14 typedef long long BLASLONG;
15 typedef unsigned long long BLASULONG;
17 typedef long BLASLONG;
18 typedef unsigned long BLASULONG;
22 typedef BLASLONG blasint;
24 #define blasabs(x) llabs(x)
26 #define blasabs(x) labs(x)
30 #define blasabs(x) abs(x)
33 typedef blasint integer;
35 typedef unsigned int uinteger;
36 typedef char *address;
37 typedef short int shortint;
39 typedef double doublereal;
40 typedef struct { real r, i; } complex;
41 typedef struct { doublereal r, i; } doublecomplex;
43 static inline _Fcomplex Cf(complex *z) {_Fcomplex zz={z->r , z->i}; return zz;}
44 static inline _Dcomplex Cd(doublecomplex *z) {_Dcomplex zz={z->r , z->i};return zz;}
45 static inline _Fcomplex * _pCf(complex *z) {return (_Fcomplex*)z;}
46 static inline _Dcomplex * _pCd(doublecomplex *z) {return (_Dcomplex*)z;}
48 static inline _Complex float Cf(complex *z) {return z->r + z->i*_Complex_I;}
49 static inline _Complex double Cd(doublecomplex *z) {return z->r + z->i*_Complex_I;}
50 static inline _Complex float * _pCf(complex *z) {return (_Complex float*)z;}
51 static inline _Complex double * _pCd(doublecomplex *z) {return (_Complex double*)z;}
53 #define pCf(z) (*_pCf(z))
54 #define pCd(z) (*_pCd(z))
56 typedef short int shortlogical;
57 typedef char logical1;
58 typedef char integer1;
63 /* Extern is for use with -E */
74 /*external read, write*/
83 /*internal read, write*/
113 /*rewind, backspace, endfile*/
125 ftnint *inex; /*parameters in standard's order*/
151 union Multitype { /* for multiple entry points */
162 typedef union Multitype Multitype;
164 struct Vardesc { /* for Namelist */
170 typedef struct Vardesc Vardesc;
177 typedef struct Namelist Namelist;
179 #define abs(x) ((x) >= 0 ? (x) : -(x))
180 #define dabs(x) (fabs(x))
181 #define f2cmin(a,b) ((a) <= (b) ? (a) : (b))
182 #define f2cmax(a,b) ((a) >= (b) ? (a) : (b))
183 #define dmin(a,b) (f2cmin(a,b))
184 #define dmax(a,b) (f2cmax(a,b))
185 #define bit_test(a,b) ((a) >> (b) & 1)
186 #define bit_clear(a,b) ((a) & ~((uinteger)1 << (b)))
187 #define bit_set(a,b) ((a) | ((uinteger)1 << (b)))
189 #define abort_() { sig_die("Fortran abort routine called", 1); }
190 #define c_abs(z) (cabsf(Cf(z)))
191 #define c_cos(R,Z) { pCf(R)=ccos(Cf(Z)); }
193 #define c_div(c, a, b) {Cf(c)._Val[0] = (Cf(a)._Val[0]/Cf(b)._Val[0]); Cf(c)._Val[1]=(Cf(a)._Val[1]/Cf(b)._Val[1]);}
194 #define z_div(c, a, b) {Cd(c)._Val[0] = (Cd(a)._Val[0]/Cd(b)._Val[0]); Cd(c)._Val[1]=(Cd(a)._Val[1]/df(b)._Val[1]);}
196 #define c_div(c, a, b) {pCf(c) = Cf(a)/Cf(b);}
197 #define z_div(c, a, b) {pCd(c) = Cd(a)/Cd(b);}
199 #define c_exp(R, Z) {pCf(R) = cexpf(Cf(Z));}
200 #define c_log(R, Z) {pCf(R) = clogf(Cf(Z));}
201 #define c_sin(R, Z) {pCf(R) = csinf(Cf(Z));}
202 //#define c_sqrt(R, Z) {*(R) = csqrtf(Cf(Z));}
203 #define c_sqrt(R, Z) {pCf(R) = csqrtf(Cf(Z));}
204 #define d_abs(x) (fabs(*(x)))
205 #define d_acos(x) (acos(*(x)))
206 #define d_asin(x) (asin(*(x)))
207 #define d_atan(x) (atan(*(x)))
208 #define d_atn2(x, y) (atan2(*(x),*(y)))
209 #define d_cnjg(R, Z) { pCd(R) = conj(Cd(Z)); }
210 #define r_cnjg(R, Z) { pCf(R) = conjf(Cf(Z)); }
211 #define d_cos(x) (cos(*(x)))
212 #define d_cosh(x) (cosh(*(x)))
213 #define d_dim(__a, __b) ( *(__a) > *(__b) ? *(__a) - *(__b) : 0.0 )
214 #define d_exp(x) (exp(*(x)))
215 #define d_imag(z) (cimag(Cd(z)))
216 #define r_imag(z) (cimagf(Cf(z)))
217 #define d_int(__x) (*(__x)>0 ? floor(*(__x)) : -floor(- *(__x)))
218 #define r_int(__x) (*(__x)>0 ? floor(*(__x)) : -floor(- *(__x)))
219 #define d_lg10(x) ( 0.43429448190325182765 * log(*(x)) )
220 #define r_lg10(x) ( 0.43429448190325182765 * log(*(x)) )
221 #define d_log(x) (log(*(x)))
222 #define d_mod(x, y) (fmod(*(x), *(y)))
223 #define u_nint(__x) ((__x)>=0 ? floor((__x) + .5) : -floor(.5 - (__x)))
224 #define d_nint(x) u_nint(*(x))
225 #define u_sign(__a,__b) ((__b) >= 0 ? ((__a) >= 0 ? (__a) : -(__a)) : -((__a) >= 0 ? (__a) : -(__a)))
226 #define d_sign(a,b) u_sign(*(a),*(b))
227 #define r_sign(a,b) u_sign(*(a),*(b))
228 #define d_sin(x) (sin(*(x)))
229 #define d_sinh(x) (sinh(*(x)))
230 #define d_sqrt(x) (sqrt(*(x)))
231 #define d_tan(x) (tan(*(x)))
232 #define d_tanh(x) (tanh(*(x)))
233 #define i_abs(x) abs(*(x))
234 #define i_dnnt(x) ((integer)u_nint(*(x)))
235 #define i_len(s, n) (n)
236 #define i_nint(x) ((integer)u_nint(*(x)))
237 #define i_sign(a,b) ((integer)u_sign((integer)*(a),(integer)*(b)))
238 #define pow_dd(ap, bp) ( pow(*(ap), *(bp)))
239 #define pow_si(B,E) spow_ui(*(B),*(E))
240 #define pow_ri(B,E) spow_ui(*(B),*(E))
241 #define pow_di(B,E) dpow_ui(*(B),*(E))
242 #define pow_zi(p, a, b) {pCd(p) = zpow_ui(Cd(a), *(b));}
243 #define pow_ci(p, a, b) {pCf(p) = cpow_ui(Cf(a), *(b));}
244 #define pow_zz(R,A,B) {pCd(R) = cpow(Cd(A),*(B));}
245 #define s_cat(lpp, rpp, rnp, np, llp) { ftnlen i, nc, ll; char *f__rp, *lp; ll = (llp); lp = (lpp); for(i=0; i < (int)*(np); ++i) { nc = ll; if((rnp)[i] < nc) nc = (rnp)[i]; ll -= nc; f__rp = (rpp)[i]; while(--nc >= 0) *lp++ = *(f__rp)++; } while(--ll >= 0) *lp++ = ' '; }
246 #define s_cmp(a,b,c,d) ((integer)strncmp((a),(b),f2cmin((c),(d))))
247 #define s_copy(A,B,C,D) { int __i,__m; for (__i=0, __m=f2cmin((C),(D)); __i<__m && (B)[__i] != 0; ++__i) (A)[__i] = (B)[__i]; }
248 #define sig_die(s, kill) { exit(1); }
249 #define s_stop(s, n) {exit(0);}
250 static char junk[] = "\n@(#)LIBF77 VERSION 19990503\n";
251 #define z_abs(z) (cabs(Cd(z)))
252 #define z_exp(R, Z) {pCd(R) = cexp(Cd(Z));}
253 #define z_sqrt(R, Z) {pCd(R) = csqrt(Cd(Z));}
254 #define myexit_() break;
255 #define mycycle() continue;
256 #define myceiling(w) {ceil(w)}
257 #define myhuge(w) {HUGE_VAL}
258 //#define mymaxloc_(w,s,e,n) {if (sizeof(*(w)) == sizeof(double)) dmaxloc_((w),*(s),*(e),n); else dmaxloc_((w),*(s),*(e),n);}
259 #define mymaxloc(w,s,e,n) {dmaxloc_(w,*(s),*(e),n)}
261 /* procedure parameter types for -A and -C++ */
263 #define F2C_proc_par_types 1
265 typedef logical (*L_fp)(...);
267 typedef logical (*L_fp)();
270 static float spow_ui(float x, integer n) {
271 float pow=1.0; unsigned long int u;
273 if(n < 0) n = -n, x = 1/x;
282 static double dpow_ui(double x, integer n) {
283 double pow=1.0; unsigned long int u;
285 if(n < 0) n = -n, x = 1/x;
295 static _Fcomplex cpow_ui(complex x, integer n) {
296 complex pow={1.0,0.0}; unsigned long int u;
298 if(n < 0) n = -n, x.r = 1/x.r, x.i=1/x.i;
300 if(u & 01) pow.r *= x.r, pow.i *= x.i;
301 if(u >>= 1) x.r *= x.r, x.i *= x.i;
305 _Fcomplex p={pow.r, pow.i};
309 static _Complex float cpow_ui(_Complex float x, integer n) {
310 _Complex float pow=1.0; unsigned long int u;
312 if(n < 0) n = -n, x = 1/x;
323 static _Dcomplex zpow_ui(_Dcomplex x, integer n) {
324 _Dcomplex pow={1.0,0.0}; unsigned long int u;
326 if(n < 0) n = -n, x._Val[0] = 1/x._Val[0], x._Val[1] =1/x._Val[1];
328 if(u & 01) pow._Val[0] *= x._Val[0], pow._Val[1] *= x._Val[1];
329 if(u >>= 1) x._Val[0] *= x._Val[0], x._Val[1] *= x._Val[1];
333 _Dcomplex p = {pow._Val[0], pow._Val[1]};
337 static _Complex double zpow_ui(_Complex double x, integer n) {
338 _Complex double pow=1.0; unsigned long int u;
340 if(n < 0) n = -n, x = 1/x;
350 static integer pow_ii(integer x, integer n) {
351 integer pow; unsigned long int u;
353 if (n == 0 || x == 1) pow = 1;
354 else if (x != -1) pow = x == 0 ? 1/x : 0;
357 if ((n > 0) || !(n == 0 || x == 1 || x != -1)) {
367 static integer dmaxloc_(double *w, integer s, integer e, integer *n)
369 double m; integer i, mi;
370 for(m=w[s-1], mi=s, i=s+1; i<=e; i++)
371 if (w[i-1]>m) mi=i ,m=w[i-1];
374 static integer smaxloc_(float *w, integer s, integer e, integer *n)
376 float m; integer i, mi;
377 for(m=w[s-1], mi=s, i=s+1; i<=e; i++)
378 if (w[i-1]>m) mi=i ,m=w[i-1];
381 static inline void cdotc_(complex *z, integer *n_, complex *x, integer *incx_, complex *y, integer *incy_) {
382 integer n = *n_, incx = *incx_, incy = *incy_, i;
384 _Fcomplex zdotc = {0.0, 0.0};
385 if (incx == 1 && incy == 1) {
386 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
387 zdotc._Val[0] += conjf(Cf(&x[i]))._Val[0] * Cf(&y[i])._Val[0];
388 zdotc._Val[1] += conjf(Cf(&x[i]))._Val[1] * Cf(&y[i])._Val[1];
391 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
392 zdotc._Val[0] += conjf(Cf(&x[i*incx]))._Val[0] * Cf(&y[i*incy])._Val[0];
393 zdotc._Val[1] += conjf(Cf(&x[i*incx]))._Val[1] * Cf(&y[i*incy])._Val[1];
399 _Complex float zdotc = 0.0;
400 if (incx == 1 && incy == 1) {
401 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
402 zdotc += conjf(Cf(&x[i])) * Cf(&y[i]);
405 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
406 zdotc += conjf(Cf(&x[i*incx])) * Cf(&y[i*incy]);
412 static inline void zdotc_(doublecomplex *z, integer *n_, doublecomplex *x, integer *incx_, doublecomplex *y, integer *incy_) {
413 integer n = *n_, incx = *incx_, incy = *incy_, i;
415 _Dcomplex zdotc = {0.0, 0.0};
416 if (incx == 1 && incy == 1) {
417 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
418 zdotc._Val[0] += conj(Cd(&x[i]))._Val[0] * Cd(&y[i])._Val[0];
419 zdotc._Val[1] += conj(Cd(&x[i]))._Val[1] * Cd(&y[i])._Val[1];
422 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
423 zdotc._Val[0] += conj(Cd(&x[i*incx]))._Val[0] * Cd(&y[i*incy])._Val[0];
424 zdotc._Val[1] += conj(Cd(&x[i*incx]))._Val[1] * Cd(&y[i*incy])._Val[1];
430 _Complex double zdotc = 0.0;
431 if (incx == 1 && incy == 1) {
432 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
433 zdotc += conj(Cd(&x[i])) * Cd(&y[i]);
436 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
437 zdotc += conj(Cd(&x[i*incx])) * Cd(&y[i*incy]);
443 static inline void cdotu_(complex *z, integer *n_, complex *x, integer *incx_, complex *y, integer *incy_) {
444 integer n = *n_, incx = *incx_, incy = *incy_, i;
446 _Fcomplex zdotc = {0.0, 0.0};
447 if (incx == 1 && incy == 1) {
448 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
449 zdotc._Val[0] += Cf(&x[i])._Val[0] * Cf(&y[i])._Val[0];
450 zdotc._Val[1] += Cf(&x[i])._Val[1] * Cf(&y[i])._Val[1];
453 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
454 zdotc._Val[0] += Cf(&x[i*incx])._Val[0] * Cf(&y[i*incy])._Val[0];
455 zdotc._Val[1] += Cf(&x[i*incx])._Val[1] * Cf(&y[i*incy])._Val[1];
461 _Complex float zdotc = 0.0;
462 if (incx == 1 && incy == 1) {
463 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
464 zdotc += Cf(&x[i]) * Cf(&y[i]);
467 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
468 zdotc += Cf(&x[i*incx]) * Cf(&y[i*incy]);
474 static inline void zdotu_(doublecomplex *z, integer *n_, doublecomplex *x, integer *incx_, doublecomplex *y, integer *incy_) {
475 integer n = *n_, incx = *incx_, incy = *incy_, i;
477 _Dcomplex zdotc = {0.0, 0.0};
478 if (incx == 1 && incy == 1) {
479 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
480 zdotc._Val[0] += Cd(&x[i])._Val[0] * Cd(&y[i])._Val[0];
481 zdotc._Val[1] += Cd(&x[i])._Val[1] * Cd(&y[i])._Val[1];
484 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
485 zdotc._Val[0] += Cd(&x[i*incx])._Val[0] * Cd(&y[i*incy])._Val[0];
486 zdotc._Val[1] += Cd(&x[i*incx])._Val[1] * Cd(&y[i*incy])._Val[1];
492 _Complex double zdotc = 0.0;
493 if (incx == 1 && incy == 1) {
494 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
495 zdotc += Cd(&x[i]) * Cd(&y[i]);
498 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
499 zdotc += Cd(&x[i*incx]) * Cd(&y[i*incy]);
505 /* -- translated by f2c (version 20000121).
506 You must link the resulting object file with the libraries:
507 -lf2c -lm (in that order)
513 /* > \brief <b> DPOSVXX computes the solution to system of linear equations A * X = B for PO matrices</b> */
515 /* =========== DOCUMENTATION =========== */
517 /* Online html documentation available at */
518 /* http://www.netlib.org/lapack/explore-html/ */
521 /* > Download DPOSVXX + dependencies */
522 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dposvxx
525 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dposvxx
528 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dposvxx
536 /* SUBROUTINE DPOSVXX( FACT, UPLO, N, NRHS, A, LDA, AF, LDAF, EQUED, */
537 /* S, B, LDB, X, LDX, RCOND, RPVGRW, BERR, */
538 /* N_ERR_BNDS, ERR_BNDS_NORM, ERR_BNDS_COMP, */
539 /* NPARAMS, PARAMS, WORK, IWORK, INFO ) */
541 /* CHARACTER EQUED, FACT, UPLO */
542 /* INTEGER INFO, LDA, LDAF, LDB, LDX, N, NRHS, NPARAMS, */
544 /* DOUBLE PRECISION RCOND, RPVGRW */
545 /* INTEGER IWORK( * ) */
546 /* DOUBLE PRECISION A( LDA, * ), AF( LDAF, * ), B( LDB, * ), */
547 /* $ X( LDX, * ), WORK( * ) */
548 /* DOUBLE PRECISION S( * ), PARAMS( * ), BERR( * ), */
549 /* $ ERR_BNDS_NORM( NRHS, * ), */
550 /* $ ERR_BNDS_COMP( NRHS, * ) */
553 /* > \par Purpose: */
558 /* > DPOSVXX uses the Cholesky factorization A = U**T*U or A = L*L**T */
559 /* > to compute the solution to a double precision system of linear equations */
560 /* > A * X = B, where A is an N-by-N symmetric positive definite matrix */
561 /* > and X and B are N-by-NRHS matrices. */
563 /* > If requested, both normwise and maximum componentwise error bounds */
564 /* > are returned. DPOSVXX will return a solution with a tiny */
565 /* > guaranteed error (O(eps) where eps is the working machine */
566 /* > precision) unless the matrix is very ill-conditioned, in which */
567 /* > case a warning is returned. Relevant condition numbers also are */
568 /* > calculated and returned. */
570 /* > DPOSVXX accepts user-provided factorizations and equilibration */
571 /* > factors; see the definitions of the FACT and EQUED options. */
572 /* > Solving with refinement and using a factorization from a previous */
573 /* > DPOSVXX call will also produce a solution with either O(eps) */
574 /* > errors or warnings, but we cannot make that claim for general */
575 /* > user-provided factorizations and equilibration factors if they */
576 /* > differ from what DPOSVXX would itself produce. */
579 /* > \par Description: */
580 /* ================= */
584 /* > The following steps are performed: */
586 /* > 1. If FACT = 'E', double precision scaling factors are computed to equilibrate */
589 /* > diag(S)*A*diag(S) *inv(diag(S))*X = diag(S)*B */
591 /* > Whether or not the system will be equilibrated depends on the */
592 /* > scaling of the matrix A, but if equilibration is used, A is */
593 /* > overwritten by diag(S)*A*diag(S) and B by diag(S)*B. */
595 /* > 2. If FACT = 'N' or 'E', the Cholesky decomposition is used to */
596 /* > factor the matrix A (after equilibration if FACT = 'E') as */
597 /* > A = U**T* U, if UPLO = 'U', or */
598 /* > A = L * L**T, if UPLO = 'L', */
599 /* > where U is an upper triangular matrix and L is a lower triangular */
602 /* > 3. If the leading i-by-i principal minor is not positive definite, */
603 /* > then the routine returns with INFO = i. Otherwise, the factored */
604 /* > form of A is used to estimate the condition number of the matrix */
605 /* > A (see argument RCOND). If the reciprocal of the condition number */
606 /* > is less than machine precision, the routine still goes on to solve */
607 /* > for X and compute error bounds as described below. */
609 /* > 4. The system of equations is solved for X using the factored form */
612 /* > 5. By default (unless PARAMS(LA_LINRX_ITREF_I) is set to zero), */
613 /* > the routine will use iterative refinement to try to get a small */
614 /* > error and error bounds. Refinement calculates the residual to at */
615 /* > least twice the working precision. */
617 /* > 6. If equilibration was used, the matrix X is premultiplied by */
618 /* > diag(S) so that it solves the original system before */
619 /* > equilibration. */
626 /* > Some optional parameters are bundled in the PARAMS array. These */
627 /* > settings determine how refinement is performed, but often the */
628 /* > defaults are acceptable. If the defaults are acceptable, users */
629 /* > can pass NPARAMS = 0 which prevents the source code from accessing */
630 /* > the PARAMS argument. */
633 /* > \param[in] FACT */
635 /* > FACT is CHARACTER*1 */
636 /* > Specifies whether or not the factored form of the matrix A is */
637 /* > supplied on entry, and if not, whether the matrix A should be */
638 /* > equilibrated before it is factored. */
639 /* > = 'F': On entry, AF contains the factored form of A. */
640 /* > If EQUED is not 'N', the matrix A has been */
641 /* > equilibrated with scaling factors given by S. */
642 /* > A and AF are not modified. */
643 /* > = 'N': The matrix A will be copied to AF and factored. */
644 /* > = 'E': The matrix A will be equilibrated if necessary, then */
645 /* > copied to AF and factored. */
648 /* > \param[in] UPLO */
650 /* > UPLO is CHARACTER*1 */
651 /* > = 'U': Upper triangle of A is stored; */
652 /* > = 'L': Lower triangle of A is stored. */
658 /* > The number of linear equations, i.e., the order of the */
659 /* > matrix A. N >= 0. */
662 /* > \param[in] NRHS */
664 /* > NRHS is INTEGER */
665 /* > The number of right hand sides, i.e., the number of columns */
666 /* > of the matrices B and X. NRHS >= 0. */
669 /* > \param[in,out] A */
671 /* > A is DOUBLE PRECISION array, dimension (LDA,N) */
672 /* > On entry, the symmetric matrix A, except if FACT = 'F' and EQUED = */
673 /* > 'Y', then A must contain the equilibrated matrix */
674 /* > diag(S)*A*diag(S). If UPLO = 'U', the leading N-by-N upper */
675 /* > triangular part of A contains the upper triangular part of the */
676 /* > matrix A, and the strictly lower triangular part of A is not */
677 /* > referenced. If UPLO = 'L', the leading N-by-N lower triangular */
678 /* > part of A contains the lower triangular part of the matrix A, and */
679 /* > the strictly upper triangular part of A is not referenced. A is */
680 /* > not modified if FACT = 'F' or 'N', or if FACT = 'E' and EQUED = */
683 /* > On exit, if FACT = 'E' and EQUED = 'Y', A is overwritten by */
684 /* > diag(S)*A*diag(S). */
687 /* > \param[in] LDA */
689 /* > LDA is INTEGER */
690 /* > The leading dimension of the array A. LDA >= f2cmax(1,N). */
693 /* > \param[in,out] AF */
695 /* > AF is DOUBLE PRECISION array, dimension (LDAF,N) */
696 /* > If FACT = 'F', then AF is an input argument and on entry */
697 /* > contains the triangular factor U or L from the Cholesky */
698 /* > factorization A = U**T*U or A = L*L**T, in the same storage */
699 /* > format as A. If EQUED .ne. 'N', then AF is the factored */
700 /* > form of the equilibrated matrix diag(S)*A*diag(S). */
702 /* > If FACT = 'N', then AF is an output argument and on exit */
703 /* > returns the triangular factor U or L from the Cholesky */
704 /* > factorization A = U**T*U or A = L*L**T of the original */
707 /* > If FACT = 'E', then AF is an output argument and on exit */
708 /* > returns the triangular factor U or L from the Cholesky */
709 /* > factorization A = U**T*U or A = L*L**T of the equilibrated */
710 /* > matrix A (see the description of A for the form of the */
711 /* > equilibrated matrix). */
714 /* > \param[in] LDAF */
716 /* > LDAF is INTEGER */
717 /* > The leading dimension of the array AF. LDAF >= f2cmax(1,N). */
720 /* > \param[in,out] EQUED */
722 /* > EQUED is CHARACTER*1 */
723 /* > Specifies the form of equilibration that was done. */
724 /* > = 'N': No equilibration (always true if FACT = 'N'). */
725 /* > = 'Y': Both row and column equilibration, i.e., A has been */
726 /* > replaced by diag(S) * A * diag(S). */
727 /* > EQUED is an input argument if FACT = 'F'; otherwise, it is an */
728 /* > output argument. */
731 /* > \param[in,out] S */
733 /* > S is DOUBLE PRECISION array, dimension (N) */
734 /* > The row scale factors for A. If EQUED = 'Y', A is multiplied on */
735 /* > the left and right by diag(S). S is an input argument if FACT = */
736 /* > 'F'; otherwise, S is an output argument. If FACT = 'F' and EQUED */
737 /* > = 'Y', each element of S must be positive. If S is output, each */
738 /* > element of S is a power of the radix. If S is input, each element */
739 /* > of S should be a power of the radix to ensure a reliable solution */
740 /* > and error estimates. Scaling by powers of the radix does not cause */
741 /* > rounding errors unless the result underflows or overflows. */
742 /* > Rounding errors during scaling lead to refining with a matrix that */
743 /* > is not equivalent to the input matrix, producing error estimates */
744 /* > that may not be reliable. */
747 /* > \param[in,out] B */
749 /* > B is DOUBLE PRECISION array, dimension (LDB,NRHS) */
750 /* > On entry, the N-by-NRHS right hand side matrix B. */
752 /* > if EQUED = 'N', B is not modified; */
753 /* > if EQUED = 'Y', B is overwritten by diag(S)*B; */
756 /* > \param[in] LDB */
758 /* > LDB is INTEGER */
759 /* > The leading dimension of the array B. LDB >= f2cmax(1,N). */
762 /* > \param[out] X */
764 /* > X is DOUBLE PRECISION array, dimension (LDX,NRHS) */
765 /* > If INFO = 0, the N-by-NRHS solution matrix X to the original */
766 /* > system of equations. Note that A and B are modified on exit if */
767 /* > EQUED .ne. 'N', and the solution to the equilibrated system is */
768 /* > inv(diag(S))*X. */
771 /* > \param[in] LDX */
773 /* > LDX is INTEGER */
774 /* > The leading dimension of the array X. LDX >= f2cmax(1,N). */
777 /* > \param[out] RCOND */
779 /* > RCOND is DOUBLE PRECISION */
780 /* > Reciprocal scaled condition number. This is an estimate of the */
781 /* > reciprocal Skeel condition number of the matrix A after */
782 /* > equilibration (if done). If this is less than the machine */
783 /* > precision (in particular, if it is zero), the matrix is singular */
784 /* > to working precision. Note that the error may still be small even */
785 /* > if this number is very small and the matrix appears ill- */
789 /* > \param[out] RPVGRW */
791 /* > RPVGRW is DOUBLE PRECISION */
792 /* > Reciprocal pivot growth. On exit, this contains the reciprocal */
793 /* > pivot growth factor norm(A)/norm(U). The "f2cmax absolute element" */
794 /* > norm is used. If this is much less than 1, then the stability of */
795 /* > the LU factorization of the (equilibrated) matrix A could be poor. */
796 /* > This also means that the solution X, estimated condition numbers, */
797 /* > and error bounds could be unreliable. If factorization fails with */
798 /* > 0<INFO<=N, then this contains the reciprocal pivot growth factor */
799 /* > for the leading INFO columns of A. */
802 /* > \param[out] BERR */
804 /* > BERR is DOUBLE PRECISION array, dimension (NRHS) */
805 /* > Componentwise relative backward error. This is the */
806 /* > componentwise relative backward error of each solution vector X(j) */
807 /* > (i.e., the smallest relative change in any element of A or B that */
808 /* > makes X(j) an exact solution). */
811 /* > \param[in] N_ERR_BNDS */
813 /* > N_ERR_BNDS is INTEGER */
814 /* > Number of error bounds to return for each right hand side */
815 /* > and each type (normwise or componentwise). See ERR_BNDS_NORM and */
816 /* > ERR_BNDS_COMP below. */
819 /* > \param[out] ERR_BNDS_NORM */
821 /* > ERR_BNDS_NORM is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS) */
822 /* > For each right-hand side, this array contains information about */
823 /* > various error bounds and condition numbers corresponding to the */
824 /* > normwise relative error, which is defined as follows: */
826 /* > Normwise relative error in the ith solution vector: */
827 /* > max_j (abs(XTRUE(j,i) - X(j,i))) */
828 /* > ------------------------------ */
829 /* > max_j abs(X(j,i)) */
831 /* > The array is indexed by the type of error information as described */
832 /* > below. There currently are up to three pieces of information */
835 /* > The first index in ERR_BNDS_NORM(i,:) corresponds to the ith */
836 /* > right-hand side. */
838 /* > The second index in ERR_BNDS_NORM(:,err) contains the following */
839 /* > three fields: */
840 /* > err = 1 "Trust/don't trust" boolean. Trust the answer if the */
841 /* > reciprocal condition number is less than the threshold */
842 /* > sqrt(n) * dlamch('Epsilon'). */
844 /* > err = 2 "Guaranteed" error bound: The estimated forward error, */
845 /* > almost certainly within a factor of 10 of the true error */
846 /* > so long as the next entry is greater than the threshold */
847 /* > sqrt(n) * dlamch('Epsilon'). This error bound should only */
848 /* > be trusted if the previous boolean is true. */
850 /* > err = 3 Reciprocal condition number: Estimated normwise */
851 /* > reciprocal condition number. Compared with the threshold */
852 /* > sqrt(n) * dlamch('Epsilon') to determine if the error */
853 /* > estimate is "guaranteed". These reciprocal condition */
854 /* > numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some */
855 /* > appropriately scaled matrix Z. */
856 /* > Let Z = S*A, where S scales each row by a power of the */
857 /* > radix so all absolute row sums of Z are approximately 1. */
859 /* > See Lapack Working Note 165 for further details and extra */
863 /* > \param[out] ERR_BNDS_COMP */
865 /* > ERR_BNDS_COMP is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS) */
866 /* > For each right-hand side, this array contains information about */
867 /* > various error bounds and condition numbers corresponding to the */
868 /* > componentwise relative error, which is defined as follows: */
870 /* > Componentwise relative error in the ith solution vector: */
871 /* > abs(XTRUE(j,i) - X(j,i)) */
872 /* > max_j ---------------------- */
875 /* > The array is indexed by the right-hand side i (on which the */
876 /* > componentwise relative error depends), and the type of error */
877 /* > information as described below. There currently are up to three */
878 /* > pieces of information returned for each right-hand side. If */
879 /* > componentwise accuracy is not requested (PARAMS(3) = 0.0), then */
880 /* > ERR_BNDS_COMP is not accessed. If N_ERR_BNDS < 3, then at most */
881 /* > the first (:,N_ERR_BNDS) entries are returned. */
883 /* > The first index in ERR_BNDS_COMP(i,:) corresponds to the ith */
884 /* > right-hand side. */
886 /* > The second index in ERR_BNDS_COMP(:,err) contains the following */
887 /* > three fields: */
888 /* > err = 1 "Trust/don't trust" boolean. Trust the answer if the */
889 /* > reciprocal condition number is less than the threshold */
890 /* > sqrt(n) * dlamch('Epsilon'). */
892 /* > err = 2 "Guaranteed" error bound: The estimated forward error, */
893 /* > almost certainly within a factor of 10 of the true error */
894 /* > so long as the next entry is greater than the threshold */
895 /* > sqrt(n) * dlamch('Epsilon'). This error bound should only */
896 /* > be trusted if the previous boolean is true. */
898 /* > err = 3 Reciprocal condition number: Estimated componentwise */
899 /* > reciprocal condition number. Compared with the threshold */
900 /* > sqrt(n) * dlamch('Epsilon') to determine if the error */
901 /* > estimate is "guaranteed". These reciprocal condition */
902 /* > numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some */
903 /* > appropriately scaled matrix Z. */
904 /* > Let Z = S*(A*diag(x)), where x is the solution for the */
905 /* > current right-hand side and S scales each row of */
906 /* > A*diag(x) by a power of the radix so all absolute row */
907 /* > sums of Z are approximately 1. */
909 /* > See Lapack Working Note 165 for further details and extra */
913 /* > \param[in] NPARAMS */
915 /* > NPARAMS is INTEGER */
916 /* > Specifies the number of parameters set in PARAMS. If <= 0, the */
917 /* > PARAMS array is never referenced and default values are used. */
920 /* > \param[in,out] PARAMS */
922 /* > PARAMS is DOUBLE PRECISION array, dimension NPARAMS */
923 /* > Specifies algorithm parameters. If an entry is < 0.0, then */
924 /* > that entry will be filled with default value used for that */
925 /* > parameter. Only positions up to NPARAMS are accessed; defaults */
926 /* > are used for higher-numbered parameters. */
928 /* > PARAMS(LA_LINRX_ITREF_I = 1) : Whether to perform iterative */
929 /* > refinement or not. */
930 /* > Default: 1.0D+0 */
931 /* > = 0.0: No refinement is performed, and no error bounds are */
933 /* > = 1.0: Use the extra-precise refinement algorithm. */
934 /* > (other values are reserved for future use) */
936 /* > PARAMS(LA_LINRX_ITHRESH_I = 2) : Maximum number of residual */
937 /* > computations allowed for refinement. */
939 /* > Aggressive: Set to 100 to permit convergence using approximate */
940 /* > factorizations or factorizations other than LU. If */
941 /* > the factorization uses a technique other than */
942 /* > Gaussian elimination, the guarantees in */
943 /* > err_bnds_norm and err_bnds_comp may no longer be */
946 /* > PARAMS(LA_LINRX_CWISE_I = 3) : Flag determining if the code */
947 /* > will attempt to find a solution with small componentwise */
948 /* > relative error in the double-precision algorithm. Positive */
949 /* > is true, 0.0 is false. */
950 /* > Default: 1.0 (attempt componentwise convergence) */
953 /* > \param[out] WORK */
955 /* > WORK is DOUBLE PRECISION array, dimension (4*N) */
958 /* > \param[out] IWORK */
960 /* > IWORK is INTEGER array, dimension (N) */
963 /* > \param[out] INFO */
965 /* > INFO is INTEGER */
966 /* > = 0: Successful exit. The solution to every right-hand side is */
968 /* > < 0: If INFO = -i, the i-th argument had an illegal value */
969 /* > > 0 and <= N: U(INFO,INFO) is exactly zero. The factorization */
970 /* > has been completed, but the factor U is exactly singular, so */
971 /* > the solution and error bounds could not be computed. RCOND = 0 */
973 /* > = N+J: The solution corresponding to the Jth right-hand side is */
974 /* > not guaranteed. The solutions corresponding to other right- */
975 /* > hand sides K with K > J may not be guaranteed as well, but */
976 /* > only the first such right-hand side is reported. If a small */
977 /* > componentwise error is not requested (PARAMS(3) = 0.0) then */
978 /* > the Jth right-hand side is the first with a normwise error */
979 /* > bound that is not guaranteed (the smallest J such */
980 /* > that ERR_BNDS_NORM(J,1) = 0.0). By default (PARAMS(3) = 1.0) */
981 /* > the Jth right-hand side is the first with either a normwise or */
982 /* > componentwise error bound that is not guaranteed (the smallest */
983 /* > J such that either ERR_BNDS_NORM(J,1) = 0.0 or */
984 /* > ERR_BNDS_COMP(J,1) = 0.0). See the definition of */
985 /* > ERR_BNDS_NORM(:,1) and ERR_BNDS_COMP(:,1). To get information */
986 /* > about all of the right-hand sides check ERR_BNDS_NORM or */
987 /* > ERR_BNDS_COMP. */
993 /* > \author Univ. of Tennessee */
994 /* > \author Univ. of California Berkeley */
995 /* > \author Univ. of Colorado Denver */
996 /* > \author NAG Ltd. */
998 /* > \date April 2012 */
1000 /* > \ingroup doublePOsolve */
1002 /* ===================================================================== */
1003 /* Subroutine */ int dposvxx_(char *fact, char *uplo, integer *n, integer *
1004 nrhs, doublereal *a, integer *lda, doublereal *af, integer *ldaf,
1005 char *equed, doublereal *s, doublereal *b, integer *ldb, doublereal *
1006 x, integer *ldx, doublereal *rcond, doublereal *rpvgrw, doublereal *
1007 berr, integer *n_err_bnds__, doublereal *err_bnds_norm__, doublereal *
1008 err_bnds_comp__, integer *nparams, doublereal *params, doublereal *
1009 work, integer *iwork, integer *info)
1011 /* System generated locals */
1012 integer a_dim1, a_offset, af_dim1, af_offset, b_dim1, b_offset, x_dim1,
1013 x_offset, err_bnds_norm_dim1, err_bnds_norm_offset,
1014 err_bnds_comp_dim1, err_bnds_comp_offset, i__1;
1015 doublereal d__1, d__2;
1017 /* Local variables */
1018 doublereal amax, smin, smax;
1019 extern doublereal dla_porpvgrw_(char *, integer *, doublereal *, integer
1020 *, doublereal *, integer *, doublereal *);
1022 extern logical lsame_(char *, char *);
1024 logical equil, rcequ;
1025 extern doublereal dlamch_(char *);
1027 extern /* Subroutine */ int dlacpy_(char *, integer *, integer *,
1028 doublereal *, integer *, doublereal *, integer *),
1029 xerbla_(char *, integer *, ftnlen);
1032 extern /* Subroutine */ int dlaqsy_(char *, integer *, doublereal *,
1033 integer *, doublereal *, doublereal *, doublereal *, char *), dpotrf_(char *, integer *, doublereal *, integer
1036 extern /* Subroutine */ int dpotrs_(char *, integer *, integer *,
1037 doublereal *, integer *, doublereal *, integer *, integer *), dlascl2_(integer *, integer *, doublereal *, doublereal *
1038 , integer *), dpoequb_(integer *, doublereal *, integer *,
1039 doublereal *, doublereal *, doublereal *, integer *), dporfsx_(
1040 char *, char *, integer *, integer *, doublereal *, integer *,
1041 doublereal *, integer *, doublereal *, doublereal *, integer *,
1042 doublereal *, integer *, doublereal *, doublereal *, integer *,
1043 doublereal *, doublereal *, integer *, doublereal *, doublereal *,
1044 integer *, integer *);
1047 /* -- LAPACK driver routine (version 3.7.0) -- */
1048 /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
1049 /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
1053 /* ================================================================== */
1056 /* Parameter adjustments */
1057 err_bnds_comp_dim1 = *nrhs;
1058 err_bnds_comp_offset = 1 + err_bnds_comp_dim1 * 1;
1059 err_bnds_comp__ -= err_bnds_comp_offset;
1060 err_bnds_norm_dim1 = *nrhs;
1061 err_bnds_norm_offset = 1 + err_bnds_norm_dim1 * 1;
1062 err_bnds_norm__ -= err_bnds_norm_offset;
1064 a_offset = 1 + a_dim1 * 1;
1067 af_offset = 1 + af_dim1 * 1;
1071 b_offset = 1 + b_dim1 * 1;
1074 x_offset = 1 + x_dim1 * 1;
1083 nofact = lsame_(fact, "N");
1084 equil = lsame_(fact, "E");
1085 smlnum = dlamch_("Safe minimum");
1086 bignum = 1. / smlnum;
1087 if (nofact || equil) {
1088 *(unsigned char *)equed = 'N';
1091 rcequ = lsame_(equed, "Y");
1094 /* Default is failure. If an input parameter is wrong or */
1095 /* factorization fails, make everything look horrible. Only the */
1096 /* pivot growth is set here, the rest is initialized in DPORFSX. */
1100 /* Test the input parameters. PARAMS is not tested until DPORFSX. */
1102 if (! nofact && ! equil && ! lsame_(fact, "F")) {
1104 } else if (! lsame_(uplo, "U") && ! lsame_(uplo,
1107 } else if (*n < 0) {
1109 } else if (*nrhs < 0) {
1111 } else if (*lda < f2cmax(1,*n)) {
1113 } else if (*ldaf < f2cmax(1,*n)) {
1115 } else if (lsame_(fact, "F") && ! (rcequ || lsame_(
1123 for (j = 1; j <= i__1; ++j) {
1125 d__1 = smin, d__2 = s[j];
1126 smin = f2cmin(d__1,d__2);
1128 d__1 = smax, d__2 = s[j];
1129 smax = f2cmax(d__1,d__2);
1134 } else if (*n > 0) {
1135 scond = f2cmax(smin,smlnum) / f2cmin(smax,bignum);
1141 if (*ldb < f2cmax(1,*n)) {
1143 } else if (*ldx < f2cmax(1,*n)) {
1151 xerbla_("DPOSVXX", &i__1, (ftnlen)7);
1157 /* Compute row and column scalings to equilibrate the matrix A. */
1159 dpoequb_(n, &a[a_offset], lda, &s[1], &scond, &amax, &infequ);
1162 /* Equilibrate the matrix. */
1164 dlaqsy_(uplo, n, &a[a_offset], lda, &s[1], &scond, &amax, equed);
1165 rcequ = lsame_(equed, "Y");
1169 /* Scale the right-hand side. */
1172 dlascl2_(n, nrhs, &s[1], &b[b_offset], ldb);
1175 if (nofact || equil) {
1177 /* Compute the Cholesky factorization of A. */
1179 dlacpy_(uplo, n, n, &a[a_offset], lda, &af[af_offset], ldaf);
1180 dpotrf_(uplo, n, &af[af_offset], ldaf, info);
1182 /* Return if INFO is non-zero. */
1186 /* Pivot in column INFO is exactly 0 */
1187 /* Compute the reciprocal pivot growth factor of the */
1188 /* leading rank-deficient INFO columns of A. */
1190 *rpvgrw = dla_porpvgrw_(uplo, info, &a[a_offset], lda, &af[
1191 af_offset], ldaf, &work[1]);
1196 /* Compute the reciprocal growth factor RPVGRW. */
1198 *rpvgrw = dla_porpvgrw_(uplo, n, &a[a_offset], lda, &af[af_offset], ldaf,
1201 /* Compute the solution matrix X. */
1203 dlacpy_("Full", n, nrhs, &b[b_offset], ldb, &x[x_offset], ldx);
1204 dpotrs_(uplo, n, nrhs, &af[af_offset], ldaf, &x[x_offset], ldx, info);
1206 /* Use iterative refinement to improve the computed solution and */
1207 /* compute error bounds and backward error estimates for it. */
1209 dporfsx_(uplo, equed, n, nrhs, &a[a_offset], lda, &af[af_offset], ldaf, &
1210 s[1], &b[b_offset], ldb, &x[x_offset], ldx, rcond, &berr[1],
1211 n_err_bnds__, &err_bnds_norm__[err_bnds_norm_offset], &
1212 err_bnds_comp__[err_bnds_comp_offset], nparams, ¶ms[1], &work[
1213 1], &iwork[1], info);
1215 /* Scale solutions. */
1218 dlascl2_(n, nrhs, &s[1], &x[x_offset], ldx);
1223 /* End of DPOSVXX */