Annotation of rpl/lapack/lapack/zlaqr0.f, revision 1.8
1.8 ! bertrand 1: *> \brief \b ZLAQR0
! 2: *
! 3: * =========== DOCUMENTATION ===========
! 4: *
! 5: * Online html documentation available at
! 6: * http://www.netlib.org/lapack/explore-html/
! 7: *
! 8: *> \htmlonly
! 9: *> Download ZLAQR0 + dependencies
! 10: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/zlaqr0.f">
! 11: *> [TGZ]</a>
! 12: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/zlaqr0.f">
! 13: *> [ZIP]</a>
! 14: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/zlaqr0.f">
! 15: *> [TXT]</a>
! 16: *> \endhtmlonly
! 17: *
! 18: * Definition:
! 19: * ===========
! 20: *
! 21: * SUBROUTINE ZLAQR0( 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*16 H( LDH, * ), W( * ), WORK( * ), Z( LDZ, * )
! 30: * ..
! 31: *
! 32: *
! 33: *> \par Purpose:
! 34: * =============
! 35: *>
! 36: *> \verbatim
! 37: *>
! 38: *> ZLAQR0 computes the eigenvalues of a Hessenberg matrix H
! 39: *> and, optionally, the matrices T and Z from the Schur decomposition
! 40: *> H = Z T Z**H, where T is an upper triangular matrix (the
! 41: *> Schur form), and Z is the unitary matrix of Schur vectors.
! 42: *>
! 43: *> Optionally Z may be postmultiplied into an input unitary
! 44: *> matrix Q so that this routine can give the Schur factorization
! 45: *> of a matrix A which has been reduced to the Hessenberg form H
! 46: *> by the unitary matrix Q: A = Q*H*Q**H = (QZ)*H*(QZ)**H.
! 47: *> \endverbatim
! 48: *
! 49: * Arguments:
! 50: * ==========
! 51: *
! 52: *> \param[in] WANTT
! 53: *> \verbatim
! 54: *> WANTT is LOGICAL
! 55: *> = .TRUE. : the full Schur form T is required;
! 56: *> = .FALSE.: only eigenvalues are required.
! 57: *> \endverbatim
! 58: *>
! 59: *> \param[in] WANTZ
! 60: *> \verbatim
! 61: *> WANTZ is LOGICAL
! 62: *> = .TRUE. : the matrix of Schur vectors Z is required;
! 63: *> = .FALSE.: Schur vectors are not required.
! 64: *> \endverbatim
! 65: *>
! 66: *> \param[in] N
! 67: *> \verbatim
! 68: *> N is INTEGER
! 69: *> The order of the matrix H. N .GE. 0.
! 70: *> \endverbatim
! 71: *>
! 72: *> \param[in] ILO
! 73: *> \verbatim
! 74: *> ILO is INTEGER
! 75: *> \endverbatim
! 76: *>
! 77: *> \param[in] IHI
! 78: *> \verbatim
! 79: *> IHI is INTEGER
! 80: *>
! 81: *> It is assumed that H is already upper triangular in rows
! 82: *> and columns 1:ILO-1 and IHI+1:N and, if ILO.GT.1,
! 83: *> H(ILO,ILO-1) is zero. ILO and IHI are normally set by a
! 84: *> previous call to ZGEBAL, and then passed to ZGEHRD when the
! 85: *> matrix output by ZGEBAL is reduced to Hessenberg form.
! 86: *> Otherwise, ILO and IHI should be set to 1 and N,
! 87: *> respectively. If N.GT.0, then 1.LE.ILO.LE.IHI.LE.N.
! 88: *> If N = 0, then ILO = 1 and IHI = 0.
! 89: *> \endverbatim
! 90: *>
! 91: *> \param[in,out] H
! 92: *> \verbatim
! 93: *> H is COMPLEX*16 array, dimension (LDH,N)
! 94: *> On entry, the upper Hessenberg matrix H.
! 95: *> On exit, if INFO = 0 and WANTT is .TRUE., then H
! 96: *> contains the upper triangular matrix T from the Schur
! 97: *> decomposition (the Schur form). If INFO = 0 and WANT is
! 98: *> .FALSE., then the contents of H are unspecified on exit.
! 99: *> (The output value of H when INFO.GT.0 is given under the
! 100: *> description of INFO below.)
! 101: *>
! 102: *> This subroutine may explicitly set H(i,j) = 0 for i.GT.j and
! 103: *> j = 1, 2, ... ILO-1 or j = IHI+1, IHI+2, ... N.
! 104: *> \endverbatim
! 105: *>
! 106: *> \param[in] LDH
! 107: *> \verbatim
! 108: *> LDH is INTEGER
! 109: *> The leading dimension of the array H. LDH .GE. max(1,N).
! 110: *> \endverbatim
! 111: *>
! 112: *> \param[out] W
! 113: *> \verbatim
! 114: *> W is COMPLEX*16 array, dimension (N)
! 115: *> The computed eigenvalues of H(ILO:IHI,ILO:IHI) are stored
! 116: *> in W(ILO:IHI). If WANTT is .TRUE., then the eigenvalues are
! 117: *> stored in the same order as on the diagonal of the Schur
! 118: *> form returned in H, with W(i) = H(i,i).
! 119: *> \endverbatim
! 120: *>
! 121: *> \param[in] ILOZ
! 122: *> \verbatim
! 123: *> ILOZ is INTEGER
! 124: *> \endverbatim
! 125: *>
! 126: *> \param[in] IHIZ
! 127: *> \verbatim
! 128: *> IHIZ is INTEGER
! 129: *> Specify the rows of Z to which transformations must be
! 130: *> applied if WANTZ is .TRUE..
! 131: *> 1 .LE. ILOZ .LE. ILO; IHI .LE. IHIZ .LE. N.
! 132: *> \endverbatim
! 133: *>
! 134: *> \param[in,out] Z
! 135: *> \verbatim
! 136: *> Z is COMPLEX*16 array, dimension (LDZ,IHI)
! 137: *> If WANTZ is .FALSE., then Z is not referenced.
! 138: *> If WANTZ is .TRUE., then Z(ILO:IHI,ILOZ:IHIZ) is
! 139: *> replaced by Z(ILO:IHI,ILOZ:IHIZ)*U where U is the
! 140: *> orthogonal Schur factor of H(ILO:IHI,ILO:IHI).
! 141: *> (The output value of Z when INFO.GT.0 is given under
! 142: *> the description of INFO below.)
! 143: *> \endverbatim
! 144: *>
! 145: *> \param[in] LDZ
! 146: *> \verbatim
! 147: *> LDZ is INTEGER
! 148: *> The leading dimension of the array Z. if WANTZ is .TRUE.
! 149: *> then LDZ.GE.MAX(1,IHIZ). Otherwize, LDZ.GE.1.
! 150: *> \endverbatim
! 151: *>
! 152: *> \param[out] WORK
! 153: *> \verbatim
! 154: *> WORK is COMPLEX*16 array, dimension LWORK
! 155: *> On exit, if LWORK = -1, WORK(1) returns an estimate of
! 156: *> the optimal value for LWORK.
! 157: *> \endverbatim
! 158: *>
! 159: *> \param[in] LWORK
! 160: *> \verbatim
! 161: *> LWORK is INTEGER
! 162: *> The dimension of the array WORK. LWORK .GE. max(1,N)
! 163: *> is sufficient, but LWORK typically as large as 6*N may
! 164: *> be required for optimal performance. A workspace query
! 165: *> to determine the optimal workspace size is recommended.
! 166: *>
! 167: *> If LWORK = -1, then ZLAQR0 does a workspace query.
! 168: *> In this case, ZLAQR0 checks the input parameters and
! 169: *> estimates the optimal workspace size for the given
! 170: *> values of N, ILO and IHI. The estimate is returned
! 171: *> in WORK(1). No error message related to LWORK is
! 172: *> issued by XERBLA. Neither H nor Z are accessed.
! 173: *> \endverbatim
! 174: *>
! 175: *> \param[out] INFO
! 176: *> \verbatim
! 177: *> INFO is INTEGER
! 178: *> = 0: successful exit
! 179: *> .GT. 0: if INFO = i, ZLAQR0 failed to compute all of
! 180: *> the eigenvalues. Elements 1:ilo-1 and i+1:n of WR
! 181: *> and WI contain those eigenvalues which have been
! 182: *> successfully computed. (Failures are rare.)
! 183: *>
! 184: *> If INFO .GT. 0 and WANT is .FALSE., then on exit,
! 185: *> the remaining unconverged eigenvalues are the eigen-
! 186: *> values of the upper Hessenberg matrix rows and
! 187: *> columns ILO through INFO of the final, output
! 188: *> value of H.
! 189: *>
! 190: *> If INFO .GT. 0 and WANTT is .TRUE., then on exit
! 191: *>
! 192: *> (*) (initial value of H)*U = U*(final value of H)
! 193: *>
! 194: *> where U is a unitary matrix. The final
! 195: *> value of H is upper Hessenberg and triangular in
! 196: *> rows and columns INFO+1 through IHI.
! 197: *>
! 198: *> If INFO .GT. 0 and WANTZ is .TRUE., then on exit
! 199: *>
! 200: *> (final value of Z(ILO:IHI,ILOZ:IHIZ)
! 201: *> = (initial value of Z(ILO:IHI,ILOZ:IHIZ)*U
! 202: *>
! 203: *> where U is the unitary matrix in (*) (regard-
! 204: *> less of the value of WANTT.)
! 205: *>
! 206: *> If INFO .GT. 0 and WANTZ is .FALSE., then Z is not
! 207: *> accessed.
! 208: *> \endverbatim
! 209: *
! 210: * Authors:
! 211: * ========
! 212: *
! 213: *> \author Univ. of Tennessee
! 214: *> \author Univ. of California Berkeley
! 215: *> \author Univ. of Colorado Denver
! 216: *> \author NAG Ltd.
! 217: *
! 218: *> \date November 2011
! 219: *
! 220: *> \ingroup complex16OTHERauxiliary
! 221: *
! 222: *> \par Contributors:
! 223: * ==================
! 224: *>
! 225: *> Karen Braman and Ralph Byers, Department of Mathematics,
! 226: *> University of Kansas, USA
! 227: *
! 228: *> \par References:
! 229: * ================
! 230: *>
! 231: *> K. Braman, R. Byers and R. Mathias, The Multi-Shift QR
! 232: *> Algorithm Part I: Maintaining Well Focused Shifts, and Level 3
! 233: *> Performance, SIAM Journal of Matrix Analysis, volume 23, pages
! 234: *> 929--947, 2002.
! 235: *> \n
! 236: *> K. Braman, R. Byers and R. Mathias, The Multi-Shift QR
! 237: *> Algorithm Part II: Aggressive Early Deflation, SIAM Journal
! 238: *> of Matrix Analysis, volume 23, pages 948--973, 2002.
! 239: *>
! 240: * =====================================================================
1.1 bertrand 241: SUBROUTINE ZLAQR0( WANTT, WANTZ, N, ILO, IHI, H, LDH, W, ILOZ,
242: $ IHIZ, Z, LDZ, WORK, LWORK, INFO )
243: *
1.8 ! bertrand 244: * -- LAPACK auxiliary routine (version 3.4.0) --
! 245: * -- LAPACK is a software package provided by Univ. of Tennessee, --
! 246: * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
! 247: * November 2011
1.1 bertrand 248: *
249: * .. Scalar Arguments ..
250: INTEGER IHI, IHIZ, ILO, ILOZ, INFO, LDH, LDZ, LWORK, N
251: LOGICAL WANTT, WANTZ
252: * ..
