1 *> \brief \b CLA_SYRFSX_EXTENDED improves the computed solution to a system of linear equations for symmetric indefinite matrices by performing extra-precise iterative refinement and provides error bounds and backward error estimates for the solution.
3 * =========== DOCUMENTATION ===========
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
9 *> Download CLA_SYRFSX_EXTENDED + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/cla_syrfsx_extended.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/cla_syrfsx_extended.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/cla_syrfsx_extended.f">
21 * SUBROUTINE CLA_SYRFSX_EXTENDED( PREC_TYPE, UPLO, N, NRHS, A, LDA,
22 * AF, LDAF, IPIV, COLEQU, C, B, LDB,
23 * Y, LDY, BERR_OUT, N_NORMS,
24 * ERR_BNDS_NORM, ERR_BNDS_COMP, RES,
25 * AYB, DY, Y_TAIL, RCOND, ITHRESH,
26 * RTHRESH, DZ_UB, IGNORE_CWISE,
29 * .. Scalar Arguments ..
30 * INTEGER INFO, LDA, LDAF, LDB, LDY, N, NRHS, PREC_TYPE,
33 * LOGICAL COLEQU, IGNORE_CWISE
36 * .. Array Arguments ..
38 * COMPLEX A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
39 * $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * )
40 * REAL C( * ), AYB( * ), RCOND, BERR_OUT( * ),
41 * $ ERR_BNDS_NORM( NRHS, * ),
42 * $ ERR_BNDS_COMP( NRHS, * )
51 *> CLA_SYRFSX_EXTENDED improves the computed solution to a system of
52 *> linear equations by performing extra-precise iterative refinement
53 *> and provides error bounds and backward error estimates for the solution.
54 *> This subroutine is called by CSYRFSX to perform iterative refinement.
55 *> In addition to normwise error bound, the code provides maximum
56 *> componentwise error bound if possible. See comments for ERR_BNDS_NORM
57 *> and ERR_BNDS_COMP for details of the error bounds. Note that this
58 *> subroutine is only resonsible for setting the second fields of
59 *> ERR_BNDS_NORM and ERR_BNDS_COMP.
65 *> \param[in] PREC_TYPE
67 *> PREC_TYPE is INTEGER
68 *> Specifies the intermediate precision to be used in refinement.
69 *> The value is defined by ILAPREC(P) where P is a CHARACTER and
78 *> UPLO is CHARACTER*1
79 *> = 'U': Upper triangle of A is stored;
80 *> = 'L': Lower triangle of A is stored.
86 *> The number of linear equations, i.e., the order of the
93 *> The number of right-hand-sides, i.e., the number of columns of the
99 *> A is COMPLEX array, dimension (LDA,N)
100 *> On entry, the N-by-N matrix A.
106 *> The leading dimension of the array A. LDA >= max(1,N).
111 *> AF is COMPLEX array, dimension (LDAF,N)
112 *> The block diagonal matrix D and the multipliers used to
113 *> obtain the factor U or L as computed by CSYTRF.
119 *> The leading dimension of the array AF. LDAF >= max(1,N).
124 *> IPIV is INTEGER array, dimension (N)
125 *> Details of the interchanges and the block structure of D
126 *> as determined by CSYTRF.
132 *> If .TRUE. then column equilibration was done to A before calling
133 *> this routine. This is needed to compute the solution and error
139 *> C is REAL array, dimension (N)
140 *> The column scale factors for A. If COLEQU = .FALSE., C
141 *> is not accessed. If C is input, each element of C should be a power
142 *> of the radix to ensure a reliable solution and error estimates.
143 *> Scaling by powers of the radix does not cause rounding errors unless
144 *> the result underflows or overflows. Rounding errors during scaling
145 *> lead to refining with a matrix that is not equivalent to the
146 *> input matrix, producing error estimates that may not be
152 *> B is COMPLEX array, dimension (LDB,NRHS)
153 *> The right-hand-side matrix B.
