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 /* Table of constant values */
515 static integer c__1 = 1;
516 static doublereal c_b9 = -1.;
517 static doublereal c_b11 = 1.;
519 /* > \brief \b DLA_PORFSX_EXTENDED improves the computed solution to a system of linear equations for symmetri
520 c or Hermitian positive-definite matrices by performing extra-precise iterative refinement and provide
521 s error bounds and backward error estimates fo */
522 /* r the solution. */
524 /* =========== DOCUMENTATION =========== */
526 /* Online html documentation available at */
527 /* http://www.netlib.org/lapack/explore-html/ */
530 /* > Download DLA_PORFSX_EXTENDED + dependencies */
531 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dla_por
534 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dla_por
537 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dla_por
545 /* SUBROUTINE DLA_PORFSX_EXTENDED( PREC_TYPE, UPLO, N, NRHS, A, LDA, */
546 /* AF, LDAF, COLEQU, C, B, LDB, Y, */
547 /* LDY, BERR_OUT, N_NORMS, */
548 /* ERR_BNDS_NORM, ERR_BNDS_COMP, RES, */
549 /* AYB, DY, Y_TAIL, RCOND, ITHRESH, */
550 /* RTHRESH, DZ_UB, IGNORE_CWISE, */
553 /* INTEGER INFO, LDA, LDAF, LDB, LDY, N, NRHS, PREC_TYPE, */
554 /* $ N_NORMS, ITHRESH */
556 /* LOGICAL COLEQU, IGNORE_CWISE */
557 /* DOUBLE PRECISION RTHRESH, DZ_UB */
558 /* DOUBLE PRECISION A( LDA, * ), AF( LDAF, * ), B( LDB, * ), */
559 /* $ Y( LDY, * ), RES( * ), DY( * ), Y_TAIL( * ) */
560 /* DOUBLE PRECISION C( * ), AYB(*), RCOND, BERR_OUT( * ), */
561 /* $ ERR_BNDS_NORM( NRHS, * ), */
562 /* $ ERR_BNDS_COMP( NRHS, * ) */
565 /* > \par Purpose: */
570 /* > DLA_PORFSX_EXTENDED improves the computed solution to a system of */
571 /* > linear equations by performing extra-precise iterative refinement */
572 /* > and provides error bounds and backward error estimates for the solution. */
573 /* > This subroutine is called by DPORFSX to perform iterative refinement. */
574 /* > In addition to normwise error bound, the code provides maximum */
575 /* > componentwise error bound if possible. See comments for ERR_BNDS_NORM */
576 /* > and ERR_BNDS_COMP for details of the error bounds. Note that this */
577 /* > subroutine is only resonsible for setting the second fields of */
578 /* > ERR_BNDS_NORM and ERR_BNDS_COMP. */
584 /* > \param[in] PREC_TYPE */
586 /* > PREC_TYPE is INTEGER */
587 /* > Specifies the intermediate precision to be used in refinement. */
588 /* > The value is defined by ILAPREC(P) where P is a CHARACTER and P */
589 /* > = 'S': Single */
590 /* > = 'D': Double */
591 /* > = 'I': Indigenous */
592 /* > = 'X' or 'E': Extra */
595 /* > \param[in] UPLO */
597 /* > UPLO is CHARACTER*1 */
598 /* > = 'U': Upper triangle of A is stored; */
599 /* > = 'L': Lower triangle of A is stored. */
605 /* > The number of linear equations, i.e., the order of the */
606 /* > matrix A. N >= 0. */
609 /* > \param[in] NRHS */
611 /* > NRHS is INTEGER */
612 /* > The number of right-hand-sides, i.e., the number of columns of the */
618 /* > A is DOUBLE PRECISION array, dimension (LDA,N) */
619 /* > On entry, the N-by-N matrix A. */
622 /* > \param[in] LDA */
624 /* > LDA is INTEGER */
625 /* > The leading dimension of the array A. LDA >= f2cmax(1,N). */
628 /* > \param[in] AF */
630 /* > AF is DOUBLE PRECISION array, dimension (LDAF,N) */
631 /* > The triangular factor U or L from the Cholesky factorization */
632 /* > A = U**T*U or A = L*L**T, as computed by DPOTRF. */
635 /* > \param[in] LDAF */
637 /* > LDAF is INTEGER */
638 /* > The leading dimension of the array AF. LDAF >= f2cmax(1,N). */
641 /* > \param[in] COLEQU */
643 /* > COLEQU is LOGICAL */
644 /* > If .TRUE. then column equilibration was done to A before calling */
645 /* > this routine. This is needed to compute the solution and error */
646 /* > bounds correctly. */
651 /* > C is DOUBLE PRECISION array, dimension (N) */
652 /* > The column scale factors for A. If COLEQU = .FALSE., C */
653 /* > is not accessed. If C is input, each element of C should be a power */
654 /* > of the radix to ensure a reliable solution and error estimates. */
655 /* > Scaling by powers of the radix does not cause rounding errors unless */
656 /* > the result underflows or overflows. Rounding errors during scaling */
657 /* > lead to refining with a matrix that is not equivalent to the */
658 /* > input matrix, producing error estimates that may not be */
664 /* > B is DOUBLE PRECISION array, dimension (LDB,NRHS) */
665 /* > The right-hand-side matrix B. */
668 /* > \param[in] LDB */
670 /* > LDB is INTEGER */
671 /* > The leading dimension of the array B. LDB >= f2cmax(1,N). */
674 /* > \param[in,out] Y */
676 /* > Y is DOUBLE PRECISION array, dimension (LDY,NRHS) */
677 /* > On entry, the solution matrix X, as computed by DPOTRS. */
678 /* > On exit, the improved solution matrix Y. */
681 /* > \param[in] LDY */
683 /* > LDY is INTEGER */
684 /* > The leading dimension of the array Y. LDY >= f2cmax(1,N). */
687 /* > \param[out] BERR_OUT */
689 /* > BERR_OUT is DOUBLE PRECISION array, dimension (NRHS) */
690 /* > On exit, BERR_OUT(j) contains the componentwise relative backward */
