1: *> \brief <b> DGBSVXX computes the solution to system of linear equations A * X = B for GB matrices</b>
2: *
3: * =========== DOCUMENTATION ===========
4: *
5: * Online html documentation available at
6: * http://www.netlib.org/lapack/explore-html/
7: *
8: *> \htmlonly
9: *> Download DGBSVXX + dependencies
10: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dgbsvxx.f">
11: *> [TGZ]</a>
12: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dgbsvxx.f">
13: *> [ZIP]</a>
14: *> <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dgbsvxx.f">
15: *> [TXT]</a>
16: *> \endhtmlonly
17: *
18: * Definition:
19: * ===========
20: *
21: * SUBROUTINE DGBSVXX( FACT, TRANS, N, KL, KU, NRHS, AB, LDAB, AFB,
22: * LDAFB, IPIV, EQUED, R, C, B, LDB, X, LDX,
23: * RCOND, RPVGRW, BERR, N_ERR_BNDS,
24: * ERR_BNDS_NORM, ERR_BNDS_COMP, NPARAMS, PARAMS,
25: * WORK, IWORK, INFO )
26: *
27: * .. Scalar Arguments ..
28: * CHARACTER EQUED, FACT, TRANS
29: * INTEGER INFO, LDAB, LDAFB, LDB, LDX, N, NRHS, NPARAMS,
30: * $ N_ERR_BNDS, KL, KU
31: * DOUBLE PRECISION RCOND, RPVGRW
32: * ..
33: * .. Array Arguments ..
34: * INTEGER IPIV( * ), IWORK( * )
35: * DOUBLE PRECISION AB( LDAB, * ), AFB( LDAFB, * ), B( LDB, * ),
36: * $ X( LDX , * ),WORK( * )
37: * DOUBLE PRECISION R( * ), C( * ), PARAMS( * ), BERR( * ),
38: * $ ERR_BNDS_NORM( NRHS, * ),
39: * $ ERR_BNDS_COMP( NRHS, * )
40: * ..
41: *
42: *
43: *> \par Purpose:
44: * =============
45: *>
46: *> \verbatim
47: *>
48: *> DGBSVXX uses the LU factorization to compute the solution to a
49: *> double precision system of linear equations A * X = B, where A is an
50: *> N-by-N matrix and X and B are N-by-NRHS matrices.
51: *>
52: *> If requested, both normwise and maximum componentwise error bounds
53: *> are returned. DGBSVXX will return a solution with a tiny
54: *> guaranteed error (O(eps) where eps is the working machine
55: *> precision) unless the matrix is very ill-conditioned, in which
56: *> case a warning is returned. Relevant condition numbers also are
57: *> calculated and returned.
58: *>
59: *> DGBSVXX accepts user-provided factorizations and equilibration
60: *> factors; see the definitions of the FACT and EQUED options.
61: *> Solving with refinement and using a factorization from a previous
62: *> DGBSVXX call will also produce a solution with either O(eps)
63: *> errors or warnings, but we cannot make that claim for general
64: *> user-provided factorizations and equilibration factors if they
65: *> differ from what DGBSVXX would itself produce.
66: *> \endverbatim
67: *
68: *> \par Description:
69: * =================
70: *>
71: *> \verbatim
72: *>
73: *> The following steps are performed:
74: *>
75: *> 1. If FACT = 'E', double precision scaling factors are computed to equilibrate
76: *> the system:
77: *>
78: *> TRANS = 'N': diag(R)*A*diag(C) *inv(diag(C))*X = diag(R)*B
79: *> TRANS = 'T': (diag(R)*A*diag(C))**T *inv(diag(R))*X = diag(C)*B
80: *> TRANS = 'C': (diag(R)*A*diag(C))**H *inv(diag(R))*X = diag(C)*B
81: *>
82: *> Whether or not the system will be equilibrated depends on the
83: *> scaling of the matrix A, but if equilibration is used, A is
84: *> overwritten by diag(R)*A*diag(C) and B by diag(R)*B (if TRANS='N')
85: *> or diag(C)*B (if TRANS = 'T' or 'C').
86: *>
87: *> 2. If FACT = 'N' or 'E', the LU decomposition is used to factor
88: *> the matrix A (after equilibration if FACT = 'E') as
89: *>
90: *> A = P * L * U,
91: *>
92: *> where P is a permutation matrix, L is a unit lower triangular
93: *> matrix, and U is upper triangular.
94: *>
95: *> 3. If some U(i,i)=0, so that U is exactly singular, then the
96: *> routine returns with INFO = i. Otherwise, the factored form of A
97: *> is used to estimate the condition number of the matrix A (see
98: *> argument RCOND). If the reciprocal of the condition number is less
99: *> than machine precision, the routine still goes on to solve for X
100: *> and compute error bounds as described below.