159 *> The leading dimension of the array B. LDB >= max(1,N).
164 *> Y is COMPLEX array, dimension
166 *> On entry, the solution matrix X, as computed by CSYTRS.
167 *> On exit, the improved solution matrix Y.
173 *> The leading dimension of the array Y. LDY >= max(1,N).
176 *> \param[out] BERR_OUT
178 *> BERR_OUT is REAL array, dimension (NRHS)
179 *> On exit, BERR_OUT(j) contains the componentwise relative backward
180 *> error for right-hand-side j from the formula
181 *> max(i) ( abs(RES(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
182 *> where abs(Z) is the componentwise absolute value of the matrix
183 *> or vector Z. This is computed by CLA_LIN_BERR.
186 *> \param[in] N_NORMS
188 *> N_NORMS is INTEGER
189 *> Determines which error bounds to return (see ERR_BNDS_NORM
190 *> and ERR_BNDS_COMP).
191 *> If N_NORMS >= 1 return normwise error bounds.
192 *> If N_NORMS >= 2 return componentwise error bounds.
195 *> \param[in,out] ERR_BNDS_NORM
197 *> ERR_BNDS_NORM is REAL array, dimension
198 *> (NRHS, N_ERR_BNDS)
199 *> For each right-hand side, this array contains information about
200 *> various error bounds and condition numbers corresponding to the
201 *> normwise relative error, which is defined as follows:
203 *> Normwise relative error in the ith solution vector:
204 *> max_j (abs(XTRUE(j,i) - X(j,i)))
205 *> ------------------------------
208 *> The array is indexed by the type of error information as described
209 *> below. There currently are up to three pieces of information
212 *> The first index in ERR_BNDS_NORM(i,:) corresponds to the ith
215 *> The second index in ERR_BNDS_NORM(:,err) contains the following
217 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
218 *> reciprocal condition number is less than the threshold
219 *> sqrt(n) * slamch('Epsilon').
221 *> err = 2 "Guaranteed" error bound: The estimated forward error,
222 *> almost certainly within a factor of 10 of the true error
223 *> so long as the next entry is greater than the threshold
224 *> sqrt(n) * slamch('Epsilon'). This error bound should only
225 *> be trusted if the previous boolean is true.
227 *> err = 3 Reciprocal condition number: Estimated normwise
228 *> reciprocal condition number. Compared with the threshold
229 *> sqrt(n) * slamch('Epsilon') to determine if the error
230 *> estimate is "guaranteed". These reciprocal condition
231 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
232 *> appropriately scaled matrix Z.
233 *> Let Z = S*A, where S scales each row by a power of the
234 *> radix so all absolute row sums of Z are approximately 1.
236 *> This subroutine is only responsible for setting the second field
238 *> See Lapack Working Note 165 for further details and extra
242 *> \param[in,out] ERR_BNDS_COMP
244 *> ERR_BNDS_COMP is REAL array, dimension
245 *> (NRHS, N_ERR_BNDS)
246 *> For each right-hand side, this array contains information about
247 *> various error bounds and condition numbers corresponding to the
248 *> componentwise relative error, which is defined as follows:
250 *> Componentwise relative error in the ith solution vector:
251 *> abs(XTRUE(j,i) - X(j,i))
252 *> max_j ----------------------
255 *> The array is indexed by the right-hand side i (on which the
256 *> componentwise relative error depends), and the type of error
257 *> information as described below. There currently are up to three
258 *> pieces of information returned for each right-hand side. If
259 *> componentwise accuracy is not requested (PARAMS(3) = 0.0), then
260 *> ERR_BNDS_COMP is not accessed. If N_ERR_BNDS .LT. 3, then at most
261 *> the first (:,N_ERR_BNDS) entries are returned.
