faiss 0.2.0 → 0.2.4
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +16 -0
- data/LICENSE.txt +1 -1
- data/README.md +7 -7
- data/ext/faiss/extconf.rb +6 -3
- data/ext/faiss/numo.hpp +4 -4
- data/ext/faiss/utils.cpp +1 -1
- data/ext/faiss/utils.h +1 -1
- data/lib/faiss/version.rb +1 -1
- data/vendor/faiss/faiss/AutoTune.cpp +292 -291
- data/vendor/faiss/faiss/AutoTune.h +55 -56
- data/vendor/faiss/faiss/Clustering.cpp +365 -194
- data/vendor/faiss/faiss/Clustering.h +102 -35
- data/vendor/faiss/faiss/IVFlib.cpp +171 -195
- data/vendor/faiss/faiss/IVFlib.h +48 -51
- data/vendor/faiss/faiss/Index.cpp +85 -103
- data/vendor/faiss/faiss/Index.h +54 -48
- data/vendor/faiss/faiss/Index2Layer.cpp +126 -224
- data/vendor/faiss/faiss/Index2Layer.h +22 -36
- data/vendor/faiss/faiss/IndexAdditiveQuantizer.cpp +407 -0
- data/vendor/faiss/faiss/IndexAdditiveQuantizer.h +195 -0
- data/vendor/faiss/faiss/IndexBinary.cpp +45 -37
- data/vendor/faiss/faiss/IndexBinary.h +140 -132
- data/vendor/faiss/faiss/IndexBinaryFlat.cpp +73 -53
- data/vendor/faiss/faiss/IndexBinaryFlat.h +29 -24
- data/vendor/faiss/faiss/IndexBinaryFromFloat.cpp +46 -43
- data/vendor/faiss/faiss/IndexBinaryFromFloat.h +16 -15
- data/vendor/faiss/faiss/IndexBinaryHNSW.cpp +215 -232
- data/vendor/faiss/faiss/IndexBinaryHNSW.h +25 -24
- data/vendor/faiss/faiss/IndexBinaryHash.cpp +182 -177
- data/vendor/faiss/faiss/IndexBinaryHash.h +41 -34
- data/vendor/faiss/faiss/IndexBinaryIVF.cpp +489 -461
- data/vendor/faiss/faiss/IndexBinaryIVF.h +97 -68
- data/vendor/faiss/faiss/IndexFlat.cpp +115 -176
- data/vendor/faiss/faiss/IndexFlat.h +42 -59
- data/vendor/faiss/faiss/IndexFlatCodes.cpp +67 -0
- data/vendor/faiss/faiss/IndexFlatCodes.h +47 -0
- data/vendor/faiss/faiss/IndexHNSW.cpp +372 -348
- data/vendor/faiss/faiss/IndexHNSW.h +57 -41
- data/vendor/faiss/faiss/IndexIVF.cpp +545 -453
- data/vendor/faiss/faiss/IndexIVF.h +169 -118
- data/vendor/faiss/faiss/IndexIVFAdditiveQuantizer.cpp +316 -0
- data/vendor/faiss/faiss/IndexIVFAdditiveQuantizer.h +121 -0
- data/vendor/faiss/faiss/IndexIVFFlat.cpp +247 -252
- data/vendor/faiss/faiss/IndexIVFFlat.h +48 -51
- data/vendor/faiss/faiss/IndexIVFPQ.cpp +459 -517
- data/vendor/faiss/faiss/IndexIVFPQ.h +75 -67
- data/vendor/faiss/faiss/IndexIVFPQFastScan.cpp +406 -372
- data/vendor/faiss/faiss/IndexIVFPQFastScan.h +82 -57
- data/vendor/faiss/faiss/IndexIVFPQR.cpp +104 -102
- data/vendor/faiss/faiss/IndexIVFPQR.h +33 -28
- data/vendor/faiss/faiss/IndexIVFSpectralHash.cpp +163 -150
- data/vendor/faiss/faiss/IndexIVFSpectralHash.h +38 -25
- data/vendor/faiss/faiss/IndexLSH.cpp +66 -113
- data/vendor/faiss/faiss/IndexLSH.h +20 -38
- data/vendor/faiss/faiss/IndexLattice.cpp +42 -56
- data/vendor/faiss/faiss/IndexLattice.h +11 -16
- data/vendor/faiss/faiss/IndexNNDescent.cpp +229 -0
- data/vendor/faiss/faiss/IndexNNDescent.h +72 -0
- data/vendor/faiss/faiss/IndexNSG.cpp +301 -0
- data/vendor/faiss/faiss/IndexNSG.h +85 -0
- data/vendor/faiss/faiss/IndexPQ.cpp +387 -495
- data/vendor/faiss/faiss/IndexPQ.h +64 -82
- data/vendor/faiss/faiss/IndexPQFastScan.cpp +143 -170
- data/vendor/faiss/faiss/IndexPQFastScan.h +46 -32
- data/vendor/faiss/faiss/IndexPreTransform.cpp +120 -150
- data/vendor/faiss/faiss/IndexPreTransform.h +33 -36
- data/vendor/faiss/faiss/IndexRefine.cpp +139 -127
- data/vendor/faiss/faiss/IndexRefine.h +32 -23
- data/vendor/faiss/faiss/IndexReplicas.cpp +147 -153
- data/vendor/faiss/faiss/IndexReplicas.h +62 -56
- data/vendor/faiss/faiss/IndexScalarQuantizer.cpp +111 -172
- data/vendor/faiss/faiss/IndexScalarQuantizer.h +41 -59
- data/vendor/faiss/faiss/IndexShards.cpp +256 -240
- data/vendor/faiss/faiss/IndexShards.h +85 -73
- data/vendor/faiss/faiss/MatrixStats.cpp +112 -97
- data/vendor/faiss/faiss/MatrixStats.h +7 -10
- data/vendor/faiss/faiss/MetaIndexes.cpp +135 -157
- data/vendor/faiss/faiss/MetaIndexes.h +40 -34
- data/vendor/faiss/faiss/MetricType.h +7 -7
- data/vendor/faiss/faiss/VectorTransform.cpp +654 -475
