LAPACK  3.6.1
LAPACK: Linear Algebra PACKage
cgesvj.f
Go to the documentation of this file.
1 *> \brief \b CGESVJ
2 *
3 * =========== DOCUMENTATION ===========
4 *
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
7 *
8 *> \htmlonly
9 *> Download CGESVJ + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/cgesvj.f">
11 *> [TGZ]</a>
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/cgesvj.f">
13 *> [ZIP]</a>
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/cgesvj.f">
15 *> [TXT]</a>
16 *> \endhtmlonly
17 *
18 * Definition:
19 * ===========
20 *
21 * SUBROUTINE CGESVJ( JOBA, JOBU, JOBV, M, N, A, LDA, SVA, MV, V,
22 * LDV, CWORK, LWORK, RWORK, LRWORK, INFO )
23 *
24 * .. Scalar Arguments ..
25 * INTEGER INFO, LDA, LDV, LWORK, LRWORK, M, MV, N
26 * CHARACTER*1 JOBA, JOBU, JOBV
27 * ..
28 * .. Array Arguments ..
29 * COMPLEX A( LDA, * ), V( LDV, * ), CWORK( LWORK )
30 * REAL RWORK( LRWORK ), SVA( N )
31 * ..
32 *
33 *
34 *> \par Purpose:
35 * =============
36 *>
37 *> \verbatim
38 *>
39 * CGESVJ computes the singular value decomposition (SVD) of a complex
40 * M-by-N matrix A, where M >= N. The SVD of A is written as
41 * [++] [xx] [x0] [xx]
42 * A = U * SIGMA * V^*, [++] = [xx] * [ox] * [xx]
43 * [++] [xx]
44 * where SIGMA is an N-by-N diagonal matrix, U is an M-by-N orthonormal
45 * matrix, and V is an N-by-N unitary matrix. The diagonal elements
46 * of SIGMA are the singular values of A. The columns of U and V are the
47 * left and the right singular vectors of A, respectively.
48 *> \endverbatim
49 *
50 * Arguments:
51 * ==========
52 *
53 *> \param[in] JOBA
54 *> \verbatim
55 *> JOBA is CHARACTER* 1
56 *> Specifies the structure of A.
57 *> = 'L': The input matrix A is lower triangular;
58 *> = 'U': The input matrix A is upper triangular;
59 *> = 'G': The input matrix A is general M-by-N matrix, M >= N.
60 *> \endverbatim
61 *>
62 *> \param[in] JOBU
63 *> \verbatim
64 *> JOBU is CHARACTER*1
65 *> Specifies whether to compute the left singular vectors
66 *> (columns of U):
67 *> = 'U': The left singular vectors corresponding to the nonzero
68 *> singular values are computed and returned in the leading
69 *> columns of A. See more details in the description of A.
70 *> The default numerical orthogonality threshold is set to
71 *> approximately TOL=CTOL*EPS, CTOL=SQRT(M), EPS=SLAMCH('E').
72 *> = 'C': Analogous to JOBU='U', except that user can control the
73 *> level of numerical orthogonality of the computed left
74 *> singular vectors. TOL can be set to TOL = CTOL*EPS, where
75 *> CTOL is given on input in the array WORK.
76 *> No CTOL smaller than ONE is allowed. CTOL greater
77 *> than 1 / EPS is meaningless. The option 'C'
78 *> can be used if M*EPS is satisfactory orthogonality
79 *> of the computed left singular vectors, so CTOL=M could
80 *> save few sweeps of Jacobi rotations.
81 *> See the descriptions of A and WORK(1).
82 *> = 'N': The matrix U is not computed. However, see the
83 *> description of A.
84 *> \endverbatim
85 *>
86 *> \param[in] JOBV
87 *> \verbatim
88 *> JOBV is CHARACTER*1
89 *> Specifies whether to compute the right singular vectors, that
90 *> is, the matrix V:
91 *> = 'V' : the matrix V is computed and returned in the array V
92 *> = 'A' : the Jacobi rotations are applied to the MV-by-N
93 *> array V. In other words, the right singular vector
94 *> matrix V is not computed explicitly; instead it is
95 *> applied to an MV-by-N matrix initially stored in the
96 *> first MV rows of V.
97 *> = 'N' : the matrix V is not computed and the array V is not
98 *> referenced
99 *> \endverbatim
100 *>
101 *> \param[in] M
102 *> \verbatim
103 *> M is INTEGER
104 *> The number of rows of the input matrix A. 1/SLAMCH('E') > M >= 0.
105 *> \endverbatim
106 *>
107 *> \param[in] N
108 *> \verbatim
109 *> N is INTEGER
110 *> The number of columns of the input matrix A.
111 *> M >= N >= 0.
112 *> \endverbatim
113 *>
114 *> \param[in,out] A
115 *> \verbatim
116 *> A is COMPLEX array, dimension (LDA,N)
117 *> On entry, the M-by-N matrix A.
118 *> On exit,
119 *> If JOBU .EQ. 'U' .OR. JOBU .EQ. 'C':
120 *> If INFO .EQ. 0 :
121 *> RANKA orthonormal columns of U are returned in the
122 *> leading RANKA columns of the array A. Here RANKA <= N
123 *> is the number of computed singular values of A that are
124 *> above the underflow threshold SLAMCH('S'). The singular
125 *> vectors corresponding to underflowed or zero singular
126 *> values are not computed. The value of RANKA is returned
127 *> in the array RWORK as RANKA=NINT(RWORK(2)). Also see the
128 *> descriptions of SVA and RWORK. The computed columns of U
129 *> are mutually numerically orthogonal up to approximately
130 *> TOL=SQRT(M)*EPS (default); or TOL=CTOL*EPS (JOBU.EQ.'C'),
131 *> see the description of JOBU.
132 *> If INFO .GT. 0,
133 *> the procedure CGESVJ did not converge in the given number
134 *> of iterations (sweeps). In that case, the computed
135 *> columns of U may not be orthogonal up to TOL. The output
136 *> U (stored in A), SIGMA (given by the computed singular
137 *> values in SVA(1:N)) and V is still a decomposition of the
138 *> input matrix A in the sense that the residual
139 *> || A - SCALE * U * SIGMA * V^* ||_2 / ||A||_2 is small.
140 *> If JOBU .EQ. 'N':
141 *> If INFO .EQ. 0 :
142 *> Note that the left singular vectors are 'for free' in the
143 *> one-sided Jacobi SVD algorithm. However, if only the
144 *> singular values are needed, the level of numerical
145 *> orthogonality of U is not an issue and iterations are
146 *> stopped when the columns of the iterated matrix are
147 *> numerically orthogonal up to approximately M*EPS. Thus,
148 *> on exit, A contains the columns of U scaled with the
149 *> corresponding singular values.
150 *> If INFO .GT. 0 :
151 *> the procedure CGESVJ did not converge in the given number
152 *> of iterations (sweeps).
153 *> \endverbatim
154 *>
155 *> \param[in] LDA
156 *> \verbatim
157 *> LDA is INTEGER
158 *> The leading dimension of the array A. LDA >= max(1,M).
159 *> \endverbatim
160 *>
161 *> \param[out] SVA
162 *> \verbatim
163 *> SVA is REAL array, dimension (N)
164 *> On exit,
165 *> If INFO .EQ. 0 :
166 *> depending on the value SCALE = RWORK(1), we have:
167 *> If SCALE .EQ. ONE:
168 *> SVA(1:N) contains the computed singular values of A.
169 *> During the computation SVA contains the Euclidean column
170 *> norms of the iterated matrices in the array A.
171 *> If SCALE .NE. ONE:
172 *> The singular values of A are SCALE*SVA(1:N), and this
173 *> factored representation is due to the fact that some of the
174 *> singular values of A might underflow or overflow.
175 *>
176 *> If INFO .GT. 0 :
177 *> the procedure CGESVJ did not converge in the given number of
178 *> iterations (sweeps) and SCALE*SVA(1:N) may not be accurate.
179 *> \endverbatim
180 *>
181 *> \param[in] MV
182 *> \verbatim
183 *> MV is INTEGER
184 *> If JOBV .EQ. 'A', then the product of Jacobi rotations in CGESVJ
185 *> is applied to the first MV rows of V. See the description of JOBV.
186 *> \endverbatim
187 *>
188 *> \param[in,out] V
189 *> \verbatim
190 *> V is COMPLEX array, dimension (LDV,N)
191 *> If JOBV = 'V', then V contains on exit the N-by-N matrix of
192 *> the right singular vectors;
193 *> If JOBV = 'A', then V contains the product of the computed right
194 *> singular vector matrix and the initial matrix in
195 *> the array V.
196 *> If JOBV = 'N', then V is not referenced.
197 *> \endverbatim
198 *>
199 *> \param[in] LDV
200 *> \verbatim
201 *> LDV is INTEGER
202 *> The leading dimension of the array V, LDV .GE. 1.
203 *> If JOBV .EQ. 'V', then LDV .GE. max(1,N).
204 *> If JOBV .EQ. 'A', then LDV .GE. max(1,MV) .
205 *> \endverbatim
206 *>
207 *> \param[in,out] CWORK
208 *> \verbatim
209 *> CWORK is COMPLEX array, dimension M+N.
210 *> Used as work space.
211 *> \endverbatim
212 *>
213 *> \param[in] LWORK
214 *> \verbatim
215 *> LWORK is INTEGER
216 *> Length of CWORK, LWORK >= M+N.
217 *> \endverbatim
218 *>
219 *> \param[in,out] RWORK
220 *> \verbatim
221 *> RWORK is REAL array, dimension max(6,M+N).
222 *> On entry,
223 *> If JOBU .EQ. 'C' :
224 *> RWORK(1) = CTOL, where CTOL defines the threshold for convergence.
225 *> The process stops if all columns of A are mutually
226 *> orthogonal up to CTOL*EPS, EPS=SLAMCH('E').
227 *> It is required that CTOL >= ONE, i.e. it is not
228 *> allowed to force the routine to obtain orthogonality
229 *> below EPSILON.
230 *> On exit,
231 *> RWORK(1) = SCALE is the scaling factor such that SCALE*SVA(1:N)
232 *> are the computed singular values of A.
233 *> (See description of SVA().)
234 *> RWORK(2) = NINT(RWORK(2)) is the number of the computed nonzero
235 *> singular values.
236 *> RWORK(3) = NINT(RWORK(3)) is the number of the computed singular
237 *> values that are larger than the underflow threshold.
238 *> RWORK(4) = NINT(RWORK(4)) is the number of sweeps of Jacobi
239 *> rotations needed for numerical convergence.
240 *> RWORK(5) = max_{i.NE.j} |COS(A(:,i),A(:,j))| in the last sweep.
241 *> This is useful information in cases when CGESVJ did
242 *> not converge, as it can be used to estimate whether
243 *> the output is stil useful and for post festum analysis.
244 *> RWORK(6) = the largest absolute value over all sines of the
245 *> Jacobi rotation angles in the last sweep. It can be
246 *> useful for a post festum analysis.
247 *> \endverbatim
248 *>
249 *> \param[in] LRWORK
250 *> \verbatim
251 *> LRWORK is INTEGER
252 *> Length of RWORK, LRWORK >= MAX(6,N).
253 *> \endverbatim
254 *>
255 *> \param[out] INFO
256 *> \verbatim
257 *> INFO is INTEGER
258 *> = 0 : successful exit.
259 *> < 0 : if INFO = -i, then the i-th argument had an illegal value
260 *> > 0 : CGESVJ did not converge in the maximal allowed number
261 *> (NSWEEP=30) of sweeps. The output may still be useful.
262 *> See the description of RWORK.
263 *> \endverbatim
264 *
265 * Authors:
266 * ========
267 *
268 *> \author Univ. of Tennessee
269 *> \author Univ. of California Berkeley
270 *> \author Univ. of Colorado Denver
271 *> \author NAG Ltd.
272 *
273 *> \date June 2016
274 *
275 *> \ingroup complexGEcomputational
276 *
277 *> \par Further Details:
278 * =====================
279 *>
280 *> The orthogonal N-by-N matrix V is obtained as a product of Jacobi plane
281 *> rotations. In the case of underflow of the tangent of the Jacobi angle, a
282 *> modified Jacobi transformation of Drmac [3] is used. Pivot strategy uses
283 *> column interchanges of de Rijk [1]. The relative accuracy of the computed
284 *> singular values and the accuracy of the computed singular vectors (in
285 *> angle metric) is as guaranteed by the theory of Demmel and Veselic [2].
286 *> The condition number that determines the accuracy in the full rank case
287 *> is essentially min_{D=diag} kappa(A*D), where kappa(.) is the
288 *> spectral condition number. The best performance of this Jacobi SVD
289 *> procedure is achieved if used in an accelerated version of Drmac and
290 *> Veselic [4,5], and it is the kernel routine in the SIGMA library [6].
291 *> Some tunning parameters (marked with [TP]) are available for the
292 *> implementer.
293 *> The computational range for the nonzero singular values is the machine
294 *> number interval ( UNDERFLOW , OVERFLOW ). In extreme cases, even
295 *> denormalized singular values can be computed with the corresponding
296 *> gradual loss of accurate digits.
297 *>
298 *> \par Contributors:
299 * ==================
300 *>
301 *> Zlatko Drmac (Zagreb, Croatia) and Kresimir Veselic (Hagen, Germany)
302 *>
303 *> \par References:
304 * ================
305 *>
306 *> [1] P. P. M. De Rijk: A one-sided Jacobi algorithm for computing the
307 *> singular value decomposition on a vector computer.
308 *> SIAM J. Sci. Stat. Comp., Vol. 10 (1998), pp. 359-371.
309 *> [2] J. Demmel and K. Veselic: Jacobi method is more accurate than QR.
310 *> [3] Z. Drmac: Implementation of Jacobi rotations for accurate singular
311 *> value computation in floating point arithmetic.
312 *> SIAM J. Sci. Comp., Vol. 18 (1997), pp. 1200-1222.
313 *> [4] Z. Drmac and K. Veselic: New fast and accurate Jacobi SVD algorithm I.
314 *> SIAM J. Matrix Anal. Appl. Vol. 35, No. 2 (2008), pp. 1322-1342.
315 *> LAPACK Working note 169.
316 *> [5] Z. Drmac and K. Veselic: New fast and accurate Jacobi SVD algorithm II.
317 *> SIAM J. Matrix Anal. Appl. Vol. 35, No. 2 (2008), pp. 1343-1362.
318 *> LAPACK Working note 170.
319 *> [6] Z. Drmac: SIGMA - mathematical software library for accurate SVD, PSV,
320 *> QSVD, (H,K)-SVD computations.
321 *> Department of Mathematics, University of Zagreb, 2008, 2015.
322 *>
323 *> \par Bugs, Examples and Comments:
324 * =================================
325 *>
326 *> Please report all bugs and send interesting test examples and comments to
327 *> drmac@math.hr. Thank you.
328 *
329 * =====================================================================
330  SUBROUTINE cgesvj( JOBA, JOBU, JOBV, M, N, A, LDA, SVA, MV, V,
331  $ ldv, cwork, lwork, rwork, lrwork, info )
332 *
333 * -- LAPACK computational routine (version 3.6.1) --
334 * -- LAPACK is a software package provided by Univ. of Tennessee, --
335 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
336 * June 2016
337 *
338  IMPLICIT NONE
339 * .. Scalar Arguments ..
340  INTEGER INFO, LDA, LDV, LWORK, LRWORK, M, MV, N
341  CHARACTER*1 JOBA, JOBU, JOBV
342 * ..
343 * .. Array Arguments ..
