Reference documentation for deal.II version 9.5.0
|
#include <deal.II/matrix_free/cuda_matrix_free.h>
Classes | |
struct | AdditionalData |
struct | Data |
Public Types | |
using | jacobian_type = Tensor< 2, dim, Tensor< 1, dim, Number > > |
using | point_type = Point< dim, Number > |
using | CellFilter = FilteredIterator< typename DoFHandler< dim >::active_cell_iterator > |
Public Member Functions | |
MatrixFree () | |
unsigned int | get_padding_length () const |
template<typename IteratorFiltersType > | |
void | reinit (const Mapping< dim > &mapping, const DoFHandler< dim > &dof_handler, const AffineConstraints< Number > &constraints, const Quadrature< 1 > &quad, const IteratorFiltersType &iterator_filter, const AdditionalData &additional_data=AdditionalData()) |
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 &additional_data=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 | initialize_dof_vector (LinearAlgebra::CUDAWrappers::Vector< Number > &vec) const |
void | initialize_dof_vector (LinearAlgebra::distributed::Vector< Number, MemorySpace::Default > &vec) const |
const std::vector< std::vector< CellFilter > > & | get_colored_graph () const |
const std::shared_ptr< const Utilities::MPI::Partitioner > & | get_vector_partitioner () const |
const DoFHandler< dim > & | get_dof_handler () const |
std::size_t | memory_consumption () const |
template<class Archive > | |
void | serialize (Archive &ar, const unsigned int version) |
Subscriptor functionality | |
Classes derived from Subscriptor provide a facility to subscribe to this object. This is mostly used by the SmartPointer class. | |
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 |
Static Public Member Functions | |
static ::ExceptionBase & | ExcInUse (int arg1, std::string arg2, std::string arg3) |
static ::ExceptionBase & | ExcNoSubscriber (std::string arg1, std::string arg2) |
Private Types | |
using | map_value_type = decltype(counter_map)::value_type |
using | map_iterator = decltype(counter_map)::iterator |
Private Member Functions | |
template<typename IteratorFiltersType > | |
void | internal_reinit (const Mapping< dim > &mapping, const DoFHandler< dim > &dof_handler, const AffineConstraints< Number > &constraints, const Quadrature< 1 > &quad, const IteratorFiltersType &iterator_filter, const 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::Default > &src, LinearAlgebra::distributed::Vector< Number, MemorySpace::Default > &dst) const |
template<typename Functor > | |
void | distributed_cell_loop (const Functor &func, const LinearAlgebra::CUDAWrappers::Vector< Number > &src, LinearAlgebra::CUDAWrappers::Vector< Number > &dst) const |
void | check_no_subscribers () const noexcept |
Static Private Attributes | |
static std::mutex | mutex |
Friends | |
class | internal::ReinitHelper< dim, Number > |
Related Symbols | |
(Note that these are not member symbols.) | |
template<int dim, typename Number > | |
DataHost< dim, Number > | copy_mf_data_to_host (const typename::CUDAWrappers::MatrixFree< dim, Number >::Data &data, const UpdateFlags &update_flags) |
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 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::FEEvaluation.
This class traverse the cells in a different order than the usual Triangulation class in deal.II.
Definition at line 94 of file cuda_matrix_free.h.
using CUDAWrappers::MatrixFree< dim, Number >::jacobian_type = Tensor<2, dim, Tensor<1, dim, Number> > |
Definition at line 97 of file cuda_matrix_free.h.
using CUDAWrappers::MatrixFree< dim, Number >::point_type = Point<dim, Number> |
Definition at line 98 of file cuda_matrix_free.h.
using CUDAWrappers::MatrixFree< dim, Number >::CellFilter = FilteredIterator<typename DoFHandler<dim>::active_cell_iterator> |
Definition at line 99 of file cuda_matrix_free.h.
|
privateinherited |
The data type used in counter_map.
Definition at line 230 of file subscriptor.h.
|
privateinherited |
The iterator type used in counter_map.
Definition at line 235 of file subscriptor.h.
CUDAWrappers::MatrixFree< dim, Number >::MatrixFree | ( | ) |
Default constructor.
unsigned int CUDAWrappers::MatrixFree< dim, Number >::get_padding_length | ( | ) | const |
Return the length of the padding.
void CUDAWrappers::MatrixFree< dim, Number >::reinit | ( | const Mapping< dim > & | mapping, |
const DoFHandler< dim > & | dof_handler, | ||
const AffineConstraints< Number > & | constraints, | ||
const Quadrature< 1 > & | quad, | ||
const IteratorFiltersType & | iterator_filter, | ||
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. This function takes an IteratorFilters object (predicate) to loop over a subset of the active cells. When using MPI, the predicate should filter out non locally owned cells.
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() |
||
) |
Same as above using Iterators::LocallyOwnedCell() as predicate.
void CUDAWrappers::MatrixFree< dim, Number >::reinit | ( | const DoFHandler< dim > & | dof_handler, |
const AffineConstraints< Number > & | constraints, | ||
const Quadrature< 1 > & | quad, | ||
const AdditionalData & | additional_data = AdditionalData() |
||
) |
Initializes the data structures. Same as above but using a Q1 mapping.
Data CUDAWrappers::MatrixFree< dim, Number >::get_data | ( | unsigned int | color | ) | const |
Return the Data structure associated with color
.
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
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
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.
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.
void CUDAWrappers::MatrixFree< dim, Number >::initialize_dof_vector | ( | LinearAlgebra::CUDAWrappers::Vector< Number > & | vec | ) | const |
Initialize a serial vector. The size corresponds to the number of degrees of freedom in the DoFHandler object.
void CUDAWrappers::MatrixFree< dim, Number >::initialize_dof_vector | ( | LinearAlgebra::distributed::Vector< Number, MemorySpace::Default > & | vec | ) | const |
Initialize a distributed vector. The local elements correspond to the locally owned degrees of freedom and the ghost elements correspond to the (additional) locally relevant dofs.
const std::vector< std::vector< CellFilter > > & CUDAWrappers::MatrixFree< dim, Number >::get_colored_graph | ( | ) | const |
Return the colored graph of locally owned active cells.
const std::shared_ptr< const Utilities::MPI::Partitioner > & CUDAWrappers::MatrixFree< dim, Number >::get_vector_partitioner | ( | ) | const |
Return the partitioner that represents the locally owned data and the ghost indices where access is needed to for the cell loop. The partitioner is constructed from the locally owned dofs and ghost dofs given by the respective fields. If you want to have specific information about these objects, you can query them with the respective access functions. If you just want to initialize a (parallel) vector, you should usually prefer this data structure as the data exchange information can be reused from one vector to another.
const DoFHandler< dim > & CUDAWrappers::MatrixFree< dim, Number >::get_dof_handler | ( | ) | const |
Return the DoFHandler.
std::size_t CUDAWrappers::MatrixFree< dim, Number >::memory_consumption | ( | ) | const |
Return an approximation of the memory consumption of this class in bytes.
|
private |
Initializes the data structures.
|
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.
|
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.
|
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.
|
inherited |
Subscribes a user of the object by storing the pointer validity
. The subscriber may be identified by text supplied as identifier
.
Definition at line 136 of file subscriptor.cc.
|
inherited |
Unsubscribes a user from the object.
identifier
and the validity
pointer must be the same as the one supplied to subscribe(). Definition at line 156 of file subscriptor.cc.
|
inlineinherited |
Return the present number of subscriptions to this object. This allows to use this class for reference counted lifetime determination where the last one to unsubscribe also deletes the object.
Definition at line 300 of file subscriptor.h.
|
inlineinherited |
List the subscribers to the input stream
.
Definition at line 317 of file subscriptor.h.
|
inherited |
List the subscribers to deallog
.
Definition at line 204 of file subscriptor.cc.
|
inlineinherited |
Read or write the data of this object to or from a stream for the purpose of serialization using the BOOST serialization library.
This function does not actually serialize any of the member variables of this class. The reason is that what this class stores is only who subscribes to this object, but who does so at the time of storing the contents of this object does not necessarily have anything to do with who subscribes to the object when it is restored. Consequently, we do not want to overwrite the subscribers at the time of restoring, and then there is no reason to write the subscribers out in the first place.
Definition at line 309 of file subscriptor.h.
|
privatenoexceptinherited |
Check that there are no objects subscribing to this object. If this check passes then it is safe to destroy the current object. It this check fails then this function will either abort or print an error message to deallog (by using the AssertNothrow mechanism), but will not throw an exception.
