Annotation of rpl/lapack/lapack/dla_gbrfsx_extended.f, revision 1.17
1.8 bertrand 1: *> \brief \b DLA_GBRFSX_EXTENDED improves the computed solution to a system of linear equations for general banded 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_GBRFSX_EXTENDED + dependencies
10: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dla_gbrfsx_extended.f">
11: *> [TGZ]</a>
12: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dla_gbrfsx_extended.f">
13: *> [ZIP]</a>
14: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dla_gbrfsx_extended.f">
1.5 bertrand 15: *> [TXT]</a>
1.12 bertrand 16: *> \endhtmlonly
1.5 bertrand 17: *
18: * Definition:
19: * ===========
20: *
21: * SUBROUTINE DLA_GBRFSX_EXTENDED( PREC_TYPE, TRANS_TYPE, N, KL, KU,
22: * NRHS, AB, LDAB, AFB, LDAFB, IPIV,
23: * COLEQU, C, B, LDB, Y, LDY,
24: * BERR_OUT, N_NORMS, ERR_BNDS_NORM,
25: * ERR_BNDS_COMP, RES, AYB, DY,
26: * Y_TAIL, RCOND, ITHRESH, RTHRESH,
27: * DZ_UB, IGNORE_CWISE, INFO )
1.12 bertrand 28: *
1.5 bertrand 29: * .. Scalar Arguments ..
30: * INTEGER INFO, LDAB, LDAFB, LDB, LDY, N, KL, KU, NRHS,
31: * $ PREC_TYPE, TRANS_TYPE, N_NORMS, ITHRESH
32: * LOGICAL COLEQU, IGNORE_CWISE
33: * DOUBLE PRECISION RTHRESH, DZ_UB
34: * ..
35: * .. Array Arguments ..
36: * INTEGER IPIV( * )
37: * DOUBLE PRECISION AB( LDAB, * ), AFB( LDAFB, * ), 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, * )
42: * ..
1.12 bertrand 43: *
1.5 bertrand 44: *
45: *> \par Purpose:
46: * =============
47: *>
48: *> \verbatim
49: *>
1.12 bertrand 50: *>
1.5 bertrand 51: *> DLA_GBRFSX_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 DGBRFSX 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
1.17 ! bertrand 58: *> subroutine is only responsible for setting the second fields of
1.5 bertrand 59: *> ERR_BNDS_NORM and ERR_BNDS_COMP.
60: *> \endverbatim
61: *
62: * Arguments:
63: * ==========
64: *
65: *> \param[in] PREC_TYPE
66: *> \verbatim
67: *> PREC_TYPE is INTEGER
68: *> Specifies the intermediate precision to be used in refinement.
1.16 bertrand 69: *> The value is defined by ILAPREC(P) where P is a CHARACTER and P
70: *> = 'S': Single
1.5 bertrand 71: *> = 'D': Double
72: *> = 'I': Indigenous
1.16 bertrand 73: *> = 'X' or 'E': Extra
1.5 bertrand 74: *> \endverbatim
75: *>
76: *> \param[in] TRANS_TYPE
77: *> \verbatim
78: *> TRANS_TYPE is INTEGER
79: *> Specifies the transposition operation on A.
1.16 bertrand 80: *> The value is defined by ILATRANS(T) where T is a CHARACTER and T
81: *> = 'N': No transpose
1.5 bertrand 82: *> = 'T': Transpose
83: *> = 'C': Conjugate transpose
84: *> \endverbatim
85: *>
86: *> \param[in] N
87: *> \verbatim
88: *> N is INTEGER
89: *> The number of linear equations, i.e., the order of the
90: *> matrix A. N >= 0.
91: *> \endverbatim
92: *>
93: *> \param[in] KL
94: *> \verbatim
95: *> KL is INTEGER
96: *> The number of subdiagonals within the band of A. KL >= 0.
97: *> \endverbatim
98: *>
99: *> \param[in] KU
100: *> \verbatim
101: *> KU is INTEGER
102: *> The number of superdiagonals within the band of A. KU >= 0
103: *> \endverbatim
104: *>
105: *> \param[in] NRHS
106: *> \verbatim
107: *> NRHS is INTEGER
108: *> The number of right-hand-sides, i.e., the number of columns of the
109: *> matrix B.
