Implementing class for IVFPQ on the GPU. More...

Inheritance diagram for faiss::gpu::IVFPQ:

Public Member Functions
	IVFPQ (GpuResources resources, FlatIndex quantizer, int numSubQuantizers, int bitsPerSubQuantizer, float *pqCentroidData, IndicesOptions indicesOptions, bool useFloat16LookupTables, MemorySpace space)

void	setPrecomputedCodes (bool enable)
	Enable or disable pre-computed codes.

void	addCodeVectorsFromCpu (int listId, const void codes, const long indices, size_t numVecs)

int	classifyAndAddVectors (Tensor< float, 2, true > &vecs, Tensor< long, 1, true > &indices)

void	query (Tensor< float, 2, true > &queries, int nprobe, int k, Tensor< float, 2, true > &outDistances, Tensor< long, 2, true > &outIndices)

std::vector< unsigned char >	getListCodes (int listId) const
	Return the list codes of a particular list back to the CPU.

Tensor< float, 3, true >	getPQCentroids ()

Public Member Functions inherited from faiss::gpu::IVFBase
	IVFBase (GpuResources resources, FlatIndex quantizer, int bytesPerVector, IndicesOptions indicesOptions, MemorySpace space)

void	reserveMemory (size_t numVecs)
	Reserve GPU memory in our inverted lists for this number of vectors.

void	reset ()

int	getDim () const
	Return the number of dimensions we are indexing.

size_t	reclaimMemory ()

size_t	getNumLists () const
	Returns the number of inverted lists.

int	getListLength (int listId) const

std::vector< long >	getListIndices (int listId) const
	Return the list indices of a particular list back to the CPU.

Static Public Member Functions
static bool	isSupportedPQCodeLength (int size)
	Returns true if we support PQ in this size.

static bool	isSupportedNoPrecomputedSubDimSize (int dims)

Additional Inherited Members
Protected Member Functions inherited from faiss::gpu::IVFBase
size_t	reclaimMemory_ (bool exact)

void	updateDeviceListInfo_ (cudaStream_t stream)
	Update all device-side list pointer and size information.

void	updateDeviceListInfo_ (const std::vector< int > &listIds, cudaStream_t stream)

void	addIndicesFromCpu_ (int listId, const long *indices, size_t numVecs)
	Shared function to copy indices from CPU to GPU.

Protected Attributes inherited from faiss::gpu::IVFBase
GpuResources *	resources_
	Collection of GPU resources that we use.

FlatIndex *	quantizer_
	Quantizer object.

const int	dim_
	Expected dimensionality of the vectors.

const int	numLists_
	Number of inverted lists we maintain.

const int	bytesPerVector_
	Number of bytes per vector in the list.

const IndicesOptions	indicesOptions_
	How are user indices stored on the GPU?

const MemorySpace	space_
	What memory space our inverted list storage is in.

thrust::device_vector< void * >	deviceListDataPointers_

thrust::device_vector< void * >	deviceListIndexPointers_

thrust::device_vector< int >	deviceListLengths_

int	maxListLength_
	Maximum list length seen.

std::vector< std::unique_ptr < DeviceVector< unsigned char > > >	deviceListData_

std::vector< std::unique_ptr < DeviceVector< unsigned char > > >	deviceListIndices_

std::vector< std::vector< long > >	listOffsetToUserIndex_

Detailed Description

Implementing class for IVFPQ on the GPU.

Definition at line 17 of file IVFPQ.cuh.

Constructor & Destructor Documentation

faiss::gpu::IVFPQ::IVFPQ	(	GpuResources *	resources,
		FlatIndex *	quantizer,
		int	numSubQuantizers,
		int	bitsPerSubQuantizer,
		float *	pqCentroidData,
		IndicesOptions	indicesOptions,
		bool	useFloat16LookupTables,
		MemorySpace	space
	)

Parameters

quantizer We do not own this reference

Definition at line 33 of file IVFPQ.cu.

Member Function Documentation

void faiss::gpu::IVFPQ::addCodeVectorsFromCpu	(	int	listId,
		const void *	codes,
		const long *	indices,
		size_t	numVecs
	)

Adds a set of codes and indices to a list; the data can be resident on either the host or the device

Definition at line 345 of file IVFPQ.cu.

int faiss::gpu::IVFPQ::classifyAndAddVectors	(	Tensor< float, 2, true > &	vecs,
		Tensor< long, 1, true > &	indices
	)

Calcuates the residual and quantizes the vectors, adding them to this index The input data must be on our current device. Returns the number of vectors successfully added. Vectors may not be able to be added because they contain NaNs.

Definition at line 118 of file IVFPQ.cu.

Tensor< float, 3, true > faiss::gpu::IVFPQ::getPQCentroids ( )

Returns our set of sub-quantizers of the form (sub q)(code id)(sub dim)

Definition at line 591 of file IVFPQ.cu.

bool faiss::gpu::IVFPQ::isSupportedNoPrecomputedSubDimSize ( int dims )

static

For no precomputed codes, is this a supported sub-dimension size? FIXME: get MM implementation working again

Definition at line 95 of file IVFPQ.cu.

void faiss::gpu::IVFPQ::query	(	Tensor< float, 2, true > &	queries,
		int	nprobe,
		int	k,
		Tensor< float, 2, true > &	outDistances,
		Tensor< long, 2, true > &	outIndices
	)

Find the approximate k nearest neigbors for queries against our database

Definition at line 516 of file IVFPQ.cu.

The documentation for this class was generated from the following files:

/data/users/hoss/faiss/gpu/impl/IVFPQ.cuh
/data/users/hoss/faiss/gpu/impl/IVFPQ.cu

Public Member Functions

Static Public Member Functions

Additional Inherited Members

Detailed Description

Constructor & Destructor Documentation

Member Function Documentation