Mapping between a rank's main data and its ghost copies. More...

#include <IndexMapping.hpp>

Collaboration diagram for DNDS::OffsetAscendIndexMapping:

Public Member Functions
template<class TpullSet >
	OffsetAscendIndexMapping (index nmainOffset, index nmainSize, TpullSet &&pullingIndexGlobal, const GlobalOffsetsMapping &LGlobalMapping, const MPIInfo &mpi)
	Construct the mapping from a pull specification.

template<class TpushSet , class TpushStart >
	OffsetAscendIndexMapping (index nmainOffset, index nmainSize, TpushSet &&pushingIndexesLocal, TpushStart &&pushingStarts, const GlobalOffsetsMapping &LGlobalMapping, const MPIInfo &mpi)
	Construct the mapping from a push specification.

void	sort ()
	Sort `ghostIndex` into ascending order. Required invariant maintained after pull construction.

index &	ghostAt (MPI_int rank, index ighost)
	Direct mutable access to the `ighost`-th ghost global index for peer `rank`.

bool	searchInMain (index globalQuery, index &val) const
	Check whether a global index lies within the main block.

bool	searchInGhost (index globalQuery, MPI_int rank, index &val) const
	Search a single peer rank's ghost slab for a global index.

bool	searchInAllGhost (index globalQuery, MPI_int &rank, index &val) const
	Search the concatenated ghost array for a global index (O(log nGhost)).

bool	search (index globalQuery, MPI_int &rank, index &val) const
	Search main + ghost in one call. `rank == -1` signals "hit in main".

std::tuple< bool, MPI_int, index >	search (index globalQuery) const
	Tuple-return overload of the main search. `(success, rank, local)`.

bool	search_indexAppend (index globalQuery, MPI_int &rank, index &val) const
	Like search, but the returned `val` is in the concatenated father+son address space.

std::tuple< bool, MPI_int, index >	search_indexAppend (index globalQuery) const
	Tuple-return overload of search_indexAppend.

bool	search_indexRank (index globalQuery, MPI_int &rank, index &val) const
	Like search, but `val` is expressed relative to the owner rank's ghost slab (`ghostStart[rank]`-relative).

std::tuple< bool, MPI_int, index >	search_indexRank (index globalQuery) const
	Tuple-return overload of search_indexRank.

index	operator() (MPI_int rank, index val) const
	Reverse (local -> global) mapping.

Public Attributes
tMPI_intVec	ghostSizes
	Number of ghost rows pulled from each peer rank. (a.k.a. `pullIndexSizes`)

t_IndexVec	ghostIndex
	Global indices of all ghost rows on this rank, ascending. (a.k.a. `pullingIndexGlobal`)

t_MapIndexVec	ghostStart
	Per-peer prefix-sum offsets into `ghostIndex` (size == `nRanks + 1`). (a.k.a. `pullIndexStarts`)

tMPI_intVec	pushIndexSizes
	Per-peer number of rows this rank will send out on a push.

t_MapIndexVec	pushIndexStarts
	Per-peer prefix-sum offsets into `pushingIndexGlobal`.

t_IndexVec	pushingIndexGlobal
	Global indices this rank sends to others, grouped by receiver.

t_IndexVec	pullingRequestLocal
	Optional cached local-index form of ghostIndex (unused in current code).

Detailed Description

Mapping between a rank's main data and its ghost copies.

main data (owned rows) is a contiguous range [mainOffset, mainOffset + mainSize) in the global index space. ghost data (rows needed from other ranks) is stored in globally ascending order so that binary search may be used for the reverse mapping.

Two construction modes:

Pull: the caller provides the set of global indices this rank needs.
Push: the caller provides which local rows to send to which ranks. Both modes exchange the complementary information via MPI_Alltoall(v) so the resulting mapping is valid on both sides.

Internal structure (all vectors are per-peer-rank layouts, sorted by rank):

ghostSizes / ghostStart / ghostIndex: what this rank pulls from each other rank.
pushIndexSizes / pushIndexStarts / pushingIndexGlobal: what this rank sends to each other rank.

Warning: Ghost indices carry only 32-bit payload in MPI calls; the total per-rank ghost count must fit in int32_t.

Definition at line 165 of file IndexMapping.hpp.

Constructor & Destructor Documentation

◆ OffsetAscendIndexMapping() [1/2]

template<class TpullSet >

DNDS::OffsetAscendIndexMapping::OffsetAscendIndexMapping	(	index	nmainOffset,
		index	nmainSize,
		TpullSet &&	pullingIndexGlobal,
		const GlobalOffsetsMapping &	LGlobalMapping,
		const MPIInfo &	mpi
	)

inline

Construct the mapping from a pull specification.

Template Parameters

TpullSet Range of index (vector-like, supports sort/unique/erase).

Parameters

nmainOffset	Global offset of this rank's main data block.
nmainSize	Number of main rows owned by this rank.
pullingIndexGlobal	Global indices this rank wants to see as ghosts. Sorted and deduplicated in-place.
LGlobalMapping	Globally-known offsets, used to attribute each global index to its owner rank.
mpi	MPI context; the constructor is collective.

Warning: pullingIndexGlobal is mutated (sorted / uniquified / shrunk). The resulting ghostIndex ends up in ascending global-index order regardless of the caller's original ordering.

*make sorted and unique!

Definition at line 214 of file IndexMapping.hpp.

Here is the call graph for this function:

◆ OffsetAscendIndexMapping() [2/2]

template<class TpushSet , class TpushStart >

DNDS::OffsetAscendIndexMapping::OffsetAscendIndexMapping	(	index	nmainOffset,
		index	nmainSize,
		TpushSet &&	pushingIndexesLocal,
		TpushStart &&	pushingStarts,
		const GlobalOffsetsMapping &	LGlobalMapping,
		const MPIInfo &	mpi
	)

inline

Construct the mapping from a push specification.

The caller supplies, for every peer rank, a list of local row indices that this rank should send. The constructor flips this into the pull view via MPI_Alltoall(v).

Template Parameters

TpushSet	Range-like of local indices.
TpushStart	Size-`nRanks+1` prefix-sum offsets.

Parameters

nmainOffset	Global offset of this rank's main block.
nmainSize	Number of main rows owned by this rank.
pushingIndexesLocal	Local indices to push, grouped by receiver (flat array; `[pushingStarts[r], pushingStarts[r+1])` are the indices going to rank `r`).
pushingStarts	Prefix sums of per-receiver counts, size `nRanks+1`.
LGlobalMapping	Rank-offset table (to convert local -> global).
mpi	MPI context; collective.

Definition at line 297 of file IndexMapping.hpp.

Here is the call graph for this function:

Member Function Documentation

◆ ghostAt()

index & DNDS::OffsetAscendIndexMapping::ghostAt	(	MPI_int	rank,
		index	ighost
	)

inline

Direct mutable access to the ighost-th ghost global index for peer rank.

Definition at line 344 of file IndexMapping.hpp.

Here is the caller graph for this function:

◆ operator()()

index DNDS::OffsetAscendIndexMapping::operator()	(	MPI_int	rank,
		index	val
	)		const

inline

Reverse (local -> global) mapping.

rank == -1: val indexes the concatenated father+son address space. val < mainSize returns the main global index; val >= mainSize returns the corresponding ghost global index.
rank >= 0: val indexes into that peer's ghost slab (like search_indexRank's output).

if rank == -1, return the global index of local main data, or else return the ghosted global index of local ghost data.

Definition at line 509 of file IndexMapping.hpp.

◆ search() [1/2]

std::tuple< bool, MPI_int, index > DNDS::OffsetAscendIndexMapping::search ( index globalQuery ) const

inline

Tuple-return overload of the main search. (success, rank, local).

Definition at line 432 of file IndexMapping.hpp.

Here is the call graph for this function:

◆ search() [2/2]

bool DNDS::OffsetAscendIndexMapping::search	(	index	globalQuery,
		MPI_int &	rank,
		index &	val
	)		const

inline

Search main + ghost in one call. rank == -1 signals "hit in main".

Returned val is in the ghost address space (relative to ghostStart[0]) when rank != -1, else it is the main-local index. See also search_indexAppend for the combined (father+son) index space.

returns rank and place in ghost array, rank==-1 means main data

Definition at line 418 of file IndexMapping.hpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ search_indexAppend() [1/2]

std::tuple< bool, MPI_int, index > DNDS::OffsetAscendIndexMapping::search_indexAppend ( index globalQuery ) const

inline

Tuple-return overload of search_indexAppend.

Definition at line 462 of file IndexMapping.hpp.

Here is the call graph for this function:

◆ search_indexAppend() [2/2]

bool DNDS::OffsetAscendIndexMapping::search_indexAppend	(	index	globalQuery,
		MPI_int &	rank,
		index &	val
	)		const

inline

Like search, but the returned val is in the concatenated father+son address space.

val < mainSize means main row; val >= mainSize means ghost row at val - mainSize. Matches ArrayPair::operator[] layout.

returns rank and place in ghost array, rank==-1 means main data; returned val is used for pair indexing

Definition at line 446 of file IndexMapping.hpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ search_indexRank() [1/2]

std::tuple< bool, MPI_int, index > DNDS::OffsetAscendIndexMapping::search_indexRank ( index globalQuery ) const

inline

Tuple-return overload of search_indexRank.

Definition at line 492 of file IndexMapping.hpp.

