1*> \brief \b CLAQR4 computes the eigenvalues of a Hessenberg matrix, and optionally the matrices from the Schur decomposition.
2*
3*  =========== DOCUMENTATION ===========
4*
5* Online html documentation available at
6*            http://www.netlib.org/lapack/explore-html/
7*
8*> \htmlonly
9*> Download CLAQR4 + dependencies
10*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/claqr4.f">
11*> [TGZ]</a>
12*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/claqr4.f">
13*> [ZIP]</a>
14*> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/claqr4.f">
15*> [TXT]</a>
16*> \endhtmlonly
17*
18*  Definition:
19*  ===========
20*
21*       SUBROUTINE CLAQR4( WANTT, WANTZ, N, ILO, IHI, H, LDH, W, ILOZ,
22*                          IHIZ, Z, LDZ, WORK, LWORK, INFO )
23*
24*       .. Scalar Arguments ..
25*       INTEGER            IHI, IHIZ, ILO, ILOZ, INFO, LDH, LDZ, LWORK, N
26*       LOGICAL            WANTT, WANTZ
27*       ..
28*       .. Array Arguments ..
29*       COMPLEX            H( LDH, * ), W( * ), WORK( * ), Z( LDZ, * )
30*       ..
31*
32*
33*
34*> \par Purpose:
35*  =============
36*>
37*> \verbatim
38*>
39*>    CLAQR4 implements one level of recursion for CLAQR0.
40*>    It is a complete implementation of the small bulge multi-shift
41*>    QR algorithm.  It may be called by CLAQR0 and, for large enough
42*>    deflation window size, it may be called by CLAQR3.  This
43*>    subroutine is identical to CLAQR0 except that it calls CLAQR2
44*>    instead of CLAQR3.
45*>
46*>    CLAQR4 computes the eigenvalues of a Hessenberg matrix H
47*>    and, optionally, the matrices T and Z from the Schur decomposition
48*>    H = Z T Z**H, where T is an upper triangular matrix (the
49*>    Schur form), and Z is the unitary matrix of Schur vectors.
50*>
51*>    Optionally Z may be postmultiplied into an input unitary
52*>    matrix Q so that this routine can give the Schur factorization
53*>    of a matrix A which has been reduced to the Hessenberg form H
54*>    by the unitary matrix Q:  A = Q*H*Q**H = (QZ)*H*(QZ)**H.
55*> \endverbatim
56*
57*  Arguments:
58*  ==========
59*
60*> \param[in] WANTT
61*> \verbatim
62*>          WANTT is LOGICAL
63*>          = .TRUE. : the full Schur form T is required;
64*>          = .FALSE.: only eigenvalues are required.
65*> \endverbatim
66*>
67*> \param[in] WANTZ
68*> \verbatim
69*>          WANTZ is LOGICAL
70*>          = .TRUE. : the matrix of Schur vectors Z is required;
71*>          = .FALSE.: Schur vectors are not required.
72*> \endverbatim
73*>
74*> \param[in] N
75*> \verbatim
76*>          N is INTEGER
77*>           The order of the matrix H.  N .GE. 0.
78*> \endverbatim
79*>
80*> \param[in] ILO
81*> \verbatim
82*>          ILO is INTEGER
83*> \endverbatim
84*>
85*> \param[in] IHI
86*> \verbatim
87*>          IHI is INTEGER
88*>           It is assumed that H is already upper triangular in rows
89*>           and columns 1:ILO-1 and IHI+1:N and, if ILO.GT.1,
90*>           H(ILO,ILO-1) is zero. ILO and IHI are normally set by a
91*>           previous call to CGEBAL, and then passed to CGEHRD when the
92*>           matrix output by CGEBAL is reduced to Hessenberg form.
93*>           Otherwise, ILO and IHI should be set to 1 and N,
94*>           respectively.  If N.GT.0, then 1.LE.ILO.LE.IHI.LE.N.
95*>           If N = 0, then ILO = 1 and IHI = 0.
96*> \endverbatim
97*>
98*> \param[in,out] H
99*> \verbatim
100*>          H is COMPLEX array, dimension (LDH,N)
101*>           On entry, the upper Hessenberg matrix H.
102*>           On exit, if INFO = 0 and WANTT is .TRUE., then H
103*>           contains the upper triangular matrix T from the Schur
104*>           decomposition (the Schur form). If INFO = 0 and WANT is
105*>           .FALSE., then the contents of H are unspecified on exit.
106*>           (The output value of H when INFO.GT.0 is given under the
107*>           description of INFO below.)
108*>
109*>           This subroutine may explicitly set H(i,j) = 0 for i.GT.j and
110*>           j = 1, 2, ... ILO-1 or j = IHI+1, IHI+2, ... N.
111*> \endverbatim
112*>
113*> \param[in] LDH
114*> \verbatim
115*>          LDH is INTEGER
116*>           The leading dimension of the array H. LDH .GE. max(1,N).
117*> \endverbatim
118*>
119*> \param[out] W
120*> \verbatim
121*>          W is COMPLEX array, dimension (N)
122*>           The computed eigenvalues of H(ILO:IHI,ILO:IHI) are stored
123*>           in W(ILO:IHI). If WANTT is .TRUE., then the eigenvalues are
124*>           stored in the same order as on the diagonal of the Schur
125*>           form returned in H, with W(i) = H(i,i).
126*> \endverbatim
127*>
128*> \param[in] ILOZ
129*> \verbatim
130*>          ILOZ is INTEGER
131*> \endverbatim
132*>
133*> \param[in] IHIZ
134*> \verbatim
135*>          IHIZ is INTEGER
136*>           Specify the rows of Z to which transformations must be
137*>           applied if WANTZ is .TRUE..
138*>           1 .LE. ILOZ .LE. ILO; IHI .LE. IHIZ .LE. N.
139*> \endverbatim
140*>
141*> \param[in,out] Z
142*> \verbatim
143*>          Z is COMPLEX array, dimension (LDZ,IHI)
144*>           If WANTZ is .FALSE., then Z is not referenced.
145*>           If WANTZ is .TRUE., then Z(ILO:IHI,ILOZ:IHIZ) is
146*>           replaced by Z(ILO:IHI,ILOZ:IHIZ)*U where U is the
147*>           orthogonal Schur factor of H(ILO:IHI,ILO:IHI).
148*>           (The output value of Z when INFO.GT.0 is given under
149*>           the description of INFO below.)
150*> \endverbatim
151*>
152*> \param[in] LDZ
153*> \verbatim
154*>          LDZ is INTEGER
155*>           The leading dimension of the array Z.  if WANTZ is .TRUE.
156*>           then LDZ.GE.MAX(1,IHIZ).  Otherwize, LDZ.GE.1.
157*> \endverbatim
158*>
159*> \param[out] WORK
160*> \verbatim
161*>          WORK is COMPLEX array, dimension LWORK
162*>           On exit, if LWORK = -1, WORK(1) returns an estimate of
163*>           the optimal value for LWORK.
164*> \endverbatim
165*>
166*> \param[in] LWORK
167*> \verbatim
168*>          LWORK is INTEGER
169*>           The dimension of the array WORK.  LWORK .GE. max(1,N)
170*>           is sufficient, but LWORK typically as large as 6*N may
171*>           be required for optimal performance.  A workspace query
172*>           to determine the optimal workspace size is recommended.
173*>
174*>           If LWORK = -1, then CLAQR4 does a workspace query.
175*>           In this case, CLAQR4 checks the input parameters and
176*>           estimates the optimal workspace size for the given
177*>           values of N, ILO and IHI.  The estimate is returned
178*>           in WORK(1).  No error message related to LWORK is
179*>           issued by XERBLA.  Neither H nor Z are accessed.
180*> \endverbatim
181*>
182*> \param[out] INFO
183*> \verbatim
184*> \verbatim
185*>          INFO is INTEGER
186*>             =  0:  successful exit
187*>           .GT. 0:  if INFO = i, CLAQR4 failed to compute all of
188*>                the eigenvalues.  Elements 1:ilo-1 and i+1:n of WR
189*>                and WI contain those eigenvalues which have been
190*>                successfully computed.  (Failures are rare.)
191*>
192*>                If INFO .GT. 0 and WANT is .FALSE., then on exit,
193*>                the remaining unconverged eigenvalues are the eigen-
194*>                values of the upper Hessenberg matrix rows and
195*>                columns ILO through INFO of the final, output
196*>                value of H.
197*>
198*>                If INFO .GT. 0 and WANTT is .TRUE., then on exit
199*>
200*>           (*)  (initial value of H)*U  = U*(final value of H)
201*>
202*>                where U is a unitary matrix.  The final
203*>                value of  H is upper Hessenberg and triangular in
204*>                rows and columns INFO+1 through IHI.
205*>
206*>                If INFO .GT. 0 and WANTZ is .TRUE., then on exit
207*>
208*>                  (final value of Z(ILO:IHI,ILOZ:IHIZ)
209*>                   =  (initial value of Z(ILO:IHI,ILOZ:IHIZ)*U
210*>
211*>                where U is the unitary matrix in (*) (regard-
212*>                less of the value of WANTT.)
213*>
214*>                If INFO .GT. 0 and WANTZ is .FALSE., then Z is not
215*>                accessed.
216*> \endverbatim
217*
218*  Authors:
219*  ========
220*
221*> \author Univ. of Tennessee
222*> \author Univ. of California Berkeley
223*> \author Univ. of Colorado Denver
224*> \author NAG Ltd.
225*
226*> \date September 2012
227*
228*> \ingroup complexOTHERauxiliary
229*
230*> \par Contributors:
231*  ==================
232*>
233*>       Karen Braman and Ralph Byers, Department of Mathematics,
234*>       University of Kansas, USA
235*
236*> \par References:
237*  ================
238*>
239*>       K. Braman, R. Byers and R. Mathias, The Multi-Shift QR
240*>       Algorithm Part I: Maintaining Well Focused Shifts, and Level 3
241*>       Performance, SIAM Journal of Matrix Analysis, volume 23, pages
242*>       929--947, 2002.
243*> \n
244*>       K. Braman, R. Byers and R. Mathias, The Multi-Shift QR
245*>       Algorithm Part II: Aggressive Early Deflation, SIAM Journal
246*>       of Matrix Analysis, volume 23, pages 948--973, 2002.
247*>
248*  =====================================================================
249      SUBROUTINE CLAQR4( WANTT, WANTZ, N, ILO, IHI, H, LDH, W, ILOZ,
250     $                   IHIZ, Z, LDZ, WORK, LWORK, INFO )
251*
252*  -- LAPACK auxiliary routine (version 3.4.2) --
253*  -- LAPACK is a software package provided by Univ. of Tennessee,    --
254*  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
255*     September 2012
256*
257*     .. Scalar Arguments ..
258      INTEGER            IHI, IHIZ, ILO, ILOZ, INFO, LDH, LDZ, LWORK, N
259      LOGICAL            WANTT, WANTZ
260*     ..
261*     .. Array Arguments ..
262      COMPLEX            H( LDH, * ), W( * ), WORK( * ), Z( LDZ, * )
263*     ..
264*
265*
266*  ================================================================
267*
268*     .. Parameters ..
269*
270*     ==== Matrices of order NTINY or smaller must be processed by
271*     .    CLAHQR because of insufficient subdiagonal scratch space.
272*     .    (This is a hard limit.) ====
273      INTEGER            NTINY
274      PARAMETER          ( NTINY = 11 )
275*
276*     ==== Exceptional deflation windows:  try to cure rare
277*     .    slow convergence by varying the size of the
278*     .    deflation window after KEXNW iterations. ====
279      INTEGER            KEXNW
280      PARAMETER          ( KEXNW = 5 )
281*
282*     ==== Exceptional shifts: try to cure rare slow convergence
283*     .    with ad-hoc exceptional shifts every KEXSH iterations.
284*     .    ====
285      INTEGER            KEXSH
286      PARAMETER          ( KEXSH = 6 )
287*
288*     ==== The constant WILK1 is used to form the exceptional
289*     .    shifts. ====
290      REAL               WILK1
291      PARAMETER          ( WILK1 = 0.75e0 )
292      COMPLEX            ZERO, ONE
293      PARAMETER          ( ZERO = ( 0.0e0, 0.0e0 ),
294     $                   ONE = ( 1.0e0, 0.0e0 ) )
295      REAL               TWO
296      PARAMETER          ( TWO = 2.0e0 )
297*     ..
298*     .. Local Scalars ..
299      COMPLEX            AA, BB, CC, CDUM, DD, DET, RTDISC, SWAP, TR2
300      REAL               S
301      INTEGER            I, INF, IT, ITMAX, K, KACC22, KBOT, KDU, KS,
302     $                   KT, KTOP, KU, KV, KWH, KWTOP, KWV, LD, LS,
303     $                   LWKOPT, NDEC, NDFL, NH, NHO, NIBBLE, NMIN, NS,
304     $                   NSMAX, NSR, NVE, NW, NWMAX, NWR, NWUPBD
305      LOGICAL            SORTED
306      CHARACTER          JBCMPZ*2
307*     ..
308*     .. External Functions ..
309      INTEGER            ILAENV
310      EXTERNAL           ILAENV
311*     ..
312*     .. Local Arrays ..
313      COMPLEX            ZDUM( 1, 1 )
314*     ..
315*     .. External Subroutines ..
316      EXTERNAL           CLACPY, CLAHQR, CLAQR2, CLAQR5
317*     ..
318*     .. Intrinsic Functions ..
319      INTRINSIC          ABS, AIMAG, CMPLX, INT, MAX, MIN, MOD, REAL,
320     $                   SQRT
321*     ..
322*     .. Statement Functions ..
323      REAL               CABS1
324*     ..
325*     .. Statement Function definitions ..
326      CABS1( CDUM ) = ABS( REAL( CDUM ) ) + ABS( AIMAG( CDUM ) )
327*     ..
328*     .. Executable Statements ..
329      INFO = 0
330*
331*     ==== Quick return for N = 0: nothing to do. ====
332*
333      IF( N.EQ.0 ) THEN
334         WORK( 1 ) = ONE
335         RETURN
336      END IF
337*
338      IF( N.LE.NTINY ) THEN
339*
340*        ==== Tiny matrices must use CLAHQR. ====
341*
342         LWKOPT = 1
343         IF( LWORK.NE.-1 )
344     $      CALL CLAHQR( WANTT, WANTZ, N, ILO, IHI, H, LDH, W, ILOZ,
345     $                   IHIZ, Z, LDZ, INFO )
346      ELSE
347*
348*        ==== Use small bulge multi-shift QR with aggressive early
349*        .    deflation on larger-than-tiny matrices. ====
350*
351*        ==== Hope for the best. ====
352*
353         INFO = 0
354*
355*        ==== Set up job flags for ILAENV. ====
356*
357         IF( WANTT ) THEN
358            JBCMPZ( 1: 1 ) = 'S'
359         ELSE
360            JBCMPZ( 1: 1 ) = 'E'
361         END IF
362         IF( WANTZ ) THEN
363            JBCMPZ( 2: 2 ) = 'V'
364         ELSE
365            JBCMPZ( 2: 2 ) = 'N'
366         END IF
367*
368*        ==== NWR = recommended deflation window size.  At this
369*        .    point,  N .GT. NTINY = 11, so there is enough
370*        .    subdiagonal workspace for NWR.GE.2 as required.
371*        .    (In fact, there is enough subdiagonal space for
372*        .    NWR.GE.3.) ====
373*
374         NWR = ILAENV( 13, 'CLAQR4', JBCMPZ, N, ILO, IHI, LWORK )
375         NWR = MAX( 2, NWR )
376         NWR = MIN( IHI-ILO+1, ( N-1 ) / 3, NWR )
377*
378*        ==== NSR = recommended number of simultaneous shifts.
379*        .    At this point N .GT. NTINY = 11, so there is at
380*        .    enough subdiagonal workspace for NSR to be even
381*        .    and greater than or equal to two as required. ====
382*
383         NSR = ILAENV( 15, 'CLAQR4', JBCMPZ, N, ILO, IHI, LWORK )
384         NSR = MIN( NSR, ( N+6 ) / 9, IHI-ILO )
385         NSR = MAX( 2, NSR-MOD( NSR, 2 ) )
386*
387*        ==== Estimate optimal workspace ====
388*
389*        ==== Workspace query call to CLAQR2 ====
390*
391         CALL CLAQR2( WANTT, WANTZ, N, ILO, IHI, NWR+1, H, LDH, ILOZ,
392     $                IHIZ, Z, LDZ, LS, LD, W, H, LDH, N, H, LDH, N, H,
393     $                LDH, WORK, -1 )
394*
395*        ==== Optimal workspace = MAX(CLAQR5, CLAQR2) ====
396*
397         LWKOPT = MAX( 3*NSR / 2, INT( WORK( 1 ) ) )
398*
399*        ==== Quick return in case of workspace query. ====
400*
401         IF( LWORK.EQ.-1 ) THEN
402            WORK( 1 ) = CMPLX( LWKOPT, 0 )
403            RETURN
404         END IF
405*
406*        ==== CLAHQR/CLAQR0 crossover point ====
407*
408         NMIN = ILAENV( 12, 'CLAQR4', JBCMPZ, N, ILO, IHI, LWORK )
409         NMIN = MAX( NTINY, NMIN )
410*
411*        ==== Nibble crossover point ====
412*
413         NIBBLE = ILAENV( 14, 'CLAQR4', JBCMPZ, N, ILO, IHI, LWORK )
414         NIBBLE = MAX( 0, NIBBLE )
415*
416*        ==== Accumulate reflections during ttswp?  Use block
417*        .    2-by-2 structure during matrix-matrix multiply? ====
418*
419         KACC22 = ILAENV( 16, 'CLAQR4', JBCMPZ, N, ILO, IHI, LWORK )
420         KACC22 = MAX( 0, KACC22 )
421         KACC22 = MIN( 2, KACC22 )
422*
423*        ==== NWMAX = the largest possible deflation window for
424*        .    which there is sufficient workspace. ====
425*
426         NWMAX = MIN( ( N-1 ) / 3, LWORK / 2 )
427         NW = NWMAX
428*
429*        ==== NSMAX = the Largest number of simultaneous shifts
430*        .    for which there is sufficient workspace. ====
431*
432         NSMAX = MIN( ( N+6 ) / 9, 2*LWORK / 3 )
433         NSMAX = NSMAX - MOD( NSMAX, 2 )
434*
435*        ==== NDFL: an iteration count restarted at deflation. ====
436*
437         NDFL = 1
438*
439*        ==== ITMAX = iteration limit ====
440*
441         ITMAX = MAX( 30, 2*KEXSH )*MAX( 10, ( IHI-ILO+1 ) )
442*
443*        ==== Last row and column in the active block ====
444*
445         KBOT = IHI
446*
447*        ==== Main Loop ====
448*
449         DO 70 IT = 1, ITMAX
450*
451*           ==== Done when KBOT falls below ILO ====
452*
453            IF( KBOT.LT.ILO )
454     $         GO TO 80
455*
456*           ==== Locate active block ====
457*
458            DO 10 K = KBOT, ILO + 1, -1
459               IF( H( K, K-1 ).EQ.ZERO )
460     $            GO TO 20
461   10       CONTINUE
462            K = ILO
463   20       CONTINUE
464            KTOP = K
465*
466*           ==== Select deflation window size:
467*           .    Typical Case:
468*           .      If possible and advisable, nibble the entire
469*           .      active block.  If not, use size MIN(NWR,NWMAX)
470*           .      or MIN(NWR+1,NWMAX) depending upon which has
471*           .      the smaller corresponding subdiagonal entry
472*           .      (a heuristic).
473*           .
474*           .    Exceptional Case:
475*           .      If there have been no deflations in KEXNW or
476*           .      more iterations, then vary the deflation window
477*           .      size.   At first, because, larger windows are,
478*           .      in general, more powerful than smaller ones,
479*           .      rapidly increase the window to the maximum possible.
480*           .      Then, gradually reduce the window size. ====
481*
482            NH = KBOT - KTOP + 1
483            NWUPBD = MIN( NH, NWMAX )
484            IF( NDFL.LT.KEXNW ) THEN
485               NW = MIN( NWUPBD, NWR )
486            ELSE
487               NW = MIN( NWUPBD, 2*NW )
488            END IF
489            IF( NW.LT.NWMAX ) THEN
490               IF( NW.GE.NH-1 ) THEN
491                  NW = NH
492               ELSE
493                  KWTOP = KBOT - NW + 1
494                  IF( CABS1( H( KWTOP, KWTOP-1 ) ).GT.
495     $                CABS1( H( KWTOP-1, KWTOP-2 ) ) )NW = NW + 1
496               END IF
497            END IF
498            IF( NDFL.LT.KEXNW ) THEN
499               NDEC = -1
500            ELSE IF( NDEC.GE.0 .OR. NW.GE.NWUPBD ) THEN
501               NDEC = NDEC + 1
502               IF( NW-NDEC.LT.2 )
503     $            NDEC = 0
504               NW = NW - NDEC
505            END IF
506*
507*           ==== Aggressive early deflation:
508*           .    split workspace under the subdiagonal into
509*           .      - an nw-by-nw work array V in the lower
510*           .        left-hand-corner,
511*           .      - an NW-by-at-least-NW-but-more-is-better
512*           .        (NW-by-NHO) horizontal work array along
513*           .        the bottom edge,
514*           .      - an at-least-NW-but-more-is-better (NHV-by-NW)
515*           .        vertical work array along the left-hand-edge.
516*           .        ====
517*
518            KV = N - NW + 1
519            KT = NW + 1
520            NHO = ( N-NW-1 ) - KT + 1
521            KWV = NW + 2
522            NVE = ( N-NW ) - KWV + 1
523*
524*           ==== Aggressive early deflation ====
525*
526            CALL CLAQR2( WANTT, WANTZ, N, KTOP, KBOT, NW, H, LDH, ILOZ,
527     $                   IHIZ, Z, LDZ, LS, LD, W, H( KV, 1 ), LDH, NHO,
528     $                   H( KV, KT ), LDH, NVE, H( KWV, 1 ), LDH, WORK,
529     $                   LWORK )
530*
531*           ==== Adjust KBOT accounting for new deflations. ====
532*
533            KBOT = KBOT - LD
534*
535*           ==== KS points to the shifts. ====
536*
537            KS = KBOT - LS + 1
538*
539*           ==== Skip an expensive QR sweep if there is a (partly
540*           .    heuristic) reason to expect that many eigenvalues
541*           .    will deflate without it.  Here, the QR sweep is
542*           .    skipped if many eigenvalues have just been deflated
543*           .    or if the remaining active block is small.
544*
545            IF( ( LD.EQ.0 ) .OR. ( ( 100*LD.LE.NW*NIBBLE ) .AND. ( KBOT-
546     $          KTOP+1.GT.MIN( NMIN, NWMAX ) ) ) ) THEN
547*
548*              ==== NS = nominal number of simultaneous shifts.
549*              .    This may be lowered (slightly) if CLAQR2
550*              .    did not provide that many shifts. ====
551*
552               NS = MIN( NSMAX, NSR, MAX( 2, KBOT-KTOP ) )
553               NS = NS - MOD( NS, 2 )
554*
555*              ==== If there have been no deflations
556*              .    in a multiple of KEXSH iterations,
557*              .    then try exceptional shifts.
558*              .    Otherwise use shifts provided by
559*              .    CLAQR2 above or from the eigenvalues
560*              .    of a trailing principal submatrix. ====
561*
562               IF( MOD( NDFL, KEXSH ).EQ.0 ) THEN
563                  KS = KBOT - NS + 1
564                  DO 30 I = KBOT, KS + 1, -2
565                     W( I ) = H( I, I ) + WILK1*CABS1( H( I, I-1 ) )
566                     W( I-1 ) = W( I )
567   30             CONTINUE
568               ELSE
569*
570*                 ==== Got NS/2 or fewer shifts? Use CLAHQR
571*                 .    on a trailing principal submatrix to
572*                 .    get more. (Since NS.LE.NSMAX.LE.(N+6)/9,
573*                 .    there is enough space below the subdiagonal
574*                 .    to fit an NS-by-NS scratch array.) ====
575*
576                  IF( KBOT-KS+1.LE.NS / 2 ) THEN
577                     KS = KBOT - NS + 1
578                     KT = N - NS + 1
579                     CALL CLACPY( 'A', NS, NS, H( KS, KS ), LDH,
580     $                            H( KT, 1 ), LDH )
581                     CALL CLAHQR( .false., .false., NS, 1, NS,
582     $                            H( KT, 1 ), LDH, W( KS ), 1, 1, ZDUM,
583     $                            1, INF )
584                     KS = KS + INF
585*
586*                    ==== In case of a rare QR failure use
587*                    .    eigenvalues of the trailing 2-by-2
588*                    .    principal submatrix.  Scale to avoid
589*                    .    overflows, underflows and subnormals.
590*                    .    (The scale factor S can not be zero,
591*                    .    because H(KBOT,KBOT-1) is nonzero.) ====
592*
593                     IF( KS.GE.KBOT ) THEN
594                        S = CABS1( H( KBOT-1, KBOT-1 ) ) +
595     $                      CABS1( H( KBOT, KBOT-1 ) ) +
596     $                      CABS1( H( KBOT-1, KBOT ) ) +
597     $                      CABS1( H( KBOT, KBOT ) )
598                        AA = H( KBOT-1, KBOT-1 ) / S
599                        CC = H( KBOT, KBOT-1 ) / S
600                        BB = H( KBOT-1, KBOT ) / S
601                        DD = H( KBOT, KBOT ) / S
602                        TR2 = ( AA+DD ) / TWO
603                        DET = ( AA-TR2 )*( DD-TR2 ) - BB*CC
604                        RTDISC = SQRT( -DET )
605                        W( KBOT-1 ) = ( TR2+RTDISC )*S
606                        W( KBOT ) = ( TR2-RTDISC )*S
607*
608                        KS = KBOT - 1
609                     END IF
610                  END IF
611*
612                  IF( KBOT-KS+1.GT.NS ) THEN
613*
614*                    ==== Sort the shifts (Helps a little) ====
615*
616                     SORTED = .false.
617                     DO 50 K = KBOT, KS + 1, -1
618                        IF( SORTED )
619     $                     GO TO 60
620                        SORTED = .true.
621                        DO 40 I = KS, K - 1
622                           IF( CABS1( W( I ) ).LT.CABS1( W( I+1 ) ) )
623     $                          THEN
624                              SORTED = .false.
625                              SWAP = W( I )
626                              W( I ) = W( I+1 )
627                              W( I+1 ) = SWAP
628                           END IF
629   40                   CONTINUE
630   50                CONTINUE
631   60                CONTINUE
632                  END IF
633               END IF
634*
635*              ==== If there are only two shifts, then use
636*              .    only one.  ====
637*
638               IF( KBOT-KS+1.EQ.2 ) THEN
639                  IF( CABS1( W( KBOT )-H( KBOT, KBOT ) ).LT.
640     $                CABS1( W( KBOT-1 )-H( KBOT, KBOT ) ) ) THEN
641                     W( KBOT-1 ) = W( KBOT )
642                  ELSE
643                     W( KBOT ) = W( KBOT-1 )
644                  END IF
645               END IF
646*
647*              ==== Use up to NS of the the smallest magnatiude
648*              .    shifts.  If there aren't NS shifts available,
649*              .    then use them all, possibly dropping one to
650*              .    make the number of shifts even. ====
651*
652               NS = MIN( NS, KBOT-KS+1 )
653               NS = NS - MOD( NS, 2 )
654               KS = KBOT - NS + 1
655*
656*              ==== Small-bulge multi-shift QR sweep:
657*              .    split workspace under the subdiagonal into
658*              .    - a KDU-by-KDU work array U in the lower
659*              .      left-hand-corner,
660*              .    - a KDU-by-at-least-KDU-but-more-is-better
661*              .      (KDU-by-NHo) horizontal work array WH along
662*              .      the bottom edge,
663*              .    - and an at-least-KDU-but-more-is-better-by-KDU
664*              .      (NVE-by-KDU) vertical work WV arrow along
665*              .      the left-hand-edge. ====
666*
667               KDU = 3*NS - 3
668               KU = N - KDU + 1
669               KWH = KDU + 1
670               NHO = ( N-KDU+1-4 ) - ( KDU+1 ) + 1
671               KWV = KDU + 4
672               NVE = N - KDU - KWV + 1
673*
674*              ==== Small-bulge multi-shift QR sweep ====
675*
676               CALL CLAQR5( WANTT, WANTZ, KACC22, N, KTOP, KBOT, NS,
677     $                      W( KS ), H, LDH, ILOZ, IHIZ, Z, LDZ, WORK,
678     $                      3, H( KU, 1 ), LDH, NVE, H( KWV, 1 ), LDH,
679     $                      NHO, H( KU, KWH ), LDH )
680            END IF
681*
682*           ==== Note progress (or the lack of it). ====
683*
684            IF( LD.GT.0 ) THEN
685               NDFL = 1
686            ELSE
687               NDFL = NDFL + 1
688            END IF
689*
690*           ==== End of main loop ====
691   70    CONTINUE
692*
693*        ==== Iteration limit exceeded.  Set INFO to show where
694*        .    the problem occurred and exit. ====
695*
696         INFO = KBOT
697   80    CONTINUE
698      END IF
699*
700*     ==== Return the optimal value of LWORK. ====
701*
702      WORK( 1 ) = CMPLX( LWKOPT, 0 )
703*
704*     ==== End of CLAQR4 ====
705*
706      END
707