File:  [local] / rpl / lapack / lapack / dsprfs.f
Revision 1.6: download - view: text, annotated - select for diffs - revision graph
Fri Aug 13 21:03:57 2010 UTC (13 years, 9 months ago) by bertrand
Branches: MAIN
CVS tags: rpl-4_0_19, rpl-4_0_18, HEAD
Patches pour OS/2

    1:       SUBROUTINE DSPRFS( UPLO, N, NRHS, AP, AFP, IPIV, B, LDB, X, LDX,
    2:      $                   FERR, BERR, WORK, IWORK, INFO )
    3: *
    4: *  -- LAPACK routine (version 3.2) --
    5: *  -- LAPACK is a software package provided by Univ. of Tennessee,    --
    6: *  -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
    7: *     November 2006
    8: *
    9: *     Modified to call DLACN2 in place of DLACON, 5 Feb 03, SJH.
   10: *
   11: *     .. Scalar Arguments ..
   12:       CHARACTER          UPLO
   13:       INTEGER            INFO, LDB, LDX, N, NRHS
   14: *     ..
   15: *     .. Array Arguments ..
   16:       INTEGER            IPIV( * ), IWORK( * )
   17:       DOUBLE PRECISION   AFP( * ), AP( * ), B( LDB, * ), BERR( * ),
   18:      $                   FERR( * ), WORK( * ), X( LDX, * )
   19: *     ..
   20: *
   21: *  Purpose
   22: *  =======
   23: *
   24: *  DSPRFS improves the computed solution to a system of linear
   25: *  equations when the coefficient matrix is symmetric indefinite
   26: *  and packed, and provides error bounds and backward error estimates
   27: *  for the solution.
   28: *
   29: *  Arguments
   30: *  =========
   31: *
   32: *  UPLO    (input) CHARACTER*1
   33: *          = 'U':  Upper triangle of A is stored;
   34: *          = 'L':  Lower triangle of A is stored.
   35: *
   36: *  N       (input) INTEGER
   37: *          The order of the matrix A.  N >= 0.
   38: *
   39: *  NRHS    (input) INTEGER
   40: *          The number of right hand sides, i.e., the number of columns
   41: *          of the matrices B and X.  NRHS >= 0.
   42: *
   43: *  AP      (input) DOUBLE PRECISION array, dimension (N*(N+1)/2)
   44: *          The upper or lower triangle of the symmetric matrix A, packed
   45: *          columnwise in a linear array.  The j-th column of A is stored
   46: *          in the array AP as follows:
   47: *          if UPLO = 'U', AP(i + (j-1)*j/2) = A(i,j) for 1<=i<=j;
   48: *          if UPLO = 'L', AP(i + (j-1)*(2*n-j)/2) = A(i,j) for j<=i<=n.
   49: *
   50: *  AFP     (input) DOUBLE PRECISION array, dimension (N*(N+1)/2)
   51: *          The factored form of the matrix A.  AFP contains the block
   52: *          diagonal matrix D and the multipliers used to obtain the
   53: *          factor U or L from the factorization A = U*D*U**T or
   54: *          A = L*D*L**T as computed by DSPTRF, stored as a packed
   55: *          triangular matrix.
   56: *
   57: *  IPIV    (input) INTEGER array, dimension (N)
   58: *          Details of the interchanges and the block structure of D
   59: *          as determined by DSPTRF.
   60: *
   61: *  B       (input) DOUBLE PRECISION array, dimension (LDB,NRHS)
   62: *          The right hand side matrix B.
   63: *
   64: *  LDB     (input) INTEGER
   65: *          The leading dimension of the array B.  LDB >= max(1,N).
   66: *
   67: *  X       (input/output) DOUBLE PRECISION array, dimension (LDX,NRHS)
   68: *          On entry, the solution matrix X, as computed by DSPTRS.
   69: *          On exit, the improved solution matrix X.
   70: *
   71: *  LDX     (input) INTEGER
   72: *          The leading dimension of the array X.  LDX >= max(1,N).
   73: *
   74: *  FERR    (output) DOUBLE PRECISION array, dimension (NRHS)
   75: *          The estimated forward error bound for each solution vector
   76: *          X(j) (the j-th column of the solution matrix X).
   77: *          If XTRUE is the true solution corresponding to X(j), FERR(j)
   78: *          is an estimated upper bound for the magnitude of the largest
   79: *          element in (X(j) - XTRUE) divided by the magnitude of the
   80: *          largest element in X(j).  The estimate is as reliable as
   81: *          the estimate for RCOND, and is almost always a slight
   82: *          overestimate of the true error.
   83: *
   84: *  BERR    (output) DOUBLE PRECISION array, dimension (NRHS)
   85: *          The componentwise relative backward error of each solution
   86: *          vector X(j) (i.e., the smallest relative change in
   87: *          any element of A or B that makes X(j) an exact solution).
   88: *
   89: *  WORK    (workspace) DOUBLE PRECISION array, dimension (3*N)
   90: *
   91: *  IWORK   (workspace) INTEGER array, dimension (N)
   92: *
   93: *  INFO    (output) INTEGER
   94: *          = 0:  successful exit
   95: *          < 0:  if INFO = -i, the i-th argument had an illegal value
   96: *
   97: *  Internal Parameters
   98: *  ===================
   99: *
  100: *  ITMAX is the maximum number of steps of iterative refinement.
  101: *
  102: *  =====================================================================
  103: *
  104: *     .. Parameters ..
  105:       INTEGER            ITMAX
  106:       PARAMETER          ( ITMAX = 5 )
  107:       DOUBLE PRECISION   ZERO
  108:       PARAMETER          ( ZERO = 0.0D+0 )
  109:       DOUBLE PRECISION   ONE
  110:       PARAMETER          ( ONE = 1.0D+0 )
  111:       DOUBLE PRECISION   TWO
  112:       PARAMETER          ( TWO = 2.0D+0 )
  113:       DOUBLE PRECISION   THREE
  114:       PARAMETER          ( THREE = 3.0D+0 )
  115: *     ..
  116: *     .. Local Scalars ..
  117:       LOGICAL            UPPER
  118:       INTEGER            COUNT, I, IK, J, K, KASE, KK, NZ
  119:       DOUBLE PRECISION   EPS, LSTRES, S, SAFE1, SAFE2, SAFMIN, XK
  120: *     ..
  121: *     .. Local Arrays ..
  122:       INTEGER            ISAVE( 3 )
  123: *     ..
  124: *     .. External Subroutines ..
  125:       EXTERNAL           DAXPY, DCOPY, DLACN2, DSPMV, DSPTRS, XERBLA
  126: *     ..
  127: *     .. Intrinsic Functions ..
  128:       INTRINSIC          ABS, MAX
  129: *     ..
  130: *     .. External Functions ..
  131:       LOGICAL            LSAME
  132:       DOUBLE PRECISION   DLAMCH
  133:       EXTERNAL           LSAME, DLAMCH
  134: *     ..
  135: *     .. Executable Statements ..
  136: *
  137: *     Test the input parameters.
  138: *
  139:       INFO = 0
  140:       UPPER = LSAME( UPLO, 'U' )
  141:       IF( .NOT.UPPER .AND. .NOT.LSAME( UPLO, 'L' ) ) THEN
  142:          INFO = -1
  143:       ELSE IF( N.LT.0 ) THEN
  144:          INFO = -2
  145:       ELSE IF( NRHS.LT.0 ) THEN
  146:          INFO = -3
  147:       ELSE IF( LDB.LT.MAX( 1, N ) ) THEN
  148:          INFO = -8
  149:       ELSE IF( LDX.LT.MAX( 1, N ) ) THEN
  150:          INFO = -10
  151:       END IF
  152:       IF( INFO.NE.0 ) THEN
  153:          CALL XERBLA( 'DSPRFS', -INFO )
  154:          RETURN
  155:       END IF
  156: *
  157: *     Quick return if possible
  158: *
  159:       IF( N.EQ.0 .OR. NRHS.EQ.0 ) THEN
  160:          DO 10 J = 1, NRHS
  161:             FERR( J ) = ZERO
  162:             BERR( J ) = ZERO
  163:    10    CONTINUE
  164:          RETURN
  165:       END IF
  166: *
  167: *     NZ = maximum number of nonzero elements in each row of A, plus 1
  168: *
  169:       NZ = N + 1
  170:       EPS = DLAMCH( 'Epsilon' )
  171:       SAFMIN = DLAMCH( 'Safe minimum' )
  172:       SAFE1 = NZ*SAFMIN
  173:       SAFE2 = SAFE1 / EPS
  174: *
  175: *     Do for each right hand side
  176: *
  177:       DO 140 J = 1, NRHS
  178: *
  179:          COUNT = 1
  180:          LSTRES = THREE
  181:    20    CONTINUE
  182: *
  183: *        Loop until stopping criterion is satisfied.