- data/vendor/faiss/faiss/VectorTransform.h +64 -89
- data/vendor/faiss/faiss/clone_index.cpp +78 -73
- data/vendor/faiss/faiss/clone_index.h +4 -9
- data/vendor/faiss/faiss/gpu/GpuAutoTune.cpp +33 -38
- data/vendor/faiss/faiss/gpu/GpuAutoTune.h +11 -9
- data/vendor/faiss/faiss/gpu/GpuCloner.cpp +198 -171
- data/vendor/faiss/faiss/gpu/GpuCloner.h +53 -35
- data/vendor/faiss/faiss/gpu/GpuClonerOptions.cpp +12 -14
- data/vendor/faiss/faiss/gpu/GpuClonerOptions.h +27 -25
- data/vendor/faiss/faiss/gpu/GpuDistance.h +116 -112
- data/vendor/faiss/faiss/gpu/GpuFaissAssert.h +1 -2
- data/vendor/faiss/faiss/gpu/GpuIcmEncoder.h +60 -0
- data/vendor/faiss/faiss/gpu/GpuIndex.h +134 -137
- data/vendor/faiss/faiss/gpu/GpuIndexBinaryFlat.h +76 -73
- data/vendor/faiss/faiss/gpu/GpuIndexFlat.h +173 -162
- data/vendor/faiss/faiss/gpu/GpuIndexIVF.h +67 -64
- data/vendor/faiss/faiss/gpu/GpuIndexIVFFlat.h +89 -86
- data/vendor/faiss/faiss/gpu/GpuIndexIVFPQ.h +150 -141
- data/vendor/faiss/faiss/gpu/GpuIndexIVFScalarQuantizer.h +101 -103
- data/vendor/faiss/faiss/gpu/GpuIndicesOptions.h +17 -16
- data/vendor/faiss/faiss/gpu/GpuResources.cpp +116 -128
- data/vendor/faiss/faiss/gpu/GpuResources.h +182 -186
- data/vendor/faiss/faiss/gpu/StandardGpuResources.cpp +433 -422
- data/vendor/faiss/faiss/gpu/StandardGpuResources.h +131 -130
- data/vendor/faiss/faiss/gpu/impl/InterleavedCodes.cpp +468 -456
- data/vendor/faiss/faiss/gpu/impl/InterleavedCodes.h +25 -19
- data/vendor/faiss/faiss/gpu/impl/RemapIndices.cpp +22 -20
- data/vendor/faiss/faiss/gpu/impl/RemapIndices.h +9 -8
- data/vendor/faiss/faiss/gpu/perf/IndexWrapper-inl.h +39 -44
- data/vendor/faiss/faiss/gpu/perf/IndexWrapper.h +16 -14
- data/vendor/faiss/faiss/gpu/perf/PerfClustering.cpp +77 -71
- data/vendor/faiss/faiss/gpu/perf/PerfIVFPQAdd.cpp +109 -88
- data/vendor/faiss/faiss/gpu/perf/WriteIndex.cpp +75 -64
- data/vendor/faiss/faiss/gpu/test/TestCodePacking.cpp +230 -215
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexBinaryFlat.cpp +80 -86
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexFlat.cpp +284 -277
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexIVFFlat.cpp +416 -416
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexIVFPQ.cpp +611 -517
- data/vendor/faiss/faiss/gpu/test/TestGpuIndexIVFScalarQuantizer.cpp +166 -164
- data/vendor/faiss/faiss/gpu/test/TestGpuMemoryException.cpp +61 -53
- data/vendor/faiss/faiss/gpu/test/TestUtils.cpp +274 -238
- data/vendor/faiss/faiss/gpu/test/TestUtils.h +73 -57
- data/vendor/faiss/faiss/gpu/test/demo_ivfpq_indexing_gpu.cpp +47 -50
- data/vendor/faiss/faiss/gpu/utils/DeviceUtils.h +79 -72
- data/vendor/faiss/faiss/gpu/utils/StackDeviceMemory.cpp +140 -146
- data/vendor/faiss/faiss/gpu/utils/StackDeviceMemory.h +69 -71
- data/vendor/faiss/faiss/gpu/utils/StaticUtils.h +21 -16
- data/vendor/faiss/faiss/gpu/utils/Timer.cpp +25 -29
- data/vendor/faiss/faiss/gpu/utils/Timer.h +30 -29
- data/vendor/faiss/faiss/impl/AdditiveQuantizer.cpp +503 -0
- data/vendor/faiss/faiss/impl/AdditiveQuantizer.h +175 -0
- data/vendor/faiss/faiss/impl/AuxIndexStructures.cpp +90 -120
- data/vendor/faiss/faiss/impl/AuxIndexStructures.h +81 -65
- data/vendor/faiss/faiss/impl/FaissAssert.h +73 -58
- data/vendor/faiss/faiss/impl/FaissException.cpp +56 -48
- data/vendor/faiss/faiss/impl/FaissException.h +41 -29
- data/vendor/faiss/faiss/impl/HNSW.cpp +606 -617
- data/vendor/faiss/faiss/impl/HNSW.h +179 -200
- data/vendor/faiss/faiss/impl/LocalSearchQuantizer.cpp +855 -0
- data/vendor/faiss/faiss/impl/LocalSearchQuantizer.h +244 -0
- data/vendor/faiss/faiss/impl/NNDescent.cpp +487 -0
- data/vendor/faiss/faiss/impl/NNDescent.h +154 -0
- data/vendor/faiss/faiss/impl/NSG.cpp +679 -0
- data/vendor/faiss/faiss/impl/NSG.h +199 -0
- data/vendor/faiss/faiss/impl/PolysemousTraining.cpp +484 -454
- data/vendor/faiss/faiss/impl/PolysemousTraining.h +52 -55
- data/vendor/faiss/faiss/impl/ProductQuantizer-inl.h +26 -47
- data/vendor/faiss/faiss/impl/ProductQuantizer.cpp +469 -459
- data/vendor/faiss/faiss/impl/ProductQuantizer.h +76 -87
- data/vendor/faiss/faiss/impl/ResidualQuantizer.cpp +758 -0
- data/vendor/faiss/faiss/impl/ResidualQuantizer.h +188 -0
- data/vendor/faiss/faiss/impl/ResultHandler.h +96 -132
- data/vendor/faiss/faiss/impl/ScalarQuantizer.cpp +647 -707
- data/vendor/faiss/faiss/impl/ScalarQuantizer.h +48 -46
- data/vendor/faiss/faiss/impl/ThreadedIndex-inl.h +129 -131
- data/vendor/faiss/faiss/impl/ThreadedIndex.h +61 -55
- data/vendor/faiss/faiss/impl/index_read.cpp +631 -480
- data/vendor/faiss/faiss/impl/index_write.cpp +547 -407
- data/vendor/faiss/faiss/impl/io.cpp +76 -95
- data/vendor/faiss/faiss/impl/io.h +31 -41
- data/vendor/faiss/faiss/impl/io_macros.h +60 -29
- data/vendor/faiss/faiss/impl/kmeans1d.cpp +301 -0
- data/vendor/faiss/faiss/impl/kmeans1d.h +48 -0
- data/vendor/faiss/faiss/impl/lattice_Zn.cpp +137 -186
- data/vendor/faiss/faiss/impl/lattice_Zn.h +40 -51
- data/vendor/faiss/faiss/impl/platform_macros.h +29 -8
- data/vendor/faiss/faiss/impl/pq4_fast_scan.cpp +77 -124
- data/vendor/faiss/faiss/impl/pq4_fast_scan.h +39 -48
- data/vendor/faiss/faiss/impl/pq4_fast_scan_search_1.cpp +41 -52
- data/vendor/faiss/faiss/impl/pq4_fast_scan_search_qbs.cpp +80 -117
- data/vendor/faiss/faiss/impl/simd_result_handlers.h +109 -137
- data/vendor/faiss/faiss/index_factory.cpp +619 -397
- data/vendor/faiss/faiss/index_factory.h +8 -6
- data/vendor/faiss/faiss/index_io.h +23 -26
- data/vendor/faiss/faiss/invlists/BlockInvertedLists.cpp +67 -75
- data/vendor/faiss/faiss/invlists/BlockInvertedLists.h +22 -24
- data/vendor/faiss/faiss/invlists/DirectMap.cpp +96 -112
- data/vendor/faiss/faiss/invlists/DirectMap.h +29 -33
- data/vendor/faiss/faiss/invlists/InvertedLists.cpp +307 -364
- data/vendor/faiss/faiss/invlists/InvertedLists.h +151 -151
- data/vendor/faiss/faiss/invlists/InvertedListsIOHook.cpp +29 -34
- data/vendor/faiss/faiss/invlists/InvertedListsIOHook.h +17 -18
- data/vendor/faiss/faiss/invlists/OnDiskInvertedLists.cpp +257 -293
- data/vendor/faiss/faiss/invlists/OnDiskInvertedLists.h +50 -45
- data/vendor/faiss/faiss/python/python_callbacks.cpp +23 -26
- data/vendor/faiss/faiss/python/python_callbacks.h +9 -16
- data/vendor/faiss/faiss/utils/AlignedTable.h +79 -44
- data/vendor/faiss/faiss/utils/Heap.cpp +40 -48
- data/vendor/faiss/faiss/utils/Heap.h +186 -209
- data/vendor/faiss/faiss/utils/WorkerThread.cpp +67 -76
- data/vendor/faiss/faiss/utils/WorkerThread.h +32 -33
- data/vendor/faiss/faiss/utils/distances.cpp +305 -312
- data/vendor/faiss/faiss/utils/distances.h +170 -122
- data/vendor/faiss/faiss/utils/distances_simd.cpp +498 -508
- data/vendor/faiss/faiss/utils/extra_distances-inl.h +117 -0
- data/vendor/faiss/faiss/utils/extra_distances.cpp +113 -232
- data/vendor/faiss/faiss/utils/extra_distances.h +30 -29
- data/vendor/faiss/faiss/utils/hamming-inl.h +260 -209
- data/vendor/faiss/faiss/utils/hamming.cpp +375 -469
- data/vendor/faiss/faiss/utils/hamming.h +62 -85
- data/vendor/faiss/faiss/utils/ordered_key_value.h +16 -18
- data/vendor/faiss/faiss/utils/partitioning.cpp +393 -318
- data/vendor/faiss/faiss/utils/partitioning.h +26 -21
- data/vendor/faiss/faiss/utils/quantize_lut.cpp +78 -66
- data/vendor/faiss/faiss/utils/quantize_lut.h +22 -20
- data/vendor/faiss/faiss/utils/random.cpp +39 -63
- data/vendor/faiss/faiss/utils/random.h +13 -16
- data/vendor/faiss/faiss/utils/simdlib.h +4 -2
- data/vendor/faiss/faiss/utils/simdlib_avx2.h +88 -85
- data/vendor/faiss/faiss/utils/simdlib_emulated.h +226 -165
- data/vendor/faiss/faiss/utils/simdlib_neon.h +832 -0
- data/vendor/faiss/faiss/utils/utils.cpp +304 -287
- data/vendor/faiss/faiss/utils/utils.h +54 -49
- metadata +29 -4
@@ -7,15 +7,15 @@
|
|
7
7
|
|
8
8
|
namespace faiss {
|
9
9
|
|
10
|
+
extern const uint8_t hamdis_tab_ham_bytes[256];
|
10
11
|
|
11
|
-
inline BitstringWriter::BitstringWriter(uint8_t
|
12
|
-
|
13
|
-
|
14
|
-
memset (code, 0, code_size);
