faiss 0.1.0 → 0.1.1
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +5 -0
- data/README.md +103 -3
- data/ext/faiss/ext.cpp +99 -32
- data/ext/faiss/extconf.rb +12 -2
- data/lib/faiss/ext.bundle +0 -0
- data/lib/faiss/index.rb +3 -3
- data/lib/faiss/index_binary.rb +3 -3
- data/lib/faiss/kmeans.rb +1 -1
- data/lib/faiss/pca_matrix.rb +2 -2
- data/lib/faiss/product_quantizer.rb +3 -3
- data/lib/faiss/version.rb +1 -1
- data/vendor/faiss/AutoTune.cpp +719 -0
- data/vendor/faiss/AutoTune.h +212 -0
- data/vendor/faiss/Clustering.cpp +261 -0
- data/vendor/faiss/Clustering.h +101 -0
- data/vendor/faiss/IVFlib.cpp +339 -0
- data/vendor/faiss/IVFlib.h +132 -0
- data/vendor/faiss/Index.cpp +171 -0
- data/vendor/faiss/Index.h +261 -0
- data/vendor/faiss/Index2Layer.cpp +437 -0
- data/vendor/faiss/Index2Layer.h +85 -0
- data/vendor/faiss/IndexBinary.cpp +77 -0
- data/vendor/faiss/IndexBinary.h +163 -0
- data/vendor/faiss/IndexBinaryFlat.cpp +83 -0
- data/vendor/faiss/IndexBinaryFlat.h +54 -0
- data/vendor/faiss/IndexBinaryFromFloat.cpp +78 -0
- data/vendor/faiss/IndexBinaryFromFloat.h +52 -0
- data/vendor/faiss/IndexBinaryHNSW.cpp +325 -0
- data/vendor/faiss/IndexBinaryHNSW.h +56 -0
- data/vendor/faiss/IndexBinaryIVF.cpp +671 -0
- data/vendor/faiss/IndexBinaryIVF.h +211 -0
- data/vendor/faiss/IndexFlat.cpp +508 -0
- data/vendor/faiss/IndexFlat.h +175 -0
- data/vendor/faiss/IndexHNSW.cpp +1090 -0
- data/vendor/faiss/IndexHNSW.h +170 -0
- data/vendor/faiss/IndexIVF.cpp +909 -0
- data/vendor/faiss/IndexIVF.h +353 -0
- data/vendor/faiss/IndexIVFFlat.cpp +502 -0
- data/vendor/faiss/IndexIVFFlat.h +118 -0
- data/vendor/faiss/IndexIVFPQ.cpp +1207 -0
- data/vendor/faiss/IndexIVFPQ.h +161 -0
- data/vendor/faiss/IndexIVFPQR.cpp +219 -0
- data/vendor/faiss/IndexIVFPQR.h +65 -0
- data/vendor/faiss/IndexIVFSpectralHash.cpp +331 -0
- data/vendor/faiss/IndexIVFSpectralHash.h +75 -0
- data/vendor/faiss/IndexLSH.cpp +225 -0
- data/vendor/faiss/IndexLSH.h +87 -0
- data/vendor/faiss/IndexLattice.cpp +143 -0
- data/vendor/faiss/IndexLattice.h +68 -0
- data/vendor/faiss/IndexPQ.cpp +1188 -0
- data/vendor/faiss/IndexPQ.h +199 -0
- data/vendor/faiss/IndexPreTransform.cpp +288 -0
- data/vendor/faiss/IndexPreTransform.h +91 -0
- data/vendor/faiss/IndexReplicas.cpp +123 -0
- data/vendor/faiss/IndexReplicas.h +76 -0
- data/vendor/faiss/IndexScalarQuantizer.cpp +317 -0
- data/vendor/faiss/IndexScalarQuantizer.h +127 -0
- data/vendor/faiss/IndexShards.cpp +317 -0
- data/vendor/faiss/IndexShards.h +100 -0
- data/vendor/faiss/InvertedLists.cpp +623 -0
- data/vendor/faiss/InvertedLists.h +334 -0
- data/vendor/faiss/LICENSE +21 -0
- data/vendor/faiss/MatrixStats.cpp +252 -0
- data/vendor/faiss/MatrixStats.h +62 -0
- data/vendor/faiss/MetaIndexes.cpp +351 -0
- data/vendor/faiss/MetaIndexes.h +126 -0
- data/vendor/faiss/OnDiskInvertedLists.cpp +674 -0
- data/vendor/faiss/OnDiskInvertedLists.h +127 -0
- data/vendor/faiss/VectorTransform.cpp +1157 -0
- data/vendor/faiss/VectorTransform.h +322 -0
- data/vendor/faiss/c_api/AutoTune_c.cpp +83 -0
- data/vendor/faiss/c_api/AutoTune_c.h +64 -0
- data/vendor/faiss/c_api/Clustering_c.cpp +139 -0
- data/vendor/faiss/c_api/Clustering_c.h +117 -0
- data/vendor/faiss/c_api/IndexFlat_c.cpp +140 -0
- data/vendor/faiss/c_api/IndexFlat_c.h +115 -0
- data/vendor/faiss/c_api/IndexIVFFlat_c.cpp +64 -0
- data/vendor/faiss/c_api/IndexIVFFlat_c.h +58 -0
- data/vendor/faiss/c_api/IndexIVF_c.cpp +92 -0
- data/vendor/faiss/c_api/IndexIVF_c.h +135 -0
- data/vendor/faiss/c_api/IndexLSH_c.cpp +37 -0
- data/vendor/faiss/c_api/IndexLSH_c.h +40 -0
- data/vendor/faiss/c_api/IndexShards_c.cpp +44 -0
- data/vendor/faiss/c_api/IndexShards_c.h +42 -0
- data/vendor/faiss/c_api/Index_c.cpp +105 -0
- data/vendor/faiss/c_api/Index_c.h +183 -0
- data/vendor/faiss/c_api/MetaIndexes_c.cpp +49 -0
- data/vendor/faiss/c_api/MetaIndexes_c.h +49 -0
- data/vendor/faiss/c_api/clone_index_c.cpp +23 -0
