Namespace containing bld_mult update functions (CUDA/non-CUDA). More...

Namespaces
	kernel
	Namespace containing CUDA kernels used in bld_mult CUDA updates.

Functions
template<typename Real >
void	update_grad_plus_cuda (const cuda::DeviceMemory3D< Real > &S, const cuda::DeviceMemory2D< Real > &beta_eph, cuda::DeviceMemory3D< Real > &grad_plus)
	Perform grad_plus update employed in bld_mult algorithm using CUDA. More...

template<typename Real >
void	update_nom_cuda (const cuda::DeviceMemory2D< Real > &X_reciprocal, const cuda::DeviceMemory2D< Real > &grad_minus, const cuda::DeviceMemory3D< Real > &S, cuda::DeviceMemory2D< Real > &nom)
	Perform nom_mult update employed in bld_mult algorithm using CUDA. More...

template<typename Real >
void	update_denom_cuda (const cuda::DeviceMemory2D< Real > &X_reciprocal, const cuda::DeviceMemory3D< Real > &grad_plus, const cuda::DeviceMemory3D< Real > &S, cuda::DeviceMemory2D< Real > &denom)
	Perform denom update employed in bld_mult algorithm using CUDA. More...

template<typename Real >
void	update_S_cuda (const cuda::DeviceMemory2D< Real > &X, const cuda::DeviceMemory2D< Real > &nom, const cuda::DeviceMemory2D< Real > &denom, const cuda::DeviceMemory2D< Real > &grad_minus, const cuda::DeviceMemory3D< Real > &grad_plus, const cuda::DeviceMemory2D< Real > &S_ijp, cuda::DeviceMemory3D< Real > &S)
	Perform S update employed in bld_mult algorithm using CUDA. More...

template<typename T >
tensor_t< T, 3 >	init_S (const matrix_t< T > &X, size_t z)
	Initialize S tensor using a Dirichlet sample of size z with all concentration parameters set to 1. More...

template<typename T >
matrix_t< T >	X_reciprocal (const matrix_t< T > &X, double eps)
	Compute the reciprocal \(\hat{X}\) of the input matrix \(X\). More...

template<typename Scalar >
std::pair< vector_t< Scalar >, vector_t< Scalar > >	init_alpha_beta (const alloc_model::Params< Scalar > &params)
	Initialize alpha and beta vectors used in bld_mult. More...

template<typename T >
void	update_alpha_eph (const tensor_t< T, 2 > &S_ipk, const vector_t< T > &alpha, matrix_t< T > &alpha_eph)
	Update alpha_eph matrix used in bld_mult. More...

template<typename T >
void	update_beta_eph (const tensor_t< T, 2 > &S_pjk, const vector_t< T > &beta, matrix_t< T > &beta_eph)
	Update beta_eph matrix used in bld_mult. More...

template<typename T , typename PsiFunction >
void	update_grad_plus (const tensor_t< T, 3 > &S, const matrix_t< T > &beta_eph, PsiFunction psi_fn, tensor_t< T, 3 > &grad_plus)
	Update grad_plus tensor used in bld_mult. More...

template<typename T , typename PsiFunction >
void	update_grad_minus (const matrix_t< T > &alpha_eph, PsiFunction psi_fn, matrix_t< T > &grad_minus)
	Update grad_minus tensor used in bld_mult. More...

template<typename T >
void	update_nom_mult (const matrix_t< T > &X_reciprocal, const matrix_t< T > &grad_minus, const tensor_t< T, 3 > &S, matrix_t< T > &nom_mult)
	Update nom_mult matrix used in bld_mult. More...

template<typename T >
void	update_denom_mult (const matrix_t< T > &X_reciprocal, const tensor_t< T, 3 > &grad_plus, const tensor_t< T, 3 > &S, matrix_t< T > &denom_mult)
	Update denom_mult matrix used in bld_mult. More...

template<typename T >
void	update_S (const matrix_t< T > &X, const matrix_t< T > &nom, const matrix_t< T > &denom, const matrix_t< T > &grad_minus, const tensor_t< T, 3 > &grad_plus, const tensor_t< T, 2 > &S_ijp, tensor_t< T, 3 > &S, double eps)
	Update S tensor (output of bld_mult algorithm). More...