101: *>
102: *> 4. The system of equations is solved for X using the factored form
103: *> of A.
104: *>
105: *> 5. By default (unless PARAMS(LA_LINRX_ITREF_I) is set to zero),
106: *> the routine will use iterative refinement to try to get a small
107: *> error and error bounds. Refinement calculates the residual to at
108: *> least twice the working precision.
109: *>
110: *> 6. If equilibration was used, the matrix X is premultiplied by
111: *> diag(C) (if TRANS = 'N') or diag(R) (if TRANS = 'T' or 'C') so
112: *> that it solves the original system before equilibration.
113: *> \endverbatim
114: *
115: * Arguments:
116: * ==========
117: *
118: *> \verbatim
119: *> Some optional parameters are bundled in the PARAMS array. These
120: *> settings determine how refinement is performed, but often the
121: *> defaults are acceptable. If the defaults are acceptable, users
122: *> can pass NPARAMS = 0 which prevents the source code from accessing
123: *> the PARAMS argument.
124: *> \endverbatim
125: *>
126: *> \param[in] FACT
127: *> \verbatim
128: *> FACT is CHARACTER*1
129: *> Specifies whether or not the factored form of the matrix A is
130: *> supplied on entry, and if not, whether the matrix A should be
131: *> equilibrated before it is factored.
132: *> = 'F': On entry, AF and IPIV contain the factored form of A.
133: *> If EQUED is not 'N', the matrix A has been
134: *> equilibrated with scaling factors given by R and C.
135: *> A, AF, and IPIV are not modified.
136: *> = 'N': The matrix A will be copied to AF and factored.
137: *> = 'E': The matrix A will be equilibrated if necessary, then
138: *> copied to AF and factored.
139: *> \endverbatim
140: *>
141: *> \param[in] TRANS
142: *> \verbatim
143: *> TRANS is CHARACTER*1
144: *> Specifies the form of the system of equations:
145: *> = 'N': A * X = B (No transpose)
146: *> = 'T': A**T * X = B (Transpose)
147: *> = 'C': A**H * X = B (Conjugate Transpose = Transpose)
148: *> \endverbatim
149: *>
150: *> \param[in] N
151: *> \verbatim
152: *> N is INTEGER
153: *> The number of linear equations, i.e., the order of the
154: *> matrix A. N >= 0.
155: *> \endverbatim
156: *>
157: *> \param[in] KL
158: *> \verbatim
159: *> KL is INTEGER
160: *> The number of subdiagonals within the band of A. KL >= 0.
161: *> \endverbatim
162: *>
163: *> \param[in] KU
164: *> \verbatim
165: *> KU is INTEGER
166: *> The number of superdiagonals within the band of A. KU >= 0.
167: *> \endverbatim
168: *>
169: *> \param[in] NRHS
170: *> \verbatim
171: *> NRHS is INTEGER
172: *> The number of right hand sides, i.e., the number of columns
173: *> of the matrices B and X. NRHS >= 0.
174: *> \endverbatim
175: *>
176: *> \param[in,out] AB
177: *> \verbatim
178: *> AB is DOUBLE PRECISION array, dimension (LDAB,N)
179: *> On entry, the matrix A in band storage, in rows 1 to KL+KU+1.
180: *> The j-th column of A is stored in the j-th column of the
181: *> array AB as follows:
182: *> AB(KU+1+i-j,j) = A(i,j) for max(1,j-KU)<=i<=min(N,j+kl)
183: *>
184: *> If FACT = 'F' and EQUED is not 'N', then AB must have been
185: *> equilibrated by the scaling factors in R and/or C. AB is not
186: *> modified if FACT = 'F' or 'N', or if FACT = 'E' and
187: *> EQUED = 'N' on exit.
188: *>
189: *> On exit, if EQUED .ne. 'N', A is scaled as follows:
190: *> EQUED = 'R': A := diag(R) * A
191: *> EQUED = 'C': A := A * diag(C)
192: *> EQUED = 'B': A := diag(R) * A * diag(C).
193: *> \endverbatim
194: *>
195: *> \param[in] LDAB
196: *> \verbatim
197: *> LDAB is INTEGER
198: *> The leading dimension of the array AB. LDAB >= KL+KU+1.
199: *> \endverbatim
200: *>
201: *> \param[in,out] AFB
202: *> \verbatim
203: *> AFB is DOUBLE PRECISION array, dimension (LDAFB,N)
204: *> If FACT = 'F', then AFB is an input argument and on entry
205: *> contains details of the LU factorization of the band matrix
206: *> A, as computed by DGBTRF. U is stored as an upper triangular
207: *> band matrix with KL+KU superdiagonals in rows 1 to KL+KU+1,
208: *> and the multipliers used during the factorization are stored
209: *> in rows KL+KU+2 to 2*KL+KU+1. If EQUED .ne. 'N', then AFB is
210: *> the factored form of the equilibrated matrix A.
211: *>
212: *> If FACT = 'N', then AF is an output argument and on exit
213: *> returns the factors L and U from the factorization A = P*L*U
214: *> of the original matrix A.
215: *>
216: *> If FACT = 'E', then AF is an output argument and on exit
217: *> returns the factors L and U from the factorization A = P*L*U
218: *> of the equilibrated matrix A (see the description of A for
219: *> the form of the equilibrated matrix).
220: *> \endverbatim
221: *>
222: *> \param[in] LDAFB
223: *> \verbatim
224: *> LDAFB is INTEGER
225: *> The leading dimension of the array AFB. LDAFB >= 2*KL+KU+1.
226: *> \endverbatim
227: *>
228: *> \param[in,out] IPIV
229: *> \verbatim
230: *> IPIV is INTEGER array, dimension (N)
231: *> If FACT = 'F', then IPIV is an input argument and on entry
232: *> contains the pivot indices from the factorization A = P*L*U
233: *> as computed by DGETRF; row i of the matrix was interchanged
234: *> with row IPIV(i).
235: *>
236: *> If FACT = 'N', then IPIV is an output argument and on exit
237: *> contains the pivot indices from the factorization A = P*L*U
238: *> of the original matrix A.
239: *>
240: *> If FACT = 'E', then IPIV is an output argument and on exit
241: *> contains the pivot indices from the factorization A = P*L*U
242: *> of the equilibrated matrix A.
243: *> \endverbatim
244: *>
245: *> \param[in,out] EQUED
246: *> \verbatim
247: *> EQUED is CHARACTER*1
248: *> Specifies the form of equilibration that was done.
249: *> = 'N': No equilibration (always true if FACT = 'N').
250: *> = 'R': Row equilibration, i.e., A has been premultiplied by
251: *> diag(R).
252: *> = 'C': Column equilibration, i.e., A has been postmultiplied
253: *> by diag(C).
254: *> = 'B': Both row and column equilibration, i.e., A has been
255: *> replaced by diag(R) * A * diag(C).
256: *> EQUED is an input argument if FACT = 'F'; otherwise, it is an
257: *> output argument.
258: *> \endverbatim
259: *>
260: *> \param[in,out] R
261: *> \verbatim
262: *> R is DOUBLE PRECISION array, dimension (N)
263: *> The row scale factors for A. If EQUED = 'R' or 'B', A is
264: *> multiplied on the left by diag(R); if EQUED = 'N' or 'C', R
265: *> is not accessed. R is an input argument if FACT = 'F';
266: *> otherwise, R is an output argument. If FACT = 'F' and
267: *> EQUED = 'R' or 'B', each element of R must be positive.
268: *> If R is output, each element of R is a power of the radix.
269: *> If R is input, each element of R should be a power of the radix
270: *> to ensure a reliable solution and error estimates. Scaling by
271: *> powers of the radix does not cause rounding errors unless the
272: *> result underflows or overflows. Rounding errors during scaling
273: *> lead to refining with a matrix that is not equivalent to the
274: *> input matrix, producing error estimates that may not be
275: *> reliable.
276: *> \endverbatim
277: *>
278: *> \param[in,out] C
279: *> \verbatim
280: *> C is DOUBLE PRECISION array, dimension (N)
281: *> The column scale factors for A. If EQUED = 'C' or 'B', A is
282: *> multiplied on the right by diag(C); if EQUED = 'N' or 'R', C
283: *> is not accessed. C is an input argument if FACT = 'F';
284: *> otherwise, C is an output argument. If FACT = 'F' and
285: *> EQUED = 'C' or 'B', each element of C must be positive.
286: *> If C is output, each element of C is a power of the radix.
287: *> If C is input, each element of C should be a power of the radix
288: *> to ensure a reliable solution and error estimates. Scaling by
289: *> powers of the radix does not cause rounding errors unless the
290: *> result underflows or overflows. Rounding errors during scaling
291: *> lead to refining with a matrix that is not equivalent to the
292: *> input matrix, producing error estimates that may not be
293: *> reliable.
294: *> \endverbatim
295: *>
296: *> \param[in,out] B
297: *> \verbatim
298: *> B is DOUBLE PRECISION array, dimension (LDB,NRHS)
299: *> On entry, the N-by-NRHS right hand side matrix B.
300: *> On exit,
301: *> if EQUED = 'N', B is not modified;
302: *> if TRANS = 'N' and EQUED = 'R' or 'B', B is overwritten by
303: *> diag(R)*B;
304: *> if TRANS = 'T' or 'C' and EQUED = 'C' or 'B', B is
305: *> overwritten by diag(C)*B.
306: *> \endverbatim
307: *>
308: *> \param[in] LDB
309: *> \verbatim
310: *> LDB is INTEGER
311: *> The leading dimension of the array B. LDB >= max(1,N).
312: *> \endverbatim
313: *>
314: *> \param[out] X
315: *> \verbatim
316: *> X is DOUBLE PRECISION array, dimension (LDX,NRHS)
317: *> If INFO = 0, the N-by-NRHS solution matrix X to the original
318: *> system of equations. Note that A and B are modified on exit
319: *> if EQUED .ne. 'N', and the solution to the equilibrated system is
320: *> inv(diag(C))*X if TRANS = 'N' and EQUED = 'C' or 'B', or
321: *> inv(diag(R))*X if TRANS = 'T' or 'C' and EQUED = 'R' or 'B'.
322: *> \endverbatim
323: *>
324: *> \param[in] LDX
325: *> \verbatim
326: *> LDX is INTEGER
327: *> The leading dimension of the array X. LDX >= max(1,N).
328: *> \endverbatim
329: *>
330: *> \param[out] RCOND
331: *> \verbatim
332: *> RCOND is DOUBLE PRECISION
333: *> Reciprocal scaled condition number. This is an estimate of the
334: *> reciprocal Skeel condition number of the matrix A after
335: *> equilibration (if done). If this is less than the machine
336: *> precision (in particular, if it is zero), the matrix is singular
337: *> to working precision. Note that the error may still be small even
338: *> if this number is very small and the matrix appears ill-
339: *> conditioned.
340: *> \endverbatim
341: *>
342: *> \param[out] RPVGRW
343: *> \verbatim
344: *> RPVGRW is DOUBLE PRECISION
345: *> Reciprocal pivot growth. On exit, this contains the reciprocal
346: *> pivot growth factor norm(A)/norm(U). The "max absolute element"
347: *> norm is used. If this is much less than 1, then the stability of
348: *> the LU factorization of the (equilibrated) matrix A could be poor.
349: *> This also means that the solution X, estimated condition numbers,
350: *> and error bounds could be unreliable. If factorization fails with
351: *> 0<INFO<=N, then this contains the reciprocal pivot growth factor
352: *> for the leading INFO columns of A. In DGESVX, this quantity is
353: *> returned in WORK(1).
354: *> \endverbatim
355: *>
356: *> \param[out] BERR
357: *> \verbatim
358: *> BERR is DOUBLE PRECISION array, dimension (NRHS)
359: *> Componentwise relative backward error. This is the
360: *> componentwise relative backward error of each solution vector X(j)
361: *> (i.e., the smallest relative change in any element of A or B that
362: *> makes X(j) an exact solution).
363: *> \endverbatim
364: *>
365: *> \param[in] N_ERR_BNDS
366: *> \verbatim
367: *> N_ERR_BNDS is INTEGER
368: *> Number of error bounds to return for each right hand side
369: *> and each type (normwise or componentwise). See ERR_BNDS_NORM and
370: *> ERR_BNDS_COMP below.
371: *> \endverbatim
372: *>
373: *> \param[out] ERR_BNDS_NORM
374: *> \verbatim
375: *> ERR_BNDS_NORM is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS)
376: *> For each right-hand side, this array contains information about
377: *> various error bounds and condition numbers corresponding to the
378: *> normwise relative error, which is defined as follows:
379: *>
380: *> Normwise relative error in the ith solution vector:
381: *> max_j (abs(XTRUE(j,i) - X(j,i)))
382: *> ------------------------------
383: *> max_j abs(X(j,i))
384: *>
385: *> The array is indexed by the type of error information as described
386: *> below. There currently are up to three pieces of information
387: *> returned.
388: *>
389: *> The first index in ERR_BNDS_NORM(i,:) corresponds to the ith
390: *> right-hand side.
391: *>
392: *> The second index in ERR_BNDS_NORM(:,err) contains the following
393: *> three fields:
394: *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
395: *> reciprocal condition number is less than the threshold
396: *> sqrt(n) * dlamch('Epsilon').
397: *>
398: *> err = 2 "Guaranteed" error bound: The estimated forward error,
399: *> almost certainly within a factor of 10 of the true error
400: *> so long as the next entry is greater than the threshold
401: *> sqrt(n) * dlamch('Epsilon'). This error bound should only
402: *> be trusted if the previous boolean is true.