110: *> \endverbatim
111: *>
112: *> \param[in] AB
113: *> \verbatim
114: *> AB is DOUBLE PRECISION array, dimension (LDAB,N)
115: *> On entry, the N-by-N matrix AB.
116: *> \endverbatim
117: *>
118: *> \param[in] LDAB
119: *> \verbatim
120: *> LDAB is INTEGER
121: *> The leading dimension of the array AB. LDBA >= max(1,N).
122: *> \endverbatim
123: *>
124: *> \param[in] AFB
125: *> \verbatim
126: *> AFB is DOUBLE PRECISION array, dimension (LDAFB,N)
127: *> The factors L and U from the factorization
128: *> A = P*L*U as computed by DGBTRF.
129: *> \endverbatim
130: *>
131: *> \param[in] LDAFB
132: *> \verbatim
133: *> LDAFB is INTEGER
134: *> The leading dimension of the array AF. LDAFB >= max(1,N).
135: *> \endverbatim
136: *>
137: *> \param[in] IPIV
138: *> \verbatim
139: *> IPIV is INTEGER array, dimension (N)
140: *> The pivot indices from the factorization A = P*L*U
141: *> as computed by DGBTRF; row i of the matrix was interchanged
142: *> with row IPIV(i).
143: *> \endverbatim
144: *>
145: *> \param[in] COLEQU
146: *> \verbatim
147: *> COLEQU is LOGICAL
148: *> If .TRUE. then column equilibration was done to A before calling
149: *> this routine. This is needed to compute the solution and error
150: *> bounds correctly.
151: *> \endverbatim
152: *>
153: *> \param[in] C
154: *> \verbatim
155: *> C is DOUBLE PRECISION array, dimension (N)
156: *> The column scale factors for A. If COLEQU = .FALSE., C
157: *> is not accessed. If C is input, each element of C should be a power
158: *> of the radix to ensure a reliable solution and error estimates.
159: *> Scaling by powers of the radix does not cause rounding errors unless
160: *> the result underflows or overflows. Rounding errors during scaling
161: *> lead to refining with a matrix that is not equivalent to the
162: *> input matrix, producing error estimates that may not be
163: *> reliable.
164: *> \endverbatim
165: *>
166: *> \param[in] B
167: *> \verbatim
168: *> B is DOUBLE PRECISION array, dimension (LDB,NRHS)
169: *> The right-hand-side matrix B.
170: *> \endverbatim
171: *>
172: *> \param[in] LDB
173: *> \verbatim
174: *> LDB is INTEGER
175: *> The leading dimension of the array B. LDB >= max(1,N).
176: *> \endverbatim
177: *>
178: *> \param[in,out] Y
179: *> \verbatim
1.14 bertrand 180: *> Y is DOUBLE PRECISION array, dimension (LDY,NRHS)
1.5 bertrand 181: *> On entry, the solution matrix X, as computed by DGBTRS.
182: *> On exit, the improved solution matrix Y.
183: *> \endverbatim
184: *>
185: *> \param[in] LDY
186: *> \verbatim
187: *> LDY is INTEGER
188: *> The leading dimension of the array Y. LDY >= max(1,N).
189: *> \endverbatim
190: *>
191: *> \param[out] BERR_OUT
192: *> \verbatim
193: *> BERR_OUT is DOUBLE PRECISION array, dimension (NRHS)
194: *> On exit, BERR_OUT(j) contains the componentwise relative backward
195: *> error for right-hand-side j from the formula
196: *> max(i) ( abs(RES(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
197: *> where abs(Z) is the componentwise absolute value of the matrix
198: *> or vector Z. This is computed by DLA_LIN_BERR.
199: *> \endverbatim
200: *>
201: *> \param[in] N_NORMS
202: *> \verbatim
203: *> N_NORMS is INTEGER
204: *> Determines which error bounds to return (see ERR_BNDS_NORM
205: *> and ERR_BNDS_COMP).
206: *> If N_NORMS >= 1 return normwise error bounds.
207: *> If N_NORMS >= 2 return componentwise error bounds.
208: *> \endverbatim
209: *>
210: *> \param[in,out] ERR_BNDS_NORM
211: *> \verbatim
1.14 bertrand 212: *> ERR_BNDS_NORM is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS)
1.5 bertrand 213: *> For each right-hand side, this array contains information about
214: *> various error bounds and condition numbers corresponding to the
215: *> normwise relative error, which is defined as follows:
216: *>
217: *> Normwise relative error in the ith solution vector:
218: *> max_j (abs(XTRUE(j,i) - X(j,i)))
219: *> ------------------------------
220: *> max_j abs(X(j,i))
221: *>
222: *> The array is indexed by the type of error information as described
223: *> below. There currently are up to three pieces of information
224: *> returned.
225: *>
226: *> The first index in ERR_BNDS_NORM(i,:) corresponds to the ith
227: *> right-hand side.
228: *>
229: *> The second index in ERR_BNDS_NORM(:,err) contains the following
230: *> three fields:
231: *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
232: *> reciprocal condition number is less than the threshold
233: *> sqrt(n) * slamch('Epsilon').
234: *>
235: *> err = 2 "Guaranteed" error bound: The estimated forward error,
236: *> almost certainly within a factor of 10 of the true error
237: *> so long as the next entry is greater than the threshold
238: *> sqrt(n) * slamch('Epsilon'). This error bound should only
239: *> be trusted if the previous boolean is true.
240: *>
241: *> err = 3 Reciprocal condition number: Estimated normwise
242: *> reciprocal condition number. Compared with the threshold
243: *> sqrt(n) * slamch('Epsilon') to determine if the error
244: *> estimate is "guaranteed". These reciprocal condition
245: *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
246: *> appropriately scaled matrix Z.
247: *> Let Z = S*A, where S scales each row by a power of the
248: *> radix so all absolute row sums of Z are approximately 1.
249: *>
250: *> This subroutine is only responsible for setting the second field
251: *> above.
252: *> See Lapack Working Note 165 for further details and extra
253: *> cautions.
254: *> \endverbatim
255: *>
256: *> \param[in,out] ERR_BNDS_COMP
257: *> \verbatim
1.14 bertrand 258: *> ERR_BNDS_COMP is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS)
1.5 bertrand 259: *> For each right-hand side, this array contains information about
260: *> various error bounds and condition numbers corresponding to the
261: *> componentwise relative error, which is defined as follows:
262: *>
263: *> Componentwise relative error in the ith solution vector:
264: *> abs(XTRUE(j,i) - X(j,i))
265: *> max_j ----------------------
266: *> abs(X(j,i))
267: *>
268: *> The array is indexed by the right-hand side i (on which the
269: *> componentwise relative error depends), and the type of error
270: *> information as described below. There currently are up to three
271: *> pieces of information returned for each right-hand side. If
272: *> componentwise accuracy is not requested (PARAMS(3) = 0.0), then
1.16 bertrand 273: *> ERR_BNDS_COMP is not accessed. If N_ERR_BNDS < 3, then at most
1.5 bertrand 274: *> the first (:,N_ERR_BNDS) entries are returned.
275: *>
276: *> The first index in ERR_BNDS_COMP(i,:) corresponds to the ith
277: *> right-hand side.
278: *>
279: *> The second index in ERR_BNDS_COMP(:,err) contains the following
280: *> three fields:
281: *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
282: *> reciprocal condition number is less than the threshold
283: *> sqrt(n) * slamch('Epsilon').
284: *>
285: *> err = 2 "Guaranteed" error bound: The estimated forward error,
286: *> almost certainly within a factor of 10 of the true error
287: *> so long as the next entry is greater than the threshold
288: *> sqrt(n) * slamch('Epsilon'). This error bound should only
289: *> be trusted if the previous boolean is true.
