3 * =========== DOCUMENTATION ===========
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
9 *> Download ZPORFSX + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/zporfsx.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/zporfsx.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/zporfsx.f">
21 * SUBROUTINE ZPORFSX( UPLO, EQUED, N, NRHS, A, LDA, AF, LDAF, S, B,
22 * LDB, X, LDX, RCOND, BERR, N_ERR_BNDS,
23 * ERR_BNDS_NORM, ERR_BNDS_COMP, NPARAMS, PARAMS,
26 * .. Scalar Arguments ..
27 * CHARACTER UPLO, EQUED
28 * INTEGER INFO, LDA, LDAF, LDB, LDX, N, NRHS, NPARAMS,
30 * DOUBLE PRECISION RCOND
32 * .. Array Arguments ..
33 * COMPLEX*16 A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
34 * $ X( LDX, * ), WORK( * )
35 * DOUBLE PRECISION RWORK( * ), S( * ), PARAMS(*), BERR( * ),
36 * $ ERR_BNDS_NORM( NRHS, * ),
37 * $ ERR_BNDS_COMP( NRHS, * )
46 *> ZPORFSX improves the computed solution to a system of linear
47 *> equations when the coefficient matrix is symmetric positive
48 *> definite, and provides error bounds and backward error estimates
49 *> for the solution. In addition to normwise error bound, the code
50 *> provides maximum componentwise error bound if possible. See
51 *> comments for ERR_BNDS_NORM and ERR_BNDS_COMP for details of the
54 *> The original system of linear equations may have been equilibrated
55 *> before calling this routine, as described by arguments EQUED and S
56 *> below. In this case, the solution and error bounds returned are
57 *> for the original unequilibrated system.
64 *> Some optional parameters are bundled in the PARAMS array. These
65 *> settings determine how refinement is performed, but often the
66 *> defaults are acceptable. If the defaults are acceptable, users
67 *> can pass NPARAMS = 0 which prevents the source code from accessing
68 *> the PARAMS argument.
73 *> UPLO is CHARACTER*1
74 *> = 'U': Upper triangle of A is stored;
75 *> = 'L': Lower triangle of A is stored.
80 *> EQUED is CHARACTER*1
81 *> Specifies the form of equilibration that was done to A
82 *> before calling this routine. This is needed to compute
83 *> the solution and error bounds correctly.
84 *> = 'N': No equilibration
85 *> = 'Y': Both row and column equilibration, i.e., A has been
86 *> replaced by diag(S) * A * diag(S).
87 *> The right hand side B has been changed accordingly.
93 *> The order of the matrix A. N >= 0.
99 *> The number of right hand sides, i.e., the number of columns
100 *> of the matrices B and X. NRHS >= 0.
105 *> A is COMPLEX*16 array, dimension (LDA,N)
106 *> The symmetric matrix A. If UPLO = 'U', the leading N-by-N
107 *> upper triangular part of A contains the upper triangular part
108 *> of the matrix A, and the strictly lower triangular part of A
109 *> is not referenced. If UPLO = 'L', the leading N-by-N lower
110 *> triangular part of A contains the lower triangular part of
111 *> the matrix A, and the strictly upper triangular part of A is
118 *> The leading dimension of the array A. LDA >= max(1,N).
123 *> AF is COMPLEX*16 array, dimension (LDAF,N)
124 *> The triangular factor U or L from the Cholesky factorization
125 *> A = U**T*U or A = L*L**T, as computed by DPOTRF.
131 *> The leading dimension of the array AF. LDAF >= max(1,N).
136 *> S is DOUBLE PRECISION array, dimension (N)
137 *> The row scale factors for A. If EQUED = 'Y', A is multiplied on
138 *> the left and right by diag(S). S is an input argument if FACT =
139 *> 'F'; otherwise, S is an output argument. If FACT = 'F' and EQUED
140 *> = 'Y', each element of S must be positive. If S is output, each
141 *> element of S is a power of the radix. If S is input, each element
142 *> of S should be a power of the radix to ensure a reliable solution
143 *> and error estimates. Scaling by powers of the radix does not cause
144 *> rounding errors unless the result underflows or overflows.
145 *> Rounding errors during scaling lead to refining with a matrix that
146 *> is not equivalent to the input matrix, producing error estimates
147 *> that may not be reliable.
