1: SUBROUTINE DGSVJ0( JOBV, M, N, A, LDA, D, SVA, MV, V, LDV, EPS,
2: + SFMIN, TOL, NSWEEP, WORK, LWORK, INFO )
3: *
4: * -- LAPACK routine (version 3.2.2) --
5: *
6: * -- Contributed by Zlatko Drmac of the University of Zagreb and --
7: * -- Kresimir Veselic of the Fernuniversitaet Hagen --
8: * -- June 2010 --
9: *
10: * -- LAPACK is a software package provided by Univ. of Tennessee, --
11: * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
12: *
13: * This routine is also part of SIGMA (version 1.23, October 23. 2008.)
14: * SIGMA is a library of algorithms for highly accurate algorithms for
15: * computation of SVD, PSVD, QSVD, (H,K)-SVD, and for solution of the
16: * eigenvalue problems Hx = lambda M x, H M x = lambda x with H, M > 0.
17: *
18: IMPLICIT NONE
19: * .. Scalar Arguments ..
20: INTEGER INFO, LDA, LDV, LWORK, M, MV, N, NSWEEP
21: DOUBLE PRECISION EPS, SFMIN, TOL
22: CHARACTER*1 JOBV
23: * ..
24: * .. Array Arguments ..
25: DOUBLE PRECISION A( LDA, * ), SVA( N ), D( N ), V( LDV, * ),
26: + WORK( LWORK )
27: * ..
28: *
29: * Purpose
30: * =======
31: *
32: * DGSVJ0 is called from DGESVJ as a pre-processor and that is its main
33: * purpose. It applies Jacobi rotations in the same way as DGESVJ does, but
34: * it does not check convergence (stopping criterion). Few tuning
35: * parameters (marked by [TP]) are available for the implementer.
36: *
37: * Further Details
38: * ~~~~~~~~~~~~~~~
39: * DGSVJ0 is used just to enable SGESVJ to call a simplified version of
40: * itself to work on a submatrix of the original matrix.
41: *
42: * Contributors
43: * ~~~~~~~~~~~~
44: * Zlatko Drmac (Zagreb, Croatia) and Kresimir Veselic (Hagen, Germany)
45: *
46: * Bugs, Examples and Comments
47: * ~~~~~~~~~~~~~~~~~~~~~~~~~~~
48: * Please report all bugs and send interesting test examples and comments to
49: * drmac@math.hr. Thank you.
50: *
51: * Arguments
52: * =========
53: *
54: * JOBV (input) CHARACTER*1
55: * Specifies whether the output from this procedure is used
56: * to compute the matrix V:
57: * = 'V': the product of the Jacobi rotations is accumulated
58: * by postmulyiplying the N-by-N array V.
59: * (See the description of V.)
60: * = 'A': the product of the Jacobi rotations is accumulated
61: * by postmulyiplying the MV-by-N array V.
62: * (See the descriptions of MV and V.)
63: * = 'N': the Jacobi rotations are not accumulated.
64: *
65: * M (input) INTEGER
66: * The number of rows of the input matrix A. M >= 0.
67: *
68: * N (input) INTEGER
69: * The number of columns of the input matrix A.
70: * M >= N >= 0.
71: *
72: * A (input/output) DOUBLE PRECISION array, dimension (LDA,N)
73: * On entry, M-by-N matrix A, such that A*diag(D) represents
74: * the input matrix.
75: * On exit,
76: * A_onexit * D_onexit represents the input matrix A*diag(D)
77: * post-multiplied by a sequence of Jacobi rotations, where the
78: * rotation threshold and the total number of sweeps are given in
79: * TOL and NSWEEP, respectively.
80: * (See the descriptions of D, TOL and NSWEEP.)
81: *
82: * LDA (input) INTEGER
83: * The leading dimension of the array A. LDA >= max(1,M).
84: *
85: * D (input/workspace/output) DOUBLE PRECISION array, dimension (N)
86: * The array D accumulates the scaling factors from the fast scaled
87: * Jacobi rotations.
88: * On entry, A*diag(D) represents the input matrix.
89: * On exit, A_onexit*diag(D_onexit) represents the input matrix
90: * post-multiplied by a sequence of Jacobi rotations, where the
91: * rotation threshold and the total number of sweeps are given in
92: * TOL and NSWEEP, respectively.
93: * (See the descriptions of A, TOL and NSWEEP.)
94: *
95: * SVA (input/workspace/output) DOUBLE PRECISION array, dimension (N)
96: * On entry, SVA contains the Euclidean norms of the columns of
97: * the matrix A*diag(D).
98: * On exit, SVA contains the Euclidean norms of the columns of
99: * the matrix onexit*diag(D_onexit).
100: *
101: * MV (input) INTEGER
102: * If JOBV .EQ. 'A', then MV rows of V are post-multipled by a
103: * sequence of Jacobi rotations.
104: * If JOBV = 'N', then MV is not referenced.
105: *
106: * V (input/output) DOUBLE PRECISION array, dimension (LDV,N)
107: * If JOBV .EQ. 'V' then N rows of V are post-multipled by a
108: * sequence of Jacobi rotations.
109: * If JOBV .EQ. 'A' then MV rows of V are post-multipled by a
110: * sequence of Jacobi rotations.
111: * If JOBV = 'N', then V is not referenced.
112: *
113: * LDV (input) INTEGER
114: * The leading dimension of the array V, LDV >= 1.
115: * If JOBV = 'V', LDV .GE. N.
116: * If JOBV = 'A', LDV .GE. MV.
117: *
118: * EPS (input) DOUBLE PRECISION
119: * EPS = DLAMCH('Epsilon')
120: *
121: * SFMIN (input) DOUBLE PRECISION
122: * SFMIN = DLAMCH('Safe Minimum')
123: *
124: * TOL (input) DOUBLE PRECISION
125: * TOL is the threshold for Jacobi rotations. For a pair
126: * A(:,p), A(:,q) of pivot columns, the Jacobi rotation is
127: * applied only if DABS(COS(angle(A(:,p),A(:,q)))) .GT. TOL.
128: *
129: * NSWEEP (input) INTEGER
130: * NSWEEP is the number of sweeps of Jacobi rotations to be
131: * performed.
132: *
133: * WORK (workspace) DOUBLE PRECISION array, dimension (LWORK)
134: *
135: * LWORK (input) INTEGER
136: * LWORK is the dimension of WORK. LWORK .GE. M.
137: *
138: * INFO (output) INTEGER
139: * = 0 : successful exit.
140: * < 0 : if INFO = -i, then the i-th argument had an illegal value
141: *
142: * =====================================================================
143: *
144: * .. Local Parameters ..
145: DOUBLE PRECISION ZERO, HALF, ONE, TWO
146: PARAMETER ( ZERO = 0.0D0, HALF = 0.5D0, ONE = 1.0D0,
147: + TWO = 2.0D0 )
148: * ..
149: * .. Local Scalars ..
150: DOUBLE PRECISION AAPP, AAPP0, AAPQ, AAQQ, APOAQ, AQOAP, BIG,
151: + BIGTHETA, CS, MXAAPQ, MXSINJ, ROOTBIG, ROOTEPS,
152: + ROOTSFMIN, ROOTTOL, SMALL, SN, T, TEMP1, THETA,
153: + THSIGN
154: INTEGER BLSKIP, EMPTSW, i, ibr, IERR, igl, IJBLSK, ir1,
155: + ISWROT, jbc, jgl, KBL, LKAHEAD, MVL, NBL,
156: + NOTROT, p, PSKIPPED, q, ROWSKIP, SWBAND
157: LOGICAL APPLV, ROTOK, RSVEC
158: * ..
159: * .. Local Arrays ..
160: DOUBLE PRECISION FASTR( 5 )
161: * ..
162: * .. Intrinsic Functions ..
163: INTRINSIC DABS, DMAX1, DBLE, MIN0, DSIGN, DSQRT
164: * ..
165: * .. External Functions ..
166: DOUBLE PRECISION DDOT, DNRM2
167: INTEGER IDAMAX
168: LOGICAL LSAME
169: EXTERNAL IDAMAX, LSAME, DDOT, DNRM2
170: * ..
171: * .. External Subroutines ..
172: EXTERNAL DAXPY, DCOPY, DLASCL, DLASSQ, DROTM, DSWAP
173: * ..
174: * .. Executable Statements ..
175: *
176: APPLV = LSAME( JOBV, 'A' )
177: RSVEC = LSAME( JOBV, 'V' )
178: IF( .NOT.( RSVEC .OR. APPLV .OR. LSAME( JOBV, 'N' ) ) ) THEN
179: INFO = -1
180: ELSE IF( M.LT.0 ) THEN
181: INFO = -2
182: ELSE IF( ( N.LT.0 ) .OR. ( N.GT.M ) ) THEN
183: INFO = -3
184: ELSE IF( LDA.LT.M ) THEN
185: INFO = -5
186: ELSE IF( MV.LT.0 ) THEN
187: INFO = -8
188: ELSE IF( LDV.LT.M ) THEN
189: INFO = -10
190: ELSE IF( TOL.LE.EPS ) THEN
191: INFO = -13
192: ELSE IF( NSWEEP.LT.0 ) THEN
193: INFO = -14
194: ELSE IF( LWORK.LT.M ) THEN
195: INFO = -16
196: ELSE
197: INFO = 0
198: END IF
199: *
200: * #:(
201: IF( INFO.NE.0 ) THEN
202: CALL XERBLA( 'DGSVJ0', -INFO )
203: RETURN
204: END IF
205: *
206: IF( RSVEC ) THEN
207: MVL = N
208: ELSE IF( APPLV ) THEN
209: MVL = MV
210: END IF
211: RSVEC = RSVEC .OR. APPLV
212:
213: ROOTEPS = DSQRT( EPS )
214: ROOTSFMIN = DSQRT( SFMIN )
215: SMALL = SFMIN / EPS
216: BIG = ONE / SFMIN
217: ROOTBIG = ONE / ROOTSFMIN
218: BIGTHETA = ONE / ROOTEPS
219: ROOTTOL = DSQRT( TOL )
220: *
221: *
222: * -#- Row-cyclic Jacobi SVD algorithm with column pivoting -#-
223: *
224: EMPTSW = ( N*( N-1 ) ) / 2
225: NOTROT = 0
226: FASTR( 1 ) = ZERO
227: *
228: * -#- Row-cyclic pivot strategy with de Rijk's pivoting -#-
229: *
230:
231: SWBAND = 0
232: *[TP] SWBAND is a tuning parameter. It is meaningful and effective
233: * if SGESVJ is used as a computational routine in the preconditioned
234: * Jacobi SVD algorithm SGESVJ. For sweeps i=1:SWBAND the procedure
235: * ......
