DiFfRG::MPI Namespace Reference

Functions
bool	available ()
	Whether this build has MPI and MPI_Init has run. Everything below degrades to a single-rank no-op when this is false, so callers never have to branch on the build type.
uint	rank (MPI_Comm comm)
uint	size (MPI_Comm comm)
void	barrier (MPI_Comm comm)
void	sum_reduce (MPI_Comm comm, int *data, int size)
void	max_reduce (MPI_Comm comm, int *data, int size)
void	min_reduce (MPI_Comm comm, int *data, int size)
void	sum_reduce (MPI_Comm comm, double *data, int size)
void	max_reduce (MPI_Comm comm, double *data, int size)
void	min_reduce (MPI_Comm comm, double *data, int size)
int	sum_reduce (MPI_Comm comm, int value)
int	max_reduce (MPI_Comm comm, int value)
int	min_reduce (MPI_Comm comm, int value)
double	sum_reduce (MPI_Comm comm, double value)
double	max_reduce (MPI_Comm comm, double value)
double	min_reduce (MPI_Comm comm, double value)
bool	any_of (MPI_Comm comm, bool value)
	Collective agreement on a predicate. Used to turn a rank-local decision (an abort, a convergence flag) into one that every rank makes identically – ranks that disagree about control flow enter different numbers of collectives and hang.
bool	all_of (MPI_Comm comm, bool value)
bool	agree (MPI_Comm comm, uint64_t token)
	Whether every rank passed the same token. One collective.
void	bcast (MPI_Comm comm, void *data, size_t bytes, uint root=0)
	Broadcast raw bytes from root to all ranks.
void	allgatherv_bytes (MPI_Comm comm, const void *send, size_t send_bytes, void *recv, const std::vector< int > &recv_counts, const std::vector< int > &displs)
	Concatenating all-gather of byte blocks of differing size.
MPI_Comm	split_shared (MPI_Comm comm)
	A communicator over the ranks sharing this node's memory, for GPU affinity.
void	free_comm (MPI_Comm &comm)
void	abort (MPI_Comm comm, const std::string &message, int code=1)
	Tear the whole job down with a message on stderr.

Function Documentation

abort()

void DiFfRG::MPI::abort	(	MPI_Comm	comm,
		const std::string &	message,
		int	code = 1 )

Tear the whole job down with a message on stderr.

Used where a throw would deadlock: an exception unwinds one rank out of a collective the others are still waiting in, so the failure would present as a hang instead of an error.

agree()

bool DiFfRG::MPI::agree	(	MPI_Comm	comm,
		uint64_t	token )

Whether every rank passed the same token. One collective.

For checking agreement on something whose disagreement would otherwise show up as a hang (mismatched collective arguments) rather than an error.

all_of()

bool DiFfRG::MPI::all_of	(	MPI_Comm	comm,
		bool	value )

allgatherv_bytes()

void DiFfRG::MPI::allgatherv_bytes	(	MPI_Comm	comm,
		const void *	send,
		size_t	send_bytes,
		void *	recv,
		const std::vector< int > &	recv_counts,
		const std::vector< int > &	displs )

Concatenating all-gather of byte blocks of differing size.

recv_counts and displs are in bytes and must be identical on every rank; recv must hold sum(recv_counts) bytes. Used by MapScheduler to exchange map() slices. Deliberately not an Allreduce(SUM) over zero-filled buffers: that moves N times more data and is not bit-safe, since (-0.0) + 0.0 == +0.0 destroys a signed zero.

any_of()

bool DiFfRG::MPI::any_of	(	MPI_Comm	comm,
		bool	value )

Collective agreement on a predicate. Used to turn a rank-local decision (an abort, a convergence flag) into one that every rank makes identically – ranks that disagree about control flow enter different numbers of collectives and hang.

available()

bool DiFfRG::MPI::available

(

)

Whether this build has MPI and MPI_Init has run. Everything below degrades to a single-rank no-op when this is false, so callers never have to branch on the build type.

barrier()

void DiFfRG::MPI::barrier

(

MPI_Comm

comm

)

bcast()

void DiFfRG::MPI::bcast	(	MPI_Comm	comm,
		void *	data,
		size_t	bytes,
		uint	root = 0 )

Broadcast raw bytes from root to all ranks.

free_comm()

void DiFfRG::MPI::free_comm

(

MPI_Comm &

comm

)

max_reduce() [1/4]

void DiFfRG::MPI::max_reduce	(	MPI_Comm	comm,
		double *	data,
		int	size )

max_reduce() [2/4]

double DiFfRG::MPI::max_reduce	(	MPI_Comm	comm,
		double	value )

max_reduce() [3/4]

void DiFfRG::MPI::max_reduce	(	MPI_Comm	comm,
		int *	data,
		int	size )

max_reduce() [4/4]

int DiFfRG::MPI::max_reduce	(	MPI_Comm	comm,
		int	value )

min_reduce() [1/4]

void DiFfRG::MPI::min_reduce	(	MPI_Comm	comm,
		double *	data,
		int	size )

min_reduce() [2/4]

double DiFfRG::MPI::min_reduce	(	MPI_Comm	comm,
		double	value )

min_reduce() [3/4]

void DiFfRG::MPI::min_reduce	(	MPI_Comm	comm,
		int *	data,
		int	size )

min_reduce() [4/4]

int DiFfRG::MPI::min_reduce	(	MPI_Comm	comm,
		int	value )

rank()

uint DiFfRG::MPI::rank

(

MPI_Comm

comm

)

size()

uint DiFfRG::MPI::size

(

MPI_Comm

comm

)

split_shared()

MPI_Comm DiFfRG::MPI::split_shared

(

MPI_Comm

comm

)

A communicator over the ranks sharing this node's memory, for GPU affinity.

Returns MPI_COMM_SELF-equivalent behaviour (size 1) without MPI. Free it with free_comm().

sum_reduce() [1/4]

void DiFfRG::MPI::sum_reduce	(	MPI_Comm	comm,
		double *	data,
		int	size )

sum_reduce() [2/4]

double DiFfRG::MPI::sum_reduce	(	MPI_Comm	comm,
		double	value )

sum_reduce() [3/4]

void DiFfRG::MPI::sum_reduce	(	MPI_Comm	comm,
		int *	data,
		int	size )

sum_reduce() [4/4]

int DiFfRG::MPI::sum_reduce	(	MPI_Comm	comm,
		int	value )