1       SUBROUTINE DLA_LIN_BERR ( N, NZ, NRHS, RES, AYB, BERR )
 2 *
 3 *     -- LAPACK routine (version 3.2.2)                                 --
 4 *     -- Contributed by James Demmel, Deaglan Halligan, Yozo Hida and --
 5 *     -- Jason Riedy of Univ. of California Berkeley.                 --
 6 *     -- June 2010                                                    --
 7 *
 8 *     -- LAPACK is a software package provided by Univ. of Tennessee, --
 9 *     -- Univ. of California Berkeley and NAG Ltd.                    --
10 *
11       IMPLICIT NONE
12 *     ..
13 *     .. Scalar Arguments ..
14       INTEGER            N, NZ, NRHS
15 *     ..
16 *     .. Array Arguments ..
17       DOUBLE PRECISION   AYB( N, NRHS ), BERR( NRHS )
18       DOUBLE PRECISION   RES( N, NRHS )
19 *     ..
20 *
21 *  Purpose
22 *  =======
23 *
24 *     DLA_LIN_BERR computes component-wise relative backward error from
25 *     the formula
26 *         max(i) ( abs(R(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
27 *     where abs(Z) is the component-wise absolute value of the matrix
28 *     or vector Z.
29 *
30 *  Arguments
31 *  ==========
32 *
33 *     N       (input) INTEGER
34 *     The number of linear equations, i.e., the order of the
35 *     matrix A.  N >= 0.
36 *
37 *     NZ      (input) INTEGER
38 *     We add (NZ+1)*SLAMCH( 'Safe minimum' ) to R(i) in the numerator to
39 *     guard against spuriously zero residuals. Default value is N.
40 *
41 *     NRHS    (input) INTEGER
42 *     The number of right hand sides, i.e., the number of columns
43 *     of the matrices AYB, RES, and BERR.  NRHS >= 0.
44 *
45 *     RES     (input) DOUBLE PRECISION array, dimension (N,NRHS)
46 *     The residual matrix, i.e., the matrix R in the relative backward
47 *     error formula above.
48 *
49 *     AYB     (input) DOUBLE PRECISION array, dimension (N, NRHS)
50 *     The denominator in the relative backward error formula above, i.e.,
51 *     the matrix abs(op(A_s))*abs(Y) + abs(B_s). The matrices A, Y, and B
52 *     are from iterative refinement (see dla_gerfsx_extended.f).
53 *     
54 *     BERR    (output) DOUBLE PRECISION array, dimension (NRHS)
55 *     The component-wise relative backward error from the formula above.
56 *
57 *  =====================================================================
58 *
59 *     .. Local Scalars ..
60       DOUBLE PRECISION   TMP
61       INTEGER            I, J
62 *     ..
63 *     .. Intrinsic Functions ..
64       INTRINSIC          ABSMAX
65 *     ..
66 *     .. External Functions ..
67       EXTERNAL           DLAMCH
68       DOUBLE PRECISION   DLAMCH
69       DOUBLE PRECISION   SAFE1
70 *     ..
71 *     .. Executable Statements ..
72 *
73 *     Adding SAFE1 to the numerator guards against spuriously zero
74 *     residuals.  A similar safeguard is in the SLA_yyAMV routine used
75 *     to compute AYB.
76 *
77       SAFE1 = DLAMCH( 'Safe minimum' )
78       SAFE1 = (NZ+1)*SAFE1
79 
80       DO J = 1, NRHS
81          BERR(J) = 0.0D+0
82          DO I = 1, N
83             IF (AYB(I,J) .NE. 0.0D+0THEN
84                TMP = (SAFE1+ABS(RES(I,J)))/AYB(I,J)
85                BERR(J) = MAX( BERR(J), TMP )
86             END IF
87 *
88 *     If AYB is exactly 0.0 (and if computed by SLA_yyAMV), then we know
89 *     the true residual also must be exactly 0.0.
90 *
91          END DO
92       END DO
93       END