starpu-max/min-dgels/base/SRC/dlarrf.c

/* dlarrf.f -- translated by f2c (version 20061008).
   You must link the resulting object file with libf2c:
	on Microsoft Windows system, link with libf2c.lib;
	on Linux or Unix systems, link with .../path/to/libf2c.a -lm
	or, if you install libf2c.a in a standard place, with -lf2c -lm
	-- in that order, at the end of the command line, as in
		cc *.o -lf2c -lm
	Source for libf2c is in /netlib/f2c/libf2c.zip, e.g.,

		http://www.netlib.org/f2c/libf2c.zip
*/

#include "f2c.h"
#include "blaswrap.h"

/* Table of constant values */

static integer c__1 = 1;

/* Subroutine */ int _starpu_dlarrf_(integer *n, doublereal *d__, doublereal *l, 
	doublereal *ld, integer *clstrt, integer *clend, doublereal *w, 
	doublereal *wgap, doublereal *werr, doublereal *spdiam, doublereal *
	clgapl, doublereal *clgapr, doublereal *pivmin, doublereal *sigma, 
	doublereal *dplus, doublereal *lplus, doublereal *work, integer *info)
{
    /* System generated locals */
    integer i__1;
    doublereal d__1, d__2, d__3;

    /* Builtin functions */
    double sqrt(doublereal);

    /* Local variables */
    integer i__;
    doublereal s, bestshift, smlgrowth, eps, tmp, max1, max2, rrr1, rrr2, 
	    znm2, growthbound, fail, fact, oldp;
    integer indx;
    doublereal prod;
    integer ktry;
    doublereal fail2, avgap, ldmax, rdmax;
    integer shift;
    extern /* Subroutine */ int _starpu_dcopy_(integer *, doublereal *, integer *, 
	    doublereal *, integer *);
    logical dorrr1;
    extern doublereal _starpu_dlamch_(char *);
    doublereal ldelta;
    logical nofail;
    doublereal mingap, lsigma, rdelta;
    extern logical _starpu_disnan_(doublereal *);
    logical forcer;
    doublereal rsigma, clwdth;
    logical sawnan1, sawnan2, tryrrr1;


/*  -- LAPACK auxiliary routine (version 3.2) -- */
/*     Univ. of Tennessee, Univ. of California Berkeley and NAG Ltd.. */
/*     November 2006 */
/* * */
/*     .. Scalar Arguments .. */
/*     .. */
/*     .. Array Arguments .. */
/*     .. */

/*  Purpose */
/*  ======= */

/*  Given the initial representation L D L^T and its cluster of close */
/*  eigenvalues (in a relative measure), W( CLSTRT ), W( CLSTRT+1 ), ... */
/*  W( CLEND ), DLARRF finds a new relatively robust representation */
/*  L D L^T - SIGMA I = L(+) D(+) L(+)^T such that at least one of the */
/*  eigenvalues of L(+) D(+) L(+)^T is relatively isolated. */

/*  Arguments */
/*  ========= */

/*  N       (input) INTEGER */
/*          The order of the matrix (subblock, if the matrix splitted). */

/*  D       (input) DOUBLE PRECISION array, dimension (N) */
/*          The N diagonal elements of the diagonal matrix D. */

/*  L       (input) DOUBLE PRECISION array, dimension (N-1) */
/*          The (N-1) subdiagonal elements of the unit bidiagonal */
/*          matrix L. */

/*  LD      (input) DOUBLE PRECISION array, dimension (N-1) */
/*          The (N-1) elements L(i)*D(i). */

/*  CLSTRT  (input) INTEGER */
/*          The index of the first eigenvalue in the cluster. */

/*  CLEND   (input) INTEGER */
/*          The index of the last eigenvalue in the cluster. */

/*  W       (input) DOUBLE PRECISION array, dimension >=  (CLEND-CLSTRT+1) */
/*          The eigenvalue APPROXIMATIONS of L D L^T in ascending order. */
/*          W( CLSTRT ) through W( CLEND ) form the cluster of relatively */
/*          close eigenalues. */

/*  WGAP    (input/output) DOUBLE PRECISION array, dimension >=  (CLEND-CLSTRT+1) */
/*          The separation from the right neighbor eigenvalue in W. */

/*  WERR    (input) DOUBLE PRECISION array, dimension >=  (CLEND-CLSTRT+1) */
/*          WERR contain the semiwidth of the uncertainty */
/*          interval of the corresponding eigenvalue APPROXIMATION in W */

/*  SPDIAM (input) estimate of the spectral diameter obtained from the */
/*          Gerschgorin intervals */

/*  CLGAPL, CLGAPR (input) absolute gap on each end of the cluster. */
/*          Set by the calling routine to protect against shifts too close */
/*          to eigenvalues outside the cluster. */

/*  PIVMIN  (input) DOUBLE PRECISION */
/*          The minimum pivot allowed in the Sturm sequence. */

/*  SIGMA   (output) DOUBLE PRECISION */
/*          The shift used to form L(+) D(+) L(+)^T. */

/*  DPLUS   (output) DOUBLE PRECISION array, dimension (N) */
/*          The N diagonal elements of the diagonal matrix D(+). */

/*  LPLUS   (output) DOUBLE PRECISION array, dimension (N-1) */
/*          The first (N-1) elements of LPLUS contain the subdiagonal */
/*          elements of the unit bidiagonal matrix L(+). */

/*  WORK    (workspace) DOUBLE PRECISION array, dimension (2*N) */
/*          Workspace. */

/*  Further Details */
/*  =============== */

/*  Based on contributions by */
/*     Beresford Parlett, University of California, Berkeley, USA */
/*     Jim Demmel, University of California, Berkeley, USA */
/*     Inderjit Dhillon, University of Texas, Austin, USA */
/*     Osni Marques, LBNL/NERSC, USA */
/*     Christof Voemel, University of California, Berkeley, USA */

/*  ===================================================================== */

/*     .. Parameters .. */
/*     .. */
/*     .. Local Scalars .. */
/*     .. */
/*     .. External Functions .. */
/*     .. */
/*     .. External Subroutines .. */
/*     .. */
/*     .. Intrinsic Functions .. */
/*     .. */
/*     .. Executable Statements .. */

    /* Parameter adjustments */
    --work;
    --lplus;
    --dplus;
    --werr;
    --wgap;
    --w;
    --ld;
    --l;
    --d__;

    /* Function Body */
    *info = 0;
    fact = 2.;
    eps = _starpu_dlamch_("Precision");
    shift = 0;
    forcer = FALSE_;
/*     Note that we cannot guarantee that for any of the shifts tried, */
/*     the factorization has a small or even moderate element growth. */
/*     There could be Ritz values at both ends of the cluster and despite */
/*     backing off, there are examples where all factorizations tried */
/*     (in IEEE mode, allowing zero pivots & infinities) have INFINITE */
/*     element growth. */
/*     For this reason, we should use PIVMIN in this subroutine so that at */
/*     least the L D L^T factorization exists. It can be checked afterwards */
/*     whether the element growth caused bad residuals/orthogonality. */
/*     Decide whether the code should accept the best among all */
/*     representations despite large element growth or signal INFO=1 */
    nofail = TRUE_;

/*     Compute the average gap length of the cluster */
    clwdth = (d__1 = w[*clend] - w[*clstrt], abs(d__1)) + werr[*clend] + werr[
	    *clstrt];
    avgap = clwdth / (doublereal) (*clend - *clstrt);
    mingap = min(*clgapl,*clgapr);
/*     Initial values for shifts to both ends of cluster */
/* Computing MIN */
    d__1 = w[*clstrt], d__2 = w[*clend];
    lsigma = min(d__1,d__2) - werr[*clstrt];
/* Computing MAX */
    d__1 = w[*clstrt], d__2 = w[*clend];
    rsigma = max(d__1,d__2) + werr[*clend];
/*     Use a small fudge to make sure that we really shift to the outside */
    lsigma -= abs(lsigma) * 4. * eps;
    rsigma += abs(rsigma) * 4. * eps;
/*     Compute upper bounds for how much to back off the initial shifts */
    ldmax = mingap * .25 + *pivmin * 2.;
    rdmax = mingap * .25 + *pivmin * 2.;
/* Computing MAX */
    d__1 = avgap, d__2 = wgap[*clstrt];
    ldelta = max(d__1,d__2) / fact;
/* Computing MAX */
    d__1 = avgap, d__2 = wgap[*clend - 1];
    rdelta = max(d__1,d__2) / fact;

