Annotation of rpl/lapack/lapack/dla_gerfsx_extended.f, revision 1.17
1.8 bertrand 1: *> \brief \b DLA_GERFSX_EXTENDED improves the computed solution to a system of linear equations for general matrices by performing extra-precise iterative refinement and provides error bounds and backward error estimates for the solution.
1.5 bertrand 2: *
3: * =========== DOCUMENTATION ===========
4: *
1.12 bertrand 5: * Online html documentation available at
6: * http://www.netlib.org/lapack/explore-html/
1.5 bertrand 7: *
8: *> \htmlonly
1.12 bertrand 9: *> Download DLA_GERFSX_EXTENDED + dependencies
10: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dla_gerfsx_extended.f">
11: *> [TGZ]</a>
12: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dla_gerfsx_extended.f">
13: *> [ZIP]</a>
14: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dla_gerfsx_extended.f">
1.5 bertrand 15: *> [TXT]</a>
1.12 bertrand 16: *> \endhtmlonly
1.5 bertrand 17: *
18: * Definition:
19: * ===========
20: *
21: * SUBROUTINE DLA_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 )
1.12 bertrand 27: *
1.5 bertrand 28: * .. Scalar Arguments ..
29: * INTEGER INFO, LDA, LDAF, LDB, LDY, N, NRHS, PREC_TYPE,
30: * $ TRANS_TYPE, N_NORMS, ITHRESH
31: * LOGICAL COLEQU, IGNORE_CWISE
32: * DOUBLE PRECISION RTHRESH, DZ_UB
33: * ..
34: * .. Array Arguments ..
35: * INTEGER IPIV( * )
36: * DOUBLE PRECISION A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
37: * $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * )
38: * DOUBLE PRECISION C( * ), AYB( * ), RCOND, BERR_OUT( * ),
39: * $ ERRS_N( NRHS, * ), ERRS_C( NRHS, * )
40: * ..
1.12 bertrand 41: *
1.5 bertrand 42: *
43: *> \par Purpose:
44: * =============
45: *>
46: *> \verbatim
47: *>
1.12 bertrand 48: *>
1.5 bertrand 49: *> DLA_GERFSX_EXTENDED improves the computed solution to a system of
50: *> linear equations by performing extra-precise iterative refinement
51: *> and provides error bounds and backward error estimates for the solution.
52: *> This subroutine is called by DGERFSX to perform iterative refinement.
53: *> In addition to normwise error bound, the code provides maximum
54: *> componentwise error bound if possible. See comments for ERRS_N
55: *> and ERRS_C for details of the error bounds. Note that this
1.17 ! bertrand 56: *> subroutine is only responsible for setting the second fields of
1.5 bertrand 57: *> ERRS_N and ERRS_C.
58: *> \endverbatim
59: *
60: * Arguments:
61: * ==========
62: *
63: *> \param[in] PREC_TYPE
64: *> \verbatim
65: *> PREC_TYPE is INTEGER
66: *> Specifies the intermediate precision to be used in refinement.
1.16 bertrand 67: *> The value is defined by ILAPREC(P) where P is a CHARACTER and P
68: *> = 'S': Single
1.5 bertrand 69: *> = 'D': Double
70: *> = 'I': Indigenous
1.16 bertrand 71: *> = 'X' or 'E': Extra
1.5 bertrand 72: *> \endverbatim
73: *>
74: *> \param[in] TRANS_TYPE
75: *> \verbatim
76: *> TRANS_TYPE is INTEGER
77: *> Specifies the transposition operation on A.
1.16 bertrand 78: *> The value is defined by ILATRANS(T) where T is a CHARACTER and T
79: *> = 'N': No transpose
1.5 bertrand 80: *> = 'T': Transpose
81: *> = 'C': Conjugate transpose
82: *> \endverbatim
83: *>
84: *> \param[in] N
85: *> \verbatim
86: *> N is INTEGER
87: *> The number of linear equations, i.e., the order of the
88: *> matrix A. N >= 0.
89: *> \endverbatim
90: *>
91: *> \param[in] NRHS
92: *> \verbatim
93: *> NRHS is INTEGER
94: *> The number of right-hand-sides, i.e., the number of columns of the
95: *> matrix B.
96: *> \endverbatim
97: *>
98: *> \param[in] A
99: *> \verbatim
100: *> A is DOUBLE PRECISION array, dimension (LDA,N)
101: *> On entry, the N-by-N matrix A.
