faiss 0.1.0 → 0.1.1

Sign up to get free protection for your applications and to get access to all the features.
Files changed (226) hide show
  1. checksums.yaml +4 -4
  2. data/CHANGELOG.md +5 -0
  3. data/README.md +103 -3
  4. data/ext/faiss/ext.cpp +99 -32
  5. data/ext/faiss/extconf.rb +12 -2
  6. data/lib/faiss/ext.bundle +0 -0
  7. data/lib/faiss/index.rb +3 -3
  8. data/lib/faiss/index_binary.rb +3 -3
  9. data/lib/faiss/kmeans.rb +1 -1
  10. data/lib/faiss/pca_matrix.rb +2 -2
  11. data/lib/faiss/product_quantizer.rb +3 -3
  12. data/lib/faiss/version.rb +1 -1
  13. data/vendor/faiss/AutoTune.cpp +719 -0
  14. data/vendor/faiss/AutoTune.h +212 -0
  15. data/vendor/faiss/Clustering.cpp +261 -0
  16. data/vendor/faiss/Clustering.h +101 -0
  17. data/vendor/faiss/IVFlib.cpp +339 -0
  18. data/vendor/faiss/IVFlib.h +132 -0
  19. data/vendor/faiss/Index.cpp +171 -0
  20. data/vendor/faiss/Index.h +261 -0
  21. data/vendor/faiss/Index2Layer.cpp +437 -0
  22. data/vendor/faiss/Index2Layer.h +85 -0
  23. data/vendor/faiss/IndexBinary.cpp +77 -0
  24. data/vendor/faiss/IndexBinary.h +163 -0
  25. data/vendor/faiss/IndexBinaryFlat.cpp +83 -0
  26. data/vendor/faiss/IndexBinaryFlat.h +54 -0
  27. data/vendor/faiss/IndexBinaryFromFloat.cpp +78 -0
  28. data/vendor/faiss/IndexBinaryFromFloat.h +52 -0
  29. data/vendor/faiss/IndexBinaryHNSW.cpp +325 -0
  30. data/vendor/faiss/IndexBinaryHNSW.h +56 -0
  31. data/vendor/faiss/IndexBinaryIVF.cpp +671 -0
  32. data/vendor/faiss/IndexBinaryIVF.h +211 -0
  33. data/vendor/faiss/IndexFlat.cpp +508 -0
  34. data/vendor/faiss/IndexFlat.h +175 -0
  35. data/vendor/faiss/IndexHNSW.cpp +1090 -0
  36. data/vendor/faiss/IndexHNSW.h +170 -0
  37. data/vendor/faiss/IndexIVF.cpp +909 -0
  38. data/vendor/faiss/IndexIVF.h +353 -0
  39. data/vendor/faiss/IndexIVFFlat.cpp +502 -0
  40. data/vendor/faiss/IndexIVFFlat.h +118 -0
  41. data/vendor/faiss/IndexIVFPQ.cpp +1207 -0
  42. data/vendor/faiss/IndexIVFPQ.h +161 -0
  43. data/vendor/faiss/IndexIVFPQR.cpp +219 -0
  44. data/vendor/faiss/IndexIVFPQR.h +65 -0
  45. data/vendor/faiss/IndexIVFSpectralHash.cpp +331 -0
  46. data/vendor/faiss/IndexIVFSpectralHash.h +75 -0
  47. data/vendor/faiss/IndexLSH.cpp +225 -0
  48. data/vendor/faiss/IndexLSH.h +87 -0
  49. data/vendor/faiss/IndexLattice.cpp +143 -0
  50. data/vendor/faiss/IndexLattice.h +68 -0
  51. data/vendor/faiss/IndexPQ.cpp +1188 -0
  52. data/vendor/faiss/IndexPQ.h +199 -0
  53. data/vendor/faiss/IndexPreTransform.cpp +288 -0
  54. data/vendor/faiss/IndexPreTransform.h +91 -0
  55. data/vendor/faiss/IndexReplicas.cpp +123 -0
  56. data/vendor/faiss/IndexReplicas.h +76 -0
  57. data/vendor/faiss/IndexScalarQuantizer.cpp +317 -0
  58. data/vendor/faiss/IndexScalarQuantizer.h +127 -0
  59. data/vendor/faiss/IndexShards.cpp +317 -0
  60. data/vendor/faiss/IndexShards.h +100 -0
  61. data/vendor/faiss/InvertedLists.cpp +623 -0
  62. data/vendor/faiss/InvertedLists.h +334 -0
  63. data/vendor/faiss/LICENSE +21 -0
  64. data/vendor/faiss/MatrixStats.cpp +252 -0
  65. data/vendor/faiss/MatrixStats.h +62 -0
  66. data/vendor/faiss/MetaIndexes.cpp +351 -0
  67. data/vendor/faiss/MetaIndexes.h +126 -0
  68. data/vendor/faiss/OnDiskInvertedLists.cpp +674 -0
  69. data/vendor/faiss/OnDiskInvertedLists.h +127 -0
  70. data/vendor/faiss/VectorTransform.cpp +1157 -0
  71. data/vendor/faiss/VectorTransform.h +322 -0
  72. data/vendor/faiss/c_api/AutoTune_c.cpp +83 -0
  73. data/vendor/faiss/c_api/AutoTune_c.h +64 -0
  74. data/vendor/faiss/c_api/Clustering_c.cpp +139 -0
  75. data/vendor/faiss/c_api/Clustering_c.h +117 -0
  76. data/vendor/faiss/c_api/IndexFlat_c.cpp +140 -0
  77. data/vendor/faiss/c_api/IndexFlat_c.h +115 -0
  78. data/vendor/faiss/c_api/IndexIVFFlat_c.cpp +64 -0
  79. data/vendor/faiss/c_api/IndexIVFFlat_c.h +58 -0
  80. data/vendor/faiss/c_api/IndexIVF_c.cpp +92 -0
  81. data/vendor/faiss/c_api/IndexIVF_c.h +135 -0
  82. data/vendor/faiss/c_api/IndexLSH_c.cpp +37 -0
  83. data/vendor/faiss/c_api/IndexLSH_c.h +40 -0
  84. data/vendor/faiss/c_api/IndexShards_c.cpp +44 -0
  85. data/vendor/faiss/c_api/IndexShards_c.h +42 -0
  86. data/vendor/faiss/c_api/Index_c.cpp +105 -0
  87. data/vendor/faiss/c_api/Index_c.h +183 -0
  88. data/vendor/faiss/c_api/MetaIndexes_c.cpp +49 -0
  89. data/vendor/faiss/c_api/MetaIndexes_c.h +49 -0
  90. data/vendor/faiss/c_api/clone_index_c.cpp +23 -0
  91. data/vendor/faiss/c_api/clone_index_c.h +32 -0
  92. data/vendor/faiss/c_api/error_c.h +42 -0
  93. data/vendor/faiss/c_api/error_impl.cpp +27 -0
  94. data/vendor/faiss/c_api/error_impl.h +16 -0
  95. data/vendor/faiss/c_api/faiss_c.h +58 -0
  96. data/vendor/faiss/c_api/gpu/GpuAutoTune_c.cpp +96 -0
  97. data/vendor/faiss/c_api/gpu/GpuAutoTune_c.h +56 -0
  98. data/vendor/faiss/c_api/gpu/GpuClonerOptions_c.cpp +52 -0
  99. data/vendor/faiss/c_api/gpu/GpuClonerOptions_c.h +68 -0
  100. data/vendor/faiss/c_api/gpu/GpuIndex_c.cpp +17 -0
  101. data/vendor/faiss/c_api/gpu/GpuIndex_c.h +30 -0
  102. data/vendor/faiss/c_api/gpu/GpuIndicesOptions_c.h +38 -0
  103. data/vendor/faiss/c_api/gpu/GpuResources_c.cpp +86 -0
  104. data/vendor/faiss/c_api/gpu/GpuResources_c.h +66 -0
  105. data/vendor/faiss/c_api/gpu/StandardGpuResources_c.cpp +54 -0
  106. data/vendor/faiss/c_api/gpu/StandardGpuResources_c.h +53 -0
  107. data/vendor/faiss/c_api/gpu/macros_impl.h +42 -0
  108. data/vendor/faiss/c_api/impl/AuxIndexStructures_c.cpp +220 -0
  109. data/vendor/faiss/c_api/impl/AuxIndexStructures_c.h +149 -0
  110. data/vendor/faiss/c_api/index_factory_c.cpp +26 -0
  111. data/vendor/faiss/c_api/index_factory_c.h +30 -0
  112. data/vendor/faiss/c_api/index_io_c.cpp +42 -0
  113. data/vendor/faiss/c_api/index_io_c.h +50 -0
  114. data/vendor/faiss/c_api/macros_impl.h +110 -0
  115. data/vendor/faiss/clone_index.cpp +147 -0
  116. data/vendor/faiss/clone_index.h +38 -0
  117. data/vendor/faiss/demos/demo_imi_flat.cpp +151 -0
  118. data/vendor/faiss/demos/demo_imi_pq.cpp +199 -0
  119. data/vendor/faiss/demos/demo_ivfpq_indexing.cpp +146 -0
  120. data/vendor/faiss/demos/demo_sift1M.cpp +252 -0
  121. data/vendor/faiss/gpu/GpuAutoTune.cpp +95 -0
  122. data/vendor/faiss/gpu/GpuAutoTune.h +27 -0
  123. data/vendor/faiss/gpu/GpuCloner.cpp +403 -0
  124. data/vendor/faiss/gpu/GpuCloner.h +82 -0
  125. data/vendor/faiss/gpu/GpuClonerOptions.cpp +28 -0
  126. data/vendor/faiss/gpu/GpuClonerOptions.h +53 -0
  127. data/vendor/faiss/gpu/GpuDistance.h +52 -0
  128. data/vendor/faiss/gpu/GpuFaissAssert.h +29 -0
  129. data/vendor/faiss/gpu/GpuIndex.h +148 -0
  130. data/vendor/faiss/gpu/GpuIndexBinaryFlat.h +89 -0
  131. data/vendor/faiss/gpu/GpuIndexFlat.h +190 -0
  132. data/vendor/faiss/gpu/GpuIndexIVF.h +89 -0
  133. data/vendor/faiss/gpu/GpuIndexIVFFlat.h +85 -0
  134. data/vendor/faiss/gpu/GpuIndexIVFPQ.h +143 -0
  135. data/vendor/faiss/gpu/GpuIndexIVFScalarQuantizer.h +100 -0
  136. data/vendor/faiss/gpu/GpuIndicesOptions.h +30 -0
  137. data/vendor/faiss/gpu/GpuResources.cpp +52 -0
  138. data/vendor/faiss/gpu/GpuResources.h +73 -0
  139. data/vendor/faiss/gpu/StandardGpuResources.cpp +295 -0
  140. data/vendor/faiss/gpu/StandardGpuResources.h +114 -0
  141. data/vendor/faiss/gpu/impl/RemapIndices.cpp +43 -0
  142. data/vendor/faiss/gpu/impl/RemapIndices.h +24 -0
  143. data/vendor/faiss/gpu/perf/IndexWrapper-inl.h +71 -0
  144. data/vendor/faiss/gpu/perf/IndexWrapper.h +39 -0
  145. data/vendor/faiss/gpu/perf/PerfClustering.cpp +115 -0
  146. data/vendor/faiss/gpu/perf/PerfIVFPQAdd.cpp +139 -0
  147. data/vendor/faiss/gpu/perf/WriteIndex.cpp +102 -0
  148. data/vendor/faiss/gpu/test/TestGpuIndexBinaryFlat.cpp +130 -0
  149. data/vendor/faiss/gpu/test/TestGpuIndexFlat.cpp +371 -0
  150. data/vendor/faiss/gpu/test/TestGpuIndexIVFFlat.cpp +550 -0
  151. data/vendor/faiss/gpu/test/TestGpuIndexIVFPQ.cpp +450 -0
  152. data/vendor/faiss/gpu/test/TestGpuMemoryException.cpp +84 -0
  153. data/vendor/faiss/gpu/test/TestUtils.cpp +315 -0
  154. data/vendor/faiss/gpu/test/TestUtils.h +93 -0
  155. data/vendor/faiss/gpu/test/demo_ivfpq_indexing_gpu.cpp +159 -0
  156. data/vendor/faiss/gpu/utils/DeviceMemory.cpp +77 -0
  157. data/vendor/faiss/gpu/utils/DeviceMemory.h +71 -0
  158. data/vendor/faiss/gpu/utils/DeviceUtils.h +185 -0
  159. data/vendor/faiss/gpu/utils/MemorySpace.cpp +89 -0
  160. data/vendor/faiss/gpu/utils/MemorySpace.h +44 -0
  161. data/vendor/faiss/gpu/utils/StackDeviceMemory.cpp +239 -0
  162. data/vendor/faiss/gpu/utils/StackDeviceMemory.h +129 -0
  163. data/vendor/faiss/gpu/utils/StaticUtils.h +83 -0
  164. data/vendor/faiss/gpu/utils/Timer.cpp +60 -0
  165. data/vendor/faiss/gpu/utils/Timer.h +52 -0
  166. data/vendor/faiss/impl/AuxIndexStructures.cpp +305 -0
  167. data/vendor/faiss/impl/AuxIndexStructures.h +246 -0
  168. data/vendor/faiss/impl/FaissAssert.h +95 -0
  169. data/vendor/faiss/impl/FaissException.cpp +66 -0
  170. data/vendor/faiss/impl/FaissException.h +71 -0
  171. data/vendor/faiss/impl/HNSW.cpp +818 -0
  172. data/vendor/faiss/impl/HNSW.h +275 -0
  173. data/vendor/faiss/impl/PolysemousTraining.cpp +953 -0
  174. data/vendor/faiss/impl/PolysemousTraining.h +158 -0
  175. data/vendor/faiss/impl/ProductQuantizer.cpp +876 -0
  176. data/vendor/faiss/impl/ProductQuantizer.h +242 -0
  177. data/vendor/faiss/impl/ScalarQuantizer.cpp +1628 -0
  178. data/vendor/faiss/impl/ScalarQuantizer.h +120 -0
  179. data/vendor/faiss/impl/ThreadedIndex-inl.h +192 -0
  180. data/vendor/faiss/impl/ThreadedIndex.h +80 -0
  181. data/vendor/faiss/impl/index_read.cpp +793 -0
  182. data/vendor/faiss/impl/index_write.cpp +558 -0
  183. data/vendor/faiss/impl/io.cpp +142 -0
  184. data/vendor/faiss/impl/io.h +98 -0
  185. data/vendor/faiss/impl/lattice_Zn.cpp +712 -0
  186. data/vendor/faiss/impl/lattice_Zn.h +199 -0
  187. data/vendor/faiss/index_factory.cpp +392 -0
  188. data/vendor/faiss/index_factory.h +25 -0
  189. data/vendor/faiss/index_io.h +75 -0
  190. data/vendor/faiss/misc/test_blas.cpp +84 -0
  191. data/vendor/faiss/tests/test_binary_flat.cpp +64 -0
  192. data/vendor/faiss/tests/test_dealloc_invlists.cpp +183 -0
  193. data/vendor/faiss/tests/test_ivfpq_codec.cpp +67 -0
  194. data/vendor/faiss/tests/test_ivfpq_indexing.cpp +98 -0
  195. data/vendor/faiss/tests/test_lowlevel_ivf.cpp +566 -0
  196. data/vendor/faiss/tests/test_merge.cpp +258 -0
  197. data/vendor/faiss/tests/test_omp_threads.cpp +14 -0
  198. data/vendor/faiss/tests/test_ondisk_ivf.cpp +220 -0
  199. data/vendor/faiss/tests/test_pairs_decoding.cpp +189 -0
  200. data/vendor/faiss/tests/test_params_override.cpp +231 -0
  201. data/vendor/faiss/tests/test_pq_encoding.cpp +98 -0
  202. data/vendor/faiss/tests/test_sliding_ivf.cpp +240 -0
  203. data/vendor/faiss/tests/test_threaded_index.cpp +253 -0
  204. data/vendor/faiss/tests/test_transfer_invlists.cpp +159 -0
  205. data/vendor/faiss/tutorial/cpp/1-Flat.cpp +98 -0
  206. data/vendor/faiss/tutorial/cpp/2-IVFFlat.cpp +81 -0
  207. data/vendor/faiss/tutorial/cpp/3-IVFPQ.cpp +93 -0
  208. data/vendor/faiss/tutorial/cpp/4-GPU.cpp +119 -0
  209. data/vendor/faiss/tutorial/cpp/5-Multiple-GPUs.cpp +99 -0
  210. data/vendor/faiss/utils/Heap.cpp +122 -0
  211. data/vendor/faiss/utils/Heap.h +495 -0
  212. data/vendor/faiss/utils/WorkerThread.cpp +126 -0
  213. data/vendor/faiss/utils/WorkerThread.h +61 -0
  214. data/vendor/faiss/utils/distances.cpp +765 -0
  215. data/vendor/faiss/utils/distances.h +243 -0
  216. data/vendor/faiss/utils/distances_simd.cpp +809 -0
  217. data/vendor/faiss/utils/extra_distances.cpp +336 -0
  218. data/vendor/faiss/utils/extra_distances.h +54 -0
  219. data/vendor/faiss/utils/hamming-inl.h +472 -0
  220. data/vendor/faiss/utils/hamming.cpp +792 -0
  221. data/vendor/faiss/utils/hamming.h +220 -0
  222. data/vendor/faiss/utils/random.cpp +192 -0
  223. data/vendor/faiss/utils/random.h +60 -0
  224. data/vendor/faiss/utils/utils.cpp +783 -0
  225. data/vendor/faiss/utils/utils.h +181 -0
  226. metadata +216 -2
@@ -0,0 +1,336 @@
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/utils/distances.h>
11
+
12
+ #include <cmath>
13
+ #include <omp.h>
14
+
15
+
16
+ #include <faiss/utils/utils.h>
17
+ #include <faiss/impl/FaissAssert.h>
18
+ #include <faiss/impl/AuxIndexStructures.h>
19
+
20
+ namespace faiss {
