#include <deal.II/matrix_free/cuda_matrix_free.h>

Inheritance diagram for CUDAWrappers::MatrixFree< dim, Number >:

Classes
struct	Data

Public Types
enum	ParallelizationScheme

Public Member Functions
	MatrixFree ()

unsigned int	get_padding_length () const

void	reinit (const Mapping< dim > &mapping, const DoFHandler< dim > &dof_handler, const AffineConstraints< Number > &constraints, const Quadrature< 1 > &quad, const AdditionalData additional_data=AdditionalData())

void	reinit (const DoFHandler< dim > &dof_handler, const AffineConstraints< Number > &constraints, const Quadrature< 1 > &quad, const AdditionalData AdditionalData=AdditionalData())

Data	get_data (unsigned int color) const

template<typename Functor , typename VectorType >
void	cell_loop (const Functor &func, const VectorType &src, VectorType &dst) const

template<typename Functor >
void	evaluate_coefficients (Functor func) const

template<typename VectorType >
void	copy_constrained_values (const VectorType &src, VectorType &dst) const

template<typename VectorType >
void	set_constrained_values (const Number val, VectorType &dst) const

void	free ()

std::size_t	memory_consumption () const

Public Member Functions inherited from Subscriptor
	Subscriptor ()

	Subscriptor (const Subscriptor &)

	Subscriptor (Subscriptor &&) noexcept

virtual	~Subscriptor ()

Subscriptor &	operator= (const Subscriptor &)

Subscriptor &	operator= (Subscriptor &&) noexcept

void	subscribe (std::atomic< bool > *const validity, const std::string &identifier="") const

void	unsubscribe (std::atomic< bool > *const validity, const std::string &identifier="") const

unsigned int	n_subscriptions () const

template<typename StreamType >
void	list_subscribers (StreamType &stream) const

void	list_subscribers () const

template<class Archive >
void	serialize (Archive &ar, const unsigned int version)

Private Member Functions
void	internal_reinit (const Mapping< dim > &mapping, const DoFHandler< dim > &dof_handler, const AffineConstraints< Number > &constraints, const Quadrature< 1 > &quad, std::shared_ptr< const MPI_Comm > comm, const AdditionalData additional_data)

template<typename Functor , typename VectorType >
void	serial_cell_loop (const Functor &func, const VectorType &src, VectorType &dst) const

template<typename Functor >
void	distributed_cell_loop (const Functor &func, const LinearAlgebra::distributed::Vector< Number, MemorySpace::CUDA > &src, LinearAlgebra::distributed::Vector< Number, MemorySpace::CUDA > &dst) const

template<typename Functor >
void	distributed_cell_loop (const Functor &func, const LinearAlgebra::CUDAWrappers::Vector< Number > &src, LinearAlgebra::CUDAWrappers::Vector< Number > &dst) const

template<typename VectorType >
void	serial_copy_constrained_values (const VectorType &src, VectorType &dst) const

void	distributed_copy_constrained_values (const LinearAlgebra::distributed::Vector< Number, MemorySpace::CUDA > &src, LinearAlgebra::distributed::Vector< Number, MemorySpace::CUDA > &dst) const

void	distributed_copy_constrained_values (const LinearAlgebra::CUDAWrappers::Vector< Number > &src, LinearAlgebra::CUDAWrappers::Vector< Number > &dst) const

template<typename VectorType >
void	serial_set_constrained_values (const Number val, VectorType &dst) const

void	distributed_set_constrained_values (const Number val, LinearAlgebra::distributed::Vector< Number, MemorySpace::CUDA > &dst) const

void	distributed_set_constrained_values (const Number val, LinearAlgebra::CUDAWrappers::Vector< Number > &dst) const