|
12
|
+
inline BitstringWriter::BitstringWriter(uint8_t* code, size_t code_size)
|
13
|
+
: code(code), code_size(code_size), i(0) {
|
14
|
+
memset(code, 0, code_size);
|
15
15
|
}
|
16
16
|
|
17
17
|
inline void BitstringWriter::write(uint64_t x, int nbit) {
|
18
|
-
assert
|
18
|
+
assert(code_size * 8 >= nbit + i);
|
19
19
|
// nb of available bits in i / 8
|
20
20
|
int na = 8 - (i & 7);
|
21
21
|
|
@@ -35,13 +35,11 @@ inline void BitstringWriter::write(uint64_t x, int nbit) {
|
|
35
35
|
}
|
36
36
|
}
|
37
37
|
|
38
|
-
|
39
|
-
|
40
|
-
code (code), code_size (code_size), i(0)
|
41
|
-
{}
|
38
|
+
inline BitstringReader::BitstringReader(const uint8_t* code, size_t code_size)
|
39
|
+
: code(code), code_size(code_size), i(0) {}
|
42
40
|
|
43
41
|
inline uint64_t BitstringReader::read(int nbit) {
|
44
|
-
assert
|
42
|
+
assert(code_size * 8 >= nbit + i);
|
45
43
|
// nb of available bits in i / 8
|
46
44
|
int na = 8 - (i & 7);
|
47
45
|
// get available bits in current byte
|
@@ -67,7 +65,6 @@ inline uint64_t BitstringReader::read(int nbit) {
|
|
67
65
|
}
|
68
66
|
}
|
69
67
|
|
70
|
-
|
71
68
|
/******************************************************************
|
72
69
|
* The HammingComputer series of classes compares a single code of
|
73
70
|
* size 4 to 32 to incoming codes. They are intended for use as a
|
@@ -76,68 +73,64 @@ inline uint64_t BitstringReader::read(int nbit) {
|
|
76
73
|
* hamming() functions and put the a0, a1, ... in registers.
|
77
74
|
******************************************************************/
|
78
75
|
|
79
|
-
|
80
76
|
struct HammingComputer4 {
|
81
77
|
uint32_t a0;
|
82
78
|
|
83
|
-
HammingComputer4
|
79
|
+
HammingComputer4() {}
|
84
80
|
|
85
|
-
HammingComputer4
|
86
|
-
set
|
81
|
+
HammingComputer4(const uint8_t* a, int code_size) {
|
82
|
+
set(a, code_size);
|
87
83
|
}
|
88
84
|
|
89
|
-
void set
|
90
|
-
assert
|
91
|
-
a0 = *(uint32_t
|
85
|
+
void set(const uint8_t* a, int code_size) {
|
86
|
+
assert(code_size == 4);
|
87
|
+
a0 = *(uint32_t*)a;
|
92
88
|
}
|
93
89
|
|
94
|
-
inline int hamming
|
95
|
-
return popcount64
|
90
|
+
inline int hamming(const uint8_t* b) const {
|
91
|
+
return popcount64(*(uint32_t*)b ^ a0);
|
96
92
|
}
|
97
|
-
|
98
93
|
};
|
99
94
|
|
100
95
|
struct HammingComputer8 {
|
101
96
|
uint64_t a0;
|
102
97
|
|
103
|
-
HammingComputer8
|
98
|
+
HammingComputer8() {}
|
104
99
|
|
105
|
-
HammingComputer8
|
106
|
-
set
|
100
|
+
HammingComputer8(const uint8_t* a, int code_size) {
|
101
|
+
set(a, code_size);
|
107
102
|
}
|
108
103
|
|
109
|
-
void set
|
110
|
-
assert
|
111
|
-
a0 = *(uint64_t
|
104
|
+
void set(const uint8_t* a, int code_size) {
|
105
|
+
assert(code_size == 8);
|
106
|
+
a0 = *(uint64_t*)a;
|
112
107
|
}
|
113
108
|
|
114
|
-
inline int hamming
|
115
|
-
return popcount64
|
109
|
+
inline int hamming(const uint8_t* b) const {
|
110
|
+
return popcount64(*(uint64_t*)b ^ a0);
|
116
111
|
}
|
117
|
-
|
118
112
|
};
|
119
113
|
|
120
|
-
|
121
114
|
struct HammingComputer16 {
|
122
115
|
uint64_t a0, a1;
|
123
116
|
|
124
|
-
HammingComputer16
|
117
|
+
HammingComputer16() {}
|
125
118
|
|
126
|
-
HammingComputer16
|
127
|
-
set
|
119
|
+
HammingComputer16(const uint8_t* a8, int code_size) {
|
120
|
+
set(a8, code_size);
|
128
121
|
}
|
129
122
|
|
130
|
-
void set
|
131
|
-
assert
|
132
|
-
const uint64_t
|
133
|
-
a0 = a[0];
|
123
|
+
void set(const uint8_t* a8, int code_size) {
|
124
|
+
assert(code_size == 16);
|
125
|
+
const uint64_t* a = (uint64_t*)a8;
|
126
|
+
a0 = a[0];
|
127
|
+
a1 = a[1];
|
134
128
|
}
|
135
129
|
|
136
|
-
inline int hamming
|
137
|
-
const uint64_t
|
138
|
-
return popcount64
|
130
|
+
inline int hamming(const uint8_t* b8) const {
|
131
|
+
const uint64_t* b = (uint64_t*)b8;
|
132
|
+
return popcount64(b[0] ^ a0) + popcount64(b[1] ^ a1);
|
139
133
|
}
|
140
|
-
|
141
134
|
};
|
142
135
|
|
143
136
|
// when applied to an array, 1/2 of the 64-bit accesses are unaligned.
|
@@ -146,150 +139,212 @@ struct HammingComputer20 {
|
|
146
139
|
uint64_t a0, a1;
|
147
140
|
uint32_t a2;
|
148
141
|
|
149
|
-
HammingComputer20
|
142
|
+
HammingComputer20() {}
|
150
143
|
|
151
|
-
HammingComputer20
|
152
|
-
set
|
144
|
+
HammingComputer20(const uint8_t* a8, int code_size) {
|
145
|
+
set(a8, code_size);
|
153
146
|
}
|
154
147
|
|
155
|
-
void set
|
156
|
-
assert
|
157
|
-
const uint64_t
|
158
|
-
a0 = a[0];
|
148
|
+
void set(const uint8_t* a8, int code_size) {
|
149
|
+
assert(code_size == 20);
|
150
|
+
const uint64_t* a = (uint64_t*)a8;
|
151
|
+
a0 = a[0];
|
152
|
+
a1 = a[1];
|
153
|
+
a2 = a[2];
|
159
154
|
}
|
160
155
|
|
161
|
-
inline int hamming
|
162
|
-
const uint64_t
|
163
|
-
return popcount64
|
164
|
-
|
156
|
+
inline int hamming(const uint8_t* b8) const {
|
157
|
+
const uint64_t* b = (uint64_t*)b8;
|
158
|
+
return popcount64(b[0] ^ a0) + popcount64(b[1] ^ a1) +
|
159
|
+
popcount64(*(uint32_t*)(b + 2) ^ a2);
|
165
160
|
}
|
166
161
|
};
|
167
162
|
|
168
163
|
struct HammingComputer32 {
|
169
164
|
uint64_t a0, a1, a2, a3;
|
170
165
|
|
171
|
-
HammingComputer32
|
166
|
+
HammingComputer32() {}
|
172
167
|
|
173
|
-
HammingComputer32
|
174
|
-
set
|
168
|
+
HammingComputer32(const uint8_t* a8, int code_size) {
|
169
|
+
set(a8, code_size);
|
175
170
|
}
|
176
171
|
|
177
|
-
void set
|
178
|
-
assert
|
179
|
-
const uint64_t
|
180
|
-
a0 = a[0];
|
172
|
+
void set(const uint8_t* a8, int code_size) {
|
173
|
+
assert(code_size == 32);
|
174
|
+
const uint64_t* a = (uint64_t*)a8;
|
175
|
+
a0 = a[0];
|
176
|
+
a1 = a[1];
|
177
|
+
a2 = a[2];
|
178
|
+
a3 = a[3];
|
181
179
|
}
|
182
180
|
|
183
|
-
inline int hamming
|
184
|
-
const uint64_t
|
185
|
-
return popcount64
|
186
|
-
|
181
|
+
inline int hamming(const uint8_t* b8) const {
|
182
|
+
const uint64_t* b = (uint64_t*)b8;
|
183
|
+
return popcount64(b[0] ^ a0) + popcount64(b[1] ^ a1) +
|
184
|
+
popcount64(b[2] ^ a2) + popcount64(b[3] ^ a3);
|
187
185
|
}
|
188
|
-
|
189
186
|
};
|
190
187
|
|
191
188
|
struct HammingComputer64 {
|
192
189
|
uint64_t a0, a1, a2, a3, a4, a5, a6, a7;
|
193
190
|
|
194
|
-
HammingComputer64
|
191
|
+
HammingComputer64() {}
|
195
192
|
|
196
|
-
HammingComputer64
|
197
|
-
set
|
193
|
+
HammingComputer64(const uint8_t* a8, int code_size) {
|
194
|
+
set(a8, code_size);
|
198
195
|
}
|
199
196
|
|
200
|
-
void set
|
201
|
-
assert
|
202
|
-
const uint64_t
|
203
|
-
a0 = a[0];
|
204
|
-
|
197
|
+
void set(const uint8_t* a8, int code_size) {
|
198
|
+
assert(code_size == 64);
|
199
|
+
const uint64_t* a = (uint64_t*)a8;
|
200
|
+
a0 = a[0];
|
201
|
+
a1 = a[1];
|
202
|
+
a2 = a[2];
|
203
|
+
a3 = a[3];
|
204
|
+
a4 = a[4];
|
205
|
+
a5 = a[5];
|
206
|
+
a6 = a[6];
|
207
|
+
a7 = a[7];
|
205
208
|
}
|
206
209
|
|
207
|
-
inline int hamming
|
208
|
-
const uint64_t
|
209
|
-
return popcount64
|
210
|
-
|
211
|
-
|
212
|
-
|
210
|
+
inline int hamming(const uint8_t* b8) const {
|
211
|
+
const uint64_t* b = (uint64_t*)b8;
|
212
|
+
return popcount64(b[0] ^ a0) + popcount64(b[1] ^ a1) +
|
213
|
+
popcount64(b[2] ^ a2) + popcount64(b[3] ^ a3) +
|
214
|
+
popcount64(b[4] ^ a4) + popcount64(b[5] ^ a5) +
|
215
|
+
popcount64(b[6] ^ a6) + popcount64(b[7] ^ a7);
|
213
216
|
}
|
214
|
-
|
215
217
|
};
|
216
218
|
|
217
|
-
// very inefficient...
