Annotation of rpl/lapack/lapack/zla_gbrfsx_extended.f, revision 1.16
1.8 bertrand 1: *> \brief \b ZLA_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 ZLA_GBRFSX_EXTENDED + dependencies
10: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/zla_gbrfsx_extended.f">
11: *> [TGZ]</a>
12: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/zla_gbrfsx_extended.f">
13: *> [ZIP]</a>
14: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/zla_gbrfsx_extended.f">
1.5 bertrand 15: *> [TXT]</a>
1.12 bertrand 16: *> \endhtmlonly
1.5 bertrand 17: *
18: * Definition:
19: * ===========
20: *
21: * SUBROUTINE ZLA_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: * COMPLEX*16 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: *>
50: *> ZLA_GBRFSX_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 ZGBRFSX 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.
59: *> \endverbatim
60: *
61: * Arguments:
62: * ==========
63: *
64: *> \param[in] PREC_TYPE
65: *> \verbatim
66: *> PREC_TYPE is INTEGER
67: *> Specifies the intermediate precision to be used in refinement.
1.16 ! bertrand 68: *> The value is defined by ILAPREC(P) where P is a CHARACTER and P
! 69: *> = 'S': Single
1.5 bertrand 70: *> = 'D': Double
71: *> = 'I': Indigenous
1.16 ! bertrand 72: *> = 'X' or 'E': Extra
1.5 bertrand 73: *> \endverbatim
74: *>
75: *> \param[in] TRANS_TYPE
76: *> \verbatim
77: *> TRANS_TYPE is INTEGER
78: *> Specifies the transposition operation on A.
1.16 ! bertrand 79: *> The value is defined by ILATRANS(T) where T is a CHARACTER and T
! 80: *> = 'N': No transpose
1.5 bertrand 81: *> = 'T': Transpose
82: *> = 'C': Conjugate transpose
83: *> \endverbatim
84: *>
85: *> \param[in] N
86: *> \verbatim
87: *> N is INTEGER
88: *> The number of linear equations, i.e., the order of the
89: *> matrix A. N >= 0.
90: *> \endverbatim
91: *>
92: *> \param[in] KL
93: *> \verbatim
94: *> KL is INTEGER
95: *> The number of subdiagonals within the band of A. KL >= 0.
96: *> \endverbatim
97: *>
98: *> \param[in] KU
99: *> \verbatim
100: *> KU is INTEGER
101: *> The number of superdiagonals within the band of A. KU >= 0
102: *> \endverbatim
103: *>
104: *> \param[in] NRHS
105: *> \verbatim
106: *> NRHS is INTEGER
107: *> The number of right-hand-sides, i.e., the number of columns of the
108: *> matrix B.
109: *> \endverbatim
110: *>
111: *> \param[in] AB
112: *> \verbatim
113: *> AB is COMPLEX*16 array, dimension (LDAB,N)
114: *> On entry, the N-by-N matrix A.
115: *> \endverbatim
116: *>
117: *> \param[in] LDAB
118: *> \verbatim
119: *> LDAB is INTEGER
120: *> The leading dimension of the array A. LDAB >= max(1,N).
121: *> \endverbatim
122: *>
123: *> \param[in] AFB
124: *> \verbatim
125: *> AFB is COMPLEX*16 array, dimension (LDAF,N)
126: *> The factors L and U from the factorization
127: *> A = P*L*U as computed by ZGBTRF.
128: *> \endverbatim
129: *>
130: *> \param[in] LDAFB
131: *> \verbatim
132: *> LDAFB is INTEGER
133: *> The leading dimension of the array AF. LDAF >= max(1,N).
134: *> \endverbatim
135: *>
136: *> \param[in] IPIV
137: *> \verbatim
138: *> IPIV is INTEGER array, dimension (N)
139: *> The pivot indices from the factorization A = P*L*U
140: *> as computed by ZGBTRF; row i of the matrix was interchanged
141: *> with row IPIV(i).
142: *> \endverbatim
143: *>
144: *> \param[in] COLEQU
145: *> \verbatim
146: *> COLEQU is LOGICAL
147: *> If .TRUE. then column equilibration was done to A before calling
148: *> this routine. This is needed to compute the solution and error
149: *> bounds correctly.
150: *> \endverbatim
151: *>
152: *> \param[in] C
153: *> \verbatim
154: *> C is DOUBLE PRECISION array, dimension (N)
155: *> The column scale factors for A. If COLEQU = .FALSE., C
156: *> is not accessed. If C is input, each element of C should be a power
157: *> of the radix to ensure a reliable solution and error estimates.
158: *> Scaling by powers of the radix does not cause rounding errors unless
159: *> the result underflows or overflows. Rounding errors during scaling
160: *> lead to refining with a matrix that is not equivalent to the
161: *> input matrix, producing error estimates that may not be
162: *> reliable.
163: *> \endverbatim
164: *>
165: *> \param[in] B
166: *> \verbatim
167: *> B is COMPLEX*16 array, dimension (LDB,NRHS)
168: *> The right-hand-side matrix B.
169: *> \endverbatim
170: *>
171: *> \param[in] LDB
172: *> \verbatim
173: *> LDB is INTEGER
174: *> The leading dimension of the array B. LDB >= max(1,N).
175: *> \endverbatim
176: *>
177: *> \param[in,out] Y
178: *> \verbatim
179: *> Y is COMPLEX*16 array, dimension (LDY,NRHS)
180: *> On entry, the solution matrix X, as computed by ZGBTRS.
181: *> On exit, the improved solution matrix Y.
182: *> \endverbatim
183: *>
184: *> \param[in] LDY
185: *> \verbatim
186: *> LDY is INTEGER
187: *> The leading dimension of the array Y. LDY >= max(1,N).
188: *> \endverbatim
189: *>
190: *> \param[out] BERR_OUT
191: *> \verbatim
192: *> BERR_OUT is DOUBLE PRECISION array, dimension (NRHS)
193: *> On exit, BERR_OUT(j) contains the componentwise relative backward
194: *> error for right-hand-side j from the formula
195: *> max(i) ( abs(RES(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
196: *> where abs(Z) is the componentwise absolute value of the matrix
197: *> or vector Z. This is computed by ZLA_LIN_BERR.
198: *> \endverbatim
199: *>
200: *> \param[in] N_NORMS
201: *> \verbatim
202: *> N_NORMS is INTEGER
203: *> Determines which error bounds to return (see ERR_BNDS_NORM
204: *> and ERR_BNDS_COMP).
205: *> If N_NORMS >= 1 return normwise error bounds.
206: *> If N_NORMS >= 2 return componentwise error bounds.
207: *> \endverbatim
208: *>
209: *> \param[in,out] ERR_BNDS_NORM
210: *> \verbatim
1.14 bertrand 211: *> ERR_BNDS_NORM is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS)
1.5 bertrand 212: *> For each right-hand side, this array contains information about
213: *> various error bounds and condition numbers corresponding to the
214: *> normwise relative error, which is defined as follows:
215: *>
216: *> Normwise relative error in the ith solution vector:
217: *> max_j (abs(XTRUE(j,i) - X(j,i)))
218: *> ------------------------------
219: *> max_j abs(X(j,i))
220: *>
221: *> The array is indexed by the type of error information as described
222: *> below. There currently are up to three pieces of information
223: *> returned.
224: *>
225: *> The first index in ERR_BNDS_NORM(i,:) corresponds to the ith
226: *> right-hand side.
227: *>
228: *> The second index in ERR_BNDS_NORM(:,err) contains the following
229: *> three fields:
230: *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
231: *> reciprocal condition number is less than the threshold
232: *> sqrt(n) * slamch('Epsilon').
233: *>
234: *> err = 2 "Guaranteed" error bound: The estimated forward error,
235: *> almost certainly within a factor of 10 of the true error
236: *> so long as the next entry is greater than the threshold
237: *> sqrt(n) * slamch('Epsilon'). This error bound should only
238: *> be trusted if the previous boolean is true.
239: *>
240: *> err = 3 Reciprocal condition number: Estimated normwise
241: *> reciprocal condition number. Compared with the threshold
242: *> sqrt(n) * slamch('Epsilon') to determine if the error
243: *> estimate is "guaranteed". These reciprocal condition
244: *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
245: *> appropriately scaled matrix Z.
246: *> Let Z = S*A, where S scales each row by a power of the
247: *> radix so all absolute row sums of Z are approximately 1.
248: *>
249: *> This subroutine is only responsible for setting the second field
250: *> above.
251: *> See Lapack Working Note 165 for further details and extra
252: *> cautions.
253: *> \endverbatim
254: *>
255: *> \param[in,out] ERR_BNDS_COMP
256: *> \verbatim
1.14 bertrand 257: *> ERR_BNDS_COMP is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS)
1.5 bertrand 258: *> For each right-hand side, this array contains information about
259: *> various error bounds and condition numbers corresponding to the
260: *> componentwise relative error, which is defined as follows:
261: *>
262: *> Componentwise relative error in the ith solution vector:
263: *> abs(XTRUE(j,i) - X(j,i))
264: *> max_j ----------------------
265: *> abs(X(j,i))
266: *>
267: *> The array is indexed by the right-hand side i (on which the
268: *> componentwise relative error depends), and the type of error
269: *> information as described below. There currently are up to three
270: *> pieces of information returned for each right-hand side. If
271: *> componentwise accuracy is not requested (PARAMS(3) = 0.0), then
1.16 ! bertrand 272: *> ERR_BNDS_COMP is not accessed. If N_ERR_BNDS < 3, then at most
1.5 bertrand 273: *> the first (:,N_ERR_BNDS) entries are returned.
274: *>
275: *> The first index in ERR_BNDS_COMP(i,:) corresponds to the ith
276: *> right-hand side.
277: *>
278: *> The second index in ERR_BNDS_COMP(:,err) contains the following
279: *> three fields:
280: *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
281: *> reciprocal condition number is less than the threshold
282: *> sqrt(n) * slamch('Epsilon').
283: *>
284: *> err = 2 "Guaranteed" error bound: The estimated forward error,
285: *> almost certainly within a factor of 10 of the true error
286: *> so long as the next entry is greater than the threshold
287: *> sqrt(n) * slamch('Epsilon'). This error bound should only
288: *> be trusted if the previous boolean is true.
289: *>
290: *> err = 3 Reciprocal condition number: Estimated componentwise
291: *> reciprocal condition number. Compared with the threshold
292: *> sqrt(n) * slamch('Epsilon') to determine if the error
293: *> estimate is "guaranteed". These reciprocal condition
294: *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
295: *> appropriately scaled matrix Z.
296: *> Let Z = S*(A*diag(x)), where x is the solution for the
297: *> current right-hand side and S scales each row of
298: *> A*diag(x) by a power of the radix so all absolute row
299: *> sums of Z are approximately 1.
300: *>
301: *> This subroutine is only responsible for setting the second field
302: *> above.
303: *> See Lapack Working Note 165 for further details and extra
304: *> cautions.
305: *> \endverbatim
306: *>
307: *> \param[in] RES
308: *> \verbatim
309: *> RES is COMPLEX*16 array, dimension (N)
310: *> Workspace to hold the intermediate residual.
311: *> \endverbatim
312: *>
313: *> \param[in] AYB
314: *> \verbatim
315: *> AYB is DOUBLE PRECISION array, dimension (N)
316: *> Workspace.
317: *> \endverbatim
318: *>
319: *> \param[in] DY
320: *> \verbatim
321: *> DY is COMPLEX*16 array, dimension (N)
322: *> Workspace to hold the intermediate solution.
323: *> \endverbatim
324: *>
325: *> \param[in] Y_TAIL
326: *> \verbatim
327: *> Y_TAIL is COMPLEX*16 array, dimension (N)
328: *> Workspace to hold the trailing bits of the intermediate solution.
329: *> \endverbatim
330: *>
331: *> \param[in] RCOND
332: *> \verbatim
333: *> RCOND is DOUBLE PRECISION
334: *> Reciprocal scaled condition number. This is an estimate of the
335: *> reciprocal Skeel condition number of the matrix A after
336: *> equilibration (if done). If this is less than the machine
337: *> precision (in particular, if it is zero), the matrix is singular
338: *> to working precision. Note that the error may still be small even
339: *> if this number is very small and the matrix appears ill-
340: *> conditioned.
341: *> \endverbatim
342: *>
343: *> \param[in] ITHRESH
344: *> \verbatim
345: *> ITHRESH is INTEGER
346: *> The maximum number of residual computations allowed for
347: *> refinement. The default is 10. For 'aggressive' set to 100 to
348: *> permit convergence using approximate factorizations or
349: *> factorizations other than LU. If the factorization uses a
350: *> technique other than Gaussian elimination, the guarantees in
351: *> ERR_BNDS_NORM and ERR_BNDS_COMP may no longer be trustworthy.
352: *> \endverbatim
353: *>
354: *> \param[in] RTHRESH
355: *> \verbatim
356: *> RTHRESH is DOUBLE PRECISION
357: *> Determines when to stop refinement if the error estimate stops
358: *> decreasing. Refinement will stop when the next solution no longer
359: *> satisfies norm(dx_{i+1}) < RTHRESH * norm(dx_i) where norm(Z) is
360: *> the infinity norm of Z. RTHRESH satisfies 0 < RTHRESH <= 1. The
361: *> default value is 0.5. For 'aggressive' set to 0.9 to permit
362: *> convergence on extremely ill-conditioned matrices. See LAWN 165
363: *> for more details.
364: *> \endverbatim
365: *>
366: *> \param[in] DZ_UB
367: *> \verbatim
368: *> DZ_UB is DOUBLE PRECISION
369: *> Determines when to start considering componentwise convergence.
370: *> Componentwise convergence is only considered after each component
371: *> of the solution Y is stable, which we definte as the relative
372: *> change in each component being less than DZ_UB. The default value
373: *> is 0.25, requiring the first bit to be stable. See LAWN 165 for
374: *> more details.
375: *> \endverbatim
376: *>
377: *> \param[in] IGNORE_CWISE
378: *> \verbatim
379: *> IGNORE_CWISE is LOGICAL
380: *> If .TRUE. then ignore componentwise convergence. Default value
381: *> is .FALSE..
382: *> \endverbatim
383: *>
384: *> \param[out] INFO
385: *> \verbatim
386: *> INFO is INTEGER
387: *> = 0: Successful exit.
388: *> < 0: if INFO = -i, the ith argument to ZGBTRS had an illegal
389: *> value
390: *> \endverbatim
391: *
392: * Authors:
393: * ========
394: *
1.12 bertrand 395: *> \author Univ. of Tennessee
396: *> \author Univ. of California Berkeley
397: *> \author Univ. of Colorado Denver
398: *> \author NAG Ltd.
1.5 bertrand 399: *
1.14 bertrand 400: *> \date June 2017
1.5 bertrand 401: *
402: *> \ingroup complex16GBcomputational
403: *
404: * =====================================================================
1.1 bertrand 405: SUBROUTINE ZLA_GBRFSX_EXTENDED( PREC_TYPE, TRANS_TYPE, N, KL, KU,
406: $ NRHS, AB, LDAB, AFB, LDAFB, IPIV,
407: $ COLEQU, C, B, LDB, Y, LDY,
408: $ BERR_OUT, N_NORMS, ERR_BNDS_NORM,
409: $ ERR_BNDS_COMP, RES, AYB, DY,
410: $ Y_TAIL, RCOND, ITHRESH, RTHRESH,
411: $ DZ_UB, IGNORE_CWISE, INFO )
412: *
1.14 bertrand 413: * -- LAPACK computational routine (version 3.7.1) --
1.5 bertrand 414: * -- LAPACK is a software package provided by Univ. of Tennessee, --
415: * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
1.14 bertrand 416: * June 2017
1.1 bertrand 417: *
418: * .. Scalar Arguments ..
419: INTEGER INFO, LDAB, LDAFB, LDB, LDY, N, KL, KU, NRHS,
420: $ PREC_TYPE, TRANS_TYPE, N_NORMS, ITHRESH
421: LOGICAL COLEQU, IGNORE_CWISE
422: DOUBLE PRECISION RTHRESH, DZ_UB
423: * ..
424: * .. Array Arguments ..
425: INTEGER IPIV( * )
426: COMPLEX*16 AB( LDAB, * ), AFB( LDAFB, * ), B( LDB, * ),
427: $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * )
428: DOUBLE PRECISION C( * ), AYB(*), RCOND, BERR_OUT( * ),
429: $ ERR_BNDS_NORM( NRHS, * ),
430: $ ERR_BNDS_COMP( NRHS, * )
431: * ..
432: *
433: * =====================================================================
434: *
435: * .. Local Scalars ..
436: CHARACTER TRANS
437: INTEGER CNT, I, J, M, X_STATE, Z_STATE, Y_PREC_STATE
438: DOUBLE PRECISION YK, DYK, YMIN, NORMY, NORMX, NORMDX, DXRAT,
439: $ DZRAT, PREVNORMDX, PREV_DZ_Z, DXRATMAX,
440: $ DZRATMAX, DX_X, DZ_Z, FINAL_DX_X, FINAL_DZ_Z,
441: $ EPS, HUGEVAL, INCR_THRESH
442: LOGICAL INCR_PREC
443: COMPLEX*16 ZDUM
444: * ..
445: * .. Parameters ..
446: INTEGER UNSTABLE_STATE, WORKING_STATE, CONV_STATE,
447: $ NOPROG_STATE, BASE_RESIDUAL, EXTRA_RESIDUAL,
448: $ EXTRA_Y
449: PARAMETER ( UNSTABLE_STATE = 0, WORKING_STATE = 1,
450: $ CONV_STATE = 2, NOPROG_STATE = 3 )
451: PARAMETER ( BASE_RESIDUAL = 0, EXTRA_RESIDUAL = 1,
452: $ EXTRA_Y = 2 )
453: INTEGER FINAL_NRM_ERR_I, FINAL_CMP_ERR_I, BERR_I
454: INTEGER RCOND_I, NRM_RCOND_I, NRM_ERR_I, CMP_RCOND_I
455: INTEGER CMP_ERR_I, PIV_GROWTH_I
456: PARAMETER ( FINAL_NRM_ERR_I = 1, FINAL_CMP_ERR_I = 2,
457: $ BERR_I = 3 )
458: PARAMETER ( RCOND_I = 4, NRM_RCOND_I = 5, NRM_ERR_I = 6 )
459: PARAMETER ( CMP_RCOND_I = 7, CMP_ERR_I = 8,
460: $ PIV_GROWTH_I = 9 )
461: INTEGER LA_LINRX_ITREF_I, LA_LINRX_ITHRESH_I,
462: $ LA_LINRX_CWISE_I
463: PARAMETER ( LA_LINRX_ITREF_I = 1,
464: $ LA_LINRX_ITHRESH_I = 2 )
465: PARAMETER ( LA_LINRX_CWISE_I = 3 )
466: INTEGER LA_LINRX_TRUST_I, LA_LINRX_ERR_I,
467: $ LA_LINRX_RCOND_I
468: PARAMETER ( LA_LINRX_TRUST_I = 1, LA_LINRX_ERR_I = 2 )
469: PARAMETER ( LA_LINRX_RCOND_I = 3 )
470: * ..
