123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955 |
- /* dlaed4.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"
- /* Subroutine */ int _starpu_dlaed4_(integer *n, integer *i__, doublereal *d__,
- doublereal *z__, doublereal *delta, doublereal *rho, doublereal *dlam,
- integer *info)
- {
- /* System generated locals */
- integer i__1;
- doublereal d__1;
- /* Builtin functions */
- double sqrt(doublereal);
- /* Local variables */
- doublereal a, b, c__;
- integer j;
- doublereal w;
- integer ii;
- doublereal dw, zz[3];
- integer ip1;
- doublereal del, eta, phi, eps, tau, psi;
- integer iim1, iip1;
- doublereal dphi, dpsi;
- integer iter;
- doublereal temp, prew, temp1, dltlb, dltub, midpt;
- integer niter;
- logical swtch;
- extern /* Subroutine */ int _starpu_dlaed5_(integer *, doublereal *, doublereal *,
- doublereal *, doublereal *, doublereal *), _starpu_dlaed6_(integer *,
- logical *, doublereal *, doublereal *, doublereal *, doublereal *,
- doublereal *, integer *);
- logical swtch3;
- extern doublereal _starpu_dlamch_(char *);
- logical orgati;
- doublereal erretm, rhoinv;
- /* -- LAPACK routine (version 3.2) -- */
- /* Univ. of Tennessee, Univ. of California Berkeley and NAG Ltd.. */
- /* November 2006 */
- /* .. Scalar Arguments .. */
- /* .. */
- /* .. Array Arguments .. */
- /* .. */
- /* Purpose */
- /* ======= */
- /* This subroutine computes the I-th updated eigenvalue of a symmetric */
- /* rank-one modification to a diagonal matrix whose elements are */
- /* given in the array d, and that */
- /* D(i) < D(j) for i < j */
- /* and that RHO > 0. This is arranged by the calling routine, and is */
- /* no loss in generality. The rank-one modified system is thus */
- /* diag( D ) + RHO * Z * Z_transpose. */
- /* where we assume the Euclidean norm of Z is 1. */
- /* The method consists of approximating the rational functions in the */
- /* secular equation by simpler interpolating rational functions. */
- /* Arguments */
- /* ========= */
- /* N (input) INTEGER */
- /* The length of all arrays. */
- /* I (input) INTEGER */
- /* The index of the eigenvalue to be computed. 1 <= I <= N. */
- /* D (input) DOUBLE PRECISION array, dimension (N) */
- /* The original eigenvalues. It is assumed that they are in */
- /* order, D(I) < D(J) for I < J. */
- /* Z (input) DOUBLE PRECISION array, dimension (N) */
- /* The components of the updating vector. */
- /* DELTA (output) DOUBLE PRECISION array, dimension (N) */
- /* If N .GT. 2, DELTA contains (D(j) - lambda_I) in its j-th */
- /* component. If N = 1, then DELTA(1) = 1. If N = 2, see DLAED5 */
- /* for detail. The vector DELTA contains the information necessary */
- /* to construct the eigenvectors by DLAED3 and DLAED9. */
- /* RHO (input) DOUBLE PRECISION */
- /* The scalar in the symmetric updating formula. */
- /* DLAM (output) DOUBLE PRECISION */
- /* The computed lambda_I, the I-th updated eigenvalue. */
