dlarrd.c
Go to the documentation of this file.
00001 /* dlarrd.f -- translated by f2c (version 20061008).
00002    You must link the resulting object file with libf2c:
00003         on Microsoft Windows system, link with libf2c.lib;
00004         on Linux or Unix systems, link with .../path/to/libf2c.a -lm
00005         or, if you install libf2c.a in a standard place, with -lf2c -lm
00006         -- in that order, at the end of the command line, as in
00007                 cc *.o -lf2c -lm
00008         Source for libf2c is in /netlib/f2c/libf2c.zip, e.g.,
00009 
00010                 http://www.netlib.org/f2c/libf2c.zip
00011 */
00012 
00013 #include "f2c.h"
00014 #include "blaswrap.h"
00015 
00016 /* Table of constant values */
00017 
00018 static integer c__1 = 1;
00019 static integer c_n1 = -1;
00020 static integer c__3 = 3;
00021 static integer c__2 = 2;
00022 static integer c__0 = 0;
00023 
00024 /* Subroutine */ int dlarrd_(char *range, char *order, integer *n, doublereal 
00025         *vl, doublereal *vu, integer *il, integer *iu, doublereal *gers, 
00026         doublereal *reltol, doublereal *d__, doublereal *e, doublereal *e2, 
00027         doublereal *pivmin, integer *nsplit, integer *isplit, integer *m, 
00028         doublereal *w, doublereal *werr, doublereal *wl, doublereal *wu, 
00029         integer *iblock, integer *indexw, doublereal *work, integer *iwork, 
00030         integer *info)
00031 {
00032     /* System generated locals */
00033     integer i__1, i__2, i__3;
00034     doublereal d__1, d__2;
00035 
00036     /* Builtin functions */
00037     double log(doublereal);
00038 
00039     /* Local variables */
00040     integer i__, j, ib, ie, je, nb;
00041     doublereal gl;
00042     integer im, in;
00043     doublereal gu;
00044     integer iw, jee;
00045     doublereal eps;
00046     integer nwl;
00047     doublereal wlu, wul;
00048     integer nwu;
00049     doublereal tmp1, tmp2;
00050     integer iend, jblk, ioff, iout, itmp1, itmp2, jdisc;
00051     extern logical lsame_(char *, char *);
00052     integer iinfo;
00053     doublereal atoli;
00054     integer iwoff, itmax;
00055     doublereal wkill, rtoli, uflow, tnorm;
00056     extern doublereal dlamch_(char *);
00057     integer ibegin;
00058     extern /* Subroutine */ int dlaebz_(integer *, integer *, integer *, 
00059             integer *, integer *, integer *, doublereal *, doublereal *, 
00060             doublereal *, doublereal *, doublereal *, doublereal *, integer *, 
00061              doublereal *, doublereal *, integer *, integer *, doublereal *, 
00062             integer *, integer *);
00063     integer irange, idiscl, idumma[1];
00064     extern integer ilaenv_(integer *, char *, char *, integer *, integer *, 
00065             integer *, integer *);
00066     integer idiscu;
00067     logical ncnvrg, toofew;
00068 
00069 
00070 /*  -- LAPACK auxiliary routine (version 3.2.1)                        -- */
00071 /*  -- LAPACK is a software package provided by Univ. of Tennessee,    -- */
00072 /*  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
00073 /*  -- April 2009                                                      -- */
00074 
00075 /*     .. Scalar Arguments .. */
00076 /*     .. */
00077 /*     .. Array Arguments .. */
00078 /*     .. */
00079 
00080 /*  Purpose */
00081 /*  ======= */
00082 
00083 /*  DLARRD computes the eigenvalues of a symmetric tridiagonal */
00084 /*  matrix T to suitable accuracy. This is an auxiliary code to be */
00085 /*  called from DSTEMR. */
00086 /*  The user may ask for all eigenvalues, all eigenvalues */
00087 /*  in the half-open interval (VL, VU], or the IL-th through IU-th */
00088 /*  eigenvalues. */
00089 
00090 /*  To avoid overflow, the matrix must be scaled so that its */
00091 /*  largest element is no greater than overflow**(1/2) * */
00092 /*  underflow**(1/4) in absolute value, and for greatest */
00093 /*  accuracy, it should not be much smaller than that. */
00094 
00095 /*  See W. Kahan "Accurate Eigenvalues of a Symmetric Tridiagonal */
00096 /*  Matrix", Report CS41, Computer Science Dept., Stanford */
00097 /*  University, July 21, 1966. */
00098 
00099 /*  Arguments */
00100 /*  ========= */
00101 
00102 /*  RANGE   (input) CHARACTER */
