faiss 0.1.7 → 0.2.3
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +18 -0
- data/README.md +7 -7
- data/ext/faiss/ext.cpp +1 -1
- data/ext/faiss/extconf.rb +8 -2
- data/ext/faiss/index.cpp +102 -69
- data/ext/faiss/index_binary.cpp +24 -30
- data/ext/faiss/kmeans.cpp +20 -16
- data/ext/faiss/numo.hpp +867 -0
- data/ext/faiss/pca_matrix.cpp +13 -14
- data/ext/faiss/product_quantizer.cpp +23 -24
- data/ext/faiss/utils.cpp +10 -37
- data/ext/faiss/utils.h +2 -13
- data/lib/faiss/version.rb +1 -1
- data/lib/faiss.rb +0 -5
- data/vendor/faiss/faiss/AutoTune.cpp +292 -291
- data/vendor/faiss/faiss/AutoTune.h +55 -56
- data/vendor/faiss/faiss/Clustering.cpp +334 -195
- data/vendor/faiss/faiss/Clustering.h +88 -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 +139 -164
- data/vendor/faiss/faiss/Index2Layer.h +22 -22
- 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 +116 -147
- data/vendor/faiss/faiss/IndexFlat.h +35 -46
- data/vendor/faiss/faiss/IndexHNSW.cpp +372 -348
- data/vendor/faiss/faiss/IndexHNSW.h +57 -41
- data/vendor/faiss/faiss/IndexIVF.cpp +474 -454
- data/vendor/faiss/faiss/IndexIVF.h +146 -113
- data/vendor/faiss/faiss/IndexIVFFlat.cpp +248 -250
- data/vendor/faiss/faiss/IndexIVFFlat.h +48 -51
- data/vendor/faiss/faiss/IndexIVFPQ.cpp +457 -516
- data/vendor/faiss/faiss/IndexIVFPQ.h +74 -66
- 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 +125 -133
- data/vendor/faiss/faiss/IndexIVFSpectralHash.h +19 -21
- data/vendor/faiss/faiss/IndexLSH.cpp +75 -96
- data/vendor/faiss/faiss/IndexLSH.h +21 -26
- data/vendor/faiss/faiss/IndexLattice.cpp +42 -56
- data/vendor/faiss/faiss/IndexLattice.h +11 -16
- data/vendor/faiss/faiss/IndexNNDescent.cpp +231 -0
- data/vendor/faiss/faiss/IndexNNDescent.h +72 -0
- data/vendor/faiss/faiss/IndexNSG.cpp +303 -0
- data/vendor/faiss/faiss/IndexNSG.h +85 -0
- data/vendor/faiss/faiss/IndexPQ.cpp +405 -464
- data/vendor/faiss/faiss/IndexPQ.h +64 -67
- 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 +115 -131
- data/vendor/faiss/faiss/IndexRefine.h +22 -23
- data/vendor/faiss/faiss/IndexReplicas.cpp +147 -153
- data/vendor/faiss/faiss/IndexReplicas.h +62 -56
- data/vendor/faiss/faiss/IndexResidual.cpp +291 -0
- data/vendor/faiss/faiss/IndexResidual.h +152 -0
- data/vendor/faiss/faiss/IndexScalarQuantizer.cpp +120 -155
- data/vendor/faiss/faiss/IndexScalarQuantizer.h +41 -45
- 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 +652 -474
- data/vendor/faiss/faiss/VectorTransform.h +61 -89
- data/vendor/faiss/faiss/clone_index.cpp +77 -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 +197 -170
- 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/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 +270 -0
- data/vendor/faiss/faiss/impl/AdditiveQuantizer.h +115 -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 +595 -611
- data/vendor/faiss/faiss/impl/HNSW.h +179 -200
- data/vendor/faiss/faiss/impl/LocalSearchQuantizer.cpp +672 -0
- data/vendor/faiss/faiss/impl/LocalSearchQuantizer.h +172 -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 +682 -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 +448 -0
- data/vendor/faiss/faiss/impl/ResidualQuantizer.h +130 -0
- data/vendor/faiss/faiss/impl/ResultHandler.h +96 -132
- data/vendor/faiss/faiss/impl/ScalarQuantizer.cpp +648 -701
- 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 +547 -479
- data/vendor/faiss/faiss/impl/index_write.cpp +497 -407
- data/vendor/faiss/faiss/impl/io.cpp +75 -94
- data/vendor/faiss/faiss/impl/io.h +31 -41
- data/vendor/faiss/faiss/impl/io_macros.h +40 -29
- 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 +269 -218
- data/vendor/faiss/faiss/index_factory.h +6 -7
- 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 +301 -310
- data/vendor/faiss/faiss/utils/distances.h +133 -118
- data/vendor/faiss/faiss/utils/distances_simd.cpp +456 -516
- 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 +53 -48
- metadata +26 -12
- data/lib/faiss/index.rb +0 -20
- data/lib/faiss/index_binary.rb +0 -20
- data/lib/faiss/kmeans.rb +0 -15
- data/lib/faiss/pca_matrix.rb +0 -15
- data/lib/faiss/product_quantizer.rb +0 -22
@@ -5,55 +5,51 @@
|
|
5
5
|
* LICENSE file in the root directory of this source tree.
|
6
6
|
*/
|
7
7
|
|
8
|
-
|
9
|
-
#include <faiss/gpu/utils/Timer.h>
|
10
8
|
#include <faiss/gpu/utils/DeviceUtils.h>
|
9
|
+
#include <faiss/gpu/utils/Timer.h>
|
11
10
|
#include <faiss/impl/FaissAssert.h>
|
12
11
|
#include <chrono>
|
13
12
|
|
14
|
-
namespace faiss {
|
13
|
+
namespace faiss {
|
14
|
+
namespace gpu {
|
15
15
|
|
16
16
|
KernelTimer::KernelTimer(cudaStream_t stream)
|
17
|
-
|
18
|
-
|
19
|
-
|
20
|
-
|
21
|
-
|
22
|
-
CUDA_VERIFY(cudaEventCreate(&stopEvent_));
|
23
|
-
|
24
|
-
CUDA_VERIFY(cudaEventRecord(startEvent_, stream_));
|
17
|
+
: startEvent_(0), stopEvent_(0), stream_(stream), valid_(true) {
|
18
|
+
CUDA_VERIFY(cudaEventCreate(&startEvent_));
|
19
|
+
CUDA_VERIFY(cudaEventCreate(&stopEvent_));
|
20
|
+
|
21
|
+
CUDA_VERIFY(cudaEventRecord(startEvent_, stream_));
|
25
22
|
}
|
26
23
|
|
27
24
|
KernelTimer::~KernelTimer() {
|
28
|
-
|
29
|
-
|
25
|
+
CUDA_VERIFY(cudaEventDestroy(startEvent_));
|
26
|
+
CUDA_VERIFY(cudaEventDestroy(stopEvent_));
|
30
27
|
}
|
31
28
|
|
32
|
-
float
|
33
|
-
|
34
|
-
FAISS_ASSERT(valid_);
|
29
|
+
float KernelTimer::elapsedMilliseconds() {
|
30
|
+
FAISS_ASSERT(valid_);
|
35
31
|
|
36
|
-
|
37
|
-
|
32
|
+
CUDA_VERIFY(cudaEventRecord(stopEvent_, stream_));
|
33
|
+
CUDA_VERIFY(cudaEventSynchronize(stopEvent_));
|
38
34
|
|
39
|
-
|
40
|
-
|
41
|
-
|
35
|
+
auto time = 0.0f;
|
36
|
+
CUDA_VERIFY(cudaEventElapsedTime(&time, startEvent_, stopEvent_));
|
37
|
+
valid_ = false;
|
42
38
|
|
43
|
-
|
39
|
+
return time;
|
44
40
|
}
|
45
41
|
|
46
42
|
CpuTimer::CpuTimer() {
|
47
|
-
|
43
|
+
start_ = std::chrono::steady_clock::now();
|
48
44
|
}
|
49
45
|
|
50
|
-
float
|
51
|
-
|
52
|
-
auto end = std::chrono::steady_clock::now();
|
46
|
+
float CpuTimer::elapsedMilliseconds() {
|
47
|
+
auto end = std::chrono::steady_clock::now();
|
53
48
|
|
54
|
-
|
49
|
+
std::chrono::duration<float, std::milli> duration = end - start_;
|
55
50
|
|
56
|
-
|
51
|
+
return duration.count();
|
57
52
|
}
|
58
53
|
|
59
|
-
}
|
54
|
+
} // namespace gpu
|
55
|
+
} // namespace faiss
|
@@ -5,48 +5,49 @@
|
|
5
5
|
* LICENSE file in the root directory of this source tree.