263 *> The first index in ERR_BNDS_COMP(i,:) corresponds to the ith
266 *> The second index in ERR_BNDS_COMP(:,err) contains the following
268 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
269 *> reciprocal condition number is less than the threshold
270 *> sqrt(n) * slamch('Epsilon').
272 *> err = 2 "Guaranteed" error bound: The estimated forward error,
273 *> almost certainly within a factor of 10 of the true error
274 *> so long as the next entry is greater than the threshold
275 *> sqrt(n) * slamch('Epsilon'). This error bound should only
276 *> be trusted if the previous boolean is true.
278 *> err = 3 Reciprocal condition number: Estimated componentwise
279 *> reciprocal condition number. Compared with the threshold
280 *> sqrt(n) * slamch('Epsilon') to determine if the error
281 *> estimate is "guaranteed". These reciprocal condition
282 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
283 *> appropriately scaled matrix Z.
284 *> Let Z = S*(A*diag(x)), where x is the solution for the
285 *> current right-hand side and S scales each row of
286 *> A*diag(x) by a power of the radix so all absolute row
287 *> sums of Z are approximately 1.
289 *> This subroutine is only responsible for setting the second field
291 *> See Lapack Working Note 165 for further details and extra
297 *> RES is COMPLEX array, dimension (N)
298 *> Workspace to hold the intermediate residual.
303 *> AYB is REAL array, dimension (N)
309 *> DY is COMPLEX array, dimension (N)
310 *> Workspace to hold the intermediate solution.
315 *> Y_TAIL is COMPLEX array, dimension (N)
316 *> Workspace to hold the trailing bits of the intermediate solution.
322 *> Reciprocal scaled condition number. This is an estimate of the
323 *> reciprocal Skeel condition number of the matrix A after
324 *> equilibration (if done). If this is less than the machine
325 *> precision (in particular, if it is zero), the matrix is singular
326 *> to working precision. Note that the error may still be small even
327 *> if this number is very small and the matrix appears ill-
331 *> \param[in] ITHRESH
333 *> ITHRESH is INTEGER
334 *> The maximum number of residual computations allowed for
335 *> refinement. The default is 10. For 'aggressive' set to 100 to
336 *> permit convergence using approximate factorizations or
337 *> factorizations other than LU. If the factorization uses a
338 *> technique other than Gaussian elimination, the guarantees in
339 *> ERR_BNDS_NORM and ERR_BNDS_COMP may no longer be trustworthy.
342 *> \param[in] RTHRESH
345 *> Determines when to stop refinement if the error estimate stops
346 *> decreasing. Refinement will stop when the next solution no longer
347 *> satisfies norm(dx_{i+1}) < RTHRESH * norm(dx_i) where norm(Z) is
348 *> the infinity norm of Z. RTHRESH satisfies 0 < RTHRESH <= 1. The
349 *> default value is 0.5. For 'aggressive' set to 0.9 to permit
350 *> convergence on extremely ill-conditioned matrices. See LAWN 165
357 *> Determines when to start considering componentwise convergence.
358 *> Componentwise convergence is only considered after each component
359 *> of the solution Y is stable, which we definte as the relative
360 *> change in each component being less than DZ_UB. The default value
361 *> is 0.25, requiring the first bit to be stable. See LAWN 165 for
365 *> \param[in] IGNORE_CWISE
367 *> IGNORE_CWISE is LOGICAL
368 *> If .TRUE. then ignore componentwise convergence. Default value
375 *> = 0: Successful exit.
376 *> < 0: if INFO = -i, the ith argument to CLA_SYRFSX_EXTENDED had an illegal
383 *> \author Univ. of Tennessee
384 *> \author Univ. of California Berkeley
385 *> \author Univ. of Colorado Denver
388 *> \date September 2012
390 *> \ingroup complexSYcomputational
392 * =====================================================================
393 SUBROUTINE CLA_SYRFSX_EXTENDED( PREC_TYPE, UPLO, N, NRHS, A, LDA,
394 $ AF, LDAF, IPIV, COLEQU, C, B, LDB,
395 $ Y, LDY, BERR_OUT, N_NORMS,
396 $ ERR_BNDS_NORM, ERR_BNDS_COMP, RES,
397 $ AYB, DY, Y_TAIL, RCOND, ITHRESH,
398 $ RTHRESH, DZ_UB, IGNORE_CWISE,
401 * -- LAPACK computational routine (version 3.4.2) --
402 * -- LAPACK is a software package provided by Univ. of Tennessee, --
403 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
406 * .. Scalar Arguments ..
407 INTEGER INFO, LDA, LDAF, LDB, LDY, N, NRHS, PREC_TYPE,
410 LOGICAL COLEQU, IGNORE_CWISE
413 * .. Array Arguments ..
415 COMPLEX A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
416 $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * )
417 REAL C( * ), AYB( * ), RCOND, BERR_OUT( * ),
418 $ ERR_BNDS_NORM( NRHS, * ),
419 $ ERR_BNDS_COMP( NRHS, * )
422 * =====================================================================
424 * .. Local Scalars ..
425 INTEGER UPLO2, CNT, I, J, X_STATE, Z_STATE,
427 REAL YK, DYK, YMIN, NORMY, NORMX, NORMDX, DXRAT,
428 $ DZRAT, PREVNORMDX, PREV_DZ_Z, DXRATMAX,
429 $ DZRATMAX, DX_X, DZ_Z, FINAL_DX_X, FINAL_DZ_Z,
430 $ EPS, HUGEVAL, INCR_THRESH
431 LOGICAL INCR_PREC, UPPER
435 INTEGER UNSTABLE_STATE, WORKING_STATE, CONV_STATE,
436 $ NOPROG_STATE, BASE_RESIDUAL, EXTRA_RESIDUAL,
438 PARAMETER ( UNSTABLE_STATE = 0, WORKING_STATE = 1,
439 $ CONV_STATE = 2, NOPROG_STATE = 3 )
440 PARAMETER ( BASE_RESIDUAL = 0, EXTRA_RESIDUAL = 1,
442 INTEGER FINAL_NRM_ERR_I, FINAL_CMP_ERR_I, BERR_I
443 INTEGER RCOND_I, NRM_RCOND_I, NRM_ERR_I, CMP_RCOND_I
444 INTEGER CMP_ERR_I, PIV_GROWTH_I
445 PARAMETER ( FINAL_NRM_ERR_I = 1, FINAL_CMP_ERR_I = 2,
447 PARAMETER ( RCOND_I = 4, NRM_RCOND_I = 5, NRM_ERR_I = 6 )
448 PARAMETER ( CMP_RCOND_I = 7, CMP_ERR_I = 8,
450 INTEGER LA_LINRX_ITREF_I, LA_LINRX_ITHRESH_I,
452 PARAMETER ( LA_LINRX_ITREF_I = 1,
453 $ LA_LINRX_ITHRESH_I = 2 )
454 PARAMETER ( LA_LINRX_CWISE_I = 3 )
455 INTEGER LA_LINRX_TRUST_I, LA_LINRX_ERR_I,
457 PARAMETER ( LA_LINRX_TRUST_I = 1, LA_LINRX_ERR_I = 2 )
458 PARAMETER ( LA_LINRX_RCOND_I = 3 )
460 * .. External Functions ..
465 * .. External Subroutines ..
466 EXTERNAL CAXPY, CCOPY, CSYTRS, CSYMV, BLAS_CSYMV_X,
467 $ BLAS_CSYMV2_X, CLA_SYAMV, CLA_WWADDW,
471 * .. Intrinsic Functions ..
472 INTRINSIC ABS, REAL, AIMAG, MAX, MIN
474 * .. Statement Functions ..
477 * .. Statement Function Definitions ..
478 CABS1( ZDUM ) = ABS( REAL( ZDUM ) ) + ABS( AIMAG( ZDUM ) )
480 * .. Executable Statements ..
483 UPPER = LSAME( UPLO, 'U' )
484 IF( .NOT.UPPER .AND. .NOT.LSAME( UPLO, 'L' ) ) THEN
486 ELSE IF( N.LT.0 ) THEN
488 ELSE IF( NRHS.LT.0 ) THEN
490 ELSE IF( LDA.LT.MAX( 1, N ) ) THEN
492 ELSE IF( LDAF.LT.MAX( 1, N ) ) THEN
494 ELSE IF( LDB.LT.MAX( 1, N ) ) THEN
496 ELSE IF( LDY.LT.MAX( 1, N ) ) THEN
500 CALL XERBLA( 'CLA_SYRFSX_EXTENDED', -INFO )
503 EPS = SLAMCH( 'Epsilon' )
504 HUGEVAL = SLAMCH( 'Overflow' )
505 * Force HUGEVAL to Inf
506 HUGEVAL = HUGEVAL * HUGEVAL
507 * Using HUGEVAL may lead to spurious underflows.
508 INCR_THRESH = REAL( N ) * EPS
510 IF ( LSAME ( UPLO, 'L' ) ) THEN
511 UPLO2 = ILAUPLO( 'L' )
513 UPLO2 = ILAUPLO( 'U' )
517 Y_PREC_STATE = EXTRA_RESIDUAL
518 IF ( Y_PREC_STATE .EQ. EXTRA_Y ) THEN
535 X_STATE = WORKING_STATE
536 Z_STATE = UNSTABLE_STATE
541 * Compute residual RES = B_s - op(A_s) * Y,
542 * op(A) = A, A**T, or A**H depending on TRANS (and type).
544 CALL CCOPY( N, B( 1, J ), 1, RES, 1 )
545 IF ( Y_PREC_STATE .EQ. BASE_RESIDUAL ) THEN
546 CALL CSYMV( UPLO, N, CMPLX(-1.0), A, LDA, Y(1,J), 1,
547 $ CMPLX(1.0), RES, 1 )
548 ELSE IF ( Y_PREC_STATE .EQ. EXTRA_RESIDUAL ) THEN
549 CALL BLAS_CSYMV_X( UPLO2, N, CMPLX(-1.0), A, LDA,
550 $ Y( 1, J ), 1, CMPLX(1.0), RES, 1, PREC_TYPE )
552 CALL BLAS_CSYMV2_X(UPLO2, N, CMPLX(-1.0), A, LDA,
553 $ Y(1, J), Y_TAIL, 1, CMPLX(1.0), RES, 1, PREC_TYPE)
556 ! XXX: RES is no longer needed.
557 CALL CCOPY( N, RES, 1, DY, 1 )
558 CALL CSYTRS( UPLO, N, 1, AF, LDAF, IPIV, DY, N, INFO )
560 * Calculate relative changes DX_X, DZ_Z and ratios DXRAT, DZRAT.
569 YK = CABS1( Y( I, J ) )
570 DYK = CABS1( DY( I ) )
572 IF ( YK .NE. 0.0 ) THEN
573 DZ_Z = MAX( DZ_Z, DYK / YK )
574 ELSE IF ( DYK .NE. 0.0 ) THEN
578 YMIN = MIN( YMIN, YK )
580 NORMY = MAX( NORMY, YK )
583 NORMX = MAX( NORMX, YK * C( I ) )
584 NORMDX = MAX( NORMDX, DYK * C( I ) )
587 NORMDX = MAX( NORMDX, DYK )
591 IF ( NORMX .NE. 0.0 ) THEN
592 DX_X = NORMDX / NORMX
593 ELSE IF ( NORMDX .EQ. 0.0 ) THEN
599 DXRAT = NORMDX / PREVNORMDX
600 DZRAT = DZ_Z / PREV_DZ_Z
602 * Check termination criteria.
604 IF ( YMIN*RCOND .LT. INCR_THRESH*NORMY
605 $ .AND. Y_PREC_STATE .LT. EXTRA_Y )
608 IF ( X_STATE .EQ. NOPROG_STATE .AND. DXRAT .LE. RTHRESH )
609 $ X_STATE = WORKING_STATE
610 IF ( X_STATE .EQ. WORKING_STATE ) THEN
611 IF ( DX_X .LE. EPS ) THEN
613 ELSE IF ( DXRAT .GT. RTHRESH ) THEN
614 IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
617 X_STATE = NOPROG_STATE
620 IF (DXRAT .GT. DXRATMAX) DXRATMAX = DXRAT
622 IF ( X_STATE .GT. WORKING_STATE ) FINAL_DX_X = DX_X
625 IF ( Z_STATE .EQ. UNSTABLE_STATE .AND. DZ_Z .LE. DZ_UB )
626 $ Z_STATE = WORKING_STATE
627 IF ( Z_STATE .EQ. NOPROG_STATE .AND. DZRAT .LE. RTHRESH )
628 $ Z_STATE = WORKING_STATE
629 IF ( Z_STATE .EQ. WORKING_STATE ) THEN
630 IF ( DZ_Z .LE. EPS ) THEN
632 ELSE IF ( DZ_Z .GT. DZ_UB ) THEN
633 Z_STATE = UNSTABLE_STATE
636 ELSE IF ( DZRAT .GT. RTHRESH ) THEN
637 IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
640 Z_STATE = NOPROG_STATE
643 IF ( DZRAT .GT. DZRATMAX ) DZRATMAX = DZRAT
645 IF ( Z_STATE .GT. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
648 IF ( X_STATE.NE.WORKING_STATE.AND.
649 $ ( IGNORE_CWISE.OR.Z_STATE.NE.WORKING_STATE ) )
652 IF ( INCR_PREC ) THEN
654 Y_PREC_STATE = Y_PREC_STATE + 1
665 IF ( Y_PREC_STATE .LT. EXTRA_Y ) THEN
666 CALL CAXPY( N, CMPLX(1.0), DY, 1, Y(1,J), 1 )
668 CALL CLA_WWADDW( N, Y(1,J), Y_TAIL, DY )
672 * Target of "IF (Z_STOP .AND. X_STOP)". Sun's f77 won't EXIT.
675 * Set final_* when cnt hits ithresh.
677 IF ( X_STATE .EQ. WORKING_STATE ) FINAL_DX_X = DX_X
678 IF ( Z_STATE .EQ. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
680 * Compute error bounds.
682 IF ( N_NORMS .GE. 1 ) THEN
683 ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) =
684 $ FINAL_DX_X / (1 - DXRATMAX)
686 IF ( N_NORMS .GE. 2 ) THEN
687 ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) =
688 $ FINAL_DZ_Z / (1 - DZRATMAX)
691 * Compute componentwise relative backward error from formula
692 * max(i) ( abs(R(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
693 * where abs(Z) is the componentwise absolute value of the matrix
696 * Compute residual RES = B_s - op(A_s) * Y,
697 * op(A) = A, A**T, or A**H depending on TRANS (and type).
699 CALL CCOPY( N, B( 1, J ), 1, RES, 1 )
700 CALL CSYMV( UPLO, N, CMPLX(-1.0), A, LDA, Y(1,J), 1,
701 $ CMPLX(1.0), RES, 1 )
704 AYB( I ) = CABS1( B( I, J ) )
707 * Compute abs(op(A_s))*abs(Y) + abs(B_s).
709 CALL CLA_SYAMV ( UPLO2, N, 1.0,
710 $ A, LDA, Y(1, J), 1, 1.0, AYB, 1 )
712 CALL CLA_LIN_BERR ( N, N, 1, RES, AYB, BERR_OUT( J ) )
714 * End of loop for each RHS.