Detailed Description

Namespace containing bld_mult update functions (CUDA/non-CUDA).

Function Documentation

template<typename Scalar >

std::pair<vector_t<Scalar>, vector_t<Scalar> > bnmf_algs::details::bld_mult::init_alpha_beta ( const alloc_model::Params< Scalar > & params )

Initialize alpha and beta vectors used in bld_mult.

Template Parameters

Scalar Type of the entries of model parameters.

Parameters

params Allocation model parameters.

Returns: std::pair of alpha and beta vectors which are copies of the model parameter alpha and beta vectors.

template<typename T >

tensor_t<T, 3> bnmf_algs::details::bld_mult::init_S	(	const matrix_t< T > &	X,
		size_t	z
	)

Initialize S tensor using a Dirichlet sample of size z with all concentration parameters set to 1.

S tensor is initialized according to the following update

\[ S_{ijk} \leftarrow X_{ij}\mathcal{D}_{ij}(k) \]

where \(\mathcal{D}_{ij}(k)\) is the kth entry of the \(ij^{th}\) Dirichlet sample.

Template Parameters

T	Type of the entries of the matrix X and tensor S

Parameters

X	X matrix sent as input to bld algorithm.
z	Rank of the decomposition.

Returns: S_{x y z} tensor for a matrix \(X_{x \times y}\) and rank parameter z.

template<typename T >

void bnmf_algs::details::bld_mult::update_alpha_eph	(	const tensor_t< T, 2 > &	S_ipk,
		const vector_t< T > &	alpha,
		matrix_t< T > &	alpha_eph
	)

Update alpha_eph matrix used in bld_mult.

Template Parameters

T	Type of the matrix entries.

Parameters

S_ipk	Sum of S tensor along its 1st axis (0 based indexing)
alpha	alpha vector used in bld_mult
alpha_eph	alpha_eph matrix to update.

template<typename T >

void bnmf_algs::details::bld_mult::update_beta_eph	(	const tensor_t< T, 2 > &	S_pjk,
		const vector_t< T > &	beta,
		matrix_t< T > &	beta_eph
	)

Update beta_eph matrix used in bld_mult.

Template Parameters

T	Type of the matrix entries.

Parameters

S_pjk	Sum of S tensor along its 0th axis (0 based indexing)
beta	beta vector used in bld_mult
beta_eph	beta_eph matrix to update.

template<typename Real >

void bnmf_algs::details::bld_mult::update_denom_cuda	(	const cuda::DeviceMemory2D< Real > &	X_reciprocal,
		const cuda::DeviceMemory3D< Real > &	grad_plus,
		const cuda::DeviceMemory3D< Real > &	S,
		cuda::DeviceMemory2D< Real > &	denom
	)

Perform denom update employed in bld_mult algorithm using CUDA.

This function performs the denom update employed in bld_mult algorithm using CUDA. To learn more about denom update, refer to details::bld_mult::update_denom_mult function documentation.

All the given tensor and matrix objects are assumed to be in row-major order . This is the allocation order specified in defs.hpp file. All the CUDA indexing operations are performed according to the row-major allocation order.

Template Parameters

Real	Type of the entries of matrices and tensors such as double or float.

Parameters

X_reciprocal	Matrix \(\hat{X}\) where \(\hat{X}_{ij} = \frac{1}{X_{ij}}\). Should reside on the GPU.
grad_plus	grad_plus tensor used in bld_mult on the GPU.
S	\(S\) tensor on the GPU.
denom	denom_mult matrix on the GPU.

template<typename T >

void bnmf_algs::details::bld_mult::update_denom_mult	(	const matrix_t< T > &	X_reciprocal,
		const tensor_t< T, 3 > &	grad_plus,
		const tensor_t< T, 3 > &	S,
		matrix_t< T > &	denom_mult
	)

Update denom_mult matrix used in bld_mult.

Template Parameters

T	Type of the matrix entries.

Parameters

X_reciprocal	Matrix containing reciprocals of entries of input matrix X.
grad_plus	grad_plus tensor.
S	S tensor
denom_mult	denom_mult matrix to update.

template<typename T , typename PsiFunction >

void bnmf_algs::details::bld_mult::update_grad_minus	(	const matrix_t< T > &	alpha_eph,
		PsiFunction	psi_fn,
		matrix_t< T > &	grad_minus
	)

Update grad_minus tensor used in bld_mult.

Template Parameters

T	Type of the matrix/tensor entries
PsiFunction	A callable type that will return \(\psi(x)\) for an input parameter \(x\).

Parameters

alpha_eph	alpha_matrix used in bld_mult.
psi_fn	Psi computing callable.
grad_minus	grad_minus tensor to update.

template<typename T , typename PsiFunction >

void bnmf_algs::details::bld_mult::update_grad_plus	(	const tensor_t< T, 3 > &	S,
		const matrix_t< T > &	beta_eph,
		PsiFunction	psi_fn,
		tensor_t< T, 3 > &	grad_plus
	)

Update grad_plus tensor used in bld_mult.

Template Parameters

T	Type of the matrix/tensor entries
PsiFunction	A callable type that will return \(\psi(x)\) for an input parameter \(x\).

Parameters

S	S tensor (output of bld_mult).
beta_eph	beta_eph matrix used in bld_mult.
psi_fn	Psi computing callable.
grad_plus	grad_plus tensor to update.

template<typename Real >

void bnmf_algs::details::bld_mult::update_grad_plus_cuda	(	const cuda::DeviceMemory3D< Real > &	S,
		const cuda::DeviceMemory2D< Real > &	beta_eph,
		cuda::DeviceMemory3D< Real > &	grad_plus
	)

Perform grad_plus update employed in bld_mult algorithm using CUDA.

This function performs the grad_plus update employed in bld_mult algorithm using CUDA. To learn more about grad_plus update, refer to details::bld_mult::update_grad_plus function documentation.

All the given tensor and matrix objects are assumed to be in row-major order . This is the allocation order specified in defs.hpp file. All the CUDA indexing operations are performed according to the row-major allocation order.

Row-major order means that the last index of a matrix/tensor objects changes the fastest. This is easy to understand with matrices: Matrix is stored in terms of its rows (0th row, then 1st row, ...). Allocation order of a 3D tensor follows the same idea: Tensor is stored in terms of its fibers (depth rows). Therefore, for a a 2x2x3 tensor is stored as follows (only writing the indices):

(0,0,0) (0,0,1) (0,0,2) (0,1,0) (0,1,1) (0,1,2) (1,0,0) (1,0,1) (1,0,2) (1,1,0) (1,1,1) (1,1,2)

Template Parameters

Real	Type of the entries of matrices and tensors such as double or float.

Parameters

S	\(S\) tensor on the GPU.
beta_eph	beta_eph matrix on the GPU.
grad_plus	grad_plus tensor on the GPU.

template<typename Real >

void bnmf_algs::details::bld_mult::update_nom_cuda	(	const cuda::DeviceMemory2D< Real > &	X_reciprocal,
		const cuda::DeviceMemory2D< Real > &	grad_minus,
		const cuda::DeviceMemory3D< Real > &	S,
		cuda::DeviceMemory2D< Real > &	nom
	)

Perform nom_mult update employed in bld_mult algorithm using CUDA.

This function performs the nom_mult update employed in bld_mult algorithm using CUDA. To learn more about nom_mult update, refer to details::bld_mult::update_nom_mult function documentation.

All the given tensor and matrix objects are assumed to be in row-major order . This is the allocation order specified in defs.hpp file. All the CUDA indexing operations are performed according to the row-major allocation order.

Template Parameters

Real	Type of the entries of matrices and tensors such as double or float.

Parameters

X_reciprocal	Matrix \(\hat{X}\) where \(\hat{X}_{ij} = \frac{1}{X_{ij}}\). Should reside on the GPU.
grad_minus	grad_minus matrix used in bld_mult on the GPU.
S	\(S\) tensor on the GPU.
nom	nom_mult matrix on the GPU.

template<typename T >

void bnmf_algs::details::bld_mult::update_nom_mult	(	const matrix_t< T > &	X_reciprocal,
		const matrix_t< T > &	grad_minus,
		const tensor_t< T, 3 > &	S,
		matrix_t< T > &	nom_mult
	)

Update nom_mult matrix used in bld_mult.

Template Parameters

T	Type of the matrix entries.

Parameters

X_reciprocal	Matrix containing reciprocals of entries of input matrix X.
grad_minus	grad_minus matrix.
S	S tensor
nom_mult	nom_mult matrix to update.

template<typename T >

void bnmf_algs::details::bld_mult::update_S	(	const matrix_t< T > &	X,
		const matrix_t< T > &	nom,
		const matrix_t< T > &	denom,
		const matrix_t< T > &	grad_minus,
		const tensor_t< T, 3 > &	grad_plus,
		const tensor_t< T, 2 > &	S_ijp,
		tensor_t< T, 3 > &	S,
		double	eps
	)

Update S tensor (output of bld_mult algorithm).

Template Parameters

T	Type of tensor entries.

Parameters

X	Input matrix passed to bld_mult algorithm.
nom	nom_mult matrix used in bld_mult algorithm.
denom	denom_mult matrix used in bld_mult algorithm.
grad_minus	grad_minus matrix used in bld_mult algorithm.
grad_plus	grad_plus tensor used in bld_mult algorithm.
S_ijp	Matrix computed by summing tensor S along its 2nd axis (0 based indexing).
S	S tensor to update.
eps	Epsilon value to prevent division by 0 errors.

template<typename Real >

void bnmf_algs::details::bld_mult::update_S_cuda	(	const cuda::DeviceMemory2D< Real > &	X,
		const cuda::DeviceMemory2D< Real > &	nom,
		const cuda::DeviceMemory2D< Real > &	denom,
		const cuda::DeviceMemory2D< Real > &	grad_minus,
		const cuda::DeviceMemory3D< Real > &	grad_plus,
		const cuda::DeviceMemory2D< Real > &	S_ijp,
		cuda::DeviceMemory3D< Real > &	S
	)

Perform S update employed in bld_mult algorithm using CUDA.

This function performs the S update employed in bld_mult algorithm using CUDA. To learn more about S update, refer to details::bld_mult::update_S function documentation.

All the given tensor and matrix objects are assumed to be in row-major order . This is the allocation order specified in defs.hpp file. All the CUDA indexing operations are performed according to the row-major allocation order.

Template Parameters

Real	Type of the entries of matrices and tensors such as double or float.

Parameters

X	Input X matrix on the GPU.
nom	nom_mult.
denom	denom_mult.
grad_minus	grad_minus matrix.
grad_plus	grad_plus tensor used in bld_mult on the GPU.
S_ijp	Matrix formed by summing S tensor along its axis 2. (last axis)
S	\(S\) tensor on the GPU.

template<typename T >

matrix_t<T> bnmf_algs::details::bld_mult::X_reciprocal	(	const matrix_t< T > &	X,
		double	eps
	)

Compute the reciprocal \(\hat{X}\) of the input matrix \(X\).

\(\hat{X}_{ij} = \frac{1}{X_{ij} + \epsilon}\).

Template Parameters

T	Type of the entries of the input matrix.

Parameters

X	Input matrix.
eps	Epsilon value to be used to prevent division by 0 errors.

Returns: Reciprocal of the input matrix.

Namespaces

Functions

Detailed Description

Function Documentation