Annotation of rpl/lapack/lapack/dgsvj1.f, revision 1.7
1.7 ! bertrand 1: *> \brief \b DGSVJ1
! 2: *
! 3: * =========== DOCUMENTATION ===========
! 4: *
! 5: * Online html documentation available at
! 6: * http://www.netlib.org/lapack/explore-html/
! 7: *
! 8: *> \htmlonly
! 9: *> Download DGSVJ1 + dependencies
! 10: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dgsvj1.f">
! 11: *> [TGZ]</a>
! 12: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dgsvj1.f">
! 13: *> [ZIP]</a>
! 14: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dgsvj1.f">
! 15: *> [TXT]</a>
! 16: *> \endhtmlonly
! 17: *
! 18: * Definition:
! 19: * ===========
! 20: *
! 21: * SUBROUTINE DGSVJ1( JOBV, M, N, N1, A, LDA, D, SVA, MV, V, LDV,
! 22: * EPS, SFMIN, TOL, NSWEEP, WORK, LWORK, INFO )
! 23: *
! 24: * .. Scalar Arguments ..
! 25: * DOUBLE PRECISION EPS, SFMIN, TOL
! 26: * INTEGER INFO, LDA, LDV, LWORK, M, MV, N, N1, NSWEEP
! 27: * CHARACTER*1 JOBV
! 28: * ..
! 29: * .. Array Arguments ..
! 30: * DOUBLE PRECISION A( LDA, * ), D( N ), SVA( N ), V( LDV, * ),
! 31: * $ WORK( LWORK )
! 32: * ..
! 33: *
! 34: *
! 35: *> \par Purpose:
! 36: * =============
! 37: *>
! 38: *> \verbatim
! 39: *>
! 40: *> DGSVJ1 is called from SGESVJ as a pre-processor and that is its main
! 41: *> purpose. It applies Jacobi rotations in the same way as SGESVJ does, but
! 42: *> it targets only particular pivots and it does not check convergence
! 43: *> (stopping criterion). Few tunning parameters (marked by [TP]) are
! 44: *> available for the implementer.
! 45: *>
! 46: *> Further Details
! 47: *> ~~~~~~~~~~~~~~~
! 48: *> DGSVJ1 applies few sweeps of Jacobi rotations in the column space of
! 49: *> the input M-by-N matrix A. The pivot pairs are taken from the (1,2)
! 50: *> off-diagonal block in the corresponding N-by-N Gram matrix A^T * A. The
! 51: *> block-entries (tiles) of the (1,2) off-diagonal block are marked by the
! 52: *> [x]'s in the following scheme:
! 53: *>
! 54: *> | * * * [x] [x] [x]|
! 55: *> | * * * [x] [x] [x]| Row-cycling in the nblr-by-nblc [x] blocks.
! 56: *> | * * * [x] [x] [x]| Row-cyclic pivoting inside each [x] block.
! 57: *> |[x] [x] [x] * * * |
! 58: *> |[x] [x] [x] * * * |
! 59: *> |[x] [x] [x] * * * |
! 60: *>
! 61: *> In terms of the columns of A, the first N1 columns are rotated 'against'
! 62: *> the remaining N-N1 columns, trying to increase the angle between the
! 63: *> corresponding subspaces. The off-diagonal block is N1-by(N-N1) and it is
! 64: *> tiled using quadratic tiles of side KBL. Here, KBL is a tunning parmeter.
! 65: *> The number of sweeps is given in NSWEEP and the orthogonality threshold
! 66: *> is given in TOL.
! 67: *> \endverbatim
! 68: *
! 69: * Arguments:
! 70: * ==========
! 71: *
! 72: *> \param[in] JOBV
! 73: *> \verbatim
! 74: *> JOBV is CHARACTER*1
! 75: *> Specifies whether the output from this procedure is used
! 76: *> to compute the matrix V:
! 77: *> = 'V': the product of the Jacobi rotations is accumulated
! 78: *> by postmulyiplying the N-by-N array V.
! 79: *> (See the description of V.)
! 80: *> = 'A': the product of the Jacobi rotations is accumulated
! 81: *> by postmulyiplying the MV-by-N array V.
! 82: *> (See the descriptions of MV and V.)
! 83: *> = 'N': the Jacobi rotations are not accumulated.
! 84: *> \endverbatim
! 85: *>
! 86: *> \param[in] M
! 87: *> \verbatim
! 88: *> M is INTEGER
! 89: *> The number of rows of the input matrix A. M >= 0.
! 90: *> \endverbatim
! 91: *>
! 92: *> \param[in] N
! 93: *> \verbatim
! 94: *> N is INTEGER
! 95: *> The number of columns of the input matrix A.
! 96: *> M >= N >= 0.
! 97: *> \endverbatim
! 98: *>
! 99: *> \param[in] N1
! 100: *> \verbatim
! 101: *> N1 is INTEGER
! 102: *> N1 specifies the 2 x 2 block partition, the first N1 columns are
! 103: *> rotated 'against' the remaining N-N1 columns of A.
! 104: *> \endverbatim
! 105: *>
! 106: *> \param[in,out] A
! 107: *> \verbatim
! 108: *> A is DOUBLE PRECISION array, dimension (LDA,N)
! 109: *> On entry, M-by-N matrix A, such that A*diag(D) represents
! 110: *> the input matrix.
! 111: *> On exit,
! 112: *> A_onexit * D_onexit represents the input matrix A*diag(D)
! 113: *> post-multiplied by a sequence of Jacobi rotations, where the
! 114: *> rotation threshold and the total number of sweeps are given in
! 115: *> TOL and NSWEEP, respectively.
! 116: *> (See the descriptions of N1, D, TOL and NSWEEP.)
! 117: *> \endverbatim
! 118: *>
! 119: *> \param[in] LDA
! 120: *> \verbatim
! 121: *> LDA is INTEGER
! 122: *> The leading dimension of the array A. LDA >= max(1,M).
! 123: *> \endverbatim
! 124: *>
! 125: *> \param[in,out] D
! 126: *> \verbatim
! 127: *> D is DOUBLE PRECISION array, dimension (N)
! 128: *> The array D accumulates the scaling factors from the fast scaled
! 129: *> Jacobi rotations.
! 130: *> On entry, A*diag(D) represents the input matrix.
! 131: *> On exit, A_onexit*diag(D_onexit) represents the input matrix
! 132: *> post-multiplied by a sequence of Jacobi rotations, where the
! 133: *> rotation threshold and the total number of sweeps are given in
! 134: *> TOL and NSWEEP, respectively.
! 135: *> (See the descriptions of N1, A, TOL and NSWEEP.)
! 136: *> \endverbatim
! 137: *>
! 138: *> \param[in,out] SVA
! 139: *> \verbatim
! 140: *> SVA is DOUBLE PRECISION array, dimension (N)
! 141: *> On entry, SVA contains the Euclidean norms of the columns of
! 142: *> the matrix A*diag(D).
! 143: *> On exit, SVA contains the Euclidean norms of the columns of
! 144: *> the matrix onexit*diag(D_onexit).
! 145: *> \endverbatim
! 146: *>
! 147: *> \param[in] MV
! 148: *> \verbatim
! 149: *> MV is INTEGER
! 150: *> If JOBV .EQ. 'A', then MV rows of V are post-multipled by a
! 151: *> sequence of Jacobi rotations.
! 152: *> If JOBV = 'N', then MV is not referenced.
! 153: *> \endverbatim
! 154: *>
! 155: *> \param[in,out] V
! 156: *> \verbatim
! 157: *> V is DOUBLE PRECISION array, dimension (LDV,N)
! 158: *> If JOBV .EQ. 'V' then N rows of V are post-multipled by a
! 159: *> sequence of Jacobi rotations.
! 160: *> If JOBV .EQ. 'A' then MV rows of V are post-multipled by a
! 161: *> sequence of Jacobi rotations.
! 162: *> If JOBV = 'N', then V is not referenced.
! 163: *> \endverbatim
! 164: *>
! 165: *> \param[in] LDV
! 166: *> \verbatim
! 167: *> LDV is INTEGER
! 168: *> The leading dimension of the array V, LDV >= 1.
! 169: *> If JOBV = 'V', LDV .GE. N.
! 170: *> If JOBV = 'A', LDV .GE. MV.
! 171: *> \endverbatim
! 172: *>
! 173: *> \param[in] EPS
! 174: *> \verbatim
! 175: *> EPS is DOUBLE PRECISION
! 176: *> EPS = DLAMCH('Epsilon')
! 177: *> \endverbatim
! 178: *>
! 179: *> \param[in] SFMIN
! 180: *> \verbatim
! 181: *> SFMIN is DOUBLE PRECISION
! 182: *> SFMIN = DLAMCH('Safe Minimum')
! 183: *> \endverbatim
! 184: *>
! 185: *> \param[in] TOL
! 186: *> \verbatim
! 187: *> TOL is DOUBLE PRECISION
! 188: *> TOL is the threshold for Jacobi rotations. For a pair
! 189: *> A(:,p), A(:,q) of pivot columns, the Jacobi rotation is
! 190: *> applied only if DABS(COS(angle(A(:,p),A(:,q)))) .GT. TOL.
! 191: *> \endverbatim
! 192: *>
! 193: *> \param[in] NSWEEP
! 194: *> \verbatim
! 195: *> NSWEEP is INTEGER
! 196: *> NSWEEP is the number of sweeps of Jacobi rotations to be
! 197: *> performed.
! 198: *> \endverbatim
! 199: *>
! 200: *> \param[out] WORK
! 201: *> \verbatim
! 202: *> WORK is DOUBLE PRECISION array, dimension (LWORK)
! 203: *> \endverbatim
! 204: *>
! 205: *> \param[in] LWORK
! 206: *> \verbatim
! 207: *> LWORK is INTEGER
! 208: *> LWORK is the dimension of WORK. LWORK .GE. M.
! 209: *> \endverbatim
! 210: *>
! 211: *> \param[out] INFO
! 212: *> \verbatim
! 213: *> INFO is INTEGER
! 214: *> = 0 : successful exit.
! 215: *> < 0 : if INFO = -i, then the i-th argument had an illegal value
! 216: *> \endverbatim
! 217: *
! 218: * Authors:
! 219: * ========
! 220: *
! 221: *> \author Univ. of Tennessee
! 222: *> \author Univ. of California Berkeley
! 223: *> \author Univ. of Colorado Denver
! 224: *> \author NAG Ltd.
! 225: *
! 226: *> \date November 2011
! 227: *
! 228: *> \ingroup doubleOTHERcomputational
! 229: *
! 230: *> \par Contributors:
! 231: * ==================
! 232: *>
! 233: *> Zlatko Drmac (Zagreb, Croatia) and Kresimir Veselic (Hagen, Germany)
! 234: *
! 235: * =====================================================================
1.1 bertrand 236: SUBROUTINE DGSVJ1( JOBV, M, N, N1, A, LDA, D, SVA, MV, V, LDV,
1.6 bertrand 237: $ EPS, SFMIN, TOL, NSWEEP, WORK, LWORK, INFO )
1.1 bertrand 238: *
1.7 ! bertrand 239: * -- LAPACK computational routine (version 3.4.0) --
1.1 bertrand 240: * -- LAPACK is a software package provided by Univ. of Tennessee, --
241: * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
1.7 ! bertrand 242: * November 2011
1.1 bertrand 243: *
244: * .. Scalar Arguments ..
245: DOUBLE PRECISION EPS, SFMIN, TOL
246: INTEGER INFO, LDA, LDV, LWORK, M, MV, N, N1, NSWEEP
247: CHARACTER*1 JOBV
248: * ..
249: * .. Array Arguments ..
250: DOUBLE PRECISION A( LDA, * ), D( N ), SVA( N ), V( LDV, * ),
1.6 bertrand 251: $ WORK( LWORK )
1.1 bertrand 252: * ..
253: *
254: * =====================================================================
255: *
256: * .. Local Parameters ..
257: DOUBLE PRECISION ZERO, HALF, ONE, TWO
258: PARAMETER ( ZERO = 0.0D0, HALF = 0.5D0, ONE = 1.0D0,
1.6 bertrand 259: $ TWO = 2.0D0 )
1.1 bertrand 260: * ..
261: * .. Local Scalars ..
262: DOUBLE PRECISION AAPP, AAPP0, AAPQ, AAQQ, APOAQ, AQOAP, BIG,
1.6 bertrand 263: $ BIGTHETA, CS, LARGE, MXAAPQ, MXSINJ, ROOTBIG,
264: $ ROOTEPS, ROOTSFMIN, ROOTTOL, SMALL, SN, T,
265: $ TEMP1, THETA, THSIGN
1.1 bertrand 266: INTEGER BLSKIP, EMPTSW, i, ibr, igl, IERR, IJBLSK,
1.6 bertrand 267: $ ISWROT, jbc, jgl, KBL, MVL, NOTROT, nblc, nblr,
268: $ p, PSKIPPED, q, ROWSKIP, SWBAND
1.1 bertrand 269: LOGICAL APPLV, ROTOK, RSVEC
270: * ..
271: * .. Local Arrays ..
272: DOUBLE PRECISION FASTR( 5 )
273: * ..
274: * .. Intrinsic Functions ..
275: INTRINSIC DABS, DMAX1, DBLE, MIN0, DSIGN, DSQRT
276: * ..
277: * .. External Functions ..
278: DOUBLE PRECISION DDOT, DNRM2
279: INTEGER IDAMAX
280: LOGICAL LSAME
281: EXTERNAL IDAMAX, LSAME, DDOT, DNRM2
282: * ..
283: * .. External Subroutines ..
284: EXTERNAL DAXPY, DCOPY, DLASCL, DLASSQ, DROTM, DSWAP
285: * ..
286: * .. Executable Statements ..
287: *
288: * Test the input parameters.
289: *
290: APPLV = LSAME( JOBV, 'A' )
291: RSVEC = LSAME( JOBV, 'V' )
292: IF( .NOT.( RSVEC .OR. APPLV .OR. LSAME( JOBV, 'N' ) ) ) THEN
293: INFO = -1
294: ELSE IF( M.LT.0 ) THEN
295: INFO = -2
296: ELSE IF( ( N.LT.0 ) .OR. ( N.GT.M ) ) THEN
297: INFO = -3
298: ELSE IF( N1.LT.0 ) THEN
299: INFO = -4
300: ELSE IF( LDA.LT.M ) THEN
301: INFO = -6
1.4 bertrand 302: ELSE IF( ( RSVEC.OR.APPLV ) .AND. ( MV.LT.0 ) ) THEN
1.1 bertrand 303: INFO = -9
1.4 bertrand 304: ELSE IF( ( RSVEC.AND.( LDV.LT.N ) ).OR.
1.6 bertrand 305: $ ( APPLV.AND.( LDV.LT.MV ) ) ) THEN
1.1 bertrand 306: INFO = -11
307: ELSE IF( TOL.LE.EPS ) THEN
308: INFO = -14
309: ELSE IF( NSWEEP.LT.0 ) THEN
310: INFO = -15
311: ELSE IF( LWORK.LT.M ) THEN
312: INFO = -17
313: ELSE
314: INFO = 0
315: END IF
316: *
317: * #:(
318: IF( INFO.NE.0 ) THEN
319: CALL XERBLA( 'DGSVJ1', -INFO )
320: RETURN
321: END IF
322: *
323: IF( RSVEC ) THEN
324: MVL = N
325: ELSE IF( APPLV ) THEN
326: MVL = MV
327: END IF
328: RSVEC = RSVEC .OR. APPLV
329:
330: ROOTEPS = DSQRT( EPS )
331: ROOTSFMIN = DSQRT( SFMIN )
332: SMALL = SFMIN / EPS
333: BIG = ONE / SFMIN
334: ROOTBIG = ONE / ROOTSFMIN
335: LARGE = BIG / DSQRT( DBLE( M*N ) )
336: BIGTHETA = ONE / ROOTEPS
337: ROOTTOL = DSQRT( TOL )
338: *
339: * .. Initialize the right singular vector matrix ..
340: *
341: * RSVEC = LSAME( JOBV, 'Y' )
342: *
343: EMPTSW = N1*( N-N1 )
344: NOTROT = 0
345: FASTR( 1 ) = ZERO
346: *
347: * .. Row-cyclic pivot strategy with de Rijk's pivoting ..
348: *
349: KBL = MIN0( 8, N )
350: NBLR = N1 / KBL
351: IF( ( NBLR*KBL ).NE.N1 )NBLR = NBLR + 1
352:
353: * .. the tiling is nblr-by-nblc [tiles]
354:
355: NBLC = ( N-N1 ) / KBL
356: IF( ( NBLC*KBL ).NE.( N-N1 ) )NBLC = NBLC + 1
357: BLSKIP = ( KBL**2 ) + 1
358: *[TP] BLKSKIP is a tuning parameter that depends on SWBAND and KBL.
359:
360: ROWSKIP = MIN0( 5, KBL )
361: *[TP] ROWSKIP is a tuning parameter.
362: SWBAND = 0
363: *[TP] SWBAND is a tuning parameter. It is meaningful and effective
364: * if SGESVJ is used as a computational routine in the preconditioned
365: * Jacobi SVD algorithm SGESVJ.
366: *
367: *
368: * | * * * [x] [x] [x]|
369: * | * * * [x] [x] [x]| Row-cycling in the nblr-by-nblc [x] blocks.
370: * | * * * [x] [x] [x]| Row-cyclic pivoting inside each [x] block.
371: * |[x] [x] [x] * * * |
372: * |[x] [x] [x] * * * |
373: * |[x] [x] [x] * * * |
374: *
375: *
376: DO 1993 i = 1, NSWEEP
377: * .. go go go ...
378: *
379: MXAAPQ = ZERO
380: MXSINJ = ZERO
381: ISWROT = 0
382: *
383: NOTROT = 0
384: PSKIPPED = 0
385: *
386: DO 2000 ibr = 1, NBLR
387:
388: igl = ( ibr-1 )*KBL + 1
389: *
390: *
391: *........................................................
392: * ... go to the off diagonal blocks
393:
394: igl = ( ibr-1 )*KBL + 1
395:
396: DO 2010 jbc = 1, NBLC
397:
398: jgl = N1 + ( jbc-1 )*KBL + 1
399:
400: * doing the block at ( ibr, jbc )
401:
402: IJBLSK = 0
403: DO 2100 p = igl, MIN0( igl+KBL-1, N1 )
404:
405: AAPP = SVA( p )
406:
407: IF( AAPP.GT.ZERO ) THEN
408:
409: PSKIPPED = 0
410:
411: DO 2200 q = jgl, MIN0( jgl+KBL-1, N )
412: *
413: AAQQ = SVA( q )
414:
415: IF( AAQQ.GT.ZERO ) THEN
416: AAPP0 = AAPP
417: *
418: * .. M x 2 Jacobi SVD ..
419: *
420: * .. Safe Gram matrix computation ..
421: *
422: IF( AAQQ.GE.ONE ) THEN
423: IF( AAPP.GE.AAQQ ) THEN
424: ROTOK = ( SMALL*AAPP ).LE.AAQQ
425: ELSE
426: ROTOK = ( SMALL*AAQQ ).LE.AAPP
427: END IF
428: IF( AAPP.LT.( BIG / AAQQ ) ) THEN
429: AAPQ = ( DDOT( M, A( 1, p ), 1, A( 1,
1.6 bertrand 430: $ q ), 1 )*D( p )*D( q ) / AAQQ )
431: $ / AAPP
1.1 bertrand 432: ELSE
433: CALL DCOPY( M, A( 1, p ), 1, WORK, 1 )
434: CALL DLASCL( 'G', 0, 0, AAPP, D( p ),
1.6 bertrand 435: $ M, 1, WORK, LDA, IERR )
1.1 bertrand 436: AAPQ = DDOT( M, WORK, 1, A( 1, q ),
1.6 bertrand 437: $ 1 )*D( q ) / AAQQ
1.1 bertrand 438: END IF
439: ELSE
440: IF( AAPP.GE.AAQQ ) THEN
441: ROTOK = AAPP.LE.( AAQQ / SMALL )
442: ELSE
443: ROTOK = AAQQ.LE.( AAPP / SMALL )
444: END IF
445: IF( AAPP.GT.( SMALL / AAQQ ) ) THEN
446: AAPQ = ( DDOT( M, A( 1, p ), 1, A( 1,
1.6 bertrand 447: $ q ), 1 )*D( p )*D( q ) / AAQQ )
448: $ / AAPP
1.1 bertrand 449: ELSE
450: CALL DCOPY( M, A( 1, q ), 1, WORK, 1 )
451: CALL DLASCL( 'G', 0, 0, AAQQ, D( q ),
1.6 bertrand 452: $ M, 1, WORK, LDA, IERR )
1.1 bertrand 453: AAPQ = DDOT( M, WORK, 1, A( 1, p ),
1.6 bertrand 454: $ 1 )*D( p ) / AAPP
1.1 bertrand 455: END IF
456: END IF
457:
458: MXAAPQ = DMAX1( MXAAPQ, DABS( AAPQ ) )
459:
460: * TO rotate or NOT to rotate, THAT is the question ...
461: *
462: IF( DABS( AAPQ ).GT.TOL ) THEN
463: NOTROT = 0
464: * ROTATED = ROTATED + 1
465: PSKIPPED = 0
466: ISWROT = ISWROT + 1
467: *
468: IF( ROTOK ) THEN
469: *
470: AQOAP = AAQQ / AAPP
471: APOAQ = AAPP / AAQQ
1.6 bertrand 472: THETA = -HALF*DABS(AQOAP-APOAQ) / AAPQ
1.1 bertrand 473: IF( AAQQ.GT.AAPP0 )THETA = -THETA
474:
475: IF( DABS( THETA ).GT.BIGTHETA ) THEN
476: T = HALF / THETA
477: FASTR( 3 ) = T*D( p ) / D( q )
478: FASTR( 4 ) = -T*D( q ) / D( p )
479: CALL DROTM( M, A( 1, p ), 1,
1.6 bertrand 480: $ A( 1, q ), 1, FASTR )
1.1 bertrand 481: IF( RSVEC )CALL DROTM( MVL,
1.6 bertrand 482: $ V( 1, p ), 1,
483: $ V( 1, q ), 1,
484: $ FASTR )
1.1 bertrand 485: SVA( q ) = AAQQ*DSQRT( DMAX1( ZERO,
1.6 bertrand 486: $ ONE+T*APOAQ*AAPQ ) )
1.1 bertrand 487: AAPP = AAPP*DSQRT( DMAX1( ZERO,
1.6 bertrand 488: $ ONE-T*AQOAP*AAPQ ) )
1.1 bertrand 489: MXSINJ = DMAX1( MXSINJ, DABS( T ) )
490: ELSE
491: *
492: * .. choose correct signum for THETA and rotate
493: *
494: THSIGN = -DSIGN( ONE, AAPQ )
495: IF( AAQQ.GT.AAPP0 )THSIGN = -THSIGN
496: T = ONE / ( THETA+THSIGN*
1.6 bertrand 497: $ DSQRT( ONE+THETA*THETA ) )
1.1 bertrand 498: CS = DSQRT( ONE / ( ONE+T*T ) )
499: SN = T*CS
500: MXSINJ = DMAX1( MXSINJ, DABS( SN ) )
501: SVA( q ) = AAQQ*DSQRT( DMAX1( ZERO,
1.6 bertrand 502: $ ONE+T*APOAQ*AAPQ ) )
1.4 bertrand 503: AAPP = AAPP*DSQRT( DMAX1( ZERO,
1.6 bertrand 504: $ ONE-T*AQOAP*AAPQ ) )
1.1 bertrand 505:
506: APOAQ = D( p ) / D( q )
507: AQOAP = D( q ) / D( p )
508: IF( D( p ).GE.ONE ) THEN
509: *
510: IF( D( q ).GE.ONE ) THEN
511: FASTR( 3 ) = T*APOAQ
512: FASTR( 4 ) = -T*AQOAP
513: D( p ) = D( p )*CS
514: D( q ) = D( q )*CS
515: CALL DROTM( M, A( 1, p ), 1,
1.6 bertrand 516: $ A( 1, q ), 1,
517: $ FASTR )
1.1 bertrand 518: IF( RSVEC )CALL DROTM( MVL,
1.6 bertrand 519: $ V( 1, p ), 1, V( 1, q ),
520: $ 1, FASTR )
1.1 bertrand 521: ELSE
522: CALL DAXPY( M, -T*AQOAP,
1.6 bertrand 523: $ A( 1, q ), 1,
524: $ A( 1, p ), 1 )
1.1 bertrand 525: CALL DAXPY( M, CS*SN*APOAQ,
1.6 bertrand 526: $ A( 1, p ), 1,
527: $ A( 1, q ), 1 )
1.1 bertrand 528: IF( RSVEC ) THEN
529: CALL DAXPY( MVL, -T*AQOAP,
1.6 bertrand 530: $ V( 1, q ), 1,
531: $ V( 1, p ), 1 )
1.1 bertrand 532: CALL DAXPY( MVL,
1.6 bertrand 533: $ CS*SN*APOAQ,
534: $ V( 1, p ), 1,
535: $ V( 1, q ), 1 )
1.1 bertrand 536: END IF
537: D( p ) = D( p )*CS
538: D( q ) = D( q ) / CS
539: END IF
540: ELSE
541: IF( D( q ).GE.ONE ) THEN
542: CALL DAXPY( M, T*APOAQ,
1.6 bertrand 543: $ A( 1, p ), 1,
544: $ A( 1, q ), 1 )
1.1 bertrand 545: CALL DAXPY( M, -CS*SN*AQOAP,
1.6 bertrand 546: $ A( 1, q ), 1,
547: $ A( 1, p ), 1 )
1.1 bertrand 548: IF( RSVEC ) THEN
549: CALL DAXPY( MVL, T*APOAQ,
1.6 bertrand 550: $ V( 1, p ), 1,
551: $ V( 1, q ), 1 )
1.1 bertrand 552: CALL DAXPY( MVL,
1.6 bertrand 553: $ -CS*SN*AQOAP,
554: $ V( 1, q ), 1,
555: $ V( 1, p ), 1 )
1.1 bertrand 556: END IF
557: D( p ) = D( p ) / CS
558: D( q ) = D( q )*CS
559: ELSE
560: IF( D( p ).GE.D( q ) ) THEN
561: CALL DAXPY( M, -T*AQOAP,
1.6 bertrand 562: $ A( 1, q ), 1,
563: $ A( 1, p ), 1 )
1.1 bertrand 564: CALL DAXPY( M, CS*SN*APOAQ,
1.6 bertrand 565: $ A( 1, p ), 1,
566: $ A( 1, q ), 1 )
1.1 bertrand 567: D( p ) = D( p )*CS
568: D( q ) = D( q ) / CS
569: IF( RSVEC ) THEN
570: CALL DAXPY( MVL,
1.6 bertrand 571: $ -T*AQOAP,
572: $ V( 1, q ), 1,
573: $ V( 1, p ), 1 )
1.1 bertrand 574: CALL DAXPY( MVL,
1.6 bertrand 575: $ CS*SN*APOAQ,
576: $ V( 1, p ), 1,
577: $ V( 1, q ), 1 )
1.1 bertrand 578: END IF
579: ELSE
580: CALL DAXPY( M, T*APOAQ,
1.6 bertrand 581: $ A( 1, p ), 1,
582: $ A( 1, q ), 1 )
1.1 bertrand 583: CALL DAXPY( M,
1.6 bertrand 584: $ -CS*SN*AQOAP,
585: $ A( 1, q ), 1,
586: $ A( 1, p ), 1 )
1.1 bertrand 587: D( p ) = D( p ) / CS
588: D( q ) = D( q )*CS
589: IF( RSVEC ) THEN
590: CALL DAXPY( MVL,
1.6 bertrand 591: $ T*APOAQ, V( 1, p ),
592: $ 1, V( 1, q ), 1 )
1.1 bertrand 593: CALL DAXPY( MVL,
1.6 bertrand 594: $ -CS*SN*AQOAP,
595: $ V( 1, q ), 1,
596: $ V( 1, p ), 1 )
1.1 bertrand 597: END IF
598: END IF
599: END IF
600: END IF
601: END IF
602:
603: ELSE
604: IF( AAPP.GT.AAQQ ) THEN
605: CALL DCOPY( M, A( 1, p ), 1, WORK,
1.6 bertrand 606: $ 1 )
1.1 bertrand 607: CALL DLASCL( 'G', 0, 0, AAPP, ONE,
1.6 bertrand 608: $ M, 1, WORK, LDA, IERR )
1.1 bertrand 609: CALL DLASCL( 'G', 0, 0, AAQQ, ONE,
1.6 bertrand 610: $ M, 1, A( 1, q ), LDA,
611: $ IERR )
1.1 bertrand 612: TEMP1 = -AAPQ*D( p ) / D( q )
613: CALL DAXPY( M, TEMP1, WORK, 1,
1.6 bertrand 614: $ A( 1, q ), 1 )
1.1 bertrand 615: CALL DLASCL( 'G', 0, 0, ONE, AAQQ,
1.6 bertrand 616: $ M, 1, A( 1, q ), LDA,
617: $ IERR )
1.1 bertrand 618: SVA( q ) = AAQQ*DSQRT( DMAX1( ZERO,
1.6 bertrand 619: $ ONE-AAPQ*AAPQ ) )
1.1 bertrand 620: MXSINJ = DMAX1( MXSINJ, SFMIN )
621: ELSE
622: CALL DCOPY( M, A( 1, q ), 1, WORK,
1.6 bertrand 623: $ 1 )
1.1 bertrand 624: CALL DLASCL( 'G', 0, 0, AAQQ, ONE,
1.6 bertrand 625: $ M, 1, WORK, LDA, IERR )
1.1 bertrand 626: CALL DLASCL( 'G', 0, 0, AAPP, ONE,
1.6 bertrand 627: $ M, 1, A( 1, p ), LDA,
628: $ IERR )
1.1 bertrand 629: TEMP1 = -AAPQ*D( q ) / D( p )
630: CALL DAXPY( M, TEMP1, WORK, 1,
1.6 bertrand 631: $ A( 1, p ), 1 )
1.1 bertrand 632: CALL DLASCL( 'G', 0, 0, ONE, AAPP,
1.6 bertrand 633: $ M, 1, A( 1, p ), LDA,
634: $ IERR )
1.1 bertrand 635: SVA( p ) = AAPP*DSQRT( DMAX1( ZERO,
1.6 bertrand 636: $ ONE-AAPQ*AAPQ ) )
1.1 bertrand 637: MXSINJ = DMAX1( MXSINJ, SFMIN )
638: END IF
639: END IF
640: * END IF ROTOK THEN ... ELSE
641: *
642: * In the case of cancellation in updating SVA(q)
643: * .. recompute SVA(q)
644: IF( ( SVA( q ) / AAQQ )**2.LE.ROOTEPS )
1.6 bertrand 645: $ THEN
1.1 bertrand 646: IF( ( AAQQ.LT.ROOTBIG ) .AND.
1.6 bertrand 647: $ ( AAQQ.GT.ROOTSFMIN ) ) THEN
1.1 bertrand 648: SVA( q ) = DNRM2( M, A( 1, q ), 1 )*
1.6 bertrand 649: $ D( q )
1.1 bertrand 650: ELSE
651: T = ZERO
1.4 bertrand 652: AAQQ = ONE
1.1 bertrand 653: CALL DLASSQ( M, A( 1, q ), 1, T,
1.6 bertrand 654: $ AAQQ )
1.1 bertrand 655: SVA( q ) = T*DSQRT( AAQQ )*D( q )
656: END IF
657: END IF
658: IF( ( AAPP / AAPP0 )**2.LE.ROOTEPS ) THEN
659: IF( ( AAPP.LT.ROOTBIG ) .AND.
1.6 bertrand 660: $ ( AAPP.GT.ROOTSFMIN ) ) THEN
1.1 bertrand 661: AAPP = DNRM2( M, A( 1, p ), 1 )*
1.6 bertrand 662: $ D( p )
1.1 bertrand 663: ELSE
664: T = ZERO
1.4 bertrand 665: AAPP = ONE
1.1 bertrand 666: CALL DLASSQ( M, A( 1, p ), 1, T,
1.6 bertrand 667: $ AAPP )
1.1 bertrand 668: AAPP = T*DSQRT( AAPP )*D( p )
669: END IF
670: SVA( p ) = AAPP
671: END IF
672: * end of OK rotation
673: ELSE
674: NOTROT = NOTROT + 1
675: * SKIPPED = SKIPPED + 1
676: PSKIPPED = PSKIPPED + 1
677: IJBLSK = IJBLSK + 1
678: END IF
679: ELSE
680: NOTROT = NOTROT + 1
681: PSKIPPED = PSKIPPED + 1
682: IJBLSK = IJBLSK + 1
683: END IF
684:
685: * IF ( NOTROT .GE. EMPTSW ) GO TO 2011
686: IF( ( i.LE.SWBAND ) .AND. ( IJBLSK.GE.BLSKIP ) )
1.6 bertrand 687: $ THEN
1.1 bertrand 688: SVA( p ) = AAPP
689: NOTROT = 0
690: GO TO 2011
691: END IF
692: IF( ( i.LE.SWBAND ) .AND.
1.6 bertrand 693: $ ( PSKIPPED.GT.ROWSKIP ) ) THEN
1.1 bertrand 694: AAPP = -AAPP
695: NOTROT = 0
696: GO TO 2203
697: END IF
698:
699: *
700: 2200 CONTINUE
701: * end of the q-loop
702: 2203 CONTINUE
703:
704: SVA( p ) = AAPP
705: *
706: ELSE
707: IF( AAPP.EQ.ZERO )NOTROT = NOTROT +
1.6 bertrand 708: $ MIN0( jgl+KBL-1, N ) - jgl + 1
1.1 bertrand 709: IF( AAPP.LT.ZERO )NOTROT = 0
710: *** IF ( NOTROT .GE. EMPTSW ) GO TO 2011
711: END IF
712:
713: 2100 CONTINUE
714: * end of the p-loop
715: 2010 CONTINUE
716: * end of the jbc-loop
717: 2011 CONTINUE
718: *2011 bailed out of the jbc-loop
719: DO 2012 p = igl, MIN0( igl+KBL-1, N )
720: SVA( p ) = DABS( SVA( p ) )
721: 2012 CONTINUE
722: *** IF ( NOTROT .GE. EMPTSW ) GO TO 1994
723: 2000 CONTINUE
724: *2000 :: end of the ibr-loop
725: *
726: * .. update SVA(N)
727: IF( ( SVA( N ).LT.ROOTBIG ) .AND. ( SVA( N ).GT.ROOTSFMIN ) )
1.6 bertrand 728: $ THEN
1.1 bertrand 729: SVA( N ) = DNRM2( M, A( 1, N ), 1 )*D( N )
730: ELSE
731: T = ZERO
1.4 bertrand 732: AAPP = ONE
1.1 bertrand 733: CALL DLASSQ( M, A( 1, N ), 1, T, AAPP )
734: SVA( N ) = T*DSQRT( AAPP )*D( N )
735: END IF
736: *
737: * Additional steering devices
738: *
739: IF( ( i.LT.SWBAND ) .AND. ( ( MXAAPQ.LE.ROOTTOL ) .OR.
1.6 bertrand 740: $ ( ISWROT.LE.N ) ) )SWBAND = i
1.1 bertrand 741:
742: IF( ( i.GT.SWBAND+1 ) .AND. ( MXAAPQ.LT.DBLE( N )*TOL ) .AND.
1.6 bertrand 743: $ ( DBLE( N )*MXAAPQ*MXSINJ.LT.TOL ) ) THEN
1.1 bertrand 744: GO TO 1994
745: END IF
746:
747: *
748: IF( NOTROT.GE.EMPTSW )GO TO 1994
749:
750: 1993 CONTINUE
751: * end i=1:NSWEEP loop
752: * #:) Reaching this point means that the procedure has completed the given
753: * number of sweeps.
754: INFO = NSWEEP - 1
755: GO TO 1995
756: 1994 CONTINUE
757: * #:) Reaching this point means that during the i-th sweep all pivots were
758: * below the given threshold, causing early exit.
759:
760: INFO = 0
761: * #:) INFO = 0 confirms successful iterations.
762: 1995 CONTINUE
763: *
764: * Sort the vector D
765: *
766: DO 5991 p = 1, N - 1
767: q = IDAMAX( N-p+1, SVA( p ), 1 ) + p - 1
768: IF( p.NE.q ) THEN
769: TEMP1 = SVA( p )
770: SVA( p ) = SVA( q )
771: SVA( q ) = TEMP1
772: TEMP1 = D( p )
773: D( p ) = D( q )
774: D( q ) = TEMP1
775: CALL DSWAP( M, A( 1, p ), 1, A( 1, q ), 1 )
776: IF( RSVEC )CALL DSWAP( MVL, V( 1, p ), 1, V( 1, q ), 1 )
777: END IF
778: 5991 CONTINUE
779: *
780: RETURN
781: * ..
782: * .. END OF DGSVJ1
783: * ..
784: END
CVSweb interface <joel.bertrand@systella.fr>
![[BACK]](/cvsweb/icons/back.gif){kind=icon}
Return to dgsvj1.f CVS log ![[TXT]](/cvsweb/icons/text.gif){kind=icon}
![[DIR]](/cvsweb/icons/dir.gif){kind=icon}
Up to [local] / rpl / lapack / lapack