00103 /*          = 'A': ("All")   all eigenvalues will be found. */
00104 /*          = 'V': ("Value") all eigenvalues in the half-open interval */
00105 /*                           (VL, VU] will be found. */
00106 /*          = 'I': ("Index") the IL-th through IU-th eigenvalues (of the */
00107 /*                           entire matrix) will be found. */
00108 
00109 /*  ORDER   (input) CHARACTER */
00110 /*          = 'B': ("By Block") the eigenvalues will be grouped by */
00111 /*                              split-off block (see IBLOCK, ISPLIT) and */
00112 /*                              ordered from smallest to largest within */
00113 /*                              the block. */
00114 /*          = 'E': ("Entire matrix") */
00115 /*                              the eigenvalues for the entire matrix */
00116 /*                              will be ordered from smallest to */
00117 /*                              largest. */
00118 
00119 /*  N       (input) INTEGER */
00120 /*          The order of the tridiagonal matrix T.  N >= 0. */
00121 
00122 /*  VL      (input) DOUBLE PRECISION */
00123 /*  VU      (input) DOUBLE PRECISION */
00124 /*          If RANGE='V', the lower and upper bounds of the interval to */
00125 /*          be searched for eigenvalues.  Eigenvalues less than or equal */
00126 /*          to VL, or greater than VU, will not be returned.  VL < VU. */
00127 /*          Not referenced if RANGE = 'A' or 'I'. */
00128 
00129 /*  IL      (input) INTEGER */
00130 /*  IU      (input) INTEGER */
00131 /*          If RANGE='I', the indices (in ascending order) of the */
00132 /*          smallest and largest eigenvalues to be returned. */
00133 /*          1 <= IL <= IU <= N, if N > 0; IL = 1 and IU = 0 if N = 0. */
00134 /*          Not referenced if RANGE = 'A' or 'V'. */
00135 
00136 /*  GERS    (input) DOUBLE PRECISION array, dimension (2*N) */
00137 /*          The N Gerschgorin intervals (the i-th Gerschgorin interval */
00138 /*          is (GERS(2*i-1), GERS(2*i)). */
00139 
00140 /*  RELTOL  (input) DOUBLE PRECISION */
00141 /*          The minimum relative width of an interval.  When an interval */
00142 /*          is narrower than RELTOL times the larger (in */
00143 /*          magnitude) endpoint, then it is considered to be */
00144 /*          sufficiently small, i.e., converged.  Note: this should */
00145 /*          always be at least radix*machine epsilon. */
00146 
00147 /*  D       (input) DOUBLE PRECISION array, dimension (N) */
00148 /*          The n diagonal elements of the tridiagonal matrix T. */
00149 
00150 /*  E       (input) DOUBLE PRECISION array, dimension (N-1) */
00151 /*          The (n-1) off-diagonal elements of the tridiagonal matrix T. */
00152 
00153 /*  E2      (input) DOUBLE PRECISION array, dimension (N-1) */
00154 /*          The (n-1) squared off-diagonal elements of the tridiagonal matrix T. */
00155 
00156 /*  PIVMIN  (input) DOUBLE PRECISION */
00157 /*          The minimum pivot allowed in the Sturm sequence for T. */
00158 
00159 /*  NSPLIT  (input) INTEGER */
00160 /*          The number of diagonal blocks in the matrix T. */
00161 /*          1 <= NSPLIT <= N. */
00162 
00163 /*  ISPLIT  (input) INTEGER array, dimension (N) */
00164 /*          The splitting points, at which T breaks up into submatrices. */
00165 /*          The first submatrix consists of rows/columns 1 to ISPLIT(1), */
00166 /*          the second of rows/columns ISPLIT(1)+1 through ISPLIT(2), */
00167 /*          etc., and the NSPLIT-th consists of rows/columns */
00168 /*          ISPLIT(NSPLIT-1)+1 through ISPLIT(NSPLIT)=N. */
00169 /*          (Only the first NSPLIT elements will actually be used, but */
00170 /*          since the user cannot know a priori what value NSPLIT will */
00171 /*          have, N words must be reserved for ISPLIT.) */
00172 
00173 /*  M       (output) INTEGER */
00174 /*          The actual number of eigenvalues found. 0 <= M <= N. */
00175 /*          (See also the description of INFO=2,3.) */
00176 
00177 /*  W       (output) DOUBLE PRECISION array, dimension (N) */
00178 /*          On exit, the first M elements of W will contain the */
00179 /*          eigenvalue approximations. DLARRD computes an interval */
00180 /*          I_j = (a_j, b_j] that includes eigenvalue j. The eigenvalue */
00181 /*          approximation is given as the interval midpoint */
00182 /*          W(j)= ( a_j + b_j)/2. The corresponding error is bounded by */
00183 /*          WERR(j) = abs( a_j - b_j)/2 */
00184 
00185 /*  WERR    (output) DOUBLE PRECISION array, dimension (N) */
00186 /*          The error bound on the corresponding eigenvalue approximation */
00187 /*          in W. */
00188 
00189 /*  WL      (output) DOUBLE PRECISION */
00190 /*  WU      (output) DOUBLE PRECISION */
00191 /*          The interval (WL, WU] contains all the wanted eigenvalues. */
00192 /*          If RANGE='V', then WL=VL and WU=VU. */
00193 /*          If RANGE='A', then WL and WU are the global Gerschgorin bounds */
00194 /*                        on the spectrum. */
00195 /*          If RANGE='I', then WL and WU are computed by DLAEBZ from the */
00196 /*                        index range specified. */
00197 
00198 /*  IBLOCK  (output) INTEGER array, dimension (N) */
00199 /*          At each row/column j where E(j) is zero or small, the */
00200 /*          matrix T is considered to split into a block diagonal */
00201 /*          matrix.  On exit, if INFO = 0, IBLOCK(i) specifies to which */
00202 /*          block (from 1 to the number of blocks) the eigenvalue W(i) */
00203 /*          belongs.  (DLARRD may use the remaining N-M elements as */
00204 /*          workspace.) */
00205 
00206 /*  INDEXW  (output) INTEGER array, dimension (N) */
00207 /*          The indices of the eigenvalues within each block (submatrix); */
00208 /*          for example, INDEXW(i)= j and IBLOCK(i)=k imply that the */
00209 /*          i-th eigenvalue W(i) is the j-th eigenvalue in block k. */
00210 
00211 /*  WORK    (workspace) DOUBLE PRECISION array, dimension (4*N) */
00212 
00213 /*  IWORK   (workspace) INTEGER array, dimension (3*N) */
00214 
00215 /*  INFO    (output) INTEGER */
00216 /*          = 0:  successful exit */
00217 /*          < 0:  if INFO = -i, the i-th argument had an illegal value */
00218 /*          > 0:  some or all of the eigenvalues failed to converge or */
00219 /*                were not computed: */
00220 /*                =1 or 3: Bisection failed to converge for some */
00221 /*                        eigenvalues; these eigenvalues are flagged by a */
00222 /*                        negative block number.  The effect is that the */
00223 /*                        eigenvalues may not be as accurate as the */
00224 /*                        absolute and relative tolerances.  This is */
00225 /*                        generally caused by unexpectedly inaccurate */
00226 /*                        arithmetic. */
00227 /*                =2 or 3: RANGE='I' only: Not all of the eigenvalues */
00228 /*                        IL:IU were found. */
00229 /*                        Effect: M < IU+1-IL */
00230 /*                        Cause:  non-monotonic arithmetic, causing the */
00231 /*                                Sturm sequence to be non-monotonic. */
00232 /*                        Cure:   recalculate, using RANGE='A', and pick */
00233 /*                                out eigenvalues IL:IU.  In some cases, */
00234 /*                                increasing the PARAMETER "FUDGE" may */
00235 /*                                make things work. */
00236 /*                = 4:    RANGE='I', and the Gershgorin interval */
00237 /*                        initially used was too small.  No eigenvalues */
00238 /*                        were computed. */
00239 /*                        Probable cause: your machine has sloppy */
00240 /*                                        floating-point arithmetic. */
00241 /*                        Cure: Increase the PARAMETER "FUDGE", */
00242 /*                              recompile, and try again. */
00243 
00244 /*  Internal Parameters */
00245 /*  =================== */
00246 
00247 /*  FUDGE   DOUBLE PRECISION, default = 2 */
00248 /*          A "fudge factor" to widen the Gershgorin intervals.  Ideally, */
00249 /*          a value of 1 should work, but on machines with sloppy */
00250 /*          arithmetic, this needs to be larger.  The default for */
00251 /*          publicly released versions should be large enough to handle */
00252 /*          the worst machine around.  Note that this has no effect */
00253 /*          on accuracy of the solution. */
00254 
00255 /*  Based on contributions by */
00256 /*     W. Kahan, University of California, Berkeley, USA */
00257 /*     Beresford Parlett, University of California, Berkeley, USA */
00258 /*     Jim Demmel, University of California, Berkeley, USA */
00259 /*     Inderjit Dhillon, University of Texas, Austin, USA */
00260 /*     Osni Marques, LBNL/NERSC, USA */
00261 /*     Christof Voemel, University of California, Berkeley, USA */
00262 
00263 /*  ===================================================================== */
00264 
00265 /*     .. Parameters .. */
00266 /*     .. */
00267 /*     .. Local Scalars .. */
00268 /*     .. */
00269 /*     .. Local Arrays .. */
00270 /*     .. */
00271 /*     .. External Functions .. */
00272 /*     .. */
00273 /*     .. External Subroutines .. */
00274 /*     .. */
00275 /*     .. Intrinsic Functions .. */
00276 /*     .. */
00277 /*     .. Executable Statements .. */
00278 
00279     /* Parameter adjustments */
00280     --iwork;
00281     --work;
00282     --indexw;
00283     --iblock;
00284     --werr;
00285     --w;
00286     --isplit;
00287     --e2;
00288     --e;
00289     --d__;
00290     --gers;
00291 
00292     /* Function Body */
00293     *info = 0;
00294 
00295 /*     Decode RANGE */
00296 
00297     if (lsame_(range, "A")) {
00298         irange = 1;
00299     } else if (lsame_(range, "V")) {
00300         irange = 2;
00301     } else if (lsame_(range, "I")) {
00302         irange = 3;
00303     } else {
00304         irange = 0;
00305     }
00306 
00307 /*     Check for Errors */
00308 
00309     if (irange <= 0) {
00310         *info = -1;
00311     } else if (! (lsame_(order, "B") || lsame_(order, 
00312             "E"))) {
00313         *info = -2;
00314     } else if (*n < 0) {
00315         *info = -3;
00316     } else if (irange == 2) {
00317         if (*vl >= *vu) {
00318             *info = -5;
00319         }
00320     } else if (irange == 3 && (*il < 1 || *il > max(1,*n))) {
00321         *info = -6;
00322     } else if (irange == 3 && (*iu < min(*n,*il) || *iu > *n)) {
00323         *info = -7;
00324     }
00325 
00326     if (*info != 0) {
00327         return 0;
00328     }
00329 /*     Initialize error flags */
00330     *info = 0;
00331     ncnvrg = FALSE_;
00332     toofew = FALSE_;
00333 /*     Quick return if possible */
00334     *m = 0;
00335     if (*n == 0) {
00336         return 0;
00337     }
00338 /*     Simplification: */
00339     if (irange == 3 && *il == 1 && *iu == *n) {
00340         irange = 1;
00341     }
00342 /*     Get machine constants */
00343     eps = dlamch_("P");
00344     uflow = dlamch_("U");
00345 /*     Special Case when N=1 */
00346 /*     Treat case of 1x1 matrix for quick return */
00347     if (*n == 1) {
00348         if (irange == 1 || irange == 2 && d__[1] > *vl && d__[1] <= *vu || 
00349                 irange == 3 && *il == 1 && *iu == 1) {
00350             *m = 1;
00351             w[1] = d__[1];
00352 /*           The computation error of the eigenvalue is zero */
00353             werr[1] = 0.;
00354             iblock[1] = 1;
00355             indexw[1] = 1;
00356         }
00357         return 0;
00358     }
00359 /*     NB is the minimum vector length for vector bisection, or 0 */
00360 /*     if only scalar is to be done. */
00361     nb = ilaenv_(&c__1, "DSTEBZ", " ", n, &c_n1, &c_n1, &c_n1);
00362     if (nb <= 1) {
00363         nb = 0;
00364     }
00365 /*     Find global spectral radius */
00366     gl = d__[1];
00367     gu = d__[1];
00368     i__1 = *n;
00369     for (i__ = 1; i__ <= i__1; ++i__) {
00370 /* Computing MIN */
00371         d__1 = gl, d__2 = gers[(i__ << 1) - 1];
00372         gl = min(d__1,d__2);
00373 /* Computing MAX */
00374         d__1 = gu, d__2 = gers[i__ * 2];
00375         gu = max(d__1,d__2);
00376 /* L5: */
00377     }
00378 /*     Compute global Gerschgorin bounds and spectral diameter */
00379 /* Computing MAX */
00380     d__1 = abs(gl), d__2 = abs(gu);
00381     tnorm = max(d__1,d__2);
00382     gl = gl - tnorm * 2. * eps * *n - *pivmin * 4.;