21
+
22
+ /***************************************************************************
23
+ * Distance functions (other than L2 and IP)
24
+ ***************************************************************************/
25
+
26
+ struct VectorDistanceL2 {
27
+ size_t d;
28
+
29
+ float operator () (const float *x, const float *y) const {
30
+ return fvec_L2sqr (x, y, d);
31
+ }
32
+ };
33
+
34
+ struct VectorDistanceL1 {
35
+ size_t d;
36
+
37
+ float operator () (const float *x, const float *y) const {
38
+ return fvec_L1 (x, y, d);
39
+ }
40
+ };
41
+
42
+ struct VectorDistanceLinf {
43
+ size_t d;
44
+
45
+ float operator () (const float *x, const float *y) const {
46
+ return fvec_Linf (x, y, d);
47
+ /*
48
+ float vmax = 0;
49
+ for (size_t i = 0; i < d; i++) {
50
+ float diff = fabs (x[i] - y[i]);
51
+ if (diff > vmax) vmax = diff;
52
+ }
53
+ return vmax;*/
54
+ }
55
+ };
56
+
57
+ struct VectorDistanceLp {
58
+ size_t d;
59
+ const float p;
60
+
61
+ float operator () (const float *x, const float *y) const {
62
+ float accu = 0;
63
+ for (size_t i = 0; i < d; i++) {
64
+ float diff = fabs (x[i] - y[i]);
65
+ accu += powf (diff, p);
66
+ }
67
+ return accu;
68
+ }
69
+ };
70
+
71
+ struct VectorDistanceCanberra {
72
+ size_t d;
73
+
74
+ float operator () (const float *x, const float *y) const {
75
+ float accu = 0;
76
+ for (size_t i = 0; i < d; i++) {
77
+ float xi = x[i], yi = y[i];
78
+ accu += fabs (xi - yi) / (fabs(xi) + fabs(yi));
79
+ }
80
+ return accu;
81
+ }
82
+ };
83
+
84
+ struct VectorDistanceBrayCurtis {
85
+ size_t d;
86
+
87
+ float operator () (const float *x, const float *y) const {
88
+ float accu_num = 0, accu_den = 0;
89
+ for (size_t i = 0; i < d; i++) {
90
+ float xi = x[i], yi = y[i];
91
+ accu_num += fabs (xi - yi);
92
+ accu_den += fabs (xi + yi);
93
+ }
94
+ return accu_num / accu_den;
95
+ }
96
+ };
97
+
98
+ struct VectorDistanceJensenShannon {
99
+ size_t d;
100
+
101
+ float operator () (const float *x, const float *y) const {
102
+ float accu = 0;
103
+
104
+ for (size_t i = 0; i < d; i++) {
105
+ float xi = x[i], yi = y[i];
106
+ float mi = 0.5 * (xi + yi);
107
+ float kl1 = - xi * log(mi / xi);
108
+ float kl2 = - yi * log(mi / yi);
109
+ accu += kl1 + kl2;
110
+ }
111
+ return 0.5 * accu;
112
+ }
113
+ };
114
+
115
+
116
+
117
+
118
+
119
+
120
+
121
+
122
+
123
+
124
+ namespace {
125
+
126
+ template<class VD>
127
+ void pairwise_extra_distances_template (
128
+ VD vd,
129
+ int64_t nq, const float *xq,
130
+ int64_t nb, const float *xb,
131
+ float *dis,
132
+ int64_t ldq, int64_t ldb, int64_t ldd)
133
+ {
134
+
135
+ #pragma omp parallel for if(nq > 10)
136
+ for (int64_t i = 0; i < nq; i++) {
137
+ const float *xqi = xq + i * ldq;
138
+ const float *xbj = xb;
139
+ float *disi = dis + ldd * i;
140
+
141
+ for (int64_t j = 0; j < nb; j++) {
142
+ disi[j] = vd (xqi, xbj);
143
+ xbj += ldb;
144
+ }
145
+ }
146
+ }
147
+
148
+
149
+ template<class VD>
150
+ void knn_extra_metrics_template (
151
+ VD vd,
152
+ const float * x,
153
+ const float * y,
154
+ size_t nx, size_t ny,
155
+ float_maxheap_array_t * res)
156
+ {
157
+ size_t k = res->k;
158
+ size_t d = vd.d;
159
+ size_t check_period = InterruptCallback::get_period_hint (ny * d);
160
+ check_period *= omp_get_max_threads();
161
+
162
+ for (size_t i0 = 0; i0 < nx; i0 += check_period) {
163
+ size_t i1 = std::min(i0 + check_period, nx);
164
+
165
+ #pragma omp parallel for
166
+ for (size_t i = i0; i < i1; i++) {
167
+ const float * x_i = x + i * d;
168
+ const float * y_j = y;
169
+ size_t j;
170
+ float * simi = res->get_val(i);
171
+ int64_t * idxi = res->get_ids (i);
172
+
173
+ maxheap_heapify (k, simi, idxi);
174
+ for (j = 0; j < ny; j++) {
175
+ float disij = vd (x_i, y_j);
176
+
177
+ if (disij < simi[0]) {
178
+ maxheap_pop (k, simi, idxi);
179
+ maxheap_push (k, simi, idxi, disij, j);
180
+ }
181
+ y_j += d;
182
+ }
183
+ maxheap_reorder (k, simi, idxi);
184
+ }
185
+ InterruptCallback::check ();
186
+ }
187
+
188
+ }
189
+
190
+
191
+ template<class VD>
192
+ struct ExtraDistanceComputer : DistanceComputer {