- data/vendor/faiss/c_api/clone_index_c.h +32 -0
- data/vendor/faiss/c_api/error_c.h +42 -0
- data/vendor/faiss/c_api/error_impl.cpp +27 -0
- data/vendor/faiss/c_api/error_impl.h +16 -0
- data/vendor/faiss/c_api/faiss_c.h +58 -0
- data/vendor/faiss/c_api/gpu/GpuAutoTune_c.cpp +96 -0
- data/vendor/faiss/c_api/gpu/GpuAutoTune_c.h +56 -0
- data/vendor/faiss/c_api/gpu/GpuClonerOptions_c.cpp +52 -0
- data/vendor/faiss/c_api/gpu/GpuClonerOptions_c.h +68 -0
- data/vendor/faiss/c_api/gpu/GpuIndex_c.cpp +17 -0
- data/vendor/faiss/c_api/gpu/GpuIndex_c.h +30 -0
- data/vendor/faiss/c_api/gpu/GpuIndicesOptions_c.h +38 -0
- data/vendor/faiss/c_api/gpu/GpuResources_c.cpp +86 -0
- data/vendor/faiss/c_api/gpu/GpuResources_c.h +66 -0
- data/vendor/faiss/c_api/gpu/StandardGpuResources_c.cpp +54 -0
- data/vendor/faiss/c_api/gpu/StandardGpuResources_c.h +53 -0
- data/vendor/faiss/c_api/gpu/macros_impl.h +42 -0
- data/vendor/faiss/c_api/impl/AuxIndexStructures_c.cpp +220 -0
- data/vendor/faiss/c_api/impl/AuxIndexStructures_c.h +149 -0
- data/vendor/faiss/c_api/index_factory_c.cpp +26 -0
- data/vendor/faiss/c_api/index_factory_c.h +30 -0
- data/vendor/faiss/c_api/index_io_c.cpp +42 -0
- data/vendor/faiss/c_api/index_io_c.h +50 -0
- data/vendor/faiss/c_api/macros_impl.h +110 -0
- data/vendor/faiss/clone_index.cpp +147 -0
- data/vendor/faiss/clone_index.h +38 -0
- data/vendor/faiss/demos/demo_imi_flat.cpp +151 -0
- data/vendor/faiss/demos/demo_imi_pq.cpp +199 -0
- data/vendor/faiss/demos/demo_ivfpq_indexing.cpp +146 -0
- data/vendor/faiss/demos/demo_sift1M.cpp +252 -0
- data/vendor/faiss/gpu/GpuAutoTune.cpp +95 -0
- data/vendor/faiss/gpu/GpuAutoTune.h +27 -0
- data/vendor/faiss/gpu/GpuCloner.cpp +403 -0
- data/vendor/faiss/gpu/GpuCloner.h +82 -0
- data/vendor/faiss/gpu/GpuClonerOptions.cpp +28 -0
- data/vendor/faiss/gpu/GpuClonerOptions.h +53 -0
- data/vendor/faiss/gpu/GpuDistance.h +52 -0
- data/vendor/faiss/gpu/GpuFaissAssert.h +29 -0
- data/vendor/faiss/gpu/GpuIndex.h +148 -0
- data/vendor/faiss/gpu/GpuIndexBinaryFlat.h +89 -0
- data/vendor/faiss/gpu/GpuIndexFlat.h +190 -0
- data/vendor/faiss/gpu/GpuIndexIVF.h +89 -0
- data/vendor/faiss/gpu/GpuIndexIVFFlat.h +85 -0
- data/vendor/faiss/gpu/GpuIndexIVFPQ.h +143 -0
- data/vendor/faiss/gpu/GpuIndexIVFScalarQuantizer.h +100 -0
- data/vendor/faiss/gpu/GpuIndicesOptions.h +30 -0
- data/vendor/faiss/gpu/GpuResources.cpp +52 -0
- data/vendor/faiss/gpu/GpuResources.h +73 -0
- data/vendor/faiss/gpu/StandardGpuResources.cpp +295 -0
- data/vendor/faiss/gpu/StandardGpuResources.h +114 -0
- data/vendor/faiss/gpu/impl/RemapIndices.cpp +43 -0
- data/vendor/faiss/gpu/impl/RemapIndices.h +24 -0
- data/vendor/faiss/gpu/perf/IndexWrapper-inl.h +71 -0
- data/vendor/faiss/gpu/perf/IndexWrapper.h +39 -0
- data/vendor/faiss/gpu/perf/PerfClustering.cpp +115 -0
- data/vendor/faiss/gpu/perf/PerfIVFPQAdd.cpp +139 -0
- data/vendor/faiss/gpu/perf/WriteIndex.cpp +102 -0
- data/vendor/faiss/gpu/test/TestGpuIndexBinaryFlat.cpp +130 -0
- data/vendor/faiss/gpu/test/TestGpuIndexFlat.cpp +371 -0
- data/vendor/faiss/gpu/test/TestGpuIndexIVFFlat.cpp +550 -0
- data/vendor/faiss/gpu/test/TestGpuIndexIVFPQ.cpp +450 -0
- data/vendor/faiss/gpu/test/TestGpuMemoryException.cpp +84 -0
- data/vendor/faiss/gpu/test/TestUtils.cpp +315 -0
- data/vendor/faiss/gpu/test/TestUtils.h +93 -0
- data/vendor/faiss/gpu/test/demo_ivfpq_indexing_gpu.cpp +159 -0
- data/vendor/faiss/gpu/utils/DeviceMemory.cpp +77 -0
- data/vendor/faiss/gpu/utils/DeviceMemory.h +71 -0
- data/vendor/faiss/gpu/utils/DeviceUtils.h +185 -0
- data/vendor/faiss/gpu/utils/MemorySpace.cpp +89 -0
- data/vendor/faiss/gpu/utils/MemorySpace.h +44 -0
- data/vendor/faiss/gpu/utils/StackDeviceMemory.cpp +239 -0
- data/vendor/faiss/gpu/utils/StackDeviceMemory.h +129 -0
- data/vendor/faiss/gpu/utils/StaticUtils.h +83 -0
- data/vendor/faiss/gpu/utils/Timer.cpp +60 -0
- data/vendor/faiss/gpu/utils/Timer.h +52 -0
- data/vendor/faiss/impl/AuxIndexStructures.cpp +305 -0
