RDScheme< T, DS > Class Template Reference

Full parallel row-distributed SpMV, based on CSB (Morton curve + Cilk) and PThreads. More...

#include <RDScheme.hpp>

Inheritance diagram for RDScheme< T, DS >:

Collaboration diagram for RDScheme< T, DS >:

Public Member Functions
	RDScheme (const std::string file, T zero)
	Base constructor. More...
	RDScheme (std::vector< Triplet< T > > &input, ULI m, ULI n, T zero)
	Base constructor. More...
virtual	~RDScheme ()
	Base deconstructor. More...
void	wait ()
	Lets the calling thread wait for the end of the SpMV multiply.
virtual void	load (std::vector< Triplet< T > > &input, const ULI m, const ULI n, const T zero)
	Loads a sparse matrix from an input set of triplets. More...
virtual T *	mv (const T *x)
	Overloaded mv call; allocates output vector using numa_interleaved. More...
virtual void	zxa (const T *x, T *z)
virtual void	zxa (const T *x, T *z, const unsigned long int repeat)
virtual void	zax (const T *x, T *z)
	See SparseMatrix::zax.
virtual void	zax (const T *x, T *z, const unsigned long int repeat, const clockid_t clock_id, double *elapsed_time)
	See SparseMatrix::zax.
virtual size_t	bytesUsed ()
virtual void	getFirstIndexPair (ULI &i, ULI &j)
	Function disabled for parallel schemes! More...
Public Member Functions inherited from SparseMatrix< T, ULI >
	SparseMatrix ()
	Base constructor. More...
	SparseMatrix (const ULInzs, const ULInr, const ULInc, const T zero)
	Base constructor. More...
virtual	~SparseMatrix ()
	Base deconstructor. More...
void	loadFromFile (const std::string file, const T zero=0)
	Function which loads a matrix from a matrix market file. More...
virtual unsigned long int	m ()
	Queries the number of rows this matrix contains. More...
virtual unsigned long int	n ()
	Queries the number of columns this matrix contains. More...
virtual unsigned long int	nzs ()
	Queries the number of nonzeroes stored in this matrix. More...
virtual void	zax (const T *__restrict__ x, T *__restrict__ z)=0
	In-place z=Ax function. More...
virtual void	zxa (const T *__restrict__ x, T *__restrict__ z)=0
	In-place z=xA function. More...
Public Member Functions inherited from Matrix< T >
	Matrix ()
	Base constructor. More...
virtual	~Matrix ()
	Base deconstructor. More...
virtual void	zax (const T *__restrict__ x, T *__restrict__ z, const size_t k, const clockid_t clock_id=0, double *elapsed_time=NULL)
	Wrapper function to call the zax kernel multiple times successively, while timing the duration of the operation. More...
template<size_t k>
void	ZaX (const T *__restrict__ const *__restrict__ const X, T *__restrict__ const *__restrict__ const Z)
	In-place Z=AX function, where A is m x n, Z = m x k, and X is n x k. More...
template<size_t k>
void	ZXa (const T *__restrict__ const *__restrict__ const X, T *__restrict__ const *__restrict__ const Z)
	In-place Z=XA function, where A is m x n, Z = k x n, and X is k x m. More...
virtual void	zxa (const T *__restrict__ x, T *__restrict__ z, const unsigned long int repeat, const clockid_t clock_id=0, double *elapsed_time=NULL)
	Wrapper function to call the zxa kernel multiple times successively, while timing the operation duration. More...

Static Public Member Functions
static void	end (pthread_mutex_t *mutex, pthread_cond_t *cond, size_t *sync, const size_t P)
	End synchronisation code. More...
static void	synchronise (pthread_mutex_t *mutex, pthread_cond_t *cond, size_t *sync, const size_t P)
	Synchronises all threads. More...
static void *	thread (void *data)
	SPMD code for each thread involved with parallel SpMV multiplication. More...
static void	collectY (RDScheme_shared_data< T > *shared)
	Reduces a distributed output vector set into a single contiguous output vector at process 0. More...

