File:  [local] / rpl / lapack / lapack / dlarrj.f
Revision 1.4: download - view: text, annotated - select for diffs - revision graph
Fri Aug 6 15:32:29 2010 UTC (13 years, 9 months ago) by bertrand
Branches: MAIN
CVS tags: HEAD
Cohérence

    1:       SUBROUTINE DLARRJ( N, D, E2, IFIRST, ILAST,
    2:      $                   RTOL, OFFSET, W, WERR, WORK, IWORK,
    3:      $                   PIVMIN, SPDIAM, INFO )
    4: *
    5: *  -- LAPACK auxiliary routine (version 3.2) --
    6: *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
    7: *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
    8: *     November 2006
    9: *
   10: *     .. Scalar Arguments ..
   11:       INTEGER            IFIRST, ILAST, INFO, N, OFFSET
   12:       DOUBLE PRECISION   PIVMIN, RTOL, SPDIAM
   13: *     ..
   14: *     .. Array Arguments ..
   15:       INTEGER            IWORK( * )
   16:       DOUBLE PRECISION   D( * ), E2( * ), W( * ),
   17:      $                   WERR( * ), WORK( * )
   18: *     ..
   19: *
   20: *  Purpose
   21: *  =======
   22: *
   23: *  Given the initial eigenvalue approximations of T, DLARRJ
   24: *  does  bisection to refine the eigenvalues of T,
   25: *  W( IFIRST-OFFSET ) through W( ILAST-OFFSET ), to more accuracy. Initial
   26: *  guesses for these eigenvalues are input in W, the corresponding estimate
   27: *  of the error in these guesses in WERR. During bisection, intervals
   28: *  [left, right] are maintained by storing their mid-points and
   29: *  semi-widths in the arrays W and WERR respectively.
   30: *
   31: *  Arguments
   32: *  =========
   33: *
   34: *  N       (input) INTEGER
   35: *          The order of the matrix.
   36: *
   37: *  D       (input) DOUBLE PRECISION array, dimension (N)
   38: *          The N diagonal elements of T.
   39: *
   40: *  E2      (input) DOUBLE PRECISION array, dimension (N-1)
   41: *          The Squares of the (N-1) subdiagonal elements of T.
   42: *
   43: *  IFIRST  (input) INTEGER
   44: *          The index of the first eigenvalue to be computed.
   45: *
   46: *  ILAST   (input) INTEGER
   47: *          The index of the last eigenvalue to be computed.
   48: *
   49: *  RTOL   (input) DOUBLE PRECISION
   50: *          Tolerance for the convergence of the bisection intervals.
   51: *          An interval [LEFT,RIGHT] has converged if
   52: *          RIGHT-LEFT.LT.RTOL*MAX(|LEFT|,|RIGHT|).
   53: *
   54: *  OFFSET  (input) INTEGER
   55: *          Offset for the arrays W and WERR, i.e., the IFIRST-OFFSET
   56: *          through ILAST-OFFSET elements of these arrays are to be used.
   57: *
   58: *  W       (input/output) DOUBLE PRECISION array, dimension (N)
   59: *          On input, W( IFIRST-OFFSET ) through W( ILAST-OFFSET ) are
   60: *          estimates of the eigenvalues of L D L^T indexed IFIRST through
   61: *          ILAST.
   62: *          On output, these estimates are refined.
   63: *
   64: *  WERR    (input/output) DOUBLE PRECISION array, dimension (N)
   65: *          On input, WERR( IFIRST-OFFSET ) through WERR( ILAST-OFFSET ) are
   66: *          the errors in the estimates of the corresponding elements in W.
   67: *          On output, these errors are refined.
   68: *
   69: *  WORK    (workspace) DOUBLE PRECISION array, dimension (2*N)
   70: *          Workspace.
   71: *
   72: *  IWORK   (workspace) INTEGER array, dimension (2*N)
   73: *          Workspace.
   74: *
   75: *  PIVMIN  (input) DOUBLE PRECISION
   76: *          The minimum pivot in the Sturm sequence for T.
   77: *
   78: *  SPDIAM  (input) DOUBLE PRECISION
   79: *          The spectral diameter of T.
   80: *
   81: *  INFO    (output) INTEGER
   82: *          Error flag.
