Reference documentation for deal.II version Git 25f228b 2018-07-21 16:44:11 +0200
TrilinosWrappers::SparsityPattern Class Reference

#include <deal.II/lac/trilinos_sparsity_pattern.h>

Inheritance diagram for TrilinosWrappers::SparsityPattern:
[legend]

## Public Types

using size_type = ::types::global_dof_index

using const_iterator = SparsityPatternIterators::Iterator

## Public Member Functions

Basic constructors and initialization.
SparsityPattern ()

SparsityPattern (const size_type m, const size_type n, const size_type n_entries_per_row=0)

SparsityPattern (const size_type m, const size_type n, const std::vector< size_type > &n_entries_per_row)

SparsityPattern (SparsityPattern &&other) noexcept

SparsityPattern (const SparsityPattern &input_sparsity_pattern)

virtual ~SparsityPattern () override=default

void reinit (const size_type m, const size_type n, const size_type n_entries_per_row=0)

void reinit (const size_type m, const size_type n, const std::vector< size_type > &n_entries_per_row)

void copy_from (const SparsityPattern &input_sparsity_pattern)

template<typename SparsityPatternType >
void copy_from (const SparsityPatternType &nontrilinos_sparsity_pattern)

SparsityPatternoperator= (const SparsityPattern &input_sparsity_pattern)

void clear ()

void compress ()

Constructors and initialization using an Epetra_Map description
SparsityPattern (const Epetra_Map &parallel_partitioning, const size_type n_entries_per_row=0)

SparsityPattern (const Epetra_Map &parallel_partitioning, const std::vector< size_type > &n_entries_per_row)

SparsityPattern (const Epetra_Map &row_parallel_partitioning, const Epetra_Map &col_parallel_partitioning, const size_type n_entries_per_row=0)

SparsityPattern (const Epetra_Map &row_parallel_partitioning, const Epetra_Map &col_parallel_partitioning, const std::vector< size_type > &n_entries_per_row)

void reinit (const Epetra_Map &parallel_partitioning, const size_type n_entries_per_row=0)

void reinit (const Epetra_Map &parallel_partitioning, const std::vector< size_type > &n_entries_per_row)

void reinit (const Epetra_Map &row_parallel_partitioning, const Epetra_Map &col_parallel_partitioning, const size_type n_entries_per_row=0)

void reinit (const Epetra_Map &row_parallel_partitioning, const Epetra_Map &col_parallel_partitioning, const std::vector< size_type > &n_entries_per_row)

template<typename SparsityPatternType >
void reinit (const Epetra_Map &row_parallel_partitioning, const Epetra_Map &col_parallel_partitioning, const SparsityPatternType &nontrilinos_sparsity_pattern, const bool exchange_data=false)

template<typename SparsityPatternType >
void reinit (const Epetra_Map &parallel_partitioning, const SparsityPatternType &nontrilinos_sparsity_pattern, const bool exchange_data=false)

Constructors and initialization using an IndexSet description
SparsityPattern (const IndexSet &parallel_partitioning, const MPI_Comm &communicator=MPI_COMM_WORLD, const size_type n_entries_per_row=0)

SparsityPattern (const IndexSet &parallel_partitioning, const MPI_Comm &communicator, const std::vector< size_type > &n_entries_per_row)

SparsityPattern (const IndexSet &row_parallel_partitioning, const IndexSet &col_parallel_partitioning, const MPI_Comm &communicator=MPI_COMM_WORLD, const size_type n_entries_per_row=0)

SparsityPattern (const IndexSet &row_parallel_partitioning, const IndexSet &col_parallel_partitioning, const MPI_Comm &communicator, const std::vector< size_type > &n_entries_per_row)

SparsityPattern (const IndexSet &row_parallel_partitioning, const IndexSet &col_parallel_partitioning, const IndexSet &writable_rows, const MPI_Comm &communicator=MPI_COMM_WORLD, const size_type n_entries_per_row=0)

void reinit (const IndexSet &parallel_partitioning, const MPI_Comm &communicator=MPI_COMM_WORLD, const size_type n_entries_per_row=0)

void reinit (const IndexSet &parallel_partitioning, const MPI_Comm &communicator, const std::vector< size_type > &n_entries_per_row)

void reinit (const IndexSet &row_parallel_partitioning, const IndexSet &col_parallel_partitioning, const MPI_Comm &communicator=MPI_COMM_WORLD, const size_type n_entries_per_row=0)

void reinit (const IndexSet &row_parallel_partitioning, const IndexSet &col_parallel_partitioning, const IndexSet &writeable_rows, const MPI_Comm &communicator=MPI_COMM_WORLD, const size_type n_entries_per_row=0)

void reinit (const IndexSet &row_parallel_partitioning, const IndexSet &col_parallel_partitioning, const MPI_Comm &communicator, const std::vector< size_type > &n_entries_per_row)

template<typename SparsityPatternType >
void reinit (const IndexSet &row_parallel_partitioning, const IndexSet &col_parallel_partitioning, const SparsityPatternType &nontrilinos_sparsity_pattern, const MPI_Comm &communicator=MPI_COMM_WORLD, const bool exchange_data=false)

template<typename SparsityPatternType >
void reinit (const IndexSet &parallel_partitioning, const SparsityPatternType &nontrilinos_sparsity_pattern, const MPI_Comm &communicator=MPI_COMM_WORLD, const bool exchange_data=false)

Information on the sparsity pattern
bool is_compressed () const

unsigned int max_entries_per_row () const

size_type n_rows () const

size_type n_cols () const

unsigned int local_size () const

std::pair< size_type, size_typelocal_range () const

bool in_local_range (const size_type index) const

size_type n_nonzero_elements () const

size_type row_length (const size_type row) const

size_type bandwidth () const

bool empty () const

bool exists (const size_type i, const size_type j) const

std::size_t memory_consumption () const

void add (const size_type i, const size_type j)

template<typename ForwardIterator >
void add_entries (const size_type row, ForwardIterator begin, ForwardIterator end, const bool indices_are_sorted=false)

Access of underlying Trilinos data
const Epetra_FECrsGraph & trilinos_sparsity_pattern () const

const Epetra_Map & domain_partitioner () const

const Epetra_Map & range_partitioner () const

const Epetra_Map & row_partitioner () const

const Epetra_Map & col_partitioner () const

const Epetra_Comm & trilinos_communicator () const

MPI_Comm get_mpi_communicator () const

Partitioners
IndexSet locally_owned_domain_indices () const

IndexSet locally_owned_range_indices () const

Iterators
const_iterator begin () const

const_iterator end () const

const_iterator begin (const size_type r) const

const_iterator end (const size_type r) const

Input/Output
void write_ascii ()

void print (std::ostream &out, const bool write_extended_trilinos_info=false) const

void print_gnuplot (std::ostream &out) 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

void list_subscribers () const

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

## Static Public Member Functions

static::ExceptionBase & ExcTrilinosError (int arg1)

static::ExceptionBase & ExcInvalidIndex (size_type arg1, size_type arg2)

static::ExceptionBase & ExcSourceEqualsDestination ()

static::ExceptionBase & ExcAccessToNonLocalElement (size_type arg1, size_type arg2, size_type arg3, size_type arg4)

static::ExceptionBase & ExcAccessToNonPresentElement (size_type arg1, size_type arg2)

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)

## Private Attributes

std::unique_ptr< Epetra_Map > column_space_map

std::unique_ptr< Epetra_FECrsGraph > graph

std::unique_ptr< Epetra_CrsGraph > nonlocal_graph

## Friends

class TrilinosWrappers::SparseMatrix

class SparsityPatternIterators::Accessor

class SparsityPatternIterators::Iterator

## Detailed Description

This class implements a wrapper class to use the Trilinos distributed sparsity pattern class Epetra_FECrsGraph. This class is designed to be used for construction of parallel Trilinos matrices. The functionality of this class is modeled after the existing sparsity pattern classes, with the difference that this class can work fully in parallel according to a partitioning of the sparsity pattern rows.

This class has many similarities to the DynamicSparsityPattern, since it can dynamically add elements to the pattern without any memory being previously reserved for it. However, it also has a method SparsityPattern::compress(), that finalizes the pattern and enables its use with Trilinos sparse matrices.

Definition at line 275 of file trilinos_sparsity_pattern.h.

## Member Typedef Documentation

Declare type for container size.

Definition at line 281 of file trilinos_sparsity_pattern.h.

Declare an alias for the iterator class.

Definition at line 286 of file trilinos_sparsity_pattern.h.

## Constructor & Destructor Documentation

 SparsityPattern::SparsityPattern ( )

Default constructor. Generates an empty (zero-size) sparsity pattern.

Definition at line 83 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const size_type m, const size_type n, const size_type n_entries_per_row = 0 )

Generate a sparsity pattern that is completely stored locally, having $$m$$ rows and $$n$$ columns. The resulting matrix will be completely stored locally, too.

It is possible to specify the number of columns entries per row using the optional n_entries_per_row argument. However, this value does not need to be accurate or even given at all, since one does usually not have this kind of information before building the sparsity pattern (the usual case when the function DoFTools::make_sparsity_pattern() is called). The entries are allocated dynamically in a similar manner as for the deal.II DynamicSparsityPattern classes. However, a good estimate will reduce the setup time of the sparsity pattern.

Definition at line 133 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const size_type m, const size_type n, const std::vector< size_type > & n_entries_per_row )

Generate a sparsity pattern that is completely stored locally, having $$m$$ rows and $$n$$ columns. The resulting matrix will be completely stored locally, too.

The vector n_entries_per_row specifies the number of entries in each row (an information usually not available, though).

Definition at line 142 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( SparsityPattern && other )
noexcept

Move constructor. Create a new sparse matrix by stealing the internal data.

Definition at line 152 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const SparsityPattern & input_sparsity_pattern )

Copy constructor. Sets the calling sparsity pattern to be the same as the input sparsity pattern.

Definition at line 162 of file trilinos_sparsity_pattern.cc.

 virtual TrilinosWrappers::SparsityPattern::~SparsityPattern ( )
overridevirtualdefault

Destructor. Made virtual so that one can use pointers to this class.

 SparsityPattern::SparsityPattern ( const Epetra_Map & parallel_partitioning, const size_type n_entries_per_row = 0 )

Constructor for a square sparsity pattern using an Epetra_map for the description of the parallel partitioning. Moreover, the number of nonzero entries in the rows of the sparsity pattern can be specified. Note that this number does not need to be exact, and it is allowed that the actual sparsity structure has more nonzero entries than specified in the constructor (the usual case when the function DoFTools::make_sparsity_pattern() is called). However it is still advantageous to provide good estimates here since a good value will avoid repeated allocation of memory, which considerably increases the performance when creating the sparsity pattern.

Deprecated:
Use the respective method with IndexSet argument instead.

Definition at line 97 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const Epetra_Map & parallel_partitioning, const std::vector< size_type > & n_entries_per_row )

Same as before, but now use the exact number of nonzeros in each m row. Since we know the number of elements in the sparsity pattern exactly in this case, we can already allocate the right amount of memory, which makes the creation process by the respective SparsityPattern::reinit call considerably faster. However, this is a rather unusual situation, since knowing the number of entries in each row is usually connected to knowing the indices of nonzero entries, which the sparsity pattern is designed to describe.

Deprecated:
Use the respective method with IndexSet argument instead.

Definition at line 105 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const Epetra_Map & row_parallel_partitioning, const Epetra_Map & col_parallel_partitioning, const size_type n_entries_per_row = 0 )

This constructor is similar to the one above, but it now takes two different Epetra maps for rows and columns. This interface is meant to be used for generating rectangular sparsity pattern, where one map describes the parallel partitioning of the dofs associated with the sparsity pattern rows and the other one of the sparsity pattern columns. Note that there is no real parallelism along the columns – the processor that owns a certain row always owns all the column elements, no matter how far they might be spread out. The second Epetra_Map is only used to specify the number of columns and for specifying the correct domain space when performing matrix-vector products with vectors based on the same column map.

The number of columns entries per row is specified as the maximum number of entries argument.

Deprecated:
Use the respective method with IndexSet argument instead.

Definition at line 114 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const Epetra_Map & row_parallel_partitioning, const Epetra_Map & col_parallel_partitioning, const std::vector< size_type > & n_entries_per_row )

This constructor is similar to the one above, but it now takes two different Epetra maps for rows and columns. This interface is meant to be used for generating rectangular matrices, where one map specifies the parallel distribution of rows and the second one specifies the distribution of degrees of freedom associated with matrix columns. This second map is however not used for the distribution of the columns themselves – rather, all column elements of a row are stored on the same processor. The vector n_entries_per_row specifies the number of entries in each row of the newly generated matrix.

Deprecated:
Use the respective method with IndexSet argument instead.

Definition at line 123 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const IndexSet & parallel_partitioning, const MPI_Comm & communicator = MPI_COMM_WORLD, const size_type n_entries_per_row = 0 )

Constructor for a square sparsity pattern using an IndexSet and an MPI communicator for the description of the parallel partitioning. Moreover, the number of nonzero entries in the rows of the sparsity pattern can be specified. Note that this number does not need to be exact, and it is even allowed that the actual sparsity structure has more nonzero entries than specified in the constructor. However it is still advantageous to provide good estimates here since a good value will avoid repeated allocation of memory, which considerably increases the performance when creating the sparsity pattern.

Definition at line 178 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const IndexSet & parallel_partitioning, const MPI_Comm & communicator, const std::vector< size_type > & n_entries_per_row )

Same as before, but now use the exact number of nonzeros in each m row. Since we know the number of elements in the sparsity pattern exactly in this case, we can already allocate the right amount of memory, which makes the creation process by the respective SparsityPattern::reinit call considerably faster. However, this is a rather unusual situation, since knowing the number of entries in each row is usually connected to knowing the indices of nonzero entries, which the sparsity pattern is designed to describe.

Definition at line 190 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const IndexSet & row_parallel_partitioning, const IndexSet & col_parallel_partitioning, const MPI_Comm & communicator = MPI_COMM_WORLD, const size_type n_entries_per_row = 0 )

This constructor is similar to the one above, but it now takes two different index sets to describe the parallel partitioning of rows and columns. This interface is meant to be used for generating rectangular sparsity pattern. Note that there is no real parallelism along the columns – the processor that owns a certain row always owns all the column elements, no matter how far they might be spread out. The second Epetra_Map is only used to specify the number of columns and for internal arrangements when doing matrix-vector products with vectors based on that column map.

The number of columns entries per row is specified as the maximum number of entries argument.

Definition at line 203 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const IndexSet & row_parallel_partitioning, const IndexSet & col_parallel_partitioning, const MPI_Comm & communicator, const std::vector< size_type > & n_entries_per_row )

This constructor is similar to the one above, but it now takes two different index sets for rows and columns. This interface is meant to be used for generating rectangular matrices, where one map specifies the parallel distribution of rows and the second one specifies the distribution of degrees of freedom associated with matrix columns. This second map is however not used for the distribution of the columns themselves – rather, all column elements of a row are stored on the same processor. The vector n_entries_per_row specifies the number of entries in each row of the newly generated matrix.

Definition at line 216 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::SparsityPattern ( const IndexSet & row_parallel_partitioning, const IndexSet & col_parallel_partitioning, const IndexSet & writable_rows, const MPI_Comm & communicator = MPI_COMM_WORLD, const size_type n_entries_per_row = 0 )

This constructor constructs general sparsity patterns, possible non- square ones. Constructing a sparsity pattern this way allows the user to explicitly specify the rows into which we are going to add elements. This set is required to be a superset of the first index set row_parallel_partitioning that includes also rows that are owned by another processor (ghost rows). Note that elements can only be added to rows specified by writable_rows.

This method is beneficial when the rows to which a processor is going to write can be determined before actually inserting elements into the matrix. For the typical parallel::distributed::Triangulation class used in deal.II, we know that a processor only will add row elements for what we call the locally relevant dofs (see DoFTools::extract_locally_relevant_dofs). The other constructors methods use general Trilinos facilities that allow to add elements to arbitrary rows (as done by all the other reinit functions). However, this flexibility come at a cost, the most prominent being that adding elements into the same matrix from multiple threads in shared memory is not safe whenever MPI is used. For these settings, the current method is the one to choose: It will store the off-processor data as an additional sparsity pattern (that is then passed to the Trilinos matrix via the reinit mehtod) which can be organized in such a way that thread-safety can be ensured (as long as the user makes sure to never write into the same matrix row simultaneously, of course).

Definition at line 230 of file trilinos_sparsity_pattern.cc.

## Member Function Documentation

 void SparsityPattern::reinit ( const size_type m, const size_type n, const size_type n_entries_per_row = 0 )

Initialize a sparsity pattern that is completely stored locally, having $$m$$ rows and $$n$$ columns. The resulting matrix will be completely stored locally.

The number of columns entries per row is specified as the maximum number of entries argument. This does not need to be an accurate number since the entries are allocated dynamically in a similar manner as for the deal.II DynamicSparsityPattern classes, but a good estimate will reduce the setup time of the sparsity pattern.

Definition at line 246 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::reinit ( const size_type m, const size_type n, const std::vector< size_type > & n_entries_per_row )

Initialize a sparsity pattern that is completely stored locally, having $$m$$ rows and $$n$$ columns. The resulting matrix will be completely stored locally.

The vector n_entries_per_row specifies the number of entries in each row.

Definition at line 259 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::copy_from ( const SparsityPattern & input_sparsity_pattern )

Copy function. Sets the calling sparsity pattern to be the same as the input sparsity pattern.

Definition at line 737 of file trilinos_sparsity_pattern.cc.

template<typename SparsityPatternType >
 void SparsityPattern::copy_from ( const SparsityPatternType & nontrilinos_sparsity_pattern )

Copy function from one of the deal.II sparsity patterns. If used in parallel, this function uses an ad-hoc partitioning of the rows and columns.

Definition at line 753 of file trilinos_sparsity_pattern.cc.

 SparsityPattern & SparsityPattern::operator= ( const SparsityPattern & input_sparsity_pattern )

Copy operator. This operation is only allowed for empty objects, to avoid potentially very costly operations automatically synthesized by the compiler. Use copy_from() instead if you know that you really want to copy a sparsity pattern with non-trivial content.

Definition at line 728 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::clear ( )

Release all memory and return to a state just like after having called the default constructor.

This is a collective operation that needs to be called on all processors in order to avoid a dead lock.

Definition at line 769 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::compress ( )

In analogy to our own SparsityPattern class, this function compresses the sparsity pattern and allows the resulting pattern to be used for actually generating a (Trilinos-based) matrix. This function also exchanges non-local data that might have accumulated during the addition of new elements. This function must therefore be called once the structure is fixed. This is a collective operation, i.e., it needs to be run on all processors when used in parallel.

Definition at line 790 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::reinit ( const Epetra_Map & parallel_partitioning, const size_type n_entries_per_row = 0 )

Reinitialization function for generating a square sparsity pattern using an Epetra_Map for the description of the parallel partitioning and the number of nonzero entries in the rows of the sparsity pattern. Note that this number does not need to be exact, and it is even allowed that the actual sparsity structure has more nonzero entries than specified in the constructor. However it is still advantageous to provide good estimates here since this will considerably increase the performance when creating the sparsity pattern.

This function does not create any entries by itself, but provides the correct data structures that can be used by the respective add() function.

Deprecated:
Use the respective method with IndexSet argument instead.

Definition at line 478 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::reinit ( const Epetra_Map & parallel_partitioning, const std::vector< size_type > & n_entries_per_row )

Same as before, but now use the exact number of nonzeros in each m row. Since we know the number of elements in the sparsity pattern exactly in this case, we can already allocate the right amount of memory, which makes process of adding entries to the sparsity pattern considerably faster. However, this is a rather unusual situation, since knowing the number of entries in each row is usually connected to knowing the indices of nonzero entries, which the sparsity pattern is designed to describe.

Deprecated:
Use the respective method with IndexSet argument instead.

Definition at line 507 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::reinit ( const Epetra_Map & row_parallel_partitioning, const Epetra_Map & col_parallel_partitioning, const size_type n_entries_per_row = 0 )

This reinit function is similar to the one above, but it now takes two different Epetra maps for rows and columns. This interface is meant to be used for generating rectangular sparsity pattern, where one map describes the parallel partitioning of the dofs associated with the sparsity pattern rows and the other one of the sparsity pattern columns. Note that there is no real parallelism along the columns – the processor that owns a certain row always owns all the column elements, no matter how far they might be spread out. The second Epetra_Map is only used to specify the number of columns and for internal arrangements when doing matrix-vector products with vectors based on that column map.

The number of columns entries per row is specified by the argument n_entries_per_row.

Deprecated:
Use the respective method with IndexSet argument instead.

Definition at line 492 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::reinit ( const Epetra_Map & row_parallel_partitioning, const Epetra_Map & col_parallel_partitioning, const std::vector< size_type > & n_entries_per_row )

This reinit function is similar to the one above, but it now takes two different Epetra maps for rows and columns. This interface is meant to be used for generating rectangular matrices, where one map specifies the parallel distribution of rows and the second one specifies the distribution of degrees of freedom associated with matrix columns. This second map is however not used for the distribution of the columns themselves – rather, all column elements of a row are stored on the same processor. The vector n_entries_per_row specifies the number of entries in each row of the newly generated matrix.

Deprecated:
Use the respective method with IndexSet argument instead.

Definition at line 521 of file trilinos_sparsity_pattern.cc.

template<typename SparsityPatternType >
 void SparsityPattern::reinit ( const Epetra_Map & row_parallel_partitioning, const Epetra_Map & col_parallel_partitioning, const SparsityPatternType & nontrilinos_sparsity_pattern, const bool exchange_data = false )

Reinit function. Takes one of the deal.II sparsity patterns and a parallel partitioning of the rows and columns for initializing the current Trilinos sparsity pattern. The optional argument exchange_data can be used for reinitialization with a sparsity pattern that is not fully constructed. This feature is only implemented for input sparsity patterns of type DynamicSparsityPattern.

Deprecated:
Use the respective method with IndexSet argument instead.

Definition at line 709 of file trilinos_sparsity_pattern.cc.

template<typename SparsityPatternType >
 void SparsityPattern::reinit ( const Epetra_Map & parallel_partitioning, const SparsityPatternType & nontrilinos_sparsity_pattern, const bool exchange_data = false )

Reinit function. Takes one of the deal.II sparsity patterns and a parallel partitioning of the rows and columns for initializing the current Trilinos sparsity pattern. The optional argument exchange_data can be used for reinitialization with a sparsity pattern that is not fully constructed. This feature is only implemented for input sparsity patterns of type DynamicSparsityPattern.

Deprecated:
Use the respective method with IndexSet argument instead.

Definition at line 692 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::reinit ( const IndexSet & parallel_partitioning, const MPI_Comm & communicator = MPI_COMM_WORLD, const size_type n_entries_per_row = 0 )

Reinitialization function for generating a square sparsity pattern using an IndexSet and an MPI communicator for the description of the parallel partitioning and the number of nonzero entries in the rows of the sparsity pattern. Note that this number does not need to be exact, and it is even allowed that the actual sparsity structure has more nonzero entries than specified in the constructor. However it is still advantageous to provide good estimates here since this will considerably increase the performance when creating the sparsity pattern.

This function does not create any entries by itself, but provides the correct data structures that can be used by the respective add() function.

Definition at line 536 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::reinit ( const IndexSet & parallel_partitioning, const MPI_Comm & communicator, const std::vector< size_type > & n_entries_per_row )

Same as before, but now use the exact number of nonzeros in each m row. Since we know the number of elements in the sparsity pattern exactly in this case, we can already allocate the right amount of memory, which makes process of adding entries to the sparsity pattern considerably faster. However, this is a rather unusual situation, since knowing the number of entries in each row is usually connected to knowing the indices of nonzero entries, which the sparsity pattern is designed to describe.

Definition at line 549 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::reinit ( const IndexSet & row_parallel_partitioning, const IndexSet & col_parallel_partitioning, const MPI_Comm & communicator = MPI_COMM_WORLD, const size_type n_entries_per_row = 0 )

This reinit function is similar to the one above, but it now takes two different index sets for rows and columns. This interface is meant to be used for generating rectangular sparsity pattern, where one index set describes the parallel partitioning of the dofs associated with the sparsity pattern rows and the other one of the sparsity pattern columns. Note that there is no real parallelism along the columns – the processor that owns a certain row always owns all the column elements, no matter how far they might be spread out. The second IndexSet is only used to specify the number of columns and for internal arrangements when doing matrix-vector products with vectors based on an EpetraMap based on that IndexSet.

The number of columns entries per row is specified by the argument n_entries_per_row.

Definition at line 562 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::reinit ( const IndexSet & row_parallel_partitioning, const IndexSet & col_parallel_partitioning, const IndexSet & writeable_rows, const MPI_Comm & communicator = MPI_COMM_WORLD, const size_type n_entries_per_row = 0 )

This reinit function is used to specify general matrices, possibly non- square ones. In addition to the arguments of the other reinit method above, it allows the user to explicitly specify the rows into which we are going to add elements. This set is a superset of the first index set row_parallel_partitioning that includes also rows that are owned by another processor (ghost rows).

This method is beneficial when the rows to which a processor is going to write can be determined before actually inserting elements into the matrix. For the typical parallel::distributed::Triangulation class used in deal.II, we know that a processor only will add row elements for what we call the locally relevant dofs (see DoFTools::extract_locally_relevant_dofs). Trilinos matrices allow to add elements to arbitrary rows (as done by all the other reinit functions) and this is what all the other reinit methods do, too. However, this flexibility come at a cost, the most prominent being that adding elements into the same matrix from multiple threads in shared memory is not safe whenever MPI is used. For these settings, the current method is the one to choose: It will store the off-processor data as an additional sparsity pattern (that is then passed to the Trilinos matrix via the reinit method) which can be organized in such a way that thread-safety can be ensured (as long as the user makes sure to never write into the same matrix row simultaneously, of course).

Definition at line 602 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::reinit ( const IndexSet & row_parallel_partitioning, const IndexSet & col_parallel_partitioning, const MPI_Comm & communicator, const std::vector< size_type > & n_entries_per_row )

Same as before, but now using a vector n_entries_per_row for specifying the number of entries in each row of the sparsity pattern.

Definition at line 582 of file trilinos_sparsity_pattern.cc.

template<typename SparsityPatternType >
 void SparsityPattern::reinit ( const IndexSet & row_parallel_partitioning, const IndexSet & col_parallel_partitioning, const SparsityPatternType & nontrilinos_sparsity_pattern, const MPI_Comm & communicator = MPI_COMM_WORLD, const bool exchange_data = false )

Reinit function. Takes one of the deal.II sparsity patterns and the parallel partitioning of the rows and columns specified by two index sets and a parallel communicator for initializing the current Trilinos sparsity pattern. The optional argument exchange_data can be used for reinitialization with a sparsity pattern that is not fully constructed. This feature is only implemented for input sparsity patterns of type DynamicSparsityPattern.

Definition at line 647 of file trilinos_sparsity_pattern.cc.

template<typename SparsityPatternType >
 void SparsityPattern::reinit ( const IndexSet & parallel_partitioning, const SparsityPatternType & nontrilinos_sparsity_pattern, const MPI_Comm & communicator = MPI_COMM_WORLD, const bool exchange_data = false )

Reinit function. Takes one of the deal.II sparsity patterns and a parallel partitioning of the rows and columns for initializing the current Trilinos sparsity pattern. The optional argument exchange_data can be used for reinitialization with a sparsity pattern that is not fully constructed. This feature is only implemented for input sparsity patterns of type DynamicSparsityPattern.

Definition at line 671 of file trilinos_sparsity_pattern.cc.

 bool TrilinosWrappers::SparsityPattern::is_compressed ( ) const

Return the state of the sparsity pattern, i.e., whether compress() needs to be called after an operation requiring data exchange.

 unsigned int SparsityPattern::max_entries_per_row ( ) const

Return the maximum number of entries per row on the current processor.

Definition at line 1011 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::size_type SparsityPattern::n_rows ( ) const

Return the number of rows in this sparsity pattern.

Definition at line 956 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::size_type SparsityPattern::n_cols ( ) const

Return the number of columns in this sparsity pattern.

Definition at line 965 of file trilinos_sparsity_pattern.cc.

 unsigned int SparsityPattern::local_size ( ) const

Return the local dimension of the sparsity pattern, i.e. the number of rows stored on the present MPI process. In the sequential case, this number is the same as n_rows(), but for parallel matrices it may be smaller.

To figure out which elements exactly are stored locally, use local_range().

Definition at line 979 of file trilinos_sparsity_pattern.cc.

 std::pair< SparsityPattern::size_type, SparsityPattern::size_type > SparsityPattern::local_range ( ) const

Return a pair of indices indicating which rows of this sparsity pattern are stored locally. The first number is the index of the first row stored, the second the index of the one past the last one that is stored locally. If this is a sequential matrix, then the result will be the pair (0,n_rows()), otherwise it will be a pair (i,i+n), where n=local_size().

Definition at line 989 of file trilinos_sparsity_pattern.cc.

 bool TrilinosWrappers::SparsityPattern::in_local_range ( const size_type index ) const

Return whether index is in the local range or not, see also local_range().

 SparsityPattern::size_type SparsityPattern::n_nonzero_elements ( ) const

Return the number of nonzero elements of this sparsity pattern.

Definition at line 1001 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::size_type SparsityPattern::row_length ( const size_type row ) const

Return the number of entries in the given row.

In a parallel context, the row in question may of course not be stored on the current processor, and in that case it is not possible to query the number of entries in it. In that case, the returned value is static_cast<size_type>(-1).

Definition at line 1021 of file trilinos_sparsity_pattern.cc.

 SparsityPattern::size_type SparsityPattern::bandwidth ( ) const

Compute the bandwidth of the matrix represented by this structure. The bandwidth is the maximum of $$|i-j|$$ for which the index pair $$(i,j)$$ represents a nonzero entry of the matrix. Consequently, the maximum bandwidth a $$n\times m$$ matrix can have is $$\max\{n-1,m-1\}$$.

Definition at line 929 of file trilinos_sparsity_pattern.cc.

 bool TrilinosWrappers::SparsityPattern::empty ( ) const

Return whether the object is empty. It is empty if no memory is allocated, which is the same as when both dimensions are zero.

 bool SparsityPattern::exists ( const size_type i, const size_type j ) const

Return whether the index (i,j) exists in the sparsity pattern (i.e., it may be non-zero) or not.

Definition at line 836 of file trilinos_sparsity_pattern.cc.

 std::size_t SparsityPattern::memory_consumption ( ) const

Determine an estimate for the memory consumption (in bytes) of this object. Currently not implemented for this class.

Definition at line 1168 of file trilinos_sparsity_pattern.cc.

 void TrilinosWrappers::SparsityPattern::add ( const size_type i, const size_type j )

Add the element (i,j) to the sparsity pattern.

template<typename ForwardIterator >
 void TrilinosWrappers::SparsityPattern::add_entries ( const size_type row, ForwardIterator begin, ForwardIterator end, const bool indices_are_sorted = false )

Add several elements in one row to the sparsity pattern.

 const Epetra_FECrsGraph& TrilinosWrappers::SparsityPattern::trilinos_sparsity_pattern ( ) const

Return a const reference to the underlying Trilinos Epetra_CrsGraph data that stores the sparsity pattern.

 const Epetra_Map & SparsityPattern::domain_partitioner ( ) const

Return a const reference to the underlying Trilinos Epetra_Map that sets the parallel partitioning of the domain space of this sparsity pattern, i.e., the partitioning of the vectors matrices based on this sparsity pattern are multiplied with.

Deprecated:

Definition at line 1041 of file trilinos_sparsity_pattern.cc.

 const Epetra_Map & SparsityPattern::range_partitioner ( ) const

Return a const reference to the underlying Trilinos Epetra_Map that sets the partitioning of the range space of this sparsity pattern, i.e., the partitioning of the vectors that are result from matrix- vector products.

Deprecated:

Definition at line 1049 of file trilinos_sparsity_pattern.cc.

 const Epetra_Map & SparsityPattern::row_partitioner ( ) const

Return a const reference to the underlying Trilinos Epetra_Map that sets the partitioning of the sparsity pattern rows. Equal to the partitioning of the range.

Deprecated:

Definition at line 1057 of file trilinos_sparsity_pattern.cc.

 const Epetra_Map & SparsityPattern::col_partitioner ( ) const

Return a const reference to the underlying Trilinos Epetra_Map that sets the partitioning of the sparsity pattern columns. This is in general not equal to the partitioner Epetra_Map for the domain because of overlap in the matrix.

Deprecated:
Usually not necessary. If desired, access via the Epetra_FECrsGraph.

Definition at line 1065 of file trilinos_sparsity_pattern.cc.

 const Epetra_Comm & SparsityPattern::trilinos_communicator ( ) const

Return a const reference to the communicator used for this object.

Deprecated:

Definition at line 1073 of file trilinos_sparsity_pattern.cc.

 MPI_Comm SparsityPattern::get_mpi_communicator ( ) const

Return the MPI communicator object in use with this matrix.

Definition at line 1081 of file trilinos_sparsity_pattern.cc.

 IndexSet TrilinosWrappers::SparsityPattern::locally_owned_domain_indices ( ) const

Return the partitioning of the domain space of this pattern, i.e., the partitioning of the vectors a matrix based on this sparsity pattern has to be multiplied with.

 IndexSet TrilinosWrappers::SparsityPattern::locally_owned_range_indices ( ) const

Return the partitioning of the range space of this pattern, i.e., the partitioning of the vectors that are the result from matrix-vector products from a matrix based on this pattern.

 const_iterator TrilinosWrappers::SparsityPattern::begin ( ) const

Iterator starting at the first entry.

 const_iterator TrilinosWrappers::SparsityPattern::end ( ) const

Final iterator.

 const_iterator TrilinosWrappers::SparsityPattern::begin ( const size_type r ) const

Iterator starting at the first entry of row r.

Note that if the given row is empty, i.e. does not contain any nonzero entries, then the iterator returned by this function equals end(r). Note also that the iterator may not be dereferencable in that case.

 const_iterator TrilinosWrappers::SparsityPattern::end ( const size_type r ) const

Final iterator of row r. It points to the first element past the end of line r, or past the end of the entire sparsity pattern.

Note that the end iterator is not necessarily dereferencable. This is in particular the case if it is the end iterator for the last row of a matrix.

 void SparsityPattern::write_ascii ( )

Abstract Trilinos object that helps view in ASCII other Trilinos objects. Currently this function is not implemented. TODO: Not implemented.

Definition at line 1099 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::print ( std::ostream & out, const bool write_extended_trilinos_info = false ) const

Print (the locally owned part of) the sparsity pattern to the given stream, using the format (line,col). The optional flag outputs the sparsity pattern in Trilinos style, where even the according processor number is printed to the stream, as well as a summary before actually writing the entries.

Definition at line 1110 of file trilinos_sparsity_pattern.cc.

 void SparsityPattern::print_gnuplot ( std::ostream & out ) const

Print the sparsity of the matrix in a format that gnuplot understands and which can be used to plot the sparsity pattern in a graphical way. The format consists of pairs i j of nonzero elements, each representing one entry of this matrix, one per line of the output file. Indices are counted from zero on, as usual. Since sparsity patterns are printed in the same way as matrices are displayed, we print the negative of the column index, which means that the (0,0) element is in the top left rather than in the bottom left corner.

Print the sparsity pattern in gnuplot by setting the data style to dots or points and use the plot command.

Definition at line 1137 of file trilinos_sparsity_pattern.cc.

## Member Data Documentation

 std::unique_ptr TrilinosWrappers::SparsityPattern::column_space_map
private

Pointer to the user-supplied Epetra Trilinos mapping of the matrix columns that assigns parts of the matrix to the individual processes.

Definition at line 1216 of file trilinos_sparsity_pattern.h.

 std::unique_ptr TrilinosWrappers::SparsityPattern::graph
private

A sparsity pattern object in Trilinos to be used for finite element based problems which allows for adding non-local elements to the pattern.

Definition at line 1223 of file trilinos_sparsity_pattern.h.

 std::unique_ptr TrilinosWrappers::SparsityPattern::nonlocal_graph
private

A sparsity pattern object for the non-local part of the sparsity pattern that is going to be sent to the owning processor. Only used when the particular constructor or reinit method with writable_rows argument is set

Definition at line 1231 of file trilinos_sparsity_pattern.h.

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