344  COMPLEX A( lda, * ), V( ldv, * ), CWORK( lwork )
345  REAL RWORK( lrwork ), SVA( n )
346 * ..
347 *
348 * =====================================================================
349 *
350 * .. Local Parameters ..
351  REAL ZERO, HALF, ONE
352  parameter( zero = 0.0e0, half = 0.5e0, one = 1.0e0)
353  COMPLEX CZERO, CONE
354  parameter( czero = (0.0e0, 0.0e0), cone = (1.0e0, 0.0e0) )
355  INTEGER NSWEEP
356  parameter( nsweep = 30 )
357 * ..
358 * .. Local Scalars ..
359  COMPLEX AAPQ, OMPQ
360  REAL AAPP, AAPP0, AAPQ1, AAQQ, APOAQ, AQOAP, BIG,
361  $ bigtheta, cs, ctol, epsln, large, mxaapq,
362  $ mxsinj, rootbig, rooteps, rootsfmin, roottol,
363  $ skl, sfmin, small, sn, t, temp1, theta, thsign, tol
364  INTEGER BLSKIP, EMPTSW, i, ibr, IERR, igl, IJBLSK, ir1,
365  $ iswrot, jbc, jgl, kbl, lkahead, mvl, n2, n34,
366  $ n4, nbl, notrot, p, pskipped, q, rowskip, swband
367  LOGICAL APPLV, GOSCALE, LOWER, LSVEC, NOSCALE, ROTOK,
368  $ rsvec, uctol, upper
369 * ..
370 * ..
371 * .. Intrinsic Functions ..
372  INTRINSIC abs, amax1, amin1, conjg, float, min0, max0,
373  $ sign, sqrt
374 * ..
375 * .. External Functions ..
376 * ..
377 * from BLAS
378  REAL SCNRM2
379  COMPLEX CDOTC
380  EXTERNAL cdotc, scnrm2
381  INTEGER ISAMAX
382  EXTERNAL isamax
383 * from LAPACK
384  REAL SLAMCH
385  EXTERNAL slamch
386  LOGICAL LSAME
387  EXTERNAL lsame
388 * ..
389 * .. External Subroutines ..
390 * ..
391 * from BLAS
392  EXTERNAL ccopy, crot, csscal, cswap
393 * from LAPACK
394  EXTERNAL clascl, claset, classq, slascl, xerbla
395  EXTERNAL cgsvj0, cgsvj1
396 * ..
397 * .. Executable Statements ..
398 *
399 * Test the input arguments
400 *
401  lsvec = lsame( jobu, 'U' )
402  uctol = lsame( jobu, 'C' )
403  rsvec = lsame( jobv, 'V' )
404  applv = lsame( jobv, 'A' )
405  upper = lsame( joba, 'U' )
406  lower = lsame( joba, 'L' )
407 *
408  IF( .NOT.( upper .OR. lower .OR. lsame( joba, 'G' ) ) ) THEN
409  info = -1
410  ELSE IF( .NOT.( lsvec .OR. uctol .OR. lsame( jobu, 'N' ) ) ) THEN
411  info = -2
412  ELSE IF( .NOT.( rsvec .OR. applv .OR. lsame( jobv, 'N' ) ) ) THEN
413  info = -3
414  ELSE IF( m.LT.0 ) THEN
415  info = -4
416  ELSE IF( ( n.LT.0 ) .OR. ( n.GT.m ) ) THEN
417  info = -5
418  ELSE IF( lda.LT.m ) THEN
419  info = -7
420  ELSE IF( mv.LT.0 ) THEN
421  info = -9
422  ELSE IF( ( rsvec .AND. ( ldv.LT.n ) ) .OR.
423  $ ( applv .AND. ( ldv.LT.mv ) ) ) THEN
424  info = -11
425  ELSE IF( uctol .AND. ( rwork( 1 ).LE.one ) ) THEN
426  info = -12
427  ELSE IF( lwork.LT.( m+n ) ) THEN
428  info = -13
429  ELSE IF( lrwork.LT.max0( n, 6 ) ) THEN
430  info = -15
431  ELSE
432  info = 0
433  END IF
434 *
435 * #:(
436  IF( info.NE.0 ) THEN
437  CALL xerbla( 'CGESVJ', -info )
438  RETURN
439  END IF
440 *
441 * #:) Quick return for void matrix
442 *
443  IF( ( m.EQ.0 ) .OR. ( n.EQ.0 ) )RETURN
444 *
445 * Set numerical parameters
446 * The stopping criterion for Jacobi rotations is
447 *
448 * max_{i<>j}|A(:,i)^* * A(:,j)| / (||A(:,i)||*||A(:,j)||) < CTOL*EPS
449 *
450 * where EPS is the round-off and CTOL is defined as follows:
451 *
452  IF( uctol ) THEN
453 * ... user controlled
454  ctol = rwork( 1 )
455  ELSE
456 * ... default
457  IF( lsvec .OR. rsvec .OR. applv ) THEN
458  ctol = sqrt( float( m ) )
459  ELSE
460  ctol = float( m )
461  END IF
462  END IF
463 * ... and the machine dependent parameters are
464 *[!] (Make sure that SLAMCH() works properly on the target machine.)
465 *
466  epsln = slamch( 'Epsilon' )
467  rooteps = sqrt( epsln )
468  sfmin = slamch( 'SafeMinimum' )
469  rootsfmin = sqrt( sfmin )
470  small = sfmin / epsln
471  big = slamch( 'Overflow' )
472 * BIG = ONE / SFMIN
473  rootbig = one / rootsfmin
474  large = big / sqrt( float( m*n ) )
475  bigtheta = one / rooteps
476 *
477  tol = ctol*epsln
478  roottol = sqrt( tol )
479 *
480  IF( float( m )*epsln.GE.one ) THEN
481  info = -4
482  CALL xerbla( 'CGESVJ', -info )
483  RETURN
484  END IF
485 *
486 * Initialize the right singular vector matrix.
487 *
488  IF( rsvec ) THEN
489  mvl = n
490  CALL claset( 'A', mvl, n, czero, cone, v, ldv )
491  ELSE IF( applv ) THEN
492  mvl = mv
493  END IF
494  rsvec = rsvec .OR. applv
495 *
496 * Initialize SVA( 1:N ) = ( ||A e_i||_2, i = 1:N )
497 *(!) If necessary, scale A to protect the largest singular value
498 * from overflow. It is possible that saving the largest singular
499 * value destroys the information about the small ones.
500 * This initial scaling is almost minimal in the sense that the
501 * goal is to make sure that no column norm overflows, and that
502 * SQRT(N)*max_i SVA(i) does not overflow. If INFinite entries
503 * in A are detected, the procedure returns with INFO=-6.
504 *
505  skl = one / sqrt( float( m )*float( n ) )
506  noscale = .true.
507  goscale = .true.
508 *
509  IF( lower ) THEN
510 * the input matrix is M-by-N lower triangular (trapezoidal)
511  DO 1874 p = 1, n
512  aapp = zero
513  aaqq = one
514  CALL classq( m-p+1, a( p, p ), 1, aapp, aaqq )
515  IF( aapp.GT.big ) THEN
516  info = -6
517  CALL xerbla( 'CGESVJ', -info )
518  RETURN
519  END IF
520  aaqq = sqrt( aaqq )
521  IF( ( aapp.LT.( big / aaqq ) ) .AND. noscale ) THEN
522  sva( p ) = aapp*aaqq
523  ELSE
524  noscale = .false.
525  sva( p ) = aapp*( aaqq*skl )
526  IF( goscale ) THEN
527  goscale = .false.
528  DO 1873 q = 1, p - 1
529  sva( q ) = sva( q )*skl
530  1873 CONTINUE
531  END IF
532  END IF
533  1874 CONTINUE
534  ELSE IF( upper ) THEN
535 * the input matrix is M-by-N upper triangular (trapezoidal)
536  DO 2874 p = 1, n
537  aapp = zero
538  aaqq = one
539  CALL classq( p, a( 1, p ), 1, aapp, aaqq )
540  IF( aapp.GT.big ) THEN
541  info = -6
542  CALL xerbla( 'CGESVJ', -info )
543  RETURN
544  END IF
545  aaqq = sqrt( aaqq )
546  IF( ( aapp.LT.( big / aaqq ) ) .AND. noscale ) THEN
547  sva( p ) = aapp*aaqq
548  ELSE
549  noscale = .false.
550  sva( p ) = aapp*( aaqq*skl )
551  IF( goscale ) THEN
552  goscale = .false.
553  DO 2873 q = 1, p - 1
554  sva( q ) = sva( q )*skl
555  2873 CONTINUE
556  END IF
557  END IF
558  2874 CONTINUE
559  ELSE
560 * the input matrix is M-by-N general dense
561  DO 3874 p = 1, n
562  aapp = zero
563  aaqq = one
564  CALL classq( m, a( 1, p ), 1, aapp, aaqq )
565  IF( aapp.GT.big ) THEN
566  info = -6
567  CALL xerbla( 'CGESVJ', -info )
568  RETURN
569  END IF
570  aaqq = sqrt( aaqq )
571  IF( ( aapp.LT.( big / aaqq ) ) .AND. noscale ) THEN
572  sva( p ) = aapp*aaqq
573  ELSE
574  noscale = .false.
575  sva( p ) = aapp*( aaqq*skl )
576  IF( goscale ) THEN
577  goscale = .false.
578  DO 3873 q = 1, p - 1
579  sva( q ) = sva( q )*skl
580  3873 CONTINUE
581  END IF
582  END IF
583  3874 CONTINUE
584  END IF
585 *
586  IF( noscale )skl = one
587 *
588 * Move the smaller part of the spectrum from the underflow threshold
589 *(!) Start by determining the position of the nonzero entries of the
590 * array SVA() relative to ( SFMIN, BIG ).
591 *
592  aapp = zero
593  aaqq = big
594  DO 4781 p = 1, n
595  IF( sva( p ).NE.zero )aaqq = amin1( aaqq, sva( p ) )
596  aapp = amax1( aapp, sva( p ) )
597  4781 CONTINUE
598 *
599 * #:) Quick return for zero matrix
600 *
601  IF( aapp.EQ.zero ) THEN
602  IF( lsvec )CALL claset( 'G', m, n, czero, cone, a, lda )
603  rwork( 1 ) = one
604  rwork( 2 ) = zero
605  rwork( 3 ) = zero
606  rwork( 4 ) = zero
607  rwork( 5 ) = zero
608  rwork( 6 ) = zero
609  RETURN
610  END IF
611 *
612 * #:) Quick return for one-column matrix
613 *
614  IF( n.EQ.1 ) THEN
615  IF( lsvec )CALL clascl( 'G', 0, 0, sva( 1 ), skl, m, 1,
616  $ a( 1, 1 ), lda, ierr )
617  rwork( 1 ) = one / skl
618  IF( sva( 1 ).GE.sfmin ) THEN
619  rwork( 2 ) = one
620  ELSE
621  rwork( 2 ) = zero
622  END IF
623  rwork( 3 ) = zero
624  rwork( 4 ) = zero
625  rwork( 5 ) = zero
626  rwork( 6 ) = zero
627  RETURN
628  END IF
629 *
630 * Protect small singular values from underflow, and try to
631 * avoid underflows/overflows in computing Jacobi rotations.
632 *
633  sn = sqrt( sfmin / epsln )
634  temp1 = sqrt( big / float( n ) )
635  IF( ( aapp.LE.sn ) .OR. ( aaqq.GE.temp1 ) .OR.
636  $ ( ( sn.LE.aaqq ) .AND. ( aapp.LE.temp1 ) ) ) THEN
637  temp1 = amin1( big, temp1 / aapp )
638 * AAQQ = AAQQ*TEMP1
639 * AAPP = AAPP*TEMP1
640  ELSE IF( ( aaqq.LE.sn ) .AND. ( aapp.LE.temp1 ) ) THEN
641  temp1 = amin1( sn / aaqq, big / ( aapp*sqrt( float( n ) ) ) )
642 * AAQQ = AAQQ*TEMP1
643 * AAPP = AAPP*TEMP1
644  ELSE IF( ( aaqq.GE.sn ) .AND. ( aapp.GE.temp1 ) ) THEN
645  temp1 = amax1( sn / aaqq, temp1 / aapp )
646 * AAQQ = AAQQ*TEMP1
647 * AAPP = AAPP*TEMP1
648  ELSE IF( ( aaqq.LE.sn ) .AND. ( aapp.GE.temp1 ) ) THEN
649  temp1 = amin1( sn / aaqq, big / ( sqrt( float( n ) )*aapp ) )
650 * AAQQ = AAQQ*TEMP1
651 * AAPP = AAPP*TEMP1
652  ELSE
653  temp1 = one
654  END IF
655 *
656 * Scale, if necessary
657 *
658  IF( temp1.NE.one ) THEN
659  CALL slascl( 'G', 0, 0, one, temp1, n, 1, sva, n, ierr )
660  END IF
661  skl = temp1*skl
662  IF( skl.NE.one ) THEN
663  CALL clascl( joba, 0, 0, one, skl, m, n, a, lda, ierr )
664  skl = one / skl
665  END IF
666 *
667 * Row-cyclic Jacobi SVD algorithm with column pivoting
668 *
669  emptsw = ( n*( n-1 ) ) / 2
670  notrot = 0
671 
672  DO 1868 q = 1, n
673  cwork( q ) = cone
674  1868 CONTINUE
675 *
676 *
677 *
678  swband = 3
679 *[TP] SWBAND is a tuning parameter [TP]. It is meaningful and effective
680 * if CGESVJ is used as a computational routine in the preconditioned
681 * Jacobi SVD algorithm CGEJSV. For sweeps i=1:SWBAND the procedure
682 * works on pivots inside a band-like region around the diagonal.
683 * The boundaries are determined dynamically, based on the number of
684 * pivots above a threshold.
685 *
686  kbl = min0( 8, n )
687 *[TP] KBL is a tuning parameter that defines the tile size in the
688 * tiling of the p-q loops of pivot pairs. In general, an optimal
689 * value of KBL depends on the matrix dimensions and on the
690 * parameters of the computer's memory.
691 *
692  nbl = n / kbl
693  IF( ( nbl*kbl ).NE.n )nbl = nbl + 1
694 *
695  blskip = kbl**2
696 *[TP] BLKSKIP is a tuning parameter that depends on SWBAND and KBL.
697 *
698  rowskip = min0( 5, kbl )
699 *[TP] ROWSKIP is a tuning parameter.
700 *
701  lkahead = 1
702 *[TP] LKAHEAD is a tuning parameter.
703 *
704 * Quasi block transformations, using the lower (upper) triangular
705 * structure of the input matrix. The quasi-block-cycling usually
706 * invokes cubic convergence. Big part of this cycle is done inside
707 * canonical subspaces of dimensions less than M.
708 *
709  IF( ( lower .OR. upper ) .AND. ( n.GT.max0( 64, 4*kbl ) ) ) THEN
710 *[TP] The number of partition levels and the actual partition are
711 * tuning parameters.
712  n4 = n / 4
713  n2 = n / 2
714  n34 = 3*n4
715  IF( applv ) THEN
716  q = 0
717  ELSE
718  q = 1
719  END IF
720 *
721  IF( lower ) THEN
722 *
723 * This works very well on lower triangular matrices, in particular
724 * in the framework of the preconditioned Jacobi SVD (xGEJSV).
725 * The idea is simple:
726 * [+ 0 0 0] Note that Jacobi transformations of [0 0]
727 * [+ + 0 0] [0 0]
728 * [+ + x 0] actually work on [x 0] [x 0]
729 * [+ + x x] [x x]. [x x]
730 *
731  CALL cgsvj0( jobv, m-n34, n-n34, a( n34+1, n34+1 ), lda,
732  $ cwork( n34+1 ), sva( n34+1 ), mvl,
733  $ v( n34*q+1, n34+1 ), ldv, epsln, sfmin, tol,
734  $ 2, cwork( n+1 ), lwork-n, ierr )
735 
736  CALL cgsvj0( jobv, m-n2, n34-n2, a( n2+1, n2+1 ), lda,
737  $ cwork( n2+1 ), sva( n2+1 ), mvl,
738  $ v( n2*q+1, n2+1 ), ldv, epsln, sfmin, tol, 2,
739  $ cwork( n+1 ), lwork-n, ierr )
740 
741  CALL cgsvj1( jobv, m-n2, n-n2, n4, a( n2+1, n2+1 ), lda,
742  $ cwork( n2+1 ), sva( n2+1 ), mvl,
743  $ v( n2*q+1, n2+1 ), ldv, epsln, sfmin, tol, 1,
744  $ cwork( n+1 ), lwork-n, ierr )
745 *
746  CALL cgsvj0( jobv, m-n4, n2-n4, a( n4+1, n4+1 ), lda,
747  $ cwork( n4+1 ), sva( n4+1 ), mvl,
748  $ v( n4*q+1, n4+1 ), ldv, epsln, sfmin, tol, 1,
749  $ cwork( n+1 ), lwork-n, ierr )
750 *
751  CALL cgsvj0( jobv, m, n4, a, lda, cwork, sva, mvl, v, ldv,
752  $ epsln, sfmin, tol, 1, cwork( n+1 ), lwork-n,
753  $ ierr )
754 *
755  CALL cgsvj1( jobv, m, n2, n4, a, lda, cwork, sva, mvl, v,
756  $ ldv, epsln, sfmin, tol, 1, cwork( n+1 ),
757  $ lwork-n, ierr )
758 *
759 *
760  ELSE IF( upper ) THEN
761 *
762 *
763  CALL cgsvj0( jobv, n4, n4, a, lda, cwork, sva, mvl, v, ldv,
764  $ epsln, sfmin, tol, 2, cwork( n+1 ), lwork-n,
765  $ ierr )
766 *
767  CALL cgsvj0( jobv, n2, n4, a( 1, n4+1 ), lda, cwork( n4+1 ),
768  $ sva( n4+1 ), mvl, v( n4*q+1, n4+1 ), ldv,
769  $ epsln, sfmin, tol, 1, cwork( n+1 ), lwork-n,
770  $ ierr )
771 *
772  CALL cgsvj1( jobv, n2, n2, n4, a, lda, cwork, sva, mvl, v,
773  $ ldv, epsln, sfmin, tol, 1, cwork( n+1 ),
774  $ lwork-n, ierr )
775 *
776  CALL cgsvj0( jobv, n2+n4, n4, a( 1, n2+1 ), lda,
777  $ cwork( n2+1 ), sva( n2+1 ), mvl,
778  $ v( n2*q+1, n2+1 ), ldv, epsln, sfmin, tol, 1,
779  $ cwork( n+1 ), lwork-n, ierr )
780 
781  END IF
782 *
783  END IF
784 *
785 * .. Row-cyclic pivot strategy with de Rijk's pivoting ..
786 *
787  DO 1993 i = 1, nsweep
788 *
789 * .. go go go ...
790 *
791  mxaapq = zero
792  mxsinj = zero
793  iswrot = 0
794 *
795  notrot = 0
796  pskipped = 0
797 *
798 * Each sweep is unrolled using KBL-by-KBL tiles over the pivot pairs
799 * 1 <= p < q <= N. This is the first step toward a blocked implementation
800 * of the rotations. New implementation, based on block transformations,
801 * is under development.
802 *
803  DO 2000 ibr = 1, nbl
804 *
805  igl = ( ibr-1 )*kbl + 1
806 *
807  DO 1002 ir1 = 0, min0( lkahead, nbl-ibr )
808 *
809  igl = igl + ir1*kbl
810 *
811  DO 2001 p = igl, min0( igl+kbl-1, n-1 )
812 *
813 * .. de Rijk's pivoting
814 *
815  q = isamax( n-p+1, sva( p ), 1 ) + p - 1
816  IF( p.NE.q ) THEN
817  CALL cswap( m, a( 1, p ), 1, a( 1, q ), 1 )
818  IF( rsvec )CALL cswap( mvl, v( 1, p ), 1,
819  $ v( 1, q ), 1 )
820  temp1 = sva( p )
821  sva( p ) = sva( q )
822  sva( q ) = temp1
823  aapq = cwork(p)
824  cwork(p) = cwork(q)
825  cwork(q) = aapq
826  END IF
827 *
828  IF( ir1.EQ.0 ) THEN
829 *
830 * Column norms are periodically updated by explicit
831 * norm computation.
832 *[!] Caveat:
833 * Unfortunately, some BLAS implementations compute SCNRM2(M,A(1,p),1)
834 * as SQRT(S=CDOTC(M,A(1,p),1,A(1,p),1)), which may cause the result to
835 * overflow for ||A(:,p)||_2 > SQRT(overflow_threshold), and to
836 * underflow for ||A(:,p)||_2 < SQRT(underflow_threshold).
837 * Hence, SCNRM2 cannot be trusted, not even in the case when
838 * the true norm is far from the under(over)flow boundaries.
839 * If properly implemented SCNRM2 is available, the IF-THEN-ELSE-END IF
840 * below should be replaced with "AAPP = SCNRM2( M, A(1,p), 1 )".
841 *
842  IF( ( sva( p ).LT.rootbig ) .AND.
843  $ ( sva( p ).GT.rootsfmin ) ) THEN
844  sva( p ) = scnrm2( m, a( 1, p ), 1 )
845  ELSE
846  temp1 = zero
847  aapp = one
848  CALL classq( m, a( 1, p ), 1, temp1, aapp )
849  sva( p ) = temp1*sqrt( aapp )
850  END IF
851  aapp = sva( p )
852  ELSE
853  aapp = sva( p )
854  END IF
855 *
856  IF( aapp.GT.zero ) THEN
857 *
858  pskipped = 0
859 *
860  DO 2002 q = p + 1, min0( igl+kbl-1, n )
861 *
862  aaqq = sva( q )
863 *
864  IF( aaqq.GT.zero ) THEN
865 *
866  aapp0 = aapp
867  IF( aaqq.GE.one ) THEN
868  rotok = ( small*aapp ).LE.aaqq
869  IF( aapp.LT.( big / aaqq ) ) THEN
870  aapq = ( cdotc( m, a( 1, p ), 1,
871  $ a( 1, q ), 1 ) / aaqq ) / aapp
872  ELSE
873  CALL ccopy( m, a( 1, p ), 1,
874  $ cwork(n+1), 1 )
875  CALL clascl( 'G', 0, 0, aapp, one,
876  $ m, 1, cwork(n+1), lda, ierr )
877  aapq = cdotc( m, cwork(n+1), 1,
878  $ a( 1, q ), 1 ) / aaqq
879  END IF
880  ELSE
881  rotok = aapp.LE.( aaqq / small )
882  IF( aapp.GT.( small / aaqq ) ) THEN
883  aapq = ( cdotc( m, a( 1, p ), 1,
884  $ a( 1, q ), 1 ) / aaqq ) / aapp
885  ELSE
886  CALL ccopy( m, a( 1, q ), 1,
887  $ cwork(n+1), 1 )
888  CALL clascl( 'G', 0, 0, aaqq,
889  $ one, m, 1,
890  $ cwork(n+1), lda, ierr )
891  aapq = cdotc( m, a(1, p ), 1,
892  $ cwork(n+1), 1 ) / aapp
893  END IF
894  END IF
895 *
896 * AAPQ = AAPQ * CONJG( CWORK(p) ) * CWORK(q)
897  aapq1 = -abs(aapq)
898  mxaapq = amax1( mxaapq, -aapq1 )
899 *
900 * TO rotate or NOT to rotate, THAT is the question ...