Protected Attributes
pthread_t *	threads
	p_threads associated to this data strcuture
RDScheme_shared_data< T > *	thread_data
	array of initial thread data
pthread_mutex_t	mutex
	Stop/continue mechanism: mutex.
pthread_cond_t	cond
	Stop/continue mechanism: condition.
pthread_mutex_t	end_mutex
	Wait for end mechanism: mutex.
pthread_cond_t	end_cond
	Wait for end mechanism: condition.
size_t	sync
	Used for synchronising threads.
size_t	end_sync
	Used for construction end signal.
Protected Attributes inherited from SparseMatrix< T, ULI >
ULI	nor
	Number of rows. More...
ULI	noc
	Number of columns.
ULI	nnz
	Number of non-zeros. More...

Static Protected Attributes
static size_t	P = 0
	Number of threads to fire up.
static const T *	input = NULL
	Input vector.
static T *	output = NULL
	Output vector.
static clockid_t	global_clock_id = 0
	Clock type used for thread-local timing.

Additional Inherited Members
Public Attributes inherited from SparseMatrix< T, ULI >
T	zero_element
	The element considered to be zero. More...

Detailed Description

template<typename T, typename DS>
class RDScheme< T, DS >

Full parallel row-distributed SpMV, based on CSB (Morton curve + Cilk) and PThreads.

Inspired by Aydin & Gilbert's CSB, and comments by Patrick Amestoy on the BICRS Hilbert scheme.

Constructor & Destructor Documentation

template<typename T, typename DS >

RDScheme< T, DS >::RDScheme ( const std::string  file, T  zero  )

inline

Base constructor.

Reads input from file.

References SparseMatrix< T, ULI >::loadFromFile().

template<typename T, typename DS >

RDScheme< T, DS >::RDScheme ( std::vector< Triplet< T > > &  input, ULI  m, ULI  n, T  zero  )

inline

Base constructor.

Reads input from a set of triplets.

References RDScheme< T, DS >::input, and RDScheme< T, DS >::load().

template<typename T, typename DS >

virtual RDScheme< T, DS >::~RDScheme ( )

inlinevirtual

Base deconstructor.

References RDScheme< T, DS >::cond, RDScheme< T, DS >::mutex, RDScheme< T, DS >::P, RDScheme< T, DS >::thread_data, and RDScheme< T, DS >::threads.

Member Function Documentation

template<typename T, typename DS >

virtual size_t RDScheme< T, DS >::bytesUsed ( )

inlinevirtual

Returns

The memory usage of this data structure (summed over all threads).

Implements Matrix< T >.

References RDScheme< T, DS >::P, and RDScheme< T, DS >::thread_data.

template<typename T, typename DS >

static void RDScheme< T, DS >::collectY ( RDScheme_shared_data< T > *  shared )

inlinestatic

Reduces a distributed output vector set into a single contiguous output vector at process 0.

Parameters

shared

Which set of output vectors to reduce.

References RDScheme_shared_data< T >::id, RDScheme_shared_data< T >::local_y, RDScheme_shared_data< T >::output_vector_offset, and RDScheme_shared_data< T >::output_vector_size.

Referenced by RDScheme< T, DS >::thread().

template<typename T, typename DS >

static void RDScheme< T, DS >::end ( pthread_mutex_t *  mutex, pthread_cond_t *  cond, size_t *  sync, const size_t  P  )

inlinestatic

End synchronisation code.

Referenced by RDScheme< T, DS >::thread().

template<typename T, typename DS >

virtual void RDScheme< T, DS >::getFirstIndexPair ( ULI &  i, ULI &  j  )

inlinevirtual

Function disabled for parallel schemes!

See also

SparseMatrix::getFirstIndexPair

Implements SparseMatrix< T, ULI >.

template<typename T, typename DS >

virtual void RDScheme< T, DS >::load ( std::vector< Triplet< T > > &  input, const ULI  m, const ULI  n, const T  zero  )

inlinevirtual

Loads a sparse matrix from an input set of triplets.