290: *>
291: *> err = 3 Reciprocal condition number: Estimated componentwise
292: *> reciprocal condition number. Compared with the threshold
293: *> sqrt(n) * slamch('Epsilon') to determine if the error
294: *> estimate is "guaranteed". These reciprocal condition
295: *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
296: *> appropriately scaled matrix Z.
297: *> Let Z = S*(A*diag(x)), where x is the solution for the
298: *> current right-hand side and S scales each row of
299: *> A*diag(x) by a power of the radix so all absolute row
300: *> sums of Z are approximately 1.
301: *>
302: *> This subroutine is only responsible for setting the second field
303: *> above.
304: *> See Lapack Working Note 165 for further details and extra
305: *> cautions.
306: *> \endverbatim
307: *>
308: *> \param[in] RES
309: *> \verbatim
310: *> RES is DOUBLE PRECISION array, dimension (N)
311: *> Workspace to hold the intermediate residual.
312: *> \endverbatim
313: *>
314: *> \param[in] AYB
315: *> \verbatim
316: *> AYB is DOUBLE PRECISION array, dimension (N)
317: *> Workspace. This can be the same workspace passed for Y_TAIL.
318: *> \endverbatim
319: *>
320: *> \param[in] DY
321: *> \verbatim
322: *> DY is DOUBLE PRECISION array, dimension (N)
323: *> Workspace to hold the intermediate solution.
324: *> \endverbatim
325: *>
326: *> \param[in] Y_TAIL
327: *> \verbatim
328: *> Y_TAIL is DOUBLE PRECISION array, dimension (N)
329: *> Workspace to hold the trailing bits of the intermediate solution.
330: *> \endverbatim
331: *>
332: *> \param[in] RCOND
333: *> \verbatim
334: *> RCOND is DOUBLE PRECISION
335: *> Reciprocal scaled condition number. This is an estimate of the
336: *> reciprocal Skeel condition number of the matrix A after
337: *> equilibration (if done). If this is less than the machine
338: *> precision (in particular, if it is zero), the matrix is singular
339: *> to working precision. Note that the error may still be small even
340: *> if this number is very small and the matrix appears ill-
341: *> conditioned.
342: *> \endverbatim
343: *>
344: *> \param[in] ITHRESH
345: *> \verbatim
346: *> ITHRESH is INTEGER
347: *> The maximum number of residual computations allowed for
348: *> refinement. The default is 10. For 'aggressive' set to 100 to
349: *> permit convergence using approximate factorizations or
350: *> factorizations other than LU. If the factorization uses a
351: *> technique other than Gaussian elimination, the guarantees in
352: *> ERR_BNDS_NORM and ERR_BNDS_COMP may no longer be trustworthy.
353: *> \endverbatim
354: *>
355: *> \param[in] RTHRESH
356: *> \verbatim
357: *> RTHRESH is DOUBLE PRECISION
358: *> Determines when to stop refinement if the error estimate stops
359: *> decreasing. Refinement will stop when the next solution no longer
360: *> satisfies norm(dx_{i+1}) < RTHRESH * norm(dx_i) where norm(Z) is
361: *> the infinity norm of Z. RTHRESH satisfies 0 < RTHRESH <= 1. The
362: *> default value is 0.5. For 'aggressive' set to 0.9 to permit
363: *> convergence on extremely ill-conditioned matrices. See LAWN 165
364: *> for more details.
365: *> \endverbatim
366: *>
367: *> \param[in] DZ_UB
368: *> \verbatim
369: *> DZ_UB is DOUBLE PRECISION
370: *> Determines when to start considering componentwise convergence.
371: *> Componentwise convergence is only considered after each component
1.17 ! bertrand 372: *> of the solution Y is stable, which we define as the relative
1.5 bertrand 373: *> change in each component being less than DZ_UB. The default value
374: *> is 0.25, requiring the first bit to be stable. See LAWN 165 for
375: *> more details.
376: *> \endverbatim
377: *>
378: *> \param[in] IGNORE_CWISE
379: *> \verbatim
380: *> IGNORE_CWISE is LOGICAL
381: *> If .TRUE. then ignore componentwise convergence. Default value
382: *> is .FALSE..
