ReductionVector< T > Class Template Reference

#include <reductionvector.h>

Public Member Functions
ReductionVector &	allreduce (bool b=true)
	allreduce(bool) turns allreduce on or off. By default on.
T *	data ()
	data() returns ptr to the raw storage
ReductionVector &	delayed (bool b=true)
	deferred(bool) turns deferred on or off. By default turns on.
void	init_product ()
	Init is to be called before every site loop.
void	init_sum ()
	Init is to be called before every site loop.
ReductionVector &	nonblocking (bool b=true)
	nonblocking(bool) turns allreduce on or off. By default on.
template<typename S , std::enable_if_t< std::is_assignable< T &, S >::value, int > = 0>
void	operator= (const S &rhs)
void	operator= (std::nullptr_t np)
	Assignment from 0.
T &	operator[] (const int i)
	And access operators - these do in practice everything already!
void	reduce ()
	Complete non-blocking or delayed reduction.
void	reduce_product ()
void	reduce_sum ()
	ReductionVector ()=default
size_t	size () const
	methods from std::vector:
void	start_reduce ()
	For delayed reduction, reduce starts or completes the reduction operation.
void	wait ()
	Wait for MPI to complete, if it is currently going on.
	~ReductionVector ()
	Destructor cleans up communications if they are in progress.

Detailed Description

template<typename T>
class ReductionVector< T >

Special reduction class for arrays: declare a reduction array as ReductionVector<T> a(size);

This can be used within site loops as onsites(ALL ) { int i = ... a[i] += ... }

or outside site loops as usual. Reductions: += *= sum or product reduction NOTE: size of the reduction vector must be same on all nodes! By default reduction is "allreduce", i.e. all nodes get the same result.

Reduction can be customized by using methods: a.size() : return the size of the array a.resize(new_size) : change size of array (all nodes must use same size!) a[i] : get/set element i

a.allreduce(bool) : turn allreduce on/off (default=true) a.nonblocking(bool) : turn non-blocking reduction on/off a.delayed(bool) : turn delayed reduction on/off

a.wait() : for non-blocking reduction, wait() has to be called after the loop to complete the reduction a.reduce() : for delayed reduction starts/completes the reduction

a.is_allreduce() : queries the reduction status is_nonblocking() is_delayed()

a.push_back(element) : add one element to The same reduction variable can be used again

Definition at line 44 of file reductionvector.h.

Constructor & Destructor Documentation

ReductionVector()

template<typename T >

ReductionVector< T >::ReductionVector ( )

explicitdefault

Initialize to zero by default (? exception to other variables) allreduce = true by default

Member Function Documentation

operator=()

template<typename T >

template<typename S , std::enable_if_t< std::is_assignable< T &, S >::value, int > = 0>

void ReductionVector< T >::operator= ( const S &  rhs )

inline

Assignment is used only outside site loops - wait for comms if needed Make this return void, hard to imagine it is used for anything useful

Definition at line 184 of file reductionvector.h.

reduce_product()

template<typename T >

void ReductionVector< T >::reduce_product ( )

inline

Product reduction – currently works only for scalar data types. For Complex, Matrix and Vector data product is not element-wise. TODO: Array or std::array ? TODO: implement using custom MPI ops (if needed)

Definition at line 233 of file reductionvector.h.

reduce_sum()

template<typename T >

void ReductionVector< T >::reduce_sum ( )

inline

Start sum reduction – works only if the type T addition == element-wise addition. This is true for all hila predefined data types

Definition at line 217 of file reductionvector.h.

---

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

• reductionvector.h