BsrMatrix

Header file: KokkosSparse_BsrMatrix.hpp

Usage

The BsrMatrix class provides a block compressed sparse row implementation of a sparse matrix, as described, for example, in Saad (2nd ed.).

Interface

template <class ScalarType, class OrdinalType, class Device,
          class MemoryTraits = void, class SizeType = default_size_type>
class BsrMatrix;

Parameters

• ScalarType: Determines the type of the values stored in the matrix, typically int, float, double, Kokkos::complex<...>, ...
• OrdinalType: Determines the type of the column indices held in the graph of the matrix, typically int or long long
• Device:
• MemoryTraits: Sets access properties via enum parameters for the templated Kokkos::MemoryTraits<> class. This is assumed to be the shared memory trait used by the underlying graph and values.
• SizeType: Determines the type of the offsets stored in the row pointer of the underlying graph, typically int or size_t

Public class members

Typedefs

type execution_space

Execution Space associated with the view, will be used for performing view initialization, and certain deep_copy operations.

type memory_space

Memory Space associated with the underlying views.

type device_type

Combination of the execution_space and memory_space, nothing further really

type value_type

The scalar type of the data the values view is referencing

type ordinal_type

The ordinal type of the data the entries view is referencing

type memory_traits

The memory trait of the underlying views

type size_type

The ordinal type of the data the rowmap view is referencing

type HostMirror

Type of the host mirror for this BsrMatrix

type StaticCrsGraphType

The type of the graph store by the matrix

type staticcrsgraph_type

Lower case alias of the above StaticCrsGraphType

type index_type

Type of column indices in the sparse matrix.

type const_ordinal_type

Const version of the type of column indices in the sparse matrix.

type non_const_ordinal_type

Nonconst version of the type of column indices in the sparse matrix.

type row_map_type

Type of the "row map" (which contains the offset for each row's data).

type const_size_type

Const version of the type of row offsets in the sparse matrix.

type non_const_size_type

Nonconst version of the type of row offsets in the sparse matrix.

type values_type

Kokkos Array type of the entries (values) in the sparse matrix.

type const_value_type

Const version of the type of the entries in the sparse matrix.

type non_const_value_type

Nonconst version of the type of the entries in the sparse matrix.

type block_layout_type

Layout of the block_type and const_block_type views

type block_type

Type of the views returned when accessing a block of the matrix with unmanaged_block()

type const_block_type

Type of the views returned when accessing a block of the matrix with unmanaged_block_const()

Constructors

BsrMatrix()

Default constructor, constructs an empty matrix with no allocations performed

BsrMatrix(const BsrMatrix<SType, OType, DType, MTType, IType>& B)

Shallow copy constructor

BsrMatrix(const std::string& arg_label, const staticcrsgraph_type& arg_graph, const OrdinalType& blockDimIn)

Construct from graph and blockDim with label.

BsrMatrix(const std::string& label, OrdinalType nrows, OrdinalType ncols,
            size_type annz, ScalarType* vals, OrdinalType* rows,
            OrdinalType* cols, OrdinalType blockdim, bool pad = false)

Construct from host data in COO format

BsrMatrix(const std::string& label, const OrdinalType nrows, const OrdinalType ncols,
          const size_type annz, const values_type& vals,
          const row_map_type& rows, const index_type& cols,
          const OrdinalType blockDimIn)

Constructor using the underlying views for row pointer, column indices and values.

BsrMatrix(const std::string& label, const OrdinalType& ncols,
            const values_type& vals, const staticcrsgraph_type& graph_,
            const OrdinalType& blockDimIn)

Constructor using graph and values as input

BsrMatrix(const KokkosSparse::CrsMatrix<SType, OType, DType, MTType, IType>&
                crs_mtx,
            const OrdinalType blockDimIn)

Constructor using CrsMatrix as input

Member functions

BsrMatrix& operator=(const BsrMatrix<aScalarType, aOrdinalType, aDevice,
                                       aMemoryTraits, aSizeType>& mtx)

Assignment operator

ordinal_type numRows() const

Get the number of block rows in the matrix, identical to the number of rows in the underlying graph.

ordinal_type numCols() const

Get the number of block columns in the matrix, identical to the number of columns in the underlying graph.

ordinal_type blockDim() const

Get the block dimension

ordinal_type numPointRows() const

Get the number of point rows in the matrix

ordinal_type numPointCols() const

Get the number of point columns in the matrix

size_type nnz() const

Get the number of blocks in the matrix, same as the number of entries in the underlying graph.

OrdinalType replaceValues(const OrdinalType rowi, const OrdinalType cols[],
                            const OrdinalType ncol, const ScalarType vals[],
                            const bool is_sorted    = false,
                            const bool force_atomic = false)

Replace an entry in the matrix

OrdinalType sumIntoValues(const OrdinalType rowi, const OrdinalType cols[],
                            const OrdinalType ncol, const ScalarType vals[],
                            const bool is_sorted    = false,
                            const bool force_atomic = false)

Add to an entry in the matrix

BsrRowView<BsrMatrix> block_row(const ordinal_type i) const

BsrRowViewConst<BsrMatrix> block_row_const(const ordinal_type i) const

block_type unmanaged_block(const size_type i) const

return an unmanaged view of block i

const_block_type unmanaged_block_const(const size_type i) const

return an unmanaged view of constant block i

Sparse linear algebra functions that support a BsrMatrix interface

1. SpMV
2. SpADD
3. SpGEMM

SAND2024-06899W