00383     gu = gu + tnorm * 2. * eps * *n + *pivmin * 4.;
00384 /*     [JAN/28/2009] remove the line below since SPDIAM variable not use */
00385 /*     SPDIAM = GU - GL */
00386 /*     Input arguments for DLAEBZ: */
00387 /*     The relative tolerance.  An interval (a,b] lies within */
00388 /*     "relative tolerance" if  b-a < RELTOL*max(|a|,|b|), */
00389     rtoli = *reltol;
00390 /*     Set the absolute tolerance for interval convergence to zero to force */
00391 /*     interval convergence based on relative size of the interval. */
00392 /*     This is dangerous because intervals might not converge when RELTOL is */
00393 /*     small. But at least a very small number should be selected so that for */
00394 /*     strongly graded matrices, the code can get relatively accurate */
00395 /*     eigenvalues. */
00396     atoli = uflow * 4. + *pivmin * 4.;
00397     if (irange == 3) {
00398 /*        RANGE='I': Compute an interval containing eigenvalues */
00399 /*        IL through IU. The initial interval [GL,GU] from the global */
00400 /*        Gerschgorin bounds GL and GU is refined by DLAEBZ. */
00401         itmax = (integer) ((log(tnorm + *pivmin) - log(*pivmin)) / log(2.)) + 
00402                 2;
00403         work[*n + 1] = gl;
00404         work[*n + 2] = gl;
00405         work[*n + 3] = gu;
00406         work[*n + 4] = gu;
00407         work[*n + 5] = gl;
00408         work[*n + 6] = gu;
00409         iwork[1] = -1;
00410         iwork[2] = -1;
00411         iwork[3] = *n + 1;
00412         iwork[4] = *n + 1;
00413         iwork[5] = *il - 1;
00414         iwork[6] = *iu;
00415 
00416         dlaebz_(&c__3, &itmax, n, &c__2, &c__2, &nb, &atoli, &rtoli, pivmin, &
00417                 d__[1], &e[1], &e2[1], &iwork[5], &work[*n + 1], &work[*n + 5]
00418 , &iout, &iwork[1], &w[1], &iblock[1], &iinfo);
00419         if (iinfo != 0) {
00420             *info = iinfo;
00421             return 0;
00422         }
00423 /*        On exit, output intervals may not be ordered by ascending negcount */
00424         if (iwork[6] == *iu) {
00425             *wl = work[*n + 1];
00426             wlu = work[*n + 3];
00427             nwl = iwork[1];
00428             *wu = work[*n + 4];
00429             wul = work[*n + 2];
00430             nwu = iwork[4];
00431         } else {
00432             *wl = work[*n + 2];
00433             wlu = work[*n + 4];
00434             nwl = iwork[2];
00435             *wu = work[*n + 3];
00436             wul = work[*n + 1];
00437             nwu = iwork[3];
00438         }
00439 /*        On exit, the interval [WL, WLU] contains a value with negcount NWL, */
00440 /*        and [WUL, WU] contains a value with negcount NWU. */
00441         if (nwl < 0 || nwl >= *n || nwu < 1 || nwu > *n) {
00442             *info = 4;
00443             return 0;
00444         }
00445     } else if (irange == 2) {
00446         *wl = *vl;
00447         *wu = *vu;
00448     } else if (irange == 1) {
00449         *wl = gl;
00450         *wu = gu;
00451     }
00452 /*     Find Eigenvalues -- Loop Over blocks and recompute NWL and NWU. */
00453 /*     NWL accumulates the number of eigenvalues .le. WL, */
00454 /*     NWU accumulates the number of eigenvalues .le. WU */
00455     *m = 0;
00456     iend = 0;
00457     *info = 0;
00458     nwl = 0;
00459     nwu = 0;
00460 
00461     i__1 = *nsplit;
00462     for (jblk = 1; jblk <= i__1; ++jblk) {
00463         ioff = iend;
00464         ibegin = ioff + 1;
00465         iend = isplit[jblk];
00466         in = iend - ioff;
00467 
00468         if (in == 1) {
00469 /*           1x1 block */
00470             if (*wl >= d__[ibegin] - *pivmin) {
00471                 ++nwl;
00472             }
00473             if (*wu >= d__[ibegin] - *pivmin) {
00474                 ++nwu;
00475             }
00476             if (irange == 1 || *wl < d__[ibegin] - *pivmin && *wu >= d__[
00477                     ibegin] - *pivmin) {
00478                 ++(*m);
00479                 w[*m] = d__[ibegin];
00480                 werr[*m] = 0.;
00481 /*              The gap for a single block doesn't matter for the later */
00482 /*              algorithm and is assigned an arbitrary large value */
00483                 iblock[*m] = jblk;
00484                 indexw[*m] = 1;
00485             }
00486 /*        Disabled 2x2 case because of a failure on the following matrix */
00487 /*        RANGE = 'I', IL = IU = 4 */
00488 /*          Original Tridiagonal, d = [ */
00489 /*           -0.150102010615740E+00 */
00490 /*           -0.849897989384260E+00 */
00491 /*           -0.128208148052635E-15 */
00492 /*            0.128257718286320E-15 */
00493 /*          ]; */
00494 /*          e = [ */
00495 /*           -0.357171383266986E+00 */
00496 /*           -0.180411241501588E-15 */
00497 /*           -0.175152352710251E-15 */
00498 /*          ]; */
00499 
00500 /*         ELSE IF( IN.EQ.2 ) THEN */
00501 /* *           2x2 block */
00502 /*            DISC = SQRT( (HALF*(D(IBEGIN)-D(IEND)))**2 + E(IBEGIN)**2 ) */
00503 /*            TMP1 = HALF*(D(IBEGIN)+D(IEND)) */
00504 /*            L1 = TMP1 - DISC */
00505 /*            IF( WL.GE. L1-PIVMIN ) */
00506 /*     $         NWL = NWL + 1 */
00507 /*            IF( WU.GE. L1-PIVMIN ) */
00508 /*     $         NWU = NWU + 1 */
00509 /*            IF( IRANGE.EQ.ALLRNG .OR. ( WL.LT.L1-PIVMIN .AND. WU.GE. */
00510 /*     $          L1-PIVMIN ) ) THEN */
00511 /*               M = M + 1 */
00512 /*               W( M ) = L1 */
00513 /* *              The uncertainty of eigenvalues of a 2x2 matrix is very small */
00514 /*               WERR( M ) = EPS * ABS( W( M ) ) * TWO */
00515 /*               IBLOCK( M ) = JBLK */
00516 /*               INDEXW( M ) = 1 */
00517 /*            ENDIF */
00518 /*            L2 = TMP1 + DISC */
00519 /*            IF( WL.GE. L2-PIVMIN ) */
00520 /*     $         NWL = NWL + 1 */
00521 /*            IF( WU.GE. L2-PIVMIN ) */
00522 /*     $         NWU = NWU + 1 */
00523 /*            IF( IRANGE.EQ.ALLRNG .OR. ( WL.LT.L2-PIVMIN .AND. WU.GE. */
00524 /*     $          L2-PIVMIN ) ) THEN */
00525 /*               M = M + 1 */
00526 /*               W( M ) = L2 */
00527 /* *              The uncertainty of eigenvalues of a 2x2 matrix is very small */
00528 /*               WERR( M ) = EPS * ABS( W( M ) ) * TWO */
00529 /*               IBLOCK( M ) = JBLK */
00530 /*               INDEXW( M ) = 2 */
00531 /*            ENDIF */
00532         } else {
00533 /*           General Case - block of size IN >= 2 */
00534 /*           Compute local Gerschgorin interval and use it as the initial */
00535 /*           interval for DLAEBZ */
00536             gu = d__[ibegin];
00537             gl = d__[ibegin];
00538             tmp1 = 0.;
00539             i__2 = iend;
00540             for (j = ibegin; j <= i__2; ++j) {
00541 /* Computing MIN */
00542                 d__1 = gl, d__2 = gers[(j << 1) - 1];
00543                 gl = min(d__1,d__2);
00544 /* Computing MAX */
00545                 d__1 = gu, d__2 = gers[j * 2];
00546                 gu = max(d__1,d__2);
00547 /* L40: */
00548             }
00549 /*           [JAN/28/2009] */
00550 /*           change SPDIAM by TNORM in lines 2 and 3 thereafter */
00551 /*           line 1: remove computation of SPDIAM (not useful anymore) */
00552 /*           SPDIAM = GU - GL */
00553 /*           GL = GL - FUDGE*SPDIAM*EPS*IN - FUDGE*PIVMIN */
00554 /*           GU = GU + FUDGE*SPDIAM*EPS*IN + FUDGE*PIVMIN */
00555             gl = gl - tnorm * 2. * eps * in - *pivmin * 2.;
00556             gu = gu + tnorm * 2. * eps * in + *pivmin * 2.;
00557 
00558             if (irange > 1) {
00559                 if (gu < *wl) {
00560 /*                 the local block contains none of the wanted eigenvalues */
00561                     nwl += in;
00562                     nwu += in;
00563                     goto L70;
00564                 }
00565 /*              refine search interval if possible, only range (WL,WU] matters */
00566                 gl = max(gl,*wl);
00567                 gu = min(gu,*wu);
00568                 if (gl >= gu) {
00569                     goto L70;
00570                 }
00571             }
00572 /*           Find negcount of initial interval boundaries GL and GU */
00573             work[*n + 1] = gl;
00574             work[*n + in + 1] = gu;
00575             dlaebz_(&c__1, &c__0, &in, &in, &c__1, &nb, &atoli, &rtoli, 
00576                     pivmin, &d__[ibegin], &e[ibegin], &e2[ibegin], idumma, &
