MPI

Detailed Description

DIY includes a simple C++ wrapper around the MPI commands. It borrows heavily from Boost.MPI design.

#include <diy/mpi.hpp>

int main(int argc, char* argv[])
{
  diy::mpi::environment     env(argc, argv);        // RAII
  diy::mpi::communicator    world;
  ...
}

A simple wrapper diy::mpi::environment takes care of MPI initialization and finalization using RAII. diy::mpi::communicator wraps an MPI_Comm object and provides methods to send() and recv() (as well as isend()) and irecv()).

All functions map C++ types to the appropriate MPI datatypes, with a specialized interface for an std::vector of them.

Collectives

• diy::mpi::broadcast()
• diy::mpi::gather()
• diy::mpi::reduce()

  int in = comm.rank() * 3;
  if (comm.rank() == 0)
  {
      int out;
      mpi::reduce(comm, in, out, std::plus<float>());
  } else
      mpi::reduce(comm, in, std::plus<float>());

• diy::mpi::all_reduce()
• diy::mpi::scan()
• diy::mpi::all_to_all()

Operations

class	MPI operation
diy::mpi::maximum<U>	MPI_MAX
diy::mpi::minimum<U>	MPI_MIN
std::plus<U>	MPI_SUM
std::multiplies<U>	MPI_PROD
std::logical_and<U>	MPI_LAND
std::logical_or<U>	MPI_LOR

Classes
class	diy::mpi::communicator
	Simple wrapper around MPI_Comm. More...
class	diy::mpi::io::file
	Wraps MPI file IO. More...
struct	diy::mpi::maximum< U >
struct	diy::mpi::minimum< U >
struct	diy::mpi::environment

Functions
template<class T >
void	diy::mpi::broadcast (const communicator &comm, T &x, int root)
	Broadcast to all processes in comm.
template<class T >
void	diy::mpi::broadcast (const communicator &comm, std::vector< T > &x, int root)
	Broadcast for vectors.
template<class T >
request	diy::mpi::ibroadcast (const communicator &comm, T &x, int root)
	iBroadcast to all processes in comm.
template<class T >
void	diy::mpi::gather (const communicator &comm, const T &in, std::vector< T > &out, int root)
	Gather from all processes in comm. On root process, out is resized to comm.size() and filled with elements from the respective ranks.
template<class T >
void	diy::mpi::gather (const communicator &comm, const std::vector< T > &in, std::vector< std::vector< T > > &out, int root)
	Same as above, but for vectors.
template<class T >
void	diy::mpi::gather (const communicator &comm, const T &in, int root)
	Simplified version (without out) for use on non-root processes.
template<class T >
void	diy::mpi::gather (const communicator &comm, const std::vector< T > &in, int root)
	Simplified version (without out) for use on non-root processes.
template<class T >
void	diy::mpi::all_gather (const communicator &comm, const T &in, std::vector< T > &out)
	all_gather from all processes in comm. out is resized to comm.size() and filled with elements from the respective ranks.
template<class T >
void	diy::mpi::all_gather (const communicator &comm, const std::vector< T > &in, std::vector< std::vector< T > > &out)
	Same as above, but for vectors.
template<class T , class Op >
void	diy::mpi::reduce (const communicator &comm, const T &in, T &out, int root, const Op &op)
	reduce
template<class T , class Op >
void	diy::mpi::reduce (const communicator &comm, const T &in, int root, const Op &op)
	Simplified version (without out) for use on non-root processes.
template<class T , class Op >
void	diy::mpi::all_reduce (const communicator &comm, const T &in, T &out, const Op &op)
	all_reduce
template<class T , class Op >
void	diy::mpi::all_reduce (const communicator &comm, const std::vector< T > &in, std::vector< T > &out, const Op &op)
	Same as above, but for vectors.
template<class T , class Op >
void	diy::mpi::scan (const communicator &comm, const T &in, T &out, const Op &op)
	scan
template<class T >
void	diy::mpi::all_to_all (const communicator &comm, const std::vector< T > &in, std::vector< T > &out, int n=1)
	all_to_all