152 *> B is COMPLEX*16 array, dimension (LDB,NRHS)
153 *> The right hand side matrix B.
159 *> The leading dimension of the array B. LDB >= max(1,N).
164 *> X is COMPLEX*16 array, dimension (LDX,NRHS)
165 *> On entry, the solution matrix X, as computed by DGETRS.
166 *> On exit, the improved solution matrix X.
172 *> The leading dimension of the array X. LDX >= max(1,N).
177 *> RCOND is DOUBLE PRECISION
178 *> Reciprocal scaled condition number. This is an estimate of the
179 *> reciprocal Skeel condition number of the matrix A after
180 *> equilibration (if done). If this is less than the machine
181 *> precision (in particular, if it is zero), the matrix is singular
182 *> to working precision. Note that the error may still be small even
183 *> if this number is very small and the matrix appears ill-
189 *> BERR is DOUBLE PRECISION array, dimension (NRHS)
190 *> Componentwise relative backward error. This is the
191 *> componentwise relative backward error of each solution vector X(j)
192 *> (i.e., the smallest relative change in any element of A or B that
193 *> makes X(j) an exact solution).
196 *> \param[in] N_ERR_BNDS
198 *> N_ERR_BNDS is INTEGER
199 *> Number of error bounds to return for each right hand side
200 *> and each type (normwise or componentwise). See ERR_BNDS_NORM and
201 *> ERR_BNDS_COMP below.
204 *> \param[out] ERR_BNDS_NORM
206 *> ERR_BNDS_NORM is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS)
207 *> For each right-hand side, this array contains information about
208 *> various error bounds and condition numbers corresponding to the
209 *> normwise relative error, which is defined as follows:
211 *> Normwise relative error in the ith solution vector:
212 *> max_j (abs(XTRUE(j,i) - X(j,i)))
213 *> ------------------------------
216 *> The array is indexed by the type of error information as described
217 *> below. There currently are up to three pieces of information
220 *> The first index in ERR_BNDS_NORM(i,:) corresponds to the ith
223 *> The second index in ERR_BNDS_NORM(:,err) contains the following
225 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
226 *> reciprocal condition number is less than the threshold
227 *> sqrt(n) * dlamch('Epsilon').
229 *> err = 2 "Guaranteed" error bound: The estimated forward error,
230 *> almost certainly within a factor of 10 of the true error
231 *> so long as the next entry is greater than the threshold
232 *> sqrt(n) * dlamch('Epsilon'). This error bound should only
233 *> be trusted if the previous boolean is true.
235 *> err = 3 Reciprocal condition number: Estimated normwise
236 *> reciprocal condition number. Compared with the threshold
237 *> sqrt(n) * dlamch('Epsilon') to determine if the error
238 *> estimate is "guaranteed". These reciprocal condition
239 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
240 *> appropriately scaled matrix Z.
241 *> Let Z = S*A, where S scales each row by a power of the
242 *> radix so all absolute row sums of Z are approximately 1.
244 *> See Lapack Working Note 165 for further details and extra
248 *> \param[out] ERR_BNDS_COMP
250 *> ERR_BNDS_COMP is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS)
251 *> For each right-hand side, this array contains information about
252 *> various error bounds and condition numbers corresponding to the
253 *> componentwise relative error, which is defined as follows:
255 *> Componentwise relative error in the ith solution vector:
256 *> abs(XTRUE(j,i) - X(j,i))
257 *> max_j ----------------------
260 *> The array is indexed by the right-hand side i (on which the
261 *> componentwise relative error depends), and the type of error
262 *> information as described below. There currently are up to three
263 *> pieces of information returned for each right-hand side. If
264 *> componentwise accuracy is not requested (PARAMS(3) = 0.0), then
265 *> ERR_BNDS_COMP is not accessed. If N_ERR_BNDS .LT. 3, then at most
266 *> the first (:,N_ERR_BNDS) entries are returned.
268 *> The first index in ERR_BNDS_COMP(i,:) corresponds to the ith
271 *> The second index in ERR_BNDS_COMP(:,err) contains the following
273 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
274 *> reciprocal condition number is less than the threshold
275 *> sqrt(n) * dlamch('Epsilon').
277 *> err = 2 "Guaranteed" error bound: The estimated forward error,
278 *> almost certainly within a factor of 10 of the true error
279 *> so long as the next entry is greater than the threshold
280 *> sqrt(n) * dlamch('Epsilon'). This error bound should only
281 *> be trusted if the previous boolean is true.
283 *> err = 3 Reciprocal condition number: Estimated componentwise
284 *> reciprocal condition number. Compared with the threshold
285 *> sqrt(n) * dlamch('Epsilon') to determine if the error
286 *> estimate is "guaranteed". These reciprocal condition
287 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
288 *> appropriately scaled matrix Z.
289 *> Let Z = S*(A*diag(x)), where x is the solution for the
290 *> current right-hand side and S scales each row of
291 *> A*diag(x) by a power of the radix so all absolute row
292 *> sums of Z are approximately 1.
294 *> See Lapack Working Note 165 for further details and extra
298 *> \param[in] NPARAMS
300 *> NPARAMS is INTEGER
301 *> Specifies the number of parameters set in PARAMS. If .LE. 0, the
302 *> PARAMS array is never referenced and default values are used.
305 *> \param[in,out] PARAMS
307 *> PARAMS is DOUBLE PRECISION array, dimension NPARAMS
308 *> Specifies algorithm parameters. If an entry is .LT. 0.0, then
309 *> that entry will be filled with default value used for that
310 *> parameter. Only positions up to NPARAMS are accessed; defaults
311 *> are used for higher-numbered parameters.
313 *> PARAMS(LA_LINRX_ITREF_I = 1) : Whether to perform iterative
314 *> refinement or not.
316 *> = 0.0 : No refinement is performed, and no error bounds are
318 *> = 1.0 : Use the double-precision refinement algorithm,
319 *> possibly with doubled-single computations if the
320 *> compilation environment does not support DOUBLE
322 *> (other values are reserved for future use)
324 *> PARAMS(LA_LINRX_ITHRESH_I = 2) : Maximum number of residual
325 *> computations allowed for refinement.
327 *> Aggressive: Set to 100 to permit convergence using approximate
328 *> factorizations or factorizations other than LU. If
329 *> the factorization uses a technique other than
330 *> Gaussian elimination, the guarantees in
331 *> err_bnds_norm and err_bnds_comp may no longer be
334 *> PARAMS(LA_LINRX_CWISE_I = 3) : Flag determining if the code
335 *> will attempt to find a solution with small componentwise
336 *> relative error in the double-precision algorithm. Positive
337 *> is true, 0.0 is false.
338 *> Default: 1.0 (attempt componentwise convergence)
343 *> WORK is COMPLEX*16 array, dimension (2*N)
348 *> RWORK is DOUBLE PRECISION array, dimension (2*N)
354 *> = 0: Successful exit. The solution to every right-hand side is
356 *> < 0: If INFO = -i, the i-th argument had an illegal value
357 *> > 0 and <= N: U(INFO,INFO) is exactly zero. The factorization
358 *> has been completed, but the factor U is exactly singular, so
359 *> the solution and error bounds could not be computed. RCOND = 0
361 *> = N+J: The solution corresponding to the Jth right-hand side is
362 *> not guaranteed. The solutions corresponding to other right-
363 *> hand sides K with K > J may not be guaranteed as well, but
364 *> only the first such right-hand side is reported. If a small
365 *> componentwise error is not requested (PARAMS(3) = 0.0) then
366 *> the Jth right-hand side is the first with a normwise error
367 *> bound that is not guaranteed (the smallest J such
368 *> that ERR_BNDS_NORM(J,1) = 0.0). By default (PARAMS(3) = 1.0)
369 *> the Jth right-hand side is the first with either a normwise or
370 *> componentwise error bound that is not guaranteed (the smallest
371 *> J such that either ERR_BNDS_NORM(J,1) = 0.0 or
372 *> ERR_BNDS_COMP(J,1) = 0.0). See the definition of
373 *> ERR_BNDS_NORM(:,1) and ERR_BNDS_COMP(:,1). To get information
374 *> about all of the right-hand sides check ERR_BNDS_NORM or
381 *> \author Univ. of Tennessee
382 *> \author Univ. of California Berkeley
383 *> \author Univ. of Colorado Denver
388 *> \ingroup complex16POcomputational
390 * =====================================================================
391 SUBROUTINE ZPORFSX( UPLO, EQUED, N, NRHS, A, LDA, AF, LDAF, S, B,
392 $ LDB, X, LDX, RCOND, BERR, N_ERR_BNDS,
393 $ ERR_BNDS_NORM, ERR_BNDS_COMP, NPARAMS, PARAMS,
394 $ WORK, RWORK, INFO )
396 * -- LAPACK computational routine (version 3.4.1) --
397 * -- LAPACK is a software package provided by Univ. of Tennessee, --
398 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
401 * .. Scalar Arguments ..
402 CHARACTER UPLO, EQUED
403 INTEGER INFO, LDA, LDAF, LDB, LDX, N, NRHS, NPARAMS,
405 DOUBLE PRECISION RCOND
407 * .. Array Arguments ..
408 COMPLEX*16 A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
409 $ X( LDX, * ), WORK( * )
410 DOUBLE PRECISION RWORK( * ), S( * ), PARAMS(*), BERR( * ),
411 $ ERR_BNDS_NORM( NRHS, * ),
412 $ ERR_BNDS_COMP( NRHS, * )
415 * ==================================================================
418 DOUBLE PRECISION ZERO, ONE
419 PARAMETER ( ZERO = 0.0D+0, ONE = 1.0D+0 )
420 DOUBLE PRECISION ITREF_DEFAULT, ITHRESH_DEFAULT
421 DOUBLE PRECISION COMPONENTWISE_DEFAULT, RTHRESH_DEFAULT
422 DOUBLE PRECISION DZTHRESH_DEFAULT
423 PARAMETER ( ITREF_DEFAULT = 1.0D+0 )
424 PARAMETER ( ITHRESH_DEFAULT = 10.0D+0 )
425 PARAMETER ( COMPONENTWISE_DEFAULT = 1.0D+0 )
426 PARAMETER ( RTHRESH_DEFAULT = 0.5D+0 )
427 PARAMETER ( DZTHRESH_DEFAULT = 0.25D+0 )
428 INTEGER LA_LINRX_ITREF_I, LA_LINRX_ITHRESH_I,
430 PARAMETER ( LA_LINRX_ITREF_I = 1,
431 $ LA_LINRX_ITHRESH_I = 2 )
432 PARAMETER ( LA_LINRX_CWISE_I = 3 )
433 INTEGER LA_LINRX_TRUST_I, LA_LINRX_ERR_I,
435 PARAMETER ( LA_LINRX_TRUST_I = 1, LA_LINRX_ERR_I = 2 )
436 PARAMETER ( LA_LINRX_RCOND_I = 3 )
438 * .. Local Scalars ..
441 INTEGER J, PREC_TYPE, REF_TYPE
443 DOUBLE PRECISION ANORM, RCOND_TMP
444 DOUBLE PRECISION ILLRCOND_THRESH, ERR_LBND, CWISE_WRONG
447 DOUBLE PRECISION RTHRESH, UNSTABLE_THRESH
449 * .. External Subroutines ..
450 EXTERNAL XERBLA, ZPOCON, ZLA_PORFSX_EXTENDED
452 * .. Intrinsic Functions ..
453 INTRINSIC MAX, SQRT, TRANSFER
455 * .. External Functions ..
456 EXTERNAL LSAME, BLAS_FPINFO_X, ILATRANS, ILAPREC
457 EXTERNAL DLAMCH, ZLANHE, ZLA_PORCOND_X, ZLA_PORCOND_C
458 DOUBLE PRECISION DLAMCH, ZLANHE, ZLA_PORCOND_X, ZLA_PORCOND_C
460 INTEGER BLAS_FPINFO_X
461 INTEGER ILATRANS, ILAPREC
463 * .. Executable Statements ..
465 * Check the input parameters.
468 REF_TYPE = INT( ITREF_DEFAULT )
469 IF ( NPARAMS .GE. LA_LINRX_ITREF_I ) THEN
470 IF ( PARAMS( LA_LINRX_ITREF_I ) .LT. 0.0D+0 ) THEN
471 PARAMS( LA_LINRX_ITREF_I ) = ITREF_DEFAULT
473 REF_TYPE = PARAMS( LA_LINRX_ITREF_I )
477 * Set default parameters.
479 ILLRCOND_THRESH = DBLE( N ) * DLAMCH( 'Epsilon' )
480 ITHRESH = INT( ITHRESH_DEFAULT )
481 RTHRESH = RTHRESH_DEFAULT
482 UNSTABLE_THRESH = DZTHRESH_DEFAULT
483 IGNORE_CWISE = COMPONENTWISE_DEFAULT .EQ. 0.0D+0
485 IF ( NPARAMS.GE.LA_LINRX_ITHRESH_I ) THEN
486 IF ( PARAMS(LA_LINRX_ITHRESH_I ).LT.0.0D+0 ) THEN
487 PARAMS( LA_LINRX_ITHRESH_I ) = ITHRESH
489 ITHRESH = INT( PARAMS( LA_LINRX_ITHRESH_I ) )
492 IF ( NPARAMS.GE.LA_LINRX_CWISE_I ) THEN
493 IF ( PARAMS(LA_LINRX_CWISE_I ).LT.0.0D+0 ) THEN
494 IF ( IGNORE_CWISE ) THEN
495 PARAMS( LA_LINRX_CWISE_I ) = 0.0D+0
497 PARAMS( LA_LINRX_CWISE_I ) = 1.0D+0
500 IGNORE_CWISE = PARAMS( LA_LINRX_CWISE_I ) .EQ. 0.0D+0
503 IF ( REF_TYPE .EQ. 0 .OR. N_ERR_BNDS .EQ. 0 ) THEN
505 ELSE IF ( IGNORE_CWISE ) THEN
511 RCEQU = LSAME( EQUED, 'Y' )
513 * Test input parameters.
515 IF (.NOT.LSAME( UPLO, 'U' ) .AND. .NOT.LSAME( UPLO, 'L' ) ) THEN
517 ELSE IF( .NOT.RCEQU .AND. .NOT.LSAME( EQUED, 'N' ) ) THEN
519 ELSE IF( N.LT.0 ) THEN
521 ELSE IF( NRHS.LT.0 ) THEN
523 ELSE IF( LDA.LT.MAX( 1, N ) ) THEN
525 ELSE IF( LDAF.LT.MAX( 1, N ) ) THEN
527 ELSE IF( LDB.LT.MAX( 1, N ) ) THEN
529 ELSE IF( LDX.LT.MAX( 1, N ) ) THEN
533 CALL XERBLA( 'ZPORFSX', -INFO )
537 * Quick return if possible.
539 IF( N.EQ.0 .OR. NRHS.EQ.0 ) THEN
543 IF ( N_ERR_BNDS .GE. 1 ) THEN
544 ERR_BNDS_NORM( J, LA_LINRX_TRUST_I ) = 1.0D+0
545 ERR_BNDS_COMP( J, LA_LINRX_TRUST_I ) = 1.0D+0
547 IF ( N_ERR_BNDS .GE. 2 ) THEN
548 ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) = 0.0D+0
549 ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) = 0.0D+0
551 IF ( N_ERR_BNDS .GE. 3 ) THEN
552 ERR_BNDS_NORM( J, LA_LINRX_RCOND_I ) = 1.0D+0
553 ERR_BNDS_COMP( J, LA_LINRX_RCOND_I ) = 1.0D+0
559 * Default to failure.
564 IF ( N_ERR_BNDS .GE. 1 ) THEN
565 ERR_BNDS_NORM( J, LA_LINRX_TRUST_I ) = 1.0D+0
566 ERR_BNDS_COMP( J, LA_LINRX_TRUST_I ) = 1.0D+0
568 IF ( N_ERR_BNDS .GE. 2 ) THEN
569 ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) = 1.0D+0
570 ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) = 1.0D+0
572 IF ( N_ERR_BNDS .GE. 3 ) THEN
573 ERR_BNDS_NORM( J, LA_LINRX_RCOND_I ) = 0.0D+0
574 ERR_BNDS_COMP( J, LA_LINRX_RCOND_I ) = 0.0D+0
578 * Compute the norm of A and the reciprocal of the condition
582 ANORM = ZLANHE( NORM, UPLO, N, A, LDA, RWORK )
583 CALL ZPOCON( UPLO, N, AF, LDAF, ANORM, RCOND, WORK, RWORK,
586 * Perform refinement on each right-hand side
588 IF ( REF_TYPE .NE. 0 ) THEN
590 PREC_TYPE = ILAPREC( 'E' )
592 CALL ZLA_PORFSX_EXTENDED( PREC_TYPE, UPLO, N,
593 $ NRHS, A, LDA, AF, LDAF, RCEQU, S, B,
594 $ LDB, X, LDX, BERR, N_NORMS, ERR_BNDS_NORM, ERR_BNDS_COMP,
595 $ WORK, RWORK, WORK(N+1),
596 $ TRANSFER (RWORK(1:2*N), (/ (ZERO, ZERO) /), N), RCOND,
597 $ ITHRESH, RTHRESH, UNSTABLE_THRESH, IGNORE_CWISE,
601 ERR_LBND = MAX( 10.0D+0, SQRT( DBLE( N ) ) ) * DLAMCH( 'Epsilon' )
602 IF ( N_ERR_BNDS .GE. 1 .AND. N_NORMS .GE. 1 ) THEN
604 * Compute scaled normwise condition number cond(A*C).
607 RCOND_TMP = ZLA_PORCOND_C( UPLO, N, A, LDA, AF, LDAF,
608 $ S, .TRUE., INFO, WORK, RWORK )
610 RCOND_TMP = ZLA_PORCOND_C( UPLO, N, A, LDA, AF, LDAF,
611 $ S, .FALSE., INFO, WORK, RWORK )
615 * Cap the error at 1.0.
617 IF ( N_ERR_BNDS .GE. LA_LINRX_ERR_I
618 $ .AND. ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) .GT. 1.0D+0 )
619 $ ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) = 1.0D+0
621 * Threshold the error (see LAWN).
623 IF ( RCOND_TMP .LT. ILLRCOND_THRESH ) THEN
624 ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) = 1.0D+0
625 ERR_BNDS_NORM( J, LA_LINRX_TRUST_I ) = 0.0D+0
626 IF ( INFO .LE. N ) INFO = N + J
627 ELSE IF ( ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) .LT. ERR_LBND )
629 ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) = ERR_LBND
630 ERR_BNDS_NORM( J, LA_LINRX_TRUST_I ) = 1.0D+0
633 * Save the condition number.
635 IF ( N_ERR_BNDS .GE. LA_LINRX_RCOND_I ) THEN
636 ERR_BNDS_NORM( J, LA_LINRX_RCOND_I ) = RCOND_TMP
642 IF (N_ERR_BNDS .GE. 1 .AND. N_NORMS .GE. 2) THEN
644 * Compute componentwise condition number cond(A*diag(Y(:,J))) for
645 * each right-hand side using the current solution as an estimate of
646 * the true solution. If the componentwise error estimate is too
647 * large, then the solution is a lousy estimate of truth and the
648 * estimated RCOND may be too optimistic. To avoid misleading users,
649 * the inverse condition number is set to 0.0 when the estimated
650 * cwise error is at least CWISE_WRONG.
652 CWISE_WRONG = SQRT( DLAMCH( 'Epsilon' ) )
654 IF (ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) .LT. CWISE_WRONG )
656 RCOND_TMP = ZLA_PORCOND_X( UPLO, N, A, LDA, AF, LDAF,
657 $ X(1,J), INFO, WORK, RWORK )
662 * Cap the error at 1.0.
664 IF ( N_ERR_BNDS .GE. LA_LINRX_ERR_I
665 $ .AND. ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) .GT. 1.0D+0 )
666 $ ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) = 1.0D+0
668 * Threshold the error (see LAWN).
670 IF (RCOND_TMP .LT. ILLRCOND_THRESH) THEN
671 ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) = 1.0D+0
672 ERR_BNDS_COMP( J, LA_LINRX_TRUST_I ) = 0.0D+0
673 IF ( PARAMS( LA_LINRX_CWISE_I ) .EQ. 1.0D+0
674 $ .AND. INFO.LT.N + J ) INFO = N + J
675 ELSE IF ( ERR_BNDS_COMP( J, LA_LINRX_ERR_I )
676 $ .LT. ERR_LBND ) THEN
677 ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) = ERR_LBND
678 ERR_BNDS_COMP( J, LA_LINRX_TRUST_I ) = 1.0D+0
681 * Save the condition number.
683 IF ( N_ERR_BNDS .GE. LA_LINRX_RCOND_I ) THEN
684 ERR_BNDS_COMP( J, LA_LINRX_RCOND_I ) = RCOND_TMP