00577                     work[*n + 1], &work[*n + (in << 1) + 1], &im, &iwork[1], &
00578                     w[*m + 1], &iblock[*m + 1], &iinfo);
00579             if (iinfo != 0) {
00580                 *info = iinfo;
00581                 return 0;
00582             }
00583 
00584             nwl += iwork[1];
00585             nwu += iwork[in + 1];
00586             iwoff = *m - iwork[1];
00587 /*           Compute Eigenvalues */
00588             itmax = (integer) ((log(gu - gl + *pivmin) - log(*pivmin)) / log(
00589                     2.)) + 2;
00590             dlaebz_(&c__2, &itmax, &in, &in, &c__1, &nb, &atoli, &rtoli, 
00591                     pivmin, &d__[ibegin], &e[ibegin], &e2[ibegin], idumma, &
00592                     work[*n + 1], &work[*n + (in << 1) + 1], &iout, &iwork[1], 
00593                      &w[*m + 1], &iblock[*m + 1], &iinfo);
00594             if (iinfo != 0) {
00595                 *info = iinfo;
00596                 return 0;
00597             }
00598 
00599 /*           Copy eigenvalues into W and IBLOCK */
00600 /*           Use -JBLK for block number for unconverged eigenvalues. */
00601 /*           Loop over the number of output intervals from DLAEBZ */
00602             i__2 = iout;
00603             for (j = 1; j <= i__2; ++j) {
00604 /*              eigenvalue approximation is middle point of interval */
00605                 tmp1 = (work[j + *n] + work[j + in + *n]) * .5;
00606 /*              semi length of error interval */
00607                 tmp2 = (d__1 = work[j + *n] - work[j + in + *n], abs(d__1)) * 
00608                         .5;
00609                 if (j > iout - iinfo) {
00610 /*                 Flag non-convergence. */
00611                     ncnvrg = TRUE_;
00612                     ib = -jblk;
00613                 } else {
00614                     ib = jblk;
00615                 }
00616                 i__3 = iwork[j + in] + iwoff;
00617                 for (je = iwork[j] + 1 + iwoff; je <= i__3; ++je) {
00618                     w[je] = tmp1;
00619                     werr[je] = tmp2;
00620                     indexw[je] = je - iwoff;
00621                     iblock[je] = ib;
00622 /* L50: */
00623                 }
00624 /* L60: */
00625             }
00626 
00627             *m += im;
00628         }
00629 L70:
00630         ;
00631     }
00632 /*     If RANGE='I', then (WL,WU) contains eigenvalues NWL+1,...,NWU */
00633 /*     If NWL+1 < IL or NWU > IU, discard extra eigenvalues. */
00634     if (irange == 3) {
00635         idiscl = *il - 1 - nwl;
00636         idiscu = nwu - *iu;
00637 
00638         if (idiscl > 0) {
00639             im = 0;
00640             i__1 = *m;
00641             for (je = 1; je <= i__1; ++je) {
00642 /*              Remove some of the smallest eigenvalues from the left so that */
00643 /*              at the end IDISCL =0. Move all eigenvalues up to the left. */
00644                 if (w[je] <= wlu && idiscl > 0) {
00645                     --idiscl;
00646                 } else {
00647                     ++im;
00648                     w[im] = w[je];
00649                     werr[im] = werr[je];
00650                     indexw[im] = indexw[je];
00651                     iblock[im] = iblock[je];
00652                 }
00653 /* L80: */
00654             }
00655             *m = im;
00656         }
00657         if (idiscu > 0) {
00658 /*           Remove some of the largest eigenvalues from the right so that */
00659 /*           at the end IDISCU =0. Move all eigenvalues up to the left. */
00660             im = *m + 1;
00661             for (je = *m; je >= 1; --je) {
00662                 if (w[je] >= wul && idiscu > 0) {
00663                     --idiscu;
00664                 } else {
00665                     --im;
00666                     w[im] = w[je];
00667                     werr[im] = werr[je];
00668                     indexw[im] = indexw[je];
00669                     iblock[im] = iblock[je];
00670                 }
00671 /* L81: */
00672             }
00673             jee = 0;
00674             i__1 = *m;
00675             for (je = im; je <= i__1; ++je) {
00676                 ++jee;
00677                 w[jee] = w[je];
00678                 werr[jee] = werr[je];
