#include <spctr_2d_general.h>

Inheritance diagram for CTF_int::spctr_2d_general:

Collaboration diagram for CTF_int::spctr_2d_general:

[legend]

Public Member Functions
void	print ()
	print ctr object More...

void	run (char A, int nblk_A, int64_t const size_blk_A, char B, int nblk_B, int64_t const size_blk_B, char C, int nblk_C, int64_t size_blk_C, char *&new_C)
	Basically doing SUMMA, except assumes equal block size on each processor. Performs rank-b updates where b is the smallest blocking factor among A and B or A and C or B and C. More...

int64_t	spmem_fp (double nnz_frac_A, double nnz_frac_B, double nnz_frac_C)
	returns the number of bytes of buffer space we need More...

int64_t	spmem_rec (double nnz_frac_A, double nnz_frac_B, double nnz_frac_C)
	returns the number of bytes of buffer space we need recursively More...

double	est_time_fp (int nlyr, double nnz_frac_A, double nnz_frac_B, double nnz_frac_C)
	returns the time this kernel will take including calls to rec_ctr More...

double	est_time_rec (int nlyr, double nnz_frac_A, double nnz_frac_B, double nnz_frac_C)
	returns the time this kernel will take including calls to rec_ctr More...

spctr *	clone ()

void	find_bsizes (int64_t &b_A, int64_t &b_B, int64_t &b_C, int64_t &s_A, int64_t &s_B, int64_t &s_C, int64_t &aux_size)
	determines buffer and block sizes needed for spctr_2d_general More...

	spctr_2d_general (spctr *other)
	copies spctr object More...

	~spctr_2d_general ()
	deallocs spctr_2d_general object More...

	spctr_2d_general (contraction *c)
	partial constructor, most of the logic is in the spctr_2d_gen_build function More...

Public Member Functions inherited from CTF_int::spctr
	~spctr ()

	spctr (spctr *other)

double	est_time_fp (int nlyr)

double	est_time_rec (int nlyr)

void	run (char A, char B, char *C)

	spctr (contraction const *c)

Public Member Functions inherited from CTF_int::ctr
virtual int64_t	mem_fp ()

virtual int64_t	mem_rec ()

virtual	~ctr ()
	deallocates generic ctr object More...

	ctr (ctr *other)
	copies generic ctr object More...

	ctr (contraction const *c)
	main constructor, defines variable based on contraction class More...

Data Fields
int	edge_len

int64_t	ctr_lda_A

int64_t	ctr_sub_lda_A

int64_t	ctr_lda_B

int64_t	ctr_sub_lda_B

int64_t	ctr_lda_C

int64_t	ctr_sub_lda_C

int64_t	dns_vrt_sz_A

int64_t	dns_vrt_sz_B

int64_t	dns_vrt_sz_C

bool	move_A

bool	move_B

bool	move_C

CommData *	cdt_A

CommData *	cdt_B

CommData *	cdt_C

spctr *	rec_ctr

Data Fields inherited from CTF_int::spctr
bool	is_sparse_A

bool	is_sparse_B

bool	is_sparse_C

char *	new_C

Data Fields inherited from CTF_int::ctr
algstrct const *	sr_A

algstrct const *	sr_B

algstrct const *	sr_C

char const *	beta

int	num_lyr

int	idx_lyr

Detailed Description

Definition at line 47 of file spctr_2d_general.h.

Constructor & Destructor Documentation

CTF_int::spctr_2d_general::spctr_2d_general ( spctr * other )

copies spctr object

Definition at line 19 of file spctr_2d_general.cxx.

References cdt_A, cdt_B, cdt_C, CTF_int::spctr::clone(), ctr_lda_A, ctr_lda_B, ctr_lda_C, ctr_sub_lda_A, ctr_sub_lda_B, ctr_sub_lda_C, dns_vrt_sz_A, dns_vrt_sz_B, dns_vrt_sz_C, edge_len, move_A, move_B, move_C, and rec_ctr.

Referenced by clone().

CTF_int::spctr_2d_general::~spctr_2d_general ( )

deallocs spctr_2d_general object

Definition at line 11 of file spctr_2d_general.cxx.

References rec_ctr.

CTF_int::spctr_2d_general::spctr_2d_general ( contraction * c )

inline

partial constructor, most of the logic is in the spctr_2d_gen_build function

Parameters

[in] c contraction object to get info about spctr from

Definition at line 150 of file spctr_2d_general.h.

Member Function Documentation

spctr * CTF_int::spctr_2d_general::clone ( )

virtual

Reimplemented from CTF_int::spctr.

Definition at line 64 of file spctr_2d_general.cxx.

References spctr_2d_general().

double CTF_int::spctr_2d_general::est_time_fp	(	int	nlyr,
		double	nnz_frac_A,
		double	nnz_frac_B,
		double	nnz_frac_C
	)

virtual

returns the time this kernel will take including calls to rec_ctr

Returns: seconds needed for recursive contraction

Reimplemented from CTF_int::spctr.

Definition at line 92 of file spctr_2d_general.cxx.