253: * .. Array Arguments ..
254: COMPLEX*16 H( LDH, * ), W( * ), WORK( * ), Z( LDZ, * )
255: * ..
256: *
1.8 ! bertrand 257: * ================================================================
1.1 bertrand 258: *
259: * .. Parameters ..
260: *
261: * ==== Matrices of order NTINY or smaller must be processed by
262: * . ZLAHQR because of insufficient subdiagonal scratch space.
263: * . (This is a hard limit.) ====
264: INTEGER NTINY
265: PARAMETER ( NTINY = 11 )
266: *
267: * ==== Exceptional deflation windows: try to cure rare
268: * . slow convergence by varying the size of the
269: * . deflation window after KEXNW iterations. ====
270: INTEGER KEXNW
271: PARAMETER ( KEXNW = 5 )
272: *
273: * ==== Exceptional shifts: try to cure rare slow convergence
274: * . with ad-hoc exceptional shifts every KEXSH iterations.
275: * . ====
276: INTEGER KEXSH
277: PARAMETER ( KEXSH = 6 )
278: *
279: * ==== The constant WILK1 is used to form the exceptional
280: * . shifts. ====
281: DOUBLE PRECISION WILK1
282: PARAMETER ( WILK1 = 0.75d0 )
283: COMPLEX*16 ZERO, ONE
284: PARAMETER ( ZERO = ( 0.0d0, 0.0d0 ),
285: $ ONE = ( 1.0d0, 0.0d0 ) )
286: DOUBLE PRECISION TWO
287: PARAMETER ( TWO = 2.0d0 )
288: * ..
289: * .. Local Scalars ..
290: COMPLEX*16 AA, BB, CC, CDUM, DD, DET, RTDISC, SWAP, TR2
291: DOUBLE PRECISION S
292: INTEGER I, INF, IT, ITMAX, K, KACC22, KBOT, KDU, KS,
293: $ KT, KTOP, KU, KV, KWH, KWTOP, KWV, LD, LS,
294: $ LWKOPT, NDEC, NDFL, NH, NHO, NIBBLE, NMIN, NS,
295: $ NSMAX, NSR, NVE, NW, NWMAX, NWR, NWUPBD
296: LOGICAL SORTED
297: CHARACTER JBCMPZ*2
298: * ..
299: * .. External Functions ..
300: INTEGER ILAENV
301: EXTERNAL ILAENV
302: * ..
303: * .. Local Arrays ..
304: COMPLEX*16 ZDUM( 1, 1 )
305: * ..
306: * .. External Subroutines ..
307: EXTERNAL ZLACPY, ZLAHQR, ZLAQR3, ZLAQR4, ZLAQR5
308: * ..
309: * .. Intrinsic Functions ..
310: INTRINSIC ABS, DBLE, DCMPLX, DIMAG, INT, MAX, MIN, MOD,
311: $ SQRT
312: * ..
313: * .. Statement Functions ..
314: DOUBLE PRECISION CABS1
315: * ..
316: * .. Statement Function definitions ..
317: CABS1( CDUM ) = ABS( DBLE( CDUM ) ) + ABS( DIMAG( CDUM ) )
318: * ..
319: * .. Executable Statements ..
320: INFO = 0
321: *
322: * ==== Quick return for N = 0: nothing to do. ====
323: *
324: IF( N.EQ.0 ) THEN
325: WORK( 1 ) = ONE
326: RETURN
327: END IF
328: *
329: IF( N.LE.NTINY ) THEN
330: *
331: * ==== Tiny matrices must use ZLAHQR. ====
332: *
333: LWKOPT = 1
334: IF( LWORK.NE.-1 )
335: $ CALL ZLAHQR( WANTT, WANTZ, N, ILO, IHI, H, LDH, W, ILOZ,
336: $ IHIZ, Z, LDZ, INFO )
337: ELSE
338: *
339: * ==== Use small bulge multi-shift QR with aggressive early
340: * . deflation on larger-than-tiny matrices. ====
341: *
342: * ==== Hope for the best. ====
343: *
344: INFO = 0
345: *
346: * ==== Set up job flags for ILAENV. ====
347: *
348: IF( WANTT ) THEN
349: JBCMPZ( 1: 1 ) = 'S'
350: ELSE
351: JBCMPZ( 1: 1 ) = 'E'
352: END IF
353: IF( WANTZ ) THEN
354: JBCMPZ( 2: 2 ) = 'V'
355: ELSE
356: JBCMPZ( 2: 2 ) = 'N'
357: END IF
358: *
359: * ==== NWR = recommended deflation window size. At this
360: * . point, N .GT. NTINY = 11, so there is enough
361: * . subdiagonal workspace for NWR.GE.2 as required.
362: * . (In fact, there is enough subdiagonal space for
363: * . NWR.GE.3.) ====
364: *
365: NWR = ILAENV( 13, 'ZLAQR0', JBCMPZ, N, ILO, IHI, LWORK )
366: NWR = MAX( 2, NWR )
367: NWR = MIN( IHI-ILO+1, ( N-1 ) / 3, NWR )
368: *
369: * ==== NSR = recommended number of simultaneous shifts.
370: * . At this point N .GT. NTINY = 11, so there is at
371: * . enough subdiagonal workspace for NSR to be even
372: * . and greater than or equal to two as required. ====
373: *
374: NSR = ILAENV( 15, 'ZLAQR0', JBCMPZ, N, ILO, IHI, LWORK )
375: NSR = MIN( NSR, ( N+6 ) / 9, IHI-ILO )
376: NSR = MAX( 2, NSR-MOD( NSR, 2 ) )
377: *
378: * ==== Estimate optimal workspace ====
379: *
380: * ==== Workspace query call to ZLAQR3 ====
381: *
382: CALL ZLAQR3( WANTT, WANTZ, N, ILO, IHI, NWR+1, H, LDH, ILOZ,
383: $ IHIZ, Z, LDZ, LS, LD, W, H, LDH, N, H, LDH, N, H,
384: $ LDH, WORK, -1 )
385: *
386: * ==== Optimal workspace = MAX(ZLAQR5, ZLAQR3) ====
387: *
388: LWKOPT = MAX( 3*NSR / 2, INT( WORK( 1 ) ) )
389: *
390: * ==== Quick return in case of workspace query. ====
391: *
392: IF( LWORK.EQ.-1 ) THEN
393: WORK( 1 ) = DCMPLX( LWKOPT, 0 )
394: RETURN
395: END IF
396: *
397: * ==== ZLAHQR/ZLAQR0 crossover point ====
398: *
399: NMIN = ILAENV( 12, 'ZLAQR0', JBCMPZ, N, ILO, IHI, LWORK )
400: NMIN = MAX( NTINY, NMIN )
401: *
402: * ==== Nibble crossover point ====
403: *
404: NIBBLE = ILAENV( 14, 'ZLAQR0', JBCMPZ, N, ILO, IHI, LWORK )
405: NIBBLE = MAX( 0, NIBBLE )
406: *
407: * ==== Accumulate reflections during ttswp? Use block
408: * . 2-by-2 structure during matrix-matrix multiply? ====
409: *
410: KACC22 = ILAENV( 16, 'ZLAQR0', JBCMPZ, N, ILO, IHI, LWORK )
411: KACC22 = MAX( 0, KACC22 )
412: KACC22 = MIN( 2, KACC22 )
413: *
414: * ==== NWMAX = the largest possible deflation window for
415: * . which there is sufficient workspace. ====
416: *
417: NWMAX = MIN( ( N-1 ) / 3, LWORK / 2 )
418: NW = NWMAX
419: *
420: * ==== NSMAX = the Largest number of simultaneous shifts
421: * . for which there is sufficient workspace. ====
422: *
423: NSMAX = MIN( ( N+6 ) / 9, 2*LWORK / 3 )
424: NSMAX = NSMAX - MOD( NSMAX, 2 )
425: *
426: * ==== NDFL: an iteration count restarted at deflation. ====
427: *
428: NDFL = 1
429: *
430: * ==== ITMAX = iteration limit ====
431: *
432: ITMAX = MAX( 30, 2*KEXSH )*MAX( 10, ( IHI-ILO+1 ) )
433: *
434: * ==== Last row and column in the active block ====
435: *
436: KBOT = IHI
437: *
438: * ==== Main Loop ====
439: *
440: DO 70 IT = 1, ITMAX
441: *
442: * ==== Done when KBOT falls below ILO ====
443: *
444: IF( KBOT.LT.ILO )
445: $ GO TO 80
446: *
447: * ==== Locate active block ====
448: *
449: DO 10 K = KBOT, ILO + 1, -1
450: IF( H( K, K-1 ).EQ.ZERO )
451: $ GO TO 20
452: 10 CONTINUE
453: K = ILO
454: 20 CONTINUE
455: KTOP = K
456: *
457: * ==== Select deflation window size:
458: * . Typical Case:
459: * . If possible and advisable, nibble the entire
460: * . active block. If not, use size MIN(NWR,NWMAX)
461: * . or MIN(NWR+1,NWMAX) depending upon which has
462: * . the smaller corresponding subdiagonal entry
463: * . (a heuristic).
464: * .
465: * . Exceptional Case:
466: * . If there have been no deflations in KEXNW or
467: * . more iterations, then vary the deflation window
468: * . size. At first, because, larger windows are,
469: * . in general, more powerful than smaller ones,
470: * . rapidly increase the window to the maximum possible.
471: * . Then, gradually reduce the window size. ====
472: *
473: NH = KBOT - KTOP + 1
474: NWUPBD = MIN( NH, NWMAX )
475: IF( NDFL.LT.KEXNW ) THEN
476: NW = MIN( NWUPBD, NWR )
477: ELSE
478: NW = MIN( NWUPBD, 2*NW )
479: END IF
480: IF( NW.LT.NWMAX ) THEN
481: IF( NW.GE.NH-1 ) THEN
482: NW = NH
483: ELSE
484: KWTOP = KBOT - NW + 1
485: IF( CABS1( H( KWTOP, KWTOP-1 ) ).GT.
486: $ CABS1( H( KWTOP-1, KWTOP-2 ) ) )NW = NW + 1
487: END IF
488: END IF
489: IF( NDFL.LT.KEXNW ) THEN
490: NDEC = -1
491: ELSE IF( NDEC.GE.0 .OR. NW.GE.NWUPBD ) THEN
492: NDEC = NDEC + 1
493: IF( NW-NDEC.LT.2 )
494: $ NDEC = 0
495: NW = NW - NDEC
496: END IF
497: *
498: * ==== Aggressive early deflation:
499: * . split workspace under the subdiagonal into
500: * . - an nw-by-nw work array V in the lower
501: * . left-hand-corner,
502: * . - an NW-by-at-least-NW-but-more-is-better
503: * . (NW-by-NHO) horizontal work array along
504: * . the bottom edge,
505: * . - an at-least-NW-but-more-is-better (NHV-by-NW)
506: * . vertical work array along the left-hand-edge.
507: * . ====
508: *
509: KV = N - NW + 1
510: KT = NW + 1
511: NHO = ( N-NW-1 ) - KT + 1
512: KWV = NW + 2
513: NVE = ( N-NW ) - KWV + 1
514: *
515: * ==== Aggressive early deflation ====
516: *
517: CALL ZLAQR3( WANTT, WANTZ, N, KTOP, KBOT, NW, H, LDH, ILOZ,
518: $ IHIZ, Z, LDZ, LS, LD, W, H( KV, 1 ), LDH, NHO,
519: $ H( KV, KT ), LDH, NVE, H( KWV, 1 ), LDH, WORK,
520: $ LWORK )
521: *
522: * ==== Adjust KBOT accounting for new deflations. ====
523: *
524: KBOT = KBOT - LD
525: *
526: * ==== KS points to the shifts. ====
527: *
528: KS = KBOT - LS + 1
529: *
530: * ==== Skip an expensive QR sweep if there is a (partly
531: * . heuristic) reason to expect that many eigenvalues
532: * . will deflate without it. Here, the QR sweep is
533: * . skipped if many eigenvalues have just been deflated
534: * . or if the remaining active block is small.
535: *
536: IF( ( LD.EQ.0 ) .OR. ( ( 100*LD.LE.NW*NIBBLE ) .AND. ( KBOT-
537: $ KTOP+1.GT.MIN( NMIN, NWMAX ) ) ) ) THEN
538: *
539: * ==== NS = nominal number of simultaneous shifts.
540: * . This may be lowered (slightly) if ZLAQR3
541: * . did not provide that many shifts. ====
542: *
543: NS = MIN( NSMAX, NSR, MAX( 2, KBOT-KTOP ) )
544: NS = NS - MOD( NS, 2 )
545: *
546: * ==== If there have been no deflations
547: * . in a multiple of KEXSH iterations,
548: * . then try exceptional shifts.
549: * . Otherwise use shifts provided by
550: * . ZLAQR3 above or from the eigenvalues
551: * . of a trailing principal submatrix. ====
552: *
553: IF( MOD( NDFL, KEXSH ).EQ.0 ) THEN
554: KS = KBOT - NS + 1
555: DO 30 I = KBOT, KS + 1, -2
556: W( I ) = H( I, I ) + WILK1*CABS1( H( I, I-1 ) )
557: W( I-1 ) = W( I )
558: 30 CONTINUE
559: ELSE
560: *
561: * ==== Got NS/2 or fewer shifts? Use ZLAQR4 or
562: * . ZLAHQR on a trailing principal submatrix to
563: * . get more. (Since NS.LE.NSMAX.LE.(N+6)/9,
564: * . there is enough space below the subdiagonal
565: * . to fit an NS-by-NS scratch array.) ====
566: *
567: IF( KBOT-KS+1.LE.NS / 2 ) THEN
568: KS = KBOT - NS + 1
569: KT = N - NS + 1
570: CALL ZLACPY( 'A', NS, NS, H( KS, KS ), LDH,
571: $ H( KT, 1 ), LDH )
572: IF( NS.GT.NMIN ) THEN
573: CALL ZLAQR4( .false., .false., NS, 1, NS,
574: $ H( KT, 1 ), LDH, W( KS ), 1, 1,
575: $ ZDUM, 1, WORK, LWORK, INF )
576: ELSE
577: CALL ZLAHQR( .false., .false., NS, 1, NS,
578: $ H( KT, 1 ), LDH, W( KS ), 1, 1,
579: $ ZDUM, 1, INF )
580: END IF
581: KS = KS + INF
582: *
583: * ==== In case of a rare QR failure use
584: * . eigenvalues of the trailing 2-by-2
585: * . principal submatrix. Scale to avoid
586: * . overflows, underflows and subnormals.
587: * . (The scale factor S can not be zero,
588: * . because H(KBOT,KBOT-1) is nonzero.) ====
589: *
590: IF( KS.GE.KBOT ) THEN
591: S = CABS1( H( KBOT-1, KBOT-1 ) ) +
592: $ CABS1( H( KBOT, KBOT-1 ) ) +
593: $ CABS1( H( KBOT-1, KBOT ) ) +
594: $ CABS1( H( KBOT, KBOT ) )
595: AA = H( KBOT-1, KBOT-1 ) / S
596: CC = H( KBOT, KBOT-1 ) / S
597: BB = H( KBOT-1, KBOT ) / S
598: DD = H( KBOT, KBOT ) / S
599: TR2 = ( AA+DD ) / TWO
600: DET = ( AA-TR2 )*( DD-TR2 ) - BB*CC
601: RTDISC = SQRT( -DET )
602: W( KBOT-1 ) = ( TR2+RTDISC )*S
603: W( KBOT ) = ( TR2-RTDISC )*S
604: *
605: KS = KBOT - 1
606: END IF
607: END IF
608: *
609: IF( KBOT-KS+1.GT.NS ) THEN
610: *
611: * ==== Sort the shifts (Helps a little) ====
612: *
613: SORTED = .false.
614: DO 50 K = KBOT, KS + 1, -1
615: IF( SORTED )
616: $ GO TO 60
617: SORTED = .true.
618: DO 40 I = KS, K - 1
619: IF( CABS1( W( I ) ).LT.CABS1( W( I+1 ) ) )
620: $ THEN
621: SORTED = .false.
622: SWAP = W( I )
623: W( I ) = W( I+1 )
624: W( I+1 ) = SWAP
625: END IF
626: 40 CONTINUE
627: 50 CONTINUE
628: 60 CONTINUE
629: END IF
630: END IF
631: *
632: * ==== If there are only two shifts, then use
633: * . only one. ====
634: *
635: IF( KBOT-KS+1.EQ.2 ) THEN
636: IF( CABS1( W( KBOT )-H( KBOT, KBOT ) ).LT.
637: $ CABS1( W( KBOT-1 )-H( KBOT, KBOT ) ) ) THEN
638: W( KBOT-1 ) = W( KBOT )
639: ELSE
640: W( KBOT ) = W( KBOT-1 )
641: END IF
642: END IF
643: *
644: * ==== Use up to NS of the the smallest magnatiude
645: * . shifts. If there aren't NS shifts available,
646: * . then use them all, possibly dropping one to
647: * . make the number of shifts even. ====
648: *
649: NS = MIN( NS, KBOT-KS+1 )
650: NS = NS - MOD( NS, 2 )
651: KS = KBOT - NS + 1
652: *
653: * ==== Small-bulge multi-shift QR sweep:
654: * . split workspace under the subdiagonal into
655: * . - a KDU-by-KDU work array U in the lower
656: * . left-hand-corner,
657: * . - a KDU-by-at-least-KDU-but-more-is-better
658: * . (KDU-by-NHo) horizontal work array WH along
659: * . the bottom edge,
660: * . - and an at-least-KDU-but-more-is-better-by-KDU
661: * . (NVE-by-KDU) vertical work WV arrow along
662: * . the left-hand-edge. ====
663: *
664: KDU = 3*NS - 3
665: KU = N - KDU + 1
666: KWH = KDU + 1
667: NHO = ( N-KDU+1-4 ) - ( KDU+1 ) + 1
668: KWV = KDU + 4
669: NVE = N - KDU - KWV + 1
670: *
671: * ==== Small-bulge multi-shift QR sweep ====
672: *
673: CALL ZLAQR5( WANTT, WANTZ, KACC22, N, KTOP, KBOT, NS,
674: $ W( KS ), H, LDH, ILOZ, IHIZ, Z, LDZ, WORK,
675: $ 3, H( KU, 1 ), LDH, NVE, H( KWV, 1 ), LDH,
676: $ NHO, H( KU, KWH ), LDH )
677: END IF
678: *
679: * ==== Note progress (or the lack of it). ====
680: *
681: IF( LD.GT.0 ) THEN
682: NDFL = 1
683: ELSE
684: NDFL = NDFL + 1
685: END IF
686: *
687: * ==== End of main loop ====
688: 70 CONTINUE
689: *
690: * ==== Iteration limit exceeded. Set INFO to show where
691: * . the problem occurred and exit. ====
692: *
693: INFO = KBOT
694: 80 CONTINUE
695: END IF
696: *
697: * ==== Return the optimal value of LWORK. ====
698: *
699: WORK( 1 ) = DCMPLX( LWKOPT, 0 )
700: *
701: * ==== End of ZLAQR0 ====
702: *
703: END
CVSweb interface <joel.bertrand@systella.fr>