faiss 0.1.3 → 0.1.4
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +4 -0
- data/LICENSE.txt +1 -1
- data/README.md +1 -1
- data/ext/faiss/extconf.rb +1 -1
- data/lib/faiss/version.rb +1 -1
- data/vendor/faiss/faiss/AutoTune.cpp +36 -33
- data/vendor/faiss/faiss/AutoTune.h +6 -3
- data/vendor/faiss/faiss/Clustering.cpp +16 -12
- data/vendor/faiss/faiss/Index.cpp +3 -4
- data/vendor/faiss/faiss/Index.h +3 -3
- data/vendor/faiss/faiss/IndexBinary.cpp +3 -4
- data/vendor/faiss/faiss/IndexBinary.h +1 -1
- data/vendor/faiss/faiss/IndexBinaryHash.cpp +2 -12
- data/vendor/faiss/faiss/IndexBinaryIVF.cpp +1 -2
- data/vendor/faiss/faiss/IndexFlat.cpp +0 -148
- data/vendor/faiss/faiss/IndexFlat.h +0 -51
- data/vendor/faiss/faiss/IndexHNSW.cpp +4 -5
- data/vendor/faiss/faiss/IndexIVF.cpp +118 -31
- data/vendor/faiss/faiss/IndexIVF.h +22 -15
- data/vendor/faiss/faiss/IndexIVFFlat.cpp +3 -3
- data/vendor/faiss/faiss/IndexIVFFlat.h +2 -1
- data/vendor/faiss/faiss/IndexIVFPQ.cpp +39 -15
- data/vendor/faiss/faiss/IndexIVFPQ.h +25 -9
- data/vendor/faiss/faiss/IndexIVFPQFastScan.cpp +1116 -0
- data/vendor/faiss/faiss/IndexIVFPQFastScan.h +166 -0
- data/vendor/faiss/faiss/IndexIVFPQR.cpp +8 -9
- data/vendor/faiss/faiss/IndexIVFPQR.h +2 -1
- data/vendor/faiss/faiss/IndexIVFSpectralHash.cpp +1 -2
- data/vendor/faiss/faiss/IndexPQ.cpp +34 -18
- data/vendor/faiss/faiss/IndexPQFastScan.cpp +536 -0
- data/vendor/faiss/faiss/IndexPQFastScan.h +111 -0
- data/vendor/faiss/faiss/IndexPreTransform.cpp +47 -0
- data/vendor/faiss/faiss/IndexPreTransform.h +2 -0
- data/vendor/faiss/faiss/IndexRefine.cpp +256 -0
- data/vendor/faiss/faiss/IndexRefine.h +73 -0
- data/vendor/faiss/faiss/IndexScalarQuantizer.cpp +2 -2
- data/vendor/faiss/faiss/IndexScalarQuantizer.h +1 -1
- data/vendor/faiss/faiss/gpu/GpuDistance.h +1 -1
- data/vendor/faiss/faiss/gpu/GpuIndex.h +16 -9
- data/vendor/faiss/faiss/gpu/GpuIndexBinaryFlat.h +8 -1
- data/vendor/faiss/faiss/gpu/GpuIndexFlat.h +11 -11
- data/vendor/faiss/faiss/gpu/GpuIndexIVF.h +19 -2
- data/vendor/faiss/faiss/gpu/GpuIndexIVFFlat.h +28 -2
- data/vendor/faiss/faiss/gpu/GpuIndexIVFPQ.h +24 -14
- data/vendor/faiss/faiss/gpu/GpuIndexIVFScalarQuantizer.h +29 -2
- data/vendor/faiss/faiss/gpu/GpuResources.h +4 -0
- data/vendor/faiss/faiss/gpu/StandardGpuResources.cpp +60 -27
- data/vendor/faiss/faiss/gpu/StandardGpuResources.h +28 -6
- data/vendor/faiss/faiss/gpu/impl/InterleavedCodes.cpp +547 -0
- data/vendor/faiss/faiss/gpu/impl/InterleavedCodes.h +51 -0
- data/vendor/faiss/faiss/gpu/impl/RemapIndices.cpp +3 -3
- data/vendor/faiss/faiss/gpu/impl/RemapIndices.h +3 -2
- data/vendor/faiss/faiss/gpu/test/TestCodePacking.cpp +274 -0
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexIVFFlat.cpp +7 -2
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexIVFPQ.cpp +5 -1
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexIVFScalarQuantizer.cpp +231 -0
- data/vendor/faiss/faiss/gpu/test/TestUtils.h +33 -0
- data/vendor/faiss/faiss/gpu/utils/StackDeviceMemory.cpp +1 -0
- data/vendor/faiss/faiss/gpu/utils/StaticUtils.h +6 -0
- data/vendor/faiss/faiss/gpu/utils/Timer.cpp +5 -6
- data/vendor/faiss/faiss/gpu/utils/Timer.h +2 -2
- data/vendor/faiss/faiss/impl/AuxIndexStructures.h +5 -4
- data/vendor/faiss/faiss/impl/HNSW.cpp +2 -4
- data/vendor/faiss/faiss/impl/PolysemousTraining.h +4 -4
- data/vendor/faiss/faiss/impl/ProductQuantizer.cpp +22 -12
- data/vendor/faiss/faiss/impl/ProductQuantizer.h +2 -0
- data/vendor/faiss/faiss/impl/ResultHandler.h +452 -0
- data/vendor/faiss/faiss/impl/ScalarQuantizer.cpp +29 -19
- data/vendor/faiss/faiss/impl/ScalarQuantizer.h +6 -0
- data/vendor/faiss/faiss/impl/index_read.cpp +64 -96
- data/vendor/faiss/faiss/impl/index_write.cpp +34 -25
- data/vendor/faiss/faiss/impl/io.cpp +33 -2
- data/vendor/faiss/faiss/impl/io.h +7 -2
- data/vendor/faiss/faiss/impl/lattice_Zn.cpp +1 -15
- data/vendor/faiss/faiss/impl/platform_macros.h +44 -0
- data/vendor/faiss/faiss/impl/pq4_fast_scan.cpp +272 -0
- data/vendor/faiss/faiss/impl/pq4_fast_scan.h +169 -0
- data/vendor/faiss/faiss/impl/pq4_fast_scan_search_1.cpp +180 -0
- data/vendor/faiss/faiss/impl/pq4_fast_scan_search_qbs.cpp +354 -0
- data/vendor/faiss/faiss/impl/simd_result_handlers.h +559 -0
- data/vendor/faiss/faiss/index_factory.cpp +112 -7
- data/vendor/faiss/faiss/index_io.h +1 -48
- data/vendor/faiss/faiss/invlists/BlockInvertedLists.cpp +151 -0
- data/vendor/faiss/faiss/invlists/BlockInvertedLists.h +76 -0
- data/vendor/faiss/faiss/{DirectMap.cpp → invlists/DirectMap.cpp} +1 -1
- data/vendor/faiss/faiss/{DirectMap.h → invlists/DirectMap.h} +1 -1
- data/vendor/faiss/faiss/{InvertedLists.cpp → invlists/InvertedLists.cpp} +72 -1
- data/vendor/faiss/faiss/{InvertedLists.h → invlists/InvertedLists.h} +32 -1
- data/vendor/faiss/faiss/invlists/InvertedListsIOHook.cpp +107 -0
- data/vendor/faiss/faiss/invlists/InvertedListsIOHook.h +63 -0
- data/vendor/faiss/faiss/{OnDiskInvertedLists.cpp → invlists/OnDiskInvertedLists.cpp} +21 -6
- data/vendor/faiss/faiss/{OnDiskInvertedLists.h → invlists/OnDiskInvertedLists.h} +5 -2
- data/vendor/faiss/faiss/python/python_callbacks.h +8 -1
- data/vendor/faiss/faiss/utils/AlignedTable.h +141 -0
- data/vendor/faiss/faiss/utils/Heap.cpp +2 -4
- data/vendor/faiss/faiss/utils/Heap.h +61 -50
- data/vendor/faiss/faiss/utils/distances.cpp +164 -319
- data/vendor/faiss/faiss/utils/distances.h +28 -20
- data/vendor/faiss/faiss/utils/distances_simd.cpp +277 -49
- data/vendor/faiss/faiss/utils/extra_distances.cpp +1 -2
- data/vendor/faiss/faiss/utils/hamming-inl.h +4 -4
- data/vendor/faiss/faiss/utils/hamming.cpp +3 -6
- data/vendor/faiss/faiss/utils/hamming.h +2 -7
- data/vendor/faiss/faiss/utils/ordered_key_value.h +98 -0
- data/vendor/faiss/faiss/utils/partitioning.cpp +1256 -0
- data/vendor/faiss/faiss/utils/partitioning.h +69 -0
- data/vendor/faiss/faiss/utils/quantize_lut.cpp +277 -0
- data/vendor/faiss/faiss/utils/quantize_lut.h +80 -0
- data/vendor/faiss/faiss/utils/simdlib.h +31 -0
- data/vendor/faiss/faiss/utils/simdlib_avx2.h +461 -0
- data/vendor/faiss/faiss/utils/simdlib_emulated.h +589 -0
- metadata +43 -141
- data/vendor/faiss/benchs/bench_6bit_codec.cpp +0 -80
- data/vendor/faiss/c_api/AutoTune_c.cpp +0 -83
- data/vendor/faiss/c_api/AutoTune_c.h +0 -66
- data/vendor/faiss/c_api/Clustering_c.cpp +0 -145
- data/vendor/faiss/c_api/Clustering_c.h +0 -123
- data/vendor/faiss/c_api/IndexFlat_c.cpp +0 -140
- data/vendor/faiss/c_api/IndexFlat_c.h +0 -115
- data/vendor/faiss/c_api/IndexIVFFlat_c.cpp +0 -64
- data/vendor/faiss/c_api/IndexIVFFlat_c.h +0 -58
- data/vendor/faiss/c_api/IndexIVF_c.cpp +0 -99
- data/vendor/faiss/c_api/IndexIVF_c.h +0 -142
- data/vendor/faiss/c_api/IndexLSH_c.cpp +0 -37
- data/vendor/faiss/c_api/IndexLSH_c.h +0 -40
- data/vendor/faiss/c_api/IndexPreTransform_c.cpp +0 -21
- data/vendor/faiss/c_api/IndexPreTransform_c.h +0 -32
- data/vendor/faiss/c_api/IndexShards_c.cpp +0 -38
- data/vendor/faiss/c_api/IndexShards_c.h +0 -39
- data/vendor/faiss/c_api/Index_c.cpp +0 -105
- data/vendor/faiss/c_api/Index_c.h +0 -183
- data/vendor/faiss/c_api/MetaIndexes_c.cpp +0 -49
- data/vendor/faiss/c_api/MetaIndexes_c.h +0 -49
