AllReduce Class Reference

Global reduction operations.

#include <allreduce.hpp>

Public Member Functions
	AllReduce (Communicator const &communicator, size_type const sizeByte)
	Setup environment for global reduction. More...
template<class Iterator >
void	insert (Iterator inputIt, size_type const count)
	Copy data to internal buffer. More...
template<class Iterator >
void	extract (Iterator outputIt, size_type const count)
	Copy data from internal buffer. More...
template<typename T = void>
T *	data ()
	Direct access to the internal buffer. More...
template<typename T = void>
T const *	data () const
	Direct access to the internal buffer. More...
size_type	size () const
	Get the size of the internal buffer, and hence the maximum message size, in Byte. More...
template<typename T = unsigned char>
size_type	count () const
	Get the size of the internal buffer, and hence the maximum message size, in count of type T. More...
template<typename T >
void	execute (Operation op, size_type const count)
	Execute the global reduction using the supplied reduce function. Afterwards all process have the result stored in their local internal buffers. More...
template<typename T >
void	execute (void(*op)(T const *in, T *inout, size_type const count), size_type const count)
	Execute the global reduction using the user-defined reduce function. Afterwards all process have the result stored in their local internal buffers. More...
template<typename T >
void	executeBit (Operation op, size_type const count)
	Execute the global reduction using the supplied reduce function. Afterwards all process have the result stored in their local internal buffers. Results are guaranteed to be bit identical. More...
template<typename T >
void	executeBit (void(*op)(T const *in, T *inout, size_type const count), size_type const count)
	Execute the global reduction using the user-defined reduce function. Afterwards all process have the result stored in their local internal buffers. Results are guaranteed to be bit identical. More...

Constructor & Destructor Documentation

AllReduce	(	Communicator const &	communicator,
		size_type const	sizeByte
	)

Parameters

communicator	Communicator used in global reduction.
sizeByte	Reduce message size.

Note

The size might be chosen greater than the actual message size to allow re-use of the same environment for various messages sizes.

Member Function Documentation

size_type count

(

)

const

Returns

Actual size of the internal buffer in count of T.

T * data

(

)

Returns

Returns pointer to the internal buffer.

T const * data

(

)

const

Returns

Returns pointer to the internal buffer.

void execute	(	Operation	op,
		size_type const	count
	)

Parameters

op	Reduce function used in reduction operation.
count	Number of element of type T to be part of global reduction.

Warning

The size of all elements of type T to be part of global reduction might not exceed the size as returned by size().

The content of the internal buffer that was not required for the global reduction is undefined.

void execute	(	void(*)(T const *in, T *inout, size_type const count)	op,
		size_type const	count
	)

Parameters

op	Reduce function used in reduction operation.
count	Number of element of type T to be part of global reduction.

Note

Example op function:

void sum(float const *in, float *inout, pMR::size_type count)
{
    for(auto i = decltype(count){0}; i != count; ++i)
    {
       inout[i] += in[i];
    }
}

Warning

The size of all elements of type T to be part of global reduction might not exceed the size as returned by size().

The content of the internal buffer that was not required for the global reduction is undefined.

User-defined reduce functions might not be able to use best performing routines.

User-defined reduce functions are assumed to be associative and commutative.

void executeBit	(	Operation	op,
		size_type const	count
	)

Parameters

op	Reduce function used in reduction operation.
count	Number of element of type T to be part of global reduction.

Note

The bit identical reduction is in general slower. Only use when it is a requirement to have bit identical results.

Warning

The size of all elements of type T to be part of global reduction might not exceed the size as returned by size().

The content of the internal buffer that was not required for the global reduction is undefined.

void executeBit	(	void(*)(T const *in, T *inout, size_type const count)	op,
		size_type const	count
	)

Parameters

op	Reduce function used in reduction operation.
count	Number of element of type T to be part of global reduction.

Note

The bit identical reduction is in general slower. Only use when it is a requirement to have bit identical results.

Example op function:

void sum(float const *in, float *inout, pMR::size_type count)
{
    for(auto i = decltype(count){0}; i != count; ++i)
    {
       inout[i] += in[i];
    }
}

Warning

The size of all elements of type T to be part of global reduction might not exceed the size as returned by size().

The content of the internal buffer that was not required for the global reduction is undefined.

User-defined reduce functions might not be able to use best performing routines.

User-defined reduce functions are assumed to be associative and commutative.

void extract	(	Iterator	outputIt,
		size_type const	count
	)

Copies data from the internal buffer to the location specified by outputIt.

Parameters

outputIt	Iterator to destination.
count	Number of elements to copy.

void insert	(	Iterator	inputIt,
		size_type const	count
	)

Copies data from storage - inputIt iterator points to - to internal buffer.

Parameters

inputIt	Iterator to source.
count	Number of elements to copy.

size_type size

(

)

const

Returns

Actual size of the internal buffer in Byte.