Private Attributes
ParallelizationScheme	parallelization_scheme

unsigned int	fe_degree

unsigned int	dofs_per_cell

unsigned int	n_constrained_dofs

unsigned int	q_points_per_cell

unsigned int	n_colors

std::vector< unsigned int >	n_cells

std::vector< point_type * >	q_points

std::vector< types::global_dof_index * >	local_to_global

std::vector< Number * >	inv_jacobian

std::vector< Number * >	JxW

std::vector< dim3 >	grid_dim

std::vector< dim3 >	block_dim

std::shared_ptr< const Utilities::MPI::Partitioner >	partitioner

Additional Inherited Members
Static Public Member Functions inherited from Subscriptor
static ::ExceptionBase &	ExcInUse (int arg1, std::string arg2, std::string arg3)

static ::ExceptionBase &	ExcNoSubscriber (std::string arg1, std::string arg2)

Detailed Description

template<int dim, typename Number = double>
class CUDAWrappers::MatrixFree< dim, Number >

This class collects all the data that is stored for the matrix free implementation. The storage scheme is tailored towards several loops performed with the same data, i.e., typically doing many matrix-vector products or residual computations on the same mesh.

This class does not implement any operations involving finite element basis functions, i.e., regarding the operation performed on the cells. For these operations, the class FEEvaluation is designed to use the data collected in this class.

This class implements a loop over all cells (cell_loop()). This loop is scheduled in such a way that cells that cells that share degrees of freedom are not worked on simultaneously, which implies that it is possible to write to vectors in parallel without having to explicitly synchronize access to these vectors and matrices. This class does not implement any shape values, all it does is to cache the respective data. To implement finite element operations, use the class CUDAWrappers::FEEvalutation.

This class traverse the cells in a different order than the usual Triangulation class in deal.II.

Note: Only float and double are supported.

Definition at line 77 of file cuda_matrix_free.h.

Member Enumeration Documentation

◆ ParallelizationScheme

template<int dim, typename Number = double>

enum CUDAWrappers::MatrixFree::ParallelizationScheme

Parallelization scheme used: parallel_in_elem (parallelism at the level of degrees of freedom) or parallel_over_elem (parallelism at the level of cells)

Definition at line 88 of file cuda_matrix_free.h.

Constructor & Destructor Documentation

◆ MatrixFree()

template<int dim, typename Number = double>

CUDAWrappers::MatrixFree< dim, Number >::MatrixFree ( )

Default constructor.

Member Function Documentation

◆ get_padding_length()

template<int dim, typename Number = double>

unsigned int CUDAWrappers::MatrixFree< dim, Number >::get_padding_length ( ) const

Return the length of the padding.

◆ reinit() [1/2]

template<int dim, typename Number = double>

void CUDAWrappers::MatrixFree< dim, Number >::reinit	(	const Mapping< dim > &	mapping,
		const DoFHandler< dim > &	dof_handler,
		const AffineConstraints< Number > &	constraints,
		const Quadrature< 1 > &	quad,
		const AdditionalData	additional_data = `AdditionalData()`
	)

Extracts the information needed to perform loops over cells. The DoFHandler and AffineConstraints objects describe the layout of degrees of freedom, the DoFHandler and the mapping describe the transformation from unit to real cell, and the finite element underlying the DoFHandler together with the quadrature formula describe the local operations.

◆ reinit() [2/2]

template<int dim, typename Number = double>

void CUDAWrappers::MatrixFree< dim, Number >::reinit	(	const DoFHandler< dim > &	dof_handler,
		const AffineConstraints< Number > &	constraints,
		const Quadrature< 1 > &	quad,
		const AdditionalData	AdditionalData = `AdditionalData()`
	)

Initializes the data structures. Same as above but using a Q1 mapping.

◆ get_data()

template<int dim, typename Number = double>

Data CUDAWrappers::MatrixFree< dim, Number >::get_data ( unsigned int color ) const

Return the Data structure associated with color.