References CTF_int::algstrct::addmop(), cdt_A, cdt_B, cdt_C, dns_vrt_sz_A, dns_vrt_sz_B, dns_vrt_sz_C, edge_len, CTF_int::accumulatable::el_size, CTF_int::CommData::estimate_bcast_time(), CTF_int::algstrct::estimate_csr_red_time(), CTF_int::CommData::estimate_red_time(), find_bsizes(), CTF_int::spctr::is_sparse_A, CTF_int::spctr::is_sparse_B, CTF_int::spctr::is_sparse_C, MIN, move_A, move_B, move_C, CTF_int::algstrct::pair_size(), CTF_int::ctr::sr_A, CTF_int::ctr::sr_B, and CTF_int::ctr::sr_C.

Referenced by est_time_rec().

double CTF_int::spctr_2d_general::est_time_rec	(	int	nlyr,
		double	nnz_frac_A,
		double	nnz_frac_B,
		double	nnz_frac_C
	)

virtual

returns the time this kernel will take including calls to rec_ctr

Returns: seconds needed for recursive contraction

Reimplemented from CTF_int::spctr.

Definition at line 117 of file spctr_2d_general.cxx.

References edge_len, est_time_fp(), CTF_int::spctr::est_time_rec(), MIN, and rec_ctr.

void CTF_int::spctr_2d_general::find_bsizes	(	int64_t &	b_A,
		int64_t &	b_B,
		int64_t &	b_C,
		int64_t &	s_A,
		int64_t &	s_B,
		int64_t &	s_C,
		int64_t &	aux_size
	)

determines buffer and block sizes needed for spctr_2d_general

Parameters

[out]	b_A	block size of A if its communicated, 0 otherwise
[out]	b_B	block size of A if its communicated, 0 otherwise
[out]	b_C	block size of A if its communicated, 0 otherwise
[out]	s_A	total size of A if its communicated, 0 otherwise
[out]	s_B	total size of B if its communicated, 0 otherwise
[out]	s_C	total size of C if its communicated, 0 otherwise
[out]	aux_size	size of auxillary buffer needed

Definition at line 68 of file spctr_2d_general.cxx.

References cdt_A, cdt_B, cdt_C, ctr_lda_A, ctr_lda_B, ctr_lda_C, ctr_sub_lda_A, ctr_sub_lda_B, ctr_sub_lda_C, edge_len, CTF_int::accumulatable::el_size, MAX, move_A, move_B, move_C, CTF_int::CommData::np, CTF_int::ctr::sr_A, CTF_int::ctr::sr_B, and CTF_int::ctr::sr_C.

Referenced by est_time_fp(), run(), and spmem_fp().

void CTF_int::spctr_2d_general::print ( )

virtual

print ctr object

Reimplemented from CTF_int::ctr.

Definition at line 44 of file spctr_2d_general.cxx.

References cdt_A, cdt_B, cdt_C, ctr_lda_A, ctr_lda_B, ctr_lda_C, ctr_sub_lda_A, ctr_sub_lda_B, ctr_sub_lda_C, edge_len, move_A, move_B, move_C, CTF_int::CommData::np, CTF_int::ctr::print(), and rec_ctr.

void CTF_int::spctr_2d_general::run	(	char *	A,
		int	nblk_A,
		int64_t const *	size_blk_A,
		char *	B,
		int	nblk_B,
		int64_t const *	size_blk_B,
		char *	C,
		int	nblk_C,
		int64_t *	size_blk_C,
		char *&	new_C
	)

virtual

Basically doing SUMMA, except assumes equal block size on each processor. Performs rank-b updates where b is the smallest blocking factor among A and B or A and C or B and C.

Reimplemented from CTF_int::spctr.

Definition at line 365 of file spctr_2d_general.cxx.

References CTF_int::alloc(), CTF_int::alloc_ptr(), ASSERT, CTF_int::bcast_step(), CTF_int::ctr::beta, CTF_int::cdealloc(), cdt_A, cdt_B, cdt_C, CTF_int::algstrct::csr_add(), ctr_lda_A, ctr_lda_B, ctr_lda_C, ctr_sub_lda_A, ctr_sub_lda_B, ctr_sub_lda_C, edge_len, CTF_int::accumulatable::el_size, find_bsizes(), CTF_int::host_pinned_alloc(), CTF_int::host_pinned_free(), CTF_int::ctr::idx_lyr, CTF_int::spctr::is_sparse_A, CTF_int::spctr::is_sparse_B, CTF_int::spctr::is_sparse_C, move_A, move_B, move_C, CTF_int::mst_alloc_ptr(), CTF_int::CommData::np, CTF_int::ctr::num_lyr, CTF_int::CommData::rank, rec_ctr, CTF_int::reduce_step_post(), CTF_int::reduce_step_pre(), CTF_int::spctr::run(), CTF_int::spcopy(), CTF_int::ctr::sr_A, CTF_int::ctr::sr_B, CTF_int::ctr::sr_C, TAU_FSTART, and TAU_FSTOP.