1 *> \brief \b SLATBS solves a triangular banded system of equations.
3 * =========== DOCUMENTATION ===========
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
9 *> Download SLATBS + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/slatbs.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/slatbs.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/slatbs.f">
21 * SUBROUTINE SLATBS( UPLO, TRANS, DIAG, NORMIN, N, KD, AB, LDAB, X,
22 * SCALE, CNORM, INFO )
24 * .. Scalar Arguments ..
25 * CHARACTER DIAG, NORMIN, TRANS, UPLO
26 * INTEGER INFO, KD, LDAB, N
29 * .. Array Arguments ..
30 * REAL AB( LDAB, * ), CNORM( * ), X( * )
39 *> SLATBS solves one of the triangular systems
41 *> A *x = s*b or A**T*x = s*b
43 *> with scaling to prevent overflow, where A is an upper or lower
44 *> triangular band matrix. Here A**T denotes the transpose of A, x and b
45 *> are n-element vectors, and s is a scaling factor, usually less than
46 *> or equal to 1, chosen so that the components of x will be less than
47 *> the overflow threshold. If the unscaled problem will not cause
48 *> overflow, the Level 2 BLAS routine STBSV is called. If the matrix A
49 *> is singular (A(j,j) = 0 for some j), then s is set to 0 and a
50 *> non-trivial solution to A*x = 0 is returned.
58 *> UPLO is CHARACTER*1
59 *> Specifies whether the matrix A is upper or lower triangular.
60 *> = 'U': Upper triangular
61 *> = 'L': Lower triangular
66 *> TRANS is CHARACTER*1
67 *> Specifies the operation applied to A.
68 *> = 'N': Solve A * x = s*b (No transpose)
69 *> = 'T': Solve A**T* x = s*b (Transpose)
70 *> = 'C': Solve A**T* x = s*b (Conjugate transpose = Transpose)
75 *> DIAG is CHARACTER*1
76 *> Specifies whether or not the matrix A is unit triangular.
77 *> = 'N': Non-unit triangular
78 *> = 'U': Unit triangular
83 *> NORMIN is CHARACTER*1
84 *> Specifies whether CNORM has been set or not.
85 *> = 'Y': CNORM contains the column norms on entry
86 *> = 'N': CNORM is not set on entry. On exit, the norms will
87 *> be computed and stored in CNORM.
93 *> The order of the matrix A. N >= 0.
99 *> The number of subdiagonals or superdiagonals in the
100 *> triangular matrix A. KD >= 0.
105 *> AB is REAL array, dimension (LDAB,N)
106 *> The upper or lower triangular band matrix A, stored in the
107 *> first KD+1 rows of the array. The j-th column of A is stored
108 *> in the j-th column of the array AB as follows:
109 *> if UPLO = 'U', AB(kd+1+i-j,j) = A(i,j) for max(1,j-kd)<=i<=j;
110 *> if UPLO = 'L', AB(1+i-j,j) = A(i,j) for j<=i<=min(n,j+kd).
116 *> The leading dimension of the array AB. LDAB >= KD+1.
121 *> X is REAL array, dimension (N)
122 *> On entry, the right hand side b of the triangular system.
123 *> On exit, X is overwritten by the solution vector x.
129 *> The scaling factor s for the triangular system
130 *> A * x = s*b or A**T* x = s*b.
131 *> If SCALE = 0, the matrix A is singular or badly scaled, and
132 *> the vector x is an exact or approximate solution to A*x = 0.
135 *> \param[in,out] CNORM
137 *> CNORM is REAL array, dimension (N)
139 *> If NORMIN = 'Y', CNORM is an input argument and CNORM(j)
140 *> contains the norm of the off-diagonal part of the j-th column
141 *> of A. If TRANS = 'N', CNORM(j) must be greater than or equal
142 *> to the infinity-norm, and if TRANS = 'T' or 'C', CNORM(j)
143 *> must be greater than or equal to the 1-norm.
145 *> If NORMIN = 'N', CNORM is an output argument and CNORM(j)
146 *> returns the 1-norm of the offdiagonal part of the j-th column
153 *> = 0: successful exit
154 *> < 0: if INFO = -k, the k-th argument had an illegal value
160 *> \author Univ. of Tennessee
161 *> \author Univ. of California Berkeley
162 *> \author Univ. of Colorado Denver
165 *> \date September 2012
167 *> \ingroup realOTHERauxiliary
169 *> \par Further Details:
170 * =====================
174 *> A rough bound on x is computed; if that is less than overflow, STBSV
175 *> is called, otherwise, specific code is used which checks for possible
176 *> overflow or divide-by-zero at every operation.
178 *> A columnwise scheme is used for solving A*x = b. The basic algorithm
179 *> if A is lower triangular is
183 *> x(j) := x(j) / A(j,j)
184 *> x[j+1:n] := x[j+1:n] - x(j) * A[j+1:n,j]
187 *> Define bounds on the components of x after j iterations of the loop:
188 *> M(j) = bound on x[1:j]
189 *> G(j) = bound on x[j+1:n]
190 *> Initially, let M(0) = 0 and G(0) = max{x(i), i=1,...,n}.
192 *> Then for iteration j+1 we have
193 *> M(j+1) <= G(j) / | A(j+1,j+1) |
194 *> G(j+1) <= G(j) + M(j+1) * | A[j+2:n,j+1] |
195 *> <= G(j) ( 1 + CNORM(j+1) / | A(j+1,j+1) | )
197 *> where CNORM(j+1) is greater than or equal to the infinity-norm of
198 *> column j+1 of A, not counting the diagonal. Hence
200 *> G(j) <= G(0) product ( 1 + CNORM(i) / | A(i,i) | )
204 *> |x(j)| <= ( G(0) / |A(j,j)| ) product ( 1 + CNORM(i) / |A(i,i)| )
207 *> Since |x(j)| <= M(j), we use the Level 2 BLAS routine STBSV if the
208 *> reciprocal of the largest M(j), j=1,..,n, is larger than
209 *> max(underflow, 1/overflow).
211 *> The bound on x(j) is also used to determine when a step in the
212 *> columnwise method can be performed without fear of overflow. If
213 *> the computed bound is greater than a large constant, x is scaled to
214 *> prevent overflow, but if the bound overflows, x is set to 0, x(j) to
215 *> 1, and scale to 0, and a non-trivial solution to A*x = 0 is found.
217 *> Similarly, a row-wise scheme is used to solve A**T*x = b. The basic
218 *> algorithm for A upper triangular is
221 *> x(j) := ( b(j) - A[1:j-1,j]**T * x[1:j-1] ) / A(j,j)
224 *> We simultaneously compute two bounds
225 *> G(j) = bound on ( b(i) - A[1:i-1,i]**T * x[1:i-1] ), 1<=i<=j
226 *> M(j) = bound on x(i), 1<=i<=j
228 *> The initial values are G(0) = 0, M(0) = max{b(i), i=1,..,n}, and we
229 *> add the constraint G(j) >= G(j-1) and M(j) >= M(j-1) for j >= 1.
230 *> Then the bound on x(j) is
232 *> M(j) <= M(j-1) * ( 1 + CNORM(j) ) / | A(j,j) |
234 *> <= M(0) * product ( ( 1 + CNORM(i) ) / |A(i,i)| )
237 *> and we can safely call STBSV if 1/M(n) and 1/G(n) are both greater
238 *> than max(underflow, 1/overflow).
241 * =====================================================================
242 SUBROUTINE SLATBS( UPLO, TRANS, DIAG, NORMIN, N, KD, AB, LDAB, X,
243 $ SCALE, CNORM, INFO )
245 * -- LAPACK auxiliary routine (version 3.4.2) --
246 * -- LAPACK is a software package provided by Univ. of Tennessee, --
247 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
250 * .. Scalar Arguments ..
251 CHARACTER DIAG, NORMIN, TRANS, UPLO
252 INTEGER INFO, KD, LDAB, N
255 * .. Array Arguments ..
256 REAL AB( LDAB, * ), CNORM( * ), X( * )
259 * =====================================================================
263 PARAMETER ( ZERO = 0.0E+0, HALF = 0.5E+0, ONE = 1.0E+0 )
265 * .. Local Scalars ..
266 LOGICAL NOTRAN, NOUNIT, UPPER
267 INTEGER I, IMAX, J, JFIRST, JINC, JLAST, JLEN, MAIND
268 REAL BIGNUM, GROW, REC, SMLNUM, SUMJ, TJJ, TJJS,
269 $ TMAX, TSCAL, USCAL, XBND, XJ, XMAX
271 * .. External Functions ..
274 REAL SASUM, SDOT, SLAMCH
275 EXTERNAL LSAME, ISAMAX, SASUM, SDOT, SLAMCH
277 * .. External Subroutines ..
278 EXTERNAL SAXPY, SSCAL, STBSV, XERBLA
280 * .. Intrinsic Functions ..
281 INTRINSIC ABS, MAX, MIN
283 * .. Executable Statements ..
286 UPPER = LSAME( UPLO, 'U' )
287 NOTRAN = LSAME( TRANS, 'N' )
288 NOUNIT = LSAME( DIAG, 'N' )
290 * Test the input parameters.
292 IF( .NOT.UPPER .AND. .NOT.LSAME( UPLO, 'L' ) ) THEN
294 ELSE IF( .NOT.NOTRAN .AND. .NOT.LSAME( TRANS, 'T' ) .AND. .NOT.
295 $ LSAME( TRANS, 'C' ) ) THEN
297 ELSE IF( .NOT.NOUNIT .AND. .NOT.LSAME( DIAG, 'U' ) ) THEN
299 ELSE IF( .NOT.LSAME( NORMIN, 'Y' ) .AND. .NOT.
300 $ LSAME( NORMIN, 'N' ) ) THEN
302 ELSE IF( N.LT.0 ) THEN
304 ELSE IF( KD.LT.0 ) THEN
306 ELSE IF( LDAB.LT.KD+1 ) THEN
310 CALL XERBLA( 'SLATBS', -INFO )
314 * Quick return if possible
319 * Determine machine dependent parameters to control overflow.
321 SMLNUM = SLAMCH( 'Safe minimum' ) / SLAMCH( 'Precision' )
322 BIGNUM = ONE / SMLNUM
325 IF( LSAME( NORMIN, 'N' ) ) THEN
327 * Compute the 1-norm of each column, not including the diagonal.
331 * A is upper triangular.
334 JLEN = MIN( KD, J-1 )
335 CNORM( J ) = SASUM( JLEN, AB( KD+1-JLEN, J ), 1 )
339 * A is lower triangular.
342 JLEN = MIN( KD, N-J )
344 CNORM( J ) = SASUM( JLEN, AB( 2, J ), 1 )
352 * Scale the column norms by TSCAL if the maximum element in CNORM is
353 * greater than BIGNUM.
355 IMAX = ISAMAX( N, CNORM, 1 )
357 IF( TMAX.LE.BIGNUM ) THEN
360 TSCAL = ONE / ( SMLNUM*TMAX )
361 CALL SSCAL( N, TSCAL, CNORM, 1 )
364 * Compute a bound on the computed solution vector to see if the
365 * Level 2 BLAS routine STBSV can be used.
367 J = ISAMAX( N, X, 1 )
372 * Compute the growth in A * x = b.
386 IF( TSCAL.NE.ONE ) THEN
393 * A is non-unit triangular.
395 * Compute GROW = 1/G(j) and XBND = 1/M(j).
396 * Initially, G(0) = max{x(i), i=1,...,n}.
398 GROW = ONE / MAX( XBND, SMLNUM )
400 DO 30 J = JFIRST, JLAST, JINC
402 * Exit the loop if the growth factor is too small.
407 * M(j) = G(j-1) / abs(A(j,j))
409 TJJ = ABS( AB( MAIND, J ) )
410 XBND = MIN( XBND, MIN( ONE, TJJ )*GROW )
411 IF( TJJ+CNORM( J ).GE.SMLNUM ) THEN
413 * G(j) = G(j-1)*( 1 + CNORM(j) / abs(A(j,j)) )
415 GROW = GROW*( TJJ / ( TJJ+CNORM( J ) ) )
418 * G(j) could overflow, set GROW to 0.
426 * A is unit triangular.
428 * Compute GROW = 1/G(j), where G(0) = max{x(i), i=1,...,n}.
430 GROW = MIN( ONE, ONE / MAX( XBND, SMLNUM ) )
431 DO 40 J = JFIRST, JLAST, JINC
433 * Exit the loop if the growth factor is too small.
438 * G(j) = G(j-1)*( 1 + CNORM(j) )
440 GROW = GROW*( ONE / ( ONE+CNORM( J ) ) )
447 * Compute the growth in A**T * x = b.
461 IF( TSCAL.NE.ONE ) THEN
468 * A is non-unit triangular.
470 * Compute GROW = 1/G(j) and XBND = 1/M(j).
471 * Initially, M(0) = max{x(i), i=1,...,n}.
473 GROW = ONE / MAX( XBND, SMLNUM )
475 DO 60 J = JFIRST, JLAST, JINC
477 * Exit the loop if the growth factor is too small.
482 * G(j) = max( G(j-1), M(j-1)*( 1 + CNORM(j) ) )
484 XJ = ONE + CNORM( J )
485 GROW = MIN( GROW, XBND / XJ )
487 * M(j) = M(j-1)*( 1 + CNORM(j) ) / abs(A(j,j))
489 TJJ = ABS( AB( MAIND, J ) )
491 $ XBND = XBND*( TJJ / XJ )
493 GROW = MIN( GROW, XBND )
496 * A is unit triangular.
498 * Compute GROW = 1/G(j), where G(0) = max{x(i), i=1,...,n}.
500 GROW = MIN( ONE, ONE / MAX( XBND, SMLNUM ) )
501 DO 70 J = JFIRST, JLAST, JINC
503 * Exit the loop if the growth factor is too small.
508 * G(j) = ( 1 + CNORM(j) )*G(j-1)
510 XJ = ONE + CNORM( J )
517 IF( ( GROW*TSCAL ).GT.SMLNUM ) THEN
519 * Use the Level 2 BLAS solve if the reciprocal of the bound on
520 * elements of X is not too small.
522 CALL STBSV( UPLO, TRANS, DIAG, N, KD, AB, LDAB, X, 1 )
525 * Use a Level 1 BLAS solve, scaling intermediate results.
527 IF( XMAX.GT.BIGNUM ) THEN
529 * Scale X so that its components are less than or equal to
530 * BIGNUM in absolute value.
532 SCALE = BIGNUM / XMAX
533 CALL SSCAL( N, SCALE, X, 1 )
541 DO 100 J = JFIRST, JLAST, JINC
543 * Compute x(j) = b(j) / A(j,j), scaling x if necessary.
547 TJJS = AB( MAIND, J )*TSCAL
554 IF( TJJ.GT.SMLNUM ) THEN
556 * abs(A(j,j)) > SMLNUM:
558 IF( TJJ.LT.ONE ) THEN
559 IF( XJ.GT.TJJ*BIGNUM ) THEN
564 CALL SSCAL( N, REC, X, 1 )
569 X( J ) = X( J ) / TJJS
571 ELSE IF( TJJ.GT.ZERO ) THEN
573 * 0 < abs(A(j,j)) <= SMLNUM:
575 IF( XJ.GT.TJJ*BIGNUM ) THEN
577 * Scale x by (1/abs(x(j)))*abs(A(j,j))*BIGNUM
578 * to avoid overflow when dividing by A(j,j).
580 REC = ( TJJ*BIGNUM ) / XJ
581 IF( CNORM( J ).GT.ONE ) THEN
583 * Scale by 1/CNORM(j) to avoid overflow when
584 * multiplying x(j) times column j.
586 REC = REC / CNORM( J )
588 CALL SSCAL( N, REC, X, 1 )
592 X( J ) = X( J ) / TJJS
596 * A(j,j) = 0: Set x(1:n) = 0, x(j) = 1, and
597 * scale = 0, and compute a solution to A*x = 0.
609 * Scale x if necessary to avoid overflow when adding a
610 * multiple of column j of A.
614 IF( CNORM( J ).GT.( BIGNUM-XMAX )*REC ) THEN
616 * Scale x by 1/(2*abs(x(j))).
619 CALL SSCAL( N, REC, X, 1 )
622 ELSE IF( XJ*CNORM( J ).GT.( BIGNUM-XMAX ) ) THEN
626 CALL SSCAL( N, HALF, X, 1 )
634 * x(max(1,j-kd):j-1) := x(max(1,j-kd):j-1) -
635 * x(j)* A(max(1,j-kd):j-1,j)
637 JLEN = MIN( KD, J-1 )
638 CALL SAXPY( JLEN, -X( J )*TSCAL,
639 $ AB( KD+1-JLEN, J ), 1, X( J-JLEN ), 1 )
640 I = ISAMAX( J-1, X, 1 )
643 ELSE IF( J.LT.N ) THEN
646 * x(j+1:min(j+kd,n)) := x(j+1:min(j+kd,n)) -
647 * x(j) * A(j+1:min(j+kd,n),j)
649 JLEN = MIN( KD, N-J )
651 $ CALL SAXPY( JLEN, -X( J )*TSCAL, AB( 2, J ), 1,
653 I = J + ISAMAX( N-J, X( J+1 ), 1 )
662 DO 140 J = JFIRST, JLAST, JINC
664 * Compute x(j) = b(j) - sum A(k,j)*x(k).
669 REC = ONE / MAX( XMAX, ONE )
670 IF( CNORM( J ).GT.( BIGNUM-XJ )*REC ) THEN
672 * If x(j) could overflow, scale x by 1/(2*XMAX).
676 TJJS = AB( MAIND, J )*TSCAL
681 IF( TJJ.GT.ONE ) THEN
683 * Divide by A(j,j) when scaling x if A(j,j) > 1.
685 REC = MIN( ONE, REC*TJJ )
688 IF( REC.LT.ONE ) THEN
689 CALL SSCAL( N, REC, X, 1 )
696 IF( USCAL.EQ.ONE ) THEN
698 * If the scaling needed for A in the dot product is 1,
699 * call SDOT to perform the dot product.
702 JLEN = MIN( KD, J-1 )
703 SUMJ = SDOT( JLEN, AB( KD+1-JLEN, J ), 1,
706 JLEN = MIN( KD, N-J )
708 $ SUMJ = SDOT( JLEN, AB( 2, J ), 1, X( J+1 ), 1 )
712 * Otherwise, use in-line code for the dot product.
715 JLEN = MIN( KD, J-1 )
717 SUMJ = SUMJ + ( AB( KD+I-JLEN, J )*USCAL )*
721 JLEN = MIN( KD, N-J )
723 SUMJ = SUMJ + ( AB( I+1, J )*USCAL )*X( J+I )
728 IF( USCAL.EQ.TSCAL ) THEN
730 * Compute x(j) := ( x(j) - sumj ) / A(j,j) if 1/A(j,j)
731 * was not used to scale the dotproduct.
733 X( J ) = X( J ) - SUMJ
737 * Compute x(j) = x(j) / A(j,j), scaling if necessary.
739 TJJS = AB( MAIND, J )*TSCAL
746 IF( TJJ.GT.SMLNUM ) THEN
748 * abs(A(j,j)) > SMLNUM:
750 IF( TJJ.LT.ONE ) THEN
751 IF( XJ.GT.TJJ*BIGNUM ) THEN
753 * Scale X by 1/abs(x(j)).
756 CALL SSCAL( N, REC, X, 1 )
761 X( J ) = X( J ) / TJJS
762 ELSE IF( TJJ.GT.ZERO ) THEN
764 * 0 < abs(A(j,j)) <= SMLNUM:
766 IF( XJ.GT.TJJ*BIGNUM ) THEN
768 * Scale x by (1/abs(x(j)))*abs(A(j,j))*BIGNUM.
770 REC = ( TJJ*BIGNUM ) / XJ
771 CALL SSCAL( N, REC, X, 1 )
775 X( J ) = X( J ) / TJJS
778 * A(j,j) = 0: Set x(1:n) = 0, x(j) = 1, and
779 * scale = 0, and compute a solution to A**T*x = 0.
791 * Compute x(j) := x(j) / A(j,j) - sumj if the dot
792 * product has already been divided by 1/A(j,j).
794 X( J ) = X( J ) / TJJS - SUMJ
796 XMAX = MAX( XMAX, ABS( X( J ) ) )
799 SCALE = SCALE / TSCAL
802 * Scale the column norms by 1/TSCAL for return.
804 IF( TSCAL.NE.ONE ) THEN
805 CALL SSCAL( N, ONE / TSCAL, CNORM, 1 )