1 *> \brief \b DLA_PORFSX_EXTENDED improves the computed solution to a system of linear equations for symmetric or Hermitian positive-definite 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 DLA_PORFSX_EXTENDED + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dla_porfsx_extended.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dla_porfsx_extended.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dla_porfsx_extended.f">
21 * SUBROUTINE DLA_PORFSX_EXTENDED( PREC_TYPE, UPLO, N, NRHS, A, LDA,
22 * AF, LDAF, COLEQU, C, B, LDB, Y,
23 * 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
34 * DOUBLE PRECISION RTHRESH, DZ_UB
36 * .. Array Arguments ..
37 * DOUBLE PRECISION A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
38 * $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * )
39 * DOUBLE PRECISION C( * ), AYB(*), RCOND, BERR_OUT( * ),
40 * $ ERR_BNDS_NORM( NRHS, * ),
41 * $ ERR_BNDS_COMP( NRHS, * )
50 *> DLA_PORFSX_EXTENDED improves the computed solution to a system of
51 *> linear equations by performing extra-precise iterative refinement
52 *> and provides error bounds and backward error estimates for the solution.
53 *> This subroutine is called by DPORFSX to perform iterative refinement.
54 *> In addition to normwise error bound, the code provides maximum
55 *> componentwise error bound if possible. See comments for ERR_BNDS_NORM
56 *> and ERR_BNDS_COMP for details of the error bounds. Note that this
57 *> subroutine is only resonsible for setting the second fields of
58 *> ERR_BNDS_NORM and ERR_BNDS_COMP.
64 *> \param[in] PREC_TYPE
66 *> PREC_TYPE is INTEGER
67 *> Specifies the intermediate precision to be used in refinement.
68 *> The value is defined by ILAPREC(P) where P is a CHARACTER and
77 *> UPLO is CHARACTER*1
78 *> = 'U': Upper triangle of A is stored;
79 *> = 'L': Lower triangle of A is stored.
85 *> The number of linear equations, i.e., the order of the
92 *> The number of right-hand-sides, i.e., the number of columns of the
98 *> A is DOUBLE PRECISION array, dimension (LDA,N)
99 *> On entry, the N-by-N matrix A.
105 *> The leading dimension of the array A. LDA >= max(1,N).
110 *> AF is DOUBLE PRECISION array, dimension (LDAF,N)
111 *> The triangular factor U or L from the Cholesky factorization
112 *> A = U**T*U or A = L*L**T, as computed by DPOTRF.
118 *> The leading dimension of the array AF. LDAF >= max(1,N).
124 *> If .TRUE. then column equilibration was done to A before calling
125 *> this routine. This is needed to compute the solution and error
131 *> C is DOUBLE PRECISION array, dimension (N)
132 *> The column scale factors for A. If COLEQU = .FALSE., C
133 *> is not accessed. If C is input, each element of C should be a power
134 *> of the radix to ensure a reliable solution and error estimates.
135 *> Scaling by powers of the radix does not cause rounding errors unless
136 *> the result underflows or overflows. Rounding errors during scaling
137 *> lead to refining with a matrix that is not equivalent to the
138 *> input matrix, producing error estimates that may not be
144 *> B is DOUBLE PRECISION array, dimension (LDB,NRHS)
145 *> The right-hand-side matrix B.
151 *> The leading dimension of the array B. LDB >= max(1,N).
156 *> Y is DOUBLE PRECISION array, dimension
158 *> On entry, the solution matrix X, as computed by DPOTRS.
159 *> On exit, the improved solution matrix Y.
165 *> The leading dimension of the array Y. LDY >= max(1,N).
168 *> \param[out] BERR_OUT
170 *> BERR_OUT is DOUBLE PRECISION array, dimension (NRHS)
171 *> On exit, BERR_OUT(j) contains the componentwise relative backward
172 *> error for right-hand-side j from the formula
173 *> max(i) ( abs(RES(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
174 *> where abs(Z) is the componentwise absolute value of the matrix
175 *> or vector Z. This is computed by DLA_LIN_BERR.
178 *> \param[in] N_NORMS
180 *> N_NORMS is INTEGER
181 *> Determines which error bounds to return (see ERR_BNDS_NORM
182 *> and ERR_BNDS_COMP).
183 *> If N_NORMS >= 1 return normwise error bounds.
184 *> If N_NORMS >= 2 return componentwise error bounds.
187 *> \param[in,out] ERR_BNDS_NORM
189 *> ERR_BNDS_NORM is DOUBLE PRECISION array, dimension
190 *> (NRHS, N_ERR_BNDS)
191 *> For each right-hand side, this array contains information about
192 *> various error bounds and condition numbers corresponding to the
193 *> normwise relative error, which is defined as follows:
195 *> Normwise relative error in the ith solution vector:
196 *> max_j (abs(XTRUE(j,i) - X(j,i)))
197 *> ------------------------------
200 *> The array is indexed by the type of error information as described
201 *> below. There currently are up to three pieces of information
204 *> The first index in ERR_BNDS_NORM(i,:) corresponds to the ith
207 *> The second index in ERR_BNDS_NORM(:,err) contains the following
209 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
210 *> reciprocal condition number is less than the threshold
211 *> sqrt(n) * slamch('Epsilon').
213 *> err = 2 "Guaranteed" error bound: The estimated forward error,
214 *> almost certainly within a factor of 10 of the true error
215 *> so long as the next entry is greater than the threshold
216 *> sqrt(n) * slamch('Epsilon'). This error bound should only
217 *> be trusted if the previous boolean is true.
219 *> err = 3 Reciprocal condition number: Estimated normwise
220 *> reciprocal condition number. Compared with the threshold
221 *> sqrt(n) * slamch('Epsilon') to determine if the error
222 *> estimate is "guaranteed". These reciprocal condition
223 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
224 *> appropriately scaled matrix Z.
225 *> Let Z = S*A, where S scales each row by a power of the
226 *> radix so all absolute row sums of Z are approximately 1.
228 *> This subroutine is only responsible for setting the second field
230 *> See Lapack Working Note 165 for further details and extra
234 *> \param[in,out] ERR_BNDS_COMP
236 *> ERR_BNDS_COMP is DOUBLE PRECISION array, dimension
237 *> (NRHS, N_ERR_BNDS)
238 *> For each right-hand side, this array contains information about
239 *> various error bounds and condition numbers corresponding to the
240 *> componentwise relative error, which is defined as follows:
242 *> Componentwise relative error in the ith solution vector:
243 *> abs(XTRUE(j,i) - X(j,i))
244 *> max_j ----------------------
247 *> The array is indexed by the right-hand side i (on which the
248 *> componentwise relative error depends), and the type of error
249 *> information as described below. There currently are up to three
250 *> pieces of information returned for each right-hand side. If
251 *> componentwise accuracy is not requested (PARAMS(3) = 0.0), then
252 *> ERR_BNDS_COMP is not accessed. If N_ERR_BNDS .LT. 3, then at most
253 *> the first (:,N_ERR_BNDS) entries are returned.
255 *> The first index in ERR_BNDS_COMP(i,:) corresponds to the ith
258 *> The second index in ERR_BNDS_COMP(:,err) contains the following
260 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
261 *> reciprocal condition number is less than the threshold
262 *> sqrt(n) * slamch('Epsilon').
264 *> err = 2 "Guaranteed" error bound: The estimated forward error,
265 *> almost certainly within a factor of 10 of the true error
266 *> so long as the next entry is greater than the threshold
267 *> sqrt(n) * slamch('Epsilon'). This error bound should only
268 *> be trusted if the previous boolean is true.
270 *> err = 3 Reciprocal condition number: Estimated componentwise
271 *> reciprocal condition number. Compared with the threshold
272 *> sqrt(n) * slamch('Epsilon') to determine if the error
273 *> estimate is "guaranteed". These reciprocal condition
274 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
275 *> appropriately scaled matrix Z.
276 *> Let Z = S*(A*diag(x)), where x is the solution for the
277 *> current right-hand side and S scales each row of
278 *> A*diag(x) by a power of the radix so all absolute row
279 *> sums of Z are approximately 1.
281 *> This subroutine is only responsible for setting the second field
283 *> See Lapack Working Note 165 for further details and extra
289 *> RES is DOUBLE PRECISION array, dimension (N)
290 *> Workspace to hold the intermediate residual.
295 *> AYB is DOUBLE PRECISION array, dimension (N)
296 *> Workspace. This can be the same workspace passed for Y_TAIL.
301 *> DY is DOUBLE PRECISION array, dimension (N)
302 *> Workspace to hold the intermediate solution.
307 *> Y_TAIL is DOUBLE PRECISION array, dimension (N)
308 *> Workspace to hold the trailing bits of the intermediate solution.
313 *> RCOND is DOUBLE PRECISION
314 *> Reciprocal scaled condition number. This is an estimate of the
315 *> reciprocal Skeel condition number of the matrix A after
316 *> equilibration (if done). If this is less than the machine
317 *> precision (in particular, if it is zero), the matrix is singular
318 *> to working precision. Note that the error may still be small even
319 *> if this number is very small and the matrix appears ill-
323 *> \param[in] ITHRESH
325 *> ITHRESH is INTEGER
326 *> The maximum number of residual computations allowed for
327 *> refinement. The default is 10. For 'aggressive' set to 100 to
328 *> permit convergence using approximate factorizations or
329 *> factorizations other than LU. If the factorization uses a
330 *> technique other than Gaussian elimination, the guarantees in
331 *> ERR_BNDS_NORM and ERR_BNDS_COMP may no longer be trustworthy.
334 *> \param[in] RTHRESH
336 *> RTHRESH is DOUBLE PRECISION
337 *> Determines when to stop refinement if the error estimate stops
338 *> decreasing. Refinement will stop when the next solution no longer
339 *> satisfies norm(dx_{i+1}) < RTHRESH * norm(dx_i) where norm(Z) is
340 *> the infinity norm of Z. RTHRESH satisfies 0 < RTHRESH <= 1. The
341 *> default value is 0.5. For 'aggressive' set to 0.9 to permit
342 *> convergence on extremely ill-conditioned matrices. See LAWN 165
348 *> DZ_UB is DOUBLE PRECISION
349 *> Determines when to start considering componentwise convergence.
350 *> Componentwise convergence is only considered after each component
351 *> of the solution Y is stable, which we definte as the relative
352 *> change in each component being less than DZ_UB. The default value
353 *> is 0.25, requiring the first bit to be stable. See LAWN 165 for
357 *> \param[in] IGNORE_CWISE
359 *> IGNORE_CWISE is LOGICAL
360 *> If .TRUE. then ignore componentwise convergence. Default value
367 *> = 0: Successful exit.
368 *> < 0: if INFO = -i, the ith argument to DPOTRS had an illegal
375 *> \author Univ. of Tennessee
376 *> \author Univ. of California Berkeley
377 *> \author Univ. of Colorado Denver
380 *> \date September 2012
382 *> \ingroup doublePOcomputational
384 * =====================================================================
385 SUBROUTINE DLA_PORFSX_EXTENDED( PREC_TYPE, UPLO, N, NRHS, A, LDA,
386 $ AF, LDAF, COLEQU, C, B, LDB, Y,
387 $ LDY, BERR_OUT, N_NORMS,
388 $ ERR_BNDS_NORM, ERR_BNDS_COMP, RES,
389 $ AYB, DY, Y_TAIL, RCOND, ITHRESH,
390 $ RTHRESH, DZ_UB, IGNORE_CWISE,
393 * -- LAPACK computational routine (version 3.4.2) --
394 * -- LAPACK is a software package provided by Univ. of Tennessee, --
395 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
398 * .. Scalar Arguments ..
399 INTEGER INFO, LDA, LDAF, LDB, LDY, N, NRHS, PREC_TYPE,
402 LOGICAL COLEQU, IGNORE_CWISE
403 DOUBLE PRECISION RTHRESH, DZ_UB
405 * .. Array Arguments ..
406 DOUBLE PRECISION A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
407 $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * )
408 DOUBLE PRECISION C( * ), AYB(*), RCOND, BERR_OUT( * ),
409 $ ERR_BNDS_NORM( NRHS, * ),
410 $ ERR_BNDS_COMP( NRHS, * )
413 * =====================================================================
415 * .. Local Scalars ..
416 INTEGER UPLO2, CNT, I, J, X_STATE, Z_STATE
417 DOUBLE PRECISION YK, DYK, YMIN, NORMY, NORMX, NORMDX, DXRAT,
418 $ DZRAT, PREVNORMDX, PREV_DZ_Z, DXRATMAX,
419 $ DZRATMAX, DX_X, DZ_Z, FINAL_DX_X, FINAL_DZ_Z,
420 $ EPS, HUGEVAL, INCR_THRESH
424 INTEGER UNSTABLE_STATE, WORKING_STATE, CONV_STATE,
425 $ NOPROG_STATE, Y_PREC_STATE, BASE_RESIDUAL,
426 $ EXTRA_RESIDUAL, EXTRA_Y
427 PARAMETER ( UNSTABLE_STATE = 0, WORKING_STATE = 1,
428 $ CONV_STATE = 2, NOPROG_STATE = 3 )
429 PARAMETER ( BASE_RESIDUAL = 0, EXTRA_RESIDUAL = 1,
431 INTEGER FINAL_NRM_ERR_I, FINAL_CMP_ERR_I, BERR_I
432 INTEGER RCOND_I, NRM_RCOND_I, NRM_ERR_I, CMP_RCOND_I
433 INTEGER CMP_ERR_I, PIV_GROWTH_I
434 PARAMETER ( FINAL_NRM_ERR_I = 1, FINAL_CMP_ERR_I = 2,
436 PARAMETER ( RCOND_I = 4, NRM_RCOND_I = 5, NRM_ERR_I = 6 )
437 PARAMETER ( CMP_RCOND_I = 7, CMP_ERR_I = 8,
439 INTEGER LA_LINRX_ITREF_I, LA_LINRX_ITHRESH_I,
441 PARAMETER ( LA_LINRX_ITREF_I = 1,
442 $ LA_LINRX_ITHRESH_I = 2 )
443 PARAMETER ( LA_LINRX_CWISE_I = 3 )
444 INTEGER LA_LINRX_TRUST_I, LA_LINRX_ERR_I,
446 PARAMETER ( LA_LINRX_TRUST_I = 1, LA_LINRX_ERR_I = 2 )
447 PARAMETER ( LA_LINRX_RCOND_I = 3 )
449 * .. External Functions ..
454 * .. External Subroutines ..
455 EXTERNAL DAXPY, DCOPY, DPOTRS, DSYMV, BLAS_DSYMV_X,
456 $ BLAS_DSYMV2_X, DLA_SYAMV, DLA_WWADDW,
458 DOUBLE PRECISION DLAMCH
460 * .. Intrinsic Functions ..
461 INTRINSIC ABS, MAX, MIN
463 * .. Executable Statements ..
465 IF (INFO.NE.0) RETURN
466 EPS = DLAMCH( 'Epsilon' )
467 HUGEVAL = DLAMCH( 'Overflow' )
468 * Force HUGEVAL to Inf
469 HUGEVAL = HUGEVAL * HUGEVAL
470 * Using HUGEVAL may lead to spurious underflows.
471 INCR_THRESH = DBLE( N ) * EPS
473 IF ( LSAME ( UPLO, 'L' ) ) THEN
474 UPLO2 = ILAUPLO( 'L' )
476 UPLO2 = ILAUPLO( 'U' )
480 Y_PREC_STATE = EXTRA_RESIDUAL
481 IF ( Y_PREC_STATE .EQ. EXTRA_Y ) THEN
498 X_STATE = WORKING_STATE
499 Z_STATE = UNSTABLE_STATE
504 * Compute residual RES = B_s - op(A_s) * Y,
505 * op(A) = A, A**T, or A**H depending on TRANS (and type).
507 CALL DCOPY( N, B( 1, J ), 1, RES, 1 )
508 IF ( Y_PREC_STATE .EQ. BASE_RESIDUAL ) THEN
509 CALL DSYMV( UPLO, N, -1.0D+0, A, LDA, Y(1,J), 1,
511 ELSE IF ( Y_PREC_STATE .EQ. EXTRA_RESIDUAL ) THEN
512 CALL BLAS_DSYMV_X( UPLO2, N, -1.0D+0, A, LDA,
513 $ Y( 1, J ), 1, 1.0D+0, RES, 1, PREC_TYPE )
515 CALL BLAS_DSYMV2_X(UPLO2, N, -1.0D+0, A, LDA,
516 $ Y(1, J), Y_TAIL, 1, 1.0D+0, RES, 1, PREC_TYPE)
519 ! XXX: RES is no longer needed.
520 CALL DCOPY( N, RES, 1, DY, 1 )
521 CALL DPOTRS( UPLO, N, 1, AF, LDAF, DY, N, INFO )
523 * Calculate relative changes DX_X, DZ_Z and ratios DXRAT, DZRAT.
532 YK = ABS( Y( I, J ) )
535 IF ( YK .NE. 0.0D+0 ) THEN
536 DZ_Z = MAX( DZ_Z, DYK / YK )
537 ELSE IF ( DYK .NE. 0.0D+0 ) THEN
541 YMIN = MIN( YMIN, YK )
543 NORMY = MAX( NORMY, YK )
546 NORMX = MAX( NORMX, YK * C( I ) )
547 NORMDX = MAX( NORMDX, DYK * C( I ) )
550 NORMDX = MAX( NORMDX, DYK )
554 IF ( NORMX .NE. 0.0D+0 ) THEN
555 DX_X = NORMDX / NORMX
556 ELSE IF ( NORMDX .EQ. 0.0D+0 ) THEN
562 DXRAT = NORMDX / PREVNORMDX
563 DZRAT = DZ_Z / PREV_DZ_Z
565 * Check termination criteria.
567 IF ( YMIN*RCOND .LT. INCR_THRESH*NORMY
568 $ .AND. Y_PREC_STATE .LT. EXTRA_Y )
571 IF ( X_STATE .EQ. NOPROG_STATE .AND. DXRAT .LE. RTHRESH )
572 $ X_STATE = WORKING_STATE
573 IF ( X_STATE .EQ. WORKING_STATE ) THEN
574 IF ( DX_X .LE. EPS ) THEN
576 ELSE IF ( DXRAT .GT. RTHRESH ) THEN
577 IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
580 X_STATE = NOPROG_STATE
583 IF ( DXRAT .GT. DXRATMAX ) DXRATMAX = DXRAT
585 IF ( X_STATE .GT. WORKING_STATE ) FINAL_DX_X = DX_X
588 IF ( Z_STATE .EQ. UNSTABLE_STATE .AND. DZ_Z .LE. DZ_UB )
589 $ Z_STATE = WORKING_STATE
590 IF ( Z_STATE .EQ. NOPROG_STATE .AND. DZRAT .LE. RTHRESH )
591 $ Z_STATE = WORKING_STATE
592 IF ( Z_STATE .EQ. WORKING_STATE ) THEN
593 IF ( DZ_Z .LE. EPS ) THEN
595 ELSE IF ( DZ_Z .GT. DZ_UB ) THEN
596 Z_STATE = UNSTABLE_STATE
599 ELSE IF ( DZRAT .GT. RTHRESH ) THEN
600 IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
603 Z_STATE = NOPROG_STATE
606 IF ( DZRAT .GT. DZRATMAX ) DZRATMAX = DZRAT
608 IF ( Z_STATE .GT. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
611 IF ( X_STATE.NE.WORKING_STATE.AND.
612 $ ( IGNORE_CWISE.OR.Z_STATE.NE.WORKING_STATE ) )
615 IF ( INCR_PREC ) THEN
617 Y_PREC_STATE = Y_PREC_STATE + 1
628 IF (Y_PREC_STATE .LT. EXTRA_Y) THEN
629 CALL DAXPY( N, 1.0D+0, DY, 1, Y(1,J), 1 )
631 CALL DLA_WWADDW( N, Y( 1, J ), Y_TAIL, DY )
635 * Target of "IF (Z_STOP .AND. X_STOP)". Sun's f77 won't EXIT.
638 * Set final_* when cnt hits ithresh.
640 IF ( X_STATE .EQ. WORKING_STATE ) FINAL_DX_X = DX_X
641 IF ( Z_STATE .EQ. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
643 * Compute error bounds.
645 IF ( N_NORMS .GE. 1 ) THEN
646 ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) =
647 $ FINAL_DX_X / (1 - DXRATMAX)
649 IF ( N_NORMS .GE. 2 ) THEN
650 ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) =
651 $ FINAL_DZ_Z / (1 - DZRATMAX)
654 * Compute componentwise relative backward error from formula
655 * max(i) ( abs(R(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
656 * where abs(Z) is the componentwise absolute value of the matrix
659 * Compute residual RES = B_s - op(A_s) * Y,
660 * op(A) = A, A**T, or A**H depending on TRANS (and type).
662 CALL DCOPY( N, B( 1, J ), 1, RES, 1 )
663 CALL DSYMV( UPLO, N, -1.0D+0, A, LDA, Y(1,J), 1, 1.0D+0, RES,
667 AYB( I ) = ABS( B( I, J ) )
670 * Compute abs(op(A_s))*abs(Y) + abs(B_s).
672 CALL DLA_SYAMV( UPLO2, N, 1.0D+0,
673 $ A, LDA, Y(1, J), 1, 1.0D+0, AYB, 1 )
675 CALL DLA_LIN_BERR( N, N, 1, RES, AYB, BERR_OUT( J ) )
677 * End of loop for each RHS.