File:  [local] / rpl / lapack / lapack / dporfs.f
Revision 1.4: download - view: text, annotated - select for diffs - revision graph
Fri Aug 6 15:32:32 2010 UTC (13 years, 9 months ago) by bertrand
Branches: MAIN
CVS tags: HEAD
Cohérence

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

CVSweb interface <joel.bertrand@systella.fr>