Namespace for Tpetra implementation details. More...

Namespaces
	DefaultTypes
	Declarations of values of Tpetra classes' default template parameters.

Classes
struct	AbsMax
	Functor for the the ABSMAX CombineMode of Import and Export operations. More...

class	ContiguousUniformDirectory
	Implementation of Directory for a contiguous, uniformly distributed Map. More...

struct	CrsIJV
	Struct representing a sparse matrix entry as an i,j,v triplet. More...

class	Directory
	Computes the local ID and process ID corresponding to given global IDs. More...

class	DistributedContiguousDirectory
	Implementation of Directory for a distributed contiguous Map. More...

class	DistributedNoncontiguousDirectory
	Implementation of Directory for a distributed noncontiguous Map. More...

class	FixedHashTable

struct	GetLapackType
	Return the Teuchos::LAPACK specialization corresponding to the given Scalar type. More...

struct	Hash
	The hash function for FixedHashTable. More...

struct	Hash< KeyType, DeviceType, OffsetType, int >
	Specialization for ResultType = int. More...

class	HashTable

class	InvalidGlobalIndex
	Exception thrown by CrsMatrix on invalid global index. More...

class	InvalidGlobalRowIndex
	Exception thrown by CrsMatrix on invalid global row index. More...

class	MapCloner
	Implementation detail of Map::clone(). More...

struct	MultiVectorCloner
	Implementation of Tpetra::MultiVector::clone(). More...

class	MultiVectorFillerData
	Implementation of fill and local assembly for MultiVectorFiller. More...

class	MultiVectorFillerData2
	Second implementation of fill and local assembly for MultiVectorFiller. More...

class	OptColMap
	Implementation detail of makeOptimizedColMap, and makeOptimizedColMapAndImport. More...

struct	OrdinalTraits
	Traits class for "invalid" (flag) values of integer types that Tpetra uses as local ordinals or global ordinals. More...

struct	PackTraits
	Traits class for packing / unpacking data of type `T`, using Kokkos data structures that live in the given space `D`. More...

class	ReplicatedDirectory
	Implementation of Directory for a locally replicated Map. More...

class	TieBreak
	Interface for breaking ties in ownership. More...

class	Transfer
	Common base class of Import and Export. More...

Enumerations

Functions
void	gathervPrint (std::ostream &out, const std::string &s, const Teuchos::Comm< int > &comm)
	On Process 0 in the given communicator, print strings from each process in that communicator, in rank order. More...

template<class MapType >
MapType	makeOptimizedColMap (std::ostream &errStream, bool &lclErr, const MapType &domMap, const MapType &colMap)
	Return an optimized reordering of the given column Map. More...

template<class MapType >
std::pair< MapType, Teuchos::RCP< typename OptColMap< MapType >::import_type > >	makeOptimizedColMapAndImport (std::ostream &errStream, bool &lclErr, const MapType &domMap, const MapType &colMap, const typename OptColMap< MapType >::import_type *oldImport, const bool makeImport)
	Return an optimized reordering of the given column Map. Optionally, recompute an Import from the input domain Map to the new column Map. More...

std::string	DistributorSendTypeEnumToString (EDistributorSendType sendType)
	Convert an EDistributorSendType enum value to a string. More...

std::string	DistributorHowInitializedEnumToString (EDistributorHowInitialized how)
	Convert an EDistributorHowInitialized enum value to a string. More...

bool	congruent (const Teuchos::Comm< int > &comm1, const Teuchos::Comm< int > &comm2)
	Whether the two communicators are congruent. More...

Detailed Description

Namespace for Tpetra implementation details.

Warning: Do NOT rely on the contents of this namespace.

Enumeration Type Documentation

enum Tpetra::Details::EStorageStatus

Status of the graph's or matrix's storage, when not in a fill-complete state.

When a CrsGraph or CrsMatrix is not fill complete, its data live in one of three storage formats:

"2-D storage": The graph stores column indices as "array of arrays," and the matrix stores values as "array of arrays." The graph must have k_numRowEntries_ allocated. This only ever exists if the graph was created with DynamicProfile. A matrix with 2-D storage must own its graph, and the graph must have 2-D storage.
"Unpacked 1-D storage": The graph uses a row offsets array, and stores column indices in a single array. The matrix also stores values in a single array. "Unpacked" means that there may be extra space in each row: that is, the row offsets array only says how much space there is in each row. The graph must use k_numRowEntries_ to find out how many entries there actually are in the row. A matrix with unpacked 1-D storage must own its graph, and the graph must have unpacked 1-D storage.
"Packed 1-D storage": The matrix may or may not own the graph. "Packed" means that there is no extra space in each row. Thus, the k_numRowEntries_ array is not necessary and may have been deallocated. If the matrix was created with a constant ("static") graph, this must be true.

With respect to the Kokkos refactor version of Tpetra, "2-D storage" should be considered a legacy option.

The phrase "When not in a fill-complete state" is important. When the graph is fill complete, it always uses 1-D "packed" storage. However, if storage is "not optimized," we retain the 1-D unpacked or 2-D format, and thus retain this enum value.

Definition at line 160 of file Tpetra_CrsGraph_decl.hpp.

enum Tpetra::Details::EDistributorSendType

The type of MPI send that Distributor should use.

This is an implementation detail of Distributor. Please do not rely on these values in your code.

Definition at line 76 of file Tpetra_Distributor.hpp.

enum Tpetra::Details::EDistributorHowInitialized

Enum indicating how and whether a Distributor was initialized.

This is an implementation detail of Distributor. Please do not rely on these values in your code.

Definition at line 94 of file Tpetra_Distributor.hpp.

Function Documentation

void Tpetra::Details::gathervPrint	(	std::ostream &	out,
		const std::string &	s,
		const Teuchos::Comm< int > &	comm
	)

On Process 0 in the given communicator, print strings from each process in that communicator, in rank order.

For each process in the given communicator comm, send its string s to Process 0 in that communicator. Process 0 prints the strings in rank order.

This is a collective over the given communicator comm. Process 0 promises not to store all the strings in its memory. This function's total memory usage on any process is proportional to the calling process' string length, plus the max string length over any process. This does NOT depend on the number of processes in the communicator. Thus, we call this a "memory-scalable" operation. While the function's name suggests MPI_Gatherv, the implementation may NOT use MPI_Gather or MPI_Gatherv, because neither of those are not memory scalable.

Process 0 prints nothing other than what is in the string. It does not add an endline after each string, nor does it identify each string with its owning process' rank. If you want either of those in the string, you have to put it there yourself.

Parameters

out	[out] The output stream to which to write. ONLY Process 0 in the given communicator will write to this. Thus, this stream need only be valid on Process 0.
s	[in] The string to write. Each process in the given communicator has its own string. Strings may be different on different processes. Zero-length strings are OK.
comm	[in] The communicator over which this operation is a collective.

Definition at line 52 of file Tpetra_Details_gathervPrint.cpp.

template<class MapType >

MapType Tpetra::Details::makeOptimizedColMap	(	std::ostream &	errStream,
		bool &	lclErr,
		const MapType &	domMap,
		const MapType &	colMap
	)

Return an optimized reordering of the given column Map.

Template Parameters

MapType A specialization of Map.

Parameters

err	[out] Output stream for human-readable error reporting. This is local to the calling process and may differ on different processes.
lclErr	[out] On output: true if anything went wrong on the calling process. This value is local to the calling process and may differ on different processes.
domMap	[in] Domain Map of a CrsGraph or CrsMatrix.
colMap	[in] Original column Map of the same CrsGraph or CrsMatrix as `domMap`.

Returns: The possibly reordered column Map newColMap.

This is a convenience wrapper for makeOptimizedColMapAndImport(). (Please refer to that function's documentation in this file.) It does everything that that function does, except that it does not compute a new Import.

Definition at line 310 of file Tpetra_Details_makeOptimizedColMap.hpp.

template<class MapType >

std::pair<MapType, Teuchos::RCP<typename OptColMap<MapType>::import_type> > Tpetra::Details::makeOptimizedColMapAndImport	(	std::ostream &	errStream,
		bool &	lclErr,
		const MapType &	domMap,
		const MapType &	colMap,
		const typename OptColMap< MapType >::import_type *	oldImport,
		const bool	makeImport
	)

Return an optimized reordering of the given column Map. Optionally, recompute an Import from the input domain Map to the new column Map.