- data/vendor/faiss/impl/AuxIndexStructures.h +246 -0
- data/vendor/faiss/impl/FaissAssert.h +95 -0
- data/vendor/faiss/impl/FaissException.cpp +66 -0
- data/vendor/faiss/impl/FaissException.h +71 -0
- data/vendor/faiss/impl/HNSW.cpp +818 -0
- data/vendor/faiss/impl/HNSW.h +275 -0
- data/vendor/faiss/impl/PolysemousTraining.cpp +953 -0
- data/vendor/faiss/impl/PolysemousTraining.h +158 -0
- data/vendor/faiss/impl/ProductQuantizer.cpp +876 -0
- data/vendor/faiss/impl/ProductQuantizer.h +242 -0
- data/vendor/faiss/impl/ScalarQuantizer.cpp +1628 -0
- data/vendor/faiss/impl/ScalarQuantizer.h +120 -0
- data/vendor/faiss/impl/ThreadedIndex-inl.h +192 -0
- data/vendor/faiss/impl/ThreadedIndex.h +80 -0
- data/vendor/faiss/impl/index_read.cpp +793 -0
- data/vendor/faiss/impl/index_write.cpp +558 -0
- data/vendor/faiss/impl/io.cpp +142 -0
- data/vendor/faiss/impl/io.h +98 -0
- data/vendor/faiss/impl/lattice_Zn.cpp +712 -0
- data/vendor/faiss/impl/lattice_Zn.h +199 -0
- data/vendor/faiss/index_factory.cpp +392 -0
- data/vendor/faiss/index_factory.h +25 -0
- data/vendor/faiss/index_io.h +75 -0
- data/vendor/faiss/misc/test_blas.cpp +84 -0
- data/vendor/faiss/tests/test_binary_flat.cpp +64 -0
- data/vendor/faiss/tests/test_dealloc_invlists.cpp +183 -0
- data/vendor/faiss/tests/test_ivfpq_codec.cpp +67 -0
- data/vendor/faiss/tests/test_ivfpq_indexing.cpp +98 -0
- data/vendor/faiss/tests/test_lowlevel_ivf.cpp +566 -0
- data/vendor/faiss/tests/test_merge.cpp +258 -0
- data/vendor/faiss/tests/test_omp_threads.cpp +14 -0
- data/vendor/faiss/tests/test_ondisk_ivf.cpp +220 -0
- data/vendor/faiss/tests/test_pairs_decoding.cpp +189 -0
- data/vendor/faiss/tests/test_params_override.cpp +231 -0
- data/vendor/faiss/tests/test_pq_encoding.cpp +98 -0
- data/vendor/faiss/tests/test_sliding_ivf.cpp +240 -0
- data/vendor/faiss/tests/test_threaded_index.cpp +253 -0
- data/vendor/faiss/tests/test_transfer_invlists.cpp +159 -0
- data/vendor/faiss/tutorial/cpp/1-Flat.cpp +98 -0
- data/vendor/faiss/tutorial/cpp/2-IVFFlat.cpp +81 -0
- data/vendor/faiss/tutorial/cpp/3-IVFPQ.cpp +93 -0
- data/vendor/faiss/tutorial/cpp/4-GPU.cpp +119 -0
- data/vendor/faiss/tutorial/cpp/5-Multiple-GPUs.cpp +99 -0
- data/vendor/faiss/utils/Heap.cpp +122 -0
- data/vendor/faiss/utils/Heap.h +495 -0
- data/vendor/faiss/utils/WorkerThread.cpp +126 -0
- data/vendor/faiss/utils/WorkerThread.h +61 -0
- data/vendor/faiss/utils/distances.cpp +765 -0
- data/vendor/faiss/utils/distances.h +243 -0
- data/vendor/faiss/utils/distances_simd.cpp +809 -0
- data/vendor/faiss/utils/extra_distances.cpp +336 -0
- data/vendor/faiss/utils/extra_distances.h +54 -0
- data/vendor/faiss/utils/hamming-inl.h +472 -0
- data/vendor/faiss/utils/hamming.cpp +792 -0
- data/vendor/faiss/utils/hamming.h +220 -0
- data/vendor/faiss/utils/random.cpp +192 -0
- data/vendor/faiss/utils/random.h +60 -0
- data/vendor/faiss/utils/utils.cpp +783 -0
- data/vendor/faiss/utils/utils.h +181 -0
- metadata +216 -2
@@ -0,0 +1,623 @@
|
|
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
|
+
// -*- c++ -*-
|
9
|
+
|
10
|
+
#include <faiss/InvertedLists.h>
|
11
|
+
|
12
|
+
#include <cstdio>
|
13
|
+
|
14
|
+
#include <faiss/utils/utils.h>
|
15
|
+
#include <faiss/impl/FaissAssert.h>
|
16
|
+
|
17
|
+
namespace faiss {
|
18
|
+
|
19
|
+
using ScopedIds = InvertedLists::ScopedIds;
|
20
|
+
using ScopedCodes = InvertedLists::ScopedCodes;
|
21
|
+
|
22
|
+
|
23
|
+
/*****************************************
|
24
|
+
* InvertedLists implementation
|
25
|
+
******************************************/
|
26
|
+
|
27
|
+
InvertedLists::InvertedLists (size_t nlist, size_t code_size):
|
28
|
+
nlist (nlist), code_size (code_size)
|
29
|
+
{
|
30
|
+
}
|
31
|
+
|
32
|
+
InvertedLists::~InvertedLists ()
|
33
|
+
{}
|
34
|
+
|
35
|
+
InvertedLists::idx_t InvertedLists::get_single_id (
|
36
|
+
size_t list_no, size_t offset) const
|
37
|
+
{
|
38
|
+
assert (offset < list_size (list_no));