901 *
902  IF( abs( aapq1 ).GT.tol ) THEN
903 *
904 * .. rotate
905 *[RTD] ROTATED = ROTATED + ONE
906 *
907  IF( ir1.EQ.0 ) THEN
908  notrot = 0
909  pskipped = 0
910  iswrot = iswrot + 1
911  END IF
912 *
913  IF( rotok ) THEN
914 *
915  ompq = aapq / abs(aapq)
916  aqoap = aaqq / aapp
917  apoaq = aapp / aaqq
918  theta = -half*abs( aqoap-apoaq )/aapq1
919 *
920  IF( abs( theta ).GT.bigtheta ) THEN
921 *
922  t = half / theta
923  cs = one
924 
925  CALL crot( m, a(1,p), 1, a(1,q), 1,
926  $ cs, conjg(ompq)*t )
927  IF ( rsvec ) THEN
928  CALL crot( mvl, v(1,p), 1,
929  $ v(1,q), 1, cs, conjg(ompq)*t )
930  END IF
931 
932  sva( q ) = aaqq*sqrt( amax1( zero,
933  $ one+t*apoaq*aapq1 ) )
934  aapp = aapp*sqrt( amax1( zero,
935  $ one-t*aqoap*aapq1 ) )
936  mxsinj = amax1( mxsinj, abs( t ) )
937 *
938  ELSE
939 *
940 * .. choose correct signum for THETA and rotate
941 *
942  thsign = -sign( one, aapq1 )
943  t = one / ( theta+thsign*
944  $ sqrt( one+theta*theta ) )
945  cs = sqrt( one / ( one+t*t ) )
946  sn = t*cs
947 *
948  mxsinj = amax1( mxsinj, abs( sn ) )
949  sva( q ) = aaqq*sqrt( amax1( zero,
950  $ one+t*apoaq*aapq1 ) )
951  aapp = aapp*sqrt( amax1( zero,
952  $ one-t*aqoap*aapq1 ) )
953 *
954  CALL crot( m, a(1,p), 1, a(1,q), 1,
955  $ cs, conjg(ompq)*sn )
956  IF ( rsvec ) THEN
957  CALL crot( mvl, v(1,p), 1,
958  $ v(1,q), 1, cs, conjg(ompq)*sn )
959  END IF
960  END IF
961  cwork(p) = -cwork(q) * ompq
962 *
963  ELSE
964 * .. have to use modified Gram-Schmidt like transformation
965  CALL ccopy( m, a( 1, p ), 1,
966  $ cwork(n+1), 1 )
967  CALL clascl( 'G', 0, 0, aapp, one, m,
968  $ 1, cwork(n+1), lda,
969  $ ierr )
970  CALL clascl( 'G', 0, 0, aaqq, one, m,
971  $ 1, a( 1, q ), lda, ierr )
972  CALL caxpy( m, -aapq, cwork(n+1), 1,
973  $ a( 1, q ), 1 )
974  CALL clascl( 'G', 0, 0, one, aaqq, m,
975  $ 1, a( 1, q ), lda, ierr )
976  sva( q ) = aaqq*sqrt( amax1( zero,
977  $ one-aapq1*aapq1 ) )
978  mxsinj = amax1( mxsinj, sfmin )
979  END IF
980 * END IF ROTOK THEN ... ELSE
981 *
982 * In the case of cancellation in updating SVA(q), SVA(p)
983 * recompute SVA(q), SVA(p).
984 *
985  IF( ( sva( q ) / aaqq )**2.LE.rooteps )
986  $ THEN
987  IF( ( aaqq.LT.rootbig ) .AND.
988  $ ( aaqq.GT.rootsfmin ) ) THEN
989  sva( q ) = scnrm2( m, a( 1, q ), 1 )
990  ELSE
991  t = zero
992  aaqq = one
993  CALL classq( m, a( 1, q ), 1, t,
994  $ aaqq )
995  sva( q ) = t*sqrt( aaqq )
996  END IF
997  END IF
998  IF( ( aapp / aapp0 ).LE.rooteps ) THEN
999  IF( ( aapp.LT.rootbig ) .AND.
1000  $ ( aapp.GT.rootsfmin ) ) THEN
1001  aapp = scnrm2( m, a( 1, p ), 1 )
1002  ELSE
1003  t = zero
1004  aapp = one
1005  CALL classq( m, a( 1, p ), 1, t,
1006  $ aapp )
1007  aapp = t*sqrt( aapp )
1008  END IF
1009  sva( p ) = aapp
1010  END IF
1011 *
1012  ELSE
1013 * A(:,p) and A(:,q) already numerically orthogonal
1014  IF( ir1.EQ.0 )notrot = notrot + 1
1015 *[RTD] SKIPPED = SKIPPED + 1
1016  pskipped = pskipped + 1
1017  END IF
1018  ELSE
1019 * A(:,q) is zero column
1020  IF( ir1.EQ.0 )notrot = notrot + 1
1021  pskipped = pskipped + 1
1022  END IF
1023 *
1024  IF( ( i.LE.swband ) .AND.
1025  $ ( pskipped.GT.rowskip ) ) THEN
1026  IF( ir1.EQ.0 )aapp = -aapp
1027  notrot = 0
1028  GO TO 2103
1029  END IF
1030 *
1031  2002 CONTINUE
1032 * END q-LOOP
1033 *
1034  2103 CONTINUE
1035 * bailed out of q-loop
1036 *
1037  sva( p ) = aapp
1038 *
1039  ELSE
1040  sva( p ) = aapp
1041  IF( ( ir1.EQ.0 ) .AND. ( aapp.EQ.zero ) )
1042  $ notrot = notrot + min0( igl+kbl-1, n ) - p
1043  END IF
1044 *
1045  2001 CONTINUE
1046 * end of the p-loop
1047 * end of doing the block ( ibr, ibr )
1048  1002 CONTINUE
1049 * end of ir1-loop
1050 *
1051 * ... go to the off diagonal blocks
1052 *
1053  igl = ( ibr-1 )*kbl + 1
1054 *
1055  DO 2010 jbc = ibr + 1, nbl
1056 *
1057  jgl = ( jbc-1 )*kbl + 1
1058 *
1059 * doing the block at ( ibr, jbc )
1060 *
1061  ijblsk = 0
1062  DO 2100 p = igl, min0( igl+kbl-1, n )
1063 *
1064  aapp = sva( p )
1065  IF( aapp.GT.zero ) THEN
1066 *
1067  pskipped = 0
1068 *
1069  DO 2200 q = jgl, min0( jgl+kbl-1, n )
1070 *
1071  aaqq = sva( q )
1072  IF( aaqq.GT.zero ) THEN
1073  aapp0 = aapp
1074 *
1075 * .. M x 2 Jacobi SVD ..
1076 *
1077 * Safe Gram matrix computation
1078 *
1079  IF( aaqq.GE.one ) THEN
1080  IF( aapp.GE.aaqq ) THEN
1081  rotok = ( small*aapp ).LE.aaqq
1082  ELSE
1083  rotok = ( small*aaqq ).LE.aapp
1084  END IF
1085  IF( aapp.LT.( big / aaqq ) ) THEN
1086  aapq = ( cdotc( m, a( 1, p ), 1,
1087  $ a( 1, q ), 1 ) / aaqq ) / aapp
1088  ELSE
1089  CALL ccopy( m, a( 1, p ), 1,
1090  $ cwork(n+1), 1 )
1091  CALL clascl( 'G', 0, 0, aapp,
1092  $ one, m, 1,
1093  $ cwork(n+1), lda, ierr )
1094  aapq = cdotc( m, cwork(n+1), 1,
1095  $ a( 1, q ), 1 ) / aaqq
1096  END IF
1097  ELSE
1098  IF( aapp.GE.aaqq ) THEN
1099  rotok = aapp.LE.( aaqq / small )
1100  ELSE
1101  rotok = aaqq.LE.( aapp / small )
1102  END IF
1103  IF( aapp.GT.( small / aaqq ) ) THEN
1104  aapq = ( cdotc( m, a( 1, p ), 1,
1105  $ a( 1, q ), 1 ) / aaqq ) / aapp
1106  ELSE
1107  CALL ccopy( m, a( 1, q ), 1,
1108  $ cwork(n+1), 1 )
1109  CALL clascl( 'G', 0, 0, aaqq,
1110  $ one, m, 1,
1111  $ cwork(n+1), lda, ierr )
1112  aapq = cdotc( m, a( 1, p ), 1,
1113  $ cwork(n+1), 1 ) / aapp
1114  END IF
1115  END IF
1116 *
1117 * AAPQ = AAPQ * CONJG(CWORK(p))*CWORK(q)
1118  aapq1 = -abs(aapq)
1119  mxaapq = amax1( mxaapq, -aapq1 )
1120 *
1121 * TO rotate or NOT to rotate, THAT is the question ...
1122 *
1123  IF( abs( aapq1 ).GT.tol ) THEN
1124  notrot = 0
1125 *[RTD] ROTATED = ROTATED + 1
1126  pskipped = 0
1127  iswrot = iswrot + 1
1128 *
1129  IF( rotok ) THEN
1130 *
1131  ompq = aapq / abs(aapq)
1132  aqoap = aaqq / aapp
1133  apoaq = aapp / aaqq
1134  theta = -half*abs( aqoap-apoaq )/ aapq1
1135  IF( aaqq.GT.aapp0 )theta = -theta
1136 *
1137  IF( abs( theta ).GT.bigtheta ) THEN
1138  t = half / theta
1139  cs = one
1140  CALL crot( m, a(1,p), 1, a(1,q), 1,
1141  $ cs, conjg(ompq)*t )
1142  IF( rsvec ) THEN
1143  CALL crot( mvl, v(1,p), 1,
1144  $ v(1,q), 1, cs, conjg(ompq)*t )
1145  END IF
1146  sva( q ) = aaqq*sqrt( amax1( zero,
1147  $ one+t*apoaq*aapq1 ) )
1148  aapp = aapp*sqrt( amax1( zero,
1149  $ one-t*aqoap*aapq1 ) )
1150  mxsinj = amax1( mxsinj, abs( t ) )
1151  ELSE
1152 *
1153 * .. choose correct signum for THETA and rotate
1154 *
1155  thsign = -sign( one, aapq1 )
1156  IF( aaqq.GT.aapp0 )thsign = -thsign
1157  t = one / ( theta+thsign*
1158  $ sqrt( one+theta*theta ) )
1159  cs = sqrt( one / ( one+t*t ) )
1160  sn = t*cs
1161  mxsinj = amax1( mxsinj, abs( sn ) )
1162  sva( q ) = aaqq*sqrt( amax1( zero,
1163  $ one+t*apoaq*aapq1 ) )
1164  aapp = aapp*sqrt( amax1( zero,
1165  $ one-t*aqoap*aapq1 ) )
1166 *
1167  CALL crot( m, a(1,p), 1, a(1,q), 1,
1168  $ cs, conjg(ompq)*sn )
1169  IF( rsvec ) THEN
1170  CALL crot( mvl, v(1,p), 1,
1171  $ v(1,q), 1, cs, conjg(ompq)*sn )
1172  END IF
1173  END IF
1174  cwork(p) = -cwork(q) * ompq
1175 *
1176  ELSE
1177 * .. have to use modified Gram-Schmidt like transformation
1178  IF( aapp.GT.aaqq ) THEN
1179  CALL ccopy( m, a( 1, p ), 1,
1180  $ cwork(n+1), 1 )
1181  CALL clascl( 'G', 0, 0, aapp, one,
1182  $ m, 1, cwork(n+1),lda,
1183  $ ierr )
1184  CALL clascl( 'G', 0, 0, aaqq, one,
1185  $ m, 1, a( 1, q ), lda,
1186  $ ierr )
1187  CALL caxpy( m, -aapq, cwork(n+1),
1188  $ 1, a( 1, q ), 1 )
1189  CALL clascl( 'G', 0, 0, one, aaqq,
1190  $ m, 1, a( 1, q ), lda,
1191  $ ierr )
1192  sva( q ) = aaqq*sqrt( amax1( zero,
1193  $ one-aapq1*aapq1 ) )
1194  mxsinj = amax1( mxsinj, sfmin )
1195  ELSE
1196  CALL ccopy( m, a( 1, q ), 1,
1197  $ cwork(n+1), 1 )
1198  CALL clascl( 'G', 0, 0, aaqq, one,
1199  $ m, 1, cwork(n+1),lda,
1200  $ ierr )
1201  CALL clascl( 'G', 0, 0, aapp, one,
1202  $ m, 1, a( 1, p ), lda,
1203  $ ierr )
1204  CALL caxpy( m, -conjg(aapq),
1205  $ cwork(n+1), 1, a( 1, p ), 1 )
1206  CALL clascl( 'G', 0, 0, one, aapp,
1207  $ m, 1, a( 1, p ), lda,
1208  $ ierr )
1209  sva( p ) = aapp*sqrt( amax1( zero,
1210  $ one-aapq1*aapq1 ) )
1211  mxsinj = amax1( mxsinj, sfmin )
1212  END IF
1213  END IF
1214 * END IF ROTOK THEN ... ELSE
1215 *
1216 * In the case of cancellation in updating SVA(q), SVA(p)
1217 * .. recompute SVA(q), SVA(p)
1218  IF( ( sva( q ) / aaqq )**2.LE.rooteps )
1219  $ THEN
1220  IF( ( aaqq.LT.rootbig ) .AND.
1221  $ ( aaqq.GT.rootsfmin ) ) THEN
1222  sva( q ) = scnrm2( m, a( 1, q ), 1)
1223  ELSE
1224  t = zero
1225  aaqq = one
1226  CALL classq( m, a( 1, q ), 1, t,
1227  $ aaqq )
1228  sva( q ) = t*sqrt( aaqq )
1229  END IF
1230  END IF
1231  IF( ( aapp / aapp0 )**2.LE.rooteps ) THEN
1232  IF( ( aapp.LT.rootbig ) .AND.
1233  $ ( aapp.GT.rootsfmin ) ) THEN
1234  aapp = scnrm2( m, a( 1, p ), 1 )
1235  ELSE
1236  t = zero
1237  aapp = one
1238  CALL classq( m, a( 1, p ), 1, t,
1239  $ aapp )
1240  aapp = t*sqrt( aapp )