Parameters

input	The input set of triplets.
m	The number of rows.
n	The number of columns.
zero	What constitutes a zero in this sparse matrix instance.

Implements SparseMatrix< T, ULI >.

References RDScheme< T, DS >::cond, MachineInfo::cores(), RDScheme< T, DS >::end_cond, RDScheme< T, DS >::end_mutex, RDScheme< T, DS >::end_sync, MachineInfo::getInstance(), RDScheme< T, DS >::input, SparseMatrix< T, ULI >::m(), RDScheme< T, DS >::mutex, SparseMatrix< T, ULI >::n(), SparseMatrix< T, ULI >::nnz, SparseMatrix< T, ULI >::noc, SparseMatrix< T, ULI >::nor, RDScheme< T, DS >::P, RDScheme< T, DS >::sync, RDScheme< T, DS >::thread(), RDScheme< T, DS >::thread_data, RDScheme< T, DS >::threads, RDScheme< T, DS >::wait(), and SparseMatrix< T, ULI >::zero_element.

Referenced by RDScheme< T, DS >::RDScheme().

template<typename T, typename DS >

virtual T* RDScheme< T, DS >::mv ( const T *  x )

inlinevirtual

Overloaded mv call; allocates output vector using numa_interleaved.

Reimplemented from SparseMatrix< T, ULI >.

References SparseMatrix< T, ULI >::nor, RDScheme< T, DS >::zax(), and SparseMatrix< T, ULI >::zero_element.

template<typename T, typename DS >

static void RDScheme< T, DS >::synchronise ( pthread_mutex_t *  mutex, pthread_cond_t *  cond, size_t *  sync, const size_t  P  )

inlinestatic

Synchronises all threads.

Referenced by RDScheme< T, DS >::thread().

template<typename T, typename DS >

static void* RDScheme< T, DS >::thread ( void *  data )

inlinestatic

SPMD code for each thread involved with parallel SpMV multiplication.

References RDScheme_shared_data< T >::bytes, RDScheme< T, DS >::collectY(), RDScheme_shared_data< T >::cond, RDScheme< T, DS >::cond, RDScheme< T, DS >::end(), RDScheme_shared_data< T >::end_cond, RDScheme_shared_data< T >::end_mutex, RDScheme_shared_data< T >::end_sync, RDScheme< T, DS >::global_clock_id, RDScheme_shared_data< T >::id, RDScheme_shared_data< T >::local_y, SparseMatrix< T, ULI >::m(), RDScheme_shared_data< T >::mode, RDScheme_shared_data< T >::mutex, RDScheme< T, DS >::mutex, SparseMatrix< T, ULI >::n(), SparseMatrix< T, ULI >::nnz, RDScheme_shared_data< T >::nzb, RDScheme_shared_data< T >::original, RDScheme< T, DS >::output, RDScheme_shared_data< T >::output_vector_offset, RDScheme_shared_data< T >::output_vector_size, RDScheme_shared_data< T >::P, RDScheme< T, DS >::P, RDScheme_shared_data< T >::repeat, RDScheme_shared_data< T >::sync, RDScheme< T, DS >::synchronise(), and RDScheme_shared_data< T >::time.

Referenced by RDScheme< T, DS >::load().

template<typename T, typename DS >

virtual void RDScheme< T, DS >::zxa ( const T *  x, T *  z  )

inlinevirtual

See also

SparseMatrix::zxa

template<typename T, typename DS >

virtual void RDScheme< T, DS >::zxa ( const T *  x, T *  z, const unsigned long int  repeat  )

inlinevirtual

See also

SparseMatrix::zxa

References RDScheme< T, DS >::cond, RDScheme< T, DS >::end_mutex, RDScheme< T, DS >::mutex, RDScheme< T, DS >::P, RDScheme< T, DS >::thread_data, and RDScheme< T, DS >::wait().

---

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

• RDScheme.hpp