#include "dbm_multiply_comm.h"
#include "../mpiwrap/cp_mpi.h"
#include "../offload/offload_mempool.h"
#include <assert.h>
#include <limits.h>
#include <stdlib.h>
#include <string.h>

Data Structures
struct	plan_t
	Private struct used for planing during pack_matrix. More...

Macros
#define	DBM_MULTIPLY_COMM_MEMPOOL

Functions
static int	gcd (const int a, const int b)
	Private routine for computing greatest common divisor of two numbers.
static int	lcm (const int a, const int b)
	Private routine for computing least common multiple of two numbers.
static int	checked_byte_count (const int nelements, const size_t element_size)
	Private routine for converting element counts to byte counts.
static int	isum (const int n, const int input[n])
	Private routine for computing the sum of the given integers.
static void	icumsum (const int n, const int input[n], int output[n])
	Private routine for computing the cumulative sums of given numbers.
static void	compute_data_recv_count (const int nranks, const int blks_recv_count[nranks], const int blks_recv_displ[nranks], const int free_index_sizes[], const int sum_index_sizes[], const dbm_pack_block_t blks_recv[], int data_recv_count[nranks])
	Private routine computing received data counts from block metadata.
static unsigned long long	calculate_tick_index (int sum_index, int nticks)
	Private routine for calculating tick indices in pack plans.
static void	create_pack_plans (const bool trans_matrix, const bool trans_dist, const dbm_matrix_t matrix, const cp_mpi_comm_t comm, const dbm_dist_1d_t dist_indices, const dbm_dist_1d_t dist_ticks, const int nticks, const int npacks, plan_t plans_per_pack[npacks], int nblks_per_pack[npacks], int ndata_per_pack[npacks])
	Private routine for planing packs.
static void	fill_send_buffers (const dbm_matrix_t *matrix, const bool trans_matrix, const int nblks_send, const int ndata_send, plan_t plans[nblks_send], const int nranks, int blks_send_count[nranks], int data_send_count[nranks], int blks_send_displ[nranks], int data_send_displ[nranks], dbm_pack_block_t blks_send[nblks_send], double data_send[ndata_send])
	Private routine for filling send buffers.
static int	compare_pack_blocks_by_sum_index (const void a, const void b)
	Private comperator passed to qsort to compare two blocks by sum_index.
static void	postprocess_received_blocks (const int nranks, const int nshards, const int nblocks_recv, const int blks_recv_count[nranks], const int blks_recv_displ[nranks], const int data_recv_displ[nranks], dbm_pack_block_t blks_recv[nblocks_recv])
	Private routine for post-processing received blocks.
static dbm_packed_matrix_t	pack_matrix (const bool trans_matrix, const bool trans_dist, const dbm_matrix_t restrict matrix, const dbm_distribution_t restrict dist, const int nticks)
	Private routine for redistributing a matrix along selected dimensions.
static dbm_pack_t *	sendrecv_pack (const int itick, const int nticks, dbm_packed_matrix_t *packed)
	Private routine for sending and receiving the pack for the given tick.
static void	free_packed_matrix (dbm_packed_matrix_t *packed)
	Private routine for releasing a packed matrix.
dbm_comm_iterator_t *	dbm_comm_iterator_start (const bool transa, const bool transb, const dbm_matrix_t matrix_a, const dbm_matrix_t matrix_b, const dbm_matrix_t *matrix_c)
	Internal routine for creating a communication iterator.
bool	dbm_comm_iterator_next (dbm_comm_iterator_t iter, dbm_pack_t pack_a, dbm_pack_t *pack_b)
	Internal routine for retrieving next pair of packs of given iterator.
void	dbm_comm_iterator_stop (dbm_comm_iterator_t *iter)
	Internal routine for releasing the given communication iterator.

Macro Definition Documentation

◆ DBM_MULTIPLY_COMM_MEMPOOL

#define DBM_MULTIPLY_COMM_MEMPOOL

Definition at line 17 of file dbm_multiply_comm.c.

Function Documentation

◆ gcd()

int gcd	(	const int	a,
		const int	b )

static

Private routine for computing greatest common divisor of two numbers.

Author: Ole Schuett

Definition at line 24 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ lcm()

int lcm	(	const int	a,
		const int	b )

static

Private routine for computing least common multiple of two numbers.

Author: Ole Schuett

Definition at line 35 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ checked_byte_count()

int checked_byte_count	(	const int	nelements,
		const size_t	element_size )

static

Private routine for converting element counts to byte counts.

Author: Hans Pabst

Definition at line 41 of file dbm_multiply_comm.c.

Here is the caller graph for this function:

◆ isum()

int isum	(	const int	n,
		const int	input[n] )

inlinestatic

Private routine for computing the sum of the given integers.

Author: Ole Schuett

Definition at line 54 of file dbm_multiply_comm.c.

Here is the caller graph for this function:

◆ icumsum()

void icumsum	(	const int	n,
		const int	input[n],
		int	output[n] )

inlinestatic

Private routine for computing the cumulative sums of given numbers.

Author: Ole Schuett and Hans Pabst

Definition at line 66 of file dbm_multiply_comm.c.