|
6
6
|
*/
|
7
7
|
|
8
|
-
|
9
8
|
#pragma once
|
10
9
|
|
11
10
|
#include <cuda_runtime.h>
|
12
11
|
#include <chrono>
|
13
12
|
|
14
|
-
namespace faiss {
|
13
|
+
namespace faiss {
|
14
|
+
namespace gpu {
|
15
15
|
|
16
16
|
/// Utility class for timing execution of a kernel
|
17
17
|
class KernelTimer {
|
18
|
-
|
19
|
-
|
20
|
-
|
21
|
-
|
22
|
-
|
23
|
-
|
24
|
-
|
25
|
-
|
26
|
-
|
27
|
-
|
28
|
-
|
29
|
-
|
30
|
-
|
31
|
-
|
32
|
-
|
33
|
-
|
34
|
-
|
35
|
-
|
36
|
-
|
18
|
+
public:
|
19
|
+
/// Constructor starts the timer and adds an event into the current
|
20
|
+
/// device stream
|
21
|
+
KernelTimer(cudaStream_t stream = 0);
|
22
|
+
|
23
|
+
/// Destructor releases event resources
|
24
|
+
~KernelTimer();
|
25
|
+
|
26
|
+
/// Adds a stop event then synchronizes on the stop event to get the
|
27
|
+
/// actual GPU-side kernel timings for any kernels launched in the
|
28
|
+
/// current stream. Returns the number of milliseconds elapsed.
|
29
|
+
/// Can only be called once.
|
30
|
+
float elapsedMilliseconds();
|
31
|
+
|
32
|
+
private:
|
33
|
+
cudaEvent_t startEvent_;
|
34
|
+
cudaEvent_t stopEvent_;
|
35
|
+
cudaStream_t stream_;
|
36
|
+
bool valid_;
|
37
37
|
};
|
38
38
|
|
39
39
|
/// CPU wallclock elapsed timer
|
40
40
|
class CpuTimer {
|
41
|
-
|
42
|
-
|
43
|
-
|
41
|
+
public:
|
42
|
+
/// Creates and starts a new timer
|
43
|
+
CpuTimer();
|
44
44
|
|
45
|
-
|
46
|
-
|
45
|
+
/// Returns elapsed time in milliseconds
|
46
|
+
float elapsedMilliseconds();
|
47
47
|
|
48
|
-
|
49
|
-
|
48
|
+
private:
|
49
|
+
std::chrono::time_point<std::chrono::steady_clock> start_;
|
50
50
|
};
|
51
51
|
|
52
|
-
}
|
52
|
+
} // namespace gpu
|
53
|
+
} // namespace faiss
|
@@ -0,0 +1,270 @@
|
|
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/impl/AdditiveQuantizer.h>
|
11
|
+
#include <faiss/impl/FaissAssert.h>
|
12
|
+
|
13
|
+
#include <cstddef>
|
14
|
+
#include <cstdio>
|
15
|
+
#include <cstring>
|
16
|
+
#include <memory>
|
17
|
+
#include <random>
|
18
|
+
|
19
|
+
#include <algorithm>
|
20
|
+
|
21
|
+
#include <faiss/utils/Heap.h>
|
22
|
+
#include <faiss/utils/distances.h>
|
23
|
+
#include <faiss/utils/hamming.h> // BitstringWriter
|
24
|
+
#include <faiss/utils/utils.h>
|
25
|
+
|
26
|
+
extern "C" {
|
27
|
+
|
28
|
+
// general matrix multiplication
|
29
|
+
int sgemm_(
|
30
|
+
const char* transa,
|
31
|
+
const char* transb,
|
32
|
+
FINTEGER* m,
|
33
|
+
FINTEGER* n,
|
34
|
+
FINTEGER* k,
|
35
|
+
const float* alpha,
|
36
|
+
const float* a,
|
37
|
+
FINTEGER* lda,
|
38
|
+
const float* b,
|
39
|
+
FINTEGER* ldb,
|
40
|
+
float* beta,
|
41
|
+
float* c,
|
42
|
+
FINTEGER* ldc);