/*     Initialize the record of the best representation found */

    s = _starpu_dlamch_("S");
    smlgrowth = 1. / s;
    fail = (doublereal) (*n - 1) * mingap / (*spdiam * eps);
    fail2 = (doublereal) (*n - 1) * mingap / (*spdiam * sqrt(eps));
    bestshift = lsigma;

/*     while (KTRY <= KTRYMAX) */
    ktry = 0;
    growthbound = *spdiam * 8.;
L5:
    sawnan1 = FALSE_;
    sawnan2 = FALSE_;
/*     Ensure that we do not back off too much of the initial shifts */
    ldelta = min(ldmax,ldelta);
    rdelta = min(rdmax,rdelta);
/*     Compute the element growth when shifting to both ends of the cluster */
/*     accept the shift if there is no element growth at one of the two ends */
/*     Left end */
    s = -lsigma;
    dplus[1] = d__[1] + s;
    if (abs(dplus[1]) < *pivmin) {
	dplus[1] = -(*pivmin);
/*        Need to set SAWNAN1 because refined RRR test should not be used */
/*        in this case */
	sawnan1 = TRUE_;
    }
    max1 = abs(dplus[1]);
    i__1 = *n - 1;
    for (i__ = 1; i__ <= i__1; ++i__) {
	lplus[i__] = ld[i__] / dplus[i__];
	s = s * lplus[i__] * l[i__] - lsigma;
	dplus[i__ + 1] = d__[i__ + 1] + s;
	if ((d__1 = dplus[i__ + 1], abs(d__1)) < *pivmin) {
	    dplus[i__ + 1] = -(*pivmin);
/*           Need to set SAWNAN1 because refined RRR test should not be used */
/*           in this case */
	    sawnan1 = TRUE_;
	}
/* Computing MAX */
	d__2 = max1, d__3 = (d__1 = dplus[i__ + 1], abs(d__1));
	max1 = max(d__2,d__3);
/* L6: */
    }
    sawnan1 = sawnan1 || _starpu_disnan_(&max1);
    if (forcer || max1 <= growthbound && ! sawnan1) {
	*sigma = lsigma;
	shift = 1;
	goto L100;
    }
/*     Right end */
    s = -rsigma;
    work[1] = d__[1] + s;
    if (abs(work[1]) < *pivmin) {
	work[1] = -(*pivmin);
/*        Need to set SAWNAN2 because refined RRR test should not be used */
/*        in this case */
	sawnan2 = TRUE_;
    }
    max2 = abs(work[1]);
    i__1 = *n - 1;
    for (i__ = 1; i__ <= i__1; ++i__) {
	work[*n + i__] = ld[i__] / work[i__];
	s = s * work[*n + i__] * l[i__] - rsigma;
	work[i__ + 1] = d__[i__ + 1] + s;
	if ((d__1 = work[i__ + 1], abs(d__1)) < *pivmin) {
	    work[i__ + 1] = -(*pivmin);
/*           Need to set SAWNAN2 because refined RRR test should not be used */
/*           in this case */
	    sawnan2 = TRUE_;
	}
/* Computing MAX */
	d__2 = max2, d__3 = (d__1 = work[i__ + 1], abs(d__1));
	max2 = max(d__2,d__3);
/* L7: */
    }
    sawnan2 = sawnan2 || _starpu_disnan_(&max2);
    if (forcer || max2 <= growthbound && ! sawnan2) {
	*sigma = rsigma;
	shift = 2;
	goto L100;
    }
/*     If we are at this point, both shifts led to too much element growth */
/*     Record the better of the two shifts (provided it didn't lead to NaN) */
    if (sawnan1 && sawnan2) {
/*        both MAX1 and MAX2 are NaN */
	goto L50;
    } else {
	if (! sawnan1) {
	    indx = 1;
	    if (max1 <= smlgrowth) {
		smlgrowth = max1;
		bestshift = lsigma;
	    }
	}
	if (! sawnan2) {