- data/vendor/faiss/c_api/clone_index_c.cpp +0 -23
- data/vendor/faiss/c_api/clone_index_c.h +0 -32
- data/vendor/faiss/c_api/error_c.h +0 -42
- data/vendor/faiss/c_api/error_impl.cpp +0 -27
- data/vendor/faiss/c_api/error_impl.h +0 -16
- data/vendor/faiss/c_api/faiss_c.h +0 -58
- data/vendor/faiss/c_api/gpu/GpuAutoTune_c.cpp +0 -98
- data/vendor/faiss/c_api/gpu/GpuAutoTune_c.h +0 -56
- data/vendor/faiss/c_api/gpu/GpuClonerOptions_c.cpp +0 -52
- data/vendor/faiss/c_api/gpu/GpuClonerOptions_c.h +0 -68
- data/vendor/faiss/c_api/gpu/GpuIndex_c.cpp +0 -17
- data/vendor/faiss/c_api/gpu/GpuIndex_c.h +0 -30
- data/vendor/faiss/c_api/gpu/GpuIndicesOptions_c.h +0 -38
- data/vendor/faiss/c_api/gpu/GpuResources_c.cpp +0 -86
- data/vendor/faiss/c_api/gpu/GpuResources_c.h +0 -66
- data/vendor/faiss/c_api/gpu/StandardGpuResources_c.cpp +0 -54
- data/vendor/faiss/c_api/gpu/StandardGpuResources_c.h +0 -53
- data/vendor/faiss/c_api/gpu/macros_impl.h +0 -42
- data/vendor/faiss/c_api/impl/AuxIndexStructures_c.cpp +0 -220
- data/vendor/faiss/c_api/impl/AuxIndexStructures_c.h +0 -149
- data/vendor/faiss/c_api/index_factory_c.cpp +0 -26
- data/vendor/faiss/c_api/index_factory_c.h +0 -30
- data/vendor/faiss/c_api/index_io_c.cpp +0 -42
- data/vendor/faiss/c_api/index_io_c.h +0 -50
- data/vendor/faiss/c_api/macros_impl.h +0 -110
- data/vendor/faiss/demos/demo_imi_flat.cpp +0 -154
- data/vendor/faiss/demos/demo_imi_pq.cpp +0 -203
- data/vendor/faiss/demos/demo_ivfpq_indexing.cpp +0 -151
- data/vendor/faiss/demos/demo_sift1M.cpp +0 -252
- data/vendor/faiss/demos/demo_weighted_kmeans.cpp +0 -185
- data/vendor/faiss/misc/test_blas.cpp +0 -87
- data/vendor/faiss/tests/test_binary_flat.cpp +0 -62
- data/vendor/faiss/tests/test_dealloc_invlists.cpp +0 -188
- data/vendor/faiss/tests/test_ivfpq_codec.cpp +0 -70
- data/vendor/faiss/tests/test_ivfpq_indexing.cpp +0 -100
- data/vendor/faiss/tests/test_lowlevel_ivf.cpp +0 -573
- data/vendor/faiss/tests/test_merge.cpp +0 -260
- data/vendor/faiss/tests/test_omp_threads.cpp +0 -14
- data/vendor/faiss/tests/test_ondisk_ivf.cpp +0 -225
- data/vendor/faiss/tests/test_pairs_decoding.cpp +0 -193
- data/vendor/faiss/tests/test_params_override.cpp +0 -236
- data/vendor/faiss/tests/test_pq_encoding.cpp +0 -98
- data/vendor/faiss/tests/test_sliding_ivf.cpp +0 -246
- data/vendor/faiss/tests/test_threaded_index.cpp +0 -253
- data/vendor/faiss/tests/test_transfer_invlists.cpp +0 -159
- data/vendor/faiss/tutorial/cpp/1-Flat.cpp +0 -104
- data/vendor/faiss/tutorial/cpp/2-IVFFlat.cpp +0 -85
- data/vendor/faiss/tutorial/cpp/3-IVFPQ.cpp +0 -98
- data/vendor/faiss/tutorial/cpp/4-GPU.cpp +0 -122
- data/vendor/faiss/tutorial/cpp/5-Multiple-GPUs.cpp +0 -104
@@ -371,7 +371,7 @@ void IndexIVFPQ::reconstruct_from_offset (int64_t list_no, int64_t offset,
|
|
371
371
|
|
372
372
|
|
373
373
|
/// 2G by default, accommodates tables up to PQ32 w/ 65536 centroids
|
374
|
-
size_t
|
374
|
+
size_t precomputed_table_max_bytes = ((size_t)1) << 31;
|
375
375
|
|
376
376
|
/** Precomputed tables for residuals
|
377
377
|
*
|
@@ -403,10 +403,22 @@ size_t IndexIVFPQ::precomputed_table_max_bytes = ((size_t)1) << 31;
|
|
403
403
|
* is faster when the length of the lists is > ksub * M.
|
404
404
|
*/
|
405
405
|
|
406
|
-
void
|
406
|
+
void initialize_IVFPQ_precomputed_table (
|
407
|
+
int &use_precomputed_table,
|
408
|
+
const Index *quantizer,
|
409
|
+
const ProductQuantizer &pq,
|
410
|
+
AlignedTable<float> & precomputed_table,
|
411
|
+
bool verbose
|
412
|
+
)
|
407
413
|
{
|
408
|
-
|
414
|
+
size_t nlist = quantizer->ntotal;
|
415
|
+
size_t d = quantizer->d;
|
416
|
+
FAISS_THROW_IF_NOT(d == pq.d);
|
417
|
+
|
418
|
+
if (use_precomputed_table == -1) {
|
419
|
+
precomputed_table.resize (0);
|
409
420
|
return;
|
421
|
+
}
|
410
422
|
|
411
423
|
if (use_precomputed_table == 0) { // then choose the type of table
|
412
424
|
if (quantizer->metric_type == METRIC_INNER_PRODUCT) {
|
@@ -414,6 +426,7 @@ void IndexIVFPQ::precompute_table ()
|
|
414
426
|
printf("IndexIVFPQ::precompute_table: precomputed "
|
415
427
|
"tables not needed for inner product quantizers\n");
|
416
428
|
}
|
429
|
+
precomputed_table.resize (0);
|
417
430
|
return;
|
418
431
|
}
|
419
432
|
const MultiIndexQuantizer *miq =
|
@@ -492,6 +505,16 @@ void IndexIVFPQ::precompute_table ()
|
|
492
505
|
|
493
506
|
}
|
494
507
|
|
508
|
+
void IndexIVFPQ::precompute_table ()
|
509
|
+
{
|
510
|
+
initialize_IVFPQ_precomputed_table (
|
511
|
+
use_precomputed_table, quantizer, pq, precomputed_table,
|
512
|
+
verbose
|
513
|
+
);
|
514
|
+
}
|
515
|
+
|
516
|
+
|
517
|
+
|
495
518
|
namespace {
|
496
519
|
|
497
520
|
using idx_t = Index::idx_t;
|
@@ -676,11 +699,12 @@ struct QueryTables {
|
|
676
699
|
} else if (use_precomputed_table == 1) {
|
677
700
|
dis0 = coarse_dis;
|
678
701
|
|
679
|
-
fvec_madd (
|
680
|
-
|
681
|
-
|
682
|
-
|
683
|
-
|
702
|
+
fvec_madd (
|
703
|
+
pq.M * pq.ksub,
|
704
|
+
ivfpq.precomputed_table.data() + key * pq.ksub * pq.M,
|
705
|
+
-2.0, sim_table_2,
|
706
|
+
sim_table
|
707
|
+
);
|
684
708
|
|
685
709
|
if (polysemous_ht != 0) {
|
686
710
|
ivfpq.quantizer->compute_residual (qi, residual_vec, key);
|
@@ -706,8 +730,8 @@ struct QueryTables {
|
|
706
730
|
k >>= cpq.nbits;
|
707
731
|
|
708
732
|
// get corresponding table
|
709
|
-
const float *pc =
|
710
|
-
|
733
|
+
const float *pc = ivfpq.precomputed_table.data() +
|
734
|
+
(ki * pq.M + cm * Mf) * pq.ksub;
|
711
735
|
|
712
736
|
if (polysemous_ht == 0) {
|
713
737
|
|
@@ -741,7 +765,8 @@ struct QueryTables {
|
|
741
765
|
if (use_precomputed_table == 1) {
|
742
766
|
dis0 = coarse_dis;
|
743
767
|
|
744
|
-
const float * s =
|
768
|
+
const float * s = ivfpq.precomputed_table.data() +
|
769
|
+
key * pq.ksub * pq.M;
|
745
770
|
for (int m = 0; m < pq.M; m++) {
|
746
771
|
sim_table_ptrs [m] = s;
|
747
772
|
s += pq.ksub;
|
@@ -761,8 +786,8 @@ struct QueryTables {
|
|
761
786
|
int ki = k & ((uint64_t(1) << cpq.nbits) - 1);
|
762
787
|
k >>= cpq.nbits;
|
763
788
|
|
764
|
-
const float *pc =
|
765
|
-
|
789
|
+
const float *pc = ivfpq.precomputed_table.data() +
|
790
|
+
(ki * pq.M + cm * Mf) * pq.ksub;
|
766
791
|
|
767
792
|
for (int m = m0; m < m0 + Mf; m++) {
|
768
793
|
sim_table_ptrs [m] = pc;
|
@@ -803,9 +828,8 @@ struct KnnSearchResults {
|
|
803
828
|
|
804
829
|
inline void add (idx_t j, float dis) {
|
805
830
|
if (C::cmp (heap_sim[0], dis)) {
|
806
|
-
heap_pop<C> (k, heap_sim, heap_ids);
|
807
831
|
idx_t id = ids ? ids[j] : lo_build (key, j);
|
808
|
-
|
832
|
+
heap_replace_top<C> (k, heap_sim, heap_ids, dis, id);
|
809
833
|
nup++;
|
810
834
|
}
|
811
835
|
}
|
@@ -16,7 +16,7 @@
|
|
16
16
|
#include <faiss/IndexIVF.h>
|
17
17
|
#include <faiss/IndexPQ.h>
|
18
18
|
#include <faiss/impl/platform_macros.h>
|
19
|
-
|
19
|
+
#include <faiss/utils/AlignedTable.h>
|
20
20
|
|
21
21
|
namespace faiss {
|
22
22
|
|
@@ -28,10 +28,14 @@ struct IVFPQSearchParameters: IVFSearchParameters {
|
|
28
28
|
};
|
29
29
|
|
30
30
|
|
31
|
+
|
32
|
+
FAISS_API extern size_t precomputed_table_max_bytes;
|
33
|
+
|
34
|
+
|
31
35
|
/** Inverted file with Product Quantizer encoding. Each residual
|
32
36
|
* vector is encoded as a product quantizer code.