- /* INFO (output) INTEGER */
- /* = 0: successful exit */
- /* > 0: if INFO = 1, the updating process failed. */
- /* Internal Parameters */
- /* =================== */
- /* Logical variable ORGATI (origin-at-i?) is used for distinguishing */
- /* whether D(i) or D(i+1) is treated as the origin. */
- /* ORGATI = .true. origin at i */
- /* ORGATI = .false. origin at i+1 */
- /* Logical variable SWTCH3 (switch-for-3-poles?) is for noting */
- /* if we are working with THREE poles! */
- /* MAXIT is the maximum number of iterations allowed for each */
- /* eigenvalue. */
- /* Further Details */
- /* =============== */
- /* Based on contributions by */
- /* Ren-Cang Li, Computer Science Division, University of California */
- /* at Berkeley, USA */
- /* ===================================================================== */
- /* .. Parameters .. */
- /* .. */
- /* .. Local Scalars .. */
- /* .. */
- /* .. Local Arrays .. */
- /* .. */
- /* .. External Functions .. */
- /* .. */
- /* .. External Subroutines .. */
- /* .. */
- /* .. Intrinsic Functions .. */
- /* .. */
- /* .. Executable Statements .. */
- /* Since this routine is called in an inner loop, we do no argument */
- /* checking. */
- /* Quick return for N=1 and 2. */
- /* Parameter adjustments */
- --delta;
- --z__;
- --d__;
- /* Function Body */
- *info = 0;
- if (*n == 1) {
- /* Presumably, I=1 upon entry */
- *dlam = d__[1] + *rho * z__[1] * z__[1];
- delta[1] = 1.;
- return 0;
- }
- if (*n == 2) {
- _starpu_dlaed5_(i__, &d__[1], &z__[1], &delta[1], rho, dlam);
- return 0;
- }
- /* Compute machine epsilon */
- eps = _starpu_dlamch_("Epsilon");
- rhoinv = 1. / *rho;
- /* The case I = N */
- if (*i__ == *n) {
- /* Initialize some basic variables */
- ii = *n - 1;
- niter = 1;
- /* Calculate initial guess */
- midpt = *rho / 2.;
- /* If ||Z||_2 is not one, then TEMP should be set to */
- /* RHO * ||Z||_2^2 / TWO */
- i__1 = *n;
- for (j = 1; j <= i__1; ++j) {
- delta[j] = d__[j] - d__[*i__] - midpt;
- /* L10: */
- }
- psi = 0.;
- i__1 = *n - 2;
- for (j = 1; j <= i__1; ++j) {
- psi += z__[j] * z__[j] / delta[j];
- /* L20: */
- }
- c__ = rhoinv + psi;
- w = c__ + z__[ii] * z__[ii] / delta[ii] + z__[*n] * z__[*n] / delta[*
- n];
- if (w <= 0.) {
- temp = z__[*n - 1] * z__[*n - 1] / (d__[*n] - d__[*n - 1] + *rho)
- + z__[*n] * z__[*n] / *rho;
- if (c__ <= temp) {
- tau = *rho;
- } else {
- del = d__[*n] - d__[*n - 1];
- a = -c__ * del + z__[*n - 1] * z__[*n - 1] + z__[*n] * z__[*n]
- ;
- b = z__[*n] * z__[*n] * del;
- if (a < 0.) {
- tau = b * 2. / (sqrt(a * a + b * 4. * c__) - a);
- } else {
- tau = (a + sqrt(a * a + b * 4. * c__)) / (c__ * 2.);
- }
- }
- /* It can be proved that */
- /* D(N)+RHO/2 <= LAMBDA(N) < D(N)+TAU <= D(N)+RHO */
- dltlb = midpt;
- dltub = *rho;
- } else {
- del = d__[*n] - d__[*n - 1];
- a = -c__ * del + z__[*n - 1] * z__[*n - 1] + z__[*n] * z__[*n];
- b = z__[*n] * z__[*n] * del;