   83: *
   84: *  Further Details
   85: *  ===============
   86: *
   87: *  Based on contributions by
   88: *     Beresford Parlett, University of California, Berkeley, USA
   89: *     Jim Demmel, University of California, Berkeley, USA
   90: *     Inderjit Dhillon, University of Texas, Austin, USA
   91: *     Osni Marques, LBNL/NERSC, USA
   92: *     Christof Voemel, University of California, Berkeley, USA
   93: *
   94: *  =====================================================================
   95: *
   96: *     .. Parameters ..
   97:       DOUBLE PRECISION   ZERO, ONE, TWO, HALF
   98:       PARAMETER        ( ZERO = 0.0D0, ONE = 1.0D0, TWO = 2.0D0,
   99:      $                   HALF = 0.5D0 )
  100:       INTEGER   MAXITR
  101: *     ..
  102: *     .. Local Scalars ..
  103:       INTEGER            CNT, I, I1, I2, II, ITER, J, K, NEXT, NINT,
  104:      $                   OLNINT, P, PREV, SAVI1
  105:       DOUBLE PRECISION   DPLUS, FAC, LEFT, MID, RIGHT, S, TMP, WIDTH
  106: *
  107: *     ..
  108: *     .. Intrinsic Functions ..
  109:       INTRINSIC          ABS, MAX
  110: *     ..
  111: *     .. Executable Statements ..
  112: *
  113:       INFO = 0
  114: *
  115:       MAXITR = INT( ( LOG( SPDIAM+PIVMIN )-LOG( PIVMIN ) ) /
  116:      $           LOG( TWO ) ) + 2
  117: *
  118: *     Initialize unconverged intervals in [ WORK(2*I-1), WORK(2*I) ].
  119: *     The Sturm Count, Count( WORK(2*I-1) ) is arranged to be I-1, while
  120: *     Count( WORK(2*I) ) is stored in IWORK( 2*I ). The integer IWORK( 2*I-1 )
  121: *     for an unconverged interval is set to the index of the next unconverged
  122: *     interval, and is -1 or 0 for a converged interval. Thus a linked
  123: *     list of unconverged intervals is set up.
  124: *
  125: 
  126:       I1 = IFIRST
  127:       I2 = ILAST
  128: *     The number of unconverged intervals
  129:       NINT = 0
  130: *     The last unconverged interval found
  131:       PREV = 0
  132:       DO 75 I = I1, I2
  133:          K = 2*I
  134:          II = I - OFFSET
  135:          LEFT = W( II ) - WERR( II )
  136:          MID = W(II)
  137:          RIGHT = W( II ) + WERR( II )
  138:          WIDTH = RIGHT - MID
  139:          TMP = MAX( ABS( LEFT ), ABS( RIGHT ) )
  140: 
  141: *        The following test prevents the test of converged intervals
  142:          IF( WIDTH.LT.RTOL*TMP ) THEN
  143: *           This interval has already converged and does not need refinement.
  144: *           (Note that the gaps might change through refining the
  145: *            eigenvalues, however, they can only get bigger.)
  146: *           Remove it from the list.
  147:             IWORK( K-1 ) = -1
  148: *           Make sure that I1 always points to the first unconverged interval
  149:             IF((I.EQ.I1).AND.(I.LT.I2)) I1 = I + 1
  150:             IF((PREV.GE.I1).AND.(I.LE.I2)) IWORK( 2*PREV-1 ) = I + 1
  151:          ELSE
  152: *           unconverged interval found
  153:             PREV = I
  154: *           Make sure that [LEFT,RIGHT] contains the desired eigenvalue
  155: *
  156: *           Do while( CNT(LEFT).GT.I-1 )
  157: *
  158:             FAC = ONE
  159:  20         CONTINUE
  160:             CNT = 0
  161:             S = LEFT
  162:             DPLUS = D( 1 ) - S
  163:             IF( DPLUS.LT.ZERO ) CNT = CNT + 1
  164:             DO 30 J = 2, N
  165:                DPLUS = D( J ) - S - E2( J-1 )/DPLUS
  166:                IF( DPLUS.LT.ZERO ) CNT = CNT + 1
  167:  30         CONTINUE
  168:             IF( CNT.GT.I-1 ) THEN
  169:                LEFT = LEFT - WERR( II )*FAC
  170:                FAC = TWO*FAC
  171:                GO TO 20
  172:             END IF
  173: *
  174: *           Do while( CNT(RIGHT).LT.I )
  175: *
  176:             FAC = ONE
  177:  50         CONTINUE
  178:             CNT = 0
  179:             S = RIGHT
  180:             DPLUS = D( 1 ) - S
  181:             IF( DPLUS.LT.ZERO ) CNT = CNT + 1
  182:             DO 60 J = 2, N
  183:                DPLUS = D( J ) - S - E2( J-1 )/DPLUS
  184:                IF( DPLUS.LT.ZERO ) CNT = CNT + 1
  185:  60         CONTINUE
  186:             IF( CNT.LT.I ) THEN
  187:                RIGHT = RIGHT + WERR( II )*FAC
  188:                FAC = TWO*FAC
  189:                GO TO 50
  190:             END IF
  191:             NINT = NINT + 1
  192:             IWORK( K-1 ) = I + 1
  193:             IWORK( K ) = CNT
  194:          END IF
  195:          WORK( K-1 ) = LEFT
  196:          WORK( K ) = RIGHT
  197:  75   CONTINUE
  198: 
  199: 
  200:       SAVI1 = I1
  201: *
  202: *     Do while( NINT.GT.0 ), i.e. there are still unconverged intervals
  203: *     and while (ITER.LT.MAXITR)
  204: *
  205:       ITER = 0
  206:  80   CONTINUE
  207:       PREV = I1 - 1
  208:       I = I1
  209:       OLNINT = NINT
  210: 
  211:       DO 100 P = 1, OLNINT
  212:          K = 2*I
  213:          II = I - OFFSET
  214:          NEXT = IWORK( K-1 )
  215:          LEFT = WORK( K-1 )
  216:          RIGHT = WORK( K )
  217:          MID = HALF*( LEFT + RIGHT )
  218: 
  219: *        semiwidth of interval
  220:          WIDTH = RIGHT - MID
  221:          TMP = MAX( ABS( LEFT ), ABS( RIGHT ) )
  222: 
  223:          IF( ( WIDTH.LT.RTOL*TMP ) .OR.
  224:      $      (ITER.EQ.MAXITR) )THEN
  225: *           reduce number of unconverged intervals
  226:             NINT = NINT - 1
  227: *           Mark interval as converged.
  228:             IWORK( K-1 ) = 0
  229:             IF( I1.EQ.I ) THEN
  230:                I1 = NEXT
  231:             ELSE
  232: *              Prev holds the last unconverged interval previously examined
  233:                IF(PREV.GE.I1) IWORK( 2*PREV-1 ) = NEXT
  234:             END IF
  235:             I = NEXT
  236:             GO TO 100
  237:          END IF
  238:          PREV = I
  239: *
  240: *        Perform one bisection step
  241: *
  242:          CNT = 0
  243:          S = MID
  244:          DPLUS = D( 1 ) - S
  245:          IF( DPLUS.LT.ZERO ) CNT = CNT + 1
  246:          DO 90 J = 2, N
  247:             DPLUS = D( J ) - S - E2( J-1 )/DPLUS
  248:             IF( DPLUS.LT.ZERO ) CNT = CNT + 1
  249:  90      CONTINUE
  250:          IF( CNT.LE.I-1 ) THEN
  251:             WORK( K-1 ) = MID
  252:          ELSE
  253:             WORK( K ) = MID
  254:          END IF
  255:          I = NEXT
  256: 
  257:  100  CONTINUE
  258:       ITER = ITER + 1
  259: *     do another loop if there are still unconverged intervals
  260: *     However, in the last iteration, all intervals are accepted
  261: *     since this is the best we can do.
  262:       IF( ( NINT.GT.0 ).AND.(ITER.LE.MAXITR) ) GO TO 80
  263: *
  264: *
  265: *     At this point, all the intervals have converged
  266:       DO 110 I = SAVI1, ILAST
  267:          K = 2*I
  268:          II = I - OFFSET
  269: *        All intervals marked by '0' have been refined.
  270:          IF( IWORK( K-1 ).EQ.0 ) THEN
  271:             W( II ) = HALF*( WORK( K-1 )+WORK( K ) )
  272:             WERR( II ) = WORK( K ) - W( II )
  273:          END IF
  274:  110  CONTINUE
  275: *
  276: 
  277:       RETURN
  278: *
  279: *     End of DLARRJ
  280: *
  281:       END

CVSweb interface <joel.bertrand@systella.fr>