|
43
|
+
}
|
44
|
+
|
45
|
+
namespace {
|
46
|
+
|
47
|
+
// c and a and b can overlap
|
48
|
+
void fvec_add(size_t d, const float* a, const float* b, float* c) {
|
49
|
+
for (size_t i = 0; i < d; i++) {
|
50
|
+
c[i] = a[i] + b[i];
|
51
|
+
}
|
52
|
+
}
|
53
|
+
|
54
|
+
void fvec_add(size_t d, const float* a, float b, float* c) {
|
55
|
+
for (size_t i = 0; i < d; i++) {
|
56
|
+
c[i] = a[i] + b;
|
57
|
+
}
|
58
|
+
}
|
59
|
+
|
60
|
+
} // namespace
|
61
|
+
|
62
|
+
namespace faiss {
|
63
|
+
|
64
|
+
void AdditiveQuantizer::set_derived_values() {
|
65
|
+
tot_bits = 0;
|
66
|
+
is_byte_aligned = true;
|
67
|
+
codebook_offsets.resize(M + 1, 0);
|
68
|
+
for (int i = 0; i < M; i++) {
|
69
|
+
int nbit = nbits[i];
|
70
|
+
size_t k = 1 << nbit;
|
71
|
+
codebook_offsets[i + 1] = codebook_offsets[i] + k;
|
72
|
+
tot_bits += nbit;
|
73
|
+
if (nbit % 8 != 0) {
|
74
|
+
is_byte_aligned = false;
|
75
|
+
}
|
76
|
+
}
|
77
|
+
total_codebook_size = codebook_offsets[M];
|
78
|
+
// convert bits to bytes
|
79
|
+
code_size = (tot_bits + 7) / 8;
|
80
|
+
}
|
81
|
+
|
82
|
+
void AdditiveQuantizer::pack_codes(
|
83
|
+
size_t n,
|
84
|
+
const int32_t* codes,
|
85
|
+
uint8_t* packed_codes,
|
86
|
+
int64_t ld_codes) const {
|
87
|
+
if (ld_codes == -1) {
|
88
|
+
ld_codes = M;
|
89
|
+
}
|
90
|
+
#pragma omp parallel for if (n > 1000)
|
91
|
+
for (int64_t i = 0; i < n; i++) {
|
92
|
+
const int32_t* codes1 = codes + i * ld_codes;
|
93
|
+
BitstringWriter bsw(packed_codes + i * code_size, code_size);
|
94
|
+
for (int m = 0; m < M; m++) {
|
95
|
+
bsw.write(codes1[m], nbits[m]);
|
96
|
+
}
|
97
|
+
}
|
98
|
+
}
|
99
|
+
|
100
|
+
void AdditiveQuantizer::decode(const uint8_t* code, float* x, size_t n) const {
|
101
|
+
FAISS_THROW_IF_NOT_MSG(
|
102
|
+
is_trained, "The additive quantizer is not trained yet.");
|
103
|
+
|
104
|
+
// standard additive quantizer decoding
|
105
|
+
#pragma omp parallel for if (n > 1000)
|
106
|
+
for (int64_t i = 0; i < n; i++) {
|
107
|
+
BitstringReader bsr(code + i * code_size, code_size);
|
108
|
+
float* xi = x + i * d;
|
109
|
+
for (int m = 0; m < M; m++) {
|
110
|
+
int idx = bsr.read(nbits[m]);
|
111
|
+
const float* c = codebooks.data() + d * (codebook_offsets[m] + idx);
|
112
|
+
if (m == 0) {
|
113
|
+
memcpy(xi, c, sizeof(*x) * d);
|
114
|
+
} else {
|
115
|
+
fvec_add(d, xi, c, xi);
|
116
|
+
}
|
117
|
+
}
|
118
|
+
}
|
119
|
+
}
|
120
|
+
|
121
|
+
AdditiveQuantizer::~AdditiveQuantizer() {}
|
122
|
+
|
123
|
+
/****************************************************************************
|
124
|
+
* Support for fast distance computations and search with additive quantizer
