CUDAWrappers::MatrixFree< dim, Number >::AdditionalData Struct Reference

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

Public Member Functions
	AdditionalData (const UpdateFlags mapping_update_flags=update_gradients|update_JxW_values|update_quadrature_points, const bool use_coloring=false, const bool overlap_communication_computation=false)

Public Attributes
UpdateFlags	mapping_update_flags
bool	use_coloring
bool	overlap_communication_computation

Detailed Description

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

Standardized data struct to pipe additional data to MatrixFree.

Definition at line 105 of file cuda_matrix_free.h.

Constructor & Destructor Documentation

AdditionalData()

template<int dim, typename Number = double>

CUDAWrappers::MatrixFree< dim, Number >::AdditionalData::AdditionalData ( const UpdateFlags  mapping_update_flags = update_gradients | update_JxW_values | update_quadrature_points, const bool  use_coloring = false, const bool  overlap_communication_computation = false  )

inline

Constructor.

Definition at line 110 of file cuda_matrix_free.h.

Member Data Documentation

mapping_update_flags

template<int dim, typename Number = double>

UpdateFlags CUDAWrappers::MatrixFree< dim, Number >::AdditionalData::mapping_update_flags

This flag is used to determine which quantities should be cached. This class can cache data needed for gradient computations (inverse Jacobians), Jacobian determinants (JxW), quadrature points as well as data for Hessians (derivative of Jacobians). By default, only data for gradients and Jacobian determinants times quadrature weights, JxW, are cached. If quadrature points of second derivatives are needed, they must be specified by this field.

Definition at line 140 of file cuda_matrix_free.h.

use_coloring

template<int dim, typename Number = double>

bool CUDAWrappers::MatrixFree< dim, Number >::AdditionalData::use_coloring

If true, use graph coloring. Otherwise, use atomic operations. Graph coloring ensures bitwise reproducibility but is slower on Pascal and newer architectures.

Definition at line 147 of file cuda_matrix_free.h.

overlap_communication_computation

template<int dim, typename Number = double>

bool CUDAWrappers::MatrixFree< dim, Number >::AdditionalData::overlap_communication_computation

Overlap MPI communications with computation. This requires CUDA-aware MPI and use_coloring must be false.

Definition at line 153 of file cuda_matrix_free.h.

---

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

• include/deal.II/matrix_free/cuda_matrix_free.h