|
33
37
|
*/
|
34
|
-
struct
|
38
|
+
struct IndexIVFPQ: IndexIVF {
|
35
39
|
bool by_residual; ///< Encode residual or plain vector?
|
36
40
|
|
37
41
|
ProductQuantizer pq; ///< produces the codes
|
@@ -45,18 +49,12 @@ struct FAISS_API IndexIVFPQ: IndexIVF {
|
|
45
49
|
|
46
50
|
/** Precompute table that speed up query preprocessing at some
|
47
51
|
* memory cost (used only for by_residual with L2 metric)
|
48
|
-
* =-1: force disable
|
49
|
-
* =0: decide heuristically (default: use tables only if they are
|
50
|
-
* < precomputed_tables_max_bytes)
|
51
|
-
* =1: tables that work for all quantizers (size 256 * nlist * M)
|
52
|
-
* =2: specific version for MultiIndexQuantizer (much more compact)
|
53
52
|
*/
|
54
53
|
int use_precomputed_table;
|
55
|
-
static size_t precomputed_table_max_bytes;
|
56
54
|
|
57
55
|
/// if use_precompute_table
|
58
56
|
/// size nlist * pq.M * pq.ksub
|
59
|
-
|
57
|
+
AlignedTable<float> precomputed_table;
|
60
58
|
|
61
59
|
IndexIVFPQ (
|
62
60
|
Index * quantizer, size_t d, size_t nlist,
|
@@ -133,6 +131,24 @@ struct FAISS_API IndexIVFPQ: IndexIVF {
|
|
133
131
|
|
134
132
|
};
|
135
133
|
|
134
|
+
/** Pre-compute distance tables for IVFPQ with by-residual and METRIC_L2
|
135
|
+
*
|
136
|
+
* @param use_precomputed_table (I/O)
|
137
|
+
* =-1: force disable
|
138
|
+
* =0: decide heuristically (default: use tables only if they are
|
139
|
+
* < precomputed_tables_max_bytes), set use_precomputed_table on output
|
140
|
+
* =1: tables that work for all quantizers (size 256 * nlist * M)
|
141
|
+
* =2: specific version for MultiIndexQuantizer (much more compact)
|
142
|
+
* @param precomputed_table precomputed table to intialize
|
143
|
+
*/
|
144
|
+
|
145
|
+
void initialize_IVFPQ_precomputed_table(
|
146
|
+
int &use_precomputed_table,
|
147
|
+
const Index *quantizer,
|
148
|
+
const ProductQuantizer &pq,
|
149
|
+
AlignedTable<float> & precomputed_table,
|
150
|
+
bool verbose
|
151
|
+
);
|
136
152
|
|
137
153
|
/// statistics are robust to internal threading, but not if
|
138
154
|
/// IndexIVFPQ::search_preassigned is called by multiple threads
|
@@ -0,0 +1,1116 @@
|
|
1
|
+
/**
|
2
|
+
* Copyright (c) Facebook, Inc. and its affiliates.
|
3
|
+
*
|
4
|
+
* This source code is licensed under the MIT license found in the
|
5
|
+
* LICENSE file in the root directory of this source tree.
|
6
|
+
*/
|
7
|
+
|
8
|
+
#include <faiss/IndexIVFPQFastScan.h>
|
9
|
+
|
10
|
+
#include <cassert>
|
11
|
+
#include <cstdio>
|
12
|
+
#include <inttypes.h>
|
13
|
+
|
14
|
+
#include <omp.h>
|
15
|
+
|
16
|
+
#include <memory>
|
17
|
+
|
18
|
+
#include <faiss/impl/FaissAssert.h>
|
19
|
+
#include <faiss/utils/utils.h>
|
20
|
+
#include <faiss/utils/distances.h>
|
21
|
+
#include <faiss/utils/simdlib.h>
|
22
|
+
#include <faiss/impl/AuxIndexStructures.h>
|
23
|
+
|
24
|
+
#include <faiss/invlists/BlockInvertedLists.h>
|
25
|
+
|
26
|
+
#include <faiss/impl/simd_result_handlers.h>
|
27
|
+
#include <faiss/utils/quantize_lut.h>
|
28
|
+
#include <faiss/impl/pq4_fast_scan.h>
|
29
|
+
|
30
|
+
namespace faiss {
|
31
|
+
|
32
|
+
using namespace simd_result_handlers;
|
33
|
+
|
34
|
+
|
35
|
+
inline size_t roundup(size_t a, size_t b) {
|
36
|
+
return (a + b - 1) / b * b;
|
37
|
+
}
|
38
|
+
|
39
|
+
|
40
|
+
IndexIVFPQFastScan::IndexIVFPQFastScan (
|
41
|
+
Index * quantizer, size_t d, size_t nlist,
|
42
|
+
size_t M, size_t nbits_per_idx,
|
43
|
+
MetricType metric, int bbs):
|
44
|
+
IndexIVF (quantizer, d, nlist, 0, metric),
|
45
|
+
pq (d, M, nbits_per_idx),
|
46
|
+
bbs (bbs)
|
47
|
+
{
|
48
|
+
FAISS_THROW_IF_NOT(nbits_per_idx == 4);
|
49
|
+
M2 = roundup(pq.M, 2);
|
50
|
+
by_residual = false; // set to false by default because it's much faster
|
51
|
+
is_trained = false;
|
52
|
+
code_size = pq.code_size;
|
53
|
+
|
54
|
+
replace_invlists(
|
55
|
+
new BlockInvertedLists(nlist, bbs, bbs * M2 / 2),
|
56
|
+
true
|
57
|
+
);
|
58
|
+
}
|
59
|
+
|
60
|
+
IndexIVFPQFastScan::IndexIVFPQFastScan ()
|
61
|
+
{
|
62
|
+
by_residual = false;
|
63
|
+
bbs = 0;
|
64
|
+
M2 = 0;
|
65
|
+
}
|
66
|
+
|
67
|
+
|
68
|
+
IndexIVFPQFastScan::IndexIVFPQFastScan(const IndexIVFPQ & orig, int bbs):
|
69
|
+
IndexIVF(
|
70
|
+
orig.quantizer, orig.d, orig.nlist,
|
71
|
+
orig.pq.code_size, orig.metric_type),
|
72
|
+
pq(orig.pq),
|
73
|
+
bbs(bbs)
|
74
|
+
{
|
75
|
+
FAISS_THROW_IF_NOT(orig.pq.nbits == 4);
|
76
|
+
|
77
|
+
by_residual = orig.by_residual;
|
78
|
+
ntotal = orig.ntotal;
|
79
|
+
is_trained = orig.is_trained;
|
80
|
+
nprobe = orig.nprobe;
|
81
|
+
size_t M = pq.M;
|
82
|
+
|
83
|
+
M2 = roundup(M, 2);
|
84
|
+
|
85
|
+
replace_invlists(
|
86
|
+
new BlockInvertedLists(orig.nlist, bbs, bbs * M2 / 2),
|
87
|
+
true
|
88
|
+
);
|
89
|
+
|
90
|
+
precomputed_table.resize(orig.precomputed_table.size());
|
91
|
+
|
92
|
+
if (precomputed_table.nbytes() > 0) {
|
93
|
+
memcpy(precomputed_table.get(), orig.precomputed_table.data(),
|
94
|
+
precomputed_table.nbytes()
|
95
|
+
);
|
96
|
+
}
|
97
|
+
|
98
|
+
for(size_t i = 0; i < nlist; i++) {
|
99
|
+
size_t nb = orig.invlists->list_size(i);
|
100
|
+
size_t nb2 = roundup(nb, bbs);
|
101
|
+
AlignedTable<uint8_t> tmp(nb2 * M2 / 2);
|
102
|
+
pq4_pack_codes(
|
103
|
+
InvertedLists::ScopedCodes(orig.invlists, i).get(),
|
104
|
+
nb, M, nb2, bbs, M2,
|
105
|
+
tmp.get()
|
106
|
+
);
|
107
|
+
invlists->add_entries(
|
108
|
+
i, nb,
|
109
|
+
InvertedLists::ScopedIds(orig.invlists, i).get(),
|
110
|
+
tmp.get()
|
111
|
+
);
|
112
|
+
}
|
113
|
+
|
114
|
+
orig_invlists = orig.invlists;
|
115
|
+
}
|
116
|
+
|
117
|
+
|
118
|
+
|
119
|
+
/*********************************************************
|
120
|
+
* Training
|
121
|
+
*********************************************************/
|
122
|
+
|
123
|
+
void IndexIVFPQFastScan::train_residual (idx_t n, const float *x_in)
|
124
|
+
{
|
125
|
+
|
126
|
+
const float * x = fvecs_maybe_subsample (
|
127
|
+
d, (size_t*)&n, pq.cp.max_points_per_centroid * pq.ksub,
|
128
|
+
x_in, verbose, pq.cp.seed);
|
129
|
+
|
130
|
+
std::unique_ptr<float []> del_x;
|
131
|
+
if (x != x_in) {
|
132
|
+
del_x.reset((float*)x);
|
133
|
+
}
|
134
|
+
|
135
|
+
const float *trainset;
|
136
|
+
AlignedTable<float> residuals;
|
137
|
+
|
138
|
+
if (by_residual) {
|
139
|
+