- if (a < 0.) {
- tau = b * 2. / (sqrt(a * a + b * 4. * c__) - a);
- } else {
- tau = (a + sqrt(a * a + b * 4. * c__)) / (c__ * 2.);
- }
- /* It can be proved that */
- /* D(N) < D(N)+TAU < LAMBDA(N) < D(N)+RHO/2 */
- dltlb = 0.;
- dltub = midpt;
- }
- i__1 = *n;
- for (j = 1; j <= i__1; ++j) {
- delta[j] = d__[j] - d__[*i__] - tau;
- /* L30: */
- }
- /* Evaluate PSI and the derivative DPSI */
- dpsi = 0.;
- psi = 0.;
- erretm = 0.;
- i__1 = ii;
- for (j = 1; j <= i__1; ++j) {
- temp = z__[j] / delta[j];
- psi += z__[j] * temp;
- dpsi += temp * temp;
- erretm += psi;
- /* L40: */
- }
- erretm = abs(erretm);
- /* Evaluate PHI and the derivative DPHI */
- temp = z__[*n] / delta[*n];
- phi = z__[*n] * temp;
- dphi = temp * temp;
- erretm = (-phi - psi) * 8. + erretm - phi + rhoinv + abs(tau) * (dpsi
- + dphi);
- w = rhoinv + phi + psi;
- /* Test for convergence */
- if (abs(w) <= eps * erretm) {
- *dlam = d__[*i__] + tau;
- goto L250;
- }
- if (w <= 0.) {
- dltlb = max(dltlb,tau);
- } else {
- dltub = min(dltub,tau);
- }
- /* Calculate the new step */
- ++niter;
- c__ = w - delta[*n - 1] * dpsi - delta[*n] * dphi;
- a = (delta[*n - 1] + delta[*n]) * w - delta[*n - 1] * delta[*n] * (
- dpsi + dphi);
- b = delta[*n - 1] * delta[*n] * w;
- if (c__ < 0.) {
- c__ = abs(c__);
- }
- if (c__ == 0.) {
- /* ETA = B/A */
- /* ETA = RHO - TAU */
- eta = dltub - tau;
- } else if (a >= 0.) {
- eta = (a + sqrt((d__1 = a * a - b * 4. * c__, abs(d__1)))) / (c__
- * 2.);
- } else {
- eta = b * 2. / (a - sqrt((d__1 = a * a - b * 4. * c__, abs(d__1)))
- );
- }
- /* Note, eta should be positive if w is negative, and */
- /* eta should be negative otherwise. However, */
- /* if for some reason caused by roundoff, eta*w > 0, */
- /* we simply use one Newton step instead. This way */
- /* will guarantee eta*w < 0. */
- if (w * eta > 0.) {
- eta = -w / (dpsi + dphi);
- }
- temp = tau + eta;
- if (temp > dltub || temp < dltlb) {
- if (w < 0.) {
- eta = (dltub - tau) / 2.;
- } else {
- eta = (dltlb - tau) / 2.;
- }
- }
- i__1 = *n;
- for (j = 1; j <= i__1; ++j) {
- delta[j] -= eta;
- /* L50: */
- }
- tau += eta;
- /* Evaluate PSI and the derivative DPSI */
- dpsi = 0.;
- psi = 0.;
- erretm = 0.;
- i__1 = ii;
- for (j = 1; j <= i__1; ++j) {
- temp = z__[j] / delta[j];
- psi += z__[j] * temp;
- dpsi += temp * temp;
- erretm += psi;
- /* L60: */
- }
- erretm = abs(erretm);
- /* Evaluate PHI and the derivative DPHI */
- temp = z__[*n] / delta[*n];
- phi = z__[*n] * temp;
- dphi = temp * temp;
- erretm = (-phi - psi) * 8. + erretm - phi + rhoinv + abs(tau) * (dpsi
- + dphi);
- w = rhoinv + phi + psi;
- /* Main loop to update the values of the array DELTA */
- iter = niter + 1;
- for (niter = iter; niter <= 30; ++niter) {
- /* Test for convergence */
- if (abs(w) <= eps * erretm) {
- *dlam = d__[*i__] + tau;
- goto L250;
- }
- if (w <= 0.) {
- dltlb = max(dltlb,tau);
- } else {
- dltub = min(dltub,tau);
- }