Template Parameters

MapType A specialization of Map.

This function takes a domain Map and a column Map of a distributed graph (Tpetra::CrsGraph) or matrix (e.g., Tpetra::CrsMatrix). It then creates a new column Map, which optimizes the performance of an Import operation from the domain Map to the new column Map. This function also optionally creates that Import. Creating the new column Map and its Import at the same time saves some communication, since making the Import requires some of the same information that optimizing the column Map does.

Parameters

err	[out] Output stream for human-readable error reporting. This is local to the calling process and may differ on different processes.
lclErr	[out] On output: true if anything went wrong on the calling process. This value is local to the calling process and may differ on different processes.
domMap	[in] Domain Map of a CrsGraph or CrsMatrix.
colMap	[in] Original column Map of the same CrsGraph or CrsMatrix as `domMap`.
makeImport	[in] Whether to make and return an Import from the input domain Map to the new column Map.

Returns: The possibly reordered column Map newColMap, and the corresponding Import from domMap to newColMap. The latter is nonnull if and only if makeImport is true.

Precondition: domMap and colMap must have the same or congruent communicators.; On all calling processes, the indices in colMap must be a subset of the indices in domMap.

The returned column Map's global indices (GIDs) will have the following order on all calling processes:

GIDs that occur in both colMap and domMap (on the calling process) go first.
GIDs in colMap on the calling process, but not in the domain Map on the calling process, follow. They are ordered first contiguously by their owning process rank (in the domain Map), then in increasing order within that.

This imitates the ordering used by AztecOO and Epetra. Storing indices contiguously that are owned by the same process (in the domain Map) permits the use of contiguous send and receive buffers in Distributor, which is used in an Import operation.

Definition at line 383 of file Tpetra_Details_makeOptimizedColMap.hpp.

std::string Tpetra::Details::DistributorSendTypeEnumToString ( EDistributorSendType sendType )

Convert an EDistributorSendType enum value to a string.

This is an implementation detail of Distributor. Please do not rely on this function in your code.

Definition at line 50 of file Tpetra_Distributor.cpp.

std::string Tpetra::Details::DistributorHowInitializedEnumToString ( EDistributorHowInitialized how )

Convert an EDistributorHowInitialized enum value to a string.

This is an implementation detail of Distributor. Please do not rely on this function in your code.

Definition at line 71 of file Tpetra_Distributor.cpp.

bool Tpetra::Details::congruent	(	const Teuchos::Comm< int > &	comm1,
		const Teuchos::Comm< int > &	comm2
	)

Whether the two communicators are congruent.

Two communicators are congruent when they have the same number of processes, and those processes occur in the same rank order.

If both communicators are MpiComm instances, this function returns true exactly when MPI_Comm_compare returns MPI_IDENT (the communicators are handles for the same object) or MPI_CONGRUENT. SerialComm instances are always congruent. An MpiComm is congruent to a SerialComm if the MpiComm has only one process. This function is symmetric in its arguments.

If either Comm instance is neither an MpiComm nor a SerialComm, this method cannot do any better than to compare their process counts.

Two communicators are congruent when they have the same number of processes, and those processes occur in the same rank order.

If both communicators are Teuchos::MpiComm instances, this function returns true exactly when MPI_Comm_compare returns MPI_IDENT (the communicators are handles for the same object) or MPI_CONGRUENT on their MPI_Comm handles. Any two Teuchos::SerialComm instances are always congruent. An MpiComm instance is congruent to a SerialComm instance if and only if the MpiComm has one process. This function is symmetric in its arguments.

If either Teuchos::Comm instance is neither an MpiComm nor a SerialComm, this method cannot do any better than to compare their process counts.

Definition at line 65 of file Tpetra_Util.cpp.

Namespaces

Classes

Enumerations

Functions

Detailed Description

Enumeration Type Documentation

Function Documentation