1241  END IF
1242  sva( p ) = aapp
1243  END IF
1244 * end of OK rotation
1245  ELSE
1246  notrot = notrot + 1
1247 *[RTD] SKIPPED = SKIPPED + 1
1248  pskipped = pskipped + 1
1249  ijblsk = ijblsk + 1
1250  END IF
1251  ELSE
1252  notrot = notrot + 1
1253  pskipped = pskipped + 1
1254  ijblsk = ijblsk + 1
1255  END IF
1256 *
1257  IF( ( i.LE.swband ) .AND. ( ijblsk.GE.blskip ) )
1258  $ THEN
1259  sva( p ) = aapp
1260  notrot = 0
1261  GO TO 2011
1262  END IF
1263  IF( ( i.LE.swband ) .AND.
1264  $ ( pskipped.GT.rowskip ) ) THEN
1265  aapp = -aapp
1266  notrot = 0
1267  GO TO 2203
1268  END IF
1269 *
1270  2200 CONTINUE
1271 * end of the q-loop
1272  2203 CONTINUE
1273 *
1274  sva( p ) = aapp
1275 *
1276  ELSE
1277 *
1278  IF( aapp.EQ.zero )notrot = notrot +
1279  $ min0( jgl+kbl-1, n ) - jgl + 1
1280  IF( aapp.LT.zero )notrot = 0
1281 *
1282  END IF
1283 *
1284  2100 CONTINUE
1285 * end of the p-loop
1286  2010 CONTINUE
1287 * end of the jbc-loop
1288  2011 CONTINUE
1289 *2011 bailed out of the jbc-loop
1290  DO 2012 p = igl, min0( igl+kbl-1, n )
1291  sva( p ) = abs( sva( p ) )
1292  2012 CONTINUE
1293 ***
1294  2000 CONTINUE
1295 *2000 :: end of the ibr-loop
1296 *
1297 * .. update SVA(N)
1298  IF( ( sva( n ).LT.rootbig ) .AND. ( sva( n ).GT.rootsfmin ) )
1299  $ THEN
1300  sva( n ) = scnrm2( m, a( 1, n ), 1 )
1301  ELSE
1302  t = zero
1303  aapp = one
1304  CALL classq( m, a( 1, n ), 1, t, aapp )
1305  sva( n ) = t*sqrt( aapp )
1306  END IF
1307 *
1308 * Additional steering devices
1309 *
1310  IF( ( i.LT.swband ) .AND. ( ( mxaapq.LE.roottol ) .OR.
1311  $ ( iswrot.LE.n ) ) )swband = i
1312 *
1313  IF( ( i.GT.swband+1 ) .AND. ( mxaapq.LT.sqrt( float( n ) )*
1314  $ tol ) .AND. ( float( n )*mxaapq*mxsinj.LT.tol ) ) THEN
1315  GO TO 1994
1316  END IF
1317 *
1318  IF( notrot.GE.emptsw )GO TO 1994
1319 *
1320  1993 CONTINUE
1321 * end i=1:NSWEEP loop
1322 *
1323 * #:( Reaching this point means that the procedure has not converged.
1324  info = nsweep - 1
1325  GO TO 1995
1326 *
1327  1994 CONTINUE
1328 * #:) Reaching this point means numerical convergence after the i-th
1329 * sweep.
1330 *
1331  info = 0
1332 * #:) INFO = 0 confirms successful iterations.
1333  1995 CONTINUE
1334 *
1335 * Sort the singular values and find how many are above
1336 * the underflow threshold.
1337 *
1338  n2 = 0
1339  n4 = 0
1340  DO 5991 p = 1, n - 1
1341  q = isamax( n-p+1, sva( p ), 1 ) + p - 1
1342  IF( p.NE.q ) THEN
1343  temp1 = sva( p )
1344  sva( p ) = sva( q )
1345  sva( q ) = temp1
1346  CALL cswap( m, a( 1, p ), 1, a( 1, q ), 1 )
1347  IF( rsvec )CALL cswap( mvl, v( 1, p ), 1, v( 1, q ), 1 )
1348  END IF
1349  IF( sva( p ).NE.zero ) THEN
1350  n4 = n4 + 1
1351  IF( sva( p )*skl.GT.sfmin )n2 = n2 + 1
1352  END IF
1353  5991 CONTINUE
1354  IF( sva( n ).NE.zero ) THEN
1355  n4 = n4 + 1
1356  IF( sva( n )*skl.GT.sfmin )n2 = n2 + 1
1357  END IF
1358 *
1359 * Normalize the left singular vectors.
1360 *
1361  IF( lsvec .OR. uctol ) THEN
1362  DO 1998 p = 1, n2
1363  CALL csscal( m, one / sva( p ), a( 1, p ), 1 )
1364  1998 CONTINUE
1365  END IF
1366 *
1367 * Scale the product of Jacobi rotations.
1368 *
1369  IF( rsvec ) THEN
1370  DO 2399 p = 1, n
1371  temp1 = one / scnrm2( mvl, v( 1, p ), 1 )
1372  CALL csscal( mvl, temp1, v( 1, p ), 1 )
1373  2399 CONTINUE
1374  END IF
1375 *
1376 * Undo scaling, if necessary (and possible).
1377  IF( ( ( skl.GT.one ) .AND. ( sva( 1 ).LT.( big / skl ) ) )
1378  $ .OR. ( ( skl.LT.one ) .AND. ( sva( max( n2, 1 ) ) .GT.
1379  $ ( sfmin / skl ) ) ) ) THEN
1380  DO 2400 p = 1, n
1381  sva( p ) = skl*sva( p )
1382  2400 CONTINUE
1383  skl = one
1384  END IF
1385 *
1386  rwork( 1 ) = skl
1387 * The singular values of A are SKL*SVA(1:N). If SKL.NE.ONE
1388 * then some of the singular values may overflow or underflow and
1389 * the spectrum is given in this factored representation.
1390 *
1391  rwork( 2 ) = float( n4 )
1392 * N4 is the number of computed nonzero singular values of A.
1393 *
1394  rwork( 3 ) = float( n2 )
1395 * N2 is the number of singular values of A greater than SFMIN.
1396 * If N2<N, SVA(N2:N) contains ZEROS and/or denormalized numbers
1397 * that may carry some information.
1398 *
1399  rwork( 4 ) = float( i )
1400 * i is the index of the last sweep before declaring convergence.
1401 *
1402  rwork( 5 ) = mxaapq
1403 * MXAAPQ is the largest absolute value of scaled pivots in the
1404 * last sweep
1405 *
1406  rwork( 6 ) = mxsinj
1407 * MXSINJ is the largest absolute value of the sines of Jacobi angles
1408 * in the last sweep
1409 *
1410  RETURN
1411 * ..
1412 * .. END OF CGESVJ
1413 * ..
1414  END
subroutine clascl(TYPE, KL, KU, CFROM, CTO, M, N, A, LDA, INFO)
CLASCL multiplies a general rectangular matrix by a real scalar defined as cto/cfrom.
Definition: clascl.f:145
subroutine slascl(TYPE, KL, KU, CFROM, CTO, M, N, A, LDA, INFO)
SLASCL multiplies a general rectangular matrix by a real scalar defined as cto/cfrom.
Definition: slascl.f:145
subroutine classq(N, X, INCX, SCALE, SUMSQ)
CLASSQ updates a sum of squares represented in scaled form.
Definition: classq.f:108
subroutine xerbla(SRNAME, INFO)
XERBLA
Definition: xerbla.f:62
subroutine claset(UPLO, M, N, ALPHA, BETA, A, LDA)
CLASET initializes the off-diagonal elements and the diagonal elements of a matrix to given values...
Definition: claset.f:108
subroutine cgsvj0(JOBV, M, N, A, LDA, D, SVA, MV, V, LDV, EPS, SFMIN, TOL, NSWEEP, WORK, LWORK, INFO)
CGSVJ0 pre-processor for the routine cgesvj.
Definition: cgsvj0.f:220
subroutine cgesvj(JOBA, JOBU, JOBV, M, N, A, LDA, SVA, MV, V, LDV, CWORK, LWORK, RWORK, LRWORK, INFO)
CGESVJ
Definition: cgesvj.f:332
subroutine ccopy(N, CX, INCX, CY, INCY)
CCOPY
Definition: ccopy.f:52
subroutine cswap(N, CX, INCX, CY, INCY)
CSWAP
Definition: cswap.f:52
subroutine cgsvj1(JOBV, M, N, N1, A, LDA, D, SVA, MV, V, LDV, EPS, SFMIN, TOL, NSWEEP, WORK, LWORK, INFO)
CGSVJ1 pre-processor for the routine cgesvj, applies Jacobi rotations targeting only particular pivot...
Definition: cgsvj1.f:238
subroutine caxpy(N, CA, CX, INCX, CY, INCY)
CAXPY
Definition: caxpy.f:53
subroutine csscal(N, SA, CX, INCX)
CSSCAL
Definition: csscal.f:54
subroutine crot(N, CX, INCX, CY, INCY, C, S)
CROT applies a plane rotation with real cosine and complex sine to a pair of complex vectors...
Definition: crot.f:105