pcungqrrh.c File Reference

#include "common.h"

Include dependency graph for pcungqrrh.c:

Go to the source code of this file.

Macros
#define	A(m, n)   BLKADDR(A, PLASMA_Complex32_t, (m), (n))
#define	Q(m, n)   BLKADDR(Q, PLASMA_Complex32_t, (m), (n))
#define	T(m, n)   BLKADDR(T, PLASMA_Complex32_t, (m), (n))
#define	T2(m, n)   BLKADDR(T, PLASMA_Complex32_t, (m), (n)+(A.nt))

Functions
void	plasma_pcungqrrh_quark (PLASMA_desc A, PLASMA_desc Q, PLASMA_desc T, int BS, 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

Dulceneia Becker

Date:

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

Definition in file pcungqrrh.c.

---

Macro Definition Documentation

#define A	(		m,
			n
	)		BLKADDR(A, PLASMA_Complex32_t, (m), (n))

Definition at line 19 of file pcungqrrh.c.

#define Q	(		m,
			n
	)		BLKADDR(Q, PLASMA_Complex32_t, (m), (n))

Definition at line 20 of file pcungqrrh.c.

#define T	(		m,
			n
	)		BLKADDR(T, PLASMA_Complex32_t, (m), (n))

Definition at line 21 of file pcungqrrh.c.

#define T2	(		m,
			n
	)		BLKADDR(T, PLASMA_Complex32_t, (m), (n)+(A.nt))

Definition at line 22 of file pcungqrrh.c.

---

Function Documentation

void plasma_pcungqrrh_quark	(	PLASMA_desc	A,
		PLASMA_desc	Q,
		PLASMA_desc	T,
		int	BS,
		PLASMA_sequence *	sequence,
		PLASMA_request *	request
	)

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

Definition at line 27 of file pcungqrrh.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, PlasmaLeft, PlasmaNoTrans, Q, plasma_context_struct::quark, QUARK_CORE_ctsmqr(), QUARK_CORE_cttmqr(), QUARK_CORE_cunmqr(), plasma_sequence_t::quark_sequence, QUARK_Task_Flag_Set(), Quark_Task_Flags_Initializer, plasma_sequence_t::status, T, T2, and TASK_SEQUENCE.

{
    plasma_context_t *plasma;
    Quark_Task_Flags task_flags = Quark_Task_Flags_Initializer;

    int k, m, n;
    int K, M, RD, lastRD;
    int ldaM, ldam, ldaMRD;
    int ldbM, ldbm, ldbMRD;
    int tempkn, tempMm, tempnn, tempmm, tempMRDm, tempkmin;
    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;
    K = min(A.mt, A.nt);
    for (k = K-1; k >= 0; k--) {
        tempkn = k == A.nt-1 ? A.n-k*A.nb : A.nb;
        lastRD = 0;
        for (RD = BS; RD < A.mt-k; RD *= 2)
            lastRD = RD;
        for (RD = lastRD; RD >= BS; RD /= 2) {
            for (M = k; M+RD < A.mt; M += 2*RD) {
                tempMRDm = M+RD == A.mt-1 ? A.m-(M+RD)*A.mb : A.mb;
                ldbM   = BLKLDD(Q, M   );
                ldbMRD = BLKLDD(Q, M+RD);
                ldaMRD = BLKLDD(A, M+RD);
                for (n = 0; n < Q.nt; n++) {
                    tempnn = n == Q.nt-1 ? Q.n-n*Q.nb : Q.nb;
                    QUARK_CORE_cttmqr(
                        plasma->quark, &task_flags,
                        PlasmaLeft, PlasmaNoTrans,
                        A.nb, tempnn, tempMRDm, tempnn,
                        tempkn, ib, T.nb,
                        Q (M,    n), ldbM,
                        Q (M+RD, n), ldbMRD,
                        A (M+RD, k), ldaMRD,
                        T2(M+RD, k), T.mb); 
                }
            }
        }
        for (M = k; M < A.mt; M += BS) {
            tempMm   = M == A.mt-1 ? A.m-M*A.mb : A.mb;
            tempkmin = min(tempMm, tempkn);
            ldaM = BLKLDD(A, M);
            ldbM = BLKLDD(Q, M);
            for (m = min(M+BS, A.mt)-1; m > M; m--) {
                tempmm = m == A.mt-1 ? A.m-m*A.mb : A.mb;
                ldbm = BLKLDD(Q, m);
                ldam = BLKLDD(A, m);

                for (n = 0; n < Q.nt; n++) {
                    tempnn = n == Q.nt-1 ? Q.n-n*Q.nb : Q.nb;
                    QUARK_CORE_ctsmqr(
                        plasma->quark, &task_flags,
                        PlasmaLeft, PlasmaNoTrans,
                        A.nb, tempnn, tempmm, tempnn,
                        tempkn, ib, T.nb,
                        Q(M, n), ldbM,
                        Q(m, n), ldbm,
                        A(m, k), ldam,
                        T(m, k), T.mb);
                }
            }
            for (n = 0; n < Q.nt; n++) {
                tempnn = n == Q.nt-1 ? Q.n-n*Q.nb : Q.nb;
                QUARK_CORE_cunmqr(
                    plasma->quark, &task_flags,
                    PlasmaLeft, PlasmaNoTrans,
                    tempMm, tempnn,
                    tempkmin, ib, T.nb,
                    A(M, k), ldaM,
                    T(M, k), T.mb,
                    Q(M, n), ldbM);
            }
        }
    }
}

Here is the call graph for this function: