dlaqr5.f

Go to the documentation of this file.

*> \brief \b DLAQR5 performs a single small-bulge multi-shift QR sweep.
*
*  =========== DOCUMENTATION ===========
*
* Online html documentation available at 
*            http://www.netlib.org/lapack/explore-html/ 
*
*> \htmlonly
*> Download DLAQR5 + dependencies 
*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dlaqr5.f"> 
*> [TGZ]</a> 
*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dlaqr5.f"> 
*> [ZIP]</a> 
*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dlaqr5.f"> 
*> [TXT]</a>
*> \endhtmlonly 
*
*  Definition:
*  ===========
*
*       SUBROUTINE DLAQR5( WANTT, WANTZ, KACC22, N, KTOP, KBOT, NSHFTS,
*                          SR, SI, H, LDH, ILOZ, IHIZ, Z, LDZ, V, LDV, U,
*                          LDU, NV, WV, LDWV, NH, WH, LDWH )
* 
*       .. Scalar Arguments ..
*       INTEGER            IHIZ, ILOZ, KACC22, KBOT, KTOP, LDH, LDU, LDV,
*      $                   LDWH, LDWV, LDZ, N, NH, NSHFTS, NV
*       LOGICAL            WANTT, WANTZ
*       ..
*       .. Array Arguments ..
*       DOUBLE PRECISION   H( LDH, * ), SI( * ), SR( * ), U( LDU, * ),
*      $                   V( LDV, * ), WH( LDWH, * ), WV( LDWV, * ),
*      $                   Z( LDZ, * )
*       ..
*  
*
*> \par Purpose:
*  =============
*>
*> \verbatim
*>
*>    DLAQR5, called by DLAQR0, performs a
*>    single small-bulge multi-shift QR sweep.
*> \endverbatim
*
*  Arguments:
*  ==========
*
*> \param[in] WANTT
*> \verbatim
*>          WANTT is logical scalar
*>             WANTT = .true. if the quasi-triangular Schur factor
*>             is being computed.  WANTT is set to .false. otherwise.
*> \endverbatim
*>
*> \param[in] WANTZ
*> \verbatim
*>          WANTZ is logical scalar
*>             WANTZ = .true. if the orthogonal Schur factor is being
*>             computed.  WANTZ is set to .false. otherwise.
*> \endverbatim
*>
*> \param[in] KACC22
*> \verbatim
*>          KACC22 is integer with value 0, 1, or 2.
*>             Specifies the computation mode of far-from-diagonal
*>             orthogonal updates.
*>        = 0: DLAQR5 does not accumulate reflections and does not
*>             use matrix-matrix multiply to update far-from-diagonal
*>             matrix entries.
*>        = 1: DLAQR5 accumulates reflections and uses matrix-matrix
*>             multiply to update the far-from-diagonal matrix entries.
*>        = 2: DLAQR5 accumulates reflections, uses matrix-matrix
*>             multiply to update the far-from-diagonal matrix entries,
*>             and takes advantage of 2-by-2 block structure during
*>             matrix multiplies.
*> \endverbatim
*>
*> \param[in] N
*> \verbatim
*>          N is integer scalar
*>             N is the order of the Hessenberg matrix H upon which this
*>             subroutine operates.
*> \endverbatim
*>
*> \param[in] KTOP
*> \verbatim
*>          KTOP is integer scalar
*> \endverbatim
*>
*> \param[in] KBOT
*> \verbatim
*>          KBOT is integer scalar
*>             These are the first and last rows and columns of an
*>             isolated diagonal block upon which the QR sweep is to be
*>             applied. It is assumed without a check that
*>                       either KTOP = 1  or   H(KTOP,KTOP-1) = 0
*>             and
*>                       either KBOT = N  or   H(KBOT+1,KBOT) = 0.
*> \endverbatim
*>
*> \param[in] NSHFTS
*> \verbatim
*>          NSHFTS is integer scalar
*>             NSHFTS gives the number of simultaneous shifts.  NSHFTS
*>             must be positive and even.
*> \endverbatim
*>
*> \param[in,out] SR
*> \verbatim
*>          SR is DOUBLE PRECISION array of size (NSHFTS)
*> \endverbatim
*>
*> \param[in,out] SI
*> \verbatim
*>          SI is DOUBLE PRECISION array of size (NSHFTS)
*>             SR contains the real parts and SI contains the imaginary
*>             parts of the NSHFTS shifts of origin that define the
*>             multi-shift QR sweep.  On output SR and SI may be
*>             reordered.
*> \endverbatim
*>
*> \param[in,out] H
*> \verbatim
*>          H is DOUBLE PRECISION array of size (LDH,N)
*>             On input H contains a Hessenberg matrix.  On output a
*>             multi-shift QR sweep with shifts SR(J)+i*SI(J) is applied
*>             to the isolated diagonal block in rows and columns KTOP
*>             through KBOT.
*> \endverbatim
*>
*> \param[in] LDH
*> \verbatim
*>          LDH is integer scalar
*>             LDH is the leading dimension of H just as declared in the
*>             calling procedure.  LDH.GE.MAX(1,N).
*> \endverbatim
*>
*> \param[in] ILOZ
*> \verbatim
*>          ILOZ is INTEGER
*> \endverbatim
*>
*> \param[in] IHIZ
*> \verbatim
*>          IHIZ is INTEGER
*>             Specify the rows of Z to which transformations must be
*>             applied if WANTZ is .TRUE.. 1 .LE. ILOZ .LE. IHIZ .LE. N
*> \endverbatim
*>
*> \param[in,out] Z
*> \verbatim
*>          Z is DOUBLE PRECISION array of size (LDZ,IHIZ)
*>             If WANTZ = .TRUE., then the QR Sweep orthogonal
*>             similarity transformation is accumulated into
*>             Z(ILOZ:IHIZ,ILOZ:IHIZ) from the right.
*>             If WANTZ = .FALSE., then Z is unreferenced.
*> \endverbatim
*>
*> \param[in] LDZ
*> \verbatim
*>          LDZ is integer scalar
*>             LDA is the leading dimension of Z just as declared in
*>             the calling procedure. LDZ.GE.N.
*> \endverbatim
*>
*> \param[out] V
*> \verbatim
*>          V is DOUBLE PRECISION array of size (LDV,NSHFTS/2)
*> \endverbatim
*>
*> \param[in] LDV
*> \verbatim
*>          LDV is integer scalar
*>             LDV is the leading dimension of V as declared in the
*>             calling procedure.  LDV.GE.3.
*> \endverbatim
*>
*> \param[out] U
*> \verbatim
*>          U is DOUBLE PRECISION array of size
*>             (LDU,3*NSHFTS-3)
*> \endverbatim
*>
*> \param[in] LDU
*> \verbatim
*>          LDU is integer scalar
*>             LDU is the leading dimension of U just as declared in the
*>             in the calling subroutine.  LDU.GE.3*NSHFTS-3.
*> \endverbatim
*>
*> \param[in] NH
*> \verbatim
*>          NH is integer scalar
*>             NH is the number of columns in array WH available for
*>             workspace. NH.GE.1.
*> \endverbatim
*>
*> \param[out] WH
*> \verbatim
*>          WH is DOUBLE PRECISION array of size (LDWH,NH)
*> \endverbatim
*>
*> \param[in] LDWH
*> \verbatim
*>          LDWH is integer scalar
*>             Leading dimension of WH just as declared in the
*>             calling procedure.  LDWH.GE.3*NSHFTS-3.
*> \endverbatim
*>
*> \param[in] NV
*> \verbatim
*>          NV is integer scalar
*>             NV is the number of rows in WV agailable for workspace.
*>             NV.GE.1.
*> \endverbatim
*>
*> \param[out] WV
*> \verbatim
*>          WV is DOUBLE PRECISION array of size
*>             (LDWV,3*NSHFTS-3)
*> \endverbatim
*>
*> \param[in] LDWV
*> \verbatim
*>          LDWV is integer scalar
*>             LDWV is the leading dimension of WV as declared in the
*>             in the calling subroutine.  LDWV.GE.NV.
*> \endverbatim
*
*  Authors:
*  ========
*
*> \author Univ. of Tennessee 
*> \author Univ. of California Berkeley 
*> \author Univ. of Colorado Denver 
*> \author NAG Ltd. 
*
*> \date June 2016
*
*> \ingroup doubleOTHERauxiliary
*
*> \par Contributors:
*  ==================
*>
*>       Karen Braman and Ralph Byers, Department of Mathematics,
*>       University of Kansas, USA
*
*> \par References:
*  ================
*>
*>       K. Braman, R. Byers and R. Mathias, The Multi-Shift QR
*>       Algorithm Part I: Maintaining Well Focused Shifts, and Level 3
*>       Performance, SIAM Journal of Matrix Analysis, volume 23, pages
*>       929--947, 2002.
*>
*  =====================================================================
      SUBROUTINE dlaqr5( WANTT, WANTZ, KACC22, N, KTOP, KBOT, NSHFTS,
     $                   sr, si, h, ldh, iloz, ihiz, z, ldz, v, ldv, u,
     $                   ldu, nv, wv, ldwv, nh, wh, ldwh )
*
*  -- LAPACK auxiliary routine (version 3.6.1) --
*  -- LAPACK is a software package provided by Univ. of Tennessee,    --
*  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
*     June 2016
*
*     .. Scalar Arguments ..
