1       SUBROUTINE DGBEQUB( 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   AB( LDAB, * ), C( * ), R( * )
 20 *     ..
 21 *
 22 *  Purpose
 23 *  =======
 24 *
 25 *  DGBEQUB computes row and column scalings intended to equilibrate an
 26 *  M-by-N matrix A and reduce its condition number.  R returns the row
 27 *  scale factors and C the column scale factors, chosen to try to make
 28 *  the largest element in each row and column of the matrix B with
 29 *  elements B(i,j)=R(i)*A(i,j)*C(j) have an absolute value of at most
 30 *  the radix.
 31 *
 32 *  R(i) and C(j) are restricted to be a power of the radix between
 33 *  SMLNUM = smallest safe number and BIGNUM = largest safe number.  Use
 34 *  of these scaling factors is not guaranteed to reduce the condition
 35 *  number of A but works well in practice.
 36 *
 37 *  This routine differs from DGEEQU by restricting the scaling factors
 38 *  to a power of the radix.  Baring over- and underflow, scaling by
 39 *  these factors introduces no additional rounding errors.  However, the
 40 *  scaled entries' magnitured are no longer approximately 1 but lie
 41 *  between sqrt(radix) and 1/sqrt(radix).
 42 *
 43 *  Arguments
 44 *  =========
 45 *
 46 *  M       (input) INTEGER
 47 *          The number of rows of the matrix A.  M >= 0.
 48 *
 49 *  N       (input) INTEGER
 50 *          The number of columns of the matrix A.  N >= 0.
 51 *
 52 *  KL      (input) INTEGER
 53 *          The number of subdiagonals within the band of A.  KL >= 0.
 54 *
 55 *  KU      (input) INTEGER
 56 *          The number of superdiagonals within the band of A.  KU >= 0.
 57 *
 58 *  AB      (input) DOUBLE PRECISION array, dimension (LDAB,N)
 59 *          On entry, the matrix A in band storage, in rows 1 to KL+KU+1.
 60 *          The j-th column of A is stored in the j-th column of the
 61 *          array AB as follows:
 62 *          AB(KU+1+i-j,j) = A(i,j) for max(1,j-KU)<=i<=min(N,j+kl)
 63 *
 64 *  LDAB    (input) INTEGER
 65 *          The leading dimension of the array A.  LDAB >= max(1,M).
 66 *
 67 *  R       (output) DOUBLE PRECISION array, dimension (M)
 68 *          If INFO = 0 or INFO > M, R contains the row scale factors
 69 *          for A.
 70 *
 71 *  C       (output) DOUBLE PRECISION array, dimension (N)
 72 *          If INFO = 0,  C contains the column scale factors for A.
 73 *
 74 *  ROWCND  (output) DOUBLE PRECISION
 75 *          If INFO = 0 or INFO > M, ROWCND contains the ratio of the
 76 *          smallest R(i) to the largest R(i).  If ROWCND >= 0.1 and
 77 *          AMAX is neither too large nor too small, it is not worth
 78 *          scaling by R.
 79 *
 80 *  COLCND  (output) DOUBLE PRECISION
 81 *          If INFO = 0, COLCND contains the ratio of the smallest
 82 *          C(i) to the largest C(i).  If COLCND >= 0.1, it is not
 83 *          worth scaling by C.
 84 *
 85 *  AMAX    (output) DOUBLE PRECISION
 86 *          Absolute value of largest matrix element.  If AMAX is very
 87 *          close to overflow or very close to underflow, the matrix
 88 *          should be scaled.
 89 *
 90 *  INFO    (output) INTEGER
 91 *          = 0:  successful exit
 92 *          < 0:  if INFO = -i, the i-th argument had an illegal value
 93 *          > 0:  if INFO = i,  and i is
 94 *                <= M:  the i-th row of A is exactly zero
 95 *                >  M:  the (i-M)-th column of A is exactly zero
 96 *
 97 *  =====================================================================
 98 *
 99 *     .. Parameters ..
100       DOUBLE PRECISION   ONE, ZERO
101       PARAMETER          ( ONE = 1.0D+0, ZERO = 0.0D+0 )
102 *     ..
103 *     .. Local Scalars ..
104       INTEGER            I, J, KD
105       DOUBLE PRECISION   BIGNUM, RCMAX, RCMIN, SMLNUM, RADIX, LOGRDX
106 *     ..
107 *     .. External Functions ..
108       DOUBLE PRECISION   DLAMCH
109       EXTERNAL           DLAMCH
110 *     ..
111 *     .. External Subroutines ..
112       EXTERNAL           XERBLA
113 *     ..
114 *     .. Intrinsic Functions ..
115       INTRINSIC          ABSMAXMINLOG
116 *     ..
117 *     .. Executable Statements ..
118 *
119 *     Test the input parameters.