if(verbose) printf("computing residuals\n");
|
140
|
+
std::vector<idx_t> assign(n);
|
141
|
+
quantizer->assign (n, x, assign.data());
|
142
|
+
residuals.resize(n * d);
|
143
|
+
for (idx_t i = 0; i < n; i++) {
|
144
|
+
quantizer->compute_residual (
|
145
|
+
x + i * d,
|
146
|
+
residuals.data() + i * d,
|
147
|
+
assign[i]
|
148
|
+
);
|
149
|
+
}
|
150
|
+
trainset = residuals.data();
|
151
|
+
} else {
|
152
|
+
trainset = x;
|
153
|
+
}
|
154
|
+
|
155
|
+
if (verbose) {
|
156
|
+
printf ("training %zdx%zd product quantizer on %zd vectors in %dD\n",
|
157
|
+
pq.M, pq.ksub, long(n), d);
|
158
|
+
}
|
159
|
+
pq.verbose = verbose;
|
160
|
+
pq.train (n, trainset);
|
161
|
+
|
162
|
+
if (by_residual && metric_type == METRIC_L2) {
|
163
|
+
precompute_table();
|
164
|
+
}
|
165
|
+
|
166
|
+
}
|
167
|
+
|
168
|
+
void IndexIVFPQFastScan::precompute_table ()
|
169
|
+
{
|
170
|
+
initialize_IVFPQ_precomputed_table(
|
171
|
+
use_precomputed_table,
|
172
|
+
quantizer, pq, precomputed_table, verbose
|
173
|
+
);
|
174
|
+
}
|
175
|
+
|
176
|
+
|
177
|
+
/*********************************************************
|
178
|
+
* Code management functions
|
179
|
+
*********************************************************/
|
180
|
+
|
181
|
+
|
182
|
+
|
183
|
+
void IndexIVFPQFastScan::encode_vectors(
|
184
|
+
idx_t n, const float* x, const idx_t *list_nos,
|
185
|
+
uint8_t * codes, bool include_listnos) const
|
186
|
+
{
|
187
|
+
|
188
|
+
if (by_residual) {
|
189
|
+
AlignedTable<float> residuals (n * d);
|
190
|
+
for (size_t i = 0; i < n; i++) {
|
191
|
+
if (list_nos[i] < 0) {
|
192
|
+
memset (residuals.data() + i * d, 0, sizeof(residuals[0]) * d);
|
193
|
+
} else {
|
194
|
+
quantizer->compute_residual (
|
195
|
+
x + i * d, residuals.data() + i * d, list_nos[i]);
|
196
|
+
}
|
197
|
+
}
|
198
|
+
pq.compute_codes (residuals.data(), codes, n);
|
199
|
+
} else {
|
200
|
+
pq.compute_codes (x, codes, n);
|
201
|
+
}
|
202
|
+
|
203
|
+
if (include_listnos) {
|
204
|
+
size_t coarse_size = coarse_code_size();
|
205
|
+
for (idx_t i = n - 1; i >= 0; i--) {
|
206
|
+
uint8_t * code = codes + i * (coarse_size + code_size);
|
207
|
+
memmove (code + coarse_size,
|
208
|
+
codes + i * code_size, code_size);
|
209
|
+
encode_listno (list_nos[i], code);
|
210
|
+
}
|
211
|
+
}
|
212
|
+
}
|
213
|
+
|
214
|
+
|
215
|
+
|
216
|
+
void IndexIVFPQFastScan::add_with_ids (
|
217
|
+
idx_t n, const float * x, const idx_t *xids) {
|
218
|
+
|
219
|
+
// copied from IndexIVF::add_with_ids --->
|
220
|
+
|
221
|
+
// do some blocking to avoid excessive allocs
|
222
|
+
idx_t bs = 65536;
|
223
|
+
if (n > bs) {
|
224
|
+
for (idx_t i0 = 0; i0 < n; i0 += bs) {
|
225
|
+
idx_t i1 = std::min (n, i0 + bs);
|
226
|
+
if (verbose) {
|
227
|
+
printf(" IndexIVFPQFastScan::add_with_ids %zd: %zd",
|
228
|
+
size_t(i0), size_t(i1));
|
229
|
+
}
|
230
|
+
add_with_ids (i1 - i0, x + i0 * d,
|
231
|
+
xids ? xids + i0 : nullptr);
|
232
|
+
}
|
233
|
+
return;
|
234
|
+
}
|
235
|
+
InterruptCallback::check();
|
236
|
+
|
237
|
+
AlignedTable<uint8_t> codes(n * code_size);
|
238
|
+
|
239
|
+
FAISS_THROW_IF_NOT (is_trained);
|
240
|
+
direct_map.check_can_add (xids);
|
241
|
+
|
242
|
+
std::unique_ptr<idx_t []> idx(new idx_t[n]);
|
243
|
+
quantizer->assign (n, x, idx.get());
|
244
|
+
size_t nadd = 0, nminus1 = 0;
|
245
|
+
|
246
|
+
for (size_t i = 0; i < n; i++) {
|
247
|
+
if (idx[i] < 0) nminus1++;
|
248
|
+
}
|
249
|
+
|
250
|
+
AlignedTable<uint8_t> flat_codes(n * code_size);
|
251
|
+
encode_vectors (n, x, idx.get(), flat_codes.get());
|
252
|
+
|
253
|
+
DirectMapAdd dm_adder(direct_map, n, xids);
|
254
|
+
|
255
|
+
// <---
|
256
|
+
|
257
|
+
BlockInvertedLists *bil = dynamic_cast<BlockInvertedLists*>(invlists);
|
258
|
+
FAISS_THROW_IF_NOT_MSG (bil, "only block inverted lists supported");
|
259
|
+
|
260
|
+
// prepare batches
|
261
|
+
std::vector<idx_t> order(n);
|
262
|
+
for(idx_t i = 0; i < n ; i++) { order[i] = i; }
|
263
|
+
|
264
|
+
// TODO should not need stable
|
265
|
+
std::stable_sort(order.begin(), order.end(),
|
266
|
+
[&idx](idx_t a, idx_t b) {
|
267
|
+
return idx[a] < idx[b];
|
268
|
+
}
|
269
|
+
);
|
270
|
+
|
271
|
+
// TODO parallelize
|
272
|
+
idx_t i0 = 0;
|
273
|
+
while (i0 < n) {
|
274
|
+
idx_t list_no = idx[order[i0]];
|
275
|
+
idx_t i1 = i0 + 1;
|
276
|
+
while (i1 < n && idx[order[i1]] == list_no) {
|
277
|
+
i1 ++;
|
278
|
+
}
|
279
|
+
|
280
|
+
if (list_no == -1) {
|
281
|
+
i0 = i1;
|
282
|
+
continue;
|
283
|
+
}
|
284
|
+
|
285
|
+
// make linear array
|
286
|
+
AlignedTable<uint8_t> list_codes((i1 - i0) * code_size);
|
287
|
+
size_t list_size = bil->list_size(list_no);
|
288
|
+
|
289
|
+
bil->resize(list_no, list_size + i1 - i0);
|
290
|
+
|
291
|
+
for(idx_t i = i0; i < i1; i++) {
|
292
|
+
size_t ofs = list_size + i - i0;
|
293
|
+
idx_t id = xids ? xids[order[i]] : ntotal + order[i];
|
294
|
+
dm_adder.add (order[i], list_no, ofs);
|
295
|
+
bil->ids[list_no][ofs] = id;
|
296
|
+
memcpy(
|
297
|
+
list_codes.data() + (i - i0) * code_size,
|
298
|
+
flat_codes.data() + order[i] * code_size,
|
299
|
+
code_size
|
300
|
+
);
|
301
|
+
nadd++;
|
302
|
+
}
|
303
|
+
pq4_pack_codes_range(
|
304
|
+
list_codes.data(), pq.M,
|
305
|
+
list_size, list_size + i1 - i0,
|
306
|
+
bbs, M2, bil->codes[list_no].data()
|
307
|
+
);
|
308
|
+
|
309
|
+
i0 = i1;
|
310
|
+
}
|
311
|
+
|
312
|
+
ntotal += n;
|
313
|
+
|
314
|
+
}
|
315
|
+
|
316
|
+
|
317
|
+
|
318
|
+
/*********************************************************
|
319
|
+
* search
|
320
|
+
*********************************************************/
|
321
|
+
|
322
|
+
|
323
|
+
namespace {
|
324
|
+
|
325
|
+
// from impl/ProductQuantizer.cpp
|
326
|
+
template <class C, typename dis_t>
|
327
|
+
void pq_estimators_from_tables_generic(
|
328
|
+
const ProductQuantizer& pq, size_t nbits,
|
329
|
+
const uint8_t *codes, size_t ncodes,
|
330
|
+
const dis_t *dis_table, const int64_t * ids,
|
331
|
+
float dis0,
|
332
|
+
size_t k, typename C::T *heap_dis, int64_t *heap_ids)
|
333
|
+
{
|
334
|
+
using accu_t = typename C::T;
|
335
|
+
const size_t M = pq.M;
|
336
|
+
const size_t ksub = pq.ksub;
|
337
|
+
for (size_t j = 0; j < ncodes; ++j) {
|
338
|
+
PQDecoderGeneric decoder(
|
339
|
+
codes + j * pq.code_size, nbits
|
340
|
+
);
|