193
+ VD vd;
194
+ Index::idx_t nb;
195
+ const float *q;
196
+ const float *b;
197
+
198
+ float operator () (idx_t i) override {
199
+ return vd (q, b + i * vd.d);
200
+ }
201
+
202
+ float symmetric_dis(idx_t i, idx_t j) override {
203
+ return vd (b + j * vd.d, b + i * vd.d);
204
+ }
205
+
206
+ ExtraDistanceComputer(const VD & vd, const float *xb,
207
+ size_t nb, const float *q = nullptr)
208
+ : vd(vd), nb(nb), q(q), b(xb) {}
209
+
210
+ void set_query(const float *x) override {
211
+ q = x;
212
+ }
213
+ };
214
+
215
+
216
+
217
+
218
+
219
+
220
+
221
+
222
+
223
+
224
+
225
+
226
+
227
+
228
+
229
+
230
+ } // anonymous namespace
231
+
232
+ void pairwise_extra_distances (
233
+ int64_t d,
234
+ int64_t nq, const float *xq,
235
+ int64_t nb, const float *xb,
236
+ MetricType mt, float metric_arg,
237
+ float *dis,
238
+ int64_t ldq, int64_t ldb, int64_t ldd)
239
+ {
240
+ if (nq == 0 || nb == 0) return;
241
+ if (ldq == -1) ldq = d;
242
+ if (ldb == -1) ldb = d;
243
+ if (ldd == -1) ldd = nb;
244
+
245
+ switch(mt) {
246
+ #define HANDLE_VAR(kw) \
247
+ case METRIC_ ## kw: { \
248
+ VectorDistance ## kw vd({(size_t)d}); \
249
+ pairwise_extra_distances_template (vd, nq, xq, nb, xb, \
250
+ dis, ldq, ldb, ldd); \
251
+ break; \
252
+ }
253
+ HANDLE_VAR(L2);
254
+ HANDLE_VAR(L1);
255
+ HANDLE_VAR(Linf);
256
+ HANDLE_VAR(Canberra);
257
+ HANDLE_VAR(BrayCurtis);
258
+ HANDLE_VAR(JensenShannon);
259
+ #undef HANDLE_VAR
260
+ case METRIC_Lp: {
261
+ VectorDistanceLp vd({(size_t)d, metric_arg});
262
+ pairwise_extra_distances_template (vd, nq, xq, nb, xb,
263
+ dis, ldq, ldb, ldd);
264
+ break;
265
+ }
266
+ default:
267
+ FAISS_THROW_MSG ("metric type not implemented");
268
+ }
269
+
270
+ }
271
+
272
+ void knn_extra_metrics (
273
+ const float * x,
274
+ const float * y,
275
+ size_t d, size_t nx, size_t ny,
276
+ MetricType mt, float metric_arg,
277
+ float_maxheap_array_t * res)
278
+ {
279
+
280
+ switch(mt) {
281
+ #define HANDLE_VAR(kw) \
282
+ case METRIC_ ## kw: { \
283
+ VectorDistance ## kw vd({(size_t)d}); \
284
+ knn_extra_metrics_template (vd, x, y, nx, ny, res); \
285
+ break; \
286
+ }
287
+ HANDLE_VAR(L2);
288
+ HANDLE_VAR(L1);
289
+ HANDLE_VAR(Linf);
290
+ HANDLE_VAR(Canberra);
291
+ HANDLE_VAR(BrayCurtis);
292
+ HANDLE_VAR(JensenShannon);
293
+ #undef HANDLE_VAR
294
+ case METRIC_Lp: {
295
+ VectorDistanceLp vd({(size_t)d, metric_arg});
296
+ knn_extra_metrics_template (vd, x, y, nx, ny, res);
297
+ break;
298
+ }
299
+ default:
300
+ FAISS_THROW_MSG ("metric type not implemented");
301
+ }
302
+
303
+ }
304
+
305
+ DistanceComputer *get_extra_distance_computer (
306
+ size_t d,
307
+ MetricType mt, float metric_arg,
308
+ size_t nb, const float *xb)
309
+ {
310
+
311
+ switch(mt) {
312
+ #define HANDLE_VAR(kw) \
313
+ case METRIC_ ## kw: { \
314
+ VectorDistance ## kw vd({(size_t)d}); \
315
+ return new ExtraDistanceComputer<VectorDistance ## kw>(vd, xb, nb); \
316
+ }
317
+ HANDLE_VAR(L2);
318
+ HANDLE_VAR(L1);
319
+ HANDLE_VAR(Linf);
320
+ HANDLE_VAR(Canberra);
321
+ HANDLE_VAR(BrayCurtis);
322
+ HANDLE_VAR(JensenShannon);
323
+ #undef HANDLE_VAR
324
+ case METRIC_Lp: {
325
+ VectorDistanceLp vd({(size_t)d, metric_arg});
326
+ return new ExtraDistanceComputer<VectorDistanceLp> (vd, xb, nb);
327
+ break;
328
+ }
329
+ default:
330
+ FAISS_THROW_MSG ("metric type not implemented");
331
+ }
332
+
333
+ }
334
+
335
+
336
+ } // namespace faiss
@@ -0,0 +1,54 @@
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
+ #ifndef FAISS_distances_h
11
+ #define FAISS_distances_h
12
+
13
+ /** In this file are the implementations of extra metrics beyond L2
14
+ * and inner product */
15
+
16
+ #include <stdint.h>
17
+
18
+ #include <faiss/Index.h>
19
+
20
+ #include <faiss/utils/Heap.h>
21
+
22
+
23
+
24
+ namespace faiss {
25
+
26
+
27
+ void pairwise_extra_distances (
28
+ int64_t d,
29
+ int64_t nq, const float *xq,
30
+ int64_t nb, const float *xb,
31
+ MetricType mt, float metric_arg,
32
+ float *dis,
33
+ int64_t ldq = -1, int64_t ldb = -1, int64_t ldd = -1);
34
+
35
+
36
+ void knn_extra_metrics (