- /* Calculate the new step */
- c__ = w - delta[*n - 1] * dpsi - delta[*n] * dphi;
- a = (delta[*n - 1] + delta[*n]) * w - delta[*n - 1] * delta[*n] *
- (dpsi + dphi);
- b = delta[*n - 1] * delta[*n] * w;
- if (a >= 0.) {
- eta = (a + sqrt((d__1 = a * a - b * 4. * c__, abs(d__1)))) / (
- c__ * 2.);
- } else {
- eta = b * 2. / (a - sqrt((d__1 = a * a - b * 4. * c__, abs(
- d__1))));
- }
- /* Note, eta should be positive if w is negative, and */
- /* eta should be negative otherwise. However, */
- /* if for some reason caused by roundoff, eta*w > 0, */
- /* we simply use one Newton step instead. This way */
- /* will guarantee eta*w < 0. */
- if (w * eta > 0.) {
- eta = -w / (dpsi + dphi);
- }
- temp = tau + eta;
- if (temp > dltub || temp < dltlb) {
- if (w < 0.) {
- eta = (dltub - tau) / 2.;
- } else {
- eta = (dltlb - tau) / 2.;
- }
- }
- i__1 = *n;
- for (j = 1; j <= i__1; ++j) {
- delta[j] -= eta;
- /* L70: */
- }
- tau += eta;
- /* Evaluate PSI and the derivative DPSI */
- dpsi = 0.;
- psi = 0.;
- erretm = 0.;
- i__1 = ii;
- for (j = 1; j <= i__1; ++j) {
- temp = z__[j] / delta[j];
- psi += z__[j] * temp;
- dpsi += temp * temp;
- erretm += psi;
- /* L80: */
- }
- erretm = abs(erretm);
- /* Evaluate PHI and the derivative DPHI */
- temp = z__[*n] / delta[*n];
- phi = z__[*n] * temp;
- dphi = temp * temp;
- erretm = (-phi - psi) * 8. + erretm - phi + rhoinv + abs(tau) * (
- dpsi + dphi);
- w = rhoinv + phi + psi;
- /* L90: */
- }
- /* Return with INFO = 1, NITER = MAXIT and not converged */
- *info = 1;
- *dlam = d__[*i__] + tau;
- goto L250;
- /* End for the case I = N */
- } else {
- /* The case for I < N */
- niter = 1;
- ip1 = *i__ + 1;
- /* Calculate initial guess */
- del = d__[ip1] - d__[*i__];
- midpt = del / 2.;
- i__1 = *n;
- for (j = 1; j <= i__1; ++j) {
- delta[j] = d__[j] - d__[*i__] - midpt;
- /* L100: */
- }
- psi = 0.;
- i__1 = *i__ - 1;
- for (j = 1; j <= i__1; ++j) {
- psi += z__[j] * z__[j] / delta[j];
- /* L110: */
- }
- phi = 0.;
- i__1 = *i__ + 2;
- for (j = *n; j >= i__1; --j) {
- phi += z__[j] * z__[j] / delta[j];
- /* L120: */
- }
- c__ = rhoinv + psi + phi;
- w = c__ + z__[*i__] * z__[*i__] / delta[*i__] + z__[ip1] * z__[ip1] /
- delta[ip1];
- if (w > 0.) {
- /* d(i)< the ith eigenvalue < (d(i)+d(i+1))/2 */
- /* We choose d(i) as origin. */
- orgati = TRUE_;
- a = c__ * del + z__[*i__] * z__[*i__] + z__[ip1] * z__[ip1];
- b = z__[*i__] * z__[*i__] * del;
- if (a > 0.) {
- tau = b * 2. / (a + sqrt((d__1 = a * a - b * 4. * c__, abs(
- d__1))));
- } else {
- tau = (a - sqrt((d__1 = a * a - b * 4. * c__, abs(d__1)))) / (
- c__ * 2.);
- }
- dltlb = 0.;
- dltub = midpt;
- } else {
- /* (d(i)+d(i+1))/2 <= the ith eigenvalue < d(i+1) */
- /* We choose d(i+1) as origin. */
- orgati = FALSE_;
- a = c__ * del - z__[*i__] * z__[*i__] - z__[ip1] * z__[ip1];
- b = z__[ip1] * z__[ip1] * del;
- if (a < 0.) {
- tau = b * 2. / (a - sqrt((d__1 = a * a + b * 4. * c__, abs(
