1 *> \brief <b> SGBSVX computes the solution to system of linear equations A * X = B for GB matrices</b>
3 * =========== DOCUMENTATION ===========
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
9 *> Download SGBSVX + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/sgbsvx.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/sgbsvx.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/sgbsvx.f">
21 * SUBROUTINE SGBSVX( FACT, TRANS, N, KL, KU, NRHS, AB, LDAB, AFB,
22 * LDAFB, IPIV, EQUED, R, C, B, LDB, X, LDX,
23 * RCOND, FERR, BERR, WORK, IWORK, INFO )
25 * .. Scalar Arguments ..
26 * CHARACTER EQUED, FACT, TRANS
27 * INTEGER INFO, KL, KU, LDAB, LDAFB, LDB, LDX, N, NRHS
30 * .. Array Arguments ..
31 * INTEGER IPIV( * ), IWORK( * )
32 * REAL AB( LDAB, * ), AFB( LDAFB, * ), B( LDB, * ),
33 * $ BERR( * ), C( * ), FERR( * ), R( * ),
34 * $ WORK( * ), X( LDX, * )
43 *> SGBSVX uses the LU factorization to compute the solution to a real
44 *> system of linear equations A * X = B, A**T * X = B, or A**H * X = B,
45 *> where A is a band matrix of order N with KL subdiagonals and KU
46 *> superdiagonals, and X and B are N-by-NRHS matrices.
48 *> Error bounds on the solution and a condition estimate are also
57 *> The following steps are performed by this subroutine:
59 *> 1. If FACT = 'E', real scaling factors are computed to equilibrate
61 *> TRANS = 'N': diag(R)*A*diag(C) *inv(diag(C))*X = diag(R)*B
62 *> TRANS = 'T': (diag(R)*A*diag(C))**T *inv(diag(R))*X = diag(C)*B
63 *> TRANS = 'C': (diag(R)*A*diag(C))**H *inv(diag(R))*X = diag(C)*B
64 *> Whether or not the system will be equilibrated depends on the
65 *> scaling of the matrix A, but if equilibration is used, A is
66 *> overwritten by diag(R)*A*diag(C) and B by diag(R)*B (if TRANS='N')
67 *> or diag(C)*B (if TRANS = 'T' or 'C').
69 *> 2. If FACT = 'N' or 'E', the LU decomposition is used to factor the
70 *> matrix A (after equilibration if FACT = 'E') as
72 *> where L is a product of permutation and unit lower triangular
73 *> matrices with KL subdiagonals, and U is upper triangular with
74 *> KL+KU superdiagonals.
76 *> 3. If some U(i,i)=0, so that U is exactly singular, then the routine
77 *> returns with INFO = i. Otherwise, the factored form of A is used
78 *> to estimate the condition number of the matrix A. If the
79 *> reciprocal of the condition number is less than machine precision,
80 *> INFO = N+1 is returned as a warning, but the routine still goes on
81 *> to solve for X and compute error bounds as described below.
83 *> 4. The system of equations is solved for X using the factored form
86 *> 5. Iterative refinement is applied to improve the computed solution
87 *> matrix and calculate error bounds and backward error estimates
90 *> 6. If equilibration was used, the matrix X is premultiplied by
91 *> diag(C) (if TRANS = 'N') or diag(R) (if TRANS = 'T' or 'C') so
92 *> that it solves the original system before equilibration.
100 *> FACT is CHARACTER*1
101 *> Specifies whether or not the factored form of the matrix A is
102 *> supplied on entry, and if not, whether the matrix A should be
103 *> equilibrated before it is factored.
104 *> = 'F': On entry, AFB and IPIV contain the factored form of
105 *> A. If EQUED is not 'N', the matrix A has been
106 *> equilibrated with scaling factors given by R and C.
107 *> AB, AFB, and IPIV are not modified.
108 *> = 'N': The matrix A will be copied to AFB and factored.
109 *> = 'E': The matrix A will be equilibrated if necessary, then
110 *> copied to AFB and factored.
115 *> TRANS is CHARACTER*1
116 *> Specifies the form of the system of equations.
117 *> = 'N': A * X = B (No transpose)
118 *> = 'T': A**T * X = B (Transpose)
119 *> = 'C': A**H * X = B (Transpose)
125 *> The number of linear equations, i.e., the order of the
132 *> The number of subdiagonals within the band of A. KL >= 0.
138 *> The number of superdiagonals within the band of A. KU >= 0.
144 *> The number of right hand sides, i.e., the number of columns
145 *> of the matrices B and X. NRHS >= 0.
150 *> AB is REAL array, dimension (LDAB,N)
151 *> On entry, the matrix A in band storage, in rows 1 to KL+KU+1.
152 *> The j-th column of A is stored in the j-th column of the
153 *> array AB as follows:
154 *> AB(KU+1+i-j,j) = A(i,j) for max(1,j-KU)<=i<=min(N,j+kl)
156 *> If FACT = 'F' and EQUED is not 'N', then A must have been
157 *> equilibrated by the scaling factors in R and/or C. AB is not
158 *> modified if FACT = 'F' or 'N', or if FACT = 'E' and
159 *> EQUED = 'N' on exit.
161 *> On exit, if EQUED .ne. 'N', A is scaled as follows:
162 *> EQUED = 'R': A := diag(R) * A
163 *> EQUED = 'C': A := A * diag(C)
164 *> EQUED = 'B': A := diag(R) * A * diag(C).
170 *> The leading dimension of the array AB. LDAB >= KL+KU+1.
173 *> \param[in,out] AFB
175 *> AFB is REAL array, dimension (LDAFB,N)
176 *> If FACT = 'F', then AFB is an input argument and on entry
177 *> contains details of the LU factorization of the band matrix
178 *> A, as computed by SGBTRF. U is stored as an upper triangular
179 *> band matrix with KL+KU superdiagonals in rows 1 to KL+KU+1,
180 *> and the multipliers used during the factorization are stored
181 *> in rows KL+KU+2 to 2*KL+KU+1. If EQUED .ne. 'N', then AFB is
182 *> the factored form of the equilibrated matrix A.
184 *> If FACT = 'N', then AFB is an output argument and on exit
185 *> returns details of the LU factorization of A.
187 *> If FACT = 'E', then AFB is an output argument and on exit
188 *> returns details of the LU factorization of the equilibrated
189 *> matrix A (see the description of AB for the form of the
190 *> equilibrated matrix).
196 *> The leading dimension of the array AFB. LDAFB >= 2*KL+KU+1.
199 *> \param[in,out] IPIV
201 *> IPIV is INTEGER array, dimension (N)
202 *> If FACT = 'F', then IPIV is an input argument and on entry
203 *> contains the pivot indices from the factorization A = L*U
204 *> as computed by SGBTRF; row i of the matrix was interchanged
207 *> If FACT = 'N', then IPIV is an output argument and on exit
208 *> contains the pivot indices from the factorization A = L*U
209 *> of the original matrix A.
211 *> If FACT = 'E', then IPIV is an output argument and on exit
212 *> contains the pivot indices from the factorization A = L*U
213 *> of the equilibrated matrix A.
216 *> \param[in,out] EQUED
218 *> EQUED is CHARACTER*1
219 *> Specifies the form of equilibration that was done.
220 *> = 'N': No equilibration (always true if FACT = 'N').
221 *> = 'R': Row equilibration, i.e., A has been premultiplied by
223 *> = 'C': Column equilibration, i.e., A has been postmultiplied
225 *> = 'B': Both row and column equilibration, i.e., A has been
226 *> replaced by diag(R) * A * diag(C).
227 *> EQUED is an input argument if FACT = 'F'; otherwise, it is an
233 *> R is REAL array, dimension (N)
234 *> The row scale factors for A. If EQUED = 'R' or 'B', A is
235 *> multiplied on the left by diag(R); if EQUED = 'N' or 'C', R
236 *> is not accessed. R is an input argument if FACT = 'F';
237 *> otherwise, R is an output argument. If FACT = 'F' and
238 *> EQUED = 'R' or 'B', each element of R must be positive.
243 *> C is REAL array, dimension (N)
244 *> The column scale factors for A. If EQUED = 'C' or 'B', A is
245 *> multiplied on the right by diag(C); if EQUED = 'N' or 'R', C
246 *> is not accessed. C is an input argument if FACT = 'F';
247 *> otherwise, C is an output argument. If FACT = 'F' and
248 *> EQUED = 'C' or 'B', each element of C must be positive.
253 *> B is REAL array, dimension (LDB,NRHS)
254 *> On entry, the right hand side matrix B.
256 *> if EQUED = 'N', B is not modified;
257 *> if TRANS = 'N' and EQUED = 'R' or 'B', B is overwritten by
259 *> if TRANS = 'T' or 'C' and EQUED = 'C' or 'B', B is
260 *> overwritten by diag(C)*B.
266 *> The leading dimension of the array B. LDB >= max(1,N).
271 *> X is REAL array, dimension (LDX,NRHS)
272 *> If INFO = 0 or INFO = N+1, the N-by-NRHS solution matrix X
273 *> to the original system of equations. Note that A and B are
274 *> modified on exit if EQUED .ne. 'N', and the solution to the
275 *> equilibrated system is inv(diag(C))*X if TRANS = 'N' and
276 *> EQUED = 'C' or 'B', or inv(diag(R))*X if TRANS = 'T' or 'C'
277 *> and EQUED = 'R' or 'B'.
283 *> The leading dimension of the array X. LDX >= max(1,N).
289 *> The estimate of the reciprocal condition number of the matrix
290 *> A after equilibration (if done). If RCOND is less than the
291 *> machine precision (in particular, if RCOND = 0), the matrix
292 *> is singular to working precision. This condition is
293 *> indicated by a return code of INFO > 0.
298 *> FERR is REAL array, dimension (NRHS)
299 *> The estimated forward error bound for each solution vector
300 *> X(j) (the j-th column of the solution matrix X).
301 *> If XTRUE is the true solution corresponding to X(j), FERR(j)
302 *> is an estimated upper bound for the magnitude of the largest
303 *> element in (X(j) - XTRUE) divided by the magnitude of the
304 *> largest element in X(j). The estimate is as reliable as
305 *> the estimate for RCOND, and is almost always a slight
306 *> overestimate of the true error.
311 *> BERR is REAL array, dimension (NRHS)
312 *> The componentwise relative backward error of each solution
313 *> vector X(j) (i.e., the smallest relative change in
314 *> any element of A or B that makes X(j) an exact solution).
319 *> WORK is REAL array, dimension (3*N)
320 *> On exit, WORK(1) contains the reciprocal pivot growth
321 *> factor norm(A)/norm(U). The "max absolute element" norm is
322 *> used. If WORK(1) is much less than 1, then the stability
323 *> of the LU factorization of the (equilibrated) matrix A
324 *> could be poor. This also means that the solution X, condition
325 *> estimator RCOND, and forward error bound FERR could be
326 *> unreliable. If factorization fails with 0<INFO<=N, then
327 *> WORK(1) contains the reciprocal pivot growth factor for the
328 *> leading INFO columns of A.
333 *> IWORK is INTEGER array, dimension (N)
339 *> = 0: successful exit
340 *> < 0: if INFO = -i, the i-th argument had an illegal value
341 *> > 0: if INFO = i, and i is
342 *> <= N: U(i,i) is exactly zero. The factorization
343 *> has been completed, but the factor U is exactly
344 *> singular, so the solution and error bounds
345 *> could not be computed. RCOND = 0 is returned.
346 *> = N+1: U is nonsingular, but RCOND is less than machine
347 *> precision, meaning that the matrix is singular
348 *> to working precision. Nevertheless, the
349 *> solution and error bounds are computed because
350 *> there are a number of situations where the
351 *> computed solution can be more accurate than the
357 *> \author Univ. of Tennessee
358 *> \author Univ. of California Berkeley
359 *> \author Univ. of Colorado Denver
364 *> \ingroup realGBsolve
366 * =====================================================================
367 SUBROUTINE SGBSVX( FACT, TRANS, N, KL, KU, NRHS, AB, LDAB, AFB,
368 $ LDAFB, IPIV, EQUED, R, C, B, LDB, X, LDX,
369 $ RCOND, FERR, BERR, WORK, IWORK, INFO )
371 * -- LAPACK driver routine (version 3.4.1) --
372 * -- LAPACK is a software package provided by Univ. of Tennessee, --
373 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
376 * .. Scalar Arguments ..
377 CHARACTER EQUED, FACT, TRANS
378 INTEGER INFO, KL, KU, LDAB, LDAFB, LDB, LDX, N, NRHS
381 * .. Array Arguments ..
382 INTEGER IPIV( * ), IWORK( * )
383 REAL AB( LDAB, * ), AFB( LDAFB, * ), B( LDB, * ),
384 $ BERR( * ), C( * ), FERR( * ), R( * ),
385 $ WORK( * ), X( LDX, * )
388 * =====================================================================
389 * Moved setting of INFO = N+1 so INFO does not subsequently get
390 * overwritten. Sven, 17 Mar 05.
391 * =====================================================================
395 PARAMETER ( ZERO = 0.0E+0, ONE = 1.0E+0 )
397 * .. Local Scalars ..
398 LOGICAL COLEQU, EQUIL, NOFACT, NOTRAN, ROWEQU
400 INTEGER I, INFEQU, J, J1, J2
401 REAL AMAX, ANORM, BIGNUM, COLCND, RCMAX, RCMIN,
402 $ ROWCND, RPVGRW, SMLNUM
404 * .. External Functions ..
406 REAL SLAMCH, SLANGB, SLANTB
407 EXTERNAL LSAME, SLAMCH, SLANGB, SLANTB
409 * .. External Subroutines ..
410 EXTERNAL SCOPY, SGBCON, SGBEQU, SGBRFS, SGBTRF, SGBTRS,
411 $ SLACPY, SLAQGB, XERBLA
413 * .. Intrinsic Functions ..
414 INTRINSIC ABS, MAX, MIN
416 * .. Executable Statements ..
419 NOFACT = LSAME( FACT, 'N' )
420 EQUIL = LSAME( FACT, 'E' )
421 NOTRAN = LSAME( TRANS, 'N' )
422 IF( NOFACT .OR. EQUIL ) THEN
427 ROWEQU = LSAME( EQUED, 'R' ) .OR. LSAME( EQUED, 'B' )
428 COLEQU = LSAME( EQUED, 'C' ) .OR. LSAME( EQUED, 'B' )
429 SMLNUM = SLAMCH( 'Safe minimum' )
430 BIGNUM = ONE / SMLNUM
433 * Test the input parameters.
435 IF( .NOT.NOFACT .AND. .NOT.EQUIL .AND. .NOT.LSAME( FACT, 'F' ) )
438 ELSE IF( .NOT.NOTRAN .AND. .NOT.LSAME( TRANS, 'T' ) .AND. .NOT.
439 $ LSAME( TRANS, 'C' ) ) THEN
441 ELSE IF( N.LT.0 ) THEN
443 ELSE IF( KL.LT.0 ) THEN
445 ELSE IF( KU.LT.0 ) THEN
447 ELSE IF( NRHS.LT.0 ) THEN
449 ELSE IF( LDAB.LT.KL+KU+1 ) THEN
451 ELSE IF( LDAFB.LT.2*KL+KU+1 ) THEN
453 ELSE IF( LSAME( FACT, 'F' ) .AND. .NOT.
454 $ ( ROWEQU .OR. COLEQU .OR. LSAME( EQUED, 'N' ) ) ) THEN
461 RCMIN = MIN( RCMIN, R( J ) )
462 RCMAX = MAX( RCMAX, R( J ) )
464 IF( RCMIN.LE.ZERO ) THEN
466 ELSE IF( N.GT.0 ) THEN
467 ROWCND = MAX( RCMIN, SMLNUM ) / MIN( RCMAX, BIGNUM )
472 IF( COLEQU .AND. INFO.EQ.0 ) THEN
476 RCMIN = MIN( RCMIN, C( J ) )
477 RCMAX = MAX( RCMAX, C( J ) )
479 IF( RCMIN.LE.ZERO ) THEN
481 ELSE IF( N.GT.0 ) THEN
482 COLCND = MAX( RCMIN, SMLNUM ) / MIN( RCMAX, BIGNUM )
488 IF( LDB.LT.MAX( 1, N ) ) THEN
490 ELSE IF( LDX.LT.MAX( 1, N ) ) THEN
497 CALL XERBLA( 'SGBSVX', -INFO )
503 * Compute row and column scalings to equilibrate the matrix A.
505 CALL SGBEQU( N, N, KL, KU, AB, LDAB, R, C, ROWCND, COLCND,
507 IF( INFEQU.EQ.0 ) THEN
509 * Equilibrate the matrix.
511 CALL SLAQGB( N, N, KL, KU, AB, LDAB, R, C, ROWCND, COLCND,
513 ROWEQU = LSAME( EQUED, 'R' ) .OR. LSAME( EQUED, 'B' )
514 COLEQU = LSAME( EQUED, 'C' ) .OR. LSAME( EQUED, 'B' )
518 * Scale the right hand side.
524 B( I, J ) = R( I )*B( I, J )
528 ELSE IF( COLEQU ) THEN
531 B( I, J ) = C( I )*B( I, J )
536 IF( NOFACT .OR. EQUIL ) THEN
538 * Compute the LU factorization of the band matrix A.
543 CALL SCOPY( J2-J1+1, AB( KU+1-J+J1, J ), 1,
544 $ AFB( KL+KU+1-J+J1, J ), 1 )
547 CALL SGBTRF( N, N, KL, KU, AFB, LDAFB, IPIV, INFO )
549 * Return if INFO is non-zero.
553 * Compute the reciprocal pivot growth factor of the
554 * leading rank-deficient INFO columns of A.
558 DO 80 I = MAX( KU+2-J, 1 ), MIN( N+KU+1-J, KL+KU+1 )
559 ANORM = MAX( ANORM, ABS( AB( I, J ) ) )
562 RPVGRW = SLANTB( 'M', 'U', 'N', INFO, MIN( INFO-1, KL+KU ),
563 $ AFB( MAX( 1, KL+KU+2-INFO ), 1 ), LDAFB,
565 IF( RPVGRW.EQ.ZERO ) THEN
568 RPVGRW = ANORM / RPVGRW
576 * Compute the norm of the matrix A and the
577 * reciprocal pivot growth factor RPVGRW.
584 ANORM = SLANGB( NORM, N, KL, KU, AB, LDAB, WORK )
585 RPVGRW = SLANTB( 'M', 'U', 'N', N, KL+KU, AFB, LDAFB, WORK )
586 IF( RPVGRW.EQ.ZERO ) THEN
589 RPVGRW = SLANGB( 'M', N, KL, KU, AB, LDAB, WORK ) / RPVGRW
592 * Compute the reciprocal of the condition number of A.
594 CALL SGBCON( NORM, N, KL, KU, AFB, LDAFB, IPIV, ANORM, RCOND,
595 $ WORK, IWORK, INFO )
597 * Compute the solution matrix X.
599 CALL SLACPY( 'Full', N, NRHS, B, LDB, X, LDX )
600 CALL SGBTRS( TRANS, N, KL, KU, NRHS, AFB, LDAFB, IPIV, X, LDX,
603 * Use iterative refinement to improve the computed solution and
604 * compute error bounds and backward error estimates for it.
606 CALL SGBRFS( TRANS, N, KL, KU, NRHS, AB, LDAB, AFB, LDAFB, IPIV,
607 $ B, LDB, X, LDX, FERR, BERR, WORK, IWORK, INFO )
609 * Transform the solution matrix X to a solution of the original
616 X( I, J ) = C( I )*X( I, J )
620 FERR( J ) = FERR( J ) / COLCND
623 ELSE IF( ROWEQU ) THEN
626 X( I, J ) = R( I )*X( I, J )
630 FERR( J ) = FERR( J ) / ROWCND
634 * Set INFO = N+1 if the matrix is singular to working precision.
636 IF( RCOND.LT.SLAMCH( 'Epsilon' ) )