00679                 indexw[jee] = indexw[je];
00680                 iblock[jee] = iblock[je];
00681 /* L82: */
00682             }
00683             *m = *m - im + 1;
00684         }
00685         if (idiscl > 0 || idiscu > 0) {
00686 /*           Code to deal with effects of bad arithmetic. (If N(w) is */
00687 /*           monotone non-decreasing, this should never happen.) */
00688 /*           Some low eigenvalues to be discarded are not in (WL,WLU], */
00689 /*           or high eigenvalues to be discarded are not in (WUL,WU] */
00690 /*           so just kill off the smallest IDISCL/largest IDISCU */
00691 /*           eigenvalues, by marking the corresponding IBLOCK = 0 */
00692             if (idiscl > 0) {
00693                 wkill = *wu;
00694                 i__1 = idiscl;
00695                 for (jdisc = 1; jdisc <= i__1; ++jdisc) {
00696                     iw = 0;
00697                     i__2 = *m;
00698                     for (je = 1; je <= i__2; ++je) {
00699                         if (iblock[je] != 0 && (w[je] < wkill || iw == 0)) {
00700                             iw = je;
00701                             wkill = w[je];
00702                         }
00703 /* L90: */
00704                     }
00705                     iblock[iw] = 0;
00706 /* L100: */
00707                 }
00708             }
00709             if (idiscu > 0) {
00710                 wkill = *wl;
00711                 i__1 = idiscu;
00712                 for (jdisc = 1; jdisc <= i__1; ++jdisc) {
00713                     iw = 0;
00714                     i__2 = *m;
00715                     for (je = 1; je <= i__2; ++je) {
00716                         if (iblock[je] != 0 && (w[je] >= wkill || iw == 0)) {
00717                             iw = je;
00718                             wkill = w[je];
00719                         }
00720 /* L110: */
00721                     }
00722                     iblock[iw] = 0;
00723 /* L120: */
00724                 }
00725             }
00726 /*           Now erase all eigenvalues with IBLOCK set to zero */
00727             im = 0;
00728             i__1 = *m;
00729             for (je = 1; je <= i__1; ++je) {
00730                 if (iblock[je] != 0) {
00731                     ++im;
00732                     w[im] = w[je];
00733                     werr[im] = werr[je];
00734                     indexw[im] = indexw[je];
00735                     iblock[im] = iblock[je];
00736                 }
00737 /* L130: */
00738             }
00739             *m = im;
00740         }
00741         if (idiscl < 0 || idiscu < 0) {
00742             toofew = TRUE_;
00743         }
00744     }
00745 
00746     if (irange == 1 && *m != *n || irange == 3 && *m != *iu - *il + 1) {
00747         toofew = TRUE_;
00748     }
00749 /*     If ORDER='B', do nothing the eigenvalues are already sorted by */
00750 /*        block. */
00751 /*     If ORDER='E', sort the eigenvalues from smallest to largest */
00752     if (lsame_(order, "E") && *nsplit > 1) {
00753         i__1 = *m - 1;
00754         for (je = 1; je <= i__1; ++je) {
00755             ie = 0;
00756             tmp1 = w[je];
00757             i__2 = *m;
00758             for (j = je + 1; j <= i__2; ++j) {
00759                 if (w[j] < tmp1) {
00760                     ie = j;
00761                     tmp1 = w[j];
00762                 }
00763 /* L140: */
00764             }
00765             if (ie != 0) {
00766                 tmp2 = werr[ie];
00767                 itmp1 = iblock[ie];
00768                 itmp2 = indexw[ie];
00769                 w[ie] = w[je];
00770                 werr[ie] = werr[je];
00771                 iblock[ie] = iblock[je];
00772                 indexw[ie] = indexw[je];
00773                 w[je] = tmp1;
00774                 werr[je] = tmp2;
00775                 iblock[je] = itmp1;
00776                 indexw[je] = itmp2;
00777             }
00778 /* L150: */
00779         }
00780     }
00781 
00782     *info = 0;
00783     if (ncnvrg) {
00784         ++(*info);
00785     }
00786     if (toofew) {
00787         *info += 2;
00788     }
00789     return 0;
00790 
00791 /*     End of DLARRD */
00792 
00793 } /* dlarrd_ */


swiftnav
Author(s):
autogenerated on Sat Jun 8 2019 18:55:46