236:
237: KBL = MIN0( 8, N )
238: *[TP] KBL is a tuning parameter that defines the tile size in the
239: * tiling of the p-q loops of pivot pairs. In general, an optimal
240: * value of KBL depends on the matrix dimensions and on the
241: * parameters of the computer's memory.
242: *
243: NBL = N / KBL
244: IF( ( NBL*KBL ).NE.N )NBL = NBL + 1
245:
246: BLSKIP = ( KBL**2 ) + 1
247: *[TP] BLKSKIP is a tuning parameter that depends on SWBAND and KBL.
248:
249: ROWSKIP = MIN0( 5, KBL )
250: *[TP] ROWSKIP is a tuning parameter.
251:
252: LKAHEAD = 1
253: *[TP] LKAHEAD is a tuning parameter.
254: SWBAND = 0
255: PSKIPPED = 0
256: *
257: DO 1993 i = 1, NSWEEP
258: * .. go go go ...
259: *
260: MXAAPQ = ZERO
261: MXSINJ = ZERO
262: ISWROT = 0
263: *
264: NOTROT = 0
265: PSKIPPED = 0
266: *
267: DO 2000 ibr = 1, NBL
268:
269: igl = ( ibr-1 )*KBL + 1
270: *
271: DO 1002 ir1 = 0, MIN0( LKAHEAD, NBL-ibr )
272: *
273: igl = igl + ir1*KBL
274: *
275: DO 2001 p = igl, MIN0( igl+KBL-1, N-1 )
276:
277: * .. de Rijk's pivoting
278: q = IDAMAX( N-p+1, SVA( p ), 1 ) + p - 1
279: IF( p.NE.q ) THEN
280: CALL DSWAP( M, A( 1, p ), 1, A( 1, q ), 1 )
281: IF( RSVEC )CALL DSWAP( MVL, V( 1, p ), 1,
282: + V( 1, q ), 1 )
283: TEMP1 = SVA( p )
284: SVA( p ) = SVA( q )
285: SVA( q ) = TEMP1
286: TEMP1 = D( p )
287: D( p ) = D( q )
288: D( q ) = TEMP1
289: END IF
290: *
291: IF( ir1.EQ.0 ) THEN
292: *
293: * Column norms are periodically updated by explicit
294: * norm computation.
295: * Caveat:
296: * Some BLAS implementations compute DNRM2(M,A(1,p),1)
297: * as DSQRT(DDOT(M,A(1,p),1,A(1,p),1)), which may result in
298: * overflow for ||A(:,p)||_2 > DSQRT(overflow_threshold), and
299: * undeflow for ||A(:,p)||_2 < DSQRT(underflow_threshold).
300: * Hence, DNRM2 cannot be trusted, not even in the case when
301: * the true norm is far from the under(over)flow boundaries.
302: * If properly implemented DNRM2 is available, the IF-THEN-ELSE
303: * below should read "AAPP = DNRM2( M, A(1,p), 1 ) * D(p)".
304: *
305: IF( ( SVA( p ).LT.ROOTBIG ) .AND.
306: + ( SVA( p ).GT.ROOTSFMIN ) ) THEN
307: SVA( p ) = DNRM2( M, A( 1, p ), 1 )*D( p )
308: ELSE
309: TEMP1 = ZERO
310: AAPP = ZERO
311: CALL DLASSQ( M, A( 1, p ), 1, TEMP1, AAPP )
312: SVA( p ) = TEMP1*DSQRT( AAPP )*D( p )
313: END IF
314: AAPP = SVA( p )
315: ELSE
316: AAPP = SVA( p )
317: END IF
318:
319: *
320: IF( AAPP.GT.ZERO ) THEN
321: *
322: PSKIPPED = 0
323: *
324: DO 2002 q = p + 1, MIN0( igl+KBL-1, N )
325: *
326: AAQQ = SVA( q )
327:
328: IF( AAQQ.GT.ZERO ) THEN
329: *
330: AAPP0 = AAPP
331: IF( AAQQ.GE.ONE ) THEN
332: ROTOK = ( SMALL*AAPP ).LE.AAQQ
333: IF( AAPP.LT.( BIG / AAQQ ) ) THEN
334: AAPQ = ( DDOT( M, A( 1, p ), 1, A( 1,
335: + q ), 1 )*D( p )*D( q ) / AAQQ )
336: + / AAPP
337: ELSE
338: CALL DCOPY( M, A( 1, p ), 1, WORK, 1 )
339: CALL DLASCL( 'G', 0, 0, AAPP, D( p ),
340: + M, 1, WORK, LDA, IERR )
341: AAPQ = DDOT( M, WORK, 1, A( 1, q ),
342: + 1 )*D( q ) / AAQQ
343: END IF
344: ELSE
345: ROTOK = AAPP.LE.( AAQQ / SMALL )
346: IF( AAPP.GT.( SMALL / AAQQ ) ) THEN
347: AAPQ = ( DDOT( M, A( 1, p ), 1, A( 1,
348: + q ), 1 )*D( p )*D( q ) / AAQQ )
349: + / AAPP
350: ELSE
351: CALL DCOPY( M, A( 1, q ), 1, WORK, 1 )
352: CALL DLASCL( 'G', 0, 0, AAQQ, D( q ),
353: + M, 1, WORK, LDA, IERR )
354: AAPQ = DDOT( M, WORK, 1, A( 1, p ),
355: + 1 )*D( p ) / AAPP
356: END IF
357: END IF
358: *
359: MXAAPQ = DMAX1( MXAAPQ, DABS( AAPQ ) )
360: *
361: * TO rotate or NOT to rotate, THAT is the question ...
362: *
363: IF( DABS( AAPQ ).GT.TOL ) THEN
364: *
365: * .. rotate
366: * ROTATED = ROTATED + ONE
367: *
368: IF( ir1.EQ.0 ) THEN
369: NOTROT = 0
370: PSKIPPED = 0
371: ISWROT = ISWROT + 1
372: END IF
373: *
374: IF( ROTOK ) THEN
375: *
376: AQOAP = AAQQ / AAPP
377: APOAQ = AAPP / AAQQ
378: THETA = -HALF*DABS( AQOAP-APOAQ ) /
379: + AAPQ
380: *
381: IF( DABS( THETA ).GT.BIGTHETA ) THEN
382: *
383: T = HALF / THETA
384: FASTR( 3 ) = T*D( p ) / D( q )
385: FASTR( 4 ) = -T*D( q ) / D( p )
386: CALL DROTM( M, A( 1, p ), 1,
387: + A( 1, q ), 1, FASTR )
388: IF( RSVEC )CALL DROTM( MVL,
389: + V( 1, p ), 1,
390: + V( 1, q ), 1,
391: + FASTR )
392: SVA( q ) = AAQQ*DSQRT( DMAX1( ZERO,
393: + ONE+T*APOAQ*AAPQ ) )
394: AAPP = AAPP*DSQRT( ONE-T*AQOAP*
395: + AAPQ )
396: MXSINJ = DMAX1( MXSINJ, DABS( T ) )
397: *
398: ELSE
399: *
400: * .. choose correct signum for THETA and rotate
401: *
402: THSIGN = -DSIGN( ONE, AAPQ )
403: T = ONE / ( THETA+THSIGN*
404: + DSQRT( ONE+THETA*THETA ) )
405: CS = DSQRT( ONE / ( ONE+T*T ) )
406: SN = T*CS
407: *
408: MXSINJ = DMAX1( MXSINJ, DABS( SN ) )
409: SVA( q ) = AAQQ*DSQRT( DMAX1( ZERO,
410: + ONE+T*APOAQ*AAPQ ) )
411: AAPP = AAPP*DSQRT( DMAX1( ZERO,
412: + ONE-T*AQOAP*AAPQ ) )
413: *
414: APOAQ = D( p ) / D( q )
415: AQOAP = D( q ) / D( p )
416: IF( D( p ).GE.ONE ) THEN
417: IF( D( q ).GE.ONE ) THEN
418: FASTR( 3 ) = T*APOAQ
419: FASTR( 4 ) = -T*AQOAP
420: D( p ) = D( p )*CS
421: D( q ) = D( q )*CS
422: CALL DROTM( M, A( 1, p ), 1,
423: + A( 1, q ), 1,
424: + FASTR )
425: IF( RSVEC )CALL DROTM( MVL,
426: + V( 1, p ), 1, V( 1, q ),
427: + 1, FASTR )
428: ELSE
429: CALL DAXPY( M, -T*AQOAP,
430: + A( 1, q ), 1,
431: + A( 1, p ), 1 )
432: CALL DAXPY( M, CS*SN*APOAQ,
433: + A( 1, p ), 1,
434: + A( 1, q ), 1 )
435: D( p ) = D( p )*CS
436: D( q ) = D( q ) / CS
437: IF( RSVEC ) THEN
438: CALL DAXPY( MVL, -T*AQOAP,
439: + V( 1, q ), 1,
440: + V( 1, p ), 1 )
441: CALL DAXPY( MVL,
442: + CS*SN*APOAQ,
443: + V( 1, p ), 1,
444: + V( 1, q ), 1 )
445: END IF
446: END IF
447: ELSE
448: IF( D( q ).GE.ONE ) THEN
449: CALL DAXPY( M, T*APOAQ,
450: + A( 1, p ), 1,
451: + A( 1, q ), 1 )
452: CALL DAXPY( M, -CS*SN*AQOAP,
453: + A( 1, q ), 1,
454: + A( 1, p ), 1 )
455: D( p ) = D( p ) / CS
456: D( q ) = D( q )*CS
457: IF( RSVEC ) THEN
458: CALL DAXPY( MVL, T*APOAQ,
459: + V( 1, p ), 1,
460: + V( 1, q ), 1 )
461: CALL DAXPY( MVL,
462: + -CS*SN*AQOAP,
463: + V( 1, q ), 1,
464: + V( 1, p ), 1 )
465: END IF
466: ELSE
467: IF( D( p ).GE.D( q ) ) THEN
468: CALL DAXPY( M, -T*AQOAP,
469: + A( 1, q ), 1,
470: + A( 1, p ), 1 )
471: CALL DAXPY( M, CS*SN*APOAQ,
472: + A( 1, p ), 1,
473: + A( 1, q ), 1 )
474: D( p ) = D( p )*CS
475: D( q ) = D( q ) / CS
476: IF( RSVEC ) THEN
477: CALL DAXPY( MVL,
478: + -T*AQOAP,
479: + V( 1, q ), 1,
480: + V( 1, p ), 1 )
481: CALL DAXPY( MVL,
482: + CS*SN*APOAQ,
483: + V( 1, p ), 1,
484: + V( 1, q ), 1 )
485: END IF
486: ELSE
487: CALL DAXPY( M, T*APOAQ,
488: + A( 1, p ), 1,
489: + A( 1, q ), 1 )
490: CALL DAXPY( M,
491: + -CS*SN*AQOAP,
492: + A( 1, q ), 1,
493: + A( 1, p ), 1 )
494: D( p ) = D( p ) / CS
495: D( q ) = D( q )*CS
496: IF( RSVEC ) THEN
497: CALL DAXPY( MVL,
498: + T*APOAQ, V( 1, p ),
499: + 1, V( 1, q ), 1 )
500: CALL DAXPY( MVL,
501: + -CS*SN*AQOAP,
502: + V( 1, q ), 1,
503: + V( 1, p ), 1 )
504: END IF
505: END IF
506: END IF
507: END IF
508: END IF
509: *
510: ELSE
511: * .. have to use modified Gram-Schmidt like transformation
512: CALL DCOPY( M, A( 1, p ), 1, WORK, 1 )
513: CALL DLASCL( 'G', 0, 0, AAPP, ONE, M,
514: + 1, WORK, LDA, IERR )
515: CALL DLASCL( 'G', 0, 0, AAQQ, ONE, M,
516: + 1, A( 1, q ), LDA, IERR )
517: TEMP1 = -AAPQ*D( p ) / D( q )
518: CALL DAXPY( M, TEMP1, WORK, 1,
519: + A( 1, q ), 1 )
520: CALL DLASCL( 'G', 0, 0, ONE, AAQQ, M,
521: + 1, A( 1, q ), LDA, IERR )
522: SVA( q ) = AAQQ*DSQRT( DMAX1( ZERO,
523: + ONE-AAPQ*AAPQ ) )
524: MXSINJ = DMAX1( MXSINJ, SFMIN )
525: END IF
526: * END IF ROTOK THEN ... ELSE
527: *
528: * In the case of cancellation in updating SVA(q), SVA(p)
529: * recompute SVA(q), SVA(p).
530: IF( ( SVA( q ) / AAQQ )**2.LE.ROOTEPS )
531: + THEN
532: IF( ( AAQQ.LT.ROOTBIG ) .AND.
533: + ( AAQQ.GT.ROOTSFMIN ) ) THEN
534: SVA( q ) = DNRM2( M, A( 1, q ), 1 )*
535: + D( q )
536: ELSE
537: T = ZERO
538: AAQQ = ZERO
539: CALL DLASSQ( M, A( 1, q ), 1, T,
540: + AAQQ )
541: SVA( q ) = T*DSQRT( AAQQ )*D( q )
542: END IF
543: END IF
544: IF( ( AAPP / AAPP0 ).LE.ROOTEPS ) THEN
545: IF( ( AAPP.LT.ROOTBIG ) .AND.
546: + ( AAPP.GT.ROOTSFMIN ) ) THEN
547: AAPP = DNRM2( M, A( 1, p ), 1 )*
548: + D( p )
549: ELSE
550: T = ZERO
551: AAPP = ZERO
552: CALL DLASSQ( M, A( 1, p ), 1, T,
553: + AAPP )
554: AAPP = T*DSQRT( AAPP )*D( p )
555: END IF
556: SVA( p ) = AAPP
557: END IF
558: *
559: ELSE
560: * A(:,p) and A(:,q) already numerically orthogonal
561: IF( ir1.EQ.0 )NOTROT = NOTROT + 1
562: PSKIPPED = PSKIPPED + 1
563: END IF
564: ELSE
565: * A(:,q) is zero column
566: IF( ir1.EQ.0 )NOTROT = NOTROT + 1
567: PSKIPPED = PSKIPPED + 1
568: END IF
569: *
570: IF( ( i.LE.SWBAND ) .AND.
571: + ( PSKIPPED.GT.ROWSKIP ) ) THEN
572: IF( ir1.EQ.0 )AAPP = -AAPP
573: NOTROT = 0
574: GO TO 2103
575: END IF
576: *
577: 2002 CONTINUE
578: * END q-LOOP
579: *
580: 2103 CONTINUE
581: * bailed out of q-loop
582:
583: SVA( p ) = AAPP
584:
585: ELSE
586: SVA( p ) = AAPP
587: IF( ( ir1.EQ.0 ) .AND. ( AAPP.EQ.ZERO ) )
588: + NOTROT = NOTROT + MIN0( igl+KBL-1, N ) - p
589: END IF
590: *
591: 2001 CONTINUE
592: * end of the p-loop
593: * end of doing the block ( ibr, ibr )
594: 1002 CONTINUE
595: * end of ir1-loop
596: *
597: *........................................................
598: * ... go to the off diagonal blocks
599: *
600: igl = ( ibr-1 )*KBL + 1
601: *
602: DO 2010 jbc = ibr + 1, NBL
603: *
604: jgl = ( jbc-1 )*KBL + 1
605: *
606: * doing the block at ( ibr, jbc )
607: *
608: IJBLSK = 0
609: DO 2100 p = igl, MIN0( igl+KBL-1, N )
610: *
611: AAPP = SVA( p )
612: *
613: IF( AAPP.GT.ZERO ) THEN
614: *
615: PSKIPPED = 0
616: *
617: DO 2200 q = jgl, MIN0( jgl+KBL-1, N )
618: *
619: AAQQ = SVA( q )
620: *
621: IF( AAQQ.GT.ZERO ) THEN
622: AAPP0 = AAPP
623: *
624: * -#- M x 2 Jacobi SVD -#-
625: *
626: * -#- Safe Gram matrix computation -#-
627: *
628: IF( AAQQ.GE.ONE ) THEN
629: IF( AAPP.GE.AAQQ ) THEN
630: ROTOK = ( SMALL*AAPP ).LE.AAQQ
631: ELSE
632: ROTOK = ( SMALL*AAQQ ).LE.AAPP
633: END IF
634: IF( AAPP.LT.( BIG / AAQQ ) ) THEN
635: AAPQ = ( DDOT( M, A( 1, p ), 1, A( 1,
636: + q ), 1 )*D( p )*D( q ) / AAQQ )
637: + / AAPP
638: ELSE
639: CALL DCOPY( M, A( 1, p ), 1, WORK, 1 )
640: CALL DLASCL( 'G', 0, 0, AAPP, D( p ),
641: + M, 1, WORK, LDA, IERR )
642: AAPQ = DDOT( M, WORK, 1, A( 1, q ),
643: + 1 )*D( q ) / AAQQ
644: END IF
645: ELSE
646: IF( AAPP.GE.AAQQ ) THEN
647: ROTOK = AAPP.LE.( AAQQ / SMALL )
648: ELSE
649: ROTOK = AAQQ.LE.( AAPP / SMALL )
650: END IF
651: IF( AAPP.GT.( SMALL / AAQQ ) ) THEN
652: AAPQ = ( DDOT( M, A( 1, p ), 1, A( 1,
653: + q ), 1 )*D( p )*D( q ) / AAQQ )
654: + / AAPP
655: ELSE
656: CALL DCOPY( M, A( 1, q ), 1, WORK, 1 )
657: CALL DLASCL( 'G', 0, 0, AAQQ, D( q ),
658: + M, 1, WORK, LDA, IERR )
659: AAPQ = DDOT( M, WORK, 1, A( 1, p ),
660: + 1 )*D( p ) / AAPP
661: END IF
662: END IF
663: *
664: MXAAPQ = DMAX1( MXAAPQ, DABS( AAPQ ) )
665: *
666: * TO rotate or NOT to rotate, THAT is the question ...
667: *
668: IF( DABS( AAPQ ).GT.TOL ) THEN
669: NOTROT = 0
670: * ROTATED = ROTATED + 1
671: PSKIPPED = 0
672: ISWROT = ISWROT + 1
673: *
674: IF( ROTOK ) THEN
675: *
676: AQOAP = AAQQ / AAPP
677: APOAQ = AAPP / AAQQ
678: THETA = -HALF*DABS( AQOAP-APOAQ ) /
679: + AAPQ
680: IF( AAQQ.GT.AAPP0 )THETA = -THETA
681: *
682: IF( DABS( THETA ).GT.BIGTHETA ) THEN
683: T = HALF / THETA
684: FASTR( 3 ) = T*D( p ) / D( q )
685: FASTR( 4 ) = -T*D( q ) / D( p )
686: CALL DROTM( M, A( 1, p ), 1,
687: + A( 1, q ), 1, FASTR )
688: IF( RSVEC )CALL DROTM( MVL,
689: + V( 1, p ), 1,
690: + V( 1, q ), 1,
691: + FASTR )
692: SVA( q ) = AAQQ*DSQRT( DMAX1( ZERO,
693: + ONE+T*APOAQ*AAPQ ) )
694: AAPP = AAPP*DSQRT( DMAX1( ZERO,
695: + ONE-T*AQOAP*AAPQ ) )
696: MXSINJ = DMAX1( MXSINJ, DABS( T ) )
697: ELSE
698: *
699: * .. choose correct signum for THETA and rotate
700: *
701: THSIGN = -DSIGN( ONE, AAPQ )
702: IF( AAQQ.GT.AAPP0 )THSIGN = -THSIGN
703: T = ONE / ( THETA+THSIGN*
704: + DSQRT( ONE+THETA*THETA ) )
705: CS = DSQRT( ONE / ( ONE+T*T ) )
706: SN = T*CS
707: MXSINJ = DMAX1( MXSINJ, DABS( SN ) )
708: SVA( q ) = AAQQ*DSQRT( DMAX1( ZERO,
709: + ONE+T*APOAQ*AAPQ ) )
710: AAPP = AAPP*DSQRT( ONE-T*AQOAP*
711: + AAPQ )
712: *
713: APOAQ = D( p ) / D( q )
714: AQOAP = D( q ) / D( p )
715: IF( D( p ).GE.ONE ) THEN
716: *
717: IF( D( q ).GE.ONE ) THEN
718: FASTR( 3 ) = T*APOAQ
719: FASTR( 4 ) = -T*AQOAP
720: D( p ) = D( p )*CS
721: D( q ) = D( q )*CS
722: CALL DROTM( M, A( 1, p ), 1,
723: + A( 1, q ), 1,
724: + FASTR )
725: IF( RSVEC )CALL DROTM( MVL,
726: + V( 1, p ), 1, V( 1, q ),
727: + 1, FASTR )
728: ELSE
729: CALL DAXPY( M, -T*AQOAP,
730: + A( 1, q ), 1,
731: + A( 1, p ), 1 )
732: CALL DAXPY( M, CS*SN*APOAQ,
733: + A( 1, p ), 1,
734: + A( 1, q ), 1 )
735: IF( RSVEC ) THEN
736: CALL DAXPY( MVL, -T*AQOAP,
737: + V( 1, q ), 1,
738: + V( 1, p ), 1 )
739: CALL DAXPY( MVL,
740: + CS*SN*APOAQ,
741: + V( 1, p ), 1,
742: + V( 1, q ), 1 )
743: END IF
744: D( p ) = D( p )*CS
745: D( q ) = D( q ) / CS
746: END IF
747: ELSE
748: IF( D( q ).GE.ONE ) THEN
749: CALL DAXPY( M, T*APOAQ,
750: + A( 1, p ), 1,
751: + A( 1, q ), 1 )
752: CALL DAXPY( M, -CS*SN*AQOAP,
753: + A( 1, q ), 1,
754: + A( 1, p ), 1 )
755: IF( RSVEC ) THEN
756: CALL DAXPY( MVL, T*APOAQ,
757: + V( 1, p ), 1,
758: + V( 1, q ), 1 )
759: CALL DAXPY( MVL,
760: + -CS*SN*AQOAP,
761: + V( 1, q ), 1,
762: + V( 1, p ), 1 )
763: END IF
764: D( p ) = D( p ) / CS
765: D( q ) = D( q )*CS
766: ELSE
767: IF( D( p ).GE.D( q ) ) THEN
768: CALL DAXPY( M, -T*AQOAP,
769: + A( 1, q ), 1,
770: + A( 1, p ), 1 )
771: CALL DAXPY( M, CS*SN*APOAQ,
772: + A( 1, p ), 1,
773: + A( 1, q ), 1 )
774: D( p ) = D( p )*CS
775: D( q ) = D( q ) / CS
776: IF( RSVEC ) THEN
777: CALL DAXPY( MVL,
778: + -T*AQOAP,
779: + V( 1, q ), 1,
780: + V( 1, p ), 1 )
781: CALL DAXPY( MVL,
782: + CS*SN*APOAQ,
783: + V( 1, p ), 1,
784: + V( 1, q ), 1 )
785: END IF
786: ELSE
787: CALL DAXPY( M, T*APOAQ,
788: + A( 1, p ), 1,
789: + A( 1, q ), 1 )
790: CALL DAXPY( M,
791: + -CS*SN*AQOAP,
792: + A( 1, q ), 1,
793: + A( 1, p ), 1 )
794: D( p ) = D( p ) / CS
795: D( q ) = D( q )*CS
796: IF( RSVEC ) THEN
797: CALL DAXPY( MVL,
798: + T*APOAQ, V( 1, p ),
799: + 1, V( 1, q ), 1 )
800: CALL DAXPY( MVL,
801: + -CS*SN*AQOAP,
802: + V( 1, q ), 1,
803: + V( 1, p ), 1 )
804: END IF
805: END IF
806: END IF
807: END IF
808: END IF
809: *
810: ELSE
811: IF( AAPP.GT.AAQQ ) THEN
812: CALL DCOPY( M, A( 1, p ), 1, WORK,
813: + 1 )
814: CALL DLASCL( 'G', 0, 0, AAPP, ONE,
815: + M, 1, WORK, LDA, IERR )
816: CALL DLASCL( 'G', 0, 0, AAQQ, ONE,
817: + M, 1, A( 1, q ), LDA,
818: + IERR )
819: TEMP1 = -AAPQ*D( p ) / D( q )
820: CALL DAXPY( M, TEMP1, WORK, 1,
821: + A( 1, q ), 1 )
822: CALL DLASCL( 'G', 0, 0, ONE, AAQQ,
823: + M, 1, A( 1, q ), LDA,
824: + IERR )
825: SVA( q ) = AAQQ*DSQRT( DMAX1( ZERO,
826: + ONE-AAPQ*AAPQ ) )
827: MXSINJ = DMAX1( MXSINJ, SFMIN )
828: ELSE
829: CALL DCOPY( M, A( 1, q ), 1, WORK,
830: + 1 )
831: CALL DLASCL( 'G', 0, 0, AAQQ, ONE,
832: + M, 1, WORK, LDA, IERR )
833: CALL DLASCL( 'G', 0, 0, AAPP, ONE,
834: + M, 1, A( 1, p ), LDA,
835: + IERR )
836: TEMP1 = -AAPQ*D( q ) / D( p )
837: CALL DAXPY( M, TEMP1, WORK, 1,
838: + A( 1, p ), 1 )
839: CALL DLASCL( 'G', 0, 0, ONE, AAPP,
840: + M, 1, A( 1, p ), LDA,
841: + IERR )
842: SVA( p ) = AAPP*DSQRT( DMAX1( ZERO,
843: + ONE-AAPQ*AAPQ ) )
844: MXSINJ = DMAX1( MXSINJ, SFMIN )
845: END IF
846: END IF
847: * END IF ROTOK THEN ... ELSE
848: *
849: * In the case of cancellation in updating SVA(q)
850: * .. recompute SVA(q)
851: IF( ( SVA( q ) / AAQQ )**2.LE.ROOTEPS )
852: + THEN
853: IF( ( AAQQ.LT.ROOTBIG ) .AND.
854: + ( AAQQ.GT.ROOTSFMIN ) ) THEN
855: SVA( q ) = DNRM2( M, A( 1, q ), 1 )*
856: + D( q )
857: ELSE
858: T = ZERO
859: AAQQ = ZERO
860: CALL DLASSQ( M, A( 1, q ), 1, T,
861: + AAQQ )
862: SVA( q ) = T*DSQRT( AAQQ )*D( q )
863: END IF
864: END IF
865: IF( ( AAPP / AAPP0 )**2.LE.ROOTEPS ) THEN
866: IF( ( AAPP.LT.ROOTBIG ) .AND.
867: + ( AAPP.GT.ROOTSFMIN ) ) THEN
868: AAPP = DNRM2( M, A( 1, p ), 1 )*
869: + D( p )
870: ELSE
871: T = ZERO
872: AAPP = ZERO
873: CALL DLASSQ( M, A( 1, p ), 1, T,
874: + AAPP )
875: AAPP = T*DSQRT( AAPP )*D( p )
876: END IF
877: SVA( p ) = AAPP
878: END IF
879: * end of OK rotation
880: ELSE
881: NOTROT = NOTROT + 1
882: PSKIPPED = PSKIPPED + 1
883: IJBLSK = IJBLSK + 1
884: END IF
885: ELSE
886: NOTROT = NOTROT + 1
887: PSKIPPED = PSKIPPED + 1
888: IJBLSK = IJBLSK + 1
889: END IF
890: *
891: IF( ( i.LE.SWBAND ) .AND. ( IJBLSK.GE.BLSKIP ) )
892: + THEN
893: SVA( p ) = AAPP
894: NOTROT = 0
895: GO TO 2011
896: END IF
897: IF( ( i.LE.SWBAND ) .AND.
898: + ( PSKIPPED.GT.ROWSKIP ) ) THEN
899: AAPP = -AAPP
900: NOTROT = 0
901: GO TO 2203
902: END IF
903: *
904: 2200 CONTINUE
905: * end of the q-loop
906: 2203 CONTINUE
907: *
908: SVA( p ) = AAPP
909: *
910: ELSE
911: IF( AAPP.EQ.ZERO )NOTROT = NOTROT +
912: + MIN0( jgl+KBL-1, N ) - jgl + 1
913: IF( AAPP.LT.ZERO )NOTROT = 0
914: END IF
915:
916: 2100 CONTINUE
917: * end of the p-loop
918: 2010 CONTINUE
919: * end of the jbc-loop
920: 2011 CONTINUE
921: *2011 bailed out of the jbc-loop
922: DO 2012 p = igl, MIN0( igl+KBL-1, N )
923: SVA( p ) = DABS( SVA( p ) )
924: 2012 CONTINUE
925: *
926: 2000 CONTINUE
927: *2000 :: end of the ibr-loop
928: *
929: * .. update SVA(N)
930: IF( ( SVA( N ).LT.ROOTBIG ) .AND. ( SVA( N ).GT.ROOTSFMIN ) )
931: + THEN
932: SVA( N ) = DNRM2( M, A( 1, N ), 1 )*D( N )
933: ELSE
934: T = ZERO
935: AAPP = ZERO
936: CALL DLASSQ( M, A( 1, N ), 1, T, AAPP )
937: SVA( N ) = T*DSQRT( AAPP )*D( N )
938: END IF
939: *
940: * Additional steering devices
941: *
942: IF( ( i.LT.SWBAND ) .AND. ( ( MXAAPQ.LE.ROOTTOL ) .OR.
943: + ( ISWROT.LE.N ) ) )SWBAND = i
944: *
945: IF( ( i.GT.SWBAND+1 ) .AND. ( MXAAPQ.LT.DBLE( N )*TOL ) .AND.
946: + ( DBLE( N )*MXAAPQ*MXSINJ.LT.TOL ) ) THEN
947: GO TO 1994
948: END IF
949: *
950: IF( NOTROT.GE.EMPTSW )GO TO 1994
951:
952: 1993 CONTINUE
953: * end i=1:NSWEEP loop
954: * #:) Reaching this point means that the procedure has comleted the given
955: * number of iterations.
956: INFO = NSWEEP - 1
957: GO TO 1995
958: 1994 CONTINUE
959: * #:) Reaching this point means that during the i-th sweep all pivots were
960: * below the given tolerance, causing early exit.
961: *
962: INFO = 0
963: * #:) INFO = 0 confirms successful iterations.
964: 1995 CONTINUE
965: *
966: * Sort the vector D.
967: DO 5991 p = 1, N - 1
968: q = IDAMAX( N-p+1, SVA( p ), 1 ) + p - 1
969: IF( p.NE.q ) THEN
970: TEMP1 = SVA( p )
971: SVA( p ) = SVA( q )
972: SVA( q ) = TEMP1
973: TEMP1 = D( p )
974: D( p ) = D( q )
975: D( q ) = TEMP1
976: CALL DSWAP( M, A( 1, p ), 1, A( 1, q ), 1 )
977: IF( RSVEC )CALL DSWAP( MVL, V( 1, p ), 1, V( 1, q ), 1 )
978: END IF
979: 5991 CONTINUE
980: *
981: RETURN
982: * ..
983: * .. END OF DGSVJ0
984: * ..
985: END
CVSweb interface <joel.bertrand@systella.fr>