- d__1))));
- } else {
- tau = -(a + sqrt((d__1 = a * a + b * 4. * c__, abs(d__1)))) /
- (c__ * 2.);
- }
- dltlb = -midpt;
- dltub = 0.;
- }
- if (orgati) {
- i__1 = *n;
- for (j = 1; j <= i__1; ++j) {
- delta[j] = d__[j] - d__[*i__] - tau;
- /* L130: */
- }
- } else {
- i__1 = *n;
- for (j = 1; j <= i__1; ++j) {
- delta[j] = d__[j] - d__[ip1] - tau;
- /* L140: */
- }
- }
- if (orgati) {
- ii = *i__;
- } else {
- ii = *i__ + 1;
- }
- iim1 = ii - 1;
- iip1 = ii + 1;
- /* Evaluate PSI and the derivative DPSI */
- dpsi = 0.;
- psi = 0.;
- erretm = 0.;
- i__1 = iim1;
- for (j = 1; j <= i__1; ++j) {
- temp = z__[j] / delta[j];
- psi += z__[j] * temp;
- dpsi += temp * temp;
- erretm += psi;
- /* L150: */
- }
- erretm = abs(erretm);
- /* Evaluate PHI and the derivative DPHI */
- dphi = 0.;
- phi = 0.;
- i__1 = iip1;
- for (j = *n; j >= i__1; --j) {
- temp = z__[j] / delta[j];
- phi += z__[j] * temp;
- dphi += temp * temp;
- erretm += phi;
- /* L160: */
- }
- w = rhoinv + phi + psi;
- /* W is the value of the secular function with */
- /* its ii-th element removed. */
- swtch3 = FALSE_;
- if (orgati) {
- if (w < 0.) {
- swtch3 = TRUE_;
- }
- } else {
- if (w > 0.) {
- swtch3 = TRUE_;
- }
- }
- if (ii == 1 || ii == *n) {
- swtch3 = FALSE_;
- }
- temp = z__[ii] / delta[ii];
- dw = dpsi + dphi + temp * temp;
- temp = z__[ii] * temp;
- w += temp;
- erretm = (phi - psi) * 8. + erretm + rhoinv * 2. + abs(temp) * 3. +
- abs(tau) * dw;
- /* Test for convergence */
- if (abs(w) <= eps * erretm) {
- if (orgati) {
- *dlam = d__[*i__] + tau;
- } else {
- *dlam = d__[ip1] + tau;
- }
- goto L250;
- }
- if (w <= 0.) {
- dltlb = max(dltlb,tau);
- } else {
- dltub = min(dltub,tau);
- }
- /* Calculate the new step */
- ++niter;
- if (! swtch3) {
- if (orgati) {
- /* Computing 2nd power */
- d__1 = z__[*i__] / delta[*i__];
- c__ = w - delta[ip1] * dw - (d__[*i__] - d__[ip1]) * (d__1 *
- d__1);
- } else {
- /* Computing 2nd power */
- d__1 = z__[ip1] / delta[ip1];
- c__ = w - delta[*i__] * dw - (d__[ip1] - d__[*i__]) * (d__1 *
- d__1);
- }
- a = (delta[*i__] + delta[ip1]) * w - delta[*i__] * delta[ip1] *
- dw;
- b = delta[*i__] * delta[ip1] * w;
- if (c__ == 0.) {
- if (a == 0.) {
- if (orgati) {
- a = z__[*i__] * z__[*i__] + delta[ip1] * delta[ip1] *
- (dpsi + dphi);
- } else {
- a = z__[ip1] * z__[ip1] + delta[*i__] * delta[*i__] *
- (dpsi + dphi);
- }
- }
- eta = b / a;
- } else if (a <= 0.) {
- eta = (a - sqrt((d__1 = a * a - b * 4. * c__, abs(d__1)))) / (
- c__ * 2.);
- } else {
- eta = b * 2. / (a + sqrt((d__1 = a * a - b * 4. * c__, abs(
- d__1))));
- }
- } else {
- /* Interpolation using THREE most relevant poles */
- temp = rhoinv + psi + phi;
- if (orgati) {
- temp1 = z__[iim1] / delta[iim1];
- temp1 *= temp1;
- c__ = temp - delta[iip1] * (dpsi + dphi) - (d__[iim1] - d__[
- iip1]) * temp1;
- zz[0] = z__[iim1] * z__[iim1];
- zz[2] = delta[iip1] * delta[iip1] * (dpsi - temp1 + dphi);
- } else {
- temp1 = z__[iip1] / delta[iip1];
- temp1 *= temp1;
- c__ = temp - delta[iim1] * (dpsi + dphi) - (d__[iip1] - d__[
- iim1]) * temp1;
- zz[0] = delta[iim1] * delta[iim1] * (dpsi + (dphi - temp1));
- zz[2] = z__[iip1] * z__[iip1];
- }
- zz[1] = z__[ii] * z__[ii];
- _starpu_dlaed6_(&niter, &orgati, &c__, &delta[iim1], zz, &w, &eta, info);
- if (*info != 0) {
- goto L250;
- }
- }
- /* Note, eta should be positive if w is negative, and */
- /* eta should be negative otherwise. However, */
- /* if for some reason caused by roundoff, eta*w > 0, */
- /* we simply use one Newton step instead. This way */
- /* will guarantee eta*w < 0. */
- if (w * eta >= 0.) {
- eta = -w / dw;
- }
- temp = tau + eta;
- if (temp > dltub || temp < dltlb) {
- if (w < 0.) {
- eta = (dltub - tau) / 2.;
- } else {
- eta = (dltlb - tau) / 2.;
- }
- }
- prew = w;
- i__1 = *n;
- for (j = 1; j <= i__1; ++j) {
- delta[j] -= eta;
- /* L180: */
- }
- /* Evaluate PSI and the derivative DPSI */
- dpsi = 0.;
- psi = 0.;
- erretm = 0.;
- i__1 = iim1;
- for (j = 1; j <= i__1; ++j) {
- temp = z__[j] / delta[j];
- psi += z__[j] * temp;
- dpsi += temp * temp;
- erretm += psi;
- /* L190: */
- }
- erretm = abs(erretm);
- /* Evaluate PHI and the derivative DPHI */
- dphi = 0.;
- phi = 0.;
- i__1 = iip1;
- for (j = *n; j >= i__1; --j) {
- temp = z__[j] / delta[j];
- phi += z__[j] * temp;
- dphi += temp * temp;
- erretm += phi;
- /* L200: */
- }
- temp = z__[ii] / delta[ii];
- dw = dpsi + dphi + temp * temp;
- temp = z__[ii] * temp;
- w = rhoinv + phi + psi + temp;
- erretm = (phi - psi) * 8. + erretm + rhoinv * 2. + abs(temp) * 3. + (
- d__1 = tau + eta, abs(d__1)) * dw;
- swtch = FALSE_;
- if (orgati) {
- if (-w > abs(prew) / 10.) {
- swtch = TRUE_;
- }
- } else {
- if (w > abs(prew) / 10.) {
- swtch = TRUE_;
- }
- }
- tau += eta;
- /* Main loop to update the values of the array DELTA */
- iter = niter + 1;
- for (niter = iter; niter <= 30; ++niter) {
- /* Test for convergence */
- if (abs(w) <= eps * erretm) {
- if (orgati) {
- *dlam = d__[*i__] + tau;
- } else {
- *dlam = d__[ip1] + tau;
- }
- goto L250;
- }
- if (w <= 0.) {
- dltlb = max(dltlb,tau);
- } else {
- dltub = min(dltub,tau);
- }
- /* Calculate the new step */
- if (! swtch3) {
- if (! swtch) {
- if (orgati) {
- /* Computing 2nd power */
- d__1 = z__[*i__] / delta[*i__];
- c__ = w - delta[ip1] * dw - (d__[*i__] - d__[ip1]) * (
- d__1 * d__1);
- } else {
- /* Computing 2nd power */
- d__1 = z__[ip1] / delta[ip1];
- c__ = w - delta[*i__] * dw - (d__[ip1] - d__[*i__]) *
- (d__1 * d__1);
- }
- } else {
- temp = z__[ii] / delta[ii];
- if (orgati) {
- dpsi += temp * temp;
- } else {
- dphi += temp * temp;
- }
- c__ = w - delta[*i__] * dpsi - delta[ip1] * dphi;
- }
- a = (delta[*i__] + delta[ip1]) * w - delta[*i__] * delta[ip1]
- * dw;
- b = delta[*i__] * delta[ip1] * w;
- if (c__ == 0.) {
- if (a == 0.) {
- if (! swtch) {
- if (orgati) {