403: *>
404: *> err = 3 Reciprocal condition number: Estimated normwise
405: *> reciprocal condition number. Compared with the threshold
406: *> sqrt(n) * dlamch('Epsilon') to determine if the error
407: *> estimate is "guaranteed". These reciprocal condition
408: *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
409: *> appropriately scaled matrix Z.
410: *> Let Z = S*A, where S scales each row by a power of the
411: *> radix so all absolute row sums of Z are approximately 1.
412: *>
413: *> See Lapack Working Note 165 for further details and extra
414: *> cautions.
415: *> \endverbatim
416: *>
417: *> \param[out] ERR_BNDS_COMP
418: *> \verbatim
419: *> ERR_BNDS_COMP is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS)
420: *> For each right-hand side, this array contains information about
421: *> various error bounds and condition numbers corresponding to the
422: *> componentwise relative error, which is defined as follows:
423: *>
424: *> Componentwise relative error in the ith solution vector:
425: *> abs(XTRUE(j,i) - X(j,i))
426: *> max_j ----------------------
427: *> abs(X(j,i))
428: *>
429: *> The array is indexed by the right-hand side i (on which the
430: *> componentwise relative error depends), and the type of error
431: *> information as described below. There currently are up to three
432: *> pieces of information returned for each right-hand side. If
433: *> componentwise accuracy is not requested (PARAMS(3) = 0.0), then
434: *> ERR_BNDS_COMP is not accessed. If N_ERR_BNDS < 3, then at most
435: *> the first (:,N_ERR_BNDS) entries are returned.
436: *>
437: *> The first index in ERR_BNDS_COMP(i,:) corresponds to the ith
438: *> right-hand side.
439: *>
440: *> The second index in ERR_BNDS_COMP(:,err) contains the following
441: *> three fields:
442: *> err = 1 "Trust/don't trust" boolean. Trust the answer if the
443: *> reciprocal condition number is less than the threshold
444: *> sqrt(n) * dlamch('Epsilon').
445: *>
446: *> err = 2 "Guaranteed" error bound: The estimated forward error,
447: *> almost certainly within a factor of 10 of the true error
448: *> so long as the next entry is greater than the threshold
449: *> sqrt(n) * dlamch('Epsilon'). This error bound should only
450: *> be trusted if the previous boolean is true.
451: *>
452: *> err = 3 Reciprocal condition number: Estimated componentwise
453: *> reciprocal condition number. Compared with the threshold
454: *> sqrt(n) * dlamch('Epsilon') to determine if the error
455: *> estimate is "guaranteed". These reciprocal condition
456: *> numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some
457: *> appropriately scaled matrix Z.
458: *> Let Z = S*(A*diag(x)), where x is the solution for the
459: *> current right-hand side and S scales each row of
460: *> A*diag(x) by a power of the radix so all absolute row
461: *> sums of Z are approximately 1.
462: *>
463: *> See Lapack Working Note 165 for further details and extra
464: *> cautions.
465: *> \endverbatim
466: *>
467: *> \param[in] NPARAMS
468: *> \verbatim
469: *> NPARAMS is INTEGER
470: *> Specifies the number of parameters set in PARAMS. If <= 0, the
471: *> PARAMS array is never referenced and default values are used.
472: *> \endverbatim
473: *>
474: *> \param[in,out] PARAMS
475: *> \verbatim
476: *> PARAMS is DOUBLE PRECISION array, dimension (NPARAMS)
477: *> Specifies algorithm parameters. If an entry is < 0.0, then
478: *> that entry will be filled with default value used for that
479: *> parameter. Only positions up to NPARAMS are accessed; defaults
480: *> are used for higher-numbered parameters.
481: *>
482: *> PARAMS(LA_LINRX_ITREF_I = 1) : Whether to perform iterative
483: *> refinement or not.
484: *> Default: 1.0D+0
485: *> = 0.0: No refinement is performed, and no error bounds are
486: *> computed.
487: *> = 1.0: Use the extra-precise refinement algorithm.
488: *> (other values are reserved for future use)
489: *>
490: *> PARAMS(LA_LINRX_ITHRESH_I = 2) : Maximum number of residual
491: *> computations allowed for refinement.
492: *> Default: 10
493: *> Aggressive: Set to 100 to permit convergence using approximate
494: *> factorizations or factorizations other than LU. If
495: *> the factorization uses a technique other than
496: *> Gaussian elimination, the guarantees in
497: *> err_bnds_norm and err_bnds_comp may no longer be
498: *> trustworthy.
499: *>
500: *> PARAMS(LA_LINRX_CWISE_I = 3) : Flag determining if the code
501: *> will attempt to find a solution with small componentwise
502: *> relative error in the double-precision algorithm. Positive
503: *> is true, 0.0 is false.
504: *> Default: 1.0 (attempt componentwise convergence)
505: *> \endverbatim
506: *>
507: *> \param[out] WORK
508: *> \verbatim
509: *> WORK is DOUBLE PRECISION array, dimension (4*N)
510: *> \endverbatim
511: *>
512: *> \param[out] IWORK
513: *> \verbatim
514: *> IWORK is INTEGER array, dimension (N)
515: *> \endverbatim
516: *>
517: *> \param[out] INFO
518: *> \verbatim
519: *> INFO is INTEGER
520: *> = 0: Successful exit. The solution to every right-hand side is
521: *> guaranteed.
522: *> < 0: If INFO = -i, the i-th argument had an illegal value
523: *> > 0 and <= N: U(INFO,INFO) is exactly zero. The factorization
524: *> has been completed, but the factor U is exactly singular, so
525: *> the solution and error bounds could not be computed. RCOND = 0
526: *> is returned.
527: *> = N+J: The solution corresponding to the Jth right-hand side is
528: *> not guaranteed. The solutions corresponding to other right-
529: *> hand sides K with K > J may not be guaranteed as well, but
530: *> only the first such right-hand side is reported. If a small
531: *> componentwise error is not requested (PARAMS(3) = 0.0) then
532: *> the Jth right-hand side is the first with a normwise error
533: *> bound that is not guaranteed (the smallest J such
534: *> that ERR_BNDS_NORM(J,1) = 0.0). By default (PARAMS(3) = 1.0)
535: *> the Jth right-hand side is the first with either a normwise or
536: *> componentwise error bound that is not guaranteed (the smallest
537: *> J such that either ERR_BNDS_NORM(J,1) = 0.0 or
538: *> ERR_BNDS_COMP(J,1) = 0.0). See the definition of
539: *> ERR_BNDS_NORM(:,1) and ERR_BNDS_COMP(:,1). To get information
540: *> about all of the right-hand sides check ERR_BNDS_NORM or
541: *> ERR_BNDS_COMP.
542: *> \endverbatim
543: *
544: * Authors:
545: * ========
546: *
547: *> \author Univ. of Tennessee
548: *> \author Univ. of California Berkeley
549: *> \author Univ. of Colorado Denver
550: *> \author NAG Ltd.
551: *
552: *> \ingroup doubleGBsolve
553: *
554: * =====================================================================
555: SUBROUTINE DGBSVXX( FACT, TRANS, N, KL, KU, NRHS, AB, LDAB, AFB,
556: $ LDAFB, IPIV, EQUED, R, C, B, LDB, X, LDX,
557: $ RCOND, RPVGRW, BERR, N_ERR_BNDS,
558: $ ERR_BNDS_NORM, ERR_BNDS_COMP, NPARAMS, PARAMS,
559: $ WORK, IWORK, INFO )
560: *
561: * -- LAPACK driver routine --
562: * -- LAPACK is a software package provided by Univ. of Tennessee, --
563: * -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..--
564: *
565: * .. Scalar Arguments ..
566: CHARACTER EQUED, FACT, TRANS
567: INTEGER INFO, LDAB, LDAFB, LDB, LDX, N, NRHS, NPARAMS,
568: $ N_ERR_BNDS, KL, KU
569: DOUBLE PRECISION RCOND, RPVGRW
570: * ..
571: * .. Array Arguments ..
572: INTEGER IPIV( * ), IWORK( * )
573: DOUBLE PRECISION AB( LDAB, * ), AFB( LDAFB, * ), B( LDB, * ),
574: $ X( LDX , * ),WORK( * )
575: DOUBLE PRECISION R( * ), C( * ), PARAMS( * ), BERR( * ),
576: $ ERR_BNDS_NORM( NRHS, * ),
577: $ ERR_BNDS_COMP( NRHS, * )
578: * ..
579: *
580: * ==================================================================
581: *
582: * .. Parameters ..
583: DOUBLE PRECISION ZERO, ONE
584: PARAMETER ( ZERO = 0.0D+0, ONE = 1.0D+0 )
585: INTEGER FINAL_NRM_ERR_I, FINAL_CMP_ERR_I, BERR_I
586: INTEGER RCOND_I, NRM_RCOND_I, NRM_ERR_I, CMP_RCOND_I
587: INTEGER CMP_ERR_I, PIV_GROWTH_I
588: PARAMETER ( FINAL_NRM_ERR_I = 1, FINAL_CMP_ERR_I = 2,
589: $ BERR_I = 3 )
590: PARAMETER ( RCOND_I = 4, NRM_RCOND_I = 5, NRM_ERR_I = 6 )
591: PARAMETER ( CMP_RCOND_I = 7, CMP_ERR_I = 8,
592: $ PIV_GROWTH_I = 9 )
593: * ..
594: * .. Local Scalars ..
595: LOGICAL COLEQU, EQUIL, NOFACT, NOTRAN, ROWEQU
596: INTEGER INFEQU, I, J
597: DOUBLE PRECISION AMAX, BIGNUM, COLCND, RCMAX, RCMIN,
598: $ ROWCND, SMLNUM
599: * ..
600: * .. External Functions ..
601: EXTERNAL LSAME, DLAMCH, DLA_GBRPVGRW
602: LOGICAL LSAME
603: DOUBLE PRECISION DLAMCH, DLA_GBRPVGRW
604: * ..
605: * .. External Subroutines ..
606: EXTERNAL DGBEQUB, DGBTRF, DGBTRS, DLACPY, DLAQGB,
607: $ XERBLA, DLASCL2, DGBRFSX
608: * ..
609: * .. Intrinsic Functions ..
610: INTRINSIC MAX, MIN
611: * ..
612: * .. Executable Statements ..
613: *
614: INFO = 0
615: NOFACT = LSAME( FACT, 'N' )
616: EQUIL = LSAME( FACT, 'E' )
617: NOTRAN = LSAME( TRANS, 'N' )
618: SMLNUM = DLAMCH( 'Safe minimum' )
619: BIGNUM = ONE / SMLNUM
620: IF( NOFACT .OR. EQUIL ) THEN
621: EQUED = 'N'
622: ROWEQU = .FALSE.
623: COLEQU = .FALSE.
624: ELSE
625: ROWEQU = LSAME( EQUED, 'R' ) .OR. LSAME( EQUED, 'B' )
626: COLEQU = LSAME( EQUED, 'C' ) .OR. LSAME( EQUED, 'B' )
627: END IF
628: *
629: * Default is failure. If an input parameter is wrong or
630: * factorization fails, make everything look horrible. Only the
631: * pivot growth is set here, the rest is initialized in DGBRFSX.
632: *
633: RPVGRW = ZERO
634: *
635: * Test the input parameters. PARAMS is not tested until DGBRFSX.
636: *
637: IF( .NOT.NOFACT .AND. .NOT.EQUIL .AND. .NOT.
638: $ LSAME( FACT, 'F' ) ) THEN
639: INFO = -1
640: ELSE IF( .NOT.NOTRAN .AND. .NOT.LSAME( TRANS, 'T' ) .AND. .NOT.
641: $ LSAME( TRANS, 'C' ) ) THEN
642: INFO = -2
643: ELSE IF( N.LT.0 ) THEN
644: INFO = -3
645: ELSE IF( KL.LT.0 ) THEN
646: INFO = -4
647: ELSE IF( KU.LT.0 ) THEN
648: INFO = -5
649: ELSE IF( NRHS.LT.0 ) THEN
650: INFO = -6
651: ELSE IF( LDAB.LT.KL+KU+1 ) THEN
652: INFO = -8
653: ELSE IF( LDAFB.LT.2*KL+KU+1 ) THEN
654: INFO = -10
655: ELSE IF( LSAME( FACT, 'F' ) .AND. .NOT.
656: $ ( ROWEQU .OR. COLEQU .OR. LSAME( EQUED, 'N' ) ) ) THEN
657: INFO = -12
658: ELSE
659: IF( ROWEQU ) THEN
660: RCMIN = BIGNUM
661: RCMAX = ZERO
662: DO 10 J = 1, N
663: RCMIN = MIN( RCMIN, R( J ) )
664: RCMAX = MAX( RCMAX, R( J ) )
665: 10 CONTINUE
666: IF( RCMIN.LE.ZERO ) THEN
667: INFO = -13
668: ELSE IF( N.GT.0 ) THEN
669: ROWCND = MAX( RCMIN, SMLNUM ) / MIN( RCMAX, BIGNUM )
670: ELSE
671: ROWCND = ONE
672: END IF
673: END IF
674: IF( COLEQU .AND. INFO.EQ.0 ) THEN
675: RCMIN = BIGNUM
676: RCMAX = ZERO
677: DO 20 J = 1, N
678: RCMIN = MIN( RCMIN, C( J ) )
679: RCMAX = MAX( RCMAX, C( J ) )
680: 20 CONTINUE
681: IF( RCMIN.LE.ZERO ) THEN
682: INFO = -14
683: ELSE IF( N.GT.0 ) THEN
684: COLCND = MAX( RCMIN, SMLNUM ) / MIN( RCMAX, BIGNUM )
685: ELSE
686: COLCND = ONE
687: END IF
688: END IF
689: IF( INFO.EQ.0 ) THEN
690: IF( LDB.LT.MAX( 1, N ) ) THEN
691: INFO = -15
692: ELSE IF( LDX.LT.MAX( 1, N ) ) THEN
693: INFO = -16
694: END IF
695: END IF
696: END IF
697: *
698: IF( INFO.NE.0 ) THEN
699: CALL XERBLA( 'DGBSVXX', -INFO )
700: RETURN
701: END IF
702: *
703: IF( EQUIL ) THEN
704: *
705: * Compute row and column scalings to equilibrate the matrix A.