|
125
|
+
****************************************************************************/
|
126
|
+
|
127
|
+
void AdditiveQuantizer::compute_centroid_norms(float* norms) const {
|
128
|
+
size_t ntotal = (size_t)1 << tot_bits;
|
129
|
+
// TODO: make tree of partial sums
|
130
|
+
#pragma omp parallel
|
131
|
+
{
|
132
|
+
std::vector<float> tmp(d);
|
133
|
+
#pragma omp for
|
134
|
+
for (int64_t i = 0; i < ntotal; i++) {
|
135
|
+
decode_64bit(i, tmp.data());
|
136
|
+
norms[i] = fvec_norm_L2sqr(tmp.data(), d);
|
137
|
+
}
|
138
|
+
}
|
139
|
+
}
|
140
|
+
|
141
|
+
void AdditiveQuantizer::decode_64bit(idx_t bits, float* xi) const {
|
142
|
+
for (int m = 0; m < M; m++) {
|
143
|
+
idx_t idx = bits & (((size_t)1 << nbits[m]) - 1);
|
144
|
+
bits >>= nbits[m];
|
145
|
+
const float* c = codebooks.data() + d * (codebook_offsets[m] + idx);
|
146
|
+
if (m == 0) {
|
147
|
+
memcpy(xi, c, sizeof(*xi) * d);
|
148
|
+
} else {
|
149
|
+
fvec_add(d, xi, c, xi);
|
150
|
+
}
|
151
|
+
}
|
152
|
+
}
|
153
|
+
|
154
|
+
void AdditiveQuantizer::compute_LUT(size_t n, const float* xq, float* LUT)
|
155
|
+
const {
|
156
|
+
// in all cases, it is large matrix multiplication
|
157
|
+
|
158
|
+
FINTEGER ncenti = total_codebook_size;
|
159
|
+
FINTEGER di = d;
|
160
|
+
FINTEGER nqi = n;
|
161
|
+
float one = 1, zero = 0;
|
162
|
+
|
163
|
+
sgemm_("Transposed",
|
164
|
+
"Not transposed",
|
165
|
+
&ncenti,
|
166
|
+
&nqi,
|
167
|
+
&di,
|
168
|
+
&one,
|
169
|
+
codebooks.data(),
|
170
|
+
&di,
|
171
|
+
xq,
|
172
|
+
&di,
|
173
|
+
&zero,
|
174
|
+
LUT,
|
175
|
+
&ncenti);
|
176
|
+
}
|
177
|
+
|
178
|
+
namespace {
|
179
|
+
|
180
|
+
void compute_inner_prod_with_LUT(
|
181
|
+
const AdditiveQuantizer& aq,
|
182
|
+
const float* LUT,
|
183
|
+
float* ips) {
|
184
|
+
size_t prev_size = 1;
|
185
|
+
for (int m = 0; m < aq.M; m++) {
|
186
|
+
const float* LUTm = LUT + aq.codebook_offsets[m];
|
187
|
+
int nb = aq.nbits[m];
|
188
|
+
size_t nc = (size_t)1 << nb;
|
189
|
+
|
190
|
+
if (m == 0) {
|
191
|
+
memcpy(ips, LUT, sizeof(*ips) * nc);
|
192
|
+
} else {
|
193
|
+
for (int64_t i = nc - 1; i >= 0; i--) {
|
194
|
+
float v = LUTm[i];
|
195
|
+
fvec_add(prev_size, ips, v, ips + i * prev_size);
|
196
|
+
}
|
197
|
+
}
|
198
|
+
prev_size *= nc;
|
199
|
+
}
|
200
|
+
}
|
201
|
+
|
202
|
+
} // anonymous namespace
|
203
|
+
|
204
|
+
void AdditiveQuantizer::knn_exact_inner_product(
|
205
|
+
idx_t n,
|
206
|
+
const float* xq,
|
207
|
+
idx_t k,
|
208
|
+
float* distances,
|
209
|
+
idx_t* labels) const {
|
210
|
+
std::unique_ptr<float[]> LUT(new float[n * total_codebook_size]);
|
211
|
+
compute_LUT(n, xq, LUT.get());
