pslaset.c File Reference

#include "common.h"

Include dependency graph for pslaset.c:

Go to the source code of this file.

Macros
#define	A(m, n)   BLKADDR(A, float, m, n)

Functions
void	plasma_pslaset_quark (PLASMA_enum uplo, float alpha, float beta, PLASMA_desc A, PLASMA_sequence *sequence, PLASMA_request *request)

---

Detailed Description

PLASMA auxiliary routines PLASMA is a software package provided by Univ. of Tennessee, Univ. of California Berkeley and Univ. of Colorado Denver

Version:

2.4.5

Author:

Hatem Ltaief

Date:

2010-11-15 s Tue Nov 22 14:35:41 2011

Definition in file pslaset.c.

---

Macro Definition Documentation

#define A	(		m,
			n
	)		BLKADDR(A, float, m, n)

Definition at line 17 of file pslaset.c.

---

Function Documentation

void plasma_pslaset_quark	(	PLASMA_enum	uplo,
		float	alpha,
		float	beta,
		PLASMA_desc	A,
		PLASMA_sequence *	sequence,
		PLASMA_request *	request
	)

Parallel initialization a 2-D array A to BETA on the diagonal and ALPHA on the offdiagonals.

Definition at line 22 of file pslaset.c.

References BLKLDD, plasma_desc_t::m, plasma_desc_t::mb, min, plasma_desc_t::mt, plasma_desc_t::n, plasma_desc_t::nb, plasma_desc_t::nt, plasma_context_self(), PLASMA_SUCCESS, PlasmaLower, PlasmaUpper, PlasmaUpperLower, plasma_context_struct::quark, QUARK_CORE_slaset(), plasma_sequence_t::quark_sequence, QUARK_Task_Flag_Set(), Quark_Task_Flags_Initializer, plasma_sequence_t::status, and TASK_SEQUENCE.

{
    plasma_context_t *plasma;
    Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;

    int i, j;
    int ldai, ldaj;
    int tempim;
    int tempjm, tempjn;
    int minmn = min(A.mt, A.nt);

    plasma = plasma_context_self();
    if (sequence->status != PLASMA_SUCCESS)
        return;

    QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);

    if (uplo == PlasmaLower) {
       for (j = 0; j < minmn; j++){
           tempjm = j == A.mt-1 ? A.m-j*A.mb : A.mb;
           tempjn = j == A.nt-1 ? A.n-j*A.nb : A.nb;
           ldaj = BLKLDD(A, j);
           QUARK_CORE_slaset(
               plasma->quark, &task_flags,
               PlasmaLower, tempjm, tempjn, alpha, beta,
               A(j, j), ldaj);

           for (i = j+1; i < A.mt; i++){
               tempim = i == A.mt-1 ? A.m-i*A.mb : A.mb;
               ldai = BLKLDD(A, i);
               QUARK_CORE_slaset(
                   plasma->quark, &task_flags,
                   PlasmaUpperLower, tempim, tempjn, alpha, alpha,
                   A(i, j), ldai);
           }
       }
    }
    else if (uplo == PlasmaUpper) {
       for (j = 1; j < A.nt; j++){
           tempjn = j == A.nt-1 ? A.n-j*A.nb : A.nb;
           for (i = 0; i < min(j, A.mt); i++){
               tempim = i == A.mt-1 ? A.m-i*A.mb : A.mb;
               ldai = BLKLDD(A, i);
               QUARK_CORE_slaset(
                   plasma->quark, &task_flags,
                   PlasmaUpperLower, tempim, tempjn, alpha, alpha,
                   A(i, j), ldai);
           }
       }
       for (j = 0; j < minmn; j++){
           tempjm = j == A.mt-1 ? A.m-j*A.mb : A.mb;
           tempjn = j == A.nt-1 ? A.n-j*A.nb : A.nb;
           ldaj = BLKLDD(A, j);
           QUARK_CORE_slaset(
               plasma->quark, &task_flags,
               PlasmaUpper, tempjm, tempjn, alpha, beta,
               A(j, j), ldaj);
       }
    }
    else {
       for (i = 0; i < A.mt; i++){
           tempim = i == A.mt-1 ? A.m-i*A.mb : A.mb;
           ldai = BLKLDD(A, i);
           for (j = 0; j < A.nt; j++){
               tempjn = j == A.nt-1 ? A.n-j*A.nb : A.nb;
               QUARK_CORE_slaset(
                   plasma->quark, &task_flags,
                   PlasmaUpperLower, tempim, tempjn, alpha, alpha,
                   A(i, j), ldai);
           }
       }
       for (j = 0; j < minmn; j++){
           tempjm = j == A.mt-1 ? A.m-j*A.mb : A.mb;
           tempjn = j == A.nt-1 ? A.n-j*A.nb : A.nb;
           ldaj = BLKLDD(A, j);
           QUARK_CORE_slaset(
               plasma->quark, &task_flags,
               PlasmaUpperLower, tempjm, tempjn, alpha, beta,
               A(j, j), ldaj);
       }
    }
}

Here is the call graph for this function: