File:  [local] / rpl / lapack / lapack / zlahef_rk.f
Revision 1.2: download - view: text, annotated - select for diffs - revision graph
Sat Jun 17 11:06:53 2017 UTC (6 years, 11 months ago) by bertrand
Branches: MAIN
CVS tags: rpl-4_1_27, rpl-4_1_26, HEAD
Cohérence.

    1: *> \brief \b ZLAHEF_RK computes a partial factorization of a complex Hermitian indefinite matrix using bounded Bunch-Kaufman (rook) diagonal pivoting method.
    2: *
    3: *  =========== DOCUMENTATION ===========
    4: *
    5: * Online html documentation available at
    6: *            http://www.netlib.org/lapack/explore-html/
    7: *
    8: *> \htmlonly
    9: *> Download ZLAHEF_RK + dependencies
   10: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/zlahef_rk.f">
   11: *> [TGZ]</a>
   12: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/zlahef_rk.f">
   13: *> [ZIP]</a>
   14: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/zlahef_rk.f">
   15: *> [TXT]</a>
   16: *> \endhtmlonly
   17: *
   18: *  Definition:
   19: *  ===========
   20: *
   21: *       SUBROUTINE ZLAHEF_RK( UPLO, N, NB, KB, A, LDA, E, IPIV, W, LDW,
   22: *                             INFO )
   23: *
   24: *       .. Scalar Arguments ..
   25: *       CHARACTER          UPLO
   26: *       INTEGER            INFO, KB, LDA, LDW, N, NB
   27: *       ..
   28: *       .. Array Arguments ..
   29: *       INTEGER            IPIV( * )
   30: *       COMPLEX*16         A( LDA, * ), E( * ), W( LDW, * )
   31: *       ..
   32: *
   33: *
   34: *> \par Purpose:
   35: *  =============
   36: *>
   37: *> \verbatim
   38: *> ZLAHEF_RK computes a partial factorization of a complex Hermitian
   39: *> matrix A using the bounded Bunch-Kaufman (rook) diagonal
   40: *> pivoting method. The partial factorization has the form:
   41: *>
   42: *> A  =  ( I  U12 ) ( A11  0  ) (  I       0    )  if UPLO = 'U', or:
   43: *>       ( 0  U22 ) (  0   D  ) ( U12**H U22**H )
   44: *>
   45: *> A  =  ( L11  0 ) (  D   0  ) ( L11**H L21**H )  if UPLO = 'L',
   46: *>       ( L21  I ) (  0  A22 ) (  0       I    )
   47: *>
   48: *> where the order of D is at most NB. The actual order is returned in
   49: *> the argument KB, and is either NB or NB-1, or N if N <= NB.
   50: *>
   51: *> ZLAHEF_RK is an auxiliary routine called by ZHETRF_RK. It uses
   52: *> blocked code (calling Level 3 BLAS) to update the submatrix
   53: *> A11 (if UPLO = 'U') or A22 (if UPLO = 'L').
   54: *> \endverbatim
   55: *
   56: *  Arguments:
   57: *  ==========
   58: *
   59: *> \param[in] UPLO
   60: *> \verbatim
   61: *>          UPLO is CHARACTER*1
   62: *>          Specifies whether the upper or lower triangular part of the
   63: *>          Hermitian matrix A is stored:
   64: *>          = 'U':  Upper triangular
   65: *>          = 'L':  Lower triangular
   66: *> \endverbatim
   67: *>
   68: *> \param[in] N
   69: *> \verbatim
   70: *>          N is INTEGER
   71: *>          The order of the matrix A.  N >= 0.
   72: *> \endverbatim
   73: *>
   74: *> \param[in] NB
   75: *> \verbatim
   76: *>          NB is INTEGER
   77: *>          The maximum number of columns of the matrix A that should be
   78: *>          factored.  NB should be at least 2 to allow for 2-by-2 pivot
   79: *>          blocks.
   80: *> \endverbatim
   81: *>
   82: *> \param[out] KB
   83: *> \verbatim
   84: *>          KB is INTEGER
   85: *>          The number of columns of A that were actually factored.
   86: *>          KB is either NB-1 or NB, or N if N <= NB.
   87: *> \endverbatim
   88: *>
   89: *> \param[in,out] A
   90: *> \verbatim
   91: *>          A is COMPLEX*16 array, dimension (LDA,N)
   92: *>          On entry, the Hermitian matrix A.
   93: *>            If UPLO = 'U': the leading N-by-N upper triangular part
   94: *>            of A contains the upper triangular part of the matrix A,
   95: *>            and the strictly lower triangular part of A is not
   96: *>            referenced.
   97: *>
   98: *>            If UPLO = 'L': the leading N-by-N lower triangular part
   99: *>            of A contains the lower triangular part of the matrix A,
  100: *>            and the strictly upper triangular part of A is not
  101: *>            referenced.
  102: *>
  103: *>          On exit, contains:
  104: *>            a) ONLY diagonal elements of the Hermitian block diagonal
  105: *>               matrix D on the diagonal of A, i.e. D(k,k) = A(k,k);
  106: *>               (superdiagonal (or subdiagonal) elements of D
  107: *>                are stored on exit in array E), and
  108: *>            b) If UPLO = 'U': factor U in the superdiagonal part of A.
  109: *>               If UPLO = 'L': factor L in the subdiagonal part of A.
  110: *> \endverbatim
  111: *>
  112: *> \param[in] LDA
  113: *> \verbatim
  114: *>          LDA is INTEGER
  115: *>          The leading dimension of the array A.  LDA >= max(1,N).
  116: *> \endverbatim
  117: *>
  118: *> \param[out] E
  119: *> \verbatim
  120: *>          E is COMPLEX*16 array, dimension (N)
  121: *>          On exit, contains the superdiagonal (or subdiagonal)
  122: *>          elements of the Hermitian block diagonal matrix D
  123: *>          with 1-by-1 or 2-by-2 diagonal blocks, where
  124: *>          If UPLO = 'U': E(i) = D(i-1,i), i=2:N, E(1) is set to 0;
  125: *>          If UPLO = 'L': E(i) = D(i+1,i), i=1:N-1, E(N) is set to 0.
  126: *>
  127: *>          NOTE: For 1-by-1 diagonal block D(k), where
  128: *>          1 <= k <= N, the element E(k) is set to 0 in both
  129: *>          UPLO = 'U' or UPLO = 'L' cases.
  130: *> \endverbatim
  131: *>
  132: *> \param[out] IPIV
  133: *> \verbatim
  134: *>          IPIV is INTEGER array, dimension (N)
  135: *>          IPIV describes the permutation matrix P in the factorization
  136: *>          of matrix A as follows. The absolute value of IPIV(k)
  137: *>          represents the index of row and column that were
  138: *>          interchanged with the k-th row and column. The value of UPLO
  139: *>          describes the order in which the interchanges were applied.
  140: *>          Also, the sign of IPIV represents the block structure of
  141: *>          the Hermitian block diagonal matrix D with 1-by-1 or 2-by-2
  142: *>          diagonal blocks which correspond to 1 or 2 interchanges
  143: *>          at each factorization step.
  144: *>
  145: *>          If UPLO = 'U',
  146: *>          ( in factorization order, k decreases from N to 1 ):
  147: *>            a) A single positive entry IPIV(k) > 0 means:
  148: *>               D(k,k) is a 1-by-1 diagonal block.
  149: *>               If IPIV(k) != k, rows and columns k and IPIV(k) were
  150: *>               interchanged in the submatrix A(1:N,N-KB+1:N);
  151: *>               If IPIV(k) = k, no interchange occurred.
  152: *>
  153: *>
  154: *>            b) A pair of consecutive negative entries
  155: *>               IPIV(k) < 0 and IPIV(k-1) < 0 means:
  156: *>               D(k-1:k,k-1:k) is a 2-by-2 diagonal block.
  157: *>               (NOTE: negative entries in IPIV appear ONLY in pairs).
  158: *>               1) If -IPIV(k) != k, rows and columns
  159: *>                  k and -IPIV(k) were interchanged
  160: *>                  in the matrix A(1:N,N-KB+1:N).
  161: *>                  If -IPIV(k) = k, no interchange occurred.
  162: *>               2) If -IPIV(k-1) != k-1, rows and columns
  163: *>                  k-1 and -IPIV(k-1) were interchanged
  164: *>                  in the submatrix A(1:N,N-KB+1:N).
  165: *>                  If -IPIV(k-1) = k-1, no interchange occurred.
  166: *>
  167: *>            c) In both cases a) and b) is always ABS( IPIV(k) ) <= k.
  168: *>
  169: *>            d) NOTE: Any entry IPIV(k) is always NONZERO on output.
  170: *>
  171: *>          If UPLO = 'L',
  172: *>          ( in factorization order, k increases from 1 to N ):
  173: *>            a) A single positive entry IPIV(k) > 0 means:
  174: *>               D(k,k) is a 1-by-1 diagonal block.
  175: *>               If IPIV(k) != k, rows and columns k and IPIV(k) were
  176: *>               interchanged in the submatrix A(1:N,1:KB).
  177: *>               If IPIV(k) = k, no interchange occurred.
  178: *>
  179: *>            b) A pair of consecutive negative entries
  180: *>               IPIV(k) < 0 and IPIV(k+1) < 0 means:
  181: *>               D(k:k+1,k:k+1) is a 2-by-2 diagonal block.
  182: *>               (NOTE: negative entries in IPIV appear ONLY in pairs).
  183: *>               1) If -IPIV(k) != k, rows and columns
  184: *>                  k and -IPIV(k) were interchanged
  185: *>                  in the submatrix A(1:N,1:KB).
  186: *>                  If -IPIV(k) = k, no interchange occurred.
  187: *>               2) If -IPIV(k+1) != k+1, rows and columns
  188: *>                  k-1 and -IPIV(k-1) were interchanged
  189: *>                  in the submatrix A(1:N,1:KB).
  190: *>                  If -IPIV(k+1) = k+1, no interchange occurred.
  191: *>
  192: *>            c) In both cases a) and b) is always ABS( IPIV(k) ) >= k.
  193: *>
  194: *>            d) NOTE: Any entry IPIV(k) is always NONZERO on output.
  195: *> \endverbatim
  196: *>
  197: *> \param[out] W
  198: *> \verbatim
  199: *>          W is COMPLEX*16 array, dimension (LDW,NB)
  200: *> \endverbatim
  201: *>
  202: *> \param[in] LDW
  203: *> \verbatim
  204: *>          LDW is INTEGER
  205: *>          The leading dimension of the array W.  LDW >= max(1,N).
  206: *> \endverbatim
  207: *>
  208: *> \param[out] INFO
  209: *> \verbatim
  210: *>          INFO is INTEGER
  211: *>          = 0: successful exit
  212: *>
  213: *>          < 0: If INFO = -k, the k-th argument had an illegal value
  214: *>
  215: *>          > 0: If INFO = k, the matrix A is singular, because:
  216: *>                 If UPLO = 'U': column k in the upper
  217: *>                 triangular part of A contains all zeros.
  218: *>                 If UPLO = 'L': column k in the lower
  219: *>                 triangular part of A contains all zeros.
  220: *>
  221: *>               Therefore D(k,k) is exactly zero, and superdiagonal
  222: *>               elements of column k of U (or subdiagonal elements of
  223: *>               column k of L ) are all zeros. The factorization has
  224: *>               been completed, but the block diagonal matrix D is
  225: *>               exactly singular, and division by zero will occur if
  226: *>               it is used to solve a system of equations.
  227: *>
  228: *>               NOTE: INFO only stores the first occurrence of
  229: *>               a singularity, any subsequent occurrence of singularity
  230: *>               is not stored in INFO even though the factorization
  231: *>               always completes.
  232: *> \endverbatim
  233: *
  234: *  Authors:
  235: *  ========
  236: *
  237: *> \author Univ. of Tennessee
  238: *> \author Univ. of California Berkeley
  239: *> \author Univ. of Colorado Denver
  240: *> \author NAG Ltd.
  241: *
  242: *> \date December 2016
  243: *
  244: *> \ingroup complex16HEcomputational
  245: *
  246: *> \par Contributors:
  247: *  ==================
  248: *>
  249: *> \verbatim
  250: *>
  251: *>  December 2016,  Igor Kozachenko,
  252: *>                  Computer Science Division,
  253: *>                  University of California, Berkeley
  254: *>
  255: *>  September 2007, Sven Hammarling, Nicholas J. Higham, Craig Lucas,
  256: *>                  School of Mathematics,
  257: *>                  University of Manchester
  258: *>
  259: *> \endverbatim
  260: *
  261: *  =====================================================================
  262:       SUBROUTINE ZLAHEF_RK( UPLO, N, NB, KB, A, LDA, E, IPIV, W, LDW,
  263:      $                      INFO )
  264: *
  265: *  -- LAPACK computational routine (version 3.7.0) --
  266: *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
  267: *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
  268: *     December 2016
  269: *
  270: *     .. Scalar Arguments ..
  271:       CHARACTER          UPLO
  272:       INTEGER            INFO, KB, LDA, LDW, N, NB
  273: *     ..
  274: *     .. Array Arguments ..
  275:       INTEGER            IPIV( * )
  276:       COMPLEX*16         A( LDA, * ), W( LDW, * ), E( * )
  277: *     ..
  278: *
  279: *  =====================================================================
  280: *
  281: *     .. Parameters ..
  282:       DOUBLE PRECISION   ZERO, ONE
  283:       PARAMETER          ( ZERO = 0.0D+0, ONE = 1.0D+0 )
  284:       COMPLEX*16         CONE
  285:       PARAMETER          ( CONE = ( 1.0D+0, 0.0D+0 ) )
  286:       DOUBLE PRECISION   EIGHT, SEVTEN
  287:       PARAMETER          ( EIGHT = 8.0D+0, SEVTEN = 17.0D+0 )
  288:       COMPLEX*16         CZERO
  289:       PARAMETER          ( CZERO = ( 0.0D+0, 0.0D+0 ) )
  290: *     ..
  291: *     .. Local Scalars ..
  292:       LOGICAL            DONE
  293:       INTEGER            IMAX, ITEMP, II, J, JB, JJ, JMAX, K, KK, KKW,
  294:      $                   KP, KSTEP, KW, P
  295:       DOUBLE PRECISION   ABSAKK, ALPHA, COLMAX, DTEMP, R1, ROWMAX, T,
  296:      $                   SFMIN
  297:       COMPLEX*16         D11, D21, D22, Z
  298: *     ..
  299: *     .. External Functions ..
  300:       LOGICAL            LSAME
  301:       INTEGER            IZAMAX
  302:       DOUBLE PRECISION   DLAMCH
  303:       EXTERNAL           LSAME, IZAMAX, DLAMCH
  304: *     ..
  305: *     .. External Subroutines ..
  306:       EXTERNAL           ZCOPY, ZDSCAL, ZGEMM, ZGEMV, ZLACGV, ZSWAP
  307: *     ..
  308: *     .. Intrinsic Functions ..
  309:       INTRINSIC          ABS, DBLE, DCONJG, DIMAG, MAX, MIN, SQRT
  310: *     ..
  311: *     .. Statement Functions ..
  312:       DOUBLE PRECISION   CABS1
  313: *     ..
  314: *     .. Statement Function definitions ..
  315:       CABS1( Z ) = ABS( DBLE( Z ) ) + ABS( DIMAG( Z ) )
  316: *     ..
  317: *     .. Executable Statements ..
  318: *
  319:       INFO = 0
  320: *
  321: *     Initialize ALPHA for use in choosing pivot block size.
  322: *
  323:       ALPHA = ( ONE+SQRT( SEVTEN ) ) / EIGHT
  324: *
  325: *     Compute machine safe minimum
  326: *
  327:       SFMIN = DLAMCH( 'S' )
  328: *
  329:       IF( LSAME( UPLO, 'U' ) ) THEN
  330: *
  331: *        Factorize the trailing columns of A using the upper triangle
  332: *        of A and working backwards, and compute the matrix W = U12*D
  333: *        for use in updating A11 (note that conjg(W) is actually stored)
  334: *        Initilize the first entry of array E, where superdiagonal
  335: *        elements of D are stored
  336: *
  337:          E( 1 ) = CZERO
  338: *
  339: *        K is the main loop index, decreasing from N in steps of 1 or 2
  340: *
  341:          K = N
  342:    10    CONTINUE
  343: *
  344: *        KW is the column of W which corresponds to column K of A
  345: *
  346:          KW = NB + K - N
  347: *
  348: *        Exit from loop
  349: *
  350:          IF( ( K.LE.N-NB+1 .AND. NB.LT.N ) .OR. K.LT.1 )
  351:      $      GO TO 30
  352: *
  353:          KSTEP = 1
  354:          P = K
  355: *
  356: *        Copy column K of A to column KW of W and update it
  357: *
  358:          IF( K.GT.1 )
  359:      $      CALL ZCOPY( K-1, A( 1, K ), 1, W( 1, KW ), 1 )
  360:          W( K, KW ) = DBLE( A( K, K ) )
  361:          IF( K.LT.N ) THEN
  362:             CALL ZGEMV( 'No transpose', K, N-K, -CONE, A( 1, K+1 ), LDA,
  363:      $                  W( K, KW+1 ), LDW, CONE, W( 1, KW ), 1 )
  364:             W( K, KW ) = DBLE( W( K, KW ) )
  365:          END IF
  366: *
  367: *        Determine rows and columns to be interchanged and whether
  368: *        a 1-by-1 or 2-by-2 pivot block will be used
  369: *
  370:          ABSAKK = ABS( DBLE( W( K, KW ) ) )
  371: *
  372: *        IMAX is the row-index of the largest off-diagonal element in
  373: *        column K, and COLMAX is its absolute value.
  374: *        Determine both COLMAX and IMAX.
  375: *
  376:          IF( K.GT.1 ) THEN
  377:             IMAX = IZAMAX( K-1, W( 1, KW ), 1 )
  378:             COLMAX = CABS1( W( IMAX, KW ) )
  379:          ELSE
  380:             COLMAX = ZERO
  381:          END IF
  382: *
  383:          IF( MAX( ABSAKK, COLMAX ).EQ.ZERO ) THEN
  384: *
  385: *           Column K is zero or underflow: set INFO and continue
  386: *
  387:             IF( INFO.EQ.0 )
  388:      $         INFO = K
  389:             KP = K
  390:             A( K, K ) = DBLE( W( K, KW ) )
  391:             IF( K.GT.1 )
  392:      $         CALL ZCOPY( K-1, W( 1, KW ), 1, A( 1, K ), 1 )
  393: *
  394: *           Set E( K ) to zero
  395: *
  396:             IF( K.GT.1 )
  397:      $         E( K ) = CZERO
  398: *
  399:          ELSE
  400: *
  401: *           ============================================================
  402: *
  403: *           BEGIN pivot search
  404: *
  405: *           Case(1)
  406: *           Equivalent to testing for ABSAKK.GE.ALPHA*COLMAX
  407: *           (used to handle NaN and Inf)
  408:             IF( .NOT.( ABSAKK.LT.ALPHA*COLMAX ) ) THEN
  409: *
  410: *              no interchange, use 1-by-1 pivot block
  411: *
  412:                KP = K
  413: *
  414:             ELSE
  415: *
  416: *              Lop until pivot found
  417: *
  418:                DONE = .FALSE.
  419: *
  420:    12          CONTINUE
  421: *
  422: *                 BEGIN pivot search loop body
  423: *
  424: *
  425: *                 Copy column IMAX to column KW-1 of W and update it
  426: *
  427:                   IF( IMAX.GT.1 )
  428:      $               CALL ZCOPY( IMAX-1, A( 1, IMAX ), 1, W( 1, KW-1 ),
  429:      $                           1 )
  430:                   W( IMAX, KW-1 ) = DBLE( A( IMAX, IMAX ) )
  431: *
  432:                   CALL ZCOPY( K-IMAX, A( IMAX, IMAX+1 ), LDA,
  433:      $                        W( IMAX+1, KW-1 ), 1 )
  434:                   CALL ZLACGV( K-IMAX, W( IMAX+1, KW-1 ), 1 )
  435: *
  436:                   IF( K.LT.N ) THEN
  437:                      CALL ZGEMV( 'No transpose', K, N-K, -CONE,
  438:      $                           A( 1, K+1 ), LDA, W( IMAX, KW+1 ), LDW,
  439:      $                           CONE, W( 1, KW-1 ), 1 )
  440:                      W( IMAX, KW-1 ) = DBLE( W( IMAX, KW-1 ) )
  441:                   END IF
  442: *
  443: *                 JMAX is the column-index of the largest off-diagonal
  444: *                 element in row IMAX, and ROWMAX is its absolute value.
  445: *                 Determine both ROWMAX and JMAX.
  446: *
  447:                   IF( IMAX.NE.K ) THEN
  448:                      JMAX = IMAX + IZAMAX( K-IMAX, W( IMAX+1, KW-1 ),
  449:      $                                     1 )
  450:                      ROWMAX = CABS1( W( JMAX, KW-1 ) )
  451:                   ELSE
  452:                      ROWMAX = ZERO
  453:                   END IF
  454: *
  455:                   IF( IMAX.GT.1 ) THEN
  456:                      ITEMP = IZAMAX( IMAX-1, W( 1, KW-1 ), 1 )
  457:                      DTEMP = CABS1( W( ITEMP, KW-1 ) )
  458:                      IF( DTEMP.GT.ROWMAX ) THEN
  459:                         ROWMAX = DTEMP
  460:                         JMAX = ITEMP
  461:                      END IF
  462:                   END IF
  463: *
  464: *                 Case(2)
  465: *                 Equivalent to testing for
  466: *                 ABS( REAL( W( IMAX,KW-1 ) ) ).GE.ALPHA*ROWMAX
  467: *                 (used to handle NaN and Inf)
  468: *
  469:                   IF( .NOT.( ABS( DBLE( W( IMAX,KW-1 ) ) )
  470:      $                       .LT.ALPHA*ROWMAX ) ) THEN
  471: *
  472: *                    interchange rows and columns K and IMAX,
  473: *                    use 1-by-1 pivot block
  474: *
  475:                      KP = IMAX
  476: *
  477: *                    copy column KW-1 of W to column KW of W
  478: *
  479:                      CALL ZCOPY( K, W( 1, KW-1 ), 1, W( 1, KW ), 1 )
  480: *
  481:                      DONE = .TRUE.
  482: *
  483: *                 Case(3)
  484: *                 Equivalent to testing for ROWMAX.EQ.COLMAX,
  485: *                 (used to handle NaN and Inf)
  486: *
  487:                   ELSE IF( ( P.EQ.JMAX ) .OR. ( ROWMAX.LE.COLMAX ) )
  488:      $            THEN
  489: *
  490: *                    interchange rows and columns K-1 and IMAX,
  491: *                    use 2-by-2 pivot block
  492: *
  493:                      KP = IMAX
  494:                      KSTEP = 2
  495:                      DONE = .TRUE.
  496: *
  497: *                 Case(4)
  498:                   ELSE
  499: *
  500: *                    Pivot not found: set params and repeat
  501: *
  502:                      P = IMAX
  503:                      COLMAX = ROWMAX
  504:                      IMAX = JMAX
  505: *
  506: *                    Copy updated JMAXth (next IMAXth) column to Kth of W
  507: *
  508:                      CALL ZCOPY( K, W( 1, KW-1 ), 1, W( 1, KW ), 1 )
  509: *
  510:                   END IF
  511: *
  512: *
  513: *                 END pivot search loop body
  514: *
  515:                IF( .NOT.DONE ) GOTO 12
  516: *
  517:             END IF
  518: *
  519: *           END pivot search
  520: *
  521: *           ============================================================
  522: *
  523: *           KK is the column of A where pivoting step stopped
  524: *
  525:             KK = K - KSTEP + 1
  526: *
  527: *           KKW is the column of W which corresponds to column KK of A
  528: *
  529:             KKW = NB + KK - N
  530: *
  531: *           Interchange rows and columns P and K.
  532: *           Updated column P is already stored in column KW of W.
  533: *
  534:             IF( ( KSTEP.EQ.2 ) .AND. ( P.NE.K ) ) THEN
  535: *
  536: *              Copy non-updated column K to column P of submatrix A
  537: *              at step K. No need to copy element into columns
  538: *              K and K-1 of A for 2-by-2 pivot, since these columns
  539: *              will be later overwritten.
  540: *
  541:                A( P, P ) = DBLE( A( K, K ) )
  542:                CALL ZCOPY( K-1-P, A( P+1, K ), 1, A( P, P+1 ),
  543:      $                     LDA )
  544:                CALL ZLACGV( K-1-P, A( P, P+1 ), LDA )
  545:                IF( P.GT.1 )
  546:      $            CALL ZCOPY( P-1, A( 1, K ), 1, A( 1, P ), 1 )
  547: *
  548: *              Interchange rows K and P in the last K+1 to N columns of A
  549: *              (columns K and K-1 of A for 2-by-2 pivot will be
  550: *              later overwritten). Interchange rows K and P
  551: *              in last KKW to NB columns of W.
  552: *
  553:                IF( K.LT.N )
  554:      $            CALL ZSWAP( N-K, A( K, K+1 ), LDA, A( P, K+1 ),
  555:      $                        LDA )
  556:                CALL ZSWAP( N-KK+1, W( K, KKW ), LDW, W( P, KKW ),
  557:      $                     LDW )
  558:             END IF
  559: *
  560: *           Interchange rows and columns KP and KK.
  561: *           Updated column KP is already stored in column KKW of W.
  562: *
  563:             IF( KP.NE.KK ) THEN
  564: *
  565: *              Copy non-updated column KK to column KP of submatrix A
  566: *              at step K. No need to copy element into column K
  567: *              (or K and K-1 for 2-by-2 pivot) of A, since these columns
  568: *              will be later overwritten.
  569: *
  570:                A( KP, KP ) = DBLE( A( KK, KK ) )
  571:                CALL ZCOPY( KK-1-KP, A( KP+1, KK ), 1, A( KP, KP+1 ),
  572:      $                     LDA )
  573:                CALL ZLACGV( KK-1-KP, A( KP, KP+1 ), LDA )
  574:                IF( KP.GT.1 )
  575:      $            CALL ZCOPY( KP-1, A( 1, KK ), 1, A( 1, KP ), 1 )
  576: *
  577: *              Interchange rows KK and KP in last K+1 to N columns of A
  578: *              (columns K (or K and K-1 for 2-by-2 pivot) of A will be
  579: *              later overwritten). Interchange rows KK and KP
  580: *              in last KKW to NB columns of W.
  581: *
  582:                IF( K.LT.N )
  583:      $            CALL ZSWAP( N-K, A( KK, K+1 ), LDA, A( KP, K+1 ),
  584:      $                        LDA )
  585:                CALL ZSWAP( N-KK+1, W( KK, KKW ), LDW, W( KP, KKW ),
  586:      $                     LDW )
  587:             END IF
  588: *
  589:             IF( KSTEP.EQ.1 ) THEN
  590: *
  591: *              1-by-1 pivot block D(k): column kw of W now holds
  592: *
  593: *              W(kw) = U(k)*D(k),
  594: *
  595: *              where U(k) is the k-th column of U
  596: *
  597: *              (1) Store subdiag. elements of column U(k)
  598: *              and 1-by-1 block D(k) in column k of A.
  599: *              (NOTE: Diagonal element U(k,k) is a UNIT element
  600: *              and not stored)
  601: *                 A(k,k) := D(k,k) = W(k,kw)
  602: *                 A(1:k-1,k) := U(1:k-1,k) = W(1:k-1,kw)/D(k,k)
  603: *
  604: *              (NOTE: No need to use for Hermitian matrix
  605: *              A( K, K ) = REAL( W( K, K) ) to separately copy diagonal
  606: *              element D(k,k) from W (potentially saves only one load))
  607:                CALL ZCOPY( K, W( 1, KW ), 1, A( 1, K ), 1 )
  608:                IF( K.GT.1 ) THEN
  609: *
  610: *                 (NOTE: No need to check if A(k,k) is NOT ZERO,
  611: *                  since that was ensured earlier in pivot search:
  612: *                  case A(k,k) = 0 falls into 2x2 pivot case(3))
  613: *
  614: *                 Handle division by a small number
  615: *
  616:                   T = DBLE( A( K, K ) )
  617:                   IF( ABS( T ).GE.SFMIN ) THEN
  618:                      R1 = ONE / T
  619:                      CALL ZDSCAL( K-1, R1, A( 1, K ), 1 )
  620:                   ELSE
  621:                      DO 14 II = 1, K-1
  622:                         A( II, K ) = A( II, K ) / T
  623:    14                CONTINUE
  624:                   END IF
  625: *
  626: *                 (2) Conjugate column W(kw)
  627: *
  628:                   CALL ZLACGV( K-1, W( 1, KW ), 1 )
  629: *
  630: *                 Store the superdiagonal element of D in array E
  631: *
  632:                   E( K ) = CZERO
  633: *
  634:                END IF
  635: *
  636:             ELSE
  637: *
  638: *              2-by-2 pivot block D(k): columns kw and kw-1 of W now hold
  639: *
  640: *              ( W(kw-1) W(kw) ) = ( U(k-1) U(k) )*D(k)
  641: *
  642: *              where U(k) and U(k-1) are the k-th and (k-1)-th columns
  643: *              of U
  644: *
  645: *              (1) Store U(1:k-2,k-1) and U(1:k-2,k) and 2-by-2
  646: *              block D(k-1:k,k-1:k) in columns k-1 and k of A.
  647: *              (NOTE: 2-by-2 diagonal block U(k-1:k,k-1:k) is a UNIT
  648: *              block and not stored)
  649: *                 A(k-1:k,k-1:k) := D(k-1:k,k-1:k) = W(k-1:k,kw-1:kw)
  650: *                 A(1:k-2,k-1:k) := U(1:k-2,k:k-1:k) =
  651: *                 = W(1:k-2,kw-1:kw) * ( D(k-1:k,k-1:k)**(-1) )
  652: *
  653:                IF( K.GT.2 ) THEN
  654: *
  655: *                 Factor out the columns of the inverse of 2-by-2 pivot
  656: *                 block D, so that each column contains 1, to reduce the
  657: *                 number of FLOPS when we multiply panel
  658: *                 ( W(kw-1) W(kw) ) by this inverse, i.e. by D**(-1).
  659: *
  660: *                 D**(-1) = ( d11 cj(d21) )**(-1) =
  661: *                           ( d21    d22 )
  662: *
  663: *                 = 1/(d11*d22-|d21|**2) * ( ( d22) (-cj(d21) ) ) =
  664: *                                          ( (-d21) (     d11 ) )
  665: *
  666: *                 = 1/(|d21|**2) * 1/((d11/cj(d21))*(d22/d21)-1) *
  667: *
  668: *                   * ( d21*( d22/d21 ) conj(d21)*(           - 1 ) ) =
  669: *                     (     (      -1 )           ( d11/conj(d21) ) )
  670: *
  671: *                 = 1/(|d21|**2) * 1/(D22*D11-1) *
  672: *
  673: *                   * ( d21*( D11 ) conj(d21)*(  -1 ) ) =
  674: *                     (     (  -1 )           ( D22 ) )
  675: *
  676: *                 = (1/|d21|**2) * T * ( d21*( D11 ) conj(d21)*(  -1 ) ) =
  677: *                                      (     (  -1 )           ( D22 ) )
  678: *
  679: *                 = ( (T/conj(d21))*( D11 ) (T/d21)*(  -1 ) ) =
  680: *                   (               (  -1 )         ( D22 ) )
  681: *
  682: *                 Handle division by a small number. (NOTE: order of
  683: *                 operations is important)
  684: *
  685: *                 = ( T*(( D11 )/conj(D21)) T*((  -1 )/D21 ) )
  686: *                   (   ((  -1 )          )   (( D22 )     ) ),
  687: *
  688: *                 where D11 = d22/d21,
  689: *                       D22 = d11/conj(d21),
  690: *                       D21 = d21,
  691: *                       T = 1/(D22*D11-1).
  692: *
  693: *                 (NOTE: No need to check for division by ZERO,
  694: *                  since that was ensured earlier in pivot search:
  695: *                  (a) d21 != 0 in 2x2 pivot case(4),
  696: *                      since |d21| should be larger than |d11| and |d22|;
  697: *                  (b) (D22*D11 - 1) != 0, since from (a),
  698: *                      both |D11| < 1, |D22| < 1, hence |D22*D11| << 1.)
  699: *
  700:                   D21 = W( K-1, KW )
  701:                   D11 = W( K, KW ) / DCONJG( D21 )
  702:                   D22 = W( K-1, KW-1 ) / D21
  703:                   T = ONE / ( DBLE( D11*D22 )-ONE )
  704: *
  705: *                 Update elements in columns A(k-1) and A(k) as
  706: *                 dot products of rows of ( W(kw-1) W(kw) ) and columns
  707: *                 of D**(-1)
  708: *
  709:                   DO 20 J = 1, K - 2
  710:                      A( J, K-1 ) = T*( ( D11*W( J, KW-1 )-W( J, KW ) ) /
  711:      $                             D21 )
  712:                      A( J, K ) = T*( ( D22*W( J, KW )-W( J, KW-1 ) ) /
  713:      $                           DCONJG( D21 ) )
  714:    20             CONTINUE
  715:                END IF
  716: *
  717: *              Copy diagonal elements of D(K) to A,
  718: *              copy superdiagonal element of D(K) to E(K) and
  719: *              ZERO out superdiagonal entry of A
  720: *
  721:                A( K-1, K-1 ) = W( K-1, KW-1 )
  722:                A( K-1, K ) = CZERO
  723:                A( K, K ) = W( K, KW )
  724:                E( K ) = W( K-1, KW )
  725:                E( K-1 ) = CZERO
  726: *
  727: *              (2) Conjugate columns W(kw) and W(kw-1)
  728: *
  729:                CALL ZLACGV( K-1, W( 1, KW ), 1 )
  730:                CALL ZLACGV( K-2, W( 1, KW-1 ), 1 )
  731: *
  732:             END IF
  733: *
  734: *           End column K is nonsingular
  735: *
  736:          END IF
  737: *
  738: *        Store details of the interchanges in IPIV
  739: *
  740:          IF( KSTEP.EQ.1 ) THEN
  741:             IPIV( K ) = KP
  742:          ELSE
  743:             IPIV( K ) = -P
  744:             IPIV( K-1 ) = -KP
  745:          END IF
  746: *
  747: *        Decrease K and return to the start of the main loop
  748: *
  749:          K = K - KSTEP
  750:          GO TO 10
  751: *
  752:    30    CONTINUE
  753: *
  754: *        Update the upper triangle of A11 (= A(1:k,1:k)) as
  755: *
  756: *        A11 := A11 - U12*D*U12**H = A11 - U12*W**H
  757: *
  758: *        computing blocks of NB columns at a time (note that conjg(W) is
  759: *        actually stored)
  760: *
  761:          DO 50 J = ( ( K-1 ) / NB )*NB + 1, 1, -NB
  762:             JB = MIN( NB, K-J+1 )
  763: *
  764: *           Update the upper triangle of the diagonal block
  765: *
  766:             DO 40 JJ = J, J + JB - 1
  767:                A( JJ, JJ ) = DBLE( A( JJ, JJ ) )
  768:                CALL ZGEMV( 'No transpose', JJ-J+1, N-K, -CONE,
  769:      $                     A( J, K+1 ), LDA, W( JJ, KW+1 ), LDW, CONE,
  770:      $                     A( J, JJ ), 1 )
  771:                A( JJ, JJ ) = DBLE( A( JJ, JJ ) )
  772:    40       CONTINUE
  773: *
  774: *           Update the rectangular superdiagonal block
  775: *
  776:             IF( J.GE.2 )
  777:      $         CALL ZGEMM( 'No transpose', 'Transpose', J-1, JB, N-K,
  778:      $                     -CONE, A( 1, K+1 ), LDA, W( J, KW+1 ), LDW,
  779:      $                     CONE, A( 1, J ), LDA )
  780:    50    CONTINUE
  781: *
  782: *        Set KB to the number of columns factorized
  783: *
  784:          KB = N - K
  785: *
  786:       ELSE
  787: *
  788: *        Factorize the leading columns of A using the lower triangle
  789: *        of A and working forwards, and compute the matrix W = L21*D
  790: *        for use in updating A22 (note that conjg(W) is actually stored)
  791: *
  792: *        Initilize the unused last entry of the subdiagonal array E.
  793: *
  794:          E( N ) = CZERO
  795: *
  796: *        K is the main loop index, increasing from 1 in steps of 1 or 2
  797: *
  798:          K = 1
  799:    70    CONTINUE
  800: *
  801: *        Exit from loop
  802: *
  803:          IF( ( K.GE.NB .AND. NB.LT.N ) .OR. K.GT.N )
  804:      $      GO TO 90
  805: *
  806:          KSTEP = 1
  807:          P = K
  808: *
  809: *        Copy column K of A to column K of W and update column K of W
  810: *
  811:          W( K, K ) = DBLE( A( K, K ) )
  812:          IF( K.LT.N )
  813:      $      CALL ZCOPY( N-K, A( K+1, K ), 1, W( K+1, K ), 1 )
  814:          IF( K.GT.1 ) THEN
  815:             CALL ZGEMV( 'No transpose', N-K+1, K-1, -CONE, A( K, 1 ),
  816:      $                  LDA, W( K, 1 ), LDW, CONE, W( K, K ), 1 )
  817:             W( K, K ) = DBLE( W( K, K ) )
  818:          END IF
  819: *
  820: *        Determine rows and columns to be interchanged and whether
  821: *        a 1-by-1 or 2-by-2 pivot block will be used
  822: *
  823:          ABSAKK = ABS( DBLE( W( K, K ) ) )
  824: *
  825: *        IMAX is the row-index of the largest off-diagonal element in
  826: *        column K, and COLMAX is its absolute value.
  827: *        Determine both COLMAX and IMAX.
  828: *
  829:          IF( K.LT.N ) THEN
  830:             IMAX = K + IZAMAX( N-K, W( K+1, K ), 1 )
  831:             COLMAX = CABS1( W( IMAX, K ) )
  832:          ELSE
  833:             COLMAX = ZERO
  834:          END IF
  835: *
  836:          IF( MAX( ABSAKK, COLMAX ).EQ.ZERO ) THEN
  837: *
  838: *           Column K is zero or underflow: set INFO and continue
  839: *
  840:             IF( INFO.EQ.0 )
  841:      $         INFO = K
  842:             KP = K
  843:             A( K, K ) = DBLE( W( K, K ) )
  844:             IF( K.LT.N )
  845:      $         CALL ZCOPY( N-K, W( K+1, K ), 1, A( K+1, K ), 1 )
  846: *
  847: *           Set E( K ) to zero
  848: *
  849:             IF( K.LT.N )
  850:      $         E( K ) = CZERO
  851: *
  852:          ELSE
  853: *
  854: *           ============================================================
  855: *
  856: *           BEGIN pivot search
  857: *
  858: *           Case(1)
  859: *           Equivalent to testing for ABSAKK.GE.ALPHA*COLMAX
  860: *           (used to handle NaN and Inf)
  861: *
  862:             IF( .NOT.( ABSAKK.LT.ALPHA*COLMAX ) ) THEN
  863: *
  864: *              no interchange, use 1-by-1 pivot block
  865: *
  866:                KP = K
  867: *
  868:             ELSE
  869: *
  870:                DONE = .FALSE.
  871: *
  872: *              Loop until pivot found
  873: *
  874:    72          CONTINUE
  875: *
  876: *                 BEGIN pivot search loop body
  877: *
  878: *
  879: *                 Copy column IMAX to column k+1 of W and update it
  880: *
  881:                   CALL ZCOPY( IMAX-K, A( IMAX, K ), LDA, W( K, K+1 ), 1)
  882:                   CALL ZLACGV( IMAX-K, W( K, K+1 ), 1 )
  883:                   W( IMAX, K+1 ) = DBLE( A( IMAX, IMAX ) )
  884: *
  885:                   IF( IMAX.LT.N )
  886:      $               CALL ZCOPY( N-IMAX, A( IMAX+1, IMAX ), 1,
  887:      $                           W( IMAX+1, K+1 ), 1 )
  888: *
  889:                   IF( K.GT.1 ) THEN
  890:                      CALL ZGEMV( 'No transpose', N-K+1, K-1, -CONE,
  891:      $                            A( K, 1 ), LDA, W( IMAX, 1 ), LDW,
  892:      $                            CONE, W( K, K+1 ), 1 )
  893:                      W( IMAX, K+1 ) = DBLE( W( IMAX, K+1 ) )
  894:                   END IF
  895: *
  896: *                 JMAX is the column-index of the largest off-diagonal
  897: *                 element in row IMAX, and ROWMAX is its absolute value.
  898: *                 Determine both ROWMAX and JMAX.
  899: *
  900:                   IF( IMAX.NE.K ) THEN
  901:                      JMAX = K - 1 + IZAMAX( IMAX-K, W( K, K+1 ), 1 )
  902:                      ROWMAX = CABS1( W( JMAX, K+1 ) )
  903:                   ELSE
  904:                      ROWMAX = ZERO
  905:                   END IF
  906: *
  907:                   IF( IMAX.LT.N ) THEN
  908:                      ITEMP = IMAX + IZAMAX( N-IMAX, W( IMAX+1, K+1 ), 1)
  909:                      DTEMP = CABS1( W( ITEMP, K+1 ) )
  910:                      IF( DTEMP.GT.ROWMAX ) THEN
  911:                         ROWMAX = DTEMP
  912:                         JMAX = ITEMP
  913:                      END IF
  914:                   END IF
  915: *
  916: *                 Case(2)
  917: *                 Equivalent to testing for
  918: *                 ABS( REAL( W( IMAX,K+1 ) ) ).GE.ALPHA*ROWMAX
  919: *                 (used to handle NaN and Inf)
  920: *
  921:                   IF( .NOT.( ABS( DBLE( W( IMAX,K+1 ) ) )
  922:      $                       .LT.ALPHA*ROWMAX ) ) THEN
  923: *
  924: *                    interchange rows and columns K and IMAX,
  925: *                    use 1-by-1 pivot block
  926: *
  927:                      KP = IMAX
  928: *
  929: *                    copy column K+1 of W to column K of W
  930: *
  931:                      CALL ZCOPY( N-K+1, W( K, K+1 ), 1, W( K, K ), 1 )
  932: *
  933:                      DONE = .TRUE.
  934: *
  935: *                 Case(3)
  936: *                 Equivalent to testing for ROWMAX.EQ.COLMAX,
  937: *                 (used to handle NaN and Inf)
  938: *
  939:                   ELSE IF( ( P.EQ.JMAX ) .OR. ( ROWMAX.LE.COLMAX ) )
  940:      $            THEN
  941: *
  942: *                    interchange rows and columns K+1 and IMAX,
  943: *                    use 2-by-2 pivot block
  944: *
  945:                      KP = IMAX
  946:                      KSTEP = 2
  947:                      DONE = .TRUE.
  948: *
  949: *                 Case(4)
  950:                   ELSE
  951: *
  952: *                    Pivot not found: set params and repeat
  953: *
  954:                      P = IMAX
  955:                      COLMAX = ROWMAX
  956:                      IMAX = JMAX
  957: *
  958: *                    Copy updated JMAXth (next IMAXth) column to Kth of W
  959: *
  960:                      CALL ZCOPY( N-K+1, W( K, K+1 ), 1, W( K, K ), 1 )
  961: *
  962:                   END IF
  963: *
  964: *
  965: *                 End pivot search loop body
  966: *
  967:                IF( .NOT.DONE ) GOTO 72
  968: *
  969:             END IF
  970: *
  971: *           END pivot search
  972: *
  973: *           ============================================================
  974: *
  975: *           KK is the column of A where pivoting step stopped
  976: *
  977:             KK = K + KSTEP - 1
  978: *
  979: *           Interchange rows and columns P and K (only for 2-by-2 pivot).
  980: *           Updated column P is already stored in column K of W.
  981: *
  982:             IF( ( KSTEP.EQ.2 ) .AND. ( P.NE.K ) ) THEN
  983: *
  984: *              Copy non-updated column KK-1 to column P of submatrix A
  985: *              at step K. No need to copy element into columns
  986: *              K and K+1 of A for 2-by-2 pivot, since these columns
  987: *              will be later overwritten.
  988: *
  989:                A( P, P ) = DBLE( A( K, K ) )
  990:                CALL ZCOPY( P-K-1, A( K+1, K ), 1, A( P, K+1 ), LDA )
  991:                CALL ZLACGV( P-K-1, A( P, K+1 ), LDA )
  992:                IF( P.LT.N )
  993:      $            CALL ZCOPY( N-P, A( P+1, K ), 1, A( P+1, P ), 1 )
  994: *
  995: *              Interchange rows K and P in first K-1 columns of A
  996: *              (columns K and K+1 of A for 2-by-2 pivot will be
  997: *              later overwritten). Interchange rows K and P
  998: *              in first KK columns of W.
  999: *
 1000:                IF( K.GT.1 )
 1001:      $            CALL ZSWAP( K-1, A( K, 1 ), LDA, A( P, 1 ), LDA )
 1002:                CALL ZSWAP( KK, W( K, 1 ), LDW, W( P, 1 ), LDW )
 1003:             END IF
 1004: *
 1005: *           Interchange rows and columns KP and KK.
 1006: *           Updated column KP is already stored in column KK of W.
 1007: *
 1008:             IF( KP.NE.KK ) THEN
 1009: *
 1010: *              Copy non-updated column KK to column KP of submatrix A
 1011: *              at step K. No need to copy element into column K
 1012: *              (or K and K+1 for 2-by-2 pivot) of A, since these columns
 1013: *              will be later overwritten.
 1014: *
 1015:                A( KP, KP ) = DBLE( A( KK, KK ) )
 1016:                CALL ZCOPY( KP-KK-1, A( KK+1, KK ), 1, A( KP, KK+1 ),
 1017:      $                     LDA )
 1018:                CALL ZLACGV( KP-KK-1, A( KP, KK+1 ), LDA )
 1019:                IF( KP.LT.N )
 1020:      $            CALL ZCOPY( N-KP, A( KP+1, KK ), 1, A( KP+1, KP ), 1 )
 1021: *
 1022: *              Interchange rows KK and KP in first K-1 columns of A
 1023: *              (column K (or K and K+1 for 2-by-2 pivot) of A will be
 1024: *              later overwritten). Interchange rows KK and KP
 1025: *              in first KK columns of W.
 1026: *
 1027:                IF( K.GT.1 )
 1028:      $            CALL ZSWAP( K-1, A( KK, 1 ), LDA, A( KP, 1 ), LDA )
 1029:                CALL ZSWAP( KK, W( KK, 1 ), LDW, W( KP, 1 ), LDW )
 1030:             END IF
 1031: *
 1032:             IF( KSTEP.EQ.1 ) THEN
 1033: *
 1034: *              1-by-1 pivot block D(k): column k of W now holds
 1035: *
 1036: *              W(k) = L(k)*D(k),
 1037: *
 1038: *              where L(k) is the k-th column of L
 1039: *
 1040: *              (1) Store subdiag. elements of column L(k)
 1041: *              and 1-by-1 block D(k) in column k of A.
 1042: *              (NOTE: Diagonal element L(k,k) is a UNIT element
 1043: *              and not stored)
 1044: *                 A(k,k) := D(k,k) = W(k,k)
 1045: *                 A(k+1:N,k) := L(k+1:N,k) = W(k+1:N,k)/D(k,k)
 1046: *
 1047: *              (NOTE: No need to use for Hermitian matrix
 1048: *              A( K, K ) = REAL( W( K, K) ) to separately copy diagonal
 1049: *              element D(k,k) from W (potentially saves only one load))
 1050:                CALL ZCOPY( N-K+1, W( K, K ), 1, A( K, K ), 1 )
 1051:                IF( K.LT.N ) THEN
 1052: *
 1053: *                 (NOTE: No need to check if A(k,k) is NOT ZERO,
 1054: *                  since that was ensured earlier in pivot search:
 1055: *                  case A(k,k) = 0 falls into 2x2 pivot case(3))
 1056: *
 1057: *                 Handle division by a small number
 1058: *
 1059:                   T = DBLE( A( K, K ) )
 1060:                   IF( ABS( T ).GE.SFMIN ) THEN
 1061:                      R1 = ONE / T
 1062:                      CALL ZDSCAL( N-K, R1, A( K+1, K ), 1 )
 1063:                   ELSE
 1064:                      DO 74 II = K + 1, N
 1065:                         A( II, K ) = A( II, K ) / T
 1066:    74                CONTINUE
 1067:                   END IF
 1068: *
 1069: *                 (2) Conjugate column W(k)
 1070: *
 1071:                   CALL ZLACGV( N-K, W( K+1, K ), 1 )
 1072: *
 1073: *                 Store the subdiagonal element of D in array E
 1074: *
 1075:                   E( K ) = CZERO
 1076: *
 1077:                END IF
 1078: *
 1079:             ELSE
 1080: *
 1081: *              2-by-2 pivot block D(k): columns k and k+1 of W now hold
 1082: *
 1083: *              ( W(k) W(k+1) ) = ( L(k) L(k+1) )*D(k)
 1084: *
 1085: *              where L(k) and L(k+1) are the k-th and (k+1)-th columns
 1086: *              of L
 1087: *
 1088: *              (1) Store L(k+2:N,k) and L(k+2:N,k+1) and 2-by-2
 1089: *              block D(k:k+1,k:k+1) in columns k and k+1 of A.
