1       SUBROUTINE DLAQR5( WANTT, WANTZ, KACC22, N, KTOP, KBOT, NSHFTS,
  2      $                   SR, SI, H, LDH, ILOZ, IHIZ, Z, LDZ, V, LDV, U,
  3      $                   LDU, NV, WV, LDWV, NH, WH, LDWH )
  4 *
  5 *  -- LAPACK auxiliary routine (version 3.3.0) --
  6 *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
  7 *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
  8 *     November 2010
  9 *
 10 *     .. Scalar Arguments ..
 11       INTEGER            IHIZ, ILOZ, KACC22, KBOT, KTOP, LDH, LDU, LDV,
 12      $                   LDWH, LDWV, LDZ, N, NH, NSHFTS, NV
 13       LOGICAL            WANTT, WANTZ
 14 *     ..
 15 *     .. Array Arguments ..
 16       DOUBLE PRECISION   H( LDH, * ), SI( * ), SR( * ), U( LDU, * ),
 17      $                   V( LDV, * ), WH( LDWH, * ), WV( LDWV, * ),
 18      $                   Z( LDZ, * )
 19 *     ..
 20 *
 21 *     This auxiliary subroutine called by DLAQR0 performs a
 22 *     single small-bulge multi-shift QR sweep.
 23 *
 24 *      WANTT  (input) logical scalar
 25 *             WANTT = .true. if the quasi-triangular Schur factor
 26 *             is being computed.  WANTT is set to .false. otherwise.
 27 *
 28 *      WANTZ  (input) logical scalar
 29 *             WANTZ = .true. if the orthogonal Schur factor is being
 30 *             computed.  WANTZ is set to .false. otherwise.
 31 *
 32 *      KACC22 (input) integer with value 0, 1, or 2.
 33 *             Specifies the computation mode of far-from-diagonal
 34 *             orthogonal updates.
 35 *        = 0: DLAQR5 does not accumulate reflections and does not
 36 *             use matrix-matrix multiply to update far-from-diagonal
 37 *             matrix entries.
 38 *        = 1: DLAQR5 accumulates reflections and uses matrix-matrix
 39 *             multiply to update the far-from-diagonal matrix entries.
 40 *        = 2: DLAQR5 accumulates reflections, uses matrix-matrix
 41 *             multiply to update the far-from-diagonal matrix entries,
 42 *             and takes advantage of 2-by-2 block structure during
 43 *             matrix multiplies.
 44 *
 45 *      N      (input) integer scalar
 46 *             N is the order of the Hessenberg matrix H upon which this
 47 *             subroutine operates.
 48 *
 49 *      KTOP   (input) integer scalar
 50 *      KBOT   (input) integer scalar
 51 *             These are the first and last rows and columns of an
 52 *             isolated diagonal block upon which the QR sweep is to be
 53 *             applied. It is assumed without a check that
 54 *                       either KTOP = 1  or   H(KTOP,KTOP-1) = 0
 55 *             and
 56 *                       either KBOT = N  or   H(KBOT+1,KBOT) = 0.
 57 *
 58 *      NSHFTS (input) integer scalar
 59 *             NSHFTS gives the number of simultaneous shifts.  NSHFTS
 60 *             must be positive and even.
 61 *
 62 *      SR     (input/output) DOUBLE PRECISION array of size (NSHFTS)
 63 *      SI     (input/output) DOUBLE PRECISION array of size (NSHFTS)
 64 *             SR contains the real parts and SI contains the imaginary
 65 *             parts of the NSHFTS shifts of origin that define the
 66 *             multi-shift QR sweep.  On output SR and SI may be
 67 *             reordered.
 68 *
 69 *      H      (input/output) DOUBLE PRECISION array of size (LDH,N)
 70 *             On input H contains a Hessenberg matrix.  On output a
 71 *             multi-shift QR sweep with shifts SR(J)+i*SI(J) is applied
 72 *             to the isolated diagonal block in rows and columns KTOP
 73 *             through KBOT.
 74 *
 75 *      LDH    (input) integer scalar
 76 *             LDH is the leading dimension of H just as declared in the
 77 *             calling procedure.  LDH.GE.MAX(1,N).
 78 *
 79 *      ILOZ   (input) INTEGER
 80 *      IHIZ   (input) INTEGER
 81 *             Specify the rows of Z to which transformations must be
 82 *             applied if WANTZ is .TRUE.. 1 .LE. ILOZ .LE. IHIZ .LE. N
 83 *
 84 *      Z      (input/output) DOUBLE PRECISION array of size (LDZ,IHI)
 85 *             If WANTZ = .TRUE., then the QR Sweep orthogonal
 86 *             similarity transformation is accumulated into
 87 *             Z(ILOZ:IHIZ,ILO:IHI) from the right.
 88 *             If WANTZ = .FALSE., then Z is unreferenced.
 89 *
 90 *      LDZ    (input) integer scalar
 91 *             LDA is the leading dimension of Z just as declared in
 92 *             the calling procedure. LDZ.GE.N.
 93 *
 94 *      V      (workspace) DOUBLE PRECISION array of size (LDV,NSHFTS/2)
 95 *
 96 *      LDV    (input) integer scalar
 97 *             LDV is the leading dimension of V as declared in the
 98 *             calling procedure.  LDV.GE.3.
 99 *
100 *      U      (workspace) DOUBLE PRECISION array of size
101 *             (LDU,3*NSHFTS-3)
102 *
103 *      LDU    (input) integer scalar
104 *             LDU is the leading dimension of U just as declared in the
105 *             in the calling subroutine.  LDU.GE.3*NSHFTS-3.
106 *
107 *      NH     (input) integer scalar
108 *             NH is the number of columns in array WH available for
109 *             workspace. NH.GE.1.
110 *
111 *      WH     (workspace) DOUBLE PRECISION array of size (LDWH,NH)
112 *
113 *      LDWH   (input) integer scalar
114 *             Leading dimension of WH just as declared in the
115 *             calling procedure.  LDWH.GE.3*NSHFTS-3.
116 *
117 *      NV     (input) integer scalar
118 *             NV is the number of rows in WV agailable for workspace.
119 *             NV.GE.1.
120 *
121 *      WV     (workspace) DOUBLE PRECISION array of size
122 *             (LDWV,3*NSHFTS-3)
123 *
124 *      LDWV   (input) integer scalar
125 *             LDWV is the leading dimension of WV as declared in the
126 *             in the calling subroutine.  LDWV.GE.NV.
127 *
128 *     ================================================================
129 *     Based on contributions by
130 *        Karen Braman and Ralph Byers, Department of Mathematics,
131 *        University of Kansas, USA
132 *
133 *     ================================================================
134 *     Reference:
135 *
136 *     K. Braman, R. Byers and R. Mathias, The Multi-Shift QR
137 *     Algorithm Part I: Maintaining Well Focused Shifts, and
138 *     Level 3 Performance, SIAM Journal of Matrix Analysis,
139 *     volume 23, pages 929--947, 2002.
140 *
141 *     ================================================================
142 *     .. Parameters ..
143       DOUBLE PRECISION   ZERO, ONE
144       PARAMETER          ( ZERO = 0.0d0, ONE = 1.0d0 )
145 *     ..
146 *     .. Local Scalars ..
147       DOUBLE PRECISION   ALPHA, BETA, H11, H12, H21, H22, REFSUM,
148      $                   SAFMAX, SAFMIN, SCL, SMLNUM, SWAP, TST1, TST2,
149      $                   ULP
150       INTEGER            I, I2, I4, INCOL, J, J2, J4, JBOT, JCOL, JLEN,
151      $                   JROW, JTOP, K, K1, KDU, KMS, KNZ, KRCOL, KZS,
152      $                   M, M22, MBOT, MEND, MSTART, MTOP, NBMPS, NDCOL,
153      $                   NS, NU
154       LOGICAL            ACCUM, BLK22, BMP22
155 *     ..
156 *     .. External Functions ..
157       DOUBLE PRECISION   DLAMCH
158       EXTERNAL           DLAMCH
159 *     ..
160 *     .. Intrinsic Functions ..
161 *
162       INTRINSIC          ABSDBLEMAXMINMOD
163 *     ..
164 *     .. Local Arrays ..
165       DOUBLE PRECISION   VT( 3 )
166 *     ..
167 *     .. External Subroutines ..
168       EXTERNAL           DGEMM, DLABAD, DLACPY, DLAQR1, DLARFG, DLASET,
169      $                   DTRMM
170 *     ..
171 *     .. Executable Statements ..
172 *
173 *     ==== If there are no shifts, then there is nothing to do. ====
174 *
175       IF( NSHFTS.LT.2 )
176      $   RETURN
177 *
178 *     ==== If the active block is empty or 1-by-1, then there
179 *     .    is nothing to do. ====
180 *
181       IF( KTOP.GE.KBOT )
182      $   RETURN
183 *
184 *     ==== Shuffle shifts into pairs of real shifts and pairs
185 *     .    of complex conjugate shifts assuming complex
186 *     .    conjugate shifts are already adjacent to one
187 *     .    another. ====
188 *
189       DO 10 I = 1, NSHFTS - 22
190          IF( SI( I ).NE.-SI( I+1 ) ) THEN
191 *
192             SWAP = SR( I )
193             SR( I ) = SR( I+1 )
194             SR( I+1 ) = SR( I+2 )
195             SR( I+2 ) = SWAP
196 *
197             SWAP = SI( I )
198             SI( I ) = SI( I+1 )
199             SI( I+1 ) = SI( I+2 )
200             SI( I+2 ) = SWAP
201          END IF
202    10 CONTINUE
203 *
204 *     ==== NSHFTS is supposed to be even, but if it is odd,
205 *     .    then simply reduce it by one.  The shuffle above
206 *     .    ensures that the dropped shift is real and that
207 *     .    the remaining shifts are paired. ====
208 *
209       NS = NSHFTS - MOD( NSHFTS, 2 )
210 *
211 *     ==== Machine constants for deflation ====
212 *
213       SAFMIN = DLAMCH( 'SAFE MINIMUM' )
214       SAFMAX = ONE / SAFMIN
215       CALL DLABAD( SAFMIN, SAFMAX )
216       ULP = DLAMCH( 'PRECISION' )
217       SMLNUM = SAFMIN*DBLE( N ) / ULP )
218 *
219 *     ==== Use accumulated reflections to update far-from-diagonal
220 *     .    entries ? ====
221 *
222       ACCUM = ( KACC22.EQ.1 ) .OR. ( KACC22.EQ.2 )
223 *
224 *     ==== If so, exploit the 2-by-2 block structure? ====
225 *
226       BLK22 = ( NS.GT.2 ) .AND. ( KACC22.EQ.2 )
227 *
228 *     ==== clear trash ====
229 *
230       IF( KTOP+2.LE.KBOT )
231      $   H( KTOP+2, KTOP ) = ZERO
232 *
233 *     ==== NBMPS = number of 2-shift bulges in the chain ====
234 *
235       NBMPS = NS / 2
236 *
237 *     ==== KDU = width of slab ====
238 *
239       KDU = 6*NBMPS - 3
240 *
241 *     ==== Create and chase chains of NBMPS bulges ====
242 *
243       DO 220 INCOL = 3*1-NBMPS ) + KTOP - 1, KBOT - 23*NBMPS - 2
244          NDCOL = INCOL + KDU
245          IF( ACCUM )
246      $      CALL DLASET( 'ALL', KDU, KDU, ZERO, ONE, U, LDU )
247 *
248 *        ==== Near-the-diagonal bulge chase.  The following loop
249 *        .    performs the near-the-diagonal part of a small bulge
250 *        .    multi-shift QR sweep.  Each 6*NBMPS-2 column diagonal
251 *        .    chunk extends from column INCOL to column NDCOL
252 *        .    (including both column INCOL and column NDCOL). The
253 *        .    following loop chases a 3*NBMPS column long chain of
254 *        .    NBMPS bulges 3*NBMPS-2 columns to the right.  (INCOL
255 *        .    may be less than KTOP and and NDCOL may be greater than
256 *        .    KBOT indicating phantom columns from which to chase
257 *        .    bulges before they are actually introduced or to which
258 *        .    to chase bulges beyond column KBOT.)  ====
259 *
260          DO 150 KRCOL = INCOL, MIN( INCOL+3*NBMPS-3, KBOT-2 )
261 *
262 *           ==== Bulges number MTOP to MBOT are active double implicit
263 *           .    shift bulges.  There may or may not also be small
264 *           .    2-by-2 bulge, if there is room.  The inactive bulges
265 *           .    (if any) must wait until the active bulges have moved
266 *           .    down the diagonal to make room.  The phantom matrix
267 *           .    paradigm described above helps keep track.  ====
268 *
269             MTOP = MAX1, ( ( KTOP-1 )-KRCOL+2 ) / 3+1 )
270             MBOT = MIN( NBMPS, ( KBOT-KRCOL ) / 3 )
271             M22 = MBOT + 1
272             BMP22 = ( MBOT.LT.NBMPS ) .AND. ( KRCOL+3*( M22-1 ) ).EQ.
273      $              ( KBOT-2 )
274 *
275 *           ==== Generate reflections to chase the chain right
276 *           .    one column.  (The minimum value of K is KTOP-1.) ====
277 *
278             DO 20 M = MTOP, MBOT
279                K = KRCOL + 3*( M-1 )
280                IF( K.EQ.KTOP-1 ) THEN
281                   CALL DLAQR1( 3, H( KTOP, KTOP ), LDH, SR( 2*M-1 ),
282      $                         SI( 2*M-1 ), SR( 2*M ), SI( 2*M ),
283      $                         V( 1, M ) )
284                   ALPHA = V( 1, M )
285                   CALL DLARFG( 3, ALPHA, V( 2, M ), 1, V( 1, M ) )
286                ELSE
287                   BETA = H( K+1, K )
288                   V( 2, M ) = H( K+2, K )
289                   V( 3, M ) = H( K+3, K )
290                   CALL DLARFG( 3, BETA, V( 2, M ), 1, V( 1, M ) )
291 *
292 *                 ==== A Bulge may collapse because of vigilant
293 *                 .    deflation or destructive underflow.  In the
294 *                 .    underflow case, try the two-small-subdiagonals
295 *                 .    trick to try to reinflate the bulge.  ====
296 *
297                   IF( H( K+3, K ).NE.ZERO .OR. H( K+3, K+1 ).NE.
298      $                ZERO .OR. H( K+3, K+2 ).EQ.ZERO ) THEN
299 *
300 *                    ==== Typical case: not collapsed (yet). ====
301 *
302                      H( K+1, K ) = BETA
303                      H( K+2, K ) = ZERO
304                      H( K+3, K ) = ZERO
305                   ELSE
306 *
307 *                    ==== Atypical case: collapsed.  Attempt to
308 *                    .    reintroduce ignoring H(K+1,K) and H(K+2,K).
309 *                    .    If the fill resulting from the new
310 *                    .    reflector is too large, then abandon it.
311 *                    .    Otherwise, use the new one. ====
312 *
313                      CALL DLAQR1( 3, H( K+1, K+1 ), LDH, SR( 2*M-1 ),
314      $                            SI( 2*M-1 ), SR( 2*M ), SI( 2*M ),
315      $                            VT )
316                      ALPHA = VT( 1 )
317                      CALL DLARFG( 3, ALPHA, VT( 2 ), 1, VT( 1 ) )
318                      REFSUM = VT( 1 )*( H( K+1, K )+VT( 2 )*
319      $                        H( K+2, K ) )
320 *
321                      IFABS( H( K+2, K )-REFSUM*VT( 2 ) )+
322      $                   ABS( REFSUM*VT( 3 ) ).GT.ULP*
323      $                   ( ABS( H( K, K ) )+ABS( H( K+1,
324      $                   K+1 ) )+ABS( H( K+2, K+2 ) ) ) ) THEN
325 *
326 *                       ==== Starting a new bulge here would
327 *                       .    create non-negligible fill.  Use
328 *                       .    the old one with trepidation. ====
329 *
330                         H( K+1, K ) = BETA
331                         H( K+2, K ) = ZERO
332                         H( K+3, K ) = ZERO
333                      ELSE
334 *
335 *                       ==== Stating a new bulge here would
336 *                       .    create only negligible fill.
337 *                       .    Replace the old reflector with
338 *                       .    the new one. ====
339 *
340                         H( K+1, K ) = H( K+1, K ) - REFSUM
341                         H( K+2, K ) = ZERO
342                         H( K+3, K ) = ZERO
343                         V( 1, M ) = VT( 1 )
344                         V( 2, M ) = VT( 2 )
345                         V( 3, M ) = VT( 3 )
346                      END IF
347                   END IF
348                END IF
349    20       CONTINUE
350 *
351 *           ==== Generate a 2-by-2 reflection, if needed. ====
352 *
353             K = KRCOL + 3*( M22-1 )
354             IF( BMP22 ) THEN
355                IF( K.EQ.KTOP-1 ) THEN
356                   CALL DLAQR1( 2, H( K+1, K+1 ), LDH, SR( 2*M22-1 ),
357      $                         SI( 2*M22-1 ), SR( 2*M22 ), SI( 2*M22 ),
358      $                         V( 1, M22 ) )
359                   BETA = V( 1, M22 )
360                   CALL DLARFG( 2, BETA, V( 2, M22 ), 1, V( 1, M22 ) )
361                ELSE
362                   BETA = H( K+1, K )
363                   V( 2, M22 ) = H( K+2, K )
364                   CALL DLARFG( 2, BETA, V( 2, M22 ), 1, V( 1, M22 ) )
365                   H( K+1, K ) = BETA
366                   H( K+2, K ) = ZERO
367                END IF
368             END IF
369 *
370 *           ==== Multiply H by reflections from the left ====
371 *
372             IF( ACCUM ) THEN
373                JBOT = MIN( NDCOL, KBOT )
374             ELSE IF( WANTT ) THEN
375                JBOT = N
376             ELSE
377                JBOT = KBOT
378             END IF
379             DO 40 J = MAX( KTOP, KRCOL ), JBOT
380                MEND = MIN( MBOT, ( J-KRCOL+2 ) / 3 )
381                DO 30 M = MTOP, MEND
382                   K = KRCOL + 3*( M-1 )
383                   REFSUM = V( 1, M )*( H( K+1, J )+V( 2, M )*
384      $                     H( K+2, J )+V( 3, M )*H( K+3, J ) )
385                   H( K+1, J ) = H( K+1, J ) - REFSUM
386                   H( K+2, J ) = H( K+2, J ) - REFSUM*V( 2, M )
387                   H( K+3, J ) = H( K+3, J ) - REFSUM*V( 3, M )
388    30          CONTINUE
389    40       CONTINUE
390             IF( BMP22 ) THEN
391                K = KRCOL + 3*( M22-1 )
392                DO 50 J = MAX( K+1, KTOP ), JBOT
393                   REFSUM = V( 1, M22 )*( H( K+1, J )+V( 2, M22 )*
394      $                     H( K+2, J ) )
395                   H( K+1, J ) = H( K+1, J ) - REFSUM
396                   H( K+2, J ) = H( K+2, J ) - REFSUM*V( 2, M22 )
397    50          CONTINUE
398             END IF
399 *
400 *           ==== Multiply H by reflections from the right.
401 *           .    Delay filling in the last row until the
402 *           .    vigilant deflation check is complete. ====
403 *
404             IF( ACCUM ) THEN
405                JTOP = MAX( KTOP, INCOL )
406             ELSE IF( WANTT ) THEN
407                JTOP = 1
408             ELSE
409                JTOP = KTOP
410             END IF
411             DO 90 M = MTOP, MBOT
412                IF( V( 1, M ).NE.ZERO ) THEN
413                   K = KRCOL + 3*( M-1 )
414                   DO 60 J = JTOP, MIN( KBOT, K+3 )
415                      REFSUM = V( 1, M )*( H( J, K+1 )+V( 2, M )*
416      $                        H( J, K+2 )+V( 3, M )*H( J, K+3 ) )
417                      H( J, K+1 ) = H( J, K+1 ) - REFSUM
418                      H( J, K+2 ) = H( J, K+2 ) - REFSUM*V( 2, M )
419                      H( J, K+3 ) = H( J, K+3 ) - REFSUM*V( 3, M )
420    60             CONTINUE
421 *
422                   IF( ACCUM ) THEN
423 *
424 *                    ==== Accumulate U. (If necessary, update Z later
425 *                    .    with with an efficient matrix-matrix
426 *                    .    multiply.) ====
427 *
428                      KMS = K - INCOL
429                      DO 70 J = MAX1, KTOP-INCOL ), KDU
430                         REFSUM = V( 1, M )*( U( J, KMS+1 )+V( 2, M )*
431      $                           U( J, KMS+2 )+V( 3, M )*U( J, KMS+3 ) )
432                         U( J, KMS+1 ) = U( J, KMS+1 ) - REFSUM
433                         U( J, KMS+2 ) = U( J, KMS+2 ) - REFSUM*V( 2, M )
434                         U( J, KMS+3 ) = U( J, KMS+3 ) - REFSUM*V( 3, M )
435    70                CONTINUE
436                   ELSE IF( WANTZ ) THEN
437 *
438 *                    ==== U is not accumulated, so update Z
439 *                    .    now by multiplying by reflections
440 *                    .    from the right. ====
441 *
442                      DO 80 J = ILOZ, IHIZ
443                         REFSUM = V( 1, M )*( Z( J, K+1 )+V( 2, M )*
444      $                           Z( J, K+2 )+V( 3, M )*Z( J, K+3 ) )
445                         Z( J, K+1 ) = Z( J, K+1 ) - REFSUM
446                         Z( J, K+2 ) = Z( J, K+2 ) - REFSUM*V( 2, M )
447                         Z( J, K+3 ) = Z( J, K+3 ) - REFSUM*V( 3, M )
448    80                CONTINUE
449                   END IF
450                END IF
451    90       CONTINUE
452 *
453 *           ==== Special case: 2-by-2 reflection (if needed) ====
454 *
455             K = KRCOL + 3*( M22-1 )
456             IF( BMP22 ) THEN
457                IF ( V( 1, M22 ).NE.ZERO ) THEN
458                   DO 100 J = JTOP, MIN( KBOT, K+3 )
459                      REFSUM = V( 1, M22 )*( H( J, K+1 )+V( 2, M22 )*
460      $                        H( J, K+2 ) )
461                      H( J, K+1 ) = H( J, K+1 ) - REFSUM
462                      H( J, K+2 ) = H( J, K+2 ) - REFSUM*V( 2, M22 )
463   100             CONTINUE
464 *
465                   IF( ACCUM ) THEN
466                      KMS = K - INCOL
467                      DO 110 J = MAX1, KTOP-INCOL ), KDU
468                         REFSUM = V( 1, M22 )*( U( J, KMS+1 )+
469      $                           V( 2, M22 )*U( J, KMS+2 ) )
470                         U( J, KMS+1 ) = U( J, KMS+1 ) - REFSUM
471                         U( J, KMS+2 ) = U( J, KMS+2 ) -
472      $                                  REFSUM*V( 2, M22 )
473   110             CONTINUE
474                   ELSE IF( WANTZ ) THEN
475                      DO 120 J = ILOZ, IHIZ
476                         REFSUM = V( 1, M22 )*( Z( J, K+1 )+V( 2, M22 )*
477      $                           Z( J, K+2 ) )
478                         Z( J, K+1 ) = Z( J, K+1 ) - REFSUM
479                         Z( J, K+2 ) = Z( J, K+2 ) - REFSUM*V( 2, M22 )
480   120                CONTINUE
481                   END IF
482                END IF
483             END IF
484 *
485 *           ==== Vigilant deflation check ====
486 *
487             MSTART = MTOP
488             IF( KRCOL+3*( MSTART-1 ).LT.KTOP )
489      $         MSTART = MSTART + 1
490             MEND = MBOT
491             IF( BMP22 )
492      $         MEND = MEND + 1
493             IF( KRCOL.EQ.KBOT-2 )
494      $         MEND = MEND + 1
495             DO 130 M = MSTART, MEND
496                K = MIN( KBOT-1, KRCOL+3*( M-1 ) )
497 *
498 *              ==== The following convergence test requires that
499 *              .    the tradition small-compared-to-nearby-diagonals
500 *              .    criterion and the Ahues & Tisseur (LAWN 122, 1997)
501 *              .    criteria both be satisfied.  The latter improves
502 *              .    accuracy in some examples. Falling back on an
503 *              .    alternate convergence criterion when TST1 or TST2
504 *              .    is zero (as done here) is traditional but probably
505 *              .    unnecessary. ====
506 *
507                IF( H( K+1, K ).NE.ZERO ) THEN
508                   TST1 = ABS( H( K, K ) ) + ABS( H( K+1, K+1 ) )
509                   IF( TST1.EQ.ZERO ) THEN
510                      IF( K.GE.KTOP+1 )
511      $                  TST1 = TST1 + ABS( H( K, K-1 ) )
512                      IF( K.GE.KTOP+2 )
513      $                  TST1 = TST1 + ABS( H( K, K-2 ) )
514                      IF( K.GE.KTOP+3 )
515      $                  TST1 = TST1 + ABS( H( K, K-3 ) )
516                      IF( K.LE.KBOT-2 )
517      $                  TST1 = TST1 + ABS( H( K+2, K+1 ) )
518                      IF( K.LE.KBOT-3 )
519      $                  TST1 = TST1 + ABS( H( K+3, K+1 ) )
520                      IF( K.LE.KBOT-4 )
521      $                  TST1 = TST1 + ABS( H( K+4, K+1 ) )
522                   END IF
523                   IFABS( H( K+1, K ) ).LE.MAX( SMLNUM, ULP*TST1 ) )
524      $                 THEN
525                      H12 = MAXABS( H( K+1, K ) ), ABS( H( K, K+1 ) ) )
526                      H21 = MINABS( H( K+1, K ) ), ABS( H( K, K+1 ) ) )
527                      H11 = MAXABS( H( K+1, K+1 ) ),
528      $                     ABS( H( K, K )-H( K+1, K+1 ) ) )
529                      H22 = MINABS( H( K+1, K+1 ) ),
530      $                     ABS( H( K, K )-H( K+1, K+1 ) ) )
531                      SCL = H11 + H12
532                      TST2 = H22*( H11 / SCL )
533 *
534                      IF( TST2.EQ.ZERO .OR. H21*( H12 / SCL ).LE.
535      $                   MAX( SMLNUM, ULP*TST2 ) )H( K+1, K ) = ZERO
536                   END IF
537                END IF
538   130       CONTINUE
539 *
540 *           ==== Fill in the last row of each bulge. ====
541 *
542             MEND = MIN( NBMPS, ( KBOT-KRCOL-1 ) / 3 )
543             DO 140 M = MTOP, MEND
544                K = KRCOL + 3*( M-1 )
545                REFSUM = V( 1, M )*V( 3, M )*H( K+4, K+3 )
546                H( K+4, K+1 ) = -REFSUM
547                H( K+4, K+2 ) = -REFSUM*V( 2, M )
548                H( K+4, K+3 ) = H( K+4, K+3 ) - REFSUM*V( 3, M )
549   140       CONTINUE
550 *
551 *           ==== End of near-the-diagonal bulge chase. ====
552 *
553   150    CONTINUE
554 *
555 *        ==== Use U (if accumulated) to update far-from-diagonal
556 *        .    entries in H.  If required, use U to update Z as
557 *        .    well. ====
558 *
559          IF( ACCUM ) THEN
560             IF( WANTT ) THEN
561                JTOP = 1
562                JBOT = N
563             ELSE
564                JTOP = KTOP
565                JBOT = KBOT
566             END IF
567             IF( ( .NOT.BLK22 ) .OR. ( INCOL.LT.KTOP ) .OR.
568      $          ( NDCOL.GT.KBOT ) .OR. ( NS.LE.2 ) ) THEN
569 *
570 *              ==== Updates not exploiting the 2-by-2 block
571 *              .    structure of U.  K1 and NU keep track of
572 *              .    the location and size of U in the special
573 *              .    cases of introducing bulges and chasing
574 *              .    bulges off the bottom.  In these special
575 *              .    cases and in case the number of shifts
576 *              .    is NS = 2, there is no 2-by-2 block
577 *              .    structure to exploit.  ====
578 *
579                K1 = MAX1, KTOP-INCOL )
580                NU = ( KDU-MAX0, NDCOL-KBOT ) ) - K1 + 1
581 *
582 *              ==== Horizontal Multiply ====
583 *
584                DO 160 JCOL = MIN( NDCOL, KBOT ) + 1, JBOT, NH
585                   JLEN = MIN( NH, JBOT-JCOL+1 )
586                   CALL DGEMM( 'C''N', NU, JLEN, NU, ONE, U( K1, K1 ),
587      $                        LDU, H( INCOL+K1, JCOL ), LDH, ZERO, WH,
588      $                        LDWH )
589                   CALL DLACPY( 'ALL', NU, JLEN, WH, LDWH,
590      $                         H( INCOL+K1, JCOL ), LDH )
591   160          CONTINUE
592 *
593 *              ==== Vertical multiply ====
594 *
595                DO 170 JROW = JTOP, MAX( KTOP, INCOL ) - 1, NV
596                   JLEN = MIN( NV, MAX( KTOP, INCOL )-JROW )
597                   CALL DGEMM( 'N''N', JLEN, NU, NU, ONE,
598      $                        H( JROW, INCOL+K1 ), LDH, U( K1, K1 ),
599      $                        LDU, ZERO, WV, LDWV )
600                   CALL DLACPY( 'ALL', JLEN, NU, WV, LDWV,
601      $                         H( JROW, INCOL+K1 ), LDH )
602   170          CONTINUE
603 *
604 *              ==== Z multiply (also vertical) ====
605 *
606                IF( WANTZ ) THEN
607                   DO 180 JROW = ILOZ, IHIZ, NV
608                      JLEN = MIN( NV, IHIZ-JROW+1 )
609                      CALL DGEMM( 'N''N', JLEN, NU, NU, ONE,
610      $                           Z( JROW, INCOL+K1 ), LDZ, U( K1, K1 ),
611      $                           LDU, ZERO, WV, LDWV )
612                      CALL DLACPY( 'ALL', JLEN, NU, WV, LDWV,
613      $                            Z( JROW, INCOL+K1 ), LDZ )
614   180             CONTINUE
615                END IF
616             ELSE
617 *
618 *              ==== Updates exploiting U's 2-by-2 block structure.
619 *              .    (I2, I4, J2, J4 are the last rows and columns
620 *              .    of the blocks.) ====
621 *
622                I2 = ( KDU+1 ) / 2
623                I4 = KDU
624                J2 = I4 - I2
625                J4 = KDU
626 *
627 *              ==== KZS and KNZ deal with the band of zeros
628 *              .    along the diagonal of one of the triangular
629 *              .    blocks. ====
630 *
631                KZS = ( J4-J2 ) - ( NS+1 )
632                KNZ = NS + 1
633 *
634 *              ==== Horizontal multiply ====
635 *
636                DO 190 JCOL = MIN( NDCOL, KBOT ) + 1, JBOT, NH
637                   JLEN = MIN( NH, JBOT-JCOL+1 )
638 *
639 *                 ==== Copy bottom of H to top+KZS of scratch ====
640 *                  (The first KZS rows get multiplied by zero.) ====
641 *
642                   CALL DLACPY( 'ALL', KNZ, JLEN, H( INCOL+1+J2, JCOL ),
643      $                         LDH, WH( KZS+11 ), LDWH )
644 *
645 *                 ==== Multiply by U21**T ====
646 *
647                   CALL DLASET( 'ALL', KZS, JLEN, ZERO, ZERO, WH, LDWH )
648                   CALL DTRMM( 'L''U''C''N', KNZ, JLEN, ONE,
649      $                        U( J2+11+KZS ), LDU, WH( KZS+11 ),
650      $                        LDWH )
651 *
652 *                 ==== Multiply top of H by U11**T ====
653 *
654                   CALL DGEMM( 'C''N', I2, JLEN, J2, ONE, U, LDU,
655      $                        H( INCOL+1, JCOL ), LDH, ONE, WH, LDWH )
656 *
657 *                 ==== Copy top of H to bottom of WH ====
658 *
659                   CALL DLACPY( 'ALL', J2, JLEN, H( INCOL+1, JCOL ), LDH,
660      $                         WH( I2+11 ), LDWH )
661 *
662 *                 ==== Multiply by U21**T ====
663 *
664                   CALL DTRMM( 'L''L''C''N', J2, JLEN, ONE,
665      $                        U( 1, I2+1 ), LDU, WH( I2+11 ), LDWH )
666 *
667 *                 ==== Multiply by U22 ====
668 *
669                   CALL DGEMM( 'C''N', I4-I2, JLEN, J4-J2, ONE,
670      $                        U( J2+1, I2+1 ), LDU,
671      $                        H( INCOL+1+J2, JCOL ), LDH, ONE,
672      $                        WH( I2+11 ), LDWH )
673 *
674 *                 ==== Copy it back ====
675 *
676                   CALL DLACPY( 'ALL', KDU, JLEN, WH, LDWH,
677      $                         H( INCOL+1, JCOL ), LDH )
678   190          CONTINUE
679 *
680 *              ==== Vertical multiply ====
681 *
682                DO 200 JROW = JTOP, MAX( INCOL, KTOP ) - 1, NV
683                   JLEN = MIN( NV, MAX( INCOL, KTOP )-JROW )
684 *
685 *                 ==== Copy right of H to scratch (the first KZS
686 *                 .    columns get multiplied by zero) ====
687 *
688                   CALL DLACPY( 'ALL', JLEN, KNZ, H( JROW, INCOL+1+J2 ),
689      $                         LDH, WV( 11+KZS ), LDWV )
690 *
691 *                 ==== Multiply by U21 ====
692 *
693                   CALL DLASET( 'ALL', JLEN, KZS, ZERO, ZERO, WV, LDWV )
694                   CALL DTRMM( 'R''U''N''N', JLEN, KNZ, ONE,
695      $                        U( J2+11+KZS ), LDU, WV( 11+KZS ),
696      $                        LDWV )
697 *
698 *                 ==== Multiply by U11 ====
699 *
700                   CALL DGEMM( 'N''N', JLEN, I2, J2, ONE,
701      $                        H( JROW, INCOL+1 ), LDH, U, LDU, ONE, WV,
702      $                        LDWV )
703 *
704 *                 ==== Copy left of H to right of scratch ====
705 *
706                   CALL DLACPY( 'ALL', JLEN, J2, H( JROW, INCOL+1 ), LDH,
707      $                         WV( 11+I2 ), LDWV )
708 *
709 *                 ==== Multiply by U21 ====
710 *
711                   CALL DTRMM( 'R''L''N''N', JLEN, I4-I2, ONE,
712      $                        U( 1, I2+1 ), LDU, WV( 11+I2 ), LDWV )
713 *
714 *                 ==== Multiply by U22 ====
715 *
716                   CALL DGEMM( 'N''N', JLEN, I4-I2, J4-J2, ONE,
717      $                        H( JROW, INCOL+1+J2 ), LDH,
718      $                        U( J2+1, I2+1 ), LDU, ONE, WV( 11+I2 ),
719      $                        LDWV )
720 *
721 *                 ==== Copy it back ====
722 *
723                   CALL DLACPY( 'ALL', JLEN, KDU, WV, LDWV,
724      $                         H( JROW, INCOL+1 ), LDH )
725   200          CONTINUE
726 *
727 *              ==== Multiply Z (also vertical) ====
728 *
729                IF( WANTZ ) THEN
730                   DO 210 JROW = ILOZ, IHIZ, NV
731                      JLEN = MIN( NV, IHIZ-JROW+1 )
732 *
733 *                    ==== Copy right of Z to left of scratch (first
734 *                    .     KZS columns get multiplied by zero) ====
735 *
736                      CALL DLACPY( 'ALL', JLEN, KNZ,
737      $                            Z( JROW, INCOL+1+J2 ), LDZ,
738      $                            WV( 11+KZS ), LDWV )
739 *
740 *                    ==== Multiply by U12 ====
741 *
742                      CALL DLASET( 'ALL', JLEN, KZS, ZERO, ZERO, WV,
743      $                            LDWV )
744                      CALL DTRMM( 'R''U''N''N', JLEN, KNZ, ONE,
745      $                           U( J2+11+KZS ), LDU, WV( 11+KZS ),
746      $                           LDWV )
747 *
748 *                    ==== Multiply by U11 ====
749 *
750                      CALL DGEMM( 'N''N', JLEN, I2, J2, ONE,
751      $                           Z( JROW, INCOL+1 ), LDZ, U, LDU, ONE,
752      $                           WV, LDWV )
753 *
754 *                    ==== Copy left of Z to right of scratch ====
755 *
756                      CALL DLACPY( 'ALL', JLEN, J2, Z( JROW, INCOL+1 ),
757      $                            LDZ, WV( 11+I2 ), LDWV )
758 *
759 *                    ==== Multiply by U21 ====
760 *
761                      CALL DTRMM( 'R''L''N''N', JLEN, I4-I2, ONE,
762      $                           U( 1, I2+1 ), LDU, WV( 11+I2 ),
763      $                           LDWV )
764 *
765 *                    ==== Multiply by U22 ====
766 *
767                      CALL DGEMM( 'N''N', JLEN, I4-I2, J4-J2, ONE,
768      $                           Z( JROW, INCOL+1+J2 ), LDZ,
769      $                           U( J2+1, I2+1 ), LDU, ONE,
770      $                           WV( 11+I2 ), LDWV )
771 *
772 *                    ==== Copy the result back to Z ====
773 *
774                      CALL DLACPY( 'ALL', JLEN, KDU, WV, LDWV,
775      $                            Z( JROW, INCOL+1 ), LDZ )
776   210             CONTINUE
777                END IF
778             END IF
779          END IF
780   220 CONTINUE
781 *
782 *     ==== End of DLAQR5 ====
783 *
784       END