Mercurial > octave-nkf
comparison libcruft/lapack/zlaqr4.f @ 7034:68db500cb558
[project @ 2007-10-16 18:54:19 by jwe]
author | jwe |
---|---|
date | Tue, 16 Oct 2007 18:54:23 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
7033:f0142f2afdc6 | 7034:68db500cb558 |
---|---|
1 SUBROUTINE ZLAQR4( WANTT, WANTZ, N, ILO, IHI, H, LDH, W, ILOZ, | |
2 $ IHIZ, Z, LDZ, WORK, LWORK, INFO ) | |
3 * | |
4 * -- LAPACK auxiliary routine (version 3.1) -- | |
5 * Univ. of Tennessee, Univ. of California Berkeley and NAG Ltd.. | |
6 * November 2006 | |
7 * | |
8 * .. Scalar Arguments .. | |
9 INTEGER IHI, IHIZ, ILO, ILOZ, INFO, LDH, LDZ, LWORK, N | |
10 LOGICAL WANTT, WANTZ | |
11 * .. | |
12 * .. Array Arguments .. | |
13 COMPLEX*16 H( LDH, * ), W( * ), WORK( * ), Z( LDZ, * ) | |
14 * .. | |
15 * | |
16 * This subroutine implements one level of recursion for ZLAQR0. | |
17 * It is a complete implementation of the small bulge multi-shift | |
18 * QR algorithm. It may be called by ZLAQR0 and, for large enough | |
19 * deflation window size, it may be called by ZLAQR3. This | |
20 * subroutine is identical to ZLAQR0 except that it calls ZLAQR2 | |
21 * instead of ZLAQR3. | |
22 * | |
23 * Purpose | |
24 * ======= | |
25 * | |
26 * ZLAQR4 computes the eigenvalues of a Hessenberg matrix H | |
27 * and, optionally, the matrices T and Z from the Schur decomposition | |
28 * H = Z T Z**H, where T is an upper triangular matrix (the | |
29 * Schur form), and Z is the unitary matrix of Schur vectors. | |
30 * | |
31 * Optionally Z may be postmultiplied into an input unitary | |
32 * matrix Q so that this routine can give the Schur factorization | |
33 * of a matrix A which has been reduced to the Hessenberg form H | |
34 * by the unitary matrix Q: A = Q*H*Q**H = (QZ)*H*(QZ)**H. | |
35 * | |
36 * Arguments | |
37 * ========= | |
38 * | |
39 * WANTT (input) LOGICAL | |
40 * = .TRUE. : the full Schur form T is required; | |
41 * = .FALSE.: only eigenvalues are required. | |
42 * | |
43 * WANTZ (input) LOGICAL | |
44 * = .TRUE. : the matrix of Schur vectors Z is required; | |
45 * = .FALSE.: Schur vectors are not required. | |
46 * | |
47 * N (input) INTEGER | |
48 * The order of the matrix H. N .GE. 0. | |
49 * | |
50 * ILO (input) INTEGER | |
51 * IHI (input) INTEGER | |
52 * It is assumed that H is already upper triangular in rows | |
53 * and columns 1:ILO-1 and IHI+1:N and, if ILO.GT.1, | |
54 * H(ILO,ILO-1) is zero. ILO and IHI are normally set by a | |
55 * previous call to ZGEBAL, and then passed to ZGEHRD when the | |
56 * matrix output by ZGEBAL is reduced to Hessenberg form. | |
57 * Otherwise, ILO and IHI should be set to 1 and N, | |
58 * respectively. If N.GT.0, then 1.LE.ILO.LE.IHI.LE.N. | |
59 * If N = 0, then ILO = 1 and IHI = 0. | |
60 * | |
61 * H (input/output) COMPLEX*16 array, dimension (LDH,N) | |
62 * On entry, the upper Hessenberg matrix H. | |
63 * On exit, if INFO = 0 and WANTT is .TRUE., then H | |
64 * contains the upper triangular matrix T from the Schur | |
65 * decomposition (the Schur form). If INFO = 0 and WANT is | |
66 * .FALSE., then the contents of H are unspecified on exit. | |
67 * (The output value of H when INFO.GT.0 is given under the | |
68 * description of INFO below.) | |
69 * | |
70 * This subroutine may explicitly set H(i,j) = 0 for i.GT.j and | |
71 * j = 1, 2, ... ILO-1 or j = IHI+1, IHI+2, ... N. | |
72 * | |
73 * LDH (input) INTEGER | |
74 * The leading dimension of the array H. LDH .GE. max(1,N). | |
75 * | |
76 * W (output) COMPLEX*16 array, dimension (N) | |
77 * The computed eigenvalues of H(ILO:IHI,ILO:IHI) are stored | |
78 * in W(ILO:IHI). If WANTT is .TRUE., then the eigenvalues are | |
79 * stored in the same order as on the diagonal of the Schur | |
80 * form returned in H, with W(i) = H(i,i). | |
81 * | |
82 * Z (input/output) COMPLEX*16 array, dimension (LDZ,IHI) | |
83 * If WANTZ is .FALSE., then Z is not referenced. | |
84 * If WANTZ is .TRUE., then Z(ILO:IHI,ILOZ:IHIZ) is | |
85 * replaced by Z(ILO:IHI,ILOZ:IHIZ)*U where U is the | |
86 * orthogonal Schur factor of H(ILO:IHI,ILO:IHI). | |
87 * (The output value of Z when INFO.GT.0 is given under | |
88 * the description of INFO below.) | |
89 * | |
90 * LDZ (input) INTEGER | |
91 * The leading dimension of the array Z. if WANTZ is .TRUE. | |
92 * then LDZ.GE.MAX(1,IHIZ). Otherwize, LDZ.GE.1. | |
93 * | |
94 * WORK (workspace/output) COMPLEX*16 array, dimension LWORK | |
95 * On exit, if LWORK = -1, WORK(1) returns an estimate of | |
96 * the optimal value for LWORK. | |
97 * | |
98 * LWORK (input) INTEGER | |
99 * The dimension of the array WORK. LWORK .GE. max(1,N) | |
100 * is sufficient, but LWORK typically as large as 6*N may | |
101 * be required for optimal performance. A workspace query | |
102 * to determine the optimal workspace size is recommended. | |
103 * | |
104 * If LWORK = -1, then ZLAQR4 does a workspace query. | |
105 * In this case, ZLAQR4 checks the input parameters and | |
106 * estimates the optimal workspace size for the given | |
107 * values of N, ILO and IHI. The estimate is returned | |
108 * in WORK(1). No error message related to LWORK is | |
109 * issued by XERBLA. Neither H nor Z are accessed. | |
110 * | |
111 * | |
112 * INFO (output) INTEGER | |
113 * = 0: successful exit | |
114 * .GT. 0: if INFO = i, ZLAQR4 failed to compute all of | |
115 * the eigenvalues. Elements 1:ilo-1 and i+1:n of WR | |
116 * and WI contain those eigenvalues which have been | |
117 * successfully computed. (Failures are rare.) | |
118 * | |
119 * If INFO .GT. 0 and WANT is .FALSE., then on exit, | |
120 * the remaining unconverged eigenvalues are the eigen- | |
121 * values of the upper Hessenberg matrix rows and | |
122 * columns ILO through INFO of the final, output | |
123 * value of H. | |
124 * | |
125 * If INFO .GT. 0 and WANTT is .TRUE., then on exit | |
126 * | |
127 * (*) (initial value of H)*U = U*(final value of H) | |
128 * | |
129 * where U is a unitary matrix. The final | |
130 * value of H is upper Hessenberg and triangular in | |
131 * rows and columns INFO+1 through IHI. | |
132 * | |
133 * If INFO .GT. 0 and WANTZ is .TRUE., then on exit | |
134 * | |
135 * (final value of Z(ILO:IHI,ILOZ:IHIZ) | |
136 * = (initial value of Z(ILO:IHI,ILOZ:IHIZ)*U | |
137 * | |
138 * where U is the unitary matrix in (*) (regard- | |
139 * less of the value of WANTT.) | |
140 * | |
141 * If INFO .GT. 0 and WANTZ is .FALSE., then Z is not | |
142 * accessed. | |
143 * | |
144 * ================================================================ | |
145 * Based on contributions by | |
146 * Karen Braman and Ralph Byers, Department of Mathematics, | |
147 * University of Kansas, USA | |
148 * | |
149 * ================================================================ | |
150 * References: | |
151 * K. Braman, R. Byers and R. Mathias, The Multi-Shift QR | |
152 * Algorithm Part I: Maintaining Well Focused Shifts, and Level 3 | |
153 * Performance, SIAM Journal of Matrix Analysis, volume 23, pages | |
154 * 929--947, 2002. | |
155 * | |
156 * K. Braman, R. Byers and R. Mathias, The Multi-Shift QR | |
157 * Algorithm Part II: Aggressive Early Deflation, SIAM Journal | |
158 * of Matrix Analysis, volume 23, pages 948--973, 2002. | |
159 * | |
160 * ================================================================ | |
161 * .. Parameters .. | |
162 * | |
163 * ==== Matrices of order NTINY or smaller must be processed by | |
164 * . ZLAHQR because of insufficient subdiagonal scratch space. | |
165 * . (This is a hard limit.) ==== | |
166 * | |
167 * ==== Exceptional deflation windows: try to cure rare | |
168 * . slow convergence by increasing the size of the | |
169 * . deflation window after KEXNW iterations. ===== | |
170 * | |
171 * ==== Exceptional shifts: try to cure rare slow convergence | |
172 * . with ad-hoc exceptional shifts every KEXSH iterations. | |
173 * . The constants WILK1 and WILK2 are used to form the | |
174 * . exceptional shifts. ==== | |
175 * | |
176 INTEGER NTINY | |
177 PARAMETER ( NTINY = 11 ) | |
178 INTEGER KEXNW, KEXSH | |
179 PARAMETER ( KEXNW = 5, KEXSH = 6 ) | |
180 DOUBLE PRECISION WILK1 | |
181 PARAMETER ( WILK1 = 0.75d0 ) | |
182 COMPLEX*16 ZERO, ONE | |
183 PARAMETER ( ZERO = ( 0.0d0, 0.0d0 ), | |
184 $ ONE = ( 1.0d0, 0.0d0 ) ) | |
185 DOUBLE PRECISION TWO | |
186 PARAMETER ( TWO = 2.0d0 ) | |
187 * .. | |
188 * .. Local Scalars .. | |
189 COMPLEX*16 AA, BB, CC, CDUM, DD, DET, RTDISC, SWAP, TR2 | |
190 DOUBLE PRECISION S | |
191 INTEGER I, INF, IT, ITMAX, K, KACC22, KBOT, KDU, KS, | |
192 $ KT, KTOP, KU, KV, KWH, KWTOP, KWV, LD, LS, | |
193 $ LWKOPT, NDFL, NH, NHO, NIBBLE, NMIN, NS, NSMAX, | |
194 $ NSR, NVE, NW, NWMAX, NWR | |
195 LOGICAL NWINC, SORTED | |
196 CHARACTER JBCMPZ*2 | |
197 * .. | |
198 * .. External Functions .. | |
199 INTEGER ILAENV | |
200 EXTERNAL ILAENV | |
201 * .. | |
202 * .. Local Arrays .. | |
203 COMPLEX*16 ZDUM( 1, 1 ) | |
204 * .. | |
205 * .. External Subroutines .. | |
206 EXTERNAL ZLACPY, ZLAHQR, ZLAQR2, ZLAQR5 | |
207 * .. | |
208 * .. Intrinsic Functions .. | |
209 INTRINSIC ABS, DBLE, DCMPLX, DIMAG, INT, MAX, MIN, MOD, | |
210 $ SQRT | |
211 * .. | |
212 * .. Statement Functions .. | |
213 DOUBLE PRECISION CABS1 | |
214 * .. | |
215 * .. Statement Function definitions .. | |
