAT* pquery,
DISTT*& answer_dists,
uint32_t*& answer_ids,
Computer<DATAT, DATAT, DISTT>& dis_computer) {
TimeRecorder rc("refine");
std::cout << "refine parameters:" << std::endl;
std::cout << " index_path: " << index_path
<< " cluster size: " << K1
<< " number of query: " << nq
<< " dim of query: " << dq
<< " topk: " << topk
<< " refine_topk:" << refine_topk
<< " pq_offsets:" << static_cast<void *>(pq_offsets)
<< " pquery:" << static_cast<const void *>(pquery)
<< " answer_dists:" << static_cast<void *>(answer_dists)
<< " answer_ids:" << static_cast<void *>(answer_ids)
<< std::endl;
std::vector<std::vector<std::pair<uint32_t, uint32_t>>> refine_records(K1);
for (int64_t i = 0; i < nq; i ++) {
auto pq_offseti = pq_offsets + i * refine_topk;
for (int j = 0; j < refine_topk; j ++) {
if (pq_offseti[j] == (uint64_t)(-1))
continue;
uint32_t cid, off, qid;
parse_refine_id(pq_offseti[j], cid, off, qid);
refine_records[cid].emplace_back(off, qid);
}
}
rc.RecordSection("parse_refine_id done");
std::vector<std::ifstream> raw_data_file_handlers(K1);
std::vector<std::ifstream> ids_data_file_handlers(K1);
for (int i = 0; i < K1; i ++) {
std::string data_filei = index_path + CLUSTER + std::to_string(i) + RAWDATA + BIN;
std::string ids_filei = index_path + CLUSTER + std::to_string(i) + GLOBAL_IDS + BIN;
raw_data_file_handlers[i] = std::ifstream(data_filei, std::ios::binary);
ids_data_file_handlers[i] = std::ifstream(ids_filei , std::ios::binary);
uint32_t clu_size, clu_dim, clu_id_size, clu_id_dim;
raw_data_file_handlers[i].read((char*)&clu_size, sizeof(clu_size));
raw_data_file_handlers[i].read((char*)&clu_dim, sizeof(clu_dim));
ids_data_file_handlers[i].read((char*)&clu_id_size, sizeof(clu_id_size));
ids_data_file_handlers[i].read((char*)&clu_id_dim, sizeof(clu_id_dim));
std::cout << "cluster-" << i << " has " << clu_size << " vectors,"
<< " has clu_dim = " << clu_dim
<< " clu_id_size = " << clu_id_size
<< " clu_id_dim = " << clu_id_dim
<< std::endl;
}
rc.RecordSection("open rawdata and idsdata file handlers");
// init answer heap
#pragma omp parallel for schedule (static, 128)
for (int i = 0; i < nq; i ++) {
auto ans_disi = answer_dists + topk * i;
auto ans_idsi = answer_ids + topk * i;
heap_heapify<HEAPT>(topk, ans_disi, ans_idsi);
}
rc.RecordSection("heapify answers heaps");
// statistics
std::vector<int> load_vectors(K1, 0);
std::vector<std::mutex> mtx(nq);
#pragma omp parallel for
for (int i = 0; i < K1; i ++) {
if (refine_records[i].size() == 0)
continue;
std::sort(refine_records[i].begin(), refine_records[i].end(), [](const auto &l, const auto &r) {
return l.first < r.first;
});
uint64_t pre_off = refine_records[i][0].first;
uint32_t meta_bytes = 8; // pass meta
DATAT* data_bufi = new DATAT[dq];
uint32_t global_id;
raw_data_file_handlers[i].seekg(meta_bytes + pre_off * dq * sizeof(DATAT), raw_data_file_handlers[i].beg);
raw_data_file_handlers[i].read((char*)data_bufi, dq * sizeof(DATAT));
ids_data_file_handlers[i].seekg(meta_bytes + pre_off * sizeof(uint32_t), ids_data_file_handlers[i].beg);
ids_data_file_handlers[i].read((char*)&global_id, sizeof(uint32_t));
load_vectors[i] ++;
// for debug
for (int j = 0; j < refine_records[i].size(); j ++) {
if (refine_records[i][j].first != pre_off) {
pre_off = refine_records[i][j].first;
raw_data_file_handlers[i].seekg(meta_bytes + pre_off * dq * sizeof(DATAT), raw_data_file_handlers[i].beg);
raw_data_file_handlers[i].read((char*)data_bufi, dq * sizeof(DATAT));
ids_data_file_handlers[i].seekg(meta_bytes + pre_off * sizeof(uint32_t), ids_data_file_handlers[i].beg);
ids_data_file_handlers[i].read((char*)&global_id, sizeof(uint32_t));
assert(global_id >= 0);
load_vectors[i] ++;
}
uint32_t qid = refine_records[i][j].second;
auto dis = dis_computer(data_bufi, pquery + qid * dq, dq);
std::unique_lock<std::mutex> lk(mtx[qid]);
if (HEAPT::cmp(answer_dists[topk * qid], dis)) {
heap_swap_top<HEAPT>(topk, answer_dists + topk * qid, answer_ids + topk * qid, dis, global_id);
}
}
delete[] data_bufi;
}
rc.RecordSection("calculate done.");
int tot = 0;
std::cout << "show load refine vectors of each cluster:" << std::endl;
for (int i = 0; i < K1; i ++) {
std::cout << "cluster-" << i << ": " << load_vectors[i] << "/" << refine_topk << std::endl;
tot += load_vectors[i];
}
std::cout << "total load refine vectors: " << tot << "/" << refine_topk * nq << std::endl;
for (int i = 0; i < K1; i ++) {
std::string data_filei = index_path + CLUSTER + std::to_string(i) + RAWDATA + BIN;
std::string ids_filei = index_path + CLUSTER + std::to_string(i) + GLOBAL_IDS + BIN;
raw_data_file_handlers[i].close();
ids_data_file_handlers[i].close();
}
rc.ElapseFromBegin("refine done.");
}
template<typename DATAT, typename DISTT, typename HEAPT>
void aligned_refine(const std::string& index_path,
const int K1,
const uint32_t nq,
const uint32_t dq,
const int topk,
const int refine_topk,
uint64_t* pq_offsets,
const DATAT* pquery,
DISTT*& answer_dists,
uint32_t*& answer_ids,
Computer<DATAT, DATAT, DISTT>& dis_computer) {
TimeRecorder rc("aligned refine");
std::cout << "aligned refine parameters:" << std::endl;
std::cout << " index_path: " << index_path
<< " cluster size: " << K1
<< " number of query: " << nq
<< " dim of query: " << dq
<< " topk: " << topk
<< " refine_topk:" << refine_topk
<< " pq_offsets:" << pq_offsets
<< " pquery:" << static_cast<const void *>(pquery)
<< " answer_dists:" << static_cast<void *>(answer_dists)