|
218
219
|
struct HammingComputerDefault {
|
219
|
-
const uint8_t
|
220
|
-
int
|
220
|
+
const uint8_t* a8;
|
221
|
+
int quotient8;
|
222
|
+
int remainder8;
|
221
223
|
|
222
|
-
HammingComputerDefault
|
224
|
+
HammingComputerDefault() {}
|
223
225
|
|
224
|
-
HammingComputerDefault
|
225
|
-
set
|
226
|
+
HammingComputerDefault(const uint8_t* a8, int code_size) {
|
227
|
+
set(a8, code_size);
|
226
228
|
}
|
227
229
|
|
228
|
-
void set
|
229
|
-
|
230
|
-
|
230
|
+
void set(const uint8_t* a8, int code_size) {
|
231
|
+
this->a8 = a8;
|
232
|
+
quotient8 = code_size / 8;
|
233
|
+
remainder8 = code_size % 8;
|
231
234
|
}
|
232
235
|
|
233
|
-
int hamming
|
236
|
+
int hamming(const uint8_t* b8) const {
|
234
237
|
int accu = 0;
|
235
|
-
|
236
|
-
|
238
|
+
|
239
|
+
const uint64_t* a64 = reinterpret_cast<const uint64_t*>(a8);
|
240
|
+
const uint64_t* b64 = reinterpret_cast<const uint64_t*>(b8);
|
241
|
+
int i = 0, len = quotient8;
|
242
|
+
switch (len & 7) {
|
243
|
+
default:
|
244
|
+
while (len > 7) {
|
245
|
+
len -= 8;
|
246
|
+
accu += popcount64(a64[i] ^ b64[i]);
|
247
|
+
i++;
|
248
|
+
case 7:
|
249
|
+
accu += popcount64(a64[i] ^ b64[i]);
|
250
|
+
i++;
|
251
|
+
case 6:
|
252
|
+
accu += popcount64(a64[i] ^ b64[i]);
|
253
|
+
i++;
|
254
|
+
case 5:
|
255
|
+
accu += popcount64(a64[i] ^ b64[i]);
|
256
|
+
i++;
|
257
|
+
case 4:
|
258
|
+
accu += popcount64(a64[i] ^ b64[i]);
|
259
|
+
i++;
|
260
|
+
case 3:
|
261
|
+
accu += popcount64(a64[i] ^ b64[i]);
|
262
|
+
i++;
|
263
|
+
case 2:
|
264
|
+
accu += popcount64(a64[i] ^ b64[i]);
|
265
|
+
i++;
|
266
|
+
case 1:
|
267
|
+
accu += popcount64(a64[i] ^ b64[i]);
|
268
|
+
i++;
|
269
|
+
}
|
270
|
+
}
|
271
|
+
if (remainder8) {
|
272
|
+
const uint8_t* a = a8 + 8 * quotient8;
|
273
|
+
const uint8_t* b = b8 + 8 * quotient8;
|
274
|
+
switch (remainder8) {
|
275
|
+
case 7:
|
276
|
+
accu += hamdis_tab_ham_bytes[a[6] ^ b[6]];
|
277
|
+
case 6:
|
278
|
+
accu += hamdis_tab_ham_bytes[a[5] ^ b[5]];
|
279
|
+
case 5:
|
280
|
+
accu += hamdis_tab_ham_bytes[a[4] ^ b[4]];
|
281
|
+
case 4:
|
282
|
+
accu += hamdis_tab_ham_bytes[a[3] ^ b[3]];
|
283
|
+
case 3:
|
284
|
+
accu += hamdis_tab_ham_bytes[a[2] ^ b[2]];
|
285
|
+
case 2:
|
286
|
+
accu += hamdis_tab_ham_bytes[a[1] ^ b[1]];
|
287
|
+
case 1:
|
288
|
+
accu += hamdis_tab_ham_bytes[a[0] ^ b[0]];
|
289
|
+
default:
|
290
|
+
break;
|
291
|
+
}
|
292
|
+
}
|
293
|
+
|
237
294
|
return accu;
|
238
295
|
}
|
239
|
-
|
240
296
|
};
|
241
297
|
|
298
|
+
// more inefficient than HammingComputerDefault (obsolete)
|
242
299
|
struct HammingComputerM8 {
|
243
|
-
const uint64_t
|
300
|
+
const uint64_t* a;
|
244
301
|
int n;
|
245
302
|
|
246
|
-
HammingComputerM8
|
303
|
+
HammingComputerM8() {}
|
247
304
|
|
248
|
-
HammingComputerM8
|
249
|
-
set
|
305
|
+
HammingComputerM8(const uint8_t* a8, int code_size) {
|
306
|
+
set(a8, code_size);
|
250
307
|
}
|
251
308
|
|
252
|
-
void set
|
253
|
-
assert
|
254
|
-
a =
|
309
|
+
void set(const uint8_t* a8, int code_size) {
|
310
|
+
assert(code_size % 8 == 0);
|
311
|
+
a = (uint64_t*)a8;
|
255
312
|
n = code_size / 8;
|
256
313
|
}
|
257
314
|
|
258
|
-
int hamming
|
259
|
-
const uint64_t
|
315
|
+
int hamming(const uint8_t* b8) const {
|
316
|
+
const uint64_t* b = (uint64_t*)b8;
|
260
317
|
int accu = 0;
|
261
318
|
for (int i = 0; i < n; i++)
|
262
|
-
accu += popcount64
|
319
|
+
accu += popcount64(a[i] ^ b[i]);
|
263
320
|
return accu;
|
264
321
|
}
|
265
|
-
|
266
322
|
};
|
267
323
|
|
268
|
-
//
|
324
|
+
// more inefficient than HammingComputerDefault (obsolete)
|
269
325
|
struct HammingComputerM4 {
|
270
|
-
const uint32_t
|
326
|
+
const uint32_t* a;
|
271
327
|
int n;
|
272
328
|
|
273
|
-
HammingComputerM4
|
329
|
+
HammingComputerM4() {}
|
274
330
|
|
275
|
-
HammingComputerM4
|
276
|
-
set
|
331
|
+
HammingComputerM4(const uint8_t* a4, int code_size) {
|
332
|
+
set(a4, code_size);
|
277
333