Here is the call graph for this function:

◆ search_indexRank() [2/2]

bool DNDS::OffsetAscendIndexMapping::search_indexRank	(	index	globalQuery,
		MPI_int &	rank,
		index &	val
	)		const

inline

Like search, but val is expressed relative to the owner rank's ghost slab (ghostStart[rank]-relative).

Warning: Do not confuse with search, whose val is relative to the concatenated ghost array.

returns rank and place in ghost of rank, rank==-1 means main data

Warning: search returns index that applies to local ghost array, this only goes for the ith of rank

Definition at line 476 of file IndexMapping.hpp.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ searchInAllGhost()

bool DNDS::OffsetAscendIndexMapping::searchInAllGhost	(	index	globalQuery,
		MPI_int &	rank,
		index &	val
	)		const

inline

Search the concatenated ghost array for a global index (O(log nGhost)).

Parameters

[out]	rank	Peer rank that owns the found index.
[out]	val	Offset relative to `ghostStart[0]` (== 0).

Definition at line 395 of file IndexMapping.hpp.

Here is the caller graph for this function:

◆ searchInGhost()

bool DNDS::OffsetAscendIndexMapping::searchInGhost	(	index	globalQuery,
		MPI_int	rank,
		index &	val
	)		const

inline

Search a single peer rank's ghost slab for a global index.

Parameters

[out] val Offset relative to ghostStart[rank] on success.

Definition at line 375 of file IndexMapping.hpp.

Here is the caller graph for this function:

◆ searchInMain()

bool DNDS::OffsetAscendIndexMapping::searchInMain	(	index	globalQuery,
		index &	val
	)		const

inline

Check whether a global index lies within the main block.

Parameters

	globalQuery	Global index to look up.
[out]	val	On success, local offset within the main block.

Returns: true if globalQuery is owned by this rank.

Definition at line 361 of file IndexMapping.hpp.

Here is the caller graph for this function:

◆ sort()

void DNDS::OffsetAscendIndexMapping::sort ( )

inline

Sort ghostIndex into ascending order. Required invariant maintained after pull construction.

Definition at line 338 of file IndexMapping.hpp.

Here is the caller graph for this function:

Member Data Documentation

◆ ghostIndex

t_IndexVec DNDS::OffsetAscendIndexMapping::ghostIndex

Global indices of all ghost rows on this rank, ascending. (a.k.a. pullingIndexGlobal)

Definition at line 179 of file IndexMapping.hpp.

◆ ghostSizes

tMPI_intVec DNDS::OffsetAscendIndexMapping::ghostSizes

Number of ghost rows pulled from each peer rank. (a.k.a. pullIndexSizes)

Definition at line 175 of file IndexMapping.hpp.

◆ ghostStart

t_MapIndexVec DNDS::OffsetAscendIndexMapping::ghostStart

Per-peer prefix-sum offsets into ghostIndex (size == nRanks + 1). (a.k.a. pullIndexStarts)

Definition at line 183 of file IndexMapping.hpp.

◆ pullingRequestLocal

t_IndexVec DNDS::OffsetAscendIndexMapping::pullingRequestLocal

Optional cached local-index form of ghostIndex (unused in current code).

Encoding: >=0 means array[i], <0 means ghost_array[-i-1]. Retained for potential future reconstruction paths.

Definition at line 195 of file IndexMapping.hpp.

◆ pushIndexSizes

tMPI_intVec DNDS::OffsetAscendIndexMapping::pushIndexSizes

Per-peer number of rows this rank will send out on a push.

Definition at line 186 of file IndexMapping.hpp.

◆ pushIndexStarts

t_MapIndexVec DNDS::OffsetAscendIndexMapping::pushIndexStarts

Per-peer prefix-sum offsets into pushingIndexGlobal.

Definition at line 188 of file IndexMapping.hpp.

◆ pushingIndexGlobal

t_IndexVec DNDS::OffsetAscendIndexMapping::pushingIndexGlobal

Global indices this rank sends to others, grouped by receiver.

Definition at line 190 of file IndexMapping.hpp.

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

/home/runner/work/DNDSR/DNDSR/src/DNDS/IndexMapping.hpp

Public Member Functions

Public Attributes

Detailed Description

Constructor & Destructor Documentation

◆ OffsetAscendIndexMapping() [1/2]

◆ OffsetAscendIndexMapping() [2/2]

Member Function Documentation

◆ ghostAt()

◆ operator()()

◆ search() [1/2]

◆ search() [2/2]

◆ search_indexAppend() [1/2]

◆ search_indexAppend() [2/2]

◆ search_indexRank() [1/2]

◆ search_indexRank() [2/2]

◆ searchInAllGhost()

◆ searchInGhost()

◆ searchInMain()

◆ sort()

Member Data Documentation

◆ ghostIndex

◆ ghostSizes

◆ ghostStart

◆ pullingRequestLocal

◆ pushIndexSizes

◆ pushIndexStarts

◆ pushingIndexGlobal