Definition at line 53 of file subscriptor.cc.
|
friend |
Definition at line 625 of file cuda_matrix_free.h.
|
related |
Copy data
from the device to the host. update_flags
should be identical to the one used in MatrixFree::AdditionalData.
Definition at line 773 of file cuda_matrix_free.h.
|
private |
Unique ID associated with the object.
Definition at line 483 of file cuda_matrix_free.h.
|
private |
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 490 of file cuda_matrix_free.h.
|
private |
Overlap MPI communications with computation. This requires CUDA-aware MPI and use_coloring must be false.
Definition at line 496 of file cuda_matrix_free.h.
|
private |
Total number of degrees of freedom.
Definition at line 501 of file cuda_matrix_free.h.
|
private |
Degree of the finite element used.
Definition at line 506 of file cuda_matrix_free.h.
|
private |
Number of degrees of freedom per cell.
Definition at line 511 of file cuda_matrix_free.h.
|
private |
Number of constrained degrees of freedom.
Definition at line 516 of file cuda_matrix_free.h.
|
private |
Number of quadrature points per cells.
Definition at line 521 of file cuda_matrix_free.h.
|
private |
Number of colors produced by the graph coloring algorithm.
Definition at line 526 of file cuda_matrix_free.h.
|
private |
Number of cells in each color.
Definition at line 531 of file cuda_matrix_free.h.
|
private |
Vector of Kokkos::View to the quadrature points associated to the cells of each color.
Definition at line 538 of file cuda_matrix_free.h.
|
private |
Map the position in the local vector to the position in the global vector.
Definition at line 546 of file cuda_matrix_free.h.
|
private |
Vector of Kokkos::View of the inverse Jacobian associated to the cells of each color.
Definition at line 554 of file cuda_matrix_free.h.
|
private |
Vector of Kokkos::View to the Jacobian times the weights associated to the cells of each color.
Definition at line 561 of file cuda_matrix_free.h.
|
private |
Kokkos::View to the constrained degrees of freedom.
Definition at line 567 of file cuda_matrix_free.h.
|
private |
Mask deciding where constraints are set on a given cell.
Definition at line 575 of file cuda_matrix_free.h.
|
private |
Values of the shape functions.
Definition at line 580 of file cuda_matrix_free.h.
|
private |
Gradients of the shape functions.
Definition at line 585 of file cuda_matrix_free.h.
|
private |
Gradients of the shape functions for collocation methods.
Definition at line 591 of file cuda_matrix_free.h.
|
private |
Weights used when resolving hanginf nodes.
Definition at line 597 of file cuda_matrix_free.h.
|
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 603 of file cuda_matrix_free.h.
|
private |
Length of the padding (closest power of two larger than or equal to the number of thread).
Definition at line 610 of file cuda_matrix_free.h.
|
private |
Row start of each color.
Definition at line 615 of file cuda_matrix_free.h.
|
private |
Pointer to the DoFHandler associated with the object.
Definition at line 620 of file cuda_matrix_free.h.
|
private |
Colored graphed of locally owned active cells.
Definition at line 625 of file cuda_matrix_free.h.
|
mutableprivateinherited |
Store the number of objects which subscribed to this object. Initially, this number is zero, and upon destruction it shall be zero again (i.e. all objects which subscribed should have unsubscribed again).
The creator (and owner) of an object is counted in the map below if HE manages to supply identification.
We use the mutable
keyword in order to allow subscription to constant objects also.
This counter may be read from and written to concurrently in multithreaded code: hence we use the std::atomic
class template.
Definition at line 219 of file subscriptor.h.
|
mutableprivateinherited |
In this map, we count subscriptions for each different identification string supplied to subscribe().
Definition at line 225 of file subscriptor.h.
|
mutableprivateinherited |
In this vector, we store pointers to the validity bool in the SmartPointer objects that subscribe to this class.
Definition at line 241 of file subscriptor.h.
|
mutableprivateinherited |
Pointer to the typeinfo object of this object, from which we can later deduce the class name. Since this information on the derived class is neither available in the destructor, nor in the constructor, we obtain it in between and store it here.
Definition at line 249 of file subscriptor.h.
|
staticprivateinherited |
A mutex used to ensure data consistency when printing out the list of subscribers.
Definition at line 271 of file subscriptor.h.