 1090: *              NOTE: 2-by-2 diagonal block L(k:k+1,k:k+1) is a UNIT
 1091: *              block and not stored.
 1092: *                 A(k:k+1,k:k+1) := D(k:k+1,k:k+1) = W(k:k+1,k:k+1)
 1093: *                 A(k+2:N,k:k+1) := L(k+2:N,k:k+1) =
 1094: *                 = W(k+2:N,k:k+1) * ( D(k:k+1,k:k+1)**(-1) )
 1095: *
 1096:                IF( K.LT.N-1 ) THEN
 1097: *
 1098: *                 Factor out the columns of the inverse of 2-by-2 pivot
 1099: *                 block D, so that each column contains 1, to reduce the
 1100: *                 number of FLOPS when we multiply panel
 1101: *                 ( W(kw-1) W(kw) ) by this inverse, i.e. by D**(-1).
 1102: *
 1103: *                 D**(-1) = ( d11 cj(d21) )**(-1) =
 1104: *                           ( d21    d22 )
 1105: *
 1106: *                 = 1/(d11*d22-|d21|**2) * ( ( d22) (-cj(d21) ) ) =
 1107: *                                          ( (-d21) (     d11 ) )
 1108: *
 1109: *                 = 1/(|d21|**2) * 1/((d11/cj(d21))*(d22/d21)-1) *
 1110: *
 1111: *                   * ( d21*( d22/d21 ) conj(d21)*(           - 1 ) ) =
 1112: *                     (     (      -1 )           ( d11/conj(d21) ) )
 1113: *
 1114: *                 = 1/(|d21|**2) * 1/(D22*D11-1) *
 1115: *
 1116: *                   * ( d21*( D11 ) conj(d21)*(  -1 ) ) =
 1117: *                     (     (  -1 )           ( D22 ) )
 1118: *
 1119: *                 = (1/|d21|**2) * T * ( d21*( D11 ) conj(d21)*(  -1 ) ) =
 1120: *                                      (     (  -1 )           ( D22 ) )
 1121: *
 1122: *                 = ( (T/conj(d21))*( D11 ) (T/d21)*(  -1 ) ) =
 1123: *                   (               (  -1 )         ( D22 ) )
 1124: *
 1125: *                 Handle division by a small number. (NOTE: order of
 1126: *                 operations is important)
 1127: *
 1128: *                 = ( T*(( D11 )/conj(D21)) T*((  -1 )/D21 ) )
 1129: *                   (   ((  -1 )          )   (( D22 )     ) ),
 1130: *
 1131: *                 where D11 = d22/d21,
 1132: *                       D22 = d11/conj(d21),
 1133: *                       D21 = d21,
 1134: *                       T = 1/(D22*D11-1).
 1135: *
 1136: *                 (NOTE: No need to check for division by ZERO,
 1137: *                  since that was ensured earlier in pivot search:
 1138: *                  (a) d21 != 0 in 2x2 pivot case(4),
 1139: *                      since |d21| should be larger than |d11| and |d22|;
 1140: *                  (b) (D22*D11 - 1) != 0, since from (a),
 1141: *                      both |D11| < 1, |D22| < 1, hence |D22*D11| << 1.)
 1142: *
 1143:                   D21 = W( K+1, K )
 1144:                   D11 = W( K+1, K+1 ) / D21
 1145:                   D22 = W( K, K ) / DCONJG( D21 )
 1146:                   T = ONE / ( DBLE( D11*D22 )-ONE )
 1147: *
 1148: *                 Update elements in columns A(k) and A(k+1) as
 1149: *                 dot products of rows of ( W(k) W(k+1) ) and columns
 1150: *                 of D**(-1)
 1151: *
 1152:                   DO 80 J = K + 2, N
 1153:                      A( J, K ) = T*( ( D11*W( J, K )-W( J, K+1 ) ) /
 1154:      $                           DCONJG( D21 ) )
 1155:                      A( J, K+1 ) = T*( ( D22*W( J, K+1 )-W( J, K ) ) /
 1156:      $                             D21 )
 1157:    80             CONTINUE
 1158:                END IF
 1159: *
 1160: *              Copy diagonal elements of D(K) to A,
 1161: *              copy subdiagonal element of D(K) to E(K) and
 1162: *              ZERO out subdiagonal entry of A
 1163: *
 1164:                A( K, K ) = W( K, K )
 1165:                A( K+1, K ) = CZERO
 1166:                A( K+1, K+1 ) = W( K+1, K+1 )
 1167:                E( K ) = W( K+1, K )
 1168:                E( K+1 ) = CZERO
 1169: *
 1170: *              (2) Conjugate columns W(k) and W(k+1)
 1171: *
 1172:                CALL ZLACGV( N-K, W( K+1, K ), 1 )
 1173:                CALL ZLACGV( N-K-1, W( K+2, K+1 ), 1 )
 1174: *
 1175:             END IF
 1176: *
 1177: *           End column K is nonsingular
 1178: *
 1179:          END IF
 1180: *
 1181: *        Store details of the interchanges in IPIV
 1182: *
 1183:          IF( KSTEP.EQ.1 ) THEN
 1184:             IPIV( K ) = KP
 1185:          ELSE
 1186:             IPIV( K ) = -P
 1187:             IPIV( K+1 ) = -KP
 1188:          END IF
 1189: *
 1190: *        Increase K and return to the start of the main loop
 1191: *
 1192:          K = K + KSTEP
 1193:          GO TO 70
 1194: *
 1195:    90    CONTINUE
 1196: *
 1197: *        Update the lower triangle of A22 (= A(k:n,k:n)) as
 1198: *
 1199: *        A22 := A22 - L21*D*L21**H = A22 - L21*W**H
 1200: *
 1201: *        computing blocks of NB columns at a time (note that conjg(W) is
 1202: *        actually stored)
 1203: *
 1204:          DO 110 J = K, N, NB
 1205:             JB = MIN( NB, N-J+1 )
 1206: *
 1207: *           Update the lower triangle of the diagonal block
 1208: *
 1209:             DO 100 JJ = J, J + JB - 1
 1210:                A( JJ, JJ ) = DBLE( A( JJ, JJ ) )
 1211:                CALL ZGEMV( 'No transpose', J+JB-JJ, K-1, -CONE,
 1212:      $                     A( JJ, 1 ), LDA, W( JJ, 1 ), LDW, CONE,
 1213:      $                     A( JJ, JJ ), 1 )
 1214:                A( JJ, JJ ) = DBLE( A( JJ, JJ ) )
 1215:   100       CONTINUE
 1216: *
 1217: *           Update the rectangular subdiagonal block
 1218: *
 1219:             IF( J+JB.LE.N )
 1220:      $         CALL ZGEMM( 'No transpose', 'Transpose', N-J-JB+1, JB,
 1221:      $                     K-1, -CONE, A( J+JB, 1 ), LDA, W( J, 1 ),
 1222:      $                     LDW, CONE, A( J+JB, J ), LDA )
 1223:   110    CONTINUE
 1224: *
 1225: *        Set KB to the number of columns factorized
 1226: *
 1227:          KB = K - 1
 1228: *
 1229:       END IF
 1230:       RETURN
 1231: *
 1232: *     End of ZLAHEF_RK
 1233: *
 1234:       END

CVSweb interface <joel.bertrand@systella.fr>