◆ cell_loop()

template<int dim, typename Number = double>

template<typename Functor , typename VectorType >

void CUDAWrappers::MatrixFree< dim, Number >::cell_loop	(	const Functor &	func,
		const VectorType &	src,
		VectorType &	dst
	)		const

This method runs the loop over all cells and apply the local operation on each element in parallel. func is a functor which is applied on each color.

func needs to define

__device__ void operator()(
  const unsigned int                                          cell,
  const typename CUDAWrappers::MatrixFree<dim, Number>::Data *gpu_data,
  CUDAWrappers::SharedData<dim, Number> *                     shared_data,
  const Number *                                              src,
  Number *                                                    dst) const;
  static const unsigned int n_dofs_1d;
  static const unsigned int n_local_dofs;
  static const unsigned int n_q_points;

◆ evaluate_coefficients()

template<int dim, typename Number = double>

template<typename Functor >

void CUDAWrappers::MatrixFree< dim, Number >::evaluate_coefficients ( Functor func ) const

This method runs the loop over all cells and apply the local operation on each element in parallel. This function is very similar to cell_loop() but it uses a simpler functor.

func needs to define

 __device__ void operator()(
   const unsigned int                                          cell,
   const typename CUDAWrappers::MatrixFree<dim, Number>::Data *gpu_data);
static const unsigned int n_dofs_1d;
static const unsigned int n_local_dofs;
static const unsigned int n_q_points;

◆ copy_constrained_values()

template<int dim, typename Number = double>

template<typename VectorType >

void CUDAWrappers::MatrixFree< dim, Number >::copy_constrained_values	(	const VectorType &	src,
		VectorType &	dst
	)		const

Copy the values of the constrained entries from src to dst. This is used to impose zero Dirichlet boundary condition.

◆ set_constrained_values()

template<int dim, typename Number = double>

template<typename VectorType >

void CUDAWrappers::MatrixFree< dim, Number >::set_constrained_values	(	const Number	val,
		VectorType &	dst
	)		const

Set the entries in dst corresponding to constrained values to val. The main purpose of this function is to set the constrained entries of the source vector used in cell_loop() to zero.

◆ free()

template<int dim, typename Number = double>

void CUDAWrappers::MatrixFree< dim, Number >::free ( )

Free all the memory allocated.

◆ memory_consumption()

template<int dim, typename Number = double>

std::size_t CUDAWrappers::MatrixFree< dim, Number >::memory_consumption ( ) const

Return an approximation of the memory consumption of this class in bytes.

◆ internal_reinit()

template<int dim, typename Number = double>

void CUDAWrappers::MatrixFree< dim, Number >::internal_reinit	(	const Mapping< dim > &	mapping,
		const DoFHandler< dim > &	dof_handler,
		const AffineConstraints< Number > &	constraints,
		const Quadrature< 1 > &	quad,
		std::shared_ptr< const MPI_Comm >	comm,
		const AdditionalData	additional_data
	)

private

Initializes the data structures.

◆ serial_cell_loop()

template<int dim, typename Number = double>

template<typename Functor , typename VectorType >

void CUDAWrappers::MatrixFree< dim, Number >::serial_cell_loop	(	const Functor &	func,
		const VectorType &	src,
		VectorType &	dst
	)		const

private

Helper function. Loop over all the cells and apply the functor on each element in parallel. This function is used when MPI is not used.

◆ distributed_cell_loop() [1/2]

template<int dim, typename Number = double>

template<typename Functor >

void CUDAWrappers::MatrixFree< dim, Number >::distributed_cell_loop	(	const Functor &	func,
		const LinearAlgebra::distributed::Vector< Number, MemorySpace::CUDA > &	src,
		LinearAlgebra::distributed::Vector< Number, MemorySpace::CUDA > &	dst
	)		const

private

Helper function. Loop over all the cells and apply the functor on each element in parallel. This function is used when MPI is used.

◆ distributed_cell_loop() [2/2]

template<int dim, typename Number = double>

template<typename Functor >

void CUDAWrappers::MatrixFree< dim, Number >::distributed_cell_loop	(	const Functor &	func,
		const LinearAlgebra::CUDAWrappers::Vector< Number > &	src,
		LinearAlgebra::CUDAWrappers::Vector< Number > &	dst
	)		const

private

This function should never be called. Calling it results in an internal error. This function exists only because cell_loop needs distributed_cell_loop() to exist for LinearAlgebra::CUDAWrappers::Vector.