341
|
+
accu_t dis = dis0;
|
342
|
+
const dis_t * dt = dis_table;
|
343
|
+
for (size_t m = 0; m < M; m++) {
|
344
|
+
uint64_t c = decoder.decode();
|
345
|
+
dis += dt[c];
|
346
|
+
dt += ksub;
|
347
|
+
}
|
348
|
+
|
349
|
+
if (C::cmp(heap_dis[0], dis)) {
|
350
|
+
heap_pop<C>(k, heap_dis, heap_ids);
|
351
|
+
heap_push<C>(k, heap_dis, heap_ids, dis, ids[j]);
|
352
|
+
}
|
353
|
+
}
|
354
|
+
}
|
355
|
+
|
356
|
+
using idx_t = Index::idx_t;
|
357
|
+
using namespace quantize_lut;
|
358
|
+
|
359
|
+
void fvec_madd_avx (
|
360
|
+
size_t n, const float *a,
|
361
|
+
float bf, const float *b, float *c)
|
362
|
+
{
|
363
|
+
assert(is_aligned_pointer(a));
|
364
|
+
assert(is_aligned_pointer(b));
|
365
|
+
assert(is_aligned_pointer(c));
|
366
|
+
assert(n % 8 == 0);
|
367
|
+
simd8float32 bf8(bf);
|
368
|
+
n /= 8;
|
369
|
+
for(size_t i = 0; i < n; i++) {
|
370
|
+
simd8float32 ai(a);
|
371
|
+
simd8float32 bi(b);
|
372
|
+
|
373
|
+
simd8float32 ci = fmadd(bf8, bi, ai);
|
374
|
+
ci.store(c);
|
375
|
+
c += 8;
|
376
|
+
a += 8;
|
377
|
+
b += 8;
|
378
|
+
}
|
379
|
+
|
380
|
+
}
|
381
|
+
|
382
|
+
} // anonymous namespace
|
383
|
+
|
384
|
+
/*********************************************************
|
385
|
+
* Look-Up Table functions
|
386
|
+
*********************************************************/
|
387
|
+
|
388
|
+
|
389
|
+
void IndexIVFPQFastScan::compute_LUT(
|
390
|
+
size_t n, const float *x,
|
391
|
+
const idx_t *coarse_ids, const float *coarse_dis,
|
392
|
+
AlignedTable<float> & dis_tables,
|
393
|
+
AlignedTable<float> & biases
|
394
|
+
) const
|
395
|
+
{
|
396
|
+
const IndexIVFPQFastScan & ivfpq = *this;
|
397
|
+
size_t dim12 = pq.ksub * pq.M;
|
398
|
+
size_t d = pq.d;
|
399
|
+
size_t nprobe = ivfpq.nprobe;
|
400
|
+
|
401
|
+
if (ivfpq.by_residual) {
|
402
|
+
|
403
|
+
if (ivfpq.metric_type == METRIC_L2) {
|
404
|
+
|
405
|
+
dis_tables.resize(n * nprobe * dim12);
|
406
|
+
|
407
|
+
if (ivfpq.use_precomputed_table == 1) {
|
408
|
+
biases.resize(n * nprobe);
|
409
|
+
memcpy(biases.get(), coarse_dis, sizeof(float) * n * nprobe);
|
410
|
+
|
411
|
+
AlignedTable<float> ip_table(n * dim12);
|
412
|
+
pq.compute_inner_prod_tables (n, x, ip_table.get());
|
413
|
+
|
414
|
+
#pragma omp parallel for if (n * nprobe > 8000)
|
415
|
+
for(idx_t ij = 0; ij < n * nprobe; ij++) {
|
416
|
+
idx_t i = ij / nprobe;
|
417
|
+
float *tab = dis_tables.get() + ij * dim12;
|
418
|
+
idx_t cij = coarse_ids[ij];
|
419
|
+
|
420
|
+
if (cij >= 0) {
|
421
|
+
fvec_madd_avx (
|
422
|
+
dim12,
|
423
|
+
precomputed_table.get() + cij * dim12,
|
424
|
+
-2, ip_table.get() + i * dim12,
|
425
|
+
tab
|
426
|
+
);
|
427
|
+
} else {
|
428
|
+
// fill with NaNs so that they are ignored during
|
429
|
+
// LUT quantization
|
430
|
+
memset (tab, -1, sizeof(float) * dim12);
|
431
|
+
}
|
432
|
+
}
|
433
|
+
|
434
|
+
} else {
|
435
|
+
|
436
|
+
std::unique_ptr<float[]> xrel(new float[n * nprobe * d]);
|
437
|
+
biases.resize(n * nprobe);
|
438
|
+
memset(biases.get(), 0, sizeof(float) * n * nprobe);
|
439
|
+
|
440
|
+
#pragma omp parallel for if (n * nprobe > 8000)
|
441
|
+
for(idx_t ij = 0; ij < n * nprobe; ij++) {
|
442
|
+
idx_t i = ij / nprobe;
|
443
|
+
float *xij = &xrel[ij * d];
|
444
|
+
idx_t cij = coarse_ids[ij];
|
445
|
+
|
446
|
+
if (cij >= 0) {
|
447
|
+
ivfpq.quantizer->compute_residual(
|
448
|
+
x + i * d, xij, cij);
|
449
|
+
} else {
|
450
|
+
// will fill with NaNs
|
451
|
+
memset(xij, -1, sizeof(float) * d);
|
452
|
+
}
|
453
|
+
}
|
454
|
+
|
455
|
+
pq.compute_distance_tables (
|
456
|
+
n * nprobe, xrel.get(), dis_tables.get());
|
457
|
+
|
458
|
+
}
|
459
|
+
|
460
|
+
} else if (ivfpq.metric_type == METRIC_INNER_PRODUCT) {
|
461
|
+
dis_tables.resize(n * dim12);
|
462
|
+
pq.compute_inner_prod_tables (n, x, dis_tables.get());
|
463
|
+
// compute_inner_prod_tables(pq, n, x, dis_tables.get());
|
464
|
+
|
465
|
+
biases.resize(n * nprobe);
|
466
|
+
memcpy(biases.get(), coarse_dis, sizeof(float) * n * nprobe);
|
467
|
+
} else {
|
468
|
+
FAISS_THROW_FMT("metric %d not supported", ivfpq.metric_type);
|
469
|
+
}
|
470
|
+
|
471
|
+
} else {
|
472
|
+
dis_tables.resize(n * dim12);
|
473
|
+
if (ivfpq.metric_type == METRIC_L2) {
|
474
|
+
pq.compute_distance_tables (n, x, dis_tables.get());
|
475
|
+
} else if (ivfpq.metric_type == METRIC_INNER_PRODUCT) {
|
476
|
+
pq.compute_inner_prod_tables (n, x, dis_tables.get());
|
477
|
+
} else {
|
478
|
+
FAISS_THROW_FMT("metric %d not supported", ivfpq.metric_type);
|
479
|
+
}
|
480
|
+
}
|
481
|
+
|
482
|
+
}
|
483
|
+
|
484
|
+
void IndexIVFPQFastScan::compute_LUT_uint8(
|
485
|
+
size_t n, const float *x,
|
486
|
+
const idx_t *coarse_ids, const float *coarse_dis,
|
487
|
+
AlignedTable<uint8_t> & dis_tables,
|
488
|
+
AlignedTable<uint16_t> & biases,
|
489
|
+
float * normalizers
|
490
|
+
) const {
|
491
|
+
const IndexIVFPQFastScan & ivfpq = *this;
|
492
|
+
AlignedTable<float> dis_tables_float;
|
493
|
+
AlignedTable<float> biases_float;
|
494
|
+
|
495
|
+
uint64_t t0 = get_cy();
|
496
|
+
compute_LUT(
|
497
|
+
n, x,
|
498
|
+
coarse_ids, coarse_dis,
|
499
|
+
dis_tables_float, biases_float
|
500
|
+
);
|
501
|
+
IVFFastScan_stats.t_compute_distance_tables += get_cy() - t0;
|
502
|
+
|
503
|
+
bool lut_is_3d = ivfpq.by_residual && ivfpq.metric_type == METRIC_L2;
|
504
|
+
size_t dim123 = pq.ksub * pq.M;
|
505
|
+
size_t dim123_2 = pq.ksub * M2;
|
506
|
+
if (lut_is_3d) {
|
507
|
+
dim123 *= nprobe;
|
508
|
+
dim123_2 *= nprobe;
|
509
|
+
}
|
510
|
+
dis_tables.resize(n * dim123_2);
|
511
|
+
if (biases_float.get()) {
|
512
|
+
biases.resize(n * nprobe);
|
513
|
+
}
|
514
|
+
uint64_t t1 = get_cy();
|
515
|
+
|
516
|
+
#pragma omp parallel for if (n > 100)
|
517
|
+
for(int64_t i = 0; i < n; i++) {
|
518
|
+
const float *t_in = dis_tables_float.get() + i * dim123;
|
519
|
+
const float *b_in = nullptr;
|
520
|
+
uint8_t *t_out = dis_tables.get() + i * dim123_2;
|
521
|
+
uint16_t *b_out = nullptr;
|
522
|
+
if (biases_float.get()) {
|
523
|
+
b_in = biases_float.get() + i * nprobe;
|
524
|
+
b_out = biases.get() + i * nprobe;
|
525
|
+
}
|
526
|
+
|
527
|
+
quantize_LUT_and_bias(
|
528
|
+
nprobe, pq.M, pq.ksub, lut_is_3d,
|
529
|
+
t_in, b_in,
|
530
|
+
t_out, M2, b_out,
|
531
|
+
normalizers + 2 * i, normalizers + 2 * i + 1