102: *> \endverbatim
103: *>
104: *> \param[in] LDA
105: *> \verbatim
106: *> LDA is INTEGER
107: *> The leading dimension of the array A. LDA >= max(1,N).
108: *> \endverbatim
109: *>
110: *> \param[in] AF
111: *> \verbatim
112: *> AF is DOUBLE PRECISION array, dimension (LDAF,N)
113: *> The factors L and U from the factorization
114: *> A = P*L*U as computed by DGETRF.
115: *> \endverbatim
116: *>
117: *> \param[in] LDAF
118: *> \verbatim
119: *> LDAF is INTEGER
120: *> The leading dimension of the array AF. LDAF >= max(1,N).
121: *> \endverbatim
122: *>
123: *> \param[in] IPIV
124: *> \verbatim
125: *> IPIV is INTEGER array, dimension (N)
126: *> The pivot indices from the factorization A = P*L*U
127: *> as computed by DGETRF; row i of the matrix was interchanged
128: *> with row IPIV(i).
129: *> \endverbatim
130: *>
131: *> \param[in] COLEQU
132: *> \verbatim
133: *> COLEQU is LOGICAL
134: *> If .TRUE. then column equilibration was done to A before calling
135: *> this routine. This is needed to compute the solution and error
136: *> bounds correctly.
137: *> \endverbatim
138: *>
139: *> \param[in] C
140: *> \verbatim
141: *> C is DOUBLE PRECISION array, dimension (N)
142: *> The column scale factors for A. If COLEQU = .FALSE., C
143: *> is not accessed. If C is input, each element of C should be a power
144: *> of the radix to ensure a reliable solution and error estimates.
145: *> Scaling by powers of the radix does not cause rounding errors unless
146: *> the result underflows or overflows. Rounding errors during scaling
147: *> lead to refining with a matrix that is not equivalent to the
148: *> input matrix, producing error estimates that may not be
149: *> reliable.
150: *> \endverbatim
151: *>
152: *> \param[in] B
153: *> \verbatim
154: *> B is DOUBLE PRECISION array, dimension (LDB,NRHS)
155: *> The right-hand-side matrix B.
156: *> \endverbatim
157: *>
158: *> \param[in] LDB
159: *> \verbatim
160: *> LDB is INTEGER
161: *> The leading dimension of the array B. LDB >= max(1,N).
162: *> \endverbatim
163: *>
164: *> \param[in,out] Y
165: *> \verbatim
1.14 bertrand 166: *> Y is DOUBLE PRECISION array, dimension (LDY,NRHS)
1.5 bertrand 167: *> On entry, the solution matrix X, as computed by DGETRS.
168: *> On exit, the improved solution matrix Y.
169: *> \endverbatim
170: *>
171: *> \param[in] LDY
172: *> \verbatim
173: *> LDY is INTEGER
174: *> The leading dimension of the array Y. LDY >= max(1,N).
175: *> \endverbatim
176: *>
177: *> \param[out] BERR_OUT
178: *> \verbatim
179: *> BERR_OUT is DOUBLE PRECISION array, dimension (NRHS)
180: *> On exit, BERR_OUT(j) contains the componentwise relative backward
181: *> error for right-hand-side j from the formula
182: *> max(i) ( abs(RES(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
183: *> where abs(Z) is the componentwise absolute value of the matrix
184: *> or vector Z. This is computed by DLA_LIN_BERR.
185: *> \endverbatim
186: *>
187: *> \param[in] N_NORMS
188: *> \verbatim
189: *> N_NORMS is INTEGER
190: *> Determines which error bounds to return (see ERRS_N
191: *> and ERRS_C).
192: *> If N_NORMS >= 1 return normwise error bounds.
193: *> If N_NORMS >= 2 return componentwise error bounds.
194: *> \endverbatim
195: *>
196: *> \param[in,out] ERRS_N
197: *> \verbatim
1.14 bertrand 198: *> ERRS_N is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS)
1.5 bertrand 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:
202: *>
203: *> Normwise relative error in the ith solution vector:
204: *> max_j (abs(XTRUE(j,i) - X(j,i)))
205: *> ------------------------------
206: *> max_j abs(X(j,i))
207: *>
208: *> The array is indexed by the type of error information as described
209: *> below. There currently are up to three pieces of information
210: *> returned.