|
212
|
+
size_t ntotal = (size_t)1 << tot_bits;
|
213
|
+
|
214
|
+
#pragma omp parallel if (n > 100)
|
215
|
+
{
|
216
|
+
std::vector<float> dis(ntotal);
|
217
|
+
#pragma omp for
|
218
|
+
for (idx_t i = 0; i < n; i++) {
|
219
|
+
const float* LUTi = LUT.get() + i * total_codebook_size;
|
220
|
+
compute_inner_prod_with_LUT(*this, LUTi, dis.data());
|
221
|
+
float* distances_i = distances + i * k;
|
222
|
+
idx_t* labels_i = labels + i * k;
|
223
|
+
minheap_heapify(k, distances_i, labels_i);
|
224
|
+
minheap_addn(k, distances_i, labels_i, dis.data(), nullptr, ntotal);
|
225
|
+
minheap_reorder(k, distances_i, labels_i);
|
226
|
+
}
|
227
|
+
}
|
228
|
+
}
|
229
|
+
|
230
|
+
void AdditiveQuantizer::knn_exact_L2(
|
231
|
+
idx_t n,
|
232
|
+
const float* xq,
|
233
|
+
idx_t k,
|
234
|
+
float* distances,
|
235
|
+
idx_t* labels,
|
236
|
+
const float* norms) const {
|
237
|
+
std::unique_ptr<float[]> LUT(new float[n * total_codebook_size]);
|
238
|
+
compute_LUT(n, xq, LUT.get());
|
239
|
+
std::unique_ptr<float[]> q_norms(new float[n]);
|
240
|
+
fvec_norms_L2sqr(q_norms.get(), xq, d, n);
|
241
|
+
size_t ntotal = (size_t)1 << tot_bits;
|
242
|
+
|
243
|
+
#pragma omp parallel if (n > 100)
|
244
|
+
{
|
245
|
+
std::vector<float> dis(ntotal);
|
246
|
+
#pragma omp for
|
247
|
+
for (idx_t i = 0; i < n; i++) {
|
248
|
+
const float* LUTi = LUT.get() + i * total_codebook_size;
|
249
|
+
float* distances_i = distances + i * k;
|
250
|
+
idx_t* labels_i = labels + i * k;
|
251
|
+
|
252
|
+
compute_inner_prod_with_LUT(*this, LUTi, dis.data());
|
253
|
+
|
254
|
+
// update distances using
|
255
|
+
// ||x - y||^2 = ||x||^2 + ||y||^2 - 2 * <x,y>
|
256
|
+
|
257
|
+
maxheap_heapify(k, distances_i, labels_i);
|
258
|
+
for (idx_t j = 0; j < ntotal; j++) {
|
259
|
+
float disj = q_norms[i] + norms[j] - 2 * dis[j];
|
260
|
+
if (disj < distances_i[0]) {
|
261
|
+
heap_replace_top<CMax<float, int64_t>>(
|
262
|
+
k, distances_i, labels_i, disj, j);
|
263
|
+
}
|
264
|
+
}
|
265
|
+
maxheap_reorder(k, distances_i, labels_i);
|
266
|
+
}
|
267
|
+
}
|
268
|
+
}
|
269
|
+
|
270
|
+
} // namespace faiss
|
@@ -0,0 +1,115 @@
|
|
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
|
+
#pragma once
|
9
|
+
|
10
|
+
#include <cstdint>
|
11
|
+
#include <vector>
|
12
|
+
|
13
|
+
#include <faiss/Index.h>
|
14
|
+
|
15
|
+
namespace faiss {
|
16
|
+
|
17
|
+
/** Abstract structure for additive quantizers
|
18
|
+
*
|
19
|
+
* Different from the product quantizer in which the decoded vector is the
|
20
|
+
* concatenation of M sub-vectors, additive quantizers sum M sub-vectors
|
21
|
+
* to get the decoded vector.