471: * .. External Subroutines ..
472: EXTERNAL ZAXPY, ZCOPY, ZGBTRS, ZGBMV, BLAS_ZGBMV_X,
473: $ BLAS_ZGBMV2_X, ZLA_GBAMV, ZLA_WWADDW, DLAMCH,
474: $ CHLA_TRANSTYPE, ZLA_LIN_BERR
475: DOUBLE PRECISION DLAMCH
476: CHARACTER CHLA_TRANSTYPE
477: * ..
478: * .. Intrinsic Functions..
479: INTRINSIC ABS, MAX, MIN
480: * ..
481: * .. Statement Functions ..
482: DOUBLE PRECISION CABS1
483: * ..
484: * .. Statement Function Definitions ..
485: CABS1( ZDUM ) = ABS( DBLE( ZDUM ) ) + ABS( DIMAG( ZDUM ) )
486: * ..
487: * .. Executable Statements ..
488: *
489: IF (INFO.NE.0) RETURN
490: TRANS = CHLA_TRANSTYPE(TRANS_TYPE)
491: EPS = DLAMCH( 'Epsilon' )
492: HUGEVAL = DLAMCH( 'Overflow' )
493: * Force HUGEVAL to Inf
494: HUGEVAL = HUGEVAL * HUGEVAL
495: * Using HUGEVAL may lead to spurious underflows.
496: INCR_THRESH = DBLE( N ) * EPS
497: M = KL+KU+1
498:
499: DO J = 1, NRHS
500: Y_PREC_STATE = EXTRA_RESIDUAL
501: IF ( Y_PREC_STATE .EQ. EXTRA_Y ) THEN
502: DO I = 1, N
503: Y_TAIL( I ) = 0.0D+0
504: END DO
505: END IF
506:
507: DXRAT = 0.0D+0
508: DXRATMAX = 0.0D+0
509: DZRAT = 0.0D+0
510: DZRATMAX = 0.0D+0
511: FINAL_DX_X = HUGEVAL
512: FINAL_DZ_Z = HUGEVAL
513: PREVNORMDX = HUGEVAL
514: PREV_DZ_Z = HUGEVAL
515: DZ_Z = HUGEVAL
516: DX_X = HUGEVAL
517:
518: X_STATE = WORKING_STATE
519: Z_STATE = UNSTABLE_STATE
520: INCR_PREC = .FALSE.
521:
522: DO CNT = 1, ITHRESH
523: *
524: * Compute residual RES = B_s - op(A_s) * Y,
525: * op(A) = A, A**T, or A**H depending on TRANS (and type).
526: *
527: CALL ZCOPY( N, B( 1, J ), 1, RES, 1 )
528: IF ( Y_PREC_STATE .EQ. BASE_RESIDUAL ) THEN
529: CALL ZGBMV( TRANS, M, N, KL, KU, (-1.0D+0,0.0D+0), AB,
530: $ LDAB, Y( 1, J ), 1, (1.0D+0,0.0D+0), RES, 1 )
531: ELSE IF ( Y_PREC_STATE .EQ. EXTRA_RESIDUAL ) THEN
532: CALL BLAS_ZGBMV_X( TRANS_TYPE, N, N, KL, KU,
533: $ (-1.0D+0,0.0D+0), AB, LDAB, Y( 1, J ), 1,
534: $ (1.0D+0,0.0D+0), RES, 1, PREC_TYPE )
535: ELSE
536: CALL BLAS_ZGBMV2_X( TRANS_TYPE, N, N, KL, KU,
537: $ (-1.0D+0,0.0D+0), AB, LDAB, Y( 1, J ), Y_TAIL, 1,
538: $ (1.0D+0,0.0D+0), RES, 1, PREC_TYPE )
539: END IF
540:
541: ! XXX: RES is no longer needed.
542: CALL ZCOPY( N, RES, 1, DY, 1 )
543: CALL ZGBTRS( TRANS, N, KL, KU, 1, AFB, LDAFB, IPIV, DY, N,
544: $ INFO )
545: *
546: * Calculate relative changes DX_X, DZ_Z and ratios DXRAT, DZRAT.
547: *
548: NORMX = 0.0D+0
549: NORMY = 0.0D+0
550: NORMDX = 0.0D+0
551: DZ_Z = 0.0D+0
552: YMIN = HUGEVAL
553:
554: DO I = 1, N
555: YK = CABS1( Y( I, J ) )
556: DYK = CABS1( DY( I ) )
557:
558: IF (YK .NE. 0.0D+0) THEN
559: DZ_Z = MAX( DZ_Z, DYK / YK )
560: ELSE IF ( DYK .NE. 0.0D+0 ) THEN
561: DZ_Z = HUGEVAL
562: END IF
563:
564: YMIN = MIN( YMIN, YK )
565:
566: NORMY = MAX( NORMY, YK )
567:
568: IF ( COLEQU ) THEN
569: NORMX = MAX( NORMX, YK * C( I ) )
570: NORMDX = MAX(NORMDX, DYK * C(I))
571: ELSE
572: NORMX = NORMY
573: NORMDX = MAX( NORMDX, DYK )
574: END IF
575: END DO
576:
577: IF ( NORMX .NE. 0.0D+0 ) THEN
578: DX_X = NORMDX / NORMX
579: ELSE IF ( NORMDX .EQ. 0.0D+0 ) THEN
580: DX_X = 0.0D+0
581: ELSE
582: DX_X = HUGEVAL
583: END IF
584:
585: DXRAT = NORMDX / PREVNORMDX
586: DZRAT = DZ_Z / PREV_DZ_Z
587: *
588: * Check termination criteria.
589: *
590: IF (.NOT.IGNORE_CWISE
591: $ .AND. YMIN*RCOND .LT. INCR_THRESH*NORMY
592: $ .AND. Y_PREC_STATE .LT. EXTRA_Y )
593: $ INCR_PREC = .TRUE.
594:
595: IF ( X_STATE .EQ. NOPROG_STATE .AND. DXRAT .LE. RTHRESH )
596: $ X_STATE = WORKING_STATE
597: IF ( X_STATE .EQ. WORKING_STATE ) THEN
598: IF ( DX_X .LE. EPS ) THEN
599: X_STATE = CONV_STATE
600: ELSE IF ( DXRAT .GT. RTHRESH ) THEN
601: IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
602: INCR_PREC = .TRUE.
603: ELSE
604: X_STATE = NOPROG_STATE
605: END IF
606: ELSE
607: IF ( DXRAT .GT. DXRATMAX ) DXRATMAX = DXRAT
608: END IF
609: IF ( X_STATE .GT. WORKING_STATE ) FINAL_DX_X = DX_X
610: END IF
611:
612: IF ( Z_STATE .EQ. UNSTABLE_STATE .AND. DZ_Z .LE. DZ_UB )
613: $ Z_STATE = WORKING_STATE
614: IF ( Z_STATE .EQ. NOPROG_STATE .AND. DZRAT .LE. RTHRESH )
615: $ Z_STATE = WORKING_STATE
616: IF ( Z_STATE .EQ. WORKING_STATE ) THEN
617: IF ( DZ_Z .LE. EPS ) THEN
618: Z_STATE = CONV_STATE
619: ELSE IF ( DZ_Z .GT. DZ_UB ) THEN
620: Z_STATE = UNSTABLE_STATE
621: DZRATMAX = 0.0D+0
622: FINAL_DZ_Z = HUGEVAL
623: ELSE IF ( DZRAT .GT. RTHRESH ) THEN
624: IF ( Y_PREC_STATE .NE. EXTRA_Y ) THEN
625: INCR_PREC = .TRUE.
626: ELSE
627: Z_STATE = NOPROG_STATE
628: END IF
629: ELSE
630: IF ( DZRAT .GT. DZRATMAX ) DZRATMAX = DZRAT
631: END IF
632: IF ( Z_STATE .GT. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
633: END IF
634: *
635: * Exit if both normwise and componentwise stopped working,
636: * but if componentwise is unstable, let it go at least two
637: * iterations.
638: *
639: IF ( X_STATE.NE.WORKING_STATE ) THEN
640: IF ( IGNORE_CWISE ) GOTO 666
641: IF ( Z_STATE.EQ.NOPROG_STATE .OR. Z_STATE.EQ.CONV_STATE )
642: $ GOTO 666
643: IF ( Z_STATE.EQ.UNSTABLE_STATE .AND. CNT.GT.1 ) GOTO 666
644: END IF
645:
646: IF ( INCR_PREC ) THEN
647: INCR_PREC = .FALSE.
648: Y_PREC_STATE = Y_PREC_STATE + 1
649: DO I = 1, N
650: Y_TAIL( I ) = 0.0D+0
651: END DO
652: END IF
653:
654: PREVNORMDX = NORMDX
655: PREV_DZ_Z = DZ_Z
656: *
657: * Update soluton.
658: *
659: IF ( Y_PREC_STATE .LT. EXTRA_Y ) THEN
660: CALL ZAXPY( N, (1.0D+0,0.0D+0), DY, 1, Y(1,J), 1 )
661: ELSE
662: CALL ZLA_WWADDW( N, Y(1,J), Y_TAIL, DY )
663: END IF
664:
665: END DO
666: * Target of "IF (Z_STOP .AND. X_STOP)". Sun's f77 won't EXIT.
667: 666 CONTINUE
668: *
669: * Set final_* when cnt hits ithresh.
670: *
671: IF ( X_STATE .EQ. WORKING_STATE ) FINAL_DX_X = DX_X
672: IF ( Z_STATE .EQ. WORKING_STATE ) FINAL_DZ_Z = DZ_Z
673: *
674: * Compute error bounds.
675: *
676: IF ( N_NORMS .GE. 1 ) THEN
677: ERR_BNDS_NORM( J, LA_LINRX_ERR_I ) =
678: $ FINAL_DX_X / (1 - DXRATMAX)
679: END IF
680: IF ( N_NORMS .GE. 2 ) THEN
681: ERR_BNDS_COMP( J, LA_LINRX_ERR_I ) =
682: $ FINAL_DZ_Z / (1 - DZRATMAX)
683: END IF
684: *
685: * Compute componentwise relative backward error from formula
686: * max(i) ( abs(R(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) )
687: * where abs(Z) is the componentwise absolute value of the matrix
688: * or vector Z.
689: *
690: * Compute residual RES = B_s - op(A_s) * Y,
691: * op(A) = A, A**T, or A**H depending on TRANS (and type).
692: *
693: CALL ZCOPY( N, B( 1, J ), 1, RES, 1 )
694: CALL ZGBMV( TRANS, N, N, KL, KU, (-1.0D+0,0.0D+0), AB, LDAB,
695: $ Y(1,J), 1, (1.0D+0,0.0D+0), RES, 1 )
696:
697: DO I = 1, N
698: AYB( I ) = CABS1( B( I, J ) )
699: END DO
700: *
701: * Compute abs(op(A_s))*abs(Y) + abs(B_s).
702: *
703: CALL ZLA_GBAMV( TRANS_TYPE, N, N, KL, KU, 1.0D+0,
704: $ AB, LDAB, Y(1, J), 1, 1.0D+0, AYB, 1 )
705:
706: CALL ZLA_LIN_BERR( N, N, 1, RES, AYB, BERR_OUT( J ) )
707: *
708: * End of loop for each RHS.
709: *
710: END DO
711: *
712: RETURN
713: END
CVSweb interface <joel.bertrand@systella.fr>
![[BACK]](/cvsweb/icons/back.gif){kind=icon}
Return to zla_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