- a = z__[*i__] * z__[*i__] + delta[ip1] *
- delta[ip1] * (dpsi + dphi);
- } else {
- a = z__[ip1] * z__[ip1] + delta[*i__] * delta[
- *i__] * (dpsi + dphi);
- }
- } else {
- a = delta[*i__] * delta[*i__] * dpsi + delta[ip1]
- * delta[ip1] * dphi;
- }
- }
- eta = b / a;
- } else if (a <= 0.) {
- eta = (a - sqrt((d__1 = a * a - b * 4. * c__, abs(d__1))))
- / (c__ * 2.);
- } else {
- eta = b * 2. / (a + sqrt((d__1 = a * a - b * 4. * c__,
- abs(d__1))));
- }
- } else {
- /* Interpolation using THREE most relevant poles */
- temp = rhoinv + psi + phi;
- if (swtch) {
- c__ = temp - delta[iim1] * dpsi - delta[iip1] * dphi;
- zz[0] = delta[iim1] * delta[iim1] * dpsi;
- zz[2] = delta[iip1] * delta[iip1] * dphi;
- } else {
- if (orgati) {
- temp1 = z__[iim1] / delta[iim1];
- temp1 *= temp1;
- c__ = temp - delta[iip1] * (dpsi + dphi) - (d__[iim1]
- - d__[iip1]) * temp1;
- zz[0] = z__[iim1] * z__[iim1];
- zz[2] = delta[iip1] * delta[iip1] * (dpsi - temp1 +
- dphi);
- } else {
- temp1 = z__[iip1] / delta[iip1];
- temp1 *= temp1;
- c__ = temp - delta[iim1] * (dpsi + dphi) - (d__[iip1]
- - d__[iim1]) * temp1;
- zz[0] = delta[iim1] * delta[iim1] * (dpsi + (dphi -
- temp1));
- zz[2] = z__[iip1] * z__[iip1];
- }
- }
- _starpu_dlaed6_(&niter, &orgati, &c__, &delta[iim1], zz, &w, &eta,
- info);
- if (*info != 0) {
- goto L250;
- }
- }
- /* Note, eta should be positive if w is negative, and */
- /* eta should be negative otherwise. However, */
- /* if for some reason caused by roundoff, eta*w > 0, */
- /* we simply use one Newton step instead. This way */
- /* will guarantee eta*w < 0. */
- if (w * eta >= 0.) {
- eta = -w / dw;
- }
- temp = tau + eta;
- if (temp > dltub || temp < dltlb) {
- if (w < 0.) {
- eta = (dltub - tau) / 2.;
- } else {
- eta = (dltlb - tau) / 2.;
- }
- }
- i__1 = *n;
- for (j = 1; j <= i__1; ++j) {
- delta[j] -= eta;
- /* L210: */
- }
- tau += eta;
- prew = w;
- /* Evaluate PSI and the derivative DPSI */
- dpsi = 0.;
- psi = 0.;
- erretm = 0.;
- i__1 = iim1;
- for (j = 1; j <= i__1; ++j) {
- temp = z__[j] / delta[j];
- psi += z__[j] * temp;
- dpsi += temp * temp;
- erretm += psi;
- /* L220: */
- }
- erretm = abs(erretm);
- /* Evaluate PHI and the derivative DPHI */
- dphi = 0.;
- phi = 0.;
- i__1 = iip1;
- for (j = *n; j >= i__1; --j) {
- temp = z__[j] / delta[j];
- phi += z__[j] * temp;
- dphi += temp * temp;
- erretm += phi;
- /* L230: */
- }
- temp = z__[ii] / delta[ii];
- dw = dpsi + dphi + temp * temp;
- temp = z__[ii] * temp;
- w = rhoinv + phi + psi + temp;
- erretm = (phi - psi) * 8. + erretm + rhoinv * 2. + abs(temp) * 3.
- + abs(tau) * dw;
- if (w * prew > 0. && abs(w) > abs(prew) / 10.) {
- swtch = ! swtch;
- }
- /* L240: */
- }
- /* Return with INFO = 1, NITER = MAXIT and not converged */
- *info = 1;
- if (orgati) {
- *dlam = d__[*i__] + tau;
- } else {
- *dlam = d__[ip1] + tau;
- }
- }
- L250:
- return 0;
- /* End of DLAED4 */
- } /* _starpu_dlaed4_ */
|