  184: *
  185: *        Compute residual R = B - A * X
  186: *
  187:          CALL DCOPY( N, B( 1, J ), 1, WORK( N+1 ), 1 )
  188:          CALL DSPMV( UPLO, N, -ONE, AP, X( 1, J ), 1, ONE, WORK( N+1 ),
  189:      $               1 )
  190: *
  191: *        Compute componentwise relative backward error from formula
  192: *
  193: *        max(i) ( abs(R(i)) / ( abs(A)*abs(X) + abs(B) )(i) )
  194: *
  195: *        where abs(Z) is the componentwise absolute value of the matrix
  196: *        or vector Z.  If the i-th component of the denominator is less
  197: *        than SAFE2, then SAFE1 is added to the i-th components of the
  198: *        numerator and denominator before dividing.
  199: *
  200:          DO 30 I = 1, N
  201:             WORK( I ) = ABS( B( I, J ) )
  202:    30    CONTINUE
  203: *
  204: *        Compute abs(A)*abs(X) + abs(B).
  205: *
  206:          KK = 1
  207:          IF( UPPER ) THEN
  208:             DO 50 K = 1, N
  209:                S = ZERO
  210:                XK = ABS( X( K, J ) )
  211:                IK = KK
  212:                DO 40 I = 1, K - 1
  213:                   WORK( I ) = WORK( I ) + ABS( AP( IK ) )*XK
  214:                   S = S + ABS( AP( IK ) )*ABS( X( I, J ) )
  215:                   IK = IK + 1
  216:    40          CONTINUE
  217:                WORK( K ) = WORK( K ) + ABS( AP( KK+K-1 ) )*XK + S
  218:                KK = KK + K
  219:    50       CONTINUE
  220:          ELSE
  221:             DO 70 K = 1, N
  222:                S = ZERO
  223:                XK = ABS( X( K, J ) )
  224:                WORK( K ) = WORK( K ) + ABS( AP( KK ) )*XK
  225:                IK = KK + 1
  226:                DO 60 I = K + 1, N
  227:                   WORK( I ) = WORK( I ) + ABS( AP( IK ) )*XK
  228:                   S = S + ABS( AP( IK ) )*ABS( X( I, J ) )
  229:                   IK = IK + 1
  230:    60          CONTINUE
  231:                WORK( K ) = WORK( K ) + S
  232:                KK = KK + ( N-K+1 )
  233:    70       CONTINUE
  234:          END IF
  235:          S = ZERO
  236:          DO 80 I = 1, N
  237:             IF( WORK( I ).GT.SAFE2 ) THEN
  238:                S = MAX( S, ABS( WORK( N+I ) ) / WORK( I ) )
  239:             ELSE
  240:                S = MAX( S, ( ABS( WORK( N+I ) )+SAFE1 ) /
  241:      $             ( WORK( I )+SAFE1 ) )
  242:             END IF
  243:    80    CONTINUE
  244:          BERR( J ) = S
  245: *
  246: *        Test stopping criterion. Continue iterating if
  247: *           1) The residual BERR(J) is larger than machine epsilon, and
  248: *           2) BERR(J) decreased by at least a factor of 2 during the
  249: *              last iteration, and
  250: *           3) At most ITMAX iterations tried.
  251: *
  252:          IF( BERR( J ).GT.EPS .AND. TWO*BERR( J ).LE.LSTRES .AND.
  253:      $       COUNT.LE.ITMAX ) THEN
  254: *
  255: *           Update solution and try again.
  256: *
  257:             CALL DSPTRS( UPLO, N, 1, AFP, IPIV, WORK( N+1 ), N, INFO )
  258:             CALL DAXPY( N, ONE, WORK( N+1 ), 1, X( 1, J ), 1 )
  259:             LSTRES = BERR( J )
  260:             COUNT = COUNT + 1
  261:             GO TO 20
  262:          END IF
  263: *
  264: *        Bound error from formula
  265: *
  266: *        norm(X - XTRUE) / norm(X) .le. FERR =
  267: *        norm( abs(inv(A))*
  268: *           ( abs(R) + NZ*EPS*( abs(A)*abs(X)+abs(B) ))) / norm(X)
  269: *
  270: *        where
  271: *          norm(Z) is the magnitude of the largest component of Z
  272: *          inv(A) is the inverse of A
  273: *          abs(Z) is the componentwise absolute value of the matrix or
  274: *             vector Z
  275: *          NZ is the maximum number of nonzeros in any row of A, plus 1
  276: *          EPS is machine epsilon
  277: *
  278: *        The i-th component of abs(R)+NZ*EPS*(abs(A)*abs(X)+abs(B))
  279: *        is incremented by SAFE1 if the i-th component of
  280: *        abs(A)*abs(X) + abs(B) is less than SAFE2.
  281: *
  282: *        Use DLACN2 to estimate the infinity-norm of the matrix
  283: *           inv(A) * diag(W),
  284: *        where W = abs(R) + NZ*EPS*( abs(A)*abs(X)+abs(B) )))
  285: *
  286:          DO 90 I = 1, N
  287:             IF( WORK( I ).GT.SAFE2 ) THEN
  288:                WORK( I ) = ABS( WORK( N+I ) ) + NZ*EPS*WORK( I )
  289:             ELSE
  290:                WORK( I ) = ABS( WORK( N+I ) ) + NZ*EPS*WORK( I ) + SAFE1
  291:             END IF
  292:    90    CONTINUE
  293: *
  294:          KASE = 0
  295:   100    CONTINUE
  296:          CALL DLACN2( N, WORK( 2*N+1 ), WORK( N+1 ), IWORK, FERR( J ),
  297:      $                KASE, ISAVE )
  298:          IF( KASE.NE.0 ) THEN
  299:             IF( KASE.EQ.1 ) THEN
  300: *
  301: *              Multiply by diag(W)*inv(A').
  302: *
  303:                CALL DSPTRS( UPLO, N, 1, AFP, IPIV, WORK( N+1 ), N,
  304:      $                      INFO )
  305:                DO 110 I = 1, N
  306:                   WORK( N+I ) = WORK( I )*WORK( N+I )
  307:   110          CONTINUE
  308:             ELSE IF( KASE.EQ.2 ) THEN
  309: *
  310: *              Multiply by inv(A)*diag(W).
  311: *
  312:                DO 120 I = 1, N
  313:                   WORK( N+I ) = WORK( I )*WORK( N+I )
  314:   120          CONTINUE
  315:                CALL DSPTRS( UPLO, N, 1, AFP, IPIV, WORK( N+1 ), N,
  316:      $                      INFO )
  317:             END IF
  318:             GO TO 100
  319:          END IF
  320: *
  321: *        Normalize error.
  322: *
  323:          LSTRES = ZERO
  324:          DO 130 I = 1, N
  325:             LSTRES = MAX( LSTRES, ABS( X( I, J ) ) )
  326:   130    CONTINUE
  327:          IF( LSTRES.NE.ZERO )
  328:      $      FERR( J ) = FERR( J ) / LSTRES
  329: *
  330:   140 CONTINUE
  331: *
  332:       RETURN
  333: *
  334: *     End of DSPRFS
  335: *
  336:       END

CVSweb interface <joel.bertrand@systella.fr>