File IndexFastScan.h

namespace faiss

Implementation of k-means clustering with many variants.

Copyright (c) Facebook, Inc. and its affiliates.

This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree.

IDSelector is intended to define a subset of vectors to handle (for removal or as subset to search)

PQ4 SIMD packing and accumulation functions

The basic kernel accumulates nq query vectors with bbs = nb * 2 * 16 vectors and produces an output matrix for that. It is interesting for nq * nb <= 4, otherwise register spilling becomes too large.

The implementation of these functions is spread over 3 cpp files to reduce parallel compile times. Templates are instantiated explicitly.

This file contains callbacks for kernels that compute distances.

Throughout the library, vectors are provided as float * pointers. Most algorithms can be optimized when several vectors are processed (added/searched) together in a batch. In this case, they are passed in as a matrix. When n vectors of size d are provided as float * x, component j of vector i is

x[ i * d + j ]

where 0 <= i < n and 0 <= j < d. In other words, matrices are always compact. When specifying the size of the matrix, we call it an n*d matrix, which implies a row-major storage.

I/O functions can read/write to a filename, a file handle or to an object that abstracts the medium.

The read functions return objects that should be deallocated with delete. All references within these objectes are owned by the object.

Definition of inverted lists + a few common classes that implement the interface.

Since IVF (inverted file) indexes are of so much use for large-scale use cases, we group a few functions related to them in this small library. Most functions work both on IndexIVFs and IndexIVFs embedded within an IndexPreTransform.

In this file are the implementations of extra metrics beyond L2 and inner product

Implements a few neural net layers, mainly to support QINCo

Defines a few objects that apply transformations to a set of vectors Often these are pre-processing steps.

Variables

FAISS_API FastScanStats FastScan_stats
struct IndexFastScan : public faiss::Index
#include <IndexFastScan.h>

Fast scan version of IndexPQ and IndexAQ. Works for 4-bit PQ and AQ for now.

The codes are not stored sequentially but grouped in blocks of size bbs. This makes it possible to compute distances quickly with SIMD instructions. The trailing codes (padding codes that are added to complete the last code) are garbage.

Implementations: 12: blocked loop with internal loop on Q with qbs 13: same with reservoir accumulator to store results 14: no qbs with heap accumulator 15: no qbs with reservoir accumulator

Subclassed by faiss::IndexAdditiveQuantizerFastScan, faiss::IndexPQFastScan

Public Functions

void init_fastscan(int d, size_t M, size_t nbits, MetricType metric, int bbs)
IndexFastScan()
virtual void reset() override

removes all elements from the database.

virtual void search(idx_t n, const float *x, idx_t k, float *distances, idx_t *labels, const SearchParameters *params = nullptr) const override

query n vectors of dimension d to the index.

return at most k vectors. If there are not enough results for a query, the result array is padded with -1s.

Parameters:

  • n – number of vectors

  • x – input vectors to search, size n * d

  • k – number of extracted vectors

  • distances – output pairwise distances, size n*k

  • labels – output labels of the NNs, size n*k

virtual void add(idx_t n, const float *x) override

Add n vectors of dimension d to the index.

Vectors are implicitly assigned labels ntotal .. ntotal + n - 1 This function slices the input vectors in chunks smaller than blocksize_add and calls add_core.

Parameters:

  • n – number of vectors

  • x – input matrix, size n * d

virtual void compute_codes(uint8_t *codes, idx_t n, const float *x) const = 0
virtual void compute_float_LUT(float *lut, idx_t n, const float *x) const = 0
void compute_quantized_LUT(idx_t n, const float *x, uint8_t *lut, float *normalizers) const
template<bool is_max>
void search_dispatch_implem(idx_t n, const float *x, idx_t k, float *distances, idx_t *labels, const NormTableScaler *scaler) const
template<class Cfloat>
void search_implem_234(idx_t n, const float *x, idx_t k, float *distances, idx_t *labels, const NormTableScaler *scaler) const
template<class C>
void search_implem_12(idx_t n, const float *x, idx_t k, float *distances, idx_t *labels, int impl, const NormTableScaler *scaler) const
template<class C>
void search_implem_14(idx_t n, const float *x, idx_t k, float *distances, idx_t *labels, int impl, const NormTableScaler *scaler) const
virtual void reconstruct(idx_t key, float *recons) const override

Reconstruct a stored vector (or an approximation if lossy coding)

this function may not be defined for some indexes

Parameters:

  • key – id of the vector to reconstruct

  • recons – reconstucted vector (size d)

virtual size_t remove_ids(const IDSelector &sel) override

removes IDs from the index. Not supported by all indexes. Returns the number of elements removed.

CodePacker *get_CodePacker() const
virtual void merge_from(Index &otherIndex, idx_t add_id = 0) override

moves the entries from another dataset to self. On output, other is empty. add_id is added to all moved ids (for sequential ids, this would be this->ntotal)

virtual void check_compatible_for_merge(const Index &otherIndex) const override

check that the two indexes are compatible (ie, they are trained in the same way and have the same parameters). Otherwise throw.

Public Members

int implem = 0
int skip = 0
int bbs
int qbs = 0
size_t M
size_t nbits
size_t ksub
size_t code_size
size_t ntotal2
size_t M2
AlignedTable<uint8_t> codes
const uint8_t *orig_codes = nullptr
struct FastScanStats

Public Functions

inline FastScanStats()
inline void reset()

Public Members

uint64_t t0
uint64_t t1
uint64_t t2
uint64_t t3