File quantize_lut.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.

namespace quantize_lut

Functions to quantize PQ floating-point Look Up Tables (LUT) to uint8, and biases to uint16. The accumulation is supposed to take place in uint16. The quantization coefficients are float (a, b) such that 

 original_value = quantized_value * a / b
The hardest part of the quantization is with multiple LUTs that need to be added up together. In that case, coefficient a has to be chosen so that the sum fits in a uint16 accumulator.

Functions

void round_uint8_per_column(float *tab, size_t n, size_t d, float *a_out = nullptr, float *b_out = nullptr)
void round_uint8_per_column_multi(float *tab, size_t m, size_t n, size_t d, float *a_out = nullptr, float *b_out = nullptr)
void quantize_LUT_and_bias(size_t nprobe, size_t M, size_t ksub, bool lut_is_3d, const float *LUT, const float *bias, uint8_t *LUTq, size_t M2, uint16_t *biasq, float *a_out = nullptr, float *b_out = nullptr)

LUT quantization to uint8 and bias to uint16.

(nprobe, M, ksub, lut_is_3d) determine the size of the the LUT

LUT input:

  • 2D size (M, ksub): single matrix per probe (lut_is_3d=false)

  • 3D size (nprobe, M, ksub): separate LUT per probe (lut_is_3d=true) bias input:

  • nullptr: bias is 0

  • size (nprobe): one bias per probe Output:

  • LUTq uint8 version of the LUT (M size is rounded up to M2)

  • biasq (or nullptr): uint16 version of the LUT

  • a, b: scalars to approximate the true distance

void aq_quantize_LUT_and_bias(size_t nprobe, size_t M, size_t ksub, const float *LUT, const float *bias, size_t M_norm, int norm_scale, uint8_t *LUTq, size_t M2, uint16_t *biasq, float *a_out, float *b_out)
float aq_estimate_norm_scale(size_t M, size_t ksub, size_t M_norm, const float *LUT)