216 CABS1( CDUM ) = ABS( DBLE( CDUM ) ) + ABS( DIMAG( CDUM ) ) | |
217 * .. | |
218 * .. Executable Statements .. | |
219 INFO = 0 | |
220 * | |
221 * ==== Quick return for N = 0: nothing to do. ==== | |
222 * | |
223 IF( N.EQ.0 ) THEN | |
224 WORK( 1 ) = ONE | |
225 RETURN | |
226 END IF | |
227 * | |
228 * ==== Set up job flags for ILAENV. ==== | |
229 * | |
230 IF( WANTT ) THEN | |
231 JBCMPZ( 1: 1 ) = 'S' | |
232 ELSE | |
233 JBCMPZ( 1: 1 ) = 'E' | |
234 END IF | |
235 IF( WANTZ ) THEN | |
236 JBCMPZ( 2: 2 ) = 'V' | |
237 ELSE | |
238 JBCMPZ( 2: 2 ) = 'N' | |
239 END IF | |
240 * | |
241 * ==== Tiny matrices must use ZLAHQR. ==== | |
242 * | |
243 IF( N.LE.NTINY ) THEN | |
244 * | |
245 * ==== Estimate optimal workspace. ==== | |
246 * | |
247 LWKOPT = 1 | |
248 IF( LWORK.NE.-1 ) | |
249 $ CALL ZLAHQR( WANTT, WANTZ, N, ILO, IHI, H, LDH, W, ILOZ, | |
250 $ IHIZ, Z, LDZ, INFO ) | |
251 ELSE | |
252 * | |
253 * ==== Use small bulge multi-shift QR with aggressive early | |
254 * . deflation on larger-than-tiny matrices. ==== | |
255 * | |
256 * ==== Hope for the best. ==== | |
257 * | |
258 INFO = 0 | |
259 * | |
260 * ==== NWR = recommended deflation window size. At this | |
261 * . point, N .GT. NTINY = 11, so there is enough | |
262 * . subdiagonal workspace for NWR.GE.2 as required. | |
263 * . (In fact, there is enough subdiagonal space for | |
264 * . NWR.GE.3.) ==== | |
265 * | |
266 NWR = ILAENV( 13, 'ZLAQR4', JBCMPZ, N, ILO, IHI, LWORK ) | |
267 NWR = MAX( 2, NWR ) | |
268 NWR = MIN( IHI-ILO+1, ( N-1 ) / 3, NWR ) | |
269 NW = NWR | |
270 * | |
271 * ==== NSR = recommended number of simultaneous shifts. | |
272 * . At this point N .GT. NTINY = 11, so there is at | |
273 * . enough subdiagonal workspace for NSR to be even | |
274 * . and greater than or equal to two as required. ==== | |
275 * | |
276 NSR = ILAENV( 15, 'ZLAQR4', JBCMPZ, N, ILO, IHI, LWORK ) | |
277 NSR = MIN( NSR, ( N+6 ) / 9, IHI-ILO ) | |
278 NSR = MAX( 2, NSR-MOD( NSR, 2 ) ) | |
279 * | |
280 * ==== Estimate optimal workspace ==== | |
281 * | |
282 * ==== Workspace query call to ZLAQR2 ==== | |
283 * | |
284 CALL ZLAQR2( WANTT, WANTZ, N, ILO, IHI, NWR+1, H, LDH, ILOZ, | |
285 $ IHIZ, Z, LDZ, LS, LD, W, H, LDH, N, H, LDH, N, H, | |
286 $ LDH, WORK, -1 ) | |
287 * | |
288 * ==== Optimal workspace = MAX(ZLAQR5, ZLAQR2) ==== | |
289 * | |
290 LWKOPT = MAX( 3*NSR / 2, INT( WORK( 1 ) ) ) | |
291 * | |
292 * ==== Quick return in case of workspace query. ==== | |
293 * | |
294 IF( LWORK.EQ.-1 ) THEN | |
295 WORK( 1 ) = DCMPLX( LWKOPT, 0 ) | |
296 RETURN | |
297 END IF | |
298 * | |
299 * ==== ZLAHQR/ZLAQR0 crossover point ==== | |
300 * | |
301 NMIN = ILAENV( 12, 'ZLAQR4', JBCMPZ, N, ILO, IHI, LWORK ) | |
302 NMIN = MAX( NTINY, NMIN ) | |
303 * | |
304 * ==== Nibble crossover point ==== | |
305 * | |
306 NIBBLE = ILAENV( 14, 'ZLAQR4', JBCMPZ, N, ILO, IHI, LWORK ) | |
307 NIBBLE = MAX( 0, NIBBLE ) | |
308 * | |
309 * ==== Accumulate reflections during ttswp? Use block | |
310 * . 2-by-2 structure during matrix-matrix multiply? ==== | |
311 * | |
312 KACC22 = ILAENV( 16, 'ZLAQR4', JBCMPZ, N, ILO, IHI, LWORK ) | |
313 KACC22 = MAX( 0, KACC22 ) | |
314 KACC22 = MIN( 2, KACC22 ) | |
315 * | |
316 * ==== NWMAX = the largest possible deflation window for | |
317 * . which there is sufficient workspace. ==== | |
318 * | |
319 NWMAX = MIN( ( N-1 ) / 3, LWORK / 2 ) | |
320 * | |
321 * ==== NSMAX = the Largest number of simultaneous shifts | |
322 * . for which there is sufficient workspace. ==== | |
323 * | |
324 NSMAX = MIN( ( N+6 ) / 9, 2*LWORK / 3 ) | |
325 NSMAX = NSMAX - MOD( NSMAX, 2 ) | |
326 * | |
327 * ==== NDFL: an iteration count restarted at deflation. ==== | |
328 * | |
329 NDFL = 1 | |
330 * | |
331 * ==== ITMAX = iteration limit ==== | |
332 * | |
333 ITMAX = MAX( 30, 2*KEXSH )*MAX( 10, ( IHI-ILO+1 ) ) | |
334 * | |
335 * ==== Last row and column in the active block ==== | |
336 * | |
337 KBOT = IHI | |
338 * | |
339 * ==== Main Loop ==== | |
340 * | |
341 DO 70 IT = 1, ITMAX | |
342 * | |
343 * ==== Done when KBOT falls below ILO ==== | |
344 * | |
345 IF( KBOT.LT.ILO ) | |
346 $ GO TO 80 | |
347 * | |
348 * ==== Locate active block ==== | |
349 * | |
350 DO 10 K = KBOT, ILO + 1, -1 | |
351 IF( H( K, K-1 ).EQ.ZERO ) | |
352 $ GO TO 20 | |
353 10 CONTINUE | |
354 K = ILO | |
355 20 CONTINUE | |
356 KTOP = K | |
357 * | |
358 * ==== Select deflation window size ==== | |
359 * | |
360 NH = KBOT - KTOP + 1 | |
361 IF( NDFL.LT.KEXNW .OR. NH.LT.NW ) THEN | |
362 * | |
363 * ==== Typical deflation window. If possible and | |
364 * . advisable, nibble the entire active block. | |
365 * . If not, use size NWR or NWR+1 depending upon | |
366 * . which has the smaller corresponding subdiagonal | |
367 * . entry (a heuristic). ==== | |
368 * | |
369 NWINC = .TRUE. | |
370 IF( NH.LE.MIN( NMIN, NWMAX ) ) THEN | |
371 NW = NH | |
372 ELSE | |
373 NW = MIN( NWR, NH, NWMAX ) | |
374 IF( NW.LT.NWMAX ) THEN | |
375 IF( NW.GE.NH-1 ) THEN | |
376 NW = NH | |
377 ELSE | |
378 KWTOP = KBOT - NW + 1 | |
379 IF( CABS1( H( KWTOP, KWTOP-1 ) ).GT. | |
380 $ CABS1( H( KWTOP-1, KWTOP-2 ) ) )NW = NW + 1 | |
381 END IF | |
382 END IF | |
383 END IF | |
384 ELSE | |
385 * | |
386 * ==== Exceptional deflation window. If there have | |
387 * . been no deflations in KEXNW or more iterations, | |
388 * . then vary the deflation window size. At first, | |
389 * . because, larger windows are, in general, more | |
390 * . powerful than smaller ones, rapidly increase the | |
391 * . window up to the maximum reasonable and possible. | |
392 * . Then maybe try a slightly smaller window. ==== | |
393 * | |
394 IF( NWINC .AND. NW.LT.MIN( NWMAX, NH ) ) THEN | |
395 NW = MIN( NWMAX, NH, 2*NW ) | |
396 ELSE | |
397 NWINC = .FALSE. | |
398 IF( NW.EQ.NH .AND. NH.GT.2 ) | |
399 $ NW = NH - 1 | |
400 END IF | |
401 END IF | |
402 * | |
403 * ==== Aggressive early deflation: | |
404 * . split workspace under the subdiagonal into | |
405 * . - an nw-by-nw work array V in the lower | |
406 * . left-hand-corner, | |
407 * . - an NW-by-at-least-NW-but-more-is-better | |
408 * . (NW-by-NHO) horizontal work array along | |
409 * . the bottom edge, | |
410 * . - an at-least-NW-but-more-is-better (NHV-by-NW) | |
411 * . vertical work array along the left-hand-edge. | |
412 * . ==== | |
413 * | |
414 KV = N - NW + 1 | |
415 KT = NW + 1 | |
416 NHO = ( N-NW-1 ) - KT + 1 | |
417 KWV = NW + 2 | |
418 NVE = ( N-NW ) - KWV + 1 | |
419 * | |
420 * ==== Aggressive early deflation ==== | |
421 * | |
422 CALL ZLAQR2( WANTT, WANTZ, N, KTOP, KBOT, NW, H, LDH, ILOZ, | |
423 $ IHIZ, Z, LDZ, LS, LD, W, H( KV, 1 ), LDH, NHO, | |
424 $ H( KV, KT ), LDH, NVE, H( KWV, 1 ), LDH, WORK, | |
425 $ LWORK ) | |
426 * | |
427 * ==== Adjust KBOT accounting for new deflations. ==== | |
428 * | |
429 KBOT = KBOT - LD | |
430 * | |
431 * ==== KS points to the shifts. ==== | |
432 * | |
433 KS = KBOT - LS + 1 | |
434 * | |
435 * ==== Skip an expensive QR sweep if there is a (partly | |
436 * . heuristic) reason to expect that many eigenvalues | |
437 * . will deflate without it. Here, the QR sweep is | |
438 * . skipped if many eigenvalues have just been deflated | |
439 * . or if the remaining active block is small. | |
440 * | |
441 IF( ( LD.EQ.0 ) .OR. ( ( 100*LD.LE.NW*NIBBLE ) .AND. ( KBOT- | |
442 $ KTOP+1.GT.MIN( NMIN, NWMAX ) ) ) ) THEN | |
443 * | |
444 * ==== NS = nominal number of simultaneous shifts. | |
445 * . This may be lowered (slightly) if ZLAQR2 | |
446 * . did not provide that many shifts. ==== | |
447 * | |
448 NS = MIN( NSMAX, NSR, MAX( 2, KBOT-KTOP ) ) | |
449 NS = NS - MOD( NS, 2 ) | |
450 * | |
451 * ==== If there have been no deflations | |
452 * . in a multiple of KEXSH iterations, | |
453 * . then try exceptional shifts. | |
454 * . Otherwise use shifts provided by | |
455 * . ZLAQR2 above or from the eigenvalues | |
456 * . of a trailing principal submatrix. ==== | |
457 * | |
458 IF( MOD( NDFL, KEXSH ).EQ.0 ) THEN | |
459 KS = KBOT - NS + 1 | |
460 DO 30 I = KBOT, KS + 1, -2 | |
461 W( I ) = H( I, I ) + WILK1*CABS1( H( I, I-1 ) ) | |
462 W( I-1 ) = W( I ) | |
463 30 CONTINUE | |
464 ELSE | |
465 * | |
466 * ==== Got NS/2 or fewer shifts? Use ZLAHQR | |
467 * . on a trailing principal submatrix to | |
468 * . get more. (Since NS.LE.NSMAX.LE.(N+6)/9, | |
469 * . there is enough space below the subdiagonal | |
470 * . to fit an NS-by-NS scratch array.) ==== | |
471 * | |
472 IF( KBOT-KS+1.LE.NS / 2 ) THEN | |
473 KS = KBOT - NS + 1 | |
474 KT = N - NS + 1 | |
475 CALL ZLACPY( 'A', NS, NS, H( KS, KS ), LDH, | |
476 $ H( KT, 1 ), LDH ) | |
477 CALL ZLAHQR( .false., .false., NS, 1, NS, | |
478 $ H( KT, 1 ), LDH, W( KS ), 1, 1, ZDUM, | |
479 $ 1, INF ) | |
480 KS = KS + INF | |
481 * | |
482 * ==== In case of a rare QR failure use | |
483 * . eigenvalues of the trailing 2-by-2 | |
484 * . principal submatrix. Scale to avoid | |
485 * . overflows, underflows and subnormals. | |
486 * . (The scale factor S can not be zero, | |
487 * . because H(KBOT,KBOT-1) is nonzero.) ==== | |
488 * | |
489 IF( KS.GE.KBOT ) THEN | |
490 S = CABS1( H( KBOT-1, KBOT-1 ) ) + | |
491 $ CABS1( H( KBOT, KBOT-1 ) ) + | |
492 $ CABS1( H( KBOT-1, KBOT ) ) + | |
493 $ CABS1( H( KBOT, KBOT ) ) | |
494 AA = H( KBOT-1, KBOT-1 ) / S | |
495 CC = H( KBOT, KBOT-1 ) / S | |
496 BB = H( KBOT-1, KBOT ) / S | |
497 DD = H( KBOT, KBOT ) / S | |
498 TR2 = ( AA+DD ) / TWO | |
499 DET = ( AA-TR2 )*( DD-TR2 ) - BB*CC | |
500 RTDISC = SQRT( -DET ) | |
501 W( KBOT-1 ) = ( TR2+RTDISC )*S | |
502 W( KBOT ) = ( TR2-RTDISC )*S | |
503 * | |
504 KS = KBOT - 1 | |
505 END IF | |
506 END IF | |
507 * | |
508 IF( KBOT-KS+1.GT.NS ) THEN | |
509 * | |
510 * ==== Sort the shifts (Helps a little) ==== | |
511 * | |
512 SORTED = .false. | |
513 DO 50 K = KBOT, KS + 1, -1 | |
514 IF( SORTED ) | |
515 $ GO TO 60 | |
516 SORTED = .true. | |
517 DO 40 I = KS, K - 1 | |
518 IF( CABS1( W( I ) ).LT.CABS1( W( I+1 ) ) ) | |
519 $ THEN | |
520 SORTED = .false. | |
521 SWAP = W( I ) | |
522 W( I ) = W( I+1 ) | |
523 W( I+1 ) = SWAP | |
524 END IF | |
525 40 CONTINUE | |
526 50 CONTINUE | |
527 60 CONTINUE | |
528 END IF | |
529 END IF | |
530 * | |
531 * ==== If there are only two shifts, then use | |
532 * . only one. ==== | |
533 * | |
534 IF( KBOT-KS+1.EQ.2 ) THEN | |
535 IF( CABS1( W( KBOT )-H( KBOT, KBOT ) ).LT. | |
536 $ CABS1( W( KBOT-1 )-H( KBOT, KBOT ) ) ) THEN | |
537 W( KBOT-1 ) = W( KBOT ) | |
538 ELSE | |
539 W( KBOT ) = W( KBOT-1 ) | |
540 END IF | |
541 END IF | |
542 * | |
543 * ==== Use up to NS of the the smallest magnatiude | |
544 * . shifts. If there aren't NS shifts available, | |
545 * . then use them all, possibly dropping one to | |
546 * . make the number of shifts even. ==== | |
547 * | |
548 NS = MIN( NS, KBOT-KS+1 ) | |
549 NS = NS - MOD( NS, 2 ) | |
550 KS = KBOT - NS + 1 | |
551 * | |
552 * ==== Small-bulge multi-shift QR sweep: | |
553 * . split workspace under the subdiagonal into | |
554 * . - a KDU-by-KDU work array U in the lower | |
555 * . left-hand-corner, | |
556 * . - a KDU-by-at-least-KDU-but-more-is-better | |
557 * . (KDU-by-NHo) horizontal work array WH along | |
558 * . the bottom edge, | |
559 * . - and an at-least-KDU-but-more-is-better-by-KDU | |
560 * . (NVE-by-KDU) vertical work WV arrow along | |
561 * . the left-hand-edge. ==== | |
562 * | |
563 KDU = 3*NS - 3 | |
564 KU = N - KDU + 1 | |
565 KWH = KDU + 1 | |
566 NHO = ( N-KDU+1-4 ) - ( KDU+1 ) + 1 | |
567 KWV = KDU + 4 | |
568 NVE = N - KDU - KWV + 1 | |
569 * | |
570 * ==== Small-bulge multi-shift QR sweep ==== | |
571 * | |
572 CALL ZLAQR5( WANTT, WANTZ, KACC22, N, KTOP, KBOT, NS, | |
573 $ W( KS ), H, LDH, ILOZ, IHIZ, Z, LDZ, WORK, | |
574 $ 3, H( KU, 1 ), LDH, NVE, H( KWV, 1 ), LDH, | |
575 $ NHO, H( KU, KWH ), LDH ) | |
576 END IF | |
577 * | |
578 * ==== Note progress (or the lack of it). ==== | |
579 * | |
580 IF( LD.GT.0 ) THEN | |
581 NDFL = 1 | |
582 ELSE | |
583 NDFL = NDFL + 1 | |
584 END IF | |
585 * | |
586 * ==== End of main loop ==== | |
587 70 CONTINUE | |
588 * | |
589 * ==== Iteration limit exceeded. Set INFO to show where | |
590 * . the problem occurred and exit. ==== | |
591 * | |
592 INFO = KBOT | |
593 80 CONTINUE | |
594 END IF | |
595 * | |
596 * ==== Return the optimal value of LWORK. ==== | |
597 * | |
598 WORK( 1 ) = DCMPLX( LWKOPT, 0 ) | |
599 * | |
600 * ==== End of ZLAQR4 ==== | |
601 * | |
602 END |