|
}
|
278
334
|
|
279
|
-
void set
|
280
|
-
assert
|
281
|
-
a =
|
335
|
+
void set(const uint8_t* a4, int code_size) {
|
336
|
+
assert(code_size % 4 == 0);
|
337
|
+
a = (uint32_t*)a4;
|
282
338
|
n = code_size / 4;
|
283
339
|
}
|
284
340
|
|
285
|
-
int hamming
|
286
|
-
const uint32_t
|
341
|
+
int hamming(const uint8_t* b8) const {
|
342
|
+
const uint32_t* b = (uint32_t*)b8;
|
287
343
|
int accu = 0;
|
288
344
|
for (int i = 0; i < n; i++)
|
289
|
-
|
345
|
+
accu += popcount64(a[i] ^ b[i]);
|
290
346
|
return accu;
|
291
347
|
}
|
292
|
-
|
293
348
|
};
|
294
349
|
|
295
350
|
/***************************************************************************
|
@@ -297,17 +352,17 @@ struct HammingComputerM4 {
|
|
297
352
|
**************************************************************************/
|
298
353
|
|
299
354
|
// default template
|
300
|
-
template<int CODE_SIZE>
|
301
|
-
struct HammingComputer:
|
302
|
-
HammingComputer
|
303
|
-
|
355
|
+
template <int CODE_SIZE>
|
356
|
+
struct HammingComputer : HammingComputerDefault {
|
357
|
+
HammingComputer(const uint8_t* a, int code_size)
|
358
|
+
: HammingComputerDefault(a, code_size) {}
|
304
359
|
};
|
305
360
|
|
306
|
-
#define SPECIALIZED_HC(CODE_SIZE)
|
307
|
-
template<>
|
308
|
-
|
309
|
-
HammingComputer
|
310
|
-
|
361
|
+
#define SPECIALIZED_HC(CODE_SIZE) \
|
362
|
+
template <> \
|
363
|
+
struct HammingComputer<CODE_SIZE> : HammingComputer##CODE_SIZE { \
|
364
|
+
HammingComputer(const uint8_t* a) \
|
365
|
+
: HammingComputer##CODE_SIZE(a, CODE_SIZE) {} \
|
311
366
|
}
|
312
367
|
|
313
368
|
SPECIALIZED_HC(4);
|
@@ -319,105 +374,98 @@ SPECIALIZED_HC(64);
|
|
319
374
|
|
320
375
|
#undef SPECIALIZED_HC
|
321
376
|
|
322
|
-
|
323
377
|
/***************************************************************************
|
324
378
|
* generalized Hamming = number of bytes that are different between
|
325
379
|
* two codes.
|
326
380
|
***************************************************************************/
|
327
381
|
|
328
|
-
|
329
|
-
inline int generalized_hamming_64 (uint64_t a) {
|
382
|
+
inline int generalized_hamming_64(uint64_t a) {
|
330
383
|
a |= a >> 1;
|
331
384
|
a |= a >> 2;
|
332
385
|
a |= a >> 4;
|
333
386
|
a &= 0x0101010101010101UL;
|
334
|
-
return popcount64
|
387
|
+
return popcount64(a);
|
335
388
|
}
|
336
389
|
|
337
|
-
|
338
390
|
struct GenHammingComputer8 {
|
339
391
|
uint64_t a0;
|
340
392
|
|
341
|
-
GenHammingComputer8
|
342
|
-
assert
|
343
|
-
a0 = *(uint64_t
|
393
|
+
GenHammingComputer8(const uint8_t* a, int code_size) {
|
394
|
+
assert(code_size == 8);
|
395
|
+
a0 = *(uint64_t*)a;
|
344
396
|
}
|
345
397
|
|
346
|
-
inline int hamming
|
347
|
-
return generalized_hamming_64
|
398
|
+
inline int hamming(const uint8_t* b) const {
|
399
|
+
return generalized_hamming_64(*(uint64_t*)b ^ a0);
|
348
400
|
}
|
349
|
-
|
350
401
|
};
|
351
402
|
|
352
|
-
|
353
403
|
struct GenHammingComputer16 {
|
354
404
|
uint64_t a0, a1;
|
355
|
-
GenHammingComputer16
|
356
|
-
assert
|
357
|
-
const uint64_t
|
358
|
-
a0 = a[0];
|
405
|
+
GenHammingComputer16(const uint8_t* a8, int code_size) {
|
406
|
+
assert(code_size == 16);
|
407
|
+
const uint64_t* a = (uint64_t*)a8;
|
408
|
+
a0 = a[0];
|
409
|
+
a1 = a[1];
|
359
410
|
}
|
360
411
|
|
361
|
-
inline int hamming
|
362
|
-
const uint64_t
|
363
|
-
return generalized_hamming_64
|
364
|
-
|
412
|
+
inline int hamming(const uint8_t* b8) const {
|
413
|
+
const uint64_t* b = (uint64_t*)b8;
|
414
|
+
return generalized_hamming_64(b[0] ^ a0) +
|
415
|
+
generalized_hamming_64(b[1] ^ a1);
|
365
416
|
}
|
366
|
-
|
367
417
|
};
|
368
418
|
|
369
419
|
struct GenHammingComputer32 {
|
370
420
|
uint64_t a0, a1, a2, a3;
|
371
421
|
|
372
|
-
GenHammingComputer32
|
373
|
-
assert
|
374
|
-
const uint64_t
|
375
|
-
a0 = a[0];
|
422
|
+
GenHammingComputer32(const uint8_t* a8, int code_size) {
|
423
|
+
assert(code_size == 32);
|
424
