#include <ctr_2d_general.h>

Inheritance diagram for CTF_int::ctr_2d_general:

Collaboration diagram for CTF_int::ctr_2d_general:

[legend]

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

void	run (char A, char B, char *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	mem_fp ()
	returns the number of bytes of buffer space we need More...

int64_t	mem_rec ()
	returns the number of bytes of buffer space we need recursively More...

double	est_time_fp (int nlyr)
	returns the number of bytes this kernel will send per processor More...

double	est_time_rec (int nlyr)
	returns the number of bytes send by each proc recursively More...

ctr *	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 ctr_2d_general More...

	ctr_2d_general (ctr *other)
	copies ctr object More...

	~ctr_2d_general ()
	deallocs ctr_2d_general object More...

	ctr_2d_general (contraction *c)
	partial constructor, most of the logic is in the ctr_2d_gen_build function More...

Public Member Functions inherited from CTF_int::ctr
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

bool	move_A

bool	move_B

bool	move_C

CommData *	cdt_A

CommData *	cdt_B

CommData *	cdt_C

ctr *	rec_ctr

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 ctr_2d_general.h.

Constructor & Destructor Documentation

CTF_int::ctr_2d_general::ctr_2d_general ( ctr * other )

copies ctr object

Definition at line 179 of file ctr_2d_general.cxx.

References cdt_A, cdt_B, cdt_C, CTF_int::ctr::clone(), 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, and rec_ctr.

Referenced by clone().

CTF_int::ctr_2d_general::~ctr_2d_general ( )

deallocs ctr_2d_general object

Definition at line 171 of file ctr_2d_general.cxx.

References rec_ctr.

CTF_int::ctr_2d_general::ctr_2d_general ( contraction * c )

inline

partial constructor, most of the logic is in the ctr_2d_gen_build function

Parameters

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

Definition at line 137 of file ctr_2d_general.h.

Member Function Documentation

ctr * CTF_int::ctr_2d_general::clone ( )

virtual

Reimplemented from CTF_int::ctr.

Definition at line 220 of file ctr_2d_general.cxx.

References ctr_2d_general().

double CTF_int::ctr_2d_general::est_time_fp ( int nlyr )

virtual

returns the number of bytes this kernel will send per processor

Returns: bytes sent

Reimplemented from CTF_int::ctr.

Definition at line 248 of file ctr_2d_general.cxx.

References CTF_int::algstrct::addmop(), cdt_A, cdt_B, cdt_C, edge_len, CTF_int::accumulatable::el_size, CTF_int::CommData::estimate_bcast_time(), CTF_int::CommData::estimate_red_time(), find_bsizes(), MIN, move_A, move_B, move_C, CTF_int::ctr::sr_A, CTF_int::ctr::sr_B, and CTF_int::ctr::sr_C.

Referenced by est_time_rec().

double CTF_int::ctr_2d_general::est_time_rec ( int nlyr )

virtual

returns the number of bytes send by each proc recursively

Returns: bytes needed for recursive contraction

Reimplemented from CTF_int::ctr.

Definition at line 261 of file ctr_2d_general.cxx.

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

void CTF_int::ctr_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 ctr_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 224 of file ctr_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(), mem_fp(), and run().

int64_t CTF_int::ctr_2d_general::mem_fp ( )

virtual

returns the number of bytes of buffer space we need

Returns: bytes needed

Reimplemented from CTF_int::ctr.

Definition at line 265 of file ctr_2d_general.cxx.

References CTF_int::accumulatable::el_size, find_bsizes(), CTF_int::ctr::sr_A, CTF_int::ctr::sr_B, and CTF_int::ctr::sr_C.

Referenced by mem_rec().

int64_t CTF_int::ctr_2d_general::mem_rec ( )

virtual

returns the number of bytes of buffer space we need recursively

Returns: bytes needed for recursive contraction

Reimplemented from CTF_int::ctr.

Definition at line 271 of file ctr_2d_general.cxx.

References mem_fp(), CTF_int::ctr::mem_rec(), and rec_ctr.

void CTF_int::ctr_2d_general::print ( )

virtual

print ctr object

Reimplemented from CTF_int::ctr.

Definition at line 200 of file ctr_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::ctr_2d_general::run	(	char *	A,
		char *	B,
		char *	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::ctr.

Definition at line 275 of file ctr_2d_general.cxx.

References CTF_int::algstrct::addid(), CTF_int::algstrct::addmop(), ASSERT, CTF_int::CommData::bcast(), CTF_int::ctr::beta, CTF_int::cdealloc(), cdt_A, cdt_B, cdt_C, CTF_int::algstrct::copy(), 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::algstrct::mdtype(), move_A, move_B, move_C, CTF_int::mst_alloc_ptr(), CTF_int::algstrct::mulid(), CTF_int::CommData::np, CTF_int::ctr::num_lyr, CTF_int::CommData::rank, rec_ctr, CTF_int::CommData::red(), CTF_int::ctr::run(), CTF_int::ctr::sr_A, CTF_int::ctr::sr_B, CTF_int::ctr::sr_C, TAU_FSTART, and TAU_FSTOP.