706: *
707: CALL DGBEQUB( N, N, KL, KU, AB, LDAB, R, C, ROWCND, COLCND,
708: $ AMAX, INFEQU )
709: IF( INFEQU.EQ.0 ) THEN
710: *
711: * Equilibrate the matrix.
712: *
713: CALL DLAQGB( N, N, KL, KU, AB, LDAB, R, C, ROWCND, COLCND,
714: $ AMAX, EQUED )
715: ROWEQU = LSAME( EQUED, 'R' ) .OR. LSAME( EQUED, 'B' )
716: COLEQU = LSAME( EQUED, 'C' ) .OR. LSAME( EQUED, 'B' )
717: END IF
718: *
719: * If the scaling factors are not applied, set them to 1.0.
720: *
721: IF ( .NOT.ROWEQU ) THEN
722: DO J = 1, N
723: R( J ) = 1.0D+0
724: END DO
725: END IF
726: IF ( .NOT.COLEQU ) THEN
727: DO J = 1, N
728: C( J ) = 1.0D+0
729: END DO
730: END IF
731: END IF
732: *
733: * Scale the right hand side.
734: *
735: IF( NOTRAN ) THEN
736: IF( ROWEQU ) CALL DLASCL2(N, NRHS, R, B, LDB)
737: ELSE
738: IF( COLEQU ) CALL DLASCL2(N, NRHS, C, B, LDB)
739: END IF
740: *
741: IF( NOFACT .OR. EQUIL ) THEN
742: *
743: * Compute the LU factorization of A.
744: *
745: DO 40, J = 1, N
746: DO 30, I = KL+1, 2*KL+KU+1
747: AFB( I, J ) = AB( I-KL, J )
748: 30 CONTINUE
749: 40 CONTINUE
750: CALL DGBTRF( N, N, KL, KU, AFB, LDAFB, IPIV, INFO )
751: *
752: * Return if INFO is non-zero.
753: *
754: IF( INFO.GT.0 ) THEN
755: *
756: * Pivot in column INFO is exactly 0
757: * Compute the reciprocal pivot growth factor of the
758: * leading rank-deficient INFO columns of A.
759: *
760: RPVGRW = DLA_GBRPVGRW( N, KL, KU, INFO, AB, LDAB, AFB,
761: $ LDAFB )
762: RETURN
763: END IF
764: END IF
765: *
766: * Compute the reciprocal pivot growth factor RPVGRW.
767: *
768: RPVGRW = DLA_GBRPVGRW( N, KL, KU, N, AB, LDAB, AFB, LDAFB )
769: *
770: * Compute the solution matrix X.
771: *
772: CALL DLACPY( 'Full', N, NRHS, B, LDB, X, LDX )
773: CALL DGBTRS( TRANS, N, KL, KU, NRHS, AFB, LDAFB, IPIV, X, LDX,
774: $ INFO )
775: *
776: * Use iterative refinement to improve the computed solution and
777: * compute error bounds and backward error estimates for it.
778: *
779: CALL DGBRFSX( TRANS, EQUED, N, KL, KU, NRHS, AB, LDAB, AFB, LDAFB,
780: $ IPIV, R, C, B, LDB, X, LDX, RCOND, BERR,
781: $ N_ERR_BNDS, ERR_BNDS_NORM, ERR_BNDS_COMP, NPARAMS, PARAMS,
782: $ WORK, IWORK, INFO )
783: *
784: * Scale solutions.
785: *
786: IF ( COLEQU .AND. NOTRAN ) THEN
787: CALL DLASCL2 ( N, NRHS, C, X, LDX )
788: ELSE IF ( ROWEQU .AND. .NOT.NOTRAN ) THEN
789: CALL DLASCL2 ( N, NRHS, R, X, LDX )
790: END IF
791: *
792: RETURN
793: *
794: * End of DGBSVXX
795: *
796: END
CVSweb interface <joel.bertrand@systella.fr>