|
+
const uint64_t* a = (uint64_t*)a8;
|
425
|
+
a0 = a[0];
|
426
|
+
a1 = a[1];
|
427
|
+
a2 = a[2];
|
428
|
+
a3 = a[3];
|
376
429
|
}
|
377
430
|
|
378
|
-
inline int hamming
|
379
|
-
const uint64_t
|
380
|
-
return generalized_hamming_64
|
381
|
-
|
382
|
-
|
383
|
-
|
431
|
+
inline int hamming(const uint8_t* b8) const {
|
432
|
+
const uint64_t* b = (uint64_t*)b8;
|
433
|
+
return generalized_hamming_64(b[0] ^ a0) +
|
434
|
+
generalized_hamming_64(b[1] ^ a1) +
|
435
|
+
generalized_hamming_64(b[2] ^ a2) +
|
436
|
+
generalized_hamming_64(b[3] ^ a3);
|
384
437
|
}
|
385
|
-
|
386
438
|
};
|
387
439
|
|
388
440
|
struct GenHammingComputerM8 {
|
389
|
-
const uint64_t
|
441
|
+
const uint64_t* a;
|
390
442
|
int n;
|
391
443
|
|
392
|
-
GenHammingComputerM8
|
393
|
-
assert
|
394
|
-
a =
|
444
|
+
GenHammingComputerM8(const uint8_t* a8, int code_size) {
|
445
|
+
assert(code_size % 8 == 0);
|
446
|
+
a = (uint64_t*)a8;
|
395
447
|
n = code_size / 8;
|
396
448
|
}
|
397
449
|
|
398
|
-
int hamming
|
399
|
-
const uint64_t
|
450
|
+
int hamming(const uint8_t* b8) const {
|
451
|
+
const uint64_t* b = (uint64_t*)b8;
|
400
452
|
int accu = 0;
|
401
453
|
for (int i = 0; i < n; i++)
|
402
|
-
accu += generalized_hamming_64
|
454
|
+
accu += generalized_hamming_64(a[i] ^ b[i]);
|
403
455
|
return accu;
|
404
456
|
}
|
405
|
-
|
406
457
|
};
|
407
458
|
|
408
|
-
|
409
459
|
/** generalized Hamming distances (= count number of code bytes that
|
410
460
|
are the same) */
|
411
|
-
void generalized_hammings_knn_hc
|
412
|
-
int_maxheap_array_t
|
413
|
-
const uint8_t
|
414
|
-
const uint8_t
|
461
|
+
void generalized_hammings_knn_hc(
|
462
|
+
int_maxheap_array_t* ha,
|
463
|
+
const uint8_t* a,
|
464
|
+
const uint8_t* b,
|
415
465
|
size_t nb,
|
416
466
|
size_t code_size,
|
417
467
|
int ordered = true);
|
418
468
|
|
419
|
-
|
420
|
-
|
421
469
|
/** This class maintains a list of best distances seen so far.
|
422
470
|
*
|
423
471
|
* Since the distances are in a limited range (0 to nbit), the
|
@@ -425,46 +473,49 @@ void generalized_hammings_knn_hc (
|
|
425
473
|
* in only the n-first lists, such that the sum of sizes of the
|
426
474
|
* n lists is below k.
|
427
475
|
*/
|
428
|
-
template<class HammingComputer>
|
476
|
+
template <class HammingComputer>
|
429
477
|
struct HCounterState {
|
430
|
-
|
431
|
-
|
432
|
-
|
433
|
-
|
434
|
-
|
435
|
-
|
436
|
-
|
437
|
-
|
438
|
-
|
439
|
-
|
440
|
-
|
441
|
-
|
442
|
-
|
443
|
-
|
444
|
-
|
445
|
-
|
446
|
-
|
447
|
-
|
448
|
-
|
449
|
-
|
450
|
-
|
451
|
-
|
452
|
-
|
453
|
-
|
454
|
-
|
455
|
-
|
456
|
-
|
457
|
-
|
458
|
-
|
459
|
-
|
478
|
+
int* counters;
|
479
|
+
int64_t* ids_per_dis;
|
480
|
+
|
481
|
+
HammingComputer hc;
|
482
|
+
int thres;
|
483
|
+
int count_lt;
|
484
|
+
int count_eq;
|
485
|
+
int k;
|
486
|
+
|
487
|
+
HCounterState(
|
488
|
+
int* counters,
|
489
|
+
int64_t* ids_per_dis,
|
490
|
+
const uint8_t* x,
|
491
|
+
int d,
|
492
|
+
int k)
|
493
|
+
: counters(counters),
|
494
|
+
ids_per_dis(ids_per_dis),
|
495
|
+
hc(x, d / 8),
|
496
|
+
thres(d + 1),
|
497
|
+
count_lt(0),
|
498
|
+
count_eq(0),
|
499
|
+
k(k) {}
|
500
|
+
|
501
|
+
void update_counter(const uint8_t* y, size_t j) {
|
502
|
+
int32_t dis = hc.hamming(y);
|
503
|
+
|
504
|
+
if (dis <= thres) {
|
505
|
+
if (dis < thres) {
|
506
|
+
ids_per_dis[dis * k + counters[dis]++] = j;
|
507
|
+
++count_lt;
|
508
|
+
while (count_lt == k && thres > 0) {
|
509
|
+
--thres;
|
510
|
+
count_eq = counters[thres];
|
511
|
+
count_lt -= count_eq;
|
512
|
+
}
|
513
|
+
} else if (count_eq < k) {
|
514
|
+
ids_per_dis[dis * k + count_eq++] = j;
|
515
|
+
counters[dis] = count_eq;
|
516
|
+
}
|
460
517
|
}
|
461
|
-
} else if (count_eq < k) {
|
462
|
-
ids_per_dis[dis * k + count_eq++] = j;
|
463
|
-
counters[dis] = count_eq;
|
464
|
-
}
|
465
518
|
}
|
466
|
-
}
|
467
519
|
};
|
468
520
|
|
469
|
-
|
470
521
|
} // namespace faiss
|