1 *> \brief \b SLA_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 SLA_SYRFSX_EXTENDED + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/sla_syrfsx_extended.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/sla_syrfsx_extended.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/sla_syrfsx_extended.f">
21 * SUBROUTINE SLA_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 * REAL 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, * )
52 *> SLA_SYRFSX_EXTENDED improves the computed solution to a system of
53 *> linear equations by performing extra-precise iterative refinement
54 *> and provides error bounds and backward error estimates for the solution.
55 *> This subroutine is called by SSYRFSX to perform iterative refinement.
56 *> In addition to normwise error bound, the code provides maximum
57 *> componentwise error bound if possible. See comments for ERR_BNDS_NORM
58 *> and ERR_BNDS_COMP for details of the error bounds. Note that this
59 *> subroutine is only resonsible for setting the second fields of
60 *> ERR_BNDS_NORM and ERR_BNDS_COMP.
66 *> \param[in] PREC_TYPE
68 *> PREC_TYPE is INTEGER
69 *> Specifies the intermediate precision to be used in refinement.
70 *> The value is defined by ILAPREC(P) where P is a CHARACTER and
79 *> UPLO is CHARACTER*1
80 *> = 'U': Upper triangle of A is stored;
81 *> = 'L': Lower triangle of A is stored.
87 *> The number of linear equations, i.e., the order of the
94 *> The number of right-hand-sides, i.e., the number of columns of the
100 *> A is REAL array, dimension (LDA,N)
101 *> On entry, the N-by-N matrix A.
107 *> The leading dimension of the array A. LDA >= max(1,N).
112 *> AF is REAL array, dimension (LDAF,N)
113 *> The block diagonal matrix D and the multipliers used to
114 *> obtain the factor U or L as computed by SSYTRF.
120 *> The leading dimension of the array AF. LDAF >= max(1,N).
125 *> IPIV is INTEGER array, dimension (N)
126 *> Details of the interchanges and the block structure of D
127 *> as determined by SSYTRF.
133 *> If .TRUE. then column equilibration was done to A before calling
134 *> this routine. This is needed to compute the solution and error
140 *> C is REAL array, dimension (N)
141 *> The column scale factors for A. If COLEQU = .FALSE., C
142 *> is not accessed. If C is input, each element of C should be a power
143 *> of the radix to ensure a reliable solution and error estimates.
144 *> Scaling by powers of the radix does not cause rounding errors unless
145 *> the result underflows or overflows. Rounding errors during scaling
146 *> lead to refining with a matrix that is not equivalent to the
147 *> input matrix, producing error estimates that may not be
153 *> B is REAL array, dimension (LDB,NRHS)
154 *> The right-hand-side matrix B.
160 *> The leading dimension of the array B. LDB >= max(1,N).
165 *> Y is REAL array, dimension (LDY,NRHS)
166 *> On entry, the solution matrix X, as computed by SSYTRS.
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 SLA_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 (NRHS, N_ERR_BNDS)
198 *> For each right-hand side, this array contains information about
199 *> various error bounds and condition numbers corresponding to the
200 *> normwise relative error, which is defined as follows:
202 *> Normwise relative error in the ith solution vector:
203 *> max_j (abs(XTRUE(j,i) - X(j,i)))
204 *> ------------------------------
207 *> The array is indexed by the type of error information as described
208 *> below. There currently are up to three pieces of information
211 *> The first index in ERR_BNDS_NORM(i,:) corresponds to the ith
214 *> The second index in ERR_BNDS_NORM(:,err) contains the following
216 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
217 *> reciprocal condition number is less than the threshold
218 *> sqrt(n) * slamch('Epsilon').
220 *> err = 2 "Guaranteed" error bound: The estimated forward error,
221 *> almost certainly within a factor of 10 of the true error
222 *> so long as the next entry is greater than the threshold
223 *> sqrt(n) * slamch('Epsilon'). This error bound should only
224 *> be trusted if the previous boolean is true.
226 *> err = 3 Reciprocal condition number: Estimated normwise
227 *> reciprocal condition number. Compared with the threshold
228 *> sqrt(n) * slamch('Epsilon') to determine if the error
229 *> estimate is "guaranteed". These reciprocal condition
230 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
231 *> appropriately scaled matrix Z.
232 *> Let Z = S*A, where S scales each row by a power of the
233 *> radix so all absolute row sums of Z are approximately 1.
235 *> This subroutine is only responsible for setting the second field
237 *> See Lapack Working Note 165 for further details and extra
241 *> \param[in,out] ERR_BNDS_COMP
243 *> ERR_BNDS_COMP is REAL array, dimension (NRHS, N_ERR_BNDS)
244 *> For each right-hand side, this array contains information about
245 *> various error bounds and condition numbers corresponding to the
246 *> componentwise relative error, which is defined as follows:
248 *> Componentwise relative error in the ith solution vector:
249 *> abs(XTRUE(j,i) - X(j,i))
250 *> max_j ----------------------
253 *> The array is indexed by the right-hand side i (on which the
254 *> componentwise relative error depends), and the type of error
255 *> information as described below. There currently are up to three
256 *> pieces of information returned for each right-hand side. If
257 *> componentwise accuracy is not requested (PARAMS(3) = 0.0), then
258 *> ERR_BNDS_COMP is not accessed. If N_ERR_BNDS .LT. 3, then at most
259 *> the first (:,N_ERR_BNDS) entries are returned.
261 *> The first index in ERR_BNDS_COMP(i,:) corresponds to the ith
264 *> The second index in ERR_BNDS_COMP(:,err) contains the following
266 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
267 *> reciprocal condition number is less than the threshold
268 *> sqrt(n) * slamch('Epsilon').
270 *> err = 2 "Guaranteed" error bound: The estimated forward error,
271 *> almost certainly within a factor of 10 of the true error
272 *> so long as the next entry is greater than the threshold
273 *> sqrt(n) * slamch('Epsilon'). This error bound should only
274 *> be trusted if the previous boolean is true.
276 *> err = 3 Reciprocal condition number: Estimated componentwise
277 *> reciprocal condition number. Compared with the threshold
278 *> sqrt(n) * slamch('Epsilon') to determine if the error
279 *> estimate is "guaranteed". These reciprocal condition
280 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
281 *> appropriately scaled matrix Z.
282 *> Let Z = S*(A*diag(x)), where x is the solution for the
283 *> current right-hand side and S scales each row of
284 *> A*diag(x) by a power of the radix so all absolute row
285 *> sums of Z are approximately 1.
287 *> This subroutine is only responsible for setting the second field
289 *> See Lapack Working Note 165 for further details and extra
295 *> RES is REAL array, dimension (N)
296 *> Workspace to hold the intermediate residual.
301 *> AYB is REAL array, dimension (N)
302 *> Workspace. This can be the same workspace passed for Y_TAIL.
307 *> DY is REAL array, dimension (N)
308 *> Workspace to hold the intermediate solution.
313 *> Y_TAIL is REAL array, dimension (N)
314 *> Workspace to hold the trailing bits of the intermediate solution.
320 *> Reciprocal scaled condition number. This is an estimate of the
321 *> reciprocal Skeel condition number of the matrix A after
322 *> equilibration (if done). If this is less than the machine
323 *> precision (in particular, if it is zero), the matrix is singular
324 *> to working precision. Note that the error may still be small even
325 *> if this number is very small and the matrix appears ill-
329 *> \param[in] ITHRESH
331 *> ITHRESH is INTEGER
332 *> The maximum number of residual computations allowed for
333 *> refinement. The default is 10. For 'aggressive' set to 100 to
334 *> permit convergence using approximate factorizations or
335 *> factorizations other than LU. If the factorization uses a
336 *> technique other than Gaussian elimination, the guarantees in
337 *> ERR_BNDS_NORM and ERR_BNDS_COMP may no longer be trustworthy.
340 *> \param[in] RTHRESH
343 *> Determines when to stop refinement if the error estimate stops
344 *> decreasing. Refinement will stop when the next solution no longer
345 *> satisfies norm(dx_{i+1}) < RTHRESH * norm(dx_i) where norm(Z) is
346 *> the infinity norm of Z. RTHRESH satisfies 0 < RTHRESH <= 1. The
347 *> default value is 0.5. For 'aggressive' set to 0.9 to permit
348 *> convergence on extremely ill-conditioned matrices. See LAWN 165
355 *> Determines when to start considering componentwise convergence.
356 *> Componentwise convergence is only considered after each component
357 *> of the solution Y is stable, which we definte as the relative
358 *> change in each component being less than DZ_UB. The default value
359 *> is 0.25, requiring the first bit to be stable. See LAWN 165 for
363 *> \param[in] IGNORE_CWISE
365 *> IGNORE_CWISE is LOGICAL
366 *> If .TRUE. then ignore componentwise convergence. Default value
373 *> = 0: Successful exit.
374 *> < 0: if INFO = -i, the ith argument to SLA_SYRFSX_EXTENDED had an illegal
381 *> \author Univ. of Tennessee
382 *> \author Univ. of California Berkeley
383 *> \author Univ. of Colorado Denver
386 *> \date September 2012
388 *> \ingroup realSYcomputational
390 * =====================================================================
391 SUBROUTINE SLA_SYRFSX_EXTENDED( PREC_TYPE, UPLO, N, NRHS, A, LDA,
392 $ AF, LDAF, IPIV, COLEQU, C, B, LDB,
393 $ Y, LDY, BERR_OUT, N_NORMS,
394 $ ERR_BNDS_NORM, ERR_BNDS_COMP, RES,
395 $ AYB, DY, Y_TAIL, RCOND, ITHRESH,
396 $ RTHRESH, DZ_UB, IGNORE_CWISE,
399 * -- LAPACK computational routine (version 3.4.2) --
400 * -- LAPACK is a software package provided by Univ. of Tennessee, --
401 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
404 * .. Scalar Arguments ..
405 INTEGER INFO, LDA, LDAF, LDB, LDY, N, NRHS, PREC_TYPE,
408 LOGICAL COLEQU, IGNORE_CWISE
411 * .. Array Arguments ..
413 REAL A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
414 $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * )
415 REAL C( * ), AYB( * ), RCOND, BERR_OUT( * ),
416 $ ERR_BNDS_NORM( NRHS, * ),
417 $ ERR_BNDS_COMP( NRHS, * )
420 * =====================================================================
422 * .. Local Scalars ..
423 INTEGER UPLO2, CNT, I, J, X_STATE, Z_STATE
424 REAL YK, DYK, YMIN, NORMY, NORMX, NORMDX, DXRAT,
425 $ DZRAT, PREVNORMDX, PREV_DZ_Z, DXRATMAX,
426 $ DZRATMAX, DX_X, DZ_Z, FINAL_DX_X, FINAL_DZ_Z,
427 $ EPS, HUGEVAL, INCR_THRESH
428 LOGICAL INCR_PREC, UPPER
431 INTEGER UNSTABLE_STATE, WORKING_STATE, CONV_STATE,
432 $ NOPROG_STATE, Y_PREC_STATE, BASE_RESIDUAL,
433 $ EXTRA_RESIDUAL, EXTRA_Y
434 PARAMETER ( UNSTABLE_STATE = 0, WORKING_STATE = 1,
435 $ CONV_STATE = 2, NOPROG_STATE = 3 )
436 PARAMETER ( BASE_RESIDUAL = 0, EXTRA_RESIDUAL = 1,
438 INTEGER FINAL_NRM_ERR_I, FINAL_CMP_ERR_I, BERR_I
439 INTEGER RCOND_I, NRM_RCOND_I, NRM_ERR_I, CMP_RCOND_I
440 INTEGER CMP_ERR_I, PIV_GROWTH_I
441 PARAMETER ( FINAL_NRM_ERR_I = 1, FINAL_CMP_ERR_I = 2,
443 PARAMETER ( RCOND_I = 4, NRM_RCOND_I = 5, NRM_ERR_I = 6 )
444 PARAMETER ( CMP_RCOND_I = 7, CMP_ERR_I = 8,
446 INTEGER LA_LINRX_ITREF_I, LA_LINRX_ITHRESH_I,
448 PARAMETER ( LA_LINRX_ITREF_I = 1,
449 $ LA_LINRX_ITHRESH_I = 2 )
450 PARAMETER ( LA_LINRX_CWISE_I = 3 )
451 INTEGER LA_LINRX_TRUST_I, LA_LINRX_ERR_I,
453 PARAMETER ( LA_LINRX_TRUST_I = 1, LA_LINRX_ERR_I = 2 )
454 PARAMETER ( LA_LINRX_RCOND_I = 3 )
456 * .. External Functions ..
461 * .. External Subroutines ..
462 EXTERNAL SAXPY, SCOPY, SSYTRS, SSYMV, BLAS_SSYMV_X,
463 $ BLAS_SSYMV2_X, SLA_SYAMV, SLA_WWADDW,
467 * .. Intrinsic Functions ..
468 INTRINSIC ABS, MAX, MIN
470 * .. Executable Statements ..
473 UPPER = LSAME( UPLO, 'U' )
474 IF( .NOT.UPPER .AND. .NOT.LSAME( UPLO, 'L' ) ) THEN
476 ELSE IF( N.LT.0 ) THEN
478 ELSE IF( NRHS.LT.0 ) THEN
480 ELSE IF( LDA.LT.MAX( 1, N ) ) THEN
482 ELSE IF( LDAF.LT.MAX( 1, N ) ) THEN
484 ELSE IF( LDB.LT.MAX( 1, N ) ) THEN
486 ELSE IF( LDY.LT.MAX( 1, N ) ) THEN
490 CALL XERBLA( 'SLA_SYRFSX_EXTENDED', -INFO )
493 EPS = SLAMCH( 'Epsilon' )
494 HUGEVAL = SLAMCH( 'Overflow' )
495 * Force HUGEVAL to Inf
496 HUGEVAL = HUGEVAL * HUGEVAL
497 * Using HUGEVAL may lead to spurious underflows.
498 INCR_THRESH = REAL( N )*EPS
500 IF ( LSAME ( UPLO, 'L' ) ) THEN
501 UPLO2 = ILAUPLO( 'L' )
503 UPLO2 = ILAUPLO( 'U' )
507 Y_PREC_STATE = EXTRA_RESIDUAL
508 IF ( Y_PREC_STATE .EQ. EXTRA_Y ) THEN
525 X_STATE = WORKING_STATE
526 Z_STATE = UNSTABLE_STATE
531 * Compute residual RES = B_s - op(A_s) * Y,
532 * op(A) = A, A**T, or A**H depending on TRANS (and type).
534 CALL SCOPY( N, B( 1, J ), 1, RES, 1 )
535 IF (Y_PREC_STATE .EQ. BASE_RESIDUAL) THEN
536 CALL SSYMV( UPLO, N, -1.0, A, LDA, Y(1,J), 1,
538 ELSE IF (Y_PREC_STATE .EQ. EXTRA_RESIDUAL) THEN
539 CALL BLAS_SSYMV_X( UPLO2, N, -1.0, A, LDA,
540 $ Y( 1, J ), 1, 1.0, RES, 1, PREC_TYPE )
542 CALL BLAS_SSYMV2_X(UPLO2, N, -1.0, A, LDA,
543 $ Y(1, J), Y_TAIL, 1, 1.0, RES, 1, PREC_TYPE)
546 ! XXX: RES is no longer needed.
547 CALL SCOPY( N, RES, 1, DY, 1 )
548 CALL SSYTRS( UPLO, N, 1, AF, LDAF, IPIV, DY, N, INFO )
550 * Calculate relative changes DX_X, DZ_Z and ratios DXRAT, DZRAT.
559 YK = ABS( Y( I, J ) )
562 IF ( YK .NE. 0.0 ) THEN
563 DZ_Z = MAX( DZ_Z, DYK / YK )
564 ELSE IF ( DYK .NE. 0.0 ) THEN
568 YMIN = MIN( YMIN, YK )
570 NORMY = MAX( NORMY, YK )
573 NORMX = MAX( NORMX, YK * C( I ) )
574 NORMDX = MAX( NORMDX, DYK * C( I ) )
577 NORMDX = MAX(NORMDX, DYK)
581 IF ( NORMX .NE. 0.0 ) THEN
582 DX_X = NORMDX / NORMX
583 ELSE IF ( NORMDX .EQ. 0.0 ) THEN
589 DXRAT = NORMDX / PREVNORMDX
590 DZRAT = DZ_Z / PREV_DZ_Z
592 * Check termination criteria.
594 IF ( YMIN*RCOND .LT. INCR_THRESH*NORMY
595 $ .AND. Y_PREC_STATE .LT. EXTRA_Y )
598 IF ( X_STATE .EQ. NOPROG_STATE .AND. DXRAT .LE. RTHRESH )
599 $ X_STATE = WORKING_STATE
600 IF ( X_STATE .EQ. WORKING_STATE ) THEN
601 IF ( DX_X .LE. EPS ) THEN
603 ELSE IF ( DXRAT .GT. RTHRESH ) THEN
604 IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
607 X_STATE = NOPROG_STATE
610 IF ( DXRAT .GT. DXRATMAX ) DXRATMAX = DXRAT
612 IF ( X_STATE .GT. WORKING_STATE ) FINAL_DX_X = DX_X
615 IF ( Z_STATE .EQ. UNSTABLE_STATE .AND. DZ_Z .LE. DZ_UB )
616 $ Z_STATE = WORKING_STATE
617 IF ( Z_STATE .EQ. NOPROG_STATE .AND. DZRAT .LE. RTHRESH )
618 $ Z_STATE = WORKING_STATE
619 IF ( Z_STATE .EQ. WORKING_STATE ) THEN
620 IF ( DZ_Z .LE. EPS ) THEN
622 ELSE IF ( DZ_Z .GT. DZ_UB ) THEN
623 Z_STATE = UNSTABLE_STATE
626 ELSE IF ( DZRAT .GT. RTHRESH ) THEN
627 IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
630 Z_STATE = NOPROG_STATE
633 IF ( DZRAT .GT. DZRATMAX ) DZRATMAX = DZRAT
635 IF ( Z_STATE .GT. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
638 IF ( X_STATE.NE.WORKING_STATE.AND.
639 $ ( IGNORE_CWISE.OR.Z_STATE.NE.WORKING_STATE ) )
642 IF ( INCR_PREC ) THEN
644 Y_PREC_STATE = Y_PREC_STATE + 1
655 IF (Y_PREC_STATE .LT. EXTRA_Y) THEN
656 CALL SAXPY( N, 1.0, DY, 1, Y(1,J), 1 )
658 CALL SLA_WWADDW( N, Y(1,J), Y_TAIL, DY )
662 * Target of "IF (Z_STOP .AND. X_STOP)". Sun's f77 won't EXIT.
665 * Set final_* when cnt hits ithresh.
667 IF ( X_STATE .EQ. WORKING_STATE ) FINAL_DX_X = DX_X
668 IF ( Z_STATE .EQ. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
670 * Compute error bounds.
672 IF ( N_NORMS .GE. 1 ) THEN
673 ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) =
674 $ FINAL_DX_X / (1 - DXRATMAX)
676 IF ( N_NORMS .GE. 2 ) THEN
677 ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) =
678 $ FINAL_DZ_Z / (1 - DZRATMAX)
681 * Compute componentwise relative backward error from formula
682 * max(i) ( abs(R(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
683 * where abs(Z) is the componentwise absolute value of the matrix
686 * Compute residual RES = B_s - op(A_s) * Y,
687 * op(A) = A, A**T, or A**H depending on TRANS (and type).
688 CALL SCOPY( N, B( 1, J ), 1, RES, 1 )
689 CALL SSYMV( UPLO, N, -1.0, A, LDA, Y(1,J), 1, 1.0, RES, 1 )
692 AYB( I ) = ABS( B( I, J ) )
695 * Compute abs(op(A_s))*abs(Y) + abs(B_s).
697 CALL SLA_SYAMV( UPLO2, N, 1.0,
698 $ A, LDA, Y(1, J), 1, 1.0, AYB, 1 )
700 CALL SLA_LIN_BERR( N, N, 1, RES, AYB, BERR_OUT( J ) )
702 * End of loop for each RHS.