120 *
121       INFO = 0
122       IF( M.LT.0 ) THEN
123          INFO = -1
124       ELSE IF( N.LT.0 ) THEN
125          INFO = -2
126       ELSE IF( KL.LT.0 ) THEN
127          INFO = -3
128       ELSE IF( KU.LT.0 ) THEN
129          INFO = -4
130       ELSE IF( LDAB.LT.KL+KU+1 ) THEN
131          INFO = -6
132       END IF
133       IF( INFO.NE.0 ) THEN
134          CALL XERBLA( 'DGBEQUB'-INFO )
135          RETURN
136       END IF
137 *
138 *     Quick return if possible.
139 *
140       IF( M.EQ.0 .OR. N.EQ.0 ) THEN
141          ROWCND = ONE
142          COLCND = ONE
143          AMAX = ZERO
144          RETURN
145       END IF
146 *
147 *     Get machine constants.  Assume SMLNUM is a power of the radix.
148 *
149       SMLNUM = DLAMCH( 'S' )
150       BIGNUM = ONE / SMLNUM
151       RADIX = DLAMCH( 'B' )
152       LOGRDX = LOG(RADIX)
153 *
154 *     Compute row scale factors.
155 *
156       DO 10 I = 1, M
157          R( I ) = ZERO
158    10 CONTINUE
159 *
160 *     Find the maximum element in each row.
161 *
162       KD = KU + 1
163       DO 30 J = 1, N
164          DO 20 I = MAX( J-KU, 1 ), MIN( J+KL, M )
165             R( I ) = MAX( R( I ), ABS( AB( KD+I-J, J ) ) )
166    20    CONTINUE
167    30 CONTINUE
168       DO I = 1, M
169          IF( R( I ).GT.ZERO ) THEN
170             R( I ) = RADIX**INTLOG( R( I ) ) / LOGRDX )
171          END IF
172       END DO
173 *
174 *     Find the maximum and minimum scale factors.
175 *
176       RCMIN = BIGNUM
177       RCMAX = ZERO
178       DO 40 I = 1, M
179          RCMAX = MAX( RCMAX, R( I ) )
180          RCMIN = MIN( RCMIN, R( I ) )
181    40 CONTINUE
182       AMAX = RCMAX
183 *
184       IF( RCMIN.EQ.ZERO ) THEN
185 *
186 *        Find the first zero scale factor and return an error code.
187 *
188          DO 50 I = 1, M
189             IF( R( I ).EQ.ZERO ) THEN
190                INFO = I
191                RETURN
192             END IF
193    50    CONTINUE
194       ELSE
195 *
196 *        Invert the scale factors.
197 *
198          DO 60 I = 1, M
199             R( I ) = ONE / MINMAX( R( I ), SMLNUM ), BIGNUM )
200    60    CONTINUE
201 *
202 *        Compute ROWCND = min(R(I)) / max(R(I)).
203 *
204          ROWCND = MAX( RCMIN, SMLNUM ) / MIN( RCMAX, BIGNUM )
205       END IF
206 *
207 *     Compute column scale factors.
208 *
209       DO 70 J = 1, N
210          C( J ) = ZERO
211    70 CONTINUE
212 *
213 *     Find the maximum element in each column,
214 *     assuming the row scaling computed above.
215 *
216       DO 90 J = 1, N
217          DO 80 I = MAX( J-KU, 1 ), MIN( J+KL, M )
218             C( J ) = MAX( C( J ), ABS( AB( KD+I-J, J ) )*R( I ) )
219    80    CONTINUE
220          IF( C( J ).GT.ZERO ) THEN
221             C( J ) = RADIX**INTLOG( C( J ) ) / LOGRDX )
222          END IF
223    90 CONTINUE
224 *
225 *     Find the maximum and minimum scale factors.
226 *
227       RCMIN = BIGNUM
228       RCMAX = ZERO
229       DO 100 J = 1, N
230          RCMIN = MIN( RCMIN, C( J ) )
231          RCMAX = MAX( RCMAX, C( J ) )
232   100 CONTINUE
233 *
234       IF( RCMIN.EQ.ZERO ) THEN
235 *
236 *        Find the first zero scale factor and return an error code.
237 *
238          DO 110 J = 1, N
239             IF( C( J ).EQ.ZERO ) THEN
240                INFO = M + J
241                RETURN
242             END IF
243   110    CONTINUE
244       ELSE
245 *
246 *        Invert the scale factors.
247 *
248          DO 120 J = 1, N
249             C( J ) = ONE / MINMAX( C( J ), SMLNUM ), BIGNUM )
250   120    CONTINUE
251 *
252 *        Compute COLCND = min(C(J)) / max(C(J)).
253 *
254          COLCND = MAX( RCMIN, SMLNUM ) / MIN( RCMAX, BIGNUM )
255       END IF
256 *
257       RETURN
258 *
259 *     End of DGBEQUB
260 *
261       END