|
39
|
+
return get_ids(list_no)[offset];
|
40
|
+
}
|
41
|
+
|
42
|
+
|
43
|
+
void InvertedLists::release_codes (size_t, const uint8_t *) const
|
44
|
+
{}
|
45
|
+
|
46
|
+
void InvertedLists::release_ids (size_t, const idx_t *) const
|
47
|
+
{}
|
48
|
+
|
49
|
+
void InvertedLists::prefetch_lists (const idx_t *, int) const
|
50
|
+
{}
|
51
|
+
|
52
|
+
const uint8_t * InvertedLists::get_single_code (
|
53
|
+
size_t list_no, size_t offset) const
|
54
|
+
{
|
55
|
+
assert (offset < list_size (list_no));
|
56
|
+
return get_codes(list_no) + offset * code_size;
|
57
|
+
}
|
58
|
+
|
59
|
+
size_t InvertedLists::add_entry (size_t list_no, idx_t theid,
|
60
|
+
const uint8_t *code)
|
61
|
+
{
|
62
|
+
return add_entries (list_no, 1, &theid, code);
|
63
|
+
}
|
64
|
+
|
65
|
+
void InvertedLists::update_entry (size_t list_no, size_t offset,
|
66
|
+
idx_t id, const uint8_t *code)
|
67
|
+
{
|
68
|
+
update_entries (list_no, offset, 1, &id, code);
|
69
|
+
}
|
70
|
+
|
71
|
+
void InvertedLists::reset () {
|
72
|
+
for (size_t i = 0; i < nlist; i++) {
|
73
|
+
resize (i, 0);
|
74
|
+
}
|
75
|
+
}
|
76
|
+
|
77
|
+
void InvertedLists::merge_from (InvertedLists *oivf, size_t add_id) {
|
78
|
+
|
79
|
+
#pragma omp parallel for
|
80
|
+
for (idx_t i = 0; i < nlist; i++) {
|
81
|
+
size_t list_size = oivf->list_size (i);
|
82
|
+
ScopedIds ids (oivf, i);
|
83
|
+
if (add_id == 0) {
|
84
|
+
add_entries (i, list_size, ids.get (),
|
85
|
+
ScopedCodes (oivf, i).get());
|
86
|
+
} else {
|
87
|
+
std::vector <idx_t> new_ids (list_size);
|
88
|
+
|
89
|
+
for (size_t j = 0; j < list_size; j++) {
|
90
|
+
new_ids [j] = ids[j] + add_id;
|
91
|
+
}
|
92
|
+
add_entries (i, list_size, new_ids.data(),
|
93
|
+
ScopedCodes (oivf, i).get());
|
94
|
+
}
|
95
|
+
oivf->resize (i, 0);
|
96
|
+
}
|
97
|
+
}
|
98
|
+
|
99
|
+
double InvertedLists::imbalance_factor () const {
|
100
|
+
std::vector<int> hist(nlist);
|
101
|
+
|
102
|
+
for (size_t i = 0; i < nlist; i++) {
|
103
|
+
hist[i] = list_size(i);
|
104
|
+
}
|
105
|
+
|
106
|
+
return faiss::imbalance_factor(nlist, hist.data());
|
107
|
+
}
|
108
|
+
|
109
|
+
void InvertedLists::print_stats () const {
|
110
|
+
std::vector<int> sizes(40);
|
111
|
+
for (size_t i = 0; i < nlist; i++) {
|
112
|
+
for (size_t j = 0; j < sizes.size(); j++) {
|
113
|
+
if ((list_size(i) >> j) == 0) {
|
114
|
+
sizes[j]++;
|
115
|
+
break;
|
116
|
+
}
|
117
|
+
}
|
118
|
+
}
|
119
|
+
for (size_t i = 0; i < sizes.size(); i++) {
|
120
|
+
if (sizes[i]) {
|
121
|
+
printf("list size in < %d: %d instances\n", 1 << i, sizes[i]);
|
122
|
+
}
|
123
|
+
}
|
124
|
+
}
|
125
|
+
|
126
|
+
size_t InvertedLists::compute_ntotal () const {
|
127
|
+
size_t tot = 0;
|
128
|
+
for (size_t i = 0; i < nlist; i++) {
|
129
|
+
tot += list_size(i);
|
130
|
+
}
|
131
|
+
return tot;
|
132
|
+
}
|
133
|
+
|
134
|
+
/*****************************************
|
135
|
+
* ArrayInvertedLists implementation
|
136
|
+
******************************************/
|
137
|
+
|
138
|
+
ArrayInvertedLists::ArrayInvertedLists (size_t nlist, size_t code_size):
|
139
|
+
InvertedLists (nlist, code_size)
|
140
|
+
{
|
141
|
+
ids.resize (nlist);
|
142
|
+
codes.resize (nlist);
|
143
|
+
}
|
144
|
+
|
145
|
+
size_t ArrayInvertedLists::add_entries (
|
146
|
+
size_t list_no, size_t n_entry,
|
147
|
+
const idx_t* ids_in, const uint8_t *code)
|
148
|
+
{
|
149
|
+
if (n_entry == 0) return 0;
|
150
|
+
assert (list_no < nlist);
|
151
|
+
size_t o = ids [list_no].size();
|
152
|
+
ids [list_no].resize (o + n_entry);
|
153
|
+
memcpy (&ids[list_no][o], ids_in, sizeof (ids_in[0]) * n_entry);
|
154
|
+
codes [list_no].resize ((o + n_entry) * code_size);
|
155
|
+
memcpy (&codes[list_no][o * code_size], code, code_size * n_entry);
|
156
|
+
return o;
|
157
|
+
}
|
158
|
+
|
159
|
+
size_t ArrayInvertedLists::list_size(size_t list_no) const
|
160
|
+
{
|
161
|
+
assert (list_no < nlist);