◆ serial_copy_constrained_values()

template<int dim, typename Number = double>

template<typename VectorType >

void CUDAWrappers::MatrixFree< dim, Number >::serial_copy_constrained_values	(	const VectorType &	src,
		VectorType &	dst
	)		const

private

Helper function. Copy the values of the constrained entries of src to dst. This function is used when MPI is not used.

◆ distributed_copy_constrained_values() [1/2]

template<int dim, typename Number = double>

void CUDAWrappers::MatrixFree< dim, Number >::distributed_copy_constrained_values	(	const LinearAlgebra::distributed::Vector< Number, MemorySpace::CUDA > &	src,
		LinearAlgebra::distributed::Vector< Number, MemorySpace::CUDA > &	dst
	)		const

private

Helper function. Copy the values of the constrained entries of src to dst. This function is used when MPI is used.

◆ distributed_copy_constrained_values() [2/2]

template<int dim, typename Number = double>

void CUDAWrappers::MatrixFree< dim, Number >::distributed_copy_constrained_values	(	const LinearAlgebra::CUDAWrappers::Vector< Number > &	src,
		LinearAlgebra::CUDAWrappers::Vector< Number > &	dst
	)		const

private

This function should never be called. Calling it results in an internal error. This function exists only because copy_constrained_values needs distributed_copy_constrained_values() to exist for LinearAlgebra::CUDAWrappers::Vector.

◆ serial_set_constrained_values()

template<int dim, typename Number = double>

template<typename VectorType >

void CUDAWrappers::MatrixFree< dim, Number >::serial_set_constrained_values	(	const Number	val,
		VectorType &	dst
	)		const

private

Helper function. Set the constrained entries of dst to val. This function is used when MPI is not used.

◆ distributed_set_constrained_values() [1/2]

template<int dim, typename Number = double>

void CUDAWrappers::MatrixFree< dim, Number >::distributed_set_constrained_values	(	const Number	val,
		LinearAlgebra::distributed::Vector< Number, MemorySpace::CUDA > &	dst
	)		const

private

Helper function. Set the constrained entries of dst to val. This function is used when MPI is used.

◆ distributed_set_constrained_values() [2/2]

template<int dim, typename Number = double>

void CUDAWrappers::MatrixFree< dim, Number >::distributed_set_constrained_values	(	const Number	val,
		LinearAlgebra::CUDAWrappers::Vector< Number > &	dst
	)		const

private

This function should never be called. Calling it results in an internal error. This function exists only because set_constrained_values needs distributed_set_constrained_values() to exist for LinearAlgebra::CUDAWrappers::Vector.

Member Data Documentation

◆ parallelization_scheme

template<int dim, typename Number = double>

ParallelizationScheme CUDAWrappers::MatrixFree< dim, Number >::parallelization_scheme

private

Parallelization scheme used, parallelization over degrees of freedom or over cells.

Definition at line 361 of file cuda_matrix_free.h.

◆ fe_degree

template<int dim, typename Number = double>

unsigned int CUDAWrappers::MatrixFree< dim, Number >::fe_degree

private

Degree of the finite element used.

Definition at line 366 of file cuda_matrix_free.h.

◆ dofs_per_cell

template<int dim, typename Number = double>

unsigned int CUDAWrappers::MatrixFree< dim, Number >::dofs_per_cell

private

Number of degrees of freedom per cell.

Definition at line 371 of file cuda_matrix_free.h.

◆ n_constrained_dofs

template<int dim, typename Number = double>

unsigned int CUDAWrappers::MatrixFree< dim, Number >::n_constrained_dofs

private

Number of constrained degrees of freedom.

Definition at line 376 of file cuda_matrix_free.h.

◆ q_points_per_cell

template<int dim, typename Number = double>

unsigned int CUDAWrappers::MatrixFree< dim, Number >::q_points_per_cell

private

Number of quadrature points per cells.

Definition at line 381 of file cuda_matrix_free.h.

◆ n_colors

template<int dim, typename Number = double>

unsigned int CUDAWrappers::MatrixFree< dim, Number >::n_colors

private

Number of colors produced by the graph coloring algorithm.