|
532
|
+
);
|
533
|
+
}
|
534
|
+
IVFFastScan_stats.t_round += get_cy() - t1;
|
535
|
+
|
536
|
+
}
|
537
|
+
|
538
|
+
|
539
|
+
/*********************************************************
|
540
|
+
* Search functions
|
541
|
+
*********************************************************/
|
542
|
+
|
543
|
+
template<bool is_max>
|
544
|
+
void IndexIVFPQFastScan::search_dispatch_implem(
|
545
|
+
idx_t n,
|
546
|
+
const float* x,
|
547
|
+
idx_t k,
|
548
|
+
float* distances,
|
549
|
+
idx_t* labels) const
|
550
|
+
{
|
551
|
+
using Cfloat = typename std::conditional<is_max,
|
552
|
+
CMax<float, int64_t>, CMin<float, int64_t> >::type;
|
553
|
+
|
554
|
+
using C = typename std::conditional<is_max,
|
555
|
+
CMax<uint16_t, int64_t>, CMin<uint16_t, int64_t> >::type;
|
556
|
+
|
557
|
+
if (n == 0) {
|
558
|
+
return;
|
559
|
+
}
|
560
|
+
|
561
|
+
// actual implementation used
|
562
|
+
int impl = implem;
|
563
|
+
|
564
|
+
if (impl == 0) {
|
565
|
+
if (bbs == 32) {
|
566
|
+
impl = 12;
|
567
|
+
} else {
|
568
|
+
impl = 10;
|
569
|
+
}
|
570
|
+
if (k > 20) {
|
571
|
+
impl ++;
|
572
|
+
}
|
573
|
+
}
|
574
|
+
|
575
|
+
if (impl == 1) {
|
576
|
+
search_implem_1<Cfloat>(n, x, k, distances, labels);
|
577
|
+
} else if (impl == 2) {
|
578
|
+
search_implem_2<C>(n, x, k, distances, labels);
|
579
|
+
|
580
|
+
} else if (impl >= 10 && impl <= 13) {
|
581
|
+
size_t ndis = 0, nlist_visited = 0;
|
582
|
+
|
583
|
+
if (n < 2) {
|
584
|
+
if (impl == 12 || impl == 13) {
|
585
|
+
search_implem_12<C>
|
586
|
+
(n, x, k, distances, labels, impl, &ndis, &nlist_visited);
|
587
|
+
} else {
|
588
|
+
search_implem_10<C>
|
589
|
+
(n, x, k, distances, labels, impl, &ndis, &nlist_visited);
|
590
|
+
}
|
591
|
+
} else {
|
592
|
+
// explicitly slice over threads
|
593
|
+
int nslice;
|
594
|
+
if (n <= omp_get_max_threads()) {
|
595
|
+
nslice = n;
|
596
|
+
} else if (by_residual && metric_type == METRIC_L2) {
|
597
|
+
// make sure we don't make too big LUT tables
|
598
|
+
size_t lut_size_per_query =
|
599
|
+
pq.M * pq.ksub * nprobe * (sizeof(float) + sizeof(uint8_t));
|
600
|
+
|
601
|
+
size_t max_lut_size = precomputed_table_max_bytes;
|
602
|
+
// how many queries we can handle within mem budget
|
603
|
+
size_t nq_ok = std::max(max_lut_size / lut_size_per_query, size_t(1));
|
604
|
+
nslice = roundup(std::max(size_t(n / nq_ok), size_t(1)), omp_get_max_threads());
|
605
|
+
} else {
|
606
|
+
// LUTs unlikely to be a limiting factor
|
607
|
+
nslice = omp_get_max_threads();
|
608
|
+
}
|
609
|
+
|
610
|
+
#pragma omp parallel for reduction(+: ndis, nlist_visited)
|
611
|
+
for (int slice = 0; slice < nslice; slice++) {
|
612
|
+
idx_t i0 = n * slice / nslice;
|
613
|
+
idx_t i1 = n * (slice + 1) / nslice;
|
614
|
+
float *dis_i = distances + i0 * k;
|
615
|
+
idx_t *lab_i = labels + i0 * k;
|
616
|
+
if (impl == 12 || impl == 13) {
|
617
|
+
search_implem_12<C>(
|
618
|
+
i1 - i0, x + i0 * d, k, dis_i, lab_i,
|
619
|
+
impl, &ndis, &nlist_visited
|
620
|
+
);
|
621
|
+
} else {
|
622
|
+
search_implem_10<C>(
|
623
|
+
i1 - i0, x + i0 * d, k, dis_i, lab_i,
|
624
|
+
impl, &ndis, &nlist_visited
|
625
|
+
);
|
626
|
+
}
|
627
|
+
}
|
628
|
+
}
|
629
|
+
indexIVF_stats.nq += n;
|
630
|
+
indexIVF_stats.ndis += ndis;
|
631
|
+
indexIVF_stats.nlist += nlist_visited;
|
632
|
+
} else {
|
633
|
+
FAISS_THROW_FMT("implem %d does not exist", implem);
|
634
|
+
}
|
635
|
+
|
636
|
+
}
|
637
|
+
|
638
|
+
|
639
|
+
void IndexIVFPQFastScan::search(
|
640
|
+
idx_t n, const float* x, idx_t k,
|
641
|
+
float* distances, idx_t* labels) const
|
642
|
+
{
|
643
|
+
if (metric_type == METRIC_L2) {
|
644
|
+
search_dispatch_implem<true>(n, x, k, distances, labels);
|
645
|
+
} else {
|
646
|
+
search_dispatch_implem<false>(n, x, k, distances, labels);
|
647
|
+
}
|
648
|
+
}
|
649
|
+
|
650
|
+
template<class C>
|
651
|
+
void IndexIVFPQFastScan::search_implem_1(
|
652
|
+
idx_t n, const float* x, idx_t k,
|
653
|
+
float* distances, idx_t* labels) const
|
654
|
+
{
|
655
|
+
FAISS_THROW_IF_NOT(orig_invlists);
|
656
|
+
|
657
|
+
std::unique_ptr<idx_t[]> coarse_ids(new idx_t[n * nprobe]);
|
658
|
+
std::unique_ptr<float[]> coarse_dis(new float[n * nprobe]);
|
659
|
+
|
660
|
+
quantizer->search (n, x, nprobe, coarse_dis.get(), coarse_ids.get());
|
661
|
+
|
662
|
+
size_t dim12 = pq.ksub * pq.M;
|
663
|
+
AlignedTable<float> dis_tables;
|
664
|
+
AlignedTable<float> biases;
|
665
|
+
|
666
|
+
compute_LUT (
|
667
|
+
n, x,
|
668
|
+
coarse_ids.get(), coarse_dis.get(),
|
669
|
+
dis_tables, biases
|
670
|
+
);
|
671
|
+
|
672
|
+
bool single_LUT = !(by_residual && metric_type == METRIC_L2);
|
673
|
+
|
674
|
+
size_t ndis = 0, nlist_visited = 0;
|
675
|
+
|
676
|
+
#pragma omp parallel for reduction(+: ndis, nlist_visited)
|
677
|
+
for(idx_t i = 0; i < n; i++) {
|
678
|
+
int64_t *heap_ids = labels + i * k;
|
679
|
+
float *heap_dis = distances + i * k;
|
680
|
+
heap_heapify<C> (k, heap_dis, heap_ids);
|
681
|
+
float *LUT = nullptr;
|
682
|
+
|
683
|
+
if (single_LUT) {
|
684
|
+
LUT = dis_tables.get() + i * dim12;
|
685
|
+
}
|
686
|
+
for(idx_t j = 0; j < nprobe; j++) {
|
687
|
+
if (!single_LUT) {
|
688
|
+
LUT = dis_tables.get() + (i * nprobe + j) * dim12;
|
689
|
+
}
|
690
|
+
idx_t list_no = coarse_ids[i * nprobe + j];
|
691
|
+
if (list_no < 0) continue;
|
692
|
+
size_t ls = orig_invlists->list_size(list_no);
|
693
|
+
if (ls == 0) continue;
|
694
|
+
InvertedLists::ScopedCodes codes(orig_invlists, list_no);
|
695
|
+
InvertedLists::ScopedIds ids(orig_invlists, list_no);
|
696
|
+
|
697
|
+
float bias = biases.get() ? biases[i * nprobe + j] : 0;
|
698
|
+
|
699
|
+
pq_estimators_from_tables_generic<C>(
|
700
|
+
pq, pq.nbits, codes.get(), ls,
|
701
|
+
LUT, ids.get(), bias,
|
702
|
+
k, heap_dis, heap_ids
|
703
|
+
);
|
704
|
+
nlist_visited ++;
|
705
|
+
ndis ++;
|
706
|
+
}
|
707
|
+
heap_reorder<C> (k, heap_dis, heap_ids);
|
708
|
+
}
|
709
|
+
indexIVF_stats.nq += n;
|
710
|
+
indexIVF_stats.ndis += ndis;
|
711
|
+
indexIVF_stats.nlist += nlist_visited;
|
712
|
+
}
|
713
|
+
|
714
|
+
template<class C>
|
715
|
+
void IndexIVFPQFastScan::search_implem_2(
|
716
|
+
idx_t n, const float* x, idx_t k,
|
717
|
+
float* distances, idx_t* labels) const
|
718
|
+
{
|
719
|
+
FAISS_THROW_IF_NOT(orig_invlists);
|
720
|
+
|
721
|
+
std::unique_ptr<idx_t[]> coarse_ids(new idx_t[n * nprobe]);
|
722
|
+
std::unique_ptr<float[]> coarse_dis(new float[n * nprobe]);
|
723
|
+
|
724
|
+
quantizer->search (n, x, nprobe, coarse_dis.get(), coarse_ids.get());
|
725
|
+
|
726
|
+
size_t dim12 = pq.ksub * M2;
|
727
|
+
AlignedTable<uint8_t> dis_tables;
|
728
|
+
AlignedTable<uint16_t> biases;
|
729
|
+
std::unique_ptr<float[]> normalizers(new float[2 * n]);
|
730
|
+
|
731
|
+
compute_LUT_uint8 (
|
732
|
+
n, x,
|
733
|
+
coarse_ids.get(), coarse_dis.get(),
|
734
|
+
dis_tables, biases,
|
735
|
+
normalizers.get()
|
736
|
+
);
|
737
|
+
|
738
|
+
|
739
|
+
bool single_LUT = !(by_residual && metric_type == METRIC_L2);
|
740
|
+
|
741
|
+
size_t ndis = 0, nlist_visited = 0;
|
742
|
+
|
743
|
+
#pragma omp parallel for reduction(+: ndis, nlist_visited)
|
744
|
+
for(idx_t i = 0; i < n; i++) {
|
745
|
+
std::vector<uint16_t> tmp_dis(k);
|
746
|
+
int64_t *heap_ids = labels + i * k;
|
747
|
+
uint16_t *heap_dis = tmp_dis.data();
|
748
|
+
heap_heapify<C> (k, heap_dis, heap_ids);
|
749
|
+
const uint8_t *LUT = nullptr;
|
750
|
+
|
751
|
+
if (single_LUT) {
|
752
|
+
LUT = dis_tables.get() + i * dim12;
|
753
|
+
}
|
754
|
+
for(idx_t j = 0; j < nprobe; j++) {
|
755
|
+
if (!single_LUT) {
|
756
|
+
LUT = dis_tables.get() + (i * nprobe + j) * dim12;
|
757
|
+
}
|
758
|
+
idx_t list_no = coarse_ids[i * nprobe + j];
|
759
|
+
if (list_no < 0) continue;
|
760
|
+
size_t ls = orig_invlists->list_size(list_no);
|
761
|
+
if (ls == 0) continue;
|
762
|
+
InvertedLists::ScopedCodes codes(orig_invlists, list_no);
|
763
|
+
InvertedLists::ScopedIds ids(orig_invlists, list_no);
|
764
|
+
|
765
|
+
uint16_t bias = biases.get() ? biases[i * nprobe + j] : 0;
|
766
|
+
|
767
|
+
pq_estimators_from_tables_generic<C>(
|
768
|
+
pq, pq.nbits, codes.get(), ls,
|
769
|
+
LUT, ids.get(), bias,
|
770
|
+
k, heap_dis, heap_ids
|
771
|
+
);
|
772
|
+
|
773
|
+
nlist_visited++;
|
774
|
+
ndis += ls;
|
775
|
+
}
|
776
|
+
heap_reorder<C> (k, heap_dis, heap_ids);
|
777
|
+
// convert distances to float
|
778
|
+
{
|
779
|
+
float one_a = 1 / normalizers[2 * i], b = normalizers[2 * i + 1];
|
780
|
+
if (skip & 16) {
|
781
|
+
one_a = 1;
|
782
|
+
b = 0;
|
783
|
+
}
|
784
|
+
float *heap_dis_float = distances + i * k;
|
785
|
+
for (int j = 0; j < k; j++) {
|
786
|
+
heap_dis_float[j] = b + heap_dis[j] * one_a;
|
787
|
+
}
|
788
|
+
}
|
789
|
+
}
|
790
|
+
indexIVF_stats.nq += n;
|
791
|
+
indexIVF_stats.ndis += ndis;
|
792
|
+
indexIVF_stats.nlist += nlist_visited;
|
793
|
+
}
|
794
|
+
|
795
|
+
|
796
|
+
|
797
|
+
template<class C>
|
798
|
+
void IndexIVFPQFastScan::search_implem_10(
|
799
|
+
idx_t n, const float* x, idx_t k,
|
800
|
+
float* distances, idx_t* labels,
|
801
|
+
int impl, size_t *ndis_out, size_t *nlist_out) const
|
802
|
+
{
|
803
|
+
memset(distances, -1, sizeof(float) * k * n);
|
804
|
+
memset(labels, -1, sizeof(idx_t) * k * n);
|
805
|
+
|
806
|
+
using HeapHC = HeapHandler<C, true>;
|
807
|
+
using ReservoirHC = ReservoirHandler<C, true>;
|
808
|
+
using SingleResultHC = SingleResultHandler<C, true>;
|
809
|
+
|
810
|
+
|
811
|
+
std::unique_ptr<idx_t[]> coarse_ids(new idx_t[n * nprobe]);
|
812
|
+
std::unique_ptr<float[]> coarse_dis(new float[n * nprobe]);
|
813
|
+
|
814
|
+
uint64_t times[10];
|
815
|
+
memset(times, 0, sizeof(times));
|
816
|
+
int ti = 0;
|
817
|
+
#define TIC times[ti++] = get_cy()
|
818
|
+
TIC;
|
819
|
+
|
820
|
+
quantizer->search (n, x, nprobe, coarse_dis.get(), coarse_ids.get());
|
821
|
+
|
822
|
+
TIC;
|
823
|
+
|
824
|
+
size_t dim12 = pq.ksub * M2;
|
825
|
+
AlignedTable<uint8_t> dis_tables;
|
826
|
+
AlignedTable<uint16_t> biases;
|
827
|
+
std::unique_ptr<float[]> normalizers (new float[2 * n]);
|
828
|
+
|
829
|
+
compute_LUT_uint8 (
|
830
|
+
n, x,
|
831
|
+
coarse_ids.get(), coarse_dis.get(),
|
832
|
+
dis_tables, biases, normalizers.get()
|
833
|
+
);
|
834
|
+
|
835
|
+
TIC;
|
836
|
+
|
837
|
+
bool single_LUT = !(by_residual && metric_type == METRIC_L2);
|
838
|
+
|
839
|
+
TIC;
|
840
|
+
size_t ndis = 0, nlist_visited = 0;
|
841
|
+
|
842
|
+
{
|
843
|
+
AlignedTable<uint16_t> tmp_distances(k);
|
844
|
+
for(idx_t i = 0; i < n; i++) {
|
845
|
+
const uint8_t *LUT = nullptr;
|
846
|
+
int qmap1[1] = {0};
|
847
|
+
std::unique_ptr<SIMDResultHandler<C, true> > handler;
|
848
|
+
|
849
|
+
if (k == 1) {
|
850
|
+
handler.reset(new SingleResultHC(1, 0));
|
851
|
+
} else if (impl == 10) {
|
852
|
+
handler.reset(new HeapHC(1, tmp_distances.get(), labels + i * k, k, 0));
|
853
|
+
} else if (impl == 11) {
|
854
|
+
handler.reset(new ReservoirHC(1, 0, k, 2 * k));
|
855
|
+
} else {
|
856
|
+
FAISS_THROW_MSG("invalid");
|
857
|
+
}
|
858
|
+
|
859
|
+
handler->q_map = qmap1;
|
860
|
+
|
861
|
+
if (single_LUT) {
|
862
|
+
LUT = dis_tables.get() + i * dim12;
|
863
|
+
}
|
864
|
+
for(idx_t j = 0; j < nprobe; j++) {
|
865
|
+
size_t ij = i * nprobe + j;
|
866
|
+
if (!single_LUT) {
|
867
|
+
LUT = dis_tables.get() + ij * dim12;
|
868
|
+
}
|
869
|
+
if (biases.get()) {
|
870
|
+
handler->dbias = biases.get() + ij;
|
871
|
+
}
|
872
|
+
|
873
|
+
idx_t list_no = coarse_ids[ij];
|
874
|
+
if (list_no < 0) continue;
|
875
|
+
size_t ls = invlists->list_size(list_no);
|
876
|
+
if (ls == 0) continue;
|
877
|
+
|
878
|
+
InvertedLists::ScopedCodes codes(invlists, list_no);
|
879
|
+
InvertedLists::ScopedIds ids(invlists, list_no);
|
880
|
+
|
881
|
+
handler->ntotal = ls;
|
882
|
+
handler->id_map = ids.get();
|
883
|
+
|
884
|
+
#define DISPATCH(classHC) \
|
885
|
+
if(auto *res = dynamic_cast<classHC* > (handler.get())) { \
|
886
|
+
pq4_accumulate_loop( \
|
887
|
+
1, roundup(ls, bbs), bbs, M2, \
|
888
|
+
codes.get(), LUT, \
|
889
|
+
*res \
|
890
|
+
); \
|
891
|
+
}
|
892
|
+
DISPATCH(HeapHC)
|
893
|
+
else DISPATCH(ReservoirHC)
|
894
|
+
else DISPATCH(SingleResultHC)
|
895
|
+
#undef DISPATCH
|
896
|
+
|
897
|
+
nlist_visited ++;
|
898
|
+
ndis ++;
|
899
|
+
}
|
900
|
+
|
901
|
+
handler->to_flat_arrays(
|
902
|
+
distances + i * k, labels + i * k,
|
903
|
+
skip & 16 ? nullptr : normalizers.get() + i * 2
|
904
|
+
);
|
905
|
+
}
|
906
|
+
}
|
907
|
+
*ndis_out = ndis;
|
908
|
+
*nlist_out = nlist;
|
909
|
+
}
|
910
|
+
|
911
|
+
|
912
|
+
|
913
|
+
template<class C>
|
914
|
+
void IndexIVFPQFastScan::search_implem_12(
|
915
|
+
idx_t n, const float* x, idx_t k,
|
916
|
+
float* distances, idx_t* labels,
|
917
|
+
int impl, size_t *ndis_out, size_t *nlist_out) const
|
918
|
+
{
|
919
|
+
if (n == 0) { // does not work well with reservoir
|
920
|
+
return;
|
921
|
+
}
|
922
|
+
FAISS_THROW_IF_NOT(bbs == 32);
|
923
|
+
|
924
|
+
std::unique_ptr<idx_t[]> coarse_ids(new idx_t[n * nprobe]);
|
925
|
+
std::unique_ptr<float[]> coarse_dis(new float[n * nprobe]);
|
926
|
+
|
927
|
+
uint64_t times[10];
|
928
|
+
memset(times, 0, sizeof(times));
|
929
|
+
int ti = 0;
|
930
|
+
#define TIC times[ti++] = get_cy()
|
931
|
+
TIC;
|
932
|
+
|
933
|
+
quantizer->search (n, x, nprobe, coarse_dis.get(), coarse_ids.get());
|
934
|
+
|
935
|
+
TIC;
|
936
|
+
|
937
|
+
size_t dim12 = pq.ksub * M2;
|
938
|
+
AlignedTable<uint8_t> dis_tables;
|
939
|
+
AlignedTable<uint16_t> biases;
|
940
|
+
std::unique_ptr<float[]> normalizers (new float[2 * n]);
|
941
|
+
|
942
|
+
compute_LUT_uint8 (
|
943
|
+
n, x,
|
944
|
+
coarse_ids.get(), coarse_dis.get(),
|
945
|
+
dis_tables, biases, normalizers.get()
|
946
|
+
);
|
947
|
+
|
948
|
+
TIC;
|
949
|
+
|
950
|
+
struct QC {
|
951
|
+
int qno; // sequence number of the query
|
952
|
+
int list_no; // list to visit
|
953
|
+
int rank; // this is the rank'th result of the coarse quantizer
|
954
|
+
};
|
955
|
+
bool single_LUT = !(by_residual && metric_type == METRIC_L2);
|
956
|
+
|
957
|
+
std::vector<QC> qcs;
|
958
|
+
{
|
959
|
+
int ij = 0;
|
960
|
+
for(int i = 0; i < n; i++) {
|
961
|
+
for(int j = 0; j < nprobe; j++) {
|
962
|
+
if (coarse_ids[ij] >= 0) {
|
963
|
+
qcs.push_back(QC{i, int(coarse_ids[ij]), int(j)});
|
964
|
+
}
|
965
|
+
ij++;
|
966
|
+
}
|
967
|
+
}
|
968
|
+
std::sort(
|
969
|
+
qcs.begin(), qcs.end(),
|
970
|
+
[](const QC &a, const QC & b) {
|
971
|
+
return a.list_no < b.list_no;
|
972
|
+
}
|
973
|
+
);
|
974
|
+
}
|
975
|
+
TIC;
|
976
|
+
|
977
|
+
// prepare the result handlers
|
978
|
+
|
979
|
+
std::unique_ptr<SIMDResultHandler<C, true> > handler;
|
980
|
+
AlignedTable<uint16_t> tmp_distances;
|
981
|
+
|
982
|
+
using HeapHC = HeapHandler<C, true>;
|
983
|
+
using ReservoirHC = ReservoirHandler<C, true>;
|
984
|
+
using SingleResultHC = SingleResultHandler<C, true>;
|
985
|
+
|
986
|
+
if (k == 1) {
|
987
|
+
handler.reset(new SingleResultHC(n, 0));
|
988
|
+
} else if (impl == 12) {
|
989
|
+
tmp_distances.resize(n * k);
|
990
|
+
handler.reset(new HeapHC(n, tmp_distances.get(), labels, k, 0));
|
991
|
+
} else if (impl == 13) {
|
992
|
+
handler.reset(new ReservoirHC(n, 0, k, 2 * k));
|
993
|
+
}
|
994
|
+
|
995
|
+
int qbs2 = this->qbs2 ? this->qbs2 : 11;
|
996
|
+
|
997
|
+
std::vector<uint16_t> tmp_bias;
|
998
|
+
if (biases.get()) {
|
999
|
+
tmp_bias.resize(qbs2);
|
1000
|
+
handler->dbias = tmp_bias.data();
|
1001
|
+
}
|
1002
|
+
TIC;
|
1003
|
+
|
1004
|
+
size_t ndis = 0;
|
1005
|
+
|
1006
|
+
size_t i0 = 0;
|
1007
|
+
uint64_t t_copy_pack = 0, t_scan = 0;
|
1008
|
+
while (i0 < qcs.size()) {
|
1009
|
+
uint64_t tt0 = get_cy();
|
1010
|
+
|
1011
|
+
// find all queries that access this inverted list
|
1012
|
+
int list_no = qcs[i0].list_no;
|
1013
|
+
size_t i1 = i0 + 1;
|
1014
|
+
|
1015
|
+
while(i1 < qcs.size() && i1 < i0 + qbs2) {
|
1016
|
+
if (qcs[i1].list_no != list_no) {
|
1017
|
+
break;
|
1018
|
+
}
|
1019
|
+
i1++;
|
1020
|
+
}
|
1021
|
+
|
1022
|
+
size_t list_size = invlists->list_size(list_no);
|
1023
|
+
|
1024
|
+
if (list_size == 0) {
|
1025
|
+
i0 = i1;
|
1026
|
+
continue;
|
1027
|
+
}
|
1028
|
+
|
1029
|
+
// re-organize LUTs and biases into the right order
|
1030
|
+
int nc = i1 - i0;
|
1031
|
+
|
1032
|
+
std::vector<int> q_map(nc), lut_entries(nc);
|
1033
|
+
AlignedTable<uint8_t> LUT(nc * dim12);
|
1034
|
+
memset(LUT.get(), -1, nc * dim12);
|
1035
|
+
int qbs = pq4_preferred_qbs(nc);
|
1036
|
+
|
1037
|
+
for(size_t i = i0; i < i1; i++) {
|
1038
|
+
const QC & qc = qcs[i];
|
1039
|
+
q_map[i - i0] = qc.qno;
|
1040
|
+
int ij = qc.qno * nprobe + qc.rank;
|
1041
|
+
lut_entries[i - i0] = single_LUT ? qc.qno : ij;
|
1042
|
+
if (biases.get()) {
|
1043
|
+
tmp_bias[i - i0] = biases[ij];
|
1044
|
+
}
|
1045
|
+
}
|
1046
|
+
pq4_pack_LUT_qbs_q_map(
|
1047
|
+
qbs, M2, dis_tables.get(), lut_entries.data(),
|
1048
|
+
LUT.get()
|
1049
|
+
);
|
1050
|
+
|
1051
|
+
// access the inverted list
|
1052
|
+
|
1053
|
+
ndis += (i1 - i0) * list_size;
|
1054
|
+
|
1055
|
+
InvertedLists::ScopedCodes codes(invlists, list_no);
|
1056
|
+
InvertedLists::ScopedIds ids(invlists, list_no);
|
1057
|
+
|
1058
|
+
// prepare the handler
|
1059
|
+
|
1060
|
+
handler->ntotal = list_size;
|
1061
|
+
handler->q_map = q_map.data();
|
1062
|
+
handler->id_map = ids.get();
|
1063
|
+
uint64_t tt1 = get_cy();
|
1064
|
+
|
1065
|
+
#define DISPATCH(classHC) \
|
1066
|
+
if(auto *res = dynamic_cast<classHC* > (handler.get())) { \
|
1067
|
+
pq4_accumulate_loop_qbs( \
|
1068
|
+
qbs, list_size, M2, \
|
1069
|
+
codes.get(), LUT.get(), \
|
1070
|
+
*res \
|
1071
|
+
); \
|
1072
|
+
}
|
1073
|
+
DISPATCH(HeapHC)
|
1074
|
+
else DISPATCH(ReservoirHC)
|
1075
|
+
else DISPATCH(SingleResultHC)
|
1076
|
+
|
1077
|
+
// prepare for next loop
|
1078
|
+
i0 = i1;
|
1079
|
+
|
1080
|
+
uint64_t tt2 = get_cy();
|
1081
|
+
t_copy_pack += tt1 - tt0;
|
1082
|
+
t_scan += tt2 - tt1;
|
1083
|
+
}
|
1084
|
+
TIC;
|
1085
|
+
|
1086
|
+
// labels is in-place for HeapHC
|
1087
|
+
handler->to_flat_arrays(
|
1088
|
+
distances, labels,
|
1089
|
+
skip & 16 ? nullptr : normalizers.get()
|
1090
|
+
);
|
1091
|
+
|
1092
|
+
TIC;
|
1093
|
+
|
1094
|
+
// these stats are not thread-safe
|
1095
|
+
|
1096
|
+
for(int i = 1; i < ti; i++) {
|
1097
|
+
IVFFastScan_stats.times[i] += times[i] - times[i-1];
|
1098
|
+
}
|
1099
|
+
IVFFastScan_stats.t_copy_pack += t_copy_pack;
|
1100
|
+
IVFFastScan_stats.t_scan += t_scan;
|
1101
|
+
|
1102
|
+
if (auto *rh = dynamic_cast<ReservoirHC*> (handler.get())) {
|
1103
|
+
for (int i = 0; i < 4; i++) {
|
1104
|
+
IVFFastScan_stats.reservoir_times[i] += rh->times[i];
|
1105
|
+
}
|
1106
|
+
}
|
1107
|
+
|
1108
|
+
*ndis_out = ndis;
|
1109
|
+
*nlist_out = nlist;
|
1110
|
+
|
1111
|
+
}
|
1112
|
+
|
1113
|
+
|
1114
|
+
IVFFastScanStats IVFFastScan_stats;
|
1115
|
+
|
1116
|
+
} // namespace faiss
|