pzunglq.c File Reference

#include "common.h"

Include dependency graph for pzunglq.c:

Go to the source code of this file.

Macros
#define	A(m, n)   BLKADDR(A, PLASMA_Complex64_t, m, n)
#define	Q(m, n)   BLKADDR(Q, PLASMA_Complex64_t, m, n)
#define	T(m, n)   BLKADDR(T, PLASMA_Complex64_t, m, n)

Functions
void	plasma_pzunglq_quark (PLASMA_desc A, PLASMA_desc Q, PLASMA_desc T, 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

Jakub Kurzak

Mathieu Faverge

Date:

2010-11-15 normal z -> s d c

Definition in file pzunglq.c.

---

Macro Definition Documentation

#define A	(		m,
			n
	)		BLKADDR(A, PLASMA_Complex64_t, m, n)

Definition at line 19 of file pzunglq.c.

#define Q	(		m,
			n
	)		BLKADDR(Q, PLASMA_Complex64_t, m, n)

Definition at line 20 of file pzunglq.c.

#define T	(		m,
			n
	)		BLKADDR(T, PLASMA_Complex64_t, m, n)

Definition at line 21 of file pzunglq.c.

---

Function Documentation

void plasma_pzunglq_quark	(	PLASMA_desc	A,
		PLASMA_desc	Q,
		PLASMA_desc	T,
		PLASMA_sequence *	sequence,
		PLASMA_request *	request
	)

Parallel construction of Q using tile V (application to identity) - dynamic scheduling

Definition at line 25 of file pzunglq.c.

References A, 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_IB, PLASMA_SUCCESS, PlasmaNoTrans, PlasmaRight, Q, plasma_context_struct::quark, QUARK_CORE_ztsmlq(), QUARK_CORE_zunmlq(), plasma_sequence_t::quark_sequence, QUARK_Task_Flag_Set(), Quark_Task_Flags_Initializer, plasma_sequence_t::status, T, and TASK_SEQUENCE.

{
    plasma_context_t *plasma;
    Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;

    int k, m, n;
    int ldak, ldqm;
    int tempnn, tempmm, tempkmin, tempkn;
    int tempAkm, tempAkn;
    int ib;

    plasma = plasma_context_self();
    if (sequence->status != PLASMA_SUCCESS)
        return;
    QUARK_Task_Flag_Set(&task_flags, TASK_SEQUENCE, (intptr_t)sequence->quark_sequence);

    ib = PLASMA_IB;
    for (k = min(A.mt, A.nt)-1; k >= 0; k--) {
        tempAkm  = k == A.mt-1 ? A.m-k*A.mb : A.mb;
        tempAkn  = k == A.nt-1 ? A.n-k*A.nb : A.nb;
        tempkmin = min( tempAkn, tempAkm );
        tempkn   = k == Q.nt-1 ? Q.n-k*Q.nb : Q.nb;
        ldak = BLKLDD(A, k);
        for (n = Q.nt-1; n > k; n--) {
            tempnn = n == Q.nt-1 ? Q.n-n*Q.nb : Q.nb;
            for (m = 0; m < Q.mt; m++) {
                tempmm = m == Q.mt-1 ? Q.m-m*Q.mb : Q.mb;
                ldqm = BLKLDD(Q, m);
                QUARK_CORE_ztsmlq(
                    plasma->quark, &task_flags,
                    PlasmaRight, PlasmaNoTrans,
                    tempmm, Q.nb, tempmm, tempnn, tempAkm, ib, T.nb,
                    Q(m, k), ldqm,
                    Q(m, n), ldqm,
                    A(k, n), ldak,
                    T(k, n), T.mb);
            }
        }
        for (m = 0; m < Q.mt; m++) {
            tempmm = m == Q.mt-1 ? Q.m-m*Q.mb : Q.mb;
            ldqm = BLKLDD(Q, m);
            QUARK_CORE_zunmlq(
                plasma->quark, &task_flags,
                PlasmaRight, PlasmaNoTrans,
                tempmm, tempkn, tempkmin, ib, T.nb,
                A(k, k), ldak,
                T(k, k), T.mb,
                Q(m, k), ldqm);
        }
    }
}

Here is the call graph for this function: