1 *> \brief \b CLA_GERFSX_EXTENDED
3 * =========== DOCUMENTATION ===========
5 * Online html documentation available at
6 * http://www.netlib.org/lapack/explore-html/
9 *> Download CLA_GERFSX_EXTENDED + dependencies
10 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/cla_gerfsx_extended.f">
12 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/cla_gerfsx_extended.f">
14 *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/cla_gerfsx_extended.f">
21 * SUBROUTINE CLA_GERFSX_EXTENDED( PREC_TYPE, TRANS_TYPE, N, NRHS, A,
22 * LDA, AF, LDAF, IPIV, COLEQU, C, B,
23 * LDB, Y, LDY, BERR_OUT, N_NORMS,
24 * ERRS_N, ERRS_C, RES, AYB, DY,
25 * Y_TAIL, RCOND, ITHRESH, RTHRESH,
26 * DZ_UB, IGNORE_CWISE, INFO )
28 * .. Scalar Arguments ..
29 * INTEGER INFO, LDA, LDAF, LDB, LDY, N, NRHS, PREC_TYPE,
30 * $ TRANS_TYPE, N_NORMS
31 * LOGICAL COLEQU, IGNORE_CWISE
37 * COMPLEX A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
38 * $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * )
39 * REAL C( * ), AYB( * ), RCOND, BERR_OUT( * ),
40 * $ ERRS_N( NRHS, * ), ERRS_C( NRHS, * )
50 *> CLA_GERFSX_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 CGERFSX to perform iterative refinement.
54 *> In addition to normwise error bound, the code provides maximum
55 *> componentwise error bound if possible. See comments for ERRS_N
56 *> and ERRS_C for details of the error bounds. Note that this
57 *> subroutine is only resonsible for setting the second fields of
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
75 *> \param[in] TRANS_TYPE
77 *> TRANS_TYPE is INTEGER
78 *> Specifies the transposition operation on A.
79 *> The value is defined by ILATRANS(T) where T is a CHARACTER and
80 *> T = 'N': No transpose
82 *> = 'C': Conjugate transpose
88 *> The number of linear equations, i.e., the order of the
95 *> The number of right-hand-sides, i.e., the number of columns of the
101 *> A is COMPLEX array, dimension (LDA,N)
102 *> On entry, the N-by-N matrix A.
108 *> The leading dimension of the array A. LDA >= max(1,N).
113 *> AF is COMPLEX array, dimension (LDAF,N)
114 *> The factors L and U from the factorization
115 *> A = P*L*U as computed by CGETRF.
121 *> The leading dimension of the array AF. LDAF >= max(1,N).
126 *> IPIV is INTEGER array, dimension (N)
127 *> The pivot indices from the factorization A = P*L*U
128 *> as computed by CGETRF; row i of the matrix was interchanged
135 *> If .TRUE. then column equilibration was done to A before calling
136 *> this routine. This is needed to compute the solution and error
142 *> C is REAL array, dimension (N)
143 *> The column scale factors for A. If COLEQU = .FALSE., C
144 *> is not accessed. If C is input, each element of C should be a power
145 *> of the radix to ensure a reliable solution and error estimates.
146 *> Scaling by powers of the radix does not cause rounding errors unless
147 *> the result underflows or overflows. Rounding errors during scaling
148 *> lead to refining with a matrix that is not equivalent to the
149 *> input matrix, producing error estimates that may not be
155 *> B is COMPLEX array, dimension (LDB,NRHS)
156 *> The right-hand-side matrix B.
162 *> The leading dimension of the array B. LDB >= max(1,N).
167 *> Y is COMPLEX array, dimension (LDY,NRHS)
168 *> On entry, the solution matrix X, as computed by CGETRS.
169 *> On exit, the improved solution matrix Y.
175 *> The leading dimension of the array Y. LDY >= max(1,N).
178 *> \param[out] BERR_OUT
180 *> BERR_OUT is REAL array, dimension (NRHS)
181 *> On exit, BERR_OUT(j) contains the componentwise relative backward
182 *> error for right-hand-side j from the formula
183 *> max(i) ( abs(RES(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
184 *> where abs(Z) is the componentwise absolute value of the matrix
185 *> or vector Z. This is computed by CLA_LIN_BERR.
188 *> \param[in] N_NORMS
190 *> N_NORMS is INTEGER
191 *> Determines which error bounds to return (see ERRS_N
193 *> If N_NORMS >= 1 return normwise error bounds.
194 *> If N_NORMS >= 2 return componentwise error bounds.
197 *> \param[in,out] ERRS_N
199 *> ERRS_N is REAL array, dimension (NRHS, N_ERR_BNDS)
200 *> For each right-hand side, this array contains information about
201 *> various error bounds and condition numbers corresponding to the
202 *> normwise relative error, which is defined as follows:
204 *> Normwise relative error in the ith solution vector:
205 *> max_j (abs(XTRUE(j,i) - X(j,i)))
206 *> ------------------------------
209 *> The array is indexed by the type of error information as described
210 *> below. There currently are up to three pieces of information
213 *> The first index in ERRS_N(i,:) corresponds to the ith
216 *> The second index in ERRS_N(:,err) contains the following
218 *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
219 *> reciprocal condition number is less than the threshold
220 *> sqrt(n) * slamch('Epsilon').
222 *> err = 2 "Guaranteed" error bound: The estimated forward error,
223 *> almost certainly within a factor of 10 of the true error
224 *> so long as the next entry is greater than the threshold
225 *> sqrt(n) * slamch('Epsilon'). This error bound should only
226 *> be trusted if the previous boolean is true.
228 *> err = 3 Reciprocal condition number: Estimated normwise
229 *> reciprocal condition number. Compared with the threshold
230 *> sqrt(n) * slamch('Epsilon') to determine if the error
231 *> estimate is "guaranteed". These reciprocal condition
232 *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
233 *> appropriately scaled matrix Z.
234 *> Let Z = S*A, where S scales each row by a power of the
235 *> radix so all absolute row sums of Z are approximately 1.
237 *> This subroutine is only responsible for setting the second field
239 *> See Lapack Working Note 165 for further details and extra
243 *> \param[in,out] ERRS_C
245 *> ERRS_C is REAL array, dimension (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 *> ERRS_C 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 ERRS_C(i,:) corresponds to the ith
266 *> The second index in ERRS_C(:,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 *> ERRS_N and ERRS_C 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 CGETRS had an illegal
383 *> \author Univ. of Tennessee
384 *> \author Univ. of California Berkeley
385 *> \author Univ. of Colorado Denver
388 *> \date November 2011
390 *> \ingroup complexGEcomputational
392 * =====================================================================
393 SUBROUTINE CLA_GERFSX_EXTENDED( PREC_TYPE, TRANS_TYPE, N, NRHS, A,
394 $ LDA, AF, LDAF, IPIV, COLEQU, C, B,
395 $ LDB, Y, LDY, BERR_OUT, N_NORMS,
396 $ ERRS_N, ERRS_C, RES, AYB, DY,
397 $ Y_TAIL, RCOND, ITHRESH, RTHRESH,
398 $ DZ_UB, IGNORE_CWISE, INFO )
400 * -- LAPACK computational routine (version 3.4.0) --
401 * -- LAPACK is a software package provided by Univ. of Tennessee, --
402 * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
405 * .. Scalar Arguments ..
406 INTEGER INFO, LDA, LDAF, LDB, LDY, N, NRHS, PREC_TYPE,
407 $ TRANS_TYPE, N_NORMS
408 LOGICAL COLEQU, IGNORE_CWISE
414 COMPLEX A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
415 $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * )
416 REAL C( * ), AYB( * ), RCOND, BERR_OUT( * ),
417 $ ERRS_N( NRHS, * ), ERRS_C( NRHS, * )
420 * =====================================================================
422 * .. Local Scalars ..
424 INTEGER CNT, I, J, X_STATE, Z_STATE, Y_PREC_STATE
425 REAL YK, DYK, YMIN, NORMY, NORMX, NORMDX, DXRAT,
426 $ DZRAT, PREVNORMDX, PREV_DZ_Z, DXRATMAX,
427 $ DZRATMAX, DX_X, DZ_Z, FINAL_DX_X, FINAL_DZ_Z,
428 $ EPS, HUGEVAL, INCR_THRESH
433 INTEGER UNSTABLE_STATE, WORKING_STATE, CONV_STATE,
434 $ NOPROG_STATE, BASE_RESIDUAL, EXTRA_RESIDUAL,
436 PARAMETER ( UNSTABLE_STATE = 0, WORKING_STATE = 1,
439 PARAMETER ( BASE_RESIDUAL = 0, EXTRA_RESIDUAL = 1,
441 INTEGER FINAL_NRM_ERR_I, FINAL_CMP_ERR_I, BERR_I
442 INTEGER RCOND_I, NRM_RCOND_I, NRM_ERR_I, CMP_RCOND_I
443 INTEGER CMP_ERR_I, PIV_GROWTH_I
444 PARAMETER ( FINAL_NRM_ERR_I = 1, FINAL_CMP_ERR_I = 2,
446 PARAMETER ( RCOND_I = 4, NRM_RCOND_I = 5, NRM_ERR_I = 6 )
447 PARAMETER ( CMP_RCOND_I = 7, CMP_ERR_I = 8,
449 INTEGER LA_LINRX_ITREF_I, LA_LINRX_ITHRESH_I,
451 PARAMETER ( LA_LINRX_ITREF_I = 1,
452 $ LA_LINRX_ITHRESH_I = 2 )
453 PARAMETER ( LA_LINRX_CWISE_I = 3 )
454 INTEGER LA_LINRX_TRUST_I, LA_LINRX_ERR_I,
456 PARAMETER ( LA_LINRX_TRUST_I = 1, LA_LINRX_ERR_I = 2 )
457 PARAMETER ( LA_LINRX_RCOND_I = 3 )
459 * .. External Subroutines ..
460 EXTERNAL CAXPY, CCOPY, CGETRS, CGEMV, BLAS_CGEMV_X,
461 $ BLAS_CGEMV2_X, CLA_GEAMV, CLA_WWADDW, SLAMCH,
462 $ CHLA_TRANSTYPE, CLA_LIN_BERR
464 CHARACTER CHLA_TRANSTYPE
466 * .. Intrinsic Functions ..
467 INTRINSIC ABS, MAX, MIN
469 * .. Statement Functions ..
472 * .. Statement Function Definitions ..
473 CABS1( ZDUM ) = ABS( REAL( ZDUM ) ) + ABS( AIMAG( ZDUM ) )
475 * .. Executable Statements ..
477 IF ( INFO.NE.0 ) RETURN
478 TRANS = CHLA_TRANSTYPE(TRANS_TYPE)
479 EPS = SLAMCH( 'Epsilon' )
480 HUGEVAL = SLAMCH( 'Overflow' )
481 * Force HUGEVAL to Inf
482 HUGEVAL = HUGEVAL * HUGEVAL
483 * Using HUGEVAL may lead to spurious underflows.
484 INCR_THRESH = REAL( N ) * EPS
487 Y_PREC_STATE = EXTRA_RESIDUAL
488 IF ( Y_PREC_STATE .EQ. EXTRA_Y ) THEN
505 X_STATE = WORKING_STATE
506 Z_STATE = UNSTABLE_STATE
511 * Compute residual RES = B_s - op(A_s) * Y,
512 * op(A) = A, A**T, or A**H depending on TRANS (and type).
514 CALL CCOPY( N, B( 1, J ), 1, RES, 1 )
515 IF ( Y_PREC_STATE .EQ. BASE_RESIDUAL ) THEN
516 CALL CGEMV( TRANS, N, N, (-1.0E+0,0.0E+0), A, LDA,
517 $ Y( 1, J ), 1, (1.0E+0,0.0E+0), RES, 1)
518 ELSE IF (Y_PREC_STATE .EQ. EXTRA_RESIDUAL) THEN
519 CALL BLAS_CGEMV_X( TRANS_TYPE, N, N, (-1.0E+0,0.0E+0), A,
520 $ LDA, Y( 1, J ), 1, (1.0E+0,0.0E+0),
521 $ RES, 1, PREC_TYPE )
523 CALL BLAS_CGEMV2_X( TRANS_TYPE, N, N, (-1.0E+0,0.0E+0),
524 $ A, LDA, Y(1, J), Y_TAIL, 1, (1.0E+0,0.0E+0), RES, 1,
528 ! XXX: RES is no longer needed.
529 CALL CCOPY( N, RES, 1, DY, 1 )
530 CALL CGETRS( TRANS, N, 1, AF, LDAF, IPIV, DY, N, INFO )
532 * Calculate relative changes DX_X, DZ_Z and ratios DXRAT, DZRAT.
541 YK = CABS1( Y( I, J ) )
542 DYK = CABS1( DY( I ) )
544 IF ( YK .NE. 0.0E+0 ) THEN
545 DZ_Z = MAX( DZ_Z, DYK / YK )
546 ELSE IF ( DYK .NE. 0.0 ) THEN
550 YMIN = MIN( YMIN, YK )
552 NORMY = MAX( NORMY, YK )
555 NORMX = MAX( NORMX, YK * C( I ) )
556 NORMDX = MAX( NORMDX, DYK * C( I ) )
559 NORMDX = MAX(NORMDX, DYK)
563 IF ( NORMX .NE. 0.0 ) THEN
564 DX_X = NORMDX / NORMX
565 ELSE IF ( NORMDX .EQ. 0.0 ) THEN
571 DXRAT = NORMDX / PREVNORMDX
572 DZRAT = DZ_Z / PREV_DZ_Z
574 * Check termination criteria
576 IF (.NOT.IGNORE_CWISE
577 $ .AND. YMIN*RCOND .LT. INCR_THRESH*NORMY
578 $ .AND. Y_PREC_STATE .LT. EXTRA_Y )
581 IF ( X_STATE .EQ. NOPROG_STATE .AND. DXRAT .LE. RTHRESH )
582 $ X_STATE = WORKING_STATE
583 IF ( X_STATE .EQ. WORKING_STATE ) THEN
584 IF (DX_X .LE. EPS) THEN
586 ELSE IF ( DXRAT .GT. RTHRESH ) THEN
587 IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
590 X_STATE = NOPROG_STATE
593 IF ( DXRAT .GT. DXRATMAX ) DXRATMAX = DXRAT
595 IF ( X_STATE .GT. WORKING_STATE ) FINAL_DX_X = DX_X
598 IF ( Z_STATE .EQ. UNSTABLE_STATE .AND. DZ_Z .LE. DZ_UB )
599 $ Z_STATE = WORKING_STATE
600 IF ( Z_STATE .EQ. NOPROG_STATE .AND. DZRAT .LE. RTHRESH )
601 $ Z_STATE = WORKING_STATE
602 IF ( Z_STATE .EQ. WORKING_STATE ) THEN
603 IF ( DZ_Z .LE. EPS ) THEN
605 ELSE IF ( DZ_Z .GT. DZ_UB ) THEN
606 Z_STATE = UNSTABLE_STATE
609 ELSE IF ( DZRAT .GT. RTHRESH ) THEN
610 IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
613 Z_STATE = NOPROG_STATE
616 IF ( DZRAT .GT. DZRATMAX ) DZRATMAX = DZRAT
618 IF ( Z_STATE .GT. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
621 * Exit if both normwise and componentwise stopped working,
622 * but if componentwise is unstable, let it go at least two
625 IF ( X_STATE.NE.WORKING_STATE ) THEN
626 IF ( IGNORE_CWISE ) GOTO 666
627 IF ( Z_STATE.EQ.NOPROG_STATE .OR. Z_STATE.EQ.CONV_STATE )
629 IF ( Z_STATE.EQ.UNSTABLE_STATE .AND. CNT.GT.1 ) GOTO 666
632 IF ( INCR_PREC ) THEN
634 Y_PREC_STATE = Y_PREC_STATE + 1
645 IF ( Y_PREC_STATE .LT. EXTRA_Y ) THEN
646 CALL CAXPY( N, (1.0E+0,0.0E+0), DY, 1, Y(1,J), 1 )
648 CALL CLA_WWADDW( N, Y( 1, J ), Y_TAIL, DY )
652 * Target of "IF (Z_STOP .AND. X_STOP)". Sun's f77 won't EXIT.
655 * Set final_* when cnt hits ithresh
657 IF ( X_STATE .EQ. WORKING_STATE ) FINAL_DX_X = DX_X
658 IF ( Z_STATE .EQ. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
660 * Compute error bounds
662 IF (N_NORMS .GE. 1) THEN
663 ERRS_N( J, LA_LINRX_ERR_I ) = FINAL_DX_X / (1 - DXRATMAX)
666 IF ( N_NORMS .GE. 2 ) THEN
667 ERRS_C( J, LA_LINRX_ERR_I ) = FINAL_DZ_Z / (1 - DZRATMAX)
670 * Compute componentwise relative backward error from formula
671 * max(i) ( abs(R(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
672 * where abs(Z) is the componentwise absolute value of the matrix
675 * Compute residual RES = B_s - op(A_s) * Y,
676 * op(A) = A, A**T, or A**H depending on TRANS (and type).
678 CALL CCOPY( N, B( 1, J ), 1, RES, 1 )
679 CALL CGEMV( TRANS, N, N, (-1.0E+0,0.0E+0), A, LDA, Y(1,J), 1,
680 $ (1.0E+0,0.0E+0), RES, 1 )
683 AYB( I ) = CABS1( B( I, J ) )
686 * Compute abs(op(A_s))*abs(Y) + abs(B_s).
688 CALL CLA_GEAMV ( TRANS_TYPE, N, N, 1.0E+0,
689 $ A, LDA, Y(1, J), 1, 1.0E+0, AYB, 1 )
691 CALL CLA_LIN_BERR ( N, N, 1, RES, AYB, BERR_OUT( J ) )
693 * End of loop for each RHS.