Definition at line 386 of file cuda_matrix_free.h.

◆ n_cells

template<int dim, typename Number = double>

std::vector<unsigned int> CUDAWrappers::MatrixFree< dim, Number >::n_cells

private

Number of cells in each color.

Definition at line 391 of file cuda_matrix_free.h.

◆ q_points

template<int dim, typename Number = double>

std::vector<point_type *> CUDAWrappers::MatrixFree< dim, Number >::q_points

private

Vector of pointers to the quadrature points associated to the cells of each color.

Definition at line 397 of file cuda_matrix_free.h.

◆ local_to_global

template<int dim, typename Number = double>

std::vector<types::global_dof_index *> CUDAWrappers::MatrixFree< dim, Number >::local_to_global

private

Map the position in the local vector to the position in the global vector.

Definition at line 403 of file cuda_matrix_free.h.

◆ inv_jacobian

template<int dim, typename Number = double>

std::vector<Number *> CUDAWrappers::MatrixFree< dim, Number >::inv_jacobian

private

Vector of pointer to the inverse Jacobian associated to the cells of each color.

Definition at line 409 of file cuda_matrix_free.h.

◆ JxW

template<int dim, typename Number = double>

std::vector<Number *> CUDAWrappers::MatrixFree< dim, Number >::JxW

private

Vector of pointer to the Jacobian time the weights associated to the cells of each color.

Definition at line 415 of file cuda_matrix_free.h.

◆ grid_dim

template<int dim, typename Number = double>

std::vector<dim3> CUDAWrappers::MatrixFree< dim, Number >::grid_dim

private

Grid dimensions associated to the different colors. The grid dimensions are used to launch the CUDA kernels.

Definition at line 427 of file cuda_matrix_free.h.

◆ block_dim

template<int dim, typename Number = double>

std::vector<dim3> CUDAWrappers::MatrixFree< dim, Number >::block_dim

private

Block dimensions associated to the different colors. The block dimensions are used to launch the CUDA kernels.

Definition at line 433 of file cuda_matrix_free.h.

◆ partitioner

template<int dim, typename Number = double>

std::shared_ptr<const Utilities::MPI::Partitioner> CUDAWrappers::MatrixFree< dim, Number >::partitioner

private

Shared pointer to a Partitioner for distributed Vectors used in cell_loop. When MPI is not used the pointer is null.

Definition at line 439 of file cuda_matrix_free.h.

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

deal.II/matrix_free/cuda_matrix_free.h

Classes

Public Types

Public Member Functions

Private Member Functions

Private Attributes

Additional Inherited Members

Detailed Description

template<int dim, typename Number = double> class CUDAWrappers::MatrixFree< dim, Number >

Member Enumeration Documentation

◆ ParallelizationScheme

Constructor & Destructor Documentation

◆ MatrixFree()

Member Function Documentation

◆ get_padding_length()

◆ reinit() [1/2]

◆ reinit() [2/2]

◆ get_data()

◆ cell_loop()

◆ evaluate_coefficients()

◆ copy_constrained_values()

◆ set_constrained_values()

◆ free()

◆ memory_consumption()

◆ internal_reinit()

◆ serial_cell_loop()

◆ distributed_cell_loop() [1/2]

◆ distributed_cell_loop() [2/2]

◆ serial_copy_constrained_values()

◆ distributed_copy_constrained_values() [1/2]

◆ distributed_copy_constrained_values() [2/2]

◆ serial_set_constrained_values()

◆ distributed_set_constrained_values() [1/2]

◆ distributed_set_constrained_values() [2/2]

Member Data Documentation

◆ parallelization_scheme

◆ fe_degree

◆ dofs_per_cell

◆ n_constrained_dofs

◆ q_points_per_cell

◆ n_colors

◆ n_cells

◆ q_points

◆ local_to_global

◆ inv_jacobian

◆ JxW

◆ grid_dim

◆ block_dim

◆ partitioner

template<int dim, typename Number = double>
class CUDAWrappers::MatrixFree< dim, Number >