|
22
|
+
*/
|
23
|
+
struct AdditiveQuantizer {
|
24
|
+
size_t d; ///< size of the input vectors
|
25
|
+
size_t M; ///< number of codebooks
|
26
|
+
std::vector<size_t> nbits; ///< bits for each step
|
27
|
+
std::vector<float> codebooks; ///< codebooks
|
28
|
+
|
29
|
+
// derived values
|
30
|
+
std::vector<size_t> codebook_offsets;
|
31
|
+
size_t code_size; ///< code size in bytes
|
32
|
+
size_t tot_bits; ///< total number of bits
|
33
|
+
size_t total_codebook_size; ///< size of the codebook in vectors
|
34
|
+
bool is_byte_aligned;
|
35
|
+
|
36
|
+
bool verbose; ///< verbose during training?
|
37
|
+
bool is_trained; ///< is trained or not
|
38
|
+
|
39
|
+
///< compute derived values when d, M and nbits have been set
|
40
|
+
void set_derived_values();
|
41
|
+
|
42
|
+
///< Train the additive quantizer
|
43
|
+
virtual void train(size_t n, const float* x) = 0;
|
44
|
+
|
45
|
+
/** Encode a set of vectors
|
46
|
+
*
|
47
|
+
* @param x vectors to encode, size n * d
|
48
|
+
* @param codes output codes, size n * code_size
|
49
|
+
*/
|
50
|
+
virtual void compute_codes(const float* x, uint8_t* codes, size_t n)
|
51
|
+
const = 0;
|
52
|
+
|
53
|
+
/** pack a series of code to bit-compact format
|
54
|
+
*
|
55
|
+
* @param codes codes to be packed, size n * code_size
|
56
|
+
* @param packed_codes output bit-compact codes
|
57
|
+
* @param ld_codes leading dimension of codes
|
58
|
+
*/
|
59
|
+
void pack_codes(
|
60
|
+
size_t n,
|
61
|
+
const int32_t* codes,
|
62
|
+
uint8_t* packed_codes,
|
63
|
+
int64_t ld_codes = -1) const;
|
64
|
+
|
65
|
+
/** Decode a set of vectors
|
66
|
+
*
|
67
|
+
* @param codes codes to decode, size n * code_size
|
68
|
+
* @param x output vectors, size n * d
|
69
|
+
*/
|
70
|
+
void decode(const uint8_t* codes, float* x, size_t n) const;
|
71
|
+
|
72
|
+
/****************************************************************************
|
73
|
+
* Support for exhaustive distance computations with the centroids.
|
74
|
+
* Hence, the number of elements that can be enumerated is not too large.
|
75
|
+
****************************************************************************/
|
76
|
+
using idx_t = Index::idx_t;
|
77
|
+
|
78
|
+
/// decoding function for a code in a 64-bit word
|
79
|
+
void decode_64bit(idx_t n, float* x) const;
|
80
|
+
|
81
|
+
/** Compute inner-product look-up tables. Used in the centroid search
|
82
|
+
* functions.
|
83
|
+
*
|
84
|
+
* @param xq query vector, size (n, d)
|
85
|
+
* @param LUT look-up table, size (n, total_codebook_size)
|
86
|
+
*/
|
87
|
+
void compute_LUT(size_t n, const float* xq, float* LUT) const;
|
88
|
+
|
89
|
+
/// exact IP search
|
90
|
+
void knn_exact_inner_product(
|
91
|
+
idx_t n,
|
92
|
+
const float* xq,
|
93
|
+
idx_t k,
|
94
|
+
float* distances,
|
95
|
+
idx_t* labels) const;
|
96
|
+
|
97
|
+
/** For L2 search we need the L2 norms of the centroids
|
98
|
+
*
|
99
|
+
* @param norms output norms table, size total_codebook_size
|
100
|
+
*/
|
101
|
+
void compute_centroid_norms(float* norms) const;
|
102
|
+
|
103
|
+
/** Exact L2 search, with precomputed norms */
|
104
|
+
void knn_exact_L2(
|
105
|
+
idx_t n,
|
106
|
+
const float* xq,
|
107
|
+
idx_t k,
|
108
|
+
float* distances,
|
109
|
+
idx_t* labels,
|
110
|
+
const float* centroid_norms) const;
|
111
|
+
|
112
|
+
virtual ~AdditiveQuantizer();
|
113
|
+
};
|
114
|
+
|
115
|
+
}; // namespace faiss
|