Reference documentation for deal.II version 9.1.0-pre
Classes | Public Types | Public Member Functions | Private Attributes | List of all members
CUDAWrappers::MatrixFree< dim, Number > Class Template Reference

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

Inheritance diagram for CUDAWrappers::MatrixFree< dim, Number >:
[legend]

Classes

struct  Data
 

Public Types

Public Member Functions

 MatrixFree ()
 
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 >
void cell_loop (const functor &func, const CUDAVector< Number > &src, CUDAVector< Number > &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 ()
 
Subscriptoroperator= (const Subscriptor &)
 
Subscriptoroperator= (Subscriptor &&) noexcept
 
void subscribe (const char *identifier=nullptr) const
 
void unsubscribe (const char *identifier=nullptr) 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 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< unsigned int * > local_to_global
 
std::vector< Number * > inv_jacobian
 
std::vector< Number * > JxW
 
std::vector< dim3 > grid_dim
 
std::vector< dim3 > block_dim
 

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

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 93 of file cuda_matrix_free.h.

Constructor & Destructor Documentation

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

Default constructor.

Member Function Documentation

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.

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.

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

Return the Data structure associated with color.

template<int dim, typename Number = double>
template<typename functor >
void CUDAWrappers::MatrixFree< dim, Number >::cell_loop ( const functor &  func,
const CUDAVector< Number > &  src,
CUDAVector< Number > &  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 appplied on each color.

template<int dim, typename Number = double>
void CUDAWrappers::MatrixFree< dim, Number >::free ( )

Free all the memory allocated.

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.

Member Data Documentation

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 218 of file cuda_matrix_free.h.

template<int dim, typename Number = double>
unsigned int CUDAWrappers::MatrixFree< dim, Number >::fe_degree
private

Degree of the finite element used.

Definition at line 222 of file cuda_matrix_free.h.

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 226 of file cuda_matrix_free.h.

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 230 of file cuda_matrix_free.h.

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 234 of file cuda_matrix_free.h.

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 238 of file cuda_matrix_free.h.

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 242 of file cuda_matrix_free.h.

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 247 of file cuda_matrix_free.h.

template<int dim, typename Number = double>
std::vector<unsigned int *> 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 252 of file cuda_matrix_free.h.

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 257 of file cuda_matrix_free.h.

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 262 of file cuda_matrix_free.h.

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 271 of file cuda_matrix_free.h.

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 276 of file cuda_matrix_free.h.


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