Here is the caller graph for this function:

◆ compute_data_recv_count()

void compute_data_recv_count	(	const int	nranks,
		const int	blks_recv_count[nranks],
		const int	blks_recv_displ[nranks],
		const int	free_index_sizes[],
		const int	sum_index_sizes[],
		const dbm_pack_block_t	blks_recv[],
		int	data_recv_count[nranks] )

static

Private routine computing received data counts from block metadata.

Author: Hans Pabst

Definition at line 80 of file dbm_multiply_comm.c.

Here is the caller graph for this function:

◆ calculate_tick_index()

unsigned long long calculate_tick_index	(	int	sum_index,
		int	nticks )

inlinestatic

Private routine for calculating tick indices in pack plans.

Author: Maximilian Graml

Definition at line 118 of file dbm_multiply_comm.c.

Here is the caller graph for this function:

◆ create_pack_plans()

void create_pack_plans	(	const bool	trans_matrix,
		const bool	trans_dist,
		const dbm_matrix_t *	matrix,
		const cp_mpi_comm_t	comm,
		const dbm_dist_1d_t *	dist_indices,
		const dbm_dist_1d_t *	dist_ticks,
		const int	nticks,
		const int	npacks,
		plan_t *	plans_per_pack[npacks],
		int	nblks_per_pack[npacks],
		int	ndata_per_pack[npacks] )

static

Private routine for planing packs.

Author: Ole Schuett

Definition at line 128 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ fill_send_buffers()

void fill_send_buffers	(	const dbm_matrix_t *	matrix,
		const bool	trans_matrix,
		const int	nblks_send,
		const int	ndata_send,
		plan_t	plans[nblks_send],
		const int	nranks,
		int	blks_send_count[nranks],
		int	data_send_count[nranks],
		int	blks_send_displ[nranks],
		int	data_send_displ[nranks],
		dbm_pack_block_t	blks_send[nblks_send],
		double	data_send[ndata_send] )

static

Private routine for filling send buffers.

Author: Ole Schuett

Definition at line 210 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ compare_pack_blocks_by_sum_index()

int compare_pack_blocks_by_sum_index	(	const void *	a,
		const void *	b )

static

Private comperator passed to qsort to compare two blocks by sum_index.

Author: Ole Schuett

Definition at line 306 of file dbm_multiply_comm.c.

Here is the caller graph for this function:

◆ postprocess_received_blocks()

void postprocess_received_blocks	(	const int	nranks,
		const int	nshards,
		const int	nblocks_recv,
		const int	blks_recv_count[nranks],
		const int	blks_recv_displ[nranks],
		const int	data_recv_displ[nranks],
		dbm_pack_block_t	blks_recv[nblocks_recv] )

static

Private routine for post-processing received blocks.

Author: Ole Schuett

Definition at line 316 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ pack_matrix()

dbm_packed_matrix_t pack_matrix	(	const bool	trans_matrix,
		const bool	trans_dist,
		const dbm_matrix_t *restrict	matrix,
		const dbm_distribution_t *restrict	dist,
		const int	nticks )

static

Private routine for redistributing a matrix along selected dimensions.

Author: Ole Schuett

Definition at line 379 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ sendrecv_pack()

dbm_pack_t * sendrecv_pack	(	const int	itick,
		const int	nticks,
		dbm_packed_matrix_t *	packed )

static

Private routine for sending and receiving the pack for the given tick.

Author: Ole Schuett

Definition at line 519 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ free_packed_matrix()

void free_packed_matrix ( dbm_packed_matrix_t * packed )

static

Private routine for releasing a packed matrix.

Author: Ole Schuett

Definition at line 577 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ dbm_comm_iterator_start()

dbm_comm_iterator_t * dbm_comm_iterator_start	(	const bool	transa,
		const bool	transb,
		const dbm_matrix_t *	matrix_a,
		const dbm_matrix_t *	matrix_b,
		const dbm_matrix_t *	matrix_c )

Internal routine for creating a communication iterator.

Author: Ole Schuett

Definition at line 599 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ dbm_comm_iterator_next()

bool dbm_comm_iterator_next	(	dbm_comm_iterator_t *	iter,
		dbm_pack_t **	pack_a,
		dbm_pack_t **	pack_b )

Internal routine for retrieving next pair of packs of given iterator.

Internal routine for retrieving next pair of packs.

Author: Ole Schuett

Definition at line 627 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

◆ dbm_comm_iterator_stop()

void dbm_comm_iterator_stop ( dbm_comm_iterator_t * iter )

Internal routine for releasing the given communication iterator.

Author: Ole Schuett

Definition at line 647 of file dbm_multiply_comm.c.

Here is the call graph for this function:

Here is the caller graph for this function:

Data Structures

Macros

Functions

Macro Definition Documentation

◆ DBM_MULTIPLY_COMM_MEMPOOL

Function Documentation

◆ gcd()

◆ lcm()

◆ checked_byte_count()

◆ isum()

◆ icumsum()

◆ compute_data_recv_count()

◆ calculate_tick_index()

◆ create_pack_plans()

◆ fill_send_buffers()

◆ compare_pack_blocks_by_sum_index()

◆ postprocess_received_blocks()

◆ pack_matrix()

◆ sendrecv_pack()

◆ free_packed_matrix()

◆ dbm_comm_iterator_start()

◆ dbm_comm_iterator_next()

◆ dbm_comm_iterator_stop()