383: *> \endverbatim
384: *>
385: *> \param[out] INFO
386: *> \verbatim
387: *> INFO is INTEGER
388: *> = 0: Successful exit.
389: *> < 0: if INFO = -i, the ith argument to DGBTRS had an illegal
390: *> value
391: *> \endverbatim
392: *
393: * Authors:
394: * ========
395: *
1.12 bertrand 396: *> \author Univ. of Tennessee
397: *> \author Univ. of California Berkeley
398: *> \author Univ. of Colorado Denver
399: *> \author NAG Ltd.
1.5 bertrand 400: *
401: *> \ingroup doubleGBcomputational
402: *
403: * =====================================================================
1.1 bertrand 404: SUBROUTINE DLA_GBRFSX_EXTENDED( PREC_TYPE, TRANS_TYPE, N, KL, KU,
405: $ NRHS, AB, LDAB, AFB, LDAFB, IPIV,
406: $ COLEQU, C, B, LDB, Y, LDY,
407: $ BERR_OUT, N_NORMS, ERR_BNDS_NORM,
408: $ ERR_BNDS_COMP, RES, AYB, DY,
409: $ Y_TAIL, RCOND, ITHRESH, RTHRESH,
410: $ DZ_UB, IGNORE_CWISE, INFO )
411: *
1.17 ! bertrand 412: * -- LAPACK computational routine --
1.5 bertrand 413: * -- LAPACK is a software package provided by Univ. of Tennessee, --
414: * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
1.1 bertrand 415: *
416: * .. Scalar Arguments ..
417: INTEGER INFO, LDAB, LDAFB, LDB, LDY, N, KL, KU, NRHS,
418: $ PREC_TYPE, TRANS_TYPE, N_NORMS, ITHRESH
419: LOGICAL COLEQU, IGNORE_CWISE
420: DOUBLE PRECISION RTHRESH, DZ_UB
421: * ..
422: * .. Array Arguments ..
423: INTEGER IPIV( * )
424: DOUBLE PRECISION AB( LDAB, * ), AFB( LDAFB, * ), B( LDB, * ),
425: $ Y( LDY, * ), RES(*), DY(*), Y_TAIL(*)
426: DOUBLE PRECISION C( * ), AYB(*), RCOND, BERR_OUT(*),
427: $ ERR_BNDS_NORM( NRHS, * ),
428: $ ERR_BNDS_COMP( NRHS, * )
429: * ..
430: *
431: * =====================================================================
432: *
433: * .. Local Scalars ..
434: CHARACTER TRANS
435: INTEGER CNT, I, J, M, X_STATE, Z_STATE, Y_PREC_STATE
436: DOUBLE PRECISION YK, DYK, YMIN, NORMY, NORMX, NORMDX, DXRAT,
437: $ DZRAT, PREVNORMDX, PREV_DZ_Z, DXRATMAX,
438: $ DZRATMAX, DX_X, DZ_Z, FINAL_DX_X, FINAL_DZ_Z,
439: $ EPS, HUGEVAL, INCR_THRESH
440: LOGICAL INCR_PREC
441: * ..
442: * .. Parameters ..
443: INTEGER UNSTABLE_STATE, WORKING_STATE, CONV_STATE,
444: $ NOPROG_STATE, BASE_RESIDUAL, EXTRA_RESIDUAL,
445: $ EXTRA_Y
446: PARAMETER ( UNSTABLE_STATE = 0, WORKING_STATE = 1,
447: $ CONV_STATE = 2, NOPROG_STATE = 3 )
448: PARAMETER ( BASE_RESIDUAL = 0, EXTRA_RESIDUAL = 1,
449: $ EXTRA_Y = 2 )
450: INTEGER FINAL_NRM_ERR_I, FINAL_CMP_ERR_I, BERR_I
451: INTEGER RCOND_I, NRM_RCOND_I, NRM_ERR_I, CMP_RCOND_I
452: INTEGER CMP_ERR_I, PIV_GROWTH_I
453: PARAMETER ( FINAL_NRM_ERR_I = 1, FINAL_CMP_ERR_I = 2,
454: $ BERR_I = 3 )
455: PARAMETER ( RCOND_I = 4, NRM_RCOND_I = 5, NRM_ERR_I = 6 )
456: PARAMETER ( CMP_RCOND_I = 7, CMP_ERR_I = 8,
457: $ PIV_GROWTH_I = 9 )
458: INTEGER LA_LINRX_ITREF_I, LA_LINRX_ITHRESH_I,
459: $ LA_LINRX_CWISE_I
460: PARAMETER ( LA_LINRX_ITREF_I = 1,
461: $ LA_LINRX_ITHRESH_I = 2 )
462: PARAMETER ( LA_LINRX_CWISE_I = 3 )
463: INTEGER LA_LINRX_TRUST_I, LA_LINRX_ERR_I,
464: $ LA_LINRX_RCOND_I
465: PARAMETER ( LA_LINRX_TRUST_I = 1, LA_LINRX_ERR_I = 2 )
466: PARAMETER ( LA_LINRX_RCOND_I = 3 )
467: * ..
468: * .. External Subroutines ..
469: EXTERNAL DAXPY, DCOPY, DGBTRS, DGBMV, BLAS_DGBMV_X,
470: $ BLAS_DGBMV2_X, DLA_GBAMV, DLA_WWADDW, DLAMCH,
471: $ CHLA_TRANSTYPE, DLA_LIN_BERR
472: DOUBLE PRECISION DLAMCH
473: CHARACTER CHLA_TRANSTYPE
474: * ..
475: * .. Intrinsic Functions ..
476: INTRINSIC ABS, MAX, MIN
477: * ..
478: * .. Executable Statements ..
479: *
480: IF (INFO.NE.0) RETURN
481: TRANS = CHLA_TRANSTYPE(TRANS_TYPE)
482: EPS = DLAMCH( 'Epsilon' )
483: HUGEVAL = DLAMCH( 'Overflow' )
484: * Force HUGEVAL to Inf
485: HUGEVAL = HUGEVAL * HUGEVAL
486: * Using HUGEVAL may lead to spurious underflows.
487: INCR_THRESH = DBLE( N ) * EPS
488: M = KL+KU+1
489:
490: DO J = 1, NRHS
491: Y_PREC_STATE = EXTRA_RESIDUAL
492: IF ( Y_PREC_STATE .EQ. EXTRA_Y ) THEN
493: DO I = 1, N
494: Y_TAIL( I ) = 0.0D+0
495: END DO
496: END IF
497:
498: DXRAT = 0.0D+0
499: DXRATMAX = 0.0D+0
500: DZRAT = 0.0D+0
501: DZRATMAX = 0.0D+0
502: FINAL_DX_X = HUGEVAL
503: FINAL_DZ_Z = HUGEVAL
504: PREVNORMDX = HUGEVAL
505: PREV_DZ_Z = HUGEVAL
506: DZ_Z = HUGEVAL
507: DX_X = HUGEVAL
508:
509: X_STATE = WORKING_STATE
510: Z_STATE = UNSTABLE_STATE
511: INCR_PREC = .FALSE.
512:
513: DO CNT = 1, ITHRESH
514: *
515: * Compute residual RES = B_s - op(A_s) * Y,
516: * op(A) = A, A**T, or A**H depending on TRANS (and type).
517: *
518: CALL DCOPY( N, B( 1, J ), 1, RES, 1 )
519: IF ( Y_PREC_STATE .EQ. BASE_RESIDUAL ) THEN
520: CALL DGBMV( TRANS, M, N, KL, KU, -1.0D+0, AB, LDAB,
521: $ Y( 1, J ), 1, 1.0D+0, RES, 1 )
522: ELSE IF ( Y_PREC_STATE .EQ. EXTRA_RESIDUAL ) THEN
523: CALL BLAS_DGBMV_X( TRANS_TYPE, N, N, KL, KU,
524: $ -1.0D+0, AB, LDAB, Y( 1, J ), 1, 1.0D+0, RES, 1,
525: $ PREC_TYPE )
526: ELSE
527: CALL BLAS_DGBMV2_X( TRANS_TYPE, N, N, KL, KU, -1.0D+0,
528: $ AB, LDAB, Y( 1, J ), Y_TAIL, 1, 1.0D+0, RES, 1,
529: $ PREC_TYPE )
530: END IF
531:
532: ! XXX: RES is no longer needed.
533: CALL DCOPY( N, RES, 1, DY, 1 )
534: CALL DGBTRS( TRANS, N, KL, KU, 1, AFB, LDAFB, IPIV, DY, N,
535: $ INFO )
536: *
537: * Calculate relative changes DX_X, DZ_Z and ratios DXRAT, DZRAT.
538: *
539: NORMX = 0.0D+0
540: NORMY = 0.0D+0
541: NORMDX = 0.0D+0
542: DZ_Z = 0.0D+0
543: YMIN = HUGEVAL
544:
545: DO I = 1, N
546: YK = ABS( Y( I, J ) )
547: DYK = ABS( DY( I ) )
548:
549: IF ( YK .NE. 0.0D+0 ) THEN
550: DZ_Z = MAX( DZ_Z, DYK / YK )
551: ELSE IF ( DYK .NE. 0.0D+0 ) THEN
552: DZ_Z = HUGEVAL
553: END IF
554:
555: YMIN = MIN( YMIN, YK )
556:
557: NORMY = MAX( NORMY, YK )
558:
559: IF ( COLEQU ) THEN
560: NORMX = MAX( NORMX, YK * C( I ) )
561: NORMDX = MAX( NORMDX, DYK * C( I ) )
562: ELSE
563: NORMX = NORMY
564: NORMDX = MAX( NORMDX, DYK )
565: END IF
566: END DO
567:
568: IF ( NORMX .NE. 0.0D+0 ) THEN
569: DX_X = NORMDX / NORMX
570: ELSE IF ( NORMDX .EQ. 0.0D+0 ) THEN
571: DX_X = 0.0D+0
572: ELSE
573: DX_X = HUGEVAL
574: END IF
575:
576: DXRAT = NORMDX / PREVNORMDX
577: DZRAT = DZ_Z / PREV_DZ_Z
578: *
579: * Check termination criteria.
580: *
581: IF ( .NOT.IGNORE_CWISE
582: $ .AND. YMIN*RCOND .LT. INCR_THRESH*NORMY
583: $ .AND. Y_PREC_STATE .LT. EXTRA_Y )
584: $ INCR_PREC = .TRUE.
585:
586: IF ( X_STATE .EQ. NOPROG_STATE .AND. DXRAT .LE. RTHRESH )
587: $ X_STATE = WORKING_STATE
588: IF ( X_STATE .EQ. WORKING_STATE ) THEN
589: IF ( DX_X .LE. EPS ) THEN
590: X_STATE = CONV_STATE
591: ELSE IF ( DXRAT .GT. RTHRESH ) THEN
592: IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
593: INCR_PREC = .TRUE.
594: ELSE
595: X_STATE = NOPROG_STATE
596: END IF
597: ELSE
598: IF ( DXRAT .GT. DXRATMAX ) DXRATMAX = DXRAT
599: END IF
600: IF ( X_STATE .GT. WORKING_STATE ) FINAL_DX_X = DX_X
601: END IF
602:
603: IF ( Z_STATE .EQ. UNSTABLE_STATE .AND. DZ_Z .LE. DZ_UB )
604: $ Z_STATE = WORKING_STATE
605: IF ( Z_STATE .EQ. NOPROG_STATE .AND. DZRAT .LE. RTHRESH )
606: $ Z_STATE = WORKING_STATE
607: IF ( Z_STATE .EQ. WORKING_STATE ) THEN
608: IF ( DZ_Z .LE. EPS ) THEN
609: Z_STATE = CONV_STATE
610: ELSE IF ( DZ_Z .GT. DZ_UB ) THEN
611: Z_STATE = UNSTABLE_STATE
612: DZRATMAX = 0.0D+0
613: FINAL_DZ_Z = HUGEVAL
614: ELSE IF ( DZRAT .GT. RTHRESH ) THEN
615: IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
616: INCR_PREC = .TRUE.
617: ELSE
618: Z_STATE = NOPROG_STATE
619: END IF
620: ELSE
621: IF ( DZRAT .GT. DZRATMAX ) DZRATMAX = DZRAT
622: END IF
623: IF ( Z_STATE .GT. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
624: END IF
625: *
626: * Exit if both normwise and componentwise stopped working,
627: * but if componentwise is unstable, let it go at least two
628: * iterations.
629: *
630: IF ( X_STATE.NE.WORKING_STATE ) THEN
631: IF ( IGNORE_CWISE ) GOTO 666
632: IF ( Z_STATE.EQ.NOPROG_STATE .OR. Z_STATE.EQ.CONV_STATE )
633: $ GOTO 666
634: IF ( Z_STATE.EQ.UNSTABLE_STATE .AND. CNT.GT.1 ) GOTO 666
635: END IF
636:
637: IF ( INCR_PREC ) THEN
638: INCR_PREC = .FALSE.
639: Y_PREC_STATE = Y_PREC_STATE + 1
640: DO I = 1, N
641: Y_TAIL( I ) = 0.0D+0
642: END DO
643: END IF
644:
645: PREVNORMDX = NORMDX
646: PREV_DZ_Z = DZ_Z
647: *
648: * Update soluton.
649: *
650: IF (Y_PREC_STATE .LT. EXTRA_Y) THEN
651: CALL DAXPY( N, 1.0D+0, DY, 1, Y(1,J), 1 )
652: ELSE
653: CALL DLA_WWADDW( N, Y(1,J), Y_TAIL, DY )
654: END IF
655:
656: END DO
657: * Target of "IF (Z_STOP .AND. X_STOP)". Sun's f77 won't EXIT.
658: 666 CONTINUE
659: *
660: * Set final_* when cnt hits ithresh.
661: *
662: IF ( X_STATE .EQ. WORKING_STATE ) FINAL_DX_X = DX_X
663: IF ( Z_STATE .EQ. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
664: *
665: * Compute error bounds.
666: *
667: IF ( N_NORMS .GE. 1 ) THEN
668: ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) =
669: $ FINAL_DX_X / (1 - DXRATMAX)
670: END IF
671: IF (N_NORMS .GE. 2) THEN
672: ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) =
673: $ FINAL_DZ_Z / (1 - DZRATMAX)
674: END IF
675: *
676: * Compute componentwise relative backward error from formula
677: * max(i) ( abs(R(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
678: * where abs(Z) is the componentwise absolute value of the matrix
679: * or vector Z.
680: *
681: * Compute residual RES = B_s - op(A_s) * Y,
682: * op(A) = A, A**T, or A**H depending on TRANS (and type).
683: *
684: CALL DCOPY( N, B( 1, J ), 1, RES, 1 )
685: CALL DGBMV(TRANS, N, N, KL, KU, -1.0D+0, AB, LDAB, Y(1,J),
686: $ 1, 1.0D+0, RES, 1 )
687:
688: DO I = 1, N
689: AYB( I ) = ABS( B( I, J ) )
690: END DO
691: *
692: * Compute abs(op(A_s))*abs(Y) + abs(B_s).
693: *
694: CALL DLA_GBAMV( TRANS_TYPE, N, N, KL, KU, 1.0D+0,
695: $ AB, LDAB, Y(1, J), 1, 1.0D+0, AYB, 1 )
696:
697: CALL DLA_LIN_BERR( N, N, 1, RES, AYB, BERR_OUT( J ) )
698: *
699: * End of loop for each RHS
700: *
701: END DO
702: *
703: RETURN
1.17 ! bertrand 704: *
! 705: * End of DLA_GBRFSX_EXTENDED
! 706: *
1.1 bertrand 707: END
CVSweb interface <joel.bertrand@systella.fr>
![[BACK]](/cvsweb/icons/back.gif){kind=icon}
Return to dla_gbrfsx_extended.f CVS log ![[TXT]](/cvsweb/icons/text.gif){kind=icon}
![[DIR]](/cvsweb/icons/dir.gif){kind=icon}
Up to [local] / rpl / lapack / lapack