211: *>
212: *> The first index in ERRS_N(i,:) corresponds to the ith
213: *> right-hand side.
214: *>
215: *> The second index in ERRS_N(:,err) contains the following
216: *> three fields:
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').
220: *>
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.
226: *>
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.
235: *>
236: *> This subroutine is only responsible for setting the second field
237: *> above.
238: *> See Lapack Working Note 165 for further details and extra
239: *> cautions.
240: *> \endverbatim
241: *>
242: *> \param[in,out] ERRS_C
243: *> \verbatim
1.14 bertrand 244: *> ERRS_C is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS)
1.5 bertrand 245: *> For each right-hand side, this array contains information about
246: *> various error bounds and condition numbers corresponding to the
247: *> componentwise relative error, which is defined as follows:
248: *>
249: *> Componentwise relative error in the ith solution vector:
250: *> abs(XTRUE(j,i) - X(j,i))
251: *> max_j ----------------------
252: *> abs(X(j,i))
253: *>
254: *> The array is indexed by the right-hand side i (on which the
255: *> componentwise relative error depends), and the type of error
256: *> information as described below. There currently are up to three
257: *> pieces of information returned for each right-hand side. If
258: *> componentwise accuracy is not requested (PARAMS(3) = 0.0), then
1.16 bertrand 259: *> ERRS_C is not accessed. If N_ERR_BNDS < 3, then at most
1.5 bertrand 260: *> the first (:,N_ERR_BNDS) entries are returned.
261: *>
262: *> The first index in ERRS_C(i,:) corresponds to the ith
263: *> right-hand side.
264: *>
265: *> The second index in ERRS_C(:,err) contains the following
266: *> three fields:
267: *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
268: *> reciprocal condition number is less than the threshold
269: *> sqrt(n) * slamch('Epsilon').
270: *>
271: *> err = 2 "Guaranteed" error bound: The estimated forward error,
272: *> almost certainly within a factor of 10 of the true error
273: *> so long as the next entry is greater than the threshold
274: *> sqrt(n) * slamch('Epsilon'). This error bound should only
275: *> be trusted if the previous boolean is true.
276: *>
277: *> err = 3 Reciprocal condition number: Estimated componentwise
278: *> reciprocal condition number. Compared with the threshold
279: *> sqrt(n) * slamch('Epsilon') to determine if the error
280: *> estimate is "guaranteed". These reciprocal condition
281: *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
282: *> appropriately scaled matrix Z.
283: *> Let Z = S*(A*diag(x)), where x is the solution for the
284: *> current right-hand side and S scales each row of
285: *> A*diag(x) by a power of the radix so all absolute row
286: *> sums of Z are approximately 1.
287: *>
288: *> This subroutine is only responsible for setting the second field
289: *> above.
290: *> See Lapack Working Note 165 for further details and extra
291: *> cautions.
292: *> \endverbatim
293: *>
294: *> \param[in] RES
295: *> \verbatim
296: *> RES is DOUBLE PRECISION array, dimension (N)
297: *> Workspace to hold the intermediate residual.
298: *> \endverbatim
299: *>
300: *> \param[in] AYB
301: *> \verbatim
302: *> AYB is DOUBLE PRECISION array, dimension (N)
303: *> Workspace. This can be the same workspace passed for Y_TAIL.
304: *> \endverbatim
305: *>
306: *> \param[in] DY
307: *> \verbatim
308: *> DY is DOUBLE PRECISION array, dimension (N)
309: *> Workspace to hold the intermediate solution.
310: *> \endverbatim
311: *>
312: *> \param[in] Y_TAIL
313: *> \verbatim
314: *> Y_TAIL is DOUBLE PRECISION array, dimension (N)
315: *> Workspace to hold the trailing bits of the intermediate solution.
316: *> \endverbatim
317: *>
318: *> \param[in] RCOND
319: *> \verbatim
320: *> RCOND is DOUBLE PRECISION
321: *> Reciprocal scaled condition number. This is an estimate of the
322: *> reciprocal Skeel condition number of the matrix A after
323: *> equilibration (if done). If this is less than the machine
324: *> precision (in particular, if it is zero), the matrix is singular
325: *> to working precision. Note that the error may still be small even
326: *> if this number is very small and the matrix appears ill-
327: *> conditioned.