37
+ const float * x,
38
+ const float * y,
39
+ size_t d, size_t nx, size_t ny,
40
+ MetricType mt, float metric_arg,
41
+ float_maxheap_array_t * res);
42
+
43
+
44
+ /** get a DistanceComputer that refers to this type of distance and
45
+ * indexes a flat array of size nb */
46
+ DistanceComputer *get_extra_distance_computer (
47
+ size_t d,
48
+ MetricType mt, float metric_arg,
49
+ size_t nb, const float *xb);
50
+
51
+ }
52
+
53
+
54
+ #endif
@@ -0,0 +1,472 @@
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
+
9
+
10
+ namespace faiss {
11
+
12
+
13
+ inline BitstringWriter::BitstringWriter(uint8_t *code, int code_size):
14
+ code (code), code_size (code_size), i(0)
15
+ {
16
+ bzero (code, code_size);
17
+ }
18
+
19
+ inline void BitstringWriter::write(uint64_t x, int nbit) {
20
+ assert (code_size * 8 >= nbit + i);
21
+ // nb of available bits in i / 8
22
+ int na = 8 - (i & 7);
23
+
24
+ if (nbit <= na) {
25
+ code[i >> 3] |= x << (i & 7);
26
+ i += nbit;
27
+ return;
28
+ } else {
29
+ int j = i >> 3;
30
+ code[j++] |= x << (i & 7);
31
+ i += nbit;
32
+ x >>= na;
33
+ while (x != 0) {
34
+ code[j++] |= x;
35
+ x >>= 8;
36
+ }
37
+ }
38
+ }
39
+
40
+
41
+ inline BitstringReader::BitstringReader(const uint8_t *code, int code_size):
42
+ code (code), code_size (code_size), i(0)
43
+ {}
44
+
45
+ inline uint64_t BitstringReader::read(int nbit) {
46
+ assert (code_size * 8 >= nbit + i);
47
+ // nb of available bits in i / 8
48
+ int na = 8 - (i & 7);
49
+ // get available bits in current byte
50
+ uint64_t res = code[i >> 3] >> (i & 7);
51
+ if (nbit <= na) {
52
+ res &= (1 << nbit) - 1;
53
+ i += nbit;
54
+ return res;
55
+ } else {
56
+ int ofs = na;
57
+ int j = (i >> 3) + 1;
58
+ i += nbit;
59
+ nbit -= na;
60
+ while (nbit > 8) {
61
+ res |= ((uint64_t)code[j++]) << ofs;
62
+ ofs += 8;
63
+ nbit -= 8; // TODO remove nbit
64
+ }
65
+ uint64_t last_byte = code[j];
66
+ last_byte &= (1 << nbit) - 1;
67
+ res |= last_byte << ofs;
68
+ return res;
69
+ }
70
+ }
71
+
72
+
73
+ /******************************************************************
74
+ * The HammingComputer series of classes compares a single code of
75
+ * size 4 to 32 to incoming codes. They are intended for use as a
76
+ * template class where it would be inefficient to switch on the code
77
+ * size in the inner loop. Hopefully the compiler will inline the
78
+ * hamming() functions and put the a0, a1, ... in registers.
79
+ ******************************************************************/
80
+
81
+
82
+ struct HammingComputer4 {
83
+ uint32_t a0;
84
+
85
+ HammingComputer4 () {}
86
+
87
+ HammingComputer4 (const uint8_t *a, int code_size) {
88
+ set (a, code_size);
89
+ }
90
+
91
+ void set (const uint8_t *a, int code_size) {
92
+ assert (code_size == 4);
93
+ a0 = *(uint32_t *)a;
94
+ }
95
+
96
+ inline int hamming (const uint8_t *b) const {
97
+ return popcount64 (*(uint32_t *)b ^ a0);
98
+ }
99
+
100
+ };
101
+
102
+ struct HammingComputer8 {
103
+ uint64_t a0;
104
+
105
+ HammingComputer8 () {}
106
+
107
+ HammingComputer8 (const uint8_t *a, int code_size) {
108
+ set (a, code_size);
109
+ }
110
+
111
+ void set (const uint8_t *a, int code_size) {
112
+ assert (code_size == 8);
113
+ a0 = *(uint64_t *)a;
114
+ }
115
+
116
+ inline int hamming (const uint8_t *b) const {
117
+ return popcount64 (*(uint64_t *)b ^ a0);
118
+ }
119
+
120
+ };
121
+
122
+
123
+ struct HammingComputer16 {
124
+ uint64_t a0, a1;
125
+
126
+ HammingComputer16 () {}
127
+
128
+ HammingComputer16 (const uint8_t *a8, int code_size) {
129
+ set (a8, code_size);
130
+ }
131
+
132
+ void set (const uint8_t *a8, int code_size) {
133
+ assert (code_size == 16);
134
+ const uint64_t *a = (uint64_t *)a8;
135
+ a0 = a[0]; a1 = a[1];
136
+ }
137
+
138
+ inline int hamming (const uint8_t *b8) const {
139
+ const uint64_t *b = (uint64_t *)b8;
140
+ return popcount64 (b[0] ^ a0) + popcount64 (b[1] ^ a1);
141
+ }
142
+
143
+ };
144
+
145
+ // when applied to an array, 1/2 of the 64-bit accesses are unaligned.
146
+ // This incurs a penalty of ~10% wrt. fully aligned accesses.
147
+ struct HammingComputer20 {
148
+ uint64_t a0, a1;
149
+ uint32_t a2;
150
+
151
+ HammingComputer20 () {}
152
+
153
+ HammingComputer20 (const uint8_t *a8, int code_size) {
154
+ set (a8, code_size);
155
+ }
156
+
157
+ void set (const uint8_t *a8, int code_size) {
158
+ assert (code_size == 20);
159
+ const uint64_t *a = (uint64_t *)a8;
160
+ a0 = a[0]; a1 = a[1]; a2 = a[2];
161
+ }
162
+
163
+ inline int hamming (const uint8_t *b8) const {
164
+ const uint64_t *b = (uint64_t *)b8;
165
+ return popcount64 (b[0] ^ a0) + popcount64 (b[1] ^ a1) +
166
+ popcount64 (*(uint32_t*)(b + 2) ^ a2);
167
+ }
168
+ };
169
+
170
+ struct HammingComputer32 {
171
+ uint64_t a0, a1, a2, a3;
172
+
173
+ HammingComputer32 () {}
174
+
175
+ HammingComputer32 (const uint8_t *a8, int code_size) {
176
+ set (a8, code_size);
177
+ }
178
+
179
+ void set (const uint8_t *a8, int code_size) {
180
+ assert (code_size == 32);
181
+ const uint64_t *a = (uint64_t *)a8;
182
+ a0 = a[0]; a1 = a[1]; a2 = a[2]; a3 = a[3];
183
+ }
184
+
185
+ inline int hamming (const uint8_t *b8) const {
186
+ const uint64_t *b = (uint64_t *)b8;
187
+ return popcount64 (b[0] ^ a0) + popcount64 (b[1] ^ a1) +
188
+ popcount64 (b[2] ^ a2) + popcount64 (b[3] ^ a3);
189
+ }
190
+
191
+ };
192
+
193
+ struct HammingComputer64 {
194
+ uint64_t a0, a1, a2, a3, a4, a5, a6, a7;
195
+
196
+ HammingComputer64 () {}
197
+
198
+ HammingComputer64 (const uint8_t *a8, int code_size) {
199
+ set (a8, code_size);
200
+ }
201
+
202
+ void set (const uint8_t *a8, int code_size) {
203
+ assert (code_size == 64);
204
+ const uint64_t *a = (uint64_t *)a8;
205
+ a0 = a[0]; a1 = a[1]; a2 = a[2]; a3 = a[3];
206
+ a4 = a[4]; a5 = a[5]; a6 = a[6]; a7 = a[7];
207
+ }
208
+
209
+ inline int hamming (const uint8_t *b8) const {
210
+ const uint64_t *b = (uint64_t *)b8;
211
+ return popcount64 (b[0] ^ a0) + popcount64 (b[1] ^ a1) +
212
+ popcount64 (b[2] ^ a2) + popcount64 (b[3] ^ a3) +
213
+ popcount64 (b[4] ^ a4) + popcount64 (b[5] ^ a5) +
214
+ popcount64 (b[6] ^ a6) + popcount64 (b[7] ^ a7);
215
+ }
216
+
217
+ };
218
+
219
+ // very inefficient...
220
+ struct HammingComputerDefault {
221
+ const uint8_t *a;
222
+ int n;
223
+
224
+ HammingComputerDefault () {}
225
+
226
+ HammingComputerDefault (const uint8_t *a8, int code_size) {
227
+ set (a8, code_size);
228
+ }
229
+
230
+ void set (const uint8_t *a8, int code_size) {
231
+ a = a8;
232
+ n = code_size;
233
+ }
234
+
235
+ int hamming (const uint8_t *b8) const {
236
+ int accu = 0;
237
+ for (int i = 0; i < n; i++)
238
+ accu += popcount64 (a[i] ^ b8[i]);
239
+ return accu;
240
+ }
241
+
242
+ };
243
+
244
+ struct HammingComputerM8 {
245
+ const uint64_t *a;
246
+ int n;
247
+
248
+ HammingComputerM8 () {}
249
+
250
+ HammingComputerM8 (const uint8_t *a8, int code_size) {
251
+ set (a8, code_size);
252
+ }
253
+
254
+ void set (const uint8_t *a8, int code_size) {
255
+ assert (code_size % 8 == 0);
256
+ a = (uint64_t *)a8;
257
+ n = code_size / 8;
258
+ }
259
+
260
+ int hamming (const uint8_t *b8) const {
261
+ const uint64_t *b = (uint64_t *)b8;
262
+ int accu = 0;
263
+ for (int i = 0; i < n; i++)
264
+ accu += popcount64 (a[i] ^ b[i]);
265
+ return accu;
266
+ }
267
+
268
+ };
269
+
270
+ // even more inefficient!
271
+ struct HammingComputerM4 {
272
+ const uint32_t *a;
273
+ int n;
274
+
275
+ HammingComputerM4 () {}
276
+
277
+ HammingComputerM4 (const uint8_t *a4, int code_size) {
278
+ set (a4, code_size);
279
+ }
280
+
281
+ void set (const uint8_t *a4, int code_size) {
282
+ assert (code_size % 4 == 0);
283
+ a = (uint32_t *)a4;
284
+ n = code_size / 4;
285
+ }
286
+
287
+ int hamming (const uint8_t *b8) const {
288
+ const uint32_t *b = (uint32_t *)b8;
289
+ int accu = 0;
290
+ for (int i = 0; i < n; i++)
291
+ accu += popcount64 (a[i] ^ b[i]);
292
+ return accu;
293
+ }
294
+
295
+ };
296
+
297
+ /***************************************************************************
298
+ * Equivalence with a template class when code size is known at compile time
299
+ **************************************************************************/
300
+
301
+ // default template
302
+ template<int CODE_SIZE>
303
+ struct HammingComputer: HammingComputerM8 {
304
+ HammingComputer (const uint8_t *a, int code_size):
305
+ HammingComputerM8(a, code_size) {}
306
+ };
307
+
308
+ #define SPECIALIZED_HC(CODE_SIZE) \
309
+ template<> struct HammingComputer<CODE_SIZE>: \
310
+ HammingComputer ## CODE_SIZE { \
311
+ HammingComputer (const uint8_t *a): \
312
+ HammingComputer ## CODE_SIZE(a, CODE_SIZE) {} \
313
+ }
314
+
315
+ SPECIALIZED_HC(4);
316
+ SPECIALIZED_HC(8);
317
+ SPECIALIZED_HC(16);
318
+ SPECIALIZED_HC(20);
319
+ SPECIALIZED_HC(32);
320
+ SPECIALIZED_HC(64);
321
+
322
+ #undef SPECIALIZED_HC
323
+
324
+
325
+ /***************************************************************************
326
+ * generalized Hamming = number of bytes that are different between
327
+ * two codes.