|
162
|
+
return ids[list_no].size();
|
163
|
+
}
|
164
|
+
|
165
|
+
const uint8_t * ArrayInvertedLists::get_codes (size_t list_no) const
|
166
|
+
{
|
167
|
+
assert (list_no < nlist);
|
168
|
+
return codes[list_no].data();
|
169
|
+
}
|
170
|
+
|
171
|
+
|
172
|
+
const InvertedLists::idx_t * ArrayInvertedLists::get_ids (size_t list_no) const
|
173
|
+
{
|
174
|
+
assert (list_no < nlist);
|
175
|
+
return ids[list_no].data();
|
176
|
+
}
|
177
|
+
|
178
|
+
void ArrayInvertedLists::resize (size_t list_no, size_t new_size)
|
179
|
+
{
|
180
|
+
ids[list_no].resize (new_size);
|
181
|
+
codes[list_no].resize (new_size * code_size);
|
182
|
+
}
|
183
|
+
|
184
|
+
void ArrayInvertedLists::update_entries (
|
185
|
+
size_t list_no, size_t offset, size_t n_entry,
|
186
|
+
const idx_t *ids_in, const uint8_t *codes_in)
|
187
|
+
{
|
188
|
+
assert (list_no < nlist);
|
189
|
+
assert (n_entry + offset <= ids[list_no].size());
|
190
|
+
memcpy (&ids[list_no][offset], ids_in, sizeof(ids_in[0]) * n_entry);
|
191
|
+
memcpy (&codes[list_no][offset * code_size], codes_in, code_size * n_entry);
|
192
|
+
}
|
193
|
+
|
194
|
+
|
195
|
+
ArrayInvertedLists::~ArrayInvertedLists ()
|
196
|
+
{}
|
197
|
+
|
198
|
+
/*****************************************************************
|
199
|
+
* Meta-inverted list implementations
|
200
|
+
*****************************************************************/
|
201
|
+
|
202
|
+
|
203
|
+
size_t ReadOnlyInvertedLists::add_entries (
|
204
|
+
size_t , size_t ,
|
205
|
+
const idx_t* , const uint8_t *)
|
206
|
+
{
|
207
|
+
FAISS_THROW_MSG ("not implemented");
|
208
|
+
}
|
209
|
+
|
210
|
+
void ReadOnlyInvertedLists::update_entries (size_t, size_t , size_t ,
|
211
|
+
const idx_t *, const uint8_t *)
|
212
|
+
{
|
213
|
+
FAISS_THROW_MSG ("not implemented");
|
214
|
+
}
|
215
|
+
|
216
|
+
void ReadOnlyInvertedLists::resize (size_t , size_t )
|
217
|
+
{
|
218
|
+
FAISS_THROW_MSG ("not implemented");
|
219
|
+
}
|
220
|
+
|
221
|
+
|
222
|
+
|
223
|
+
/*****************************************
|
224
|
+
* HStackInvertedLists implementation
|
225
|
+
******************************************/
|
226
|
+
|
227
|
+
HStackInvertedLists::HStackInvertedLists (
|
228
|
+
int nil, const InvertedLists **ils_in):
|
229
|
+
ReadOnlyInvertedLists (nil > 0 ? ils_in[0]->nlist : 0,
|
230
|
+
nil > 0 ? ils_in[0]->code_size : 0)
|
231
|
+
{
|
232
|
+
FAISS_THROW_IF_NOT (nil > 0);
|
233
|
+
for (int i = 0; i < nil; i++) {
|
234
|
+
ils.push_back (ils_in[i]);
|
235
|
+
FAISS_THROW_IF_NOT (ils_in[i]->code_size == code_size &&
|
236
|
+
ils_in[i]->nlist == nlist);
|
237
|
+
}
|
238
|
+
}
|
239
|
+
|
240
|
+
size_t HStackInvertedLists::list_size(size_t list_no) const
|
241
|
+
{
|
242
|
+
size_t sz = 0;
|
243
|
+
for (int i = 0; i < ils.size(); i++) {
|
244
|
+
const InvertedLists *il = ils[i];
|
245
|
+
sz += il->list_size (list_no);
|
246
|
+
}
|
247
|
+
return sz;
|
248
|
+
}
|
249
|
+
|
250
|
+
const uint8_t * HStackInvertedLists::get_codes (size_t list_no) const
|
251
|
+
{
|
252
|
+
uint8_t *codes = new uint8_t [code_size * list_size(list_no)], *c = codes;
|
253
|
+
|
254
|
+
for (int i = 0; i < ils.size(); i++) {
|
255
|
+
const InvertedLists *il = ils[i];
|
256
|
+
size_t sz = il->list_size(list_no) * code_size;
|
257
|
+
if (sz > 0) {
|
258
|
+
memcpy (c, ScopedCodes (il, list_no).get(), sz);
|
259
|
+
c += sz;
|
260
|
+
}
|
261
|
+
}
|
262
|
+
return codes;
|
263
|
+
}
|
264
|
+
|
265
|
+
const uint8_t * HStackInvertedLists::get_single_code (
|
266
|
+
size_t list_no, size_t offset) const
|
267
|
+
{
|
268
|
+
for (int i = 0; i < ils.size(); i++) {
|
269
|
+
const InvertedLists *il = ils[i];
|
270
|
+
size_t sz = il->list_size (list_no);
|
271
|
+
if (offset < sz) {
|
272
|
+
// here we have to copy the code, otherwise it will crash at dealloc
|
273
|
+
uint8_t * code = new uint8_t [code_size];
|
274
|
+
memcpy (code, ScopedCodes (il, list_no, offset).get(), code_size);
|
275
|
+
return code;
|
276
|
+
}
|
277
|
+
offset -= sz;
|
278
|
+
}
|
279
|
+
FAISS_THROW_FMT ("offset %ld unknown", offset);
|
280
|
+
}
|
281
|
+
|
282
|
+
|
283
|
+
void HStackInvertedLists::release_codes (size_t, const uint8_t *codes) const {
|
284
|
+
delete [] codes;
|
285
|
+
}
|
286
|
+
|
287
|
+
const Index::idx_t * HStackInvertedLists::get_ids (size_t list_no) const
|
288
|
+
{
|
289
|
+
idx_t *ids = new idx_t [list_size(list_no)], *c = ids;
|
290
|
+
|
291
|
+
for (int i = 0; i < ils.size(); i++) {
|
292
|
+
const InvertedLists *il = ils[i];
|
293
|
+
size_t sz = il->list_size(list_no);
|
294
|
+
if (sz > 0) {
|
295
|
+
memcpy (c, ScopedIds (il, list_no).get(), sz * sizeof(idx_t));
|
296
|
+
c += sz;
|
297
|
+
}
|
298
|
+
}
|
299
|
+
return ids;
|
300
|
+
}
|
301
|
+
|
302
|
+
Index::idx_t HStackInvertedLists::get_single_id (
|
303
|
+
size_t list_no, size_t offset) const
|
304
|
+
{
|
305
|
+
|
306
|
+
for (int i = 0; i < ils.size(); i++) {
|
307
|
+
const InvertedLists *il = ils[i];
|
308
|
+
size_t sz = il->list_size (list_no);
|
309
|
+
if (offset < sz) {
|
310
|
+
return il->get_single_id (list_no, offset);
|
311
|
+
}
|
312
|
+
offset -= sz;
|
313
|
+
}
|
314
|
+
FAISS_THROW_FMT ("offset %ld unknown", offset);
|
315
|
+
}
|
316
|
+
|
317
|
+
|
318
|
+
void HStackInvertedLists::release_ids (size_t, const idx_t *ids) const {
|
319
|
+
delete [] ids;
|
320
|
+
}
|
321
|
+
|
322
|
+
void HStackInvertedLists::prefetch_lists (const idx_t *list_nos, int nlist) const
|
323
|
+
{
|
324
|
+
for (int i = 0; i < ils.size(); i++) {
|
325
|
+
const InvertedLists *il = ils[i];
|
326
|
+
il->prefetch_lists (list_nos, nlist);
|
327
|
+
}
|
328
|
+
}
|
329
|
+
|
330
|
+
/*****************************************
|
331
|
+
* SliceInvertedLists implementation
|
332
|
+
******************************************/
|
333
|
+
|
334
|
+
|
335
|
+
namespace {
|
336
|
+
|
337
|
+
using idx_t = InvertedLists::idx_t;
|
338
|
+
|
339
|
+
idx_t translate_list_no (const SliceInvertedLists *sil,
|
340
|
+
idx_t list_no) {
|
341
|
+
FAISS_THROW_IF_NOT (list_no >= 0 && list_no < sil->nlist);
|
342
|
+
return list_no + sil->i0;
|
343
|
+
}
|
344
|
+
|
345
|
+
};
|
346
|
+
|
347
|
+
|
348
|
+
|
349
|
+
SliceInvertedLists::SliceInvertedLists (
|
350
|
+
const InvertedLists *il, idx_t i0, idx_t i1):
|
351
|
+
ReadOnlyInvertedLists (i1 - i0, il->code_size),
|
352
|
+
il (il), i0(i0), i1(i1)
|
353
|
+
{
|
354
|
+
|
355
|
+
}
|
356
|
+
|
357
|
+
size_t SliceInvertedLists::list_size(size_t list_no) const
|
358
|
+
{
|
359
|
+
return il->list_size (translate_list_no (this, list_no));
|
360
|
+
}
|
361
|
+
|
362
|
+
const uint8_t * SliceInvertedLists::get_codes (size_t list_no) const
|
363
|
+
{
|
364
|
+
return il->get_codes (translate_list_no (this, list_no));
|
365
|
+
}
|
366
|
+
|
367
|
+
const uint8_t * SliceInvertedLists::get_single_code (
|
368
|
+
size_t list_no, size_t offset) const
|
369
|
+
{
|
370
|
+
return il->get_single_code (translate_list_no (this, list_no), offset);
|
371
|
+
}
|
372
|
+
|
373
|
+
|
374
|
+
void SliceInvertedLists::release_codes (
|
375
|
+
size_t list_no, const uint8_t *codes) const {
|
376
|
+
return il->release_codes (translate_list_no (this, list_no), codes);
|
377
|
+
}
|
378
|
+
|
379
|
+
const Index::idx_t * SliceInvertedLists::get_ids (size_t list_no) const
|
380
|
+
{
|
381
|
+
return il->get_ids (translate_list_no (this, list_no));
|
382
|
+
}
|
383
|
+
|
384
|
+
Index::idx_t SliceInvertedLists::get_single_id (
|
385
|
+
size_t list_no, size_t offset) const
|
386
|
+
{
|
387
|
+
return il->get_single_id (translate_list_no (this, list_no), offset);
|
388
|
+
}
|
389
|
+
|
390
|
+
|
391
|
+
void SliceInvertedLists::release_ids (size_t list_no, const idx_t *ids) const {
|
392
|
+
return il->release_ids (translate_list_no (this, list_no), ids);
|
393
|
+
}
|
394
|
+
|
395
|
+
void SliceInvertedLists::prefetch_lists (const idx_t *list_nos, int nlist) const
|
396
|
+
{
|
397
|
+
std::vector<idx_t> translated_list_nos;
|
398
|
+
for (int j = 0; j < nlist; j++) {
|
399
|
+
idx_t list_no = list_nos[j];
|
400
|
+
if (list_no < 0) continue;
|
401
|
+
translated_list_nos.push_back (translate_list_no (this, list_no));
|
402
|
+
}
|
403
|
+
il->prefetch_lists (translated_list_nos.data(),
|
404
|
+
translated_list_nos.size());
|
405
|
+
}
|
406
|
+
|
407
|
+
|
408
|
+
/*****************************************
|
409
|
+
* VStackInvertedLists implementation
|
410
|
+
******************************************/
|
411
|
+
|
412
|
+
namespace {
|
413
|
+
|
414
|
+
using idx_t = InvertedLists::idx_t;
|
415
|
+
|
416
|
+
// find the invlist this number belongs to
|
417
|
+
int translate_list_no (const VStackInvertedLists *vil,
|
418
|
+
idx_t list_no) {
|
419
|
+
FAISS_THROW_IF_NOT (list_no >= 0 && list_no < vil->nlist);
|
420
|
+
int i0 = 0, i1 = vil->ils.size();
|
421
|
+
const idx_t *cumsz = vil->cumsz.data();
|
422
|
+
while (i0 + 1 < i1) {
|
423
|
+
int imed = (i0 + i1) / 2;
|
424
|
+
if (list_no >= cumsz[imed]) {
|
425
|
+
i0 = imed;
|
426
|
+
} else {
|
427
|
+
i1 = imed;
|
428
|
+
}
|
429
|
+
}
|
430
|
+
assert(list_no >= cumsz[i0] && list_no < cumsz[i0 + 1]);
|
431
|
+
return i0;
|
432
|
+
}
|
433
|
+
|
434
|
+
idx_t sum_il_sizes (int nil, const InvertedLists **ils_in) {
|
435
|
+
idx_t tot = 0;
|
436
|
+
for (int i = 0; i < nil; i++) {
|
437
|
+
tot += ils_in[i]->nlist;
|
438
|
+
}
|
439
|
+
return tot;
|
440
|
+
}
|
441
|
+
|
442
|
+
};
|
443
|
+
|
444
|
+
|
445
|
+
|
446
|
+
VStackInvertedLists::VStackInvertedLists (
|
447
|
+
int nil, const InvertedLists **ils_in):
|
448
|
+
ReadOnlyInvertedLists (sum_il_sizes(nil, ils_in),
|
449
|
+
nil > 0 ? ils_in[0]->code_size : 0)
|
450
|
+
{
|
451
|
+
FAISS_THROW_IF_NOT (nil > 0);
|
452
|
+
cumsz.resize (nil + 1);
|
453
|
+
for (int i = 0; i < nil; i++) {
|
454
|
+
ils.push_back (ils_in[i]);
|
455
|
+
FAISS_THROW_IF_NOT (ils_in[i]->code_size == code_size);
|
456
|
+
cumsz[i + 1] = cumsz[i] + ils_in[i]->nlist;
|
457
|
+
}
|
458
|
+
}
|
459
|
+
|
460
|
+
size_t VStackInvertedLists::list_size(size_t list_no) const
|
461
|
+
{
|
462
|
+
int i = translate_list_no (this, list_no);
|
463
|
+
list_no -= cumsz[i];
|
464
|
+
return ils[i]->list_size (list_no);
|
465
|
+
}
|
466
|
+
|
467
|
+
const uint8_t * VStackInvertedLists::get_codes (size_t list_no) const
|
468
|
+
{
|
469
|
+
int i = translate_list_no (this, list_no);
|
470
|
+
list_no -= cumsz[i];
|
471
|
+
return ils[i]->get_codes (list_no);
|
472
|
+
}
|
473
|
+
|
474
|
+
const uint8_t * VStackInvertedLists::get_single_code (
|
475
|
+
size_t list_no, size_t offset) const
|
476
|
+
{
|
477
|
+
int i = translate_list_no (this, list_no);
|
478
|
+
list_no -= cumsz[i];
|
479
|
+
return ils[i]->get_single_code (list_no, offset);
|
480
|
+
}
|
481
|
+
|
482
|
+
|
483
|
+
void VStackInvertedLists::release_codes (
|
484
|
+
size_t list_no, const uint8_t *codes) const {
|
485
|
+
int i = translate_list_no (this, list_no);
|
486
|
+
list_no -= cumsz[i];
|
487
|
+
return ils[i]->release_codes (list_no, codes);
|
488
|
+
}
|
489
|
+
|
490
|
+
const Index::idx_t * VStackInvertedLists::get_ids (size_t list_no) const
|
491
|
+
{
|
492
|
+
int i = translate_list_no (this, list_no);
|
493
|
+
list_no -= cumsz[i];
|
494
|
+
return ils[i]->get_ids (list_no);
|
495
|
+
}
|
496
|
+
|
497
|
+
Index::idx_t VStackInvertedLists::get_single_id (
|
498
|
+
size_t list_no, size_t offset) const
|
499
|
+
{
|
500
|
+
int i = translate_list_no (this, list_no);
|
501
|
+
list_no -= cumsz[i];
|
502
|
+
return ils[i]->get_single_id (list_no, offset);
|
503
|
+
}
|
504
|
+
|
505
|
+
|
506
|
+
void VStackInvertedLists::release_ids (size_t list_no, const idx_t *ids) const {
|
507
|
+
int i = translate_list_no (this, list_no);
|
508
|
+
list_no -= cumsz[i];
|
509
|
+
return ils[i]->release_ids (list_no, ids);
|
510
|
+
}
|
511
|
+
|
512
|
+