	    if (sawnan1 || max2 <= max1) {
		indx = 2;
	    }
	    if (max2 <= smlgrowth) {
		smlgrowth = max2;
		bestshift = rsigma;
	    }
	}
    }
/*     If we are here, both the left and the right shift led to */
/*     element growth. If the element growth is moderate, then */
/*     we may still accept the representation, if it passes a */
/*     refined test for RRR. This test supposes that no NaN occurred. */
/*     Moreover, we use the refined RRR test only for isolated clusters. */
    if (clwdth < mingap / 128. && min(max1,max2) < fail2 && ! sawnan1 && ! 
	    sawnan2) {
	dorrr1 = TRUE_;
    } else {
	dorrr1 = FALSE_;
    }
    tryrrr1 = TRUE_;
    if (tryrrr1 && dorrr1) {
	if (indx == 1) {
	    tmp = (d__1 = dplus[*n], abs(d__1));
	    znm2 = 1.;
	    prod = 1.;
	    oldp = 1.;
	    for (i__ = *n - 1; i__ >= 1; --i__) {
		if (prod <= eps) {
		    prod = dplus[i__ + 1] * work[*n + i__ + 1] / (dplus[i__] *
			     work[*n + i__]) * oldp;
		} else {
		    prod *= (d__1 = work[*n + i__], abs(d__1));
		}
		oldp = prod;
/* Computing 2nd power */
		d__1 = prod;
		znm2 += d__1 * d__1;
/* Computing MAX */
		d__2 = tmp, d__3 = (d__1 = dplus[i__] * prod, abs(d__1));
		tmp = max(d__2,d__3);
/* L15: */
	    }
	    rrr1 = tmp / (*spdiam * sqrt(znm2));
	    if (rrr1 <= 8.) {
		*sigma = lsigma;
		shift = 1;
		goto L100;
	    }
	} else if (indx == 2) {
	    tmp = (d__1 = work[*n], abs(d__1));
	    znm2 = 1.;
	    prod = 1.;
	    oldp = 1.;
	    for (i__ = *n - 1; i__ >= 1; --i__) {
		if (prod <= eps) {
		    prod = work[i__ + 1] * lplus[i__ + 1] / (work[i__] * 
			    lplus[i__]) * oldp;
		} else {
		    prod *= (d__1 = lplus[i__], abs(d__1));
		}
		oldp = prod;
/* Computing 2nd power */
		d__1 = prod;
		znm2 += d__1 * d__1;
/* Computing MAX */
		d__2 = tmp, d__3 = (d__1 = work[i__] * prod, abs(d__1));
		tmp = max(d__2,d__3);
/* L16: */
	    }
	    rrr2 = tmp / (*spdiam * sqrt(znm2));
	    if (rrr2 <= 8.) {
		*sigma = rsigma;
		shift = 2;
		goto L100;
	    }
	}
    }
L50:
    if (ktry < 1) {
/*        If we are here, both shifts failed also the RRR test. */
/*        Back off to the outside */
/* Computing MAX */
	d__1 = lsigma - ldelta, d__2 = lsigma - ldmax;
	lsigma = max(d__1,d__2);
/* Computing MIN */
	d__1 = rsigma + rdelta, d__2 = rsigma + rdmax;
	rsigma = min(d__1,d__2);
	ldelta *= 2.;
	rdelta *= 2.;
	++ktry;
	goto L5;
    } else {
/*        None of the representations investigated satisfied our */
/*        criteria. Take the best one we found. */
	if (smlgrowth < fail || nofail) {
	    lsigma = bestshift;
	    rsigma = bestshift;
	    forcer = TRUE_;
	    goto L5;
	} else {
	    *info = 1;
	    return 0;
	}
    }
L100:
    if (shift == 1) {
    } else if (shift == 2) {
/*        store new L and D back into DPLUS, LPLUS */
	_starpu_dcopy_(n, &work[1], &c__1, &dplus[1], &c__1);
	i__1 = *n - 1;
	_starpu_dcopy_(&i__1, &work[*n + 1], &c__1, &lplus[1], &c__1);
    }
    return 0;

/*     End of DLARRF */

} /* _starpu_dlarrf_ */