328
+ ***************************************************************************/
329
+
330
+
331
+ inline int generalized_hamming_64 (uint64_t a) {
332
+ a |= a >> 1;
333
+ a |= a >> 2;
334
+ a |= a >> 4;
335
+ a &= 0x0101010101010101UL;
336
+ return popcount64 (a);
337
+ }
338
+
339
+
340
+ struct GenHammingComputer8 {
341
+ uint64_t a0;
342
+
343
+ GenHammingComputer8 (const uint8_t *a, int code_size) {
344
+ assert (code_size == 8);
345
+ a0 = *(uint64_t *)a;
346
+ }
347
+
348
+ inline int hamming (const uint8_t *b) const {
349
+ return generalized_hamming_64 (*(uint64_t *)b ^ a0);
350
+ }
351
+
352
+ };
353
+
354
+
355
+ struct GenHammingComputer16 {
356
+ uint64_t a0, a1;
357
+ GenHammingComputer16 (const uint8_t *a8, int code_size) {
358
+ assert (code_size == 16);
359
+ const uint64_t *a = (uint64_t *)a8;
360
+ a0 = a[0]; a1 = a[1];
361
+ }
362
+
363
+ inline int hamming (const uint8_t *b8) const {
364
+ const uint64_t *b = (uint64_t *)b8;
365
+ return generalized_hamming_64 (b[0] ^ a0) +
366
+ generalized_hamming_64 (b[1] ^ a1);
367
+ }
368
+
369
+ };
370
+
371
+ struct GenHammingComputer32 {
372
+ uint64_t a0, a1, a2, a3;
373
+
374
+ GenHammingComputer32 (const uint8_t *a8, int code_size) {
375
+ assert (code_size == 32);
376
+ const uint64_t *a = (uint64_t *)a8;
377
+ a0 = a[0]; a1 = a[1]; a2 = a[2]; a3 = a[3];
378
+ }
379
+
380
+ inline int hamming (const uint8_t *b8) const {
381
+ const uint64_t *b = (uint64_t *)b8;
382
+ return generalized_hamming_64 (b[0] ^ a0) +
383
+ generalized_hamming_64 (b[1] ^ a1) +
384
+ generalized_hamming_64 (b[2] ^ a2) +
385
+ generalized_hamming_64 (b[3] ^ a3);
386
+ }
387
+
388
+ };
389
+
390
+ struct GenHammingComputerM8 {
391
+ const uint64_t *a;
392
+ int n;
393
+
394
+ GenHammingComputerM8 (const uint8_t *a8, int code_size) {
395
+ assert (code_size % 8 == 0);
396
+ a = (uint64_t *)a8;
397
+ n = code_size / 8;
398
+ }
399
+
400
+ int hamming (const uint8_t *b8) const {
401
+ const uint64_t *b = (uint64_t *)b8;
402
+ int accu = 0;
403
+ for (int i = 0; i < n; i++)
404
+ accu += generalized_hamming_64 (a[i] ^ b[i]);
405
+ return accu;
406
+ }
407
+
408
+ };
409
+
410
+
411
+ /** generalized Hamming distances (= count number of code bytes that
412
+ are the same) */
413
+ void generalized_hammings_knn_hc (
414
+ int_maxheap_array_t * ha,
415
+ const uint8_t * a,
416
+ const uint8_t * b,
417
+ size_t nb,
418
+ size_t code_size,
419
+ int ordered = true);
420
+
421
+
422
+
423
+ /** This class maintains a list of best distances seen so far.
424
+ *
425
+ * Since the distances are in a limited range (0 to nbit), the
426
+ * object maintains one list per possible distance, and fills
427
+ * in only the n-first lists, such that the sum of sizes of the
428
+ * n lists is below k.
429
+ */
430
+ template<class HammingComputer>
431
+ struct HCounterState {
432
+ int *counters;
433
+ int64_t *ids_per_dis;
434
+
435
+ HammingComputer hc;
436
+ int thres;
437
+ int count_lt;
438
+ int count_eq;
439
+ int k;
440
+
441
+ HCounterState(int *counters, int64_t *ids_per_dis,
442
+ const uint8_t *x, int d, int k)
443
+ : counters(counters),
444
+ ids_per_dis(ids_per_dis),
445
+ hc(x, d / 8),
446
+ thres(d + 1),
447
+ count_lt(0),
448
+ count_eq(0),
449
+ k(k) {}
450
+
451
+ void update_counter(const uint8_t *y, size_t j) {
452
+ int32_t dis = hc.hamming(y);
453
+
454
+ if (dis <= thres) {
455
+ if (dis < thres) {
456
+ ids_per_dis[dis * k + counters[dis]++] = j;
457
+ ++count_lt;
458
+ while (count_lt == k && thres > 0) {
459
+ --thres;
460
+ count_eq = counters[thres];
461
+ count_lt -= count_eq;
462
+ }
463
+ } else if (count_eq < k) {
464
+ ids_per_dis[dis * k + count_eq++] = j;
465
+ counters[dis] = count_eq;
466
+ }
467
+ }
468
+ }
469
+ };
470
+
471
+
472
+ } // namespace faiss