void VStackInvertedLists::prefetch_lists (
|
513
|
+
const idx_t *list_nos, int nlist) const
|
514
|
+
{
|
515
|
+
std::vector<int> ilno (nlist, -1);
|
516
|
+
std::vector<int> n_per_il (ils.size(), 0);
|
517
|
+
for (int j = 0; j < nlist; j++) {
|
518
|
+
idx_t list_no = list_nos[j];
|
519
|
+
if (list_no < 0) continue;
|
520
|
+
int i = ilno[j] = translate_list_no (this, list_no);
|
521
|
+
n_per_il[i]++;
|
522
|
+
}
|
523
|
+
std::vector<int> cum_n_per_il (ils.size() + 1, 0);
|
524
|
+
for (int j = 0; j < ils.size(); j++) {
|
525
|
+
cum_n_per_il[j + 1] = cum_n_per_il[j] + n_per_il[j];
|
526
|
+
}
|
527
|
+
std::vector<idx_t> sorted_list_nos (cum_n_per_il.back());
|
528
|
+
for (int j = 0; j < nlist; j++) {
|
529
|
+
idx_t list_no = list_nos[j];
|
530
|
+
if (list_no < 0) continue;
|
531
|
+
int i = ilno[j];
|
532
|
+
list_no -= cumsz[i];
|
533
|
+
sorted_list_nos[cum_n_per_il[i]++] = list_no;
|
534
|
+
}
|
535
|
+
|
536
|
+
int i0 = 0;
|
537
|
+
for (int j = 0; j < ils.size(); j++) {
|
538
|
+
int i1 = i0 + n_per_il[j];
|
539
|
+
if (i1 > i0) {
|
540
|
+
ils[j]->prefetch_lists (sorted_list_nos.data() + i0,
|
541
|
+
i1 - i0);
|
542
|
+
}
|
543
|
+
i0 = i1;
|
544
|
+
}
|
545
|
+
}
|
546
|
+
|
547
|
+
|
548
|
+
|
549
|
+
/*****************************************
|
550
|
+
* MaskedInvertedLists implementation
|
551
|
+
******************************************/
|
552
|
+
|
553
|
+
|
554
|
+
MaskedInvertedLists::MaskedInvertedLists (const InvertedLists *il0,
|
555
|
+
const InvertedLists *il1):
|
556
|
+
ReadOnlyInvertedLists (il0->nlist, il0->code_size),
|
557
|
+
il0 (il0), il1 (il1)
|
558
|
+
{
|
559
|
+
FAISS_THROW_IF_NOT (il1->nlist == nlist);
|
560
|
+
FAISS_THROW_IF_NOT (il1->code_size == code_size);
|
561
|
+
}
|
562
|
+
|
563
|
+
size_t MaskedInvertedLists::list_size(size_t list_no) const
|
564
|
+
{
|
565
|
+
size_t sz = il0->list_size(list_no);
|
566
|
+
return sz ? sz : il1->list_size(list_no);
|
567
|
+
}
|
568
|
+
|
569
|
+
const uint8_t * MaskedInvertedLists::get_codes (size_t list_no) const
|
570
|
+
{
|
571
|
+
size_t sz = il0->list_size(list_no);
|
572
|
+
return (sz ? il0 : il1)->get_codes(list_no);
|
573
|
+
}
|
574
|
+
|
575
|
+
const idx_t * MaskedInvertedLists::get_ids (size_t list_no) const
|
576
|
+
{
|
577
|
+
size_t sz = il0->list_size (list_no);
|
578
|
+
return (sz ? il0 : il1)->get_ids (list_no);
|
579
|
+
}
|
580
|
+
|
581
|
+
void MaskedInvertedLists::release_codes (
|
582
|
+
size_t list_no, const uint8_t *codes) const
|
583
|
+
{
|
584
|
+
size_t sz = il0->list_size (list_no);
|
585
|
+
(sz ? il0 : il1)->release_codes (list_no, codes);
|
586
|
+
}
|
587
|
+
|
588
|
+
void MaskedInvertedLists::release_ids (size_t list_no, const idx_t *ids) const
|
589
|
+
{
|
590
|
+
size_t sz = il0->list_size (list_no);
|
591
|
+
(sz ? il0 : il1)->release_ids (list_no, ids);
|
592
|
+
}
|
593
|
+
|
594
|
+
idx_t MaskedInvertedLists::get_single_id (size_t list_no, size_t offset) const
|
595
|
+
{
|
596
|
+
size_t sz = il0->list_size (list_no);
|
597
|
+
return (sz ? il0 : il1)->get_single_id (list_no, offset);
|
598
|
+
}
|
599
|
+
|
600
|
+
const uint8_t * MaskedInvertedLists::get_single_code (
|
601
|
+
size_t list_no, size_t offset) const
|
602
|
+
{
|
603
|
+
size_t sz = il0->list_size (list_no);
|
604
|
+
return (sz ? il0 : il1)->get_single_code (list_no, offset);
|
605
|
+
}
|
606
|
+
|
607
|
+
void MaskedInvertedLists::prefetch_lists (
|
608
|
+
const idx_t *list_nos, int nlist) const
|
609
|
+
{
|
610
|
+
std::vector<idx_t> list0, list1;
|
611
|
+
for (int i = 0; i < nlist; i++) {
|
612
|
+
idx_t list_no = list_nos[i];
|
613
|
+
if (list_no < 0) continue;
|
614
|
+
size_t sz = il0->list_size(list_no);
|
615
|
+
(sz ? list0 : list1).push_back (list_no);
|
616
|
+
}
|
617
|
+
il0->prefetch_lists (list0.data(), list0.size());
|
618
|
+
il1->prefetch_lists (list1.data(), list1.size());
|
619
|
+
}
|
620
|
+
|
621
|
+
|
622
|
+
|
623
|
+
} // namespace faiss
|