      INTEGER            IHIZ, ILOZ, KACC22, KBOT, KTOP, LDH, LDU, LDV,
     $                   ldwh, ldwv, ldz, n, nh, nshfts, nv
      LOGICAL            WANTT, WANTZ
*     ..
*     .. Array Arguments ..
      DOUBLE PRECISION   H( ldh, * ), SI( * ), SR( * ), U( ldu, * ),
     $                   v( ldv, * ), wh( ldwh, * ), wv( ldwv, * ),
     $                   z( ldz, * )
*     ..
*
*  ================================================================
*     .. Parameters ..
      DOUBLE PRECISION   ZERO, ONE
      parameter                ( zero = 0.0d0, one = 1.0d0 )
*     ..
*     .. Local Scalars ..
      DOUBLE PRECISION   ALPHA, BETA, H11, H12, H21, H22, REFSUM,
     $                   safmax, safmin, scl, smlnum, swap, tst1, tst2,
     $                   ulp
      INTEGER            I, I2, I4, INCOL, J, J2, J4, JBOT, JCOL, JLEN,
     $                   jrow, jtop, k, k1, kdu, kms, knz, krcol, kzs,
     $                   m, m22, mbot, mend, mstart, mtop, nbmps, ndcol,
     $                   ns, nu
      LOGICAL            ACCUM, BLK22, BMP22
*     ..
*     .. External Functions ..
      DOUBLE PRECISION   DLAMCH
      EXTERNAL           dlamch
*     ..
*     .. Intrinsic Functions ..
*
      INTRINSIC          abs, dble, max, min, mod
*     ..
*     .. Local Arrays ..
      DOUBLE PRECISION   VT( 3 )
*     ..
*     .. External Subroutines ..
      EXTERNAL           dgemm, dlabad, dlacpy, dlaqr1, dlarfg, dlaset,
     $                   dtrmm
*     ..
*     .. Executable Statements ..
*
*     ==== If there are no shifts, then there is nothing to do. ====
*
      IF( nshfts.LT.2 )
     $   RETURN
*
*     ==== If the active block is empty or 1-by-1, then there
*     .    is nothing to do. ====
*
      IF( ktop.GE.kbot )
     $   RETURN
*
*     ==== Shuffle shifts into pairs of real shifts and pairs
*     .    of complex conjugate shifts assuming complex
*     .    conjugate shifts are already adjacent to one
*     .    another. ====
*
      DO 10 i = 1, nshfts - 2, 2
         IF( si( i ).NE.-si( i+1 ) ) THEN
*
            swap = sr( i )
            sr( i ) = sr( i+1 )
            sr( i+1 ) = sr( i+2 )
            sr( i+2 ) = swap
*
            swap = si( i )
            si( i ) = si( i+1 )
            si( i+1 ) = si( i+2 )
            si( i+2 ) = swap
         END IF
   10 CONTINUE
*
*     ==== NSHFTS is supposed to be even, but if it is odd,
*     .    then simply reduce it by one.  The shuffle above
*     .    ensures that the dropped shift is real and that
*     .    the remaining shifts are paired. ====
*
      ns = nshfts - mod( nshfts, 2 )
*
*     ==== Machine constants for deflation ====
*
      safmin = dlamch( 'SAFE MINIMUM' )
      safmax = one / safmin
      CALL dlabad( safmin, safmax )
      ulp = dlamch( 'PRECISION' )
      smlnum = safmin*( dble( n ) / ulp )
*
*     ==== Use accumulated reflections to update far-from-diagonal
*     .    entries ? ====
*
      accum = ( kacc22.EQ.1 ) .OR. ( kacc22.EQ.2 )
*
*     ==== If so, exploit the 2-by-2 block structure? ====
*
      blk22 = ( ns.GT.2 ) .AND. ( kacc22.EQ.2 )
*
*     ==== clear trash ====
*
      IF( ktop+2.LE.kbot )
     $   h( ktop+2, ktop ) = zero
*
*     ==== NBMPS = number of 2-shift bulges in the chain ====
*
      nbmps = ns / 2
*
*     ==== KDU = width of slab ====
*
      kdu = 6*nbmps - 3
*
*     ==== Create and chase chains of NBMPS bulges ====
*
      DO 220 incol = 3*( 1-nbmps ) + ktop - 1, kbot - 2, 3*nbmps - 2
         ndcol = incol + kdu
         IF( accum )
     $      CALL dlaset( 'ALL', kdu, kdu, zero, one, u, ldu )
*
*        ==== Near-the-diagonal bulge chase.  The following loop
*        .    performs the near-the-diagonal part of a small bulge
*        .    multi-shift QR sweep.  Each 6*NBMPS-2 column diagonal
*        .    chunk extends from column INCOL to column NDCOL
*        .    (including both column INCOL and column NDCOL). The
*        .    following loop chases a 3*NBMPS column long chain of
*        .    NBMPS bulges 3*NBMPS-2 columns to the right.  (INCOL
*        .    may be less than KTOP and and NDCOL may be greater than
*        .    KBOT indicating phantom columns from which to chase
*        .    bulges before they are actually introduced or to which
*        .    to chase bulges beyond column KBOT.)  ====
*
         DO 150 krcol = incol, min( incol+3*nbmps-3, kbot-2 )
*
*           ==== Bulges number MTOP to MBOT are active double implicit
*           .    shift bulges.  There may or may not also be small
*           .    2-by-2 bulge, if there is room.  The inactive bulges
*           .    (if any) must wait until the active bulges have moved
*           .    down the diagonal to make room.  The phantom matrix
*           .    paradigm described above helps keep track.  ====
*
            mtop = max( 1, ( ( ktop-1 )-krcol+2 ) / 3+1 )
            mbot = min( nbmps, ( kbot-krcol ) / 3 )
            m22 = mbot + 1
            bmp22 = ( mbot.LT.nbmps ) .AND. ( krcol+3*( m22-1 ) ).EQ.
     $              ( kbot-2 )
*
*           ==== Generate reflections to chase the chain right
*           .    one column.  (The minimum value of K is KTOP-1.) ====
*
            DO 20 m = mtop, mbot
               k = krcol + 3*( m-1 )
               IF( k.EQ.ktop-1 ) THEN
                  CALL dlaqr1( 3, h( ktop, ktop ), ldh, sr( 2*m-1 ),
     $                         si( 2*m-1 ), sr( 2*m ), si( 2*m ),
     $                         v( 1, m ) )
                  alpha = v( 1, m )
                  CALL dlarfg( 3, alpha, v( 2, m ), 1, v( 1, m ) )
               ELSE
                  beta = h( k+1, k )
                  v( 2, m ) = h( k+2, k )
                  v( 3, m ) = h( k+3, k )
                  CALL dlarfg( 3, beta, v( 2, m ), 1, v( 1, m ) )
*
*                 ==== A Bulge may collapse because of vigilant
*                 .    deflation or destructive underflow.  In the
*                 .    underflow case, try the two-small-subdiagonals
*                 .    trick to try to reinflate the bulge.  ====
*
                  IF( h( k+3, k ).NE.zero .OR. h( k+3, k+1 ).NE.
     $                zero .OR. h( k+3, k+2 ).EQ.zero ) THEN
*
*                    ==== Typical case: not collapsed (yet). ====
*
                     h( k+1, k ) = beta
                     h( k+2, k ) = zero
                     h( k+3, k ) = zero
                  ELSE
*
*                    ==== Atypical case: collapsed.  Attempt to
*                    .    reintroduce ignoring H(K+1,K) and H(K+2,K).
*                    .    If the fill resulting from the new
*                    .    reflector is too large, then abandon it.
*                    .    Otherwise, use the new one. ====
*
                     CALL dlaqr1( 3, h( k+1, k+1 ), ldh, sr( 2*m-1 ),
     $                            si( 2*m-1 ), sr( 2*m ), si( 2*m ),
     $                            vt )
                     alpha = vt( 1 )
                     CALL dlarfg( 3, alpha, vt( 2 ), 1, vt( 1 ) )
                     refsum = vt( 1 )*( h( k+1, k )+vt( 2 )*
     $                        h( k+2, k ) )
*
                     IF( abs( h( k+2, k )-refsum*vt( 2 ) )+
     $                   abs( refsum*vt( 3 ) ).GT.ulp*
     $                   ( abs( h( k, k ) )+abs( h( k+1,
     $                   k+1 ) )+abs( h( k+2, k+2 ) ) ) ) THEN
*
*                       ==== Starting a new bulge here would
*                       .    create non-negligible fill.  Use
*                       .    the old one with trepidation. ====
*
                        h( k+1, k ) = beta
                        h( k+2, k ) = zero
                        h( k+3, k ) = zero
                     ELSE
*
*                       ==== Stating a new bulge here would
*                       .    create only negligible fill.
*                       .    Replace the old reflector with
*                       .    the new one. ====
*
                        h( k+1, k ) = h( k+1, k ) - refsum
                        h( k+2, k ) = zero
                        h( k+3, k ) = zero
                        v( 1, m ) = vt( 1 )
                        v( 2, m ) = vt( 2 )
                        v( 3, m ) = vt( 3 )
                     END IF
                  END IF
               END IF
   20       CONTINUE
*
*           ==== Generate a 2-by-2 reflection, if needed. ====
*
            k = krcol + 3*( m22-1 )
            IF( bmp22 ) THEN
               IF( k.EQ.ktop-1 ) THEN
                  CALL dlaqr1( 2, h( k+1, k+1 ), ldh, sr( 2*m22-1 ),
     $                         si( 2*m22-1 ), sr( 2*m22 ), si( 2*m22 ),
     $                         v( 1, m22 ) )
                  beta = v( 1, m22 )
                  CALL dlarfg( 2, beta, v( 2, m22 ), 1, v( 1, m22 ) )
               ELSE
                  beta = h( k+1, k )
                  v( 2, m22 ) = h( k+2, k )
                  CALL dlarfg( 2, beta, v( 2, m22 ), 1, v( 1, m22 ) )
                  h( k+1, k ) = beta
                  h( k+2, k ) = zero
               END IF
            END IF
*
*           ==== Multiply H by reflections from the left ====
*
            IF( accum ) THEN
               jbot = min( ndcol, kbot )
            ELSE IF( wantt ) THEN
               jbot = n
            ELSE
               jbot = kbot
            END IF
            DO 40 j = max( ktop, krcol ), jbot
               mend = min( mbot, ( j-krcol+2 ) / 3 )
               DO 30 m = mtop, mend
                  k = krcol + 3*( m-1 )
                  refsum = v( 1, m )*( h( k+1, j )+v( 2, m )*
     $                     h( k+2, j )+v( 3, m )*h( k+3, j ) )
                  h( k+1, j ) = h( k+1, j ) - refsum
                  h( k+2, j ) = h( k+2, j ) - refsum*v( 2, m )
                  h( k+3, j ) = h( k+3, j ) - refsum*v( 3, m )
   30          CONTINUE
   40       CONTINUE
            IF( bmp22 ) THEN
               k = krcol + 3*( m22-1 )
               DO 50 j = max( k+1, ktop ), jbot
                  refsum = v( 1, m22 )*( h( k+1, j )+v( 2, m22 )*
     $                     h( k+2, j ) )
                  h( k+1, j ) = h( k+1, j ) - refsum
                  h( k+2, j ) = h( k+2, j ) - refsum*v( 2, m22 )
   50          CONTINUE
            END IF
*
*           ==== Multiply H by reflections from the right.
*           .    Delay filling in the last row until the
*           .    vigilant deflation check is complete. ====
*
            IF( accum ) THEN
               jtop = max( ktop, incol )
            ELSE IF( wantt ) THEN
               jtop = 1
            ELSE
               jtop = ktop
            END IF
            DO 90 m = mtop, mbot
               IF( v( 1, m ).NE.zero ) THEN
                  k = krcol + 3*( m-1 )
                  DO 60 j = jtop, min( kbot, k+3 )
                     refsum = v( 1, m )*( h( j, k+1 )+v( 2, m )*
     $                        h( j, k+2 )+v( 3, m )*h( j, k+3 ) )
                     h( j, k+1 ) = h( j, k+1 ) - refsum
                     h( j, k+2 ) = h( j, k+2 ) - refsum*v( 2, m )
                     h( j, k+3 ) = h( j, k+3 ) - refsum*v( 3, m )
   60             CONTINUE
*
                  IF( accum ) THEN
*
*                    ==== Accumulate U. (If necessary, update Z later
*                    .    with with an efficient matrix-matrix
*                    .    multiply.) ====
*
                     kms = k - incol
                     DO 70 j = max( 1, ktop-incol ), kdu
                        refsum = v( 1, m )*( u( j, kms+1 )+v( 2, m )*
     $                           u( j, kms+2 )+v( 3, m )*u( j, kms+3 ) )
                        u( j, kms+1 ) = u( j, kms+1 ) - refsum
                        u( j, kms+2 ) = u( j, kms+2 ) - refsum*v( 2, m )
                        u( j, kms+3 ) = u( j, kms+3 ) - refsum*v( 3, m )
   70                CONTINUE
                  ELSE IF( wantz ) THEN
*
*                    ==== U is not accumulated, so update Z
*                    .    now by multiplying by reflections
*                    .    from the right. ====
*
                     DO 80 j = iloz, ihiz
                        refsum = v( 1, m )*( z( j, k+1 )+v( 2, m )*
     $                           z( j, k+2 )+v( 3, m )*z( j, k+3 ) )
                        z( j, k+1 ) = z( j, k+1 ) - refsum
                        z( j, k+2 ) = z( j, k+2 ) - refsum*v( 2, m )
                        z( j, k+3 ) = z( j, k+3 ) - refsum*v( 3, m )
   80                CONTINUE
                  END IF
               END IF
   90       CONTINUE
*
*           ==== Special case: 2-by-2 reflection (if needed) ====
*
            k = krcol + 3*( m22-1 )
            IF( bmp22 ) THEN
               IF ( v( 1, m22 ).NE.zero ) THEN
                  DO 100 j = jtop, min( kbot, k+3 )
                     refsum = v( 1, m22 )*( h( j, k+1 )+v( 2, m22 )*
     $                        h( j, k+2 ) )
                     h( j, k+1 ) = h( j, k+1 ) - refsum
                     h( j, k+2 ) = h( j, k+2 ) - refsum*v( 2, m22 )
  100             CONTINUE
*
                  IF( accum ) THEN
                     kms = k - incol
                     DO 110 j = max( 1, ktop-incol ), kdu
                        refsum = v( 1, m22 )*( u( j, kms+1 )+
     $                           v( 2, m22 )*u( j, kms+2 ) )
                        u( j, kms+1 ) = u( j, kms+1 ) - refsum
                        u( j, kms+2 ) = u( j, kms+2 ) -
     $                                  refsum*v( 2, m22 )
  110             CONTINUE
                  ELSE IF( wantz ) THEN
                     DO 120 j = iloz, ihiz
                        refsum = v( 1, m22 )*( z( j, k+1 )+v( 2, m22 )*
     $                           z( j, k+2 ) )
                        z( j, k+1 ) = z( j, k+1 ) - refsum
                        z( j, k+2 ) = z( j, k+2 ) - refsum*v( 2, m22 )
  120                CONTINUE
                  END IF
               END IF
            END IF
*
*           ==== Vigilant deflation check ====
*
            mstart = mtop
            IF( krcol+3*( mstart-1 ).LT.ktop )
     $         mstart = mstart + 1
            mend = mbot
            IF( bmp22 )
     $         mend = mend + 1
            IF( krcol.EQ.kbot-2 )
     $         mend = mend + 1
            DO 130 m = mstart, mend
               k = min( kbot-1, krcol+3*( m-1 ) )
*
*              ==== The following convergence test requires that
*              .    the tradition small-compared-to-nearby-diagonals
*              .    criterion and the Ahues & Tisseur (LAWN 122, 1997)
*              .    criteria both be satisfied.  The latter improves
*              .    accuracy in some examples. Falling back on an
*              .    alternate convergence criterion when TST1 or TST2
*              .    is zero (as done here) is traditional but probably
*              .    unnecessary. ====
*
               IF( h( k+1, k ).NE.zero ) THEN
                  tst1 = abs( h( k, k ) ) + abs( h( k+1, k+1 ) )
                  IF( tst1.EQ.zero ) THEN
                     IF( k.GE.ktop+1 )
     $                  tst1 = tst1 + abs( h( k, k-1 ) )
                     IF( k.GE.ktop+2 )
     $                  tst1 = tst1 + abs( h( k, k-2 ) )
                     IF( k.GE.ktop+3 )
     $                  tst1 = tst1 + abs( h( k, k-3 ) )
                     IF( k.LE.kbot-2 )
     $                  tst1 = tst1 + abs( h( k+2, k+1 ) )
                     IF( k.LE.kbot-3 )
     $                  tst1 = tst1 + abs( h( k+3, k+1 ) )
                     IF( k.LE.kbot-4 )
     $                  tst1 = tst1 + abs( h( k+4, k+1 ) )
                  END IF
                  IF( abs( h( k+1, k ) ).LE.max( smlnum, ulp*tst1 ) )
     $                 THEN
                     h12 = max( abs( h( k+1, k ) ), abs( h( k, k+1 ) ) )
                     h21 = min( abs( h( k+1, k ) ), abs( h( k, k+1 ) ) )
                     h11 = max( abs( h( k+1, k+1 ) ),
     $                     abs( h( k, k )-h( k+1, k+1 ) ) )
                     h22 = min( abs( h( k+1, k+1 ) ),
     $                     abs( h( k, k )-h( k+1, k+1 ) ) )
                     scl = h11 + h12
                     tst2 = h22*( h11 / scl )
*
                     IF( tst2.EQ.zero .OR. h21*( h12 / scl ).LE.
     $                   max( smlnum, ulp*tst2 ) )h( k+1, k ) = zero
                  END IF
               END IF
  130       CONTINUE
*
*           ==== Fill in the last row of each bulge. ====
*
            mend = min( nbmps, ( kbot-krcol-1 ) / 3 )
            DO 140 m = mtop, mend
               k = krcol + 3*( m-1 )
               refsum = v( 1, m )*v( 3, m )*h( k+4, k+3 )
               h( k+4, k+1 ) = -refsum
               h( k+4, k+2 ) = -refsum*v( 2, m )
               h( k+4, k+3 ) = h( k+4, k+3 ) - refsum*v( 3, m )
  140       CONTINUE
*
*           ==== End of near-the-diagonal bulge chase. ====
*
  150    CONTINUE
*
*        ==== Use U (if accumulated) to update far-from-diagonal
*        .    entries in H.  If required, use U to update Z as
*        .    well. ====
*
         IF( accum ) THEN
            IF( wantt ) THEN
               jtop = 1
               jbot = n
            ELSE
               jtop = ktop
               jbot = kbot
            END IF
            IF( ( .NOT.blk22 ) .OR. ( incol.LT.ktop ) .OR.
     $          ( ndcol.GT.kbot ) .OR. ( ns.LE.2 ) ) THEN
*
*              ==== Updates not exploiting the 2-by-2 block
*              .    structure of U.  K1 and NU keep track of
*              .    the location and size of U in the special
*              .    cases of introducing bulges and chasing
*              .    bulges off the bottom.  In these special
*              .    cases and in case the number of shifts
*              .    is NS = 2, there is no 2-by-2 block
*              .    structure to exploit.  ====
*
               k1 = max( 1, ktop-incol )
               nu = ( kdu-max( 0, ndcol-kbot ) ) - k1 + 1
*
*              ==== Horizontal Multiply ====
*
               DO 160 jcol = min( ndcol, kbot ) + 1, jbot, nh
                  jlen = min( nh, jbot-jcol+1 )
                  CALL dgemm( 'C', 'N', nu, jlen, nu, one, u( k1, k1 ),
     $                        ldu, h( incol+k1, jcol ), ldh, zero, wh,
     $                        ldwh )
                  CALL dlacpy( 'ALL', nu, jlen, wh, ldwh,
     $                         h( incol+k1, jcol ), ldh )
  160          CONTINUE
*
*              ==== Vertical multiply ====
*
               DO 170 jrow = jtop, max( ktop, incol ) - 1, nv
                  jlen = min( nv, max( ktop, incol )-jrow )
                  CALL dgemm( 'N', 'N', jlen, nu, nu, one,
     $                        h( jrow, incol+k1 ), ldh, u( k1, k1 ),
     $                        ldu, zero, wv, ldwv )
                  CALL dlacpy( 'ALL', jlen, nu, wv, ldwv,
     $                         h( jrow, incol+k1 ), ldh )
  170          CONTINUE
*
*              ==== Z multiply (also vertical) ====
*
               IF( wantz ) THEN
                  DO 180 jrow = iloz, ihiz, nv
                     jlen = min( nv, ihiz-jrow+1 )
                     CALL dgemm( 'N', 'N', jlen, nu, nu, one,
     $                           z( jrow, incol+k1 ), ldz, u( k1, k1 ),
     $                           ldu, zero, wv, ldwv )
                     CALL dlacpy( 'ALL', jlen, nu, wv, ldwv,
     $                            z( jrow, incol+k1 ), ldz )
  180             CONTINUE
               END IF
            ELSE
*
*              ==== Updates exploiting U's 2-by-2 block structure.
*              .    (I2, I4, J2, J4 are the last rows and columns
*              .    of the blocks.) ====
*
               i2 = ( kdu+1 ) / 2
               i4 = kdu
               j2 = i4 - i2
               j4 = kdu
*
*              ==== KZS and KNZ deal with the band of zeros
*              .    along the diagonal of one of the triangular
*              .    blocks. ====
*
               kzs = ( j4-j2 ) - ( ns+1 )
               knz = ns + 1
*
*              ==== Horizontal multiply ====
*
               DO 190 jcol = min( ndcol, kbot ) + 1, jbot, nh
                  jlen = min( nh, jbot-jcol+1 )
*
*                 ==== Copy bottom of H to top+KZS of scratch ====
*                  (The first KZS rows get multiplied by zero.) ====
*
                  CALL dlacpy( 'ALL', knz, jlen, h( incol+1+j2, jcol ),
     $                         ldh, wh( kzs+1, 1 ), ldwh )
*
*                 ==== Multiply by U21**T ====
*
                  CALL dlaset( 'ALL', kzs, jlen, zero, zero, wh, ldwh )
                  CALL dtrmm( 'L', 'U', 'C', 'N', knz, jlen, one,
     $                        u( j2+1, 1+kzs ), ldu, wh( kzs+1, 1 ),
     $                        ldwh )
*
*                 ==== Multiply top of H by U11**T ====
*
                  CALL dgemm( 'C', 'N', i2, jlen, j2, one, u, ldu,
     $                        h( incol+1, jcol ), ldh, one, wh, ldwh )
*
*                 ==== Copy top of H to bottom of WH ====
*
                  CALL dlacpy( 'ALL', j2, jlen, h( incol+1, jcol ), ldh,
     $                         wh( i2+1, 1 ), ldwh )
*
*                 ==== Multiply by U21**T ====
*
                  CALL dtrmm( 'L', 'L', 'C', 'N', j2, jlen, one,
     $                        u( 1, i2+1 ), ldu, wh( i2+1, 1 ), ldwh )
*
*                 ==== Multiply by U22 ====
*
                  CALL dgemm( 'C', 'N', i4-i2, jlen, j4-j2, one,
     $                        u( j2+1, i2+1 ), ldu,
     $                        h( incol+1+j2, jcol ), ldh, one,
     $                        wh( i2+1, 1 ), ldwh )
*
*                 ==== Copy it back ====
*
                  CALL dlacpy( 'ALL', kdu, jlen, wh, ldwh,
     $                         h( incol+1, jcol ), ldh )
  190          CONTINUE
*
*              ==== Vertical multiply ====
*
               DO 200 jrow = jtop, max( incol, ktop ) - 1, nv
                  jlen = min( nv, max( incol, ktop )-jrow )
*
*                 ==== Copy right of H to scratch (the first KZS
*                 .    columns get multiplied by zero) ====
*
                  CALL dlacpy( 'ALL', jlen, knz, h( jrow, incol+1+j2 ),
     $                         ldh, wv( 1, 1+kzs ), ldwv )
*
*                 ==== Multiply by U21 ====
*
                  CALL dlaset( 'ALL', jlen, kzs, zero, zero, wv, ldwv )
                  CALL dtrmm( 'R', 'U', 'N', 'N', jlen, knz, one,
     $                        u( j2+1, 1+kzs ), ldu, wv( 1, 1+kzs ),
     $                        ldwv )
*
*                 ==== Multiply by U11 ====
*
                  CALL dgemm( 'N', 'N', jlen, i2, j2, one,
     $                        h( jrow, incol+1 ), ldh, u, ldu, one, wv,
     $                        ldwv )
*
*                 ==== Copy left of H to right of scratch ====
*
                  CALL dlacpy( 'ALL', jlen, j2, h( jrow, incol+1 ), ldh,
     $                         wv( 1, 1+i2 ), ldwv )
*
*                 ==== Multiply by U21 ====
*
                  CALL dtrmm( 'R', 'L', 'N', 'N', jlen, i4-i2, one,
     $                        u( 1, i2+1 ), ldu, wv( 1, 1+i2 ), ldwv )
*
*                 ==== Multiply by U22 ====
*
                  CALL dgemm( 'N', 'N', jlen, i4-i2, j4-j2, one,
     $                        h( jrow, incol+1+j2 ), ldh,
     $                        u( j2+1, i2+1 ), ldu, one, wv( 1, 1+i2 ),
     $                        ldwv )
*
*                 ==== Copy it back ====
*
                  CALL dlacpy( 'ALL', jlen, kdu, wv, ldwv,
     $                         h( jrow, incol+1 ), ldh )
  200          CONTINUE
*
*              ==== Multiply Z (also vertical) ====
*
               IF( wantz ) THEN
                  DO 210 jrow = iloz, ihiz, nv
                     jlen = min( nv, ihiz-jrow+1 )
*
*                    ==== Copy right of Z to left of scratch (first
*                    .     KZS columns get multiplied by zero) ====
*
                     CALL dlacpy( 'ALL', jlen, knz,
     $                            z( jrow, incol+1+j2 ), ldz,
     $                            wv( 1, 1+kzs ), ldwv )
*
*                    ==== Multiply by U12 ====
*
                     CALL dlaset( 'ALL', jlen, kzs, zero, zero, wv,
     $                            ldwv )
                     CALL dtrmm( 'R', 'U', 'N', 'N', jlen, knz, one,
     $                           u( j2+1, 1+kzs ), ldu, wv( 1, 1+kzs ),
     $                           ldwv )
*
*                    ==== Multiply by U11 ====
*
                     CALL dgemm( 'N', 'N', jlen, i2, j2, one,
     $                           z( jrow, incol+1 ), ldz, u, ldu, one,
     $                           wv, ldwv )
*
*                    ==== Copy left of Z to right of scratch ====
*
                     CALL dlacpy( 'ALL', jlen, j2, z( jrow, incol+1 ),
     $                            ldz, wv( 1, 1+i2 ), ldwv )
*
*                    ==== Multiply by U21 ====
*
                     CALL dtrmm( 'R', 'L', 'N', 'N', jlen, i4-i2, one,
     $                           u( 1, i2+1 ), ldu, wv( 1, 1+i2 ),
     $                           ldwv )
*
*                    ==== Multiply by U22 ====
*
                     CALL dgemm( 'N', 'N', jlen, i4-i2, j4-j2, one,
     $                           z( jrow, incol+1+j2 ), ldz,
     $                           u( j2+1, i2+1 ), ldu, one,
     $                           wv( 1, 1+i2 ), ldwv )
*
*                    ==== Copy the result back to Z ====
*
                     CALL dlacpy( 'ALL', jlen, kdu, wv, ldwv,
     $                            z( jrow, incol+1 ), ldz )
  210             CONTINUE
               END IF
            END IF
         END IF
  220 CONTINUE
*
*     ==== End of DLAQR5 ====
*
      END