328: *> \endverbatim
329: *>
330: *> \param[in] ITHRESH
331: *> \verbatim
332: *> ITHRESH is INTEGER
333: *> The maximum number of residual computations allowed for
334: *> refinement. The default is 10. For 'aggressive' set to 100 to
335: *> permit convergence using approximate factorizations or
336: *> factorizations other than LU. If the factorization uses a
337: *> technique other than Gaussian elimination, the guarantees in
338: *> ERRS_N and ERRS_C may no longer be trustworthy.
339: *> \endverbatim
340: *>
341: *> \param[in] RTHRESH
342: *> \verbatim
343: *> RTHRESH is DOUBLE PRECISION
344: *> Determines when to stop refinement if the error estimate stops
345: *> decreasing. Refinement will stop when the next solution no longer
346: *> satisfies norm(dx_{i+1}) < RTHRESH * norm(dx_i) where norm(Z) is
347: *> the infinity norm of Z. RTHRESH satisfies 0 < RTHRESH <= 1. The
348: *> default value is 0.5. For 'aggressive' set to 0.9 to permit
349: *> convergence on extremely ill-conditioned matrices. See LAWN 165
350: *> for more details.
351: *> \endverbatim
352: *>
353: *> \param[in] DZ_UB
354: *> \verbatim
355: *> DZ_UB is DOUBLE PRECISION
356: *> Determines when to start considering componentwise convergence.
357: *> Componentwise convergence is only considered after each component
1.17 ! bertrand 358: *> of the solution Y is stable, which we define as the relative
1.5 bertrand 359: *> change in each component being less than DZ_UB. The default value
360: *> is 0.25, requiring the first bit to be stable. See LAWN 165 for
361: *> more details.
362: *> \endverbatim
363: *>
364: *> \param[in] IGNORE_CWISE
365: *> \verbatim
366: *> IGNORE_CWISE is LOGICAL
367: *> If .TRUE. then ignore componentwise convergence. Default value
368: *> is .FALSE..
369: *> \endverbatim
370: *>
371: *> \param[out] INFO
372: *> \verbatim
373: *> INFO is INTEGER
374: *> = 0: Successful exit.
375: *> < 0: if INFO = -i, the ith argument to DGETRS had an illegal
376: *> value
377: *> \endverbatim
378: *
379: * Authors:
380: * ========
381: *
1.12 bertrand 382: *> \author Univ. of Tennessee
383: *> \author Univ. of California Berkeley
384: *> \author Univ. of Colorado Denver
385: *> \author NAG Ltd.
1.5 bertrand 386: *
387: *> \ingroup doubleGEcomputational
388: *
389: * =====================================================================
1.1 bertrand 390: SUBROUTINE DLA_GERFSX_EXTENDED( PREC_TYPE, TRANS_TYPE, N, NRHS, A,
391: $ LDA, AF, LDAF, IPIV, COLEQU, C, B,
392: $ LDB, Y, LDY, BERR_OUT, N_NORMS,
393: $ ERRS_N, ERRS_C, RES, AYB, DY,
394: $ Y_TAIL, RCOND, ITHRESH, RTHRESH,
395: $ DZ_UB, IGNORE_CWISE, INFO )
396: *
1.17 ! bertrand 397: * -- LAPACK computational routine --
1.5 bertrand 398: * -- LAPACK is a software package provided by Univ. of Tennessee, --
399: * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
1.1 bertrand 400: *
401: * .. Scalar Arguments ..
402: INTEGER INFO, LDA, LDAF, LDB, LDY, N, NRHS, PREC_TYPE,
403: $ TRANS_TYPE, N_NORMS, ITHRESH
404: LOGICAL COLEQU, IGNORE_CWISE
405: DOUBLE PRECISION RTHRESH, DZ_UB
406: * ..
407: * .. Array Arguments ..
408: INTEGER IPIV( * )
409: DOUBLE PRECISION A( LDA, * ), AF( LDAF, * ), B( LDB, * ),
410: $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * )
411: DOUBLE PRECISION C( * ), AYB( * ), RCOND, BERR_OUT( * ),
412: $ ERRS_N( NRHS, * ), ERRS_C( NRHS, * )
413: * ..
414: *
415: * =====================================================================
416: *
417: * .. Local Scalars ..
418: CHARACTER TRANS
419: INTEGER CNT, I, J, X_STATE, Z_STATE, Y_PREC_STATE
420: DOUBLE PRECISION YK, DYK, YMIN, NORMY, NORMX, NORMDX, DXRAT,
421: $ DZRAT, PREVNORMDX, PREV_DZ_Z, DXRATMAX,
422: $ DZRATMAX, DX_X, DZ_Z, FINAL_DX_X, FINAL_DZ_Z,
423: $ EPS, HUGEVAL, INCR_THRESH
424: LOGICAL INCR_PREC
425: * ..
426: * .. Parameters ..
427: INTEGER UNSTABLE_STATE, WORKING_STATE, CONV_STATE,
428: $ NOPROG_STATE, BASE_RESIDUAL, EXTRA_RESIDUAL,
429: $ EXTRA_Y
430: PARAMETER ( UNSTABLE_STATE = 0, WORKING_STATE = 1,
431: $ CONV_STATE = 2, NOPROG_STATE = 3 )
432: PARAMETER ( BASE_RESIDUAL = 0, EXTRA_RESIDUAL = 1,
433: $ EXTRA_Y = 2 )
434: INTEGER FINAL_NRM_ERR_I, FINAL_CMP_ERR_I, BERR_I
435: INTEGER RCOND_I, NRM_RCOND_I, NRM_ERR_I, CMP_RCOND_I
436: INTEGER CMP_ERR_I, PIV_GROWTH_I
437: PARAMETER ( FINAL_NRM_ERR_I = 1, FINAL_CMP_ERR_I = 2,
438: $ BERR_I = 3 )
439: PARAMETER ( RCOND_I = 4, NRM_RCOND_I = 5, NRM_ERR_I = 6 )
440: PARAMETER ( CMP_RCOND_I = 7, CMP_ERR_I = 8,
441: $ PIV_GROWTH_I = 9 )
442: INTEGER LA_LINRX_ITREF_I, LA_LINRX_ITHRESH_I,
443: $ LA_LINRX_CWISE_I
444: PARAMETER ( LA_LINRX_ITREF_I = 1,
445: $ LA_LINRX_ITHRESH_I = 2 )
446: PARAMETER ( LA_LINRX_CWISE_I = 3 )
447: INTEGER LA_LINRX_TRUST_I, LA_LINRX_ERR_I,
448: $ LA_LINRX_RCOND_I
449: PARAMETER ( LA_LINRX_TRUST_I = 1, LA_LINRX_ERR_I = 2 )
450: PARAMETER ( LA_LINRX_RCOND_I = 3 )
451: * ..
452: * .. External Subroutines ..
453: EXTERNAL DAXPY, DCOPY, DGETRS, DGEMV, BLAS_DGEMV_X,
454: $ BLAS_DGEMV2_X, DLA_GEAMV, DLA_WWADDW, DLAMCH,
455: $ CHLA_TRANSTYPE, DLA_LIN_BERR
456: DOUBLE PRECISION DLAMCH
457: CHARACTER CHLA_TRANSTYPE
458: * ..
459: * .. Intrinsic Functions ..
460: INTRINSIC ABS, MAX, MIN
461: * ..
462: * .. Executable Statements ..
463: *
464: IF ( INFO.NE.0 ) RETURN
465: TRANS = CHLA_TRANSTYPE(TRANS_TYPE)
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
472: *
473: DO J = 1, NRHS
474: Y_PREC_STATE = EXTRA_RESIDUAL
475: IF ( Y_PREC_STATE .EQ. EXTRA_Y ) THEN
476: DO I = 1, N
477: Y_TAIL( I ) = 0.0D+0
478: END DO
479: END IF
480:
481: DXRAT = 0.0D+0
482: DXRATMAX = 0.0D+0
483: DZRAT = 0.0D+0
484: DZRATMAX = 0.0D+0
485: FINAL_DX_X = HUGEVAL
486: FINAL_DZ_Z = HUGEVAL
487: PREVNORMDX = HUGEVAL
488: PREV_DZ_Z = HUGEVAL
489: DZ_Z = HUGEVAL
490: DX_X = HUGEVAL
491:
492: X_STATE = WORKING_STATE
493: Z_STATE = UNSTABLE_STATE
494: INCR_PREC = .FALSE.
495:
496: DO CNT = 1, ITHRESH
497: *
498: * Compute residual RES = B_s - op(A_s) * Y,
499: * op(A) = A, A**T, or A**H depending on TRANS (and type).
500: *
501: CALL DCOPY( N, B( 1, J ), 1, RES, 1 )
502: IF ( Y_PREC_STATE .EQ. BASE_RESIDUAL ) THEN
503: CALL DGEMV( TRANS, N, N, -1.0D+0, A, LDA, Y( 1, J ), 1,
504: $ 1.0D+0, RES, 1 )
505: ELSE IF ( Y_PREC_STATE .EQ. EXTRA_RESIDUAL ) THEN
506: CALL BLAS_DGEMV_X( TRANS_TYPE, N, N, -1.0D+0, A, LDA,
507: $ Y( 1, J ), 1, 1.0D+0, RES, 1, PREC_TYPE )
508: ELSE
509: CALL BLAS_DGEMV2_X( TRANS_TYPE, N, N, -1.0D+0, A, LDA,
510: $ Y( 1, J ), Y_TAIL, 1, 1.0D+0, RES, 1, PREC_TYPE )
511: END IF
512:
513: ! XXX: RES is no longer needed.
514: CALL DCOPY( N, RES, 1, DY, 1 )
515: CALL DGETRS( TRANS, N, 1, AF, LDAF, IPIV, DY, N, INFO )
516: *
517: * Calculate relative changes DX_X, DZ_Z and ratios DXRAT, DZRAT.
518: *
519: NORMX = 0.0D+0
520: NORMY = 0.0D+0
521: NORMDX = 0.0D+0
522: DZ_Z = 0.0D+0
523: YMIN = HUGEVAL
524: *
525: DO I = 1, N
526: YK = ABS( Y( I, J ) )
527: DYK = ABS( DY( I ) )
528:
529: IF ( YK .NE. 0.0D+0 ) THEN
530: DZ_Z = MAX( DZ_Z, DYK / YK )
531: ELSE IF ( DYK .NE. 0.0D+0 ) THEN
532: DZ_Z = HUGEVAL
533: END IF
534:
535: YMIN = MIN( YMIN, YK )
536:
537: NORMY = MAX( NORMY, YK )
538:
539: IF ( COLEQU ) THEN
540: NORMX = MAX( NORMX, YK * C( I ) )
541: NORMDX = MAX( NORMDX, DYK * C( I ) )
542: ELSE
543: NORMX = NORMY
544: NORMDX = MAX( NORMDX, DYK )
545: END IF
546: END DO
547:
548: IF ( NORMX .NE. 0.0D+0 ) THEN
549: DX_X = NORMDX / NORMX
550: ELSE IF ( NORMDX .EQ. 0.0D+0 ) THEN
551: DX_X = 0.0D+0
552: ELSE
553: DX_X = HUGEVAL
554: END IF
555:
556: DXRAT = NORMDX / PREVNORMDX
557: DZRAT = DZ_Z / PREV_DZ_Z
558: *
559: * Check termination criteria
560: *
561: IF (.NOT.IGNORE_CWISE
562: $ .AND. YMIN*RCOND .LT. INCR_THRESH*NORMY
563: $ .AND. Y_PREC_STATE .LT. EXTRA_Y)
564: $ INCR_PREC = .TRUE.
565:
566: IF ( X_STATE .EQ. NOPROG_STATE .AND. DXRAT .LE. RTHRESH )
567: $ X_STATE = WORKING_STATE
568: IF ( X_STATE .EQ. WORKING_STATE ) THEN
569: IF ( DX_X .LE. EPS ) THEN
570: X_STATE = CONV_STATE
571: ELSE IF ( DXRAT .GT. RTHRESH ) THEN
572: IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
573: INCR_PREC = .TRUE.
574: ELSE
575: X_STATE = NOPROG_STATE
576: END IF
577: ELSE
578: IF ( DXRAT .GT. DXRATMAX ) DXRATMAX = DXRAT
579: END IF
580: IF ( X_STATE .GT. WORKING_STATE ) FINAL_DX_X = DX_X
581: END IF
582:
583: IF ( Z_STATE .EQ. UNSTABLE_STATE .AND. DZ_Z .LE. DZ_UB )
584: $ Z_STATE = WORKING_STATE
585: IF ( Z_STATE .EQ. NOPROG_STATE .AND. DZRAT .LE. RTHRESH )
586: $ Z_STATE = WORKING_STATE
587: IF ( Z_STATE .EQ. WORKING_STATE ) THEN
588: IF ( DZ_Z .LE. EPS ) THEN
589: Z_STATE = CONV_STATE
590: ELSE IF ( DZ_Z .GT. DZ_UB ) THEN
591: Z_STATE = UNSTABLE_STATE
592: DZRATMAX = 0.0D+0
593: FINAL_DZ_Z = HUGEVAL
594: ELSE IF ( DZRAT .GT. RTHRESH ) THEN
595: IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
596: INCR_PREC = .TRUE.
597: ELSE
598: Z_STATE = NOPROG_STATE
599: END IF
600: ELSE
601: IF ( DZRAT .GT. DZRATMAX ) DZRATMAX = DZRAT
602: END IF
603: IF ( Z_STATE .GT. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
604: END IF
605: *
606: * Exit if both normwise and componentwise stopped working,
607: * but if componentwise is unstable, let it go at least two
608: * iterations.
609: *
610: IF ( X_STATE.NE.WORKING_STATE ) THEN
611: IF ( IGNORE_CWISE) GOTO 666
612: IF ( Z_STATE.EQ.NOPROG_STATE .OR. Z_STATE.EQ.CONV_STATE )
613: $ GOTO 666
614: IF ( Z_STATE.EQ.UNSTABLE_STATE .AND. CNT.GT.1 ) GOTO 666
615: END IF
616:
617: IF ( INCR_PREC ) THEN
618: INCR_PREC = .FALSE.
619: Y_PREC_STATE = Y_PREC_STATE + 1
620: DO I = 1, N
621: Y_TAIL( I ) = 0.0D+0
622: END DO
623: END IF
624:
625: PREVNORMDX = NORMDX
626: PREV_DZ_Z = DZ_Z
627: *
628: * Update soluton.
629: *
630: IF ( Y_PREC_STATE .LT. EXTRA_Y ) THEN
631: CALL DAXPY( N, 1.0D+0, DY, 1, Y( 1, J ), 1 )
632: ELSE
633: CALL DLA_WWADDW( N, Y( 1, J ), Y_TAIL, DY )
634: END IF
635:
636: END DO
637: * Target of "IF (Z_STOP .AND. X_STOP)". Sun's f77 won't EXIT.
638: 666 CONTINUE
639: *
640: * Set final_* when cnt hits ithresh.
641: *
642: IF ( X_STATE .EQ. WORKING_STATE ) FINAL_DX_X = DX_X
643: IF ( Z_STATE .EQ. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
644: *
645: * Compute error bounds
646: *
647: IF (N_NORMS .GE. 1) THEN
648: ERRS_N( J, LA_LINRX_ERR_I ) = FINAL_DX_X / (1 - DXRATMAX)
649: END IF
650: IF ( N_NORMS .GE. 2 ) THEN
651: ERRS_C( J, LA_LINRX_ERR_I ) = FINAL_DZ_Z / (1 - DZRATMAX)
652: END IF
653: *
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
657: * or vector Z.
658: *
659: * Compute residual RES = B_s - op(A_s) * Y,
660: * op(A) = A, A**T, or A**H depending on TRANS (and type).
661: *
662: CALL DCOPY( N, B( 1, J ), 1, RES, 1 )
1.12 bertrand 663: CALL DGEMV( TRANS, N, N, -1.0D+0, A, LDA, Y(1,J), 1, 1.0D+0,
1.1 bertrand 664: $ RES, 1 )
665:
666: DO I = 1, N
667: AYB( I ) = ABS( B( I, J ) )
668: END DO
669: *
670: * Compute abs(op(A_s))*abs(Y) + abs(B_s).
671: *
672: CALL DLA_GEAMV ( TRANS_TYPE, N, N, 1.0D+0,
673: $ A, LDA, Y(1, J), 1, 1.0D+0, AYB, 1 )
674:
675: CALL DLA_LIN_BERR ( N, N, 1, RES, AYB, BERR_OUT( J ) )
676: *
677: * End of loop for each RHS.
678: *
679: END DO
680: *
681: RETURN
1.17 ! bertrand 682: *
! 683: * End of DLA_GERFSX_EXTENDED
! 684: *
1.1 bertrand 685: END
CVSweb interface <joel.bertrand@systella.fr>