1       SUBROUTINE ZGBEQUB( M, N, KL, KU, AB, LDAB, R, C, ROWCND, COLCND,
  2      $                    AMAX, INFO )
  3 *
  4 *     -- LAPACK routine (version 3.2)                                 --
  5 *     -- Contributed by James Demmel, Deaglan Halligan, Yozo Hida and --
  6 *     -- Jason Riedy of Univ. of California Berkeley.                 --
  7 *     -- November 2008                                                --
  8 *
  9 *     -- LAPACK is a software package provided by Univ. of Tennessee, --
 10 *     -- Univ. of California Berkeley and NAG Ltd.                    --
 11 *
 12       IMPLICIT NONE
 13 *     ..
 14 *     .. Scalar Arguments ..
 15       INTEGER            INFO, KL, KU, LDAB, M, N
 16       DOUBLE PRECISION   AMAX, COLCND, ROWCND
 17 *     ..
 18 *     .. Array Arguments ..
 19       DOUBLE PRECISION   C( * ), R( * )
 20       COMPLEX*16         AB( LDAB, * )
 21 *     ..
 22 *
 23 *  Purpose
 24 *  =======
 25 *
 26 *  ZGBEQUB computes row and column scalings intended to equilibrate an
 27 *  M-by-N matrix A and reduce its condition number.  R returns the row
 28 *  scale factors and C the column scale factors, chosen to try to make
 29 *  the largest element in each row and column of the matrix B with
 30 *  elements B(i,j)=R(i)*A(i,j)*C(j) have an absolute value of at most
 31 *  the radix.
 32 *
 33 *  R(i) and C(j) are restricted to be a power of the radix between
 34 *  SMLNUM = smallest safe number and BIGNUM = largest safe number.  Use
 35 *  of these scaling factors is not guaranteed to reduce the condition
 36 *  number of A but works well in practice.
 37 *
 38 *  This routine differs from ZGEEQU by restricting the scaling factors
 39 *  to a power of the radix.  Baring over- and underflow, scaling by
 40 *  these factors introduces no additional rounding errors.  However, the
 41 *  scaled entries' magnitured are no longer approximately 1 but lie
 42 *  between sqrt(radix) and 1/sqrt(radix).
 43 *
 44 *  Arguments
 45 *  =========
 46 *
 47 *  M       (input) INTEGER
 48 *          The number of rows of the matrix A.  M >= 0.
 49 *
 50 *  N       (input) INTEGER
 51 *          The number of columns of the matrix A.  N >= 0.
 52 *
 53 *  KL      (input) INTEGER
 54 *          The number of subdiagonals within the band of A.  KL >= 0.
 55 *
 56 *  KU      (input) INTEGER
 57 *          The number of superdiagonals within the band of A.  KU >= 0.
 58 *
 59 *  AB      (input) DOUBLE PRECISION array, dimension (LDAB,N)
 60 *          On entry, the matrix A in band storage, in rows 1 to KL+KU+1.
 61 *          The j-th column of A is stored in the j-th column of the
 62 *          array AB as follows:
 63 *          AB(KU+1+i-j,j) = A(i,j) for max(1,j-KU)<=i<=min(N,j+kl)
 64 *
 65 *  LDAB    (input) INTEGER
 66 *          The leading dimension of the array A.  LDAB >= max(1,M).
 67 *
 68 *  R       (output) DOUBLE PRECISION array, dimension (M)
 69 *          If INFO = 0 or INFO > M, R contains the row scale factors
 70 *          for A.
 71 *
 72 *  C       (output) DOUBLE PRECISION array, dimension (N)
 73 *          If INFO = 0,  C contains the column scale factors for A.
 74 *
 75 *  ROWCND  (output) DOUBLE PRECISION
 76 *          If INFO = 0 or INFO > M, ROWCND contains the ratio of the
 77 *          smallest R(i) to the largest R(i).  If ROWCND >= 0.1 and
 78 *          AMAX is neither too large nor too small, it is not worth
 79 *          scaling by R.
 80 *
 81 *  COLCND  (output) DOUBLE PRECISION
 82 *          If INFO = 0, COLCND contains the ratio of the smallest
 83 *          C(i) to the largest C(i).  If COLCND >= 0.1, it is not
 84 *          worth scaling by C.
 85 *
 86 *  AMAX    (output) DOUBLE PRECISION
 87 *          Absolute value of largest matrix element.  If AMAX is very
 88 *          close to overflow or very close to underflow, the matrix
 89 *          should be scaled.
 90 *
 91 *  INFO    (output) INTEGER
 92 *          = 0:  successful exit
 93 *          < 0:  if INFO = -i, the i-th argument had an illegal value
 94 *          > 0:  if INFO = i,  and i is
 95 *                <= M:  the i-th row of A is exactly zero
 96 *                >  M:  the (i-M)-th column of A is exactly zero
 97 *
 98 *  =====================================================================
 99 *
100 *     .. Parameters ..
101       DOUBLE PRECISION   ONE, ZERO
102       PARAMETER          ( ONE = 1.0D+0, ZERO = 0.0D+0 )
103 *     ..
104 *     .. Local Scalars ..
105       INTEGER            I, J, KD
106       DOUBLE PRECISION   BIGNUM, RCMAX, RCMIN, SMLNUM, RADIX,
107      $                   LOGRDX
108       COMPLEX*16         ZDUM
109 *     ..
110 *     .. External Functions ..
111       DOUBLE PRECISION   DLAMCH
112       EXTERNAL           DLAMCH
113 *     ..
114 *     .. External Subroutines ..
115       EXTERNAL           XERBLA
116 *     ..
117 *     .. Intrinsic Functions ..
118       INTRINSIC          ABSMAXMINLOG, REAL, DIMAG
119 *     ..
120 *     .. Statement Functions ..
121       DOUBLE PRECISION   CABS1
122 *     ..
123 *     .. Statement Function definitions ..
124       CABS1( ZDUM ) = ABSDBLE( ZDUM ) ) + ABSDIMAG( ZDUM ) )
125 *     ..
126 *     .. Executable Statements ..
127 *
128 *     Test the input parameters.
129 *
130       INFO = 0
131       IF( M.LT.0 ) THEN
132          INFO = -1
133       ELSE IF( N.LT.0 ) THEN
134          INFO = -2
135       ELSE IF( KL.LT.0 ) THEN
136          INFO = -3
137       ELSE IF( KU.LT.0 ) THEN
138          INFO = -4
139       ELSE IF( LDAB.LT.KL+KU+1 ) THEN
140          INFO = -6
141       END IF
142       IF( INFO.NE.0 ) THEN
143          CALL XERBLA( 'ZGBEQUB'-INFO )
144          RETURN
145       END IF
146 *
147 *     Quick return if possible.
148 *
149       IF( M.EQ.0 .OR. N.EQ.0 ) THEN
150          ROWCND = ONE
151          COLCND = ONE
152          AMAX = ZERO
153          RETURN
154       END IF
155 *
156 *     Get machine constants.  Assume SMLNUM is a power of the radix.
157 *
158       SMLNUM = DLAMCH( 'S' )
159       BIGNUM = ONE / SMLNUM
160       RADIX = DLAMCH( 'B' )
161       LOGRDX = LOG(RADIX)
162 *
163 *     Compute row scale factors.
164 *
165       DO 10 I = 1, M
166          R( I ) = ZERO
167    10 CONTINUE
168 *
169 *     Find the maximum element in each row.
170 *
171       KD = KU + 1
172       DO 30 J = 1, N
173          DO 20 I = MAX( J-KU, 1 ), MIN( J+KL, M )
174             R( I ) = MAX( R( I ), CABS1( AB( KD+I-J, J ) ) )
175    20    CONTINUE
176    30 CONTINUE
177       DO I = 1, M
178          IF( R( I ).GT.ZERO ) THEN
179             R( I ) = RADIX**INTLOG( R( I ) ) / LOGRDX )
180          END IF
181       END DO
182 *
183 *     Find the maximum and minimum scale factors.
184 *
185       RCMIN = BIGNUM
186       RCMAX = ZERO
187       DO 40 I = 1, M
188          RCMAX = MAX( RCMAX, R( I ) )
189          RCMIN = MIN( RCMIN, R( I ) )
190    40 CONTINUE
191       AMAX = RCMAX
192 *
193       IF( RCMIN.EQ.ZERO ) THEN
194 *
195 *        Find the first zero scale factor and return an error code.
196 *
197          DO 50 I = 1, M
198             IF( R( I ).EQ.ZERO ) THEN
199                INFO = I
200                RETURN
201             END IF
202    50    CONTINUE
203       ELSE
204 *
205 *        Invert the scale factors.
206 *
207          DO 60 I = 1, M
208             R( I ) = ONE / MINMAX( R( I ), SMLNUM ), BIGNUM )
209    60    CONTINUE
210 *
211 *        Compute ROWCND = min(R(I)) / max(R(I)).
212 *
213          ROWCND = MAX( RCMIN, SMLNUM ) / MIN( RCMAX, BIGNUM )
214       END IF
215 *
216 *     Compute column scale factors.
217 *
218       DO 70 J = 1, N
219          C( J ) = ZERO
220    70 CONTINUE
221 *
222 *     Find the maximum element in each column,
223 *     assuming the row scaling computed above.
224 *
225       DO 90 J = 1, N
226          DO 80 I = MAX( J-KU, 1 ), MIN( J+KL, M )
227             C( J ) = MAX( C( J ), CABS1( AB( KD+I-J, J ) )*R( I ) )
228    80    CONTINUE
229          IF( C( J ).GT.ZERO ) THEN
230             C( J ) = RADIX**INTLOG( C( J ) ) / LOGRDX )
231          END IF
232    90 CONTINUE
233 *
234 *     Find the maximum and minimum scale factors.
235 *
236       RCMIN = BIGNUM
237       RCMAX = ZERO
238       DO 100 J = 1, N
239          RCMIN = MIN( RCMIN, C( J ) )
240          RCMAX = MAX( RCMAX, C( J ) )
241   100 CONTINUE
242 *
243       IF( RCMIN.EQ.ZERO ) THEN
244 *
245 *        Find the first zero scale factor and return an error code.
246 *
247          DO 110 J = 1, N
248             IF( C( J ).EQ.ZERO ) THEN
249                INFO = M + J
250                RETURN
251             END IF
252   110    CONTINUE
253       ELSE
254 *
255 *        Invert the scale factors.
256 *
257          DO 120 J = 1, N
258             C( J ) = ONE / MINMAX( C( J ), SMLNUM ), BIGNUM )
259   120    CONTINUE
260 *
261 *        Compute COLCND = min(C(J)) / max(C(J)).
262 *
263          COLCND = MAX( RCMIN, SMLNUM ) / MIN( RCMAX, BIGNUM )
264       END IF
265 *
266       RETURN
267 *
268 *     End of ZGBEQUB
269 *
270       END