691 /* > error for right-hand-side j from the formula */
692 /* > f2cmax(i) ( abs(RES(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) ) */
693 /* > where abs(Z) is the componentwise absolute value of the matrix */
694 /* > or vector Z. This is computed by DLA_LIN_BERR. */
697 /* > \param[in] N_NORMS */
699 /* > N_NORMS is INTEGER */
700 /* > Determines which error bounds to return (see ERR_BNDS_NORM */
701 /* > and ERR_BNDS_COMP). */
702 /* > If N_NORMS >= 1 return normwise error bounds. */
703 /* > If N_NORMS >= 2 return componentwise error bounds. */
706 /* > \param[in,out] ERR_BNDS_NORM */
708 /* > ERR_BNDS_NORM is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS) */
709 /* > For each right-hand side, this array contains information about */
710 /* > various error bounds and condition numbers corresponding to the */
711 /* > normwise relative error, which is defined as follows: */
713 /* > Normwise relative error in the ith solution vector: */
714 /* > max_j (abs(XTRUE(j,i) - X(j,i))) */
715 /* > ------------------------------ */
716 /* > max_j abs(X(j,i)) */
718 /* > The array is indexed by the type of error information as described */
719 /* > below. There currently are up to three pieces of information */
722 /* > The first index in ERR_BNDS_NORM(i,:) corresponds to the ith */
723 /* > right-hand side. */
725 /* > The second index in ERR_BNDS_NORM(:,err) contains the following */
726 /* > three fields: */
727 /* > err = 1 "Trust/don't trust" boolean. Trust the answer if the */
728 /* > reciprocal condition number is less than the threshold */
729 /* > sqrt(n) * slamch('Epsilon'). */
731 /* > err = 2 "Guaranteed" error bound: The estimated forward error, */
732 /* > almost certainly within a factor of 10 of the true error */
733 /* > so long as the next entry is greater than the threshold */
734 /* > sqrt(n) * slamch('Epsilon'). This error bound should only */
735 /* > be trusted if the previous boolean is true. */
737 /* > err = 3 Reciprocal condition number: Estimated normwise */
738 /* > reciprocal condition number. Compared with the threshold */
739 /* > sqrt(n) * slamch('Epsilon') to determine if the error */
740 /* > estimate is "guaranteed". These reciprocal condition */
741 /* > numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some */
742 /* > appropriately scaled matrix Z. */
743 /* > Let Z = S*A, where S scales each row by a power of the */
744 /* > radix so all absolute row sums of Z are approximately 1. */
746 /* > This subroutine is only responsible for setting the second field */
748 /* > See Lapack Working Note 165 for further details and extra */
752 /* > \param[in,out] ERR_BNDS_COMP */
754 /* > ERR_BNDS_COMP is DOUBLE PRECISION array, dimension (NRHS, N_ERR_BNDS) */
755 /* > For each right-hand side, this array contains information about */
756 /* > various error bounds and condition numbers corresponding to the */
757 /* > componentwise relative error, which is defined as follows: */
759 /* > Componentwise relative error in the ith solution vector: */
760 /* > abs(XTRUE(j,i) - X(j,i)) */
761 /* > max_j ---------------------- */
764 /* > The array is indexed by the right-hand side i (on which the */
765 /* > componentwise relative error depends), and the type of error */
766 /* > information as described below. There currently are up to three */
767 /* > pieces of information returned for each right-hand side. If */
768 /* > componentwise accuracy is not requested (PARAMS(3) = 0.0), then */
769 /* > ERR_BNDS_COMP is not accessed. If N_ERR_BNDS < 3, then at most */
770 /* > the first (:,N_ERR_BNDS) entries are returned. */
772 /* > The first index in ERR_BNDS_COMP(i,:) corresponds to the ith */
773 /* > right-hand side. */
775 /* > The second index in ERR_BNDS_COMP(:,err) contains the following */
776 /* > three fields: */
777 /* > err = 1 "Trust/don't trust" boolean. Trust the answer if the */
778 /* > reciprocal condition number is less than the threshold */
779 /* > sqrt(n) * slamch('Epsilon'). */
781 /* > err = 2 "Guaranteed" error bound: The estimated forward error, */
782 /* > almost certainly within a factor of 10 of the true error */
783 /* > so long as the next entry is greater than the threshold */
784 /* > sqrt(n) * slamch('Epsilon'). This error bound should only */
785 /* > be trusted if the previous boolean is true. */
787 /* > err = 3 Reciprocal condition number: Estimated componentwise */
788 /* > reciprocal condition number. Compared with the threshold */
789 /* > sqrt(n) * slamch('Epsilon') to determine if the error */
790 /* > estimate is "guaranteed". These reciprocal condition */
791 /* > numbers are 1 / (norm(Z^{-1},inf) * norm(Z,inf)) for some */
792 /* > appropriately scaled matrix Z. */
793 /* > Let Z = S*(A*diag(x)), where x is the solution for the */
794 /* > current right-hand side and S scales each row of */
795 /* > A*diag(x) by a power of the radix so all absolute row */
796 /* > sums of Z are approximately 1. */
798 /* > This subroutine is only responsible for setting the second field */
800 /* > See Lapack Working Note 165 for further details and extra */
804 /* > \param[in] RES */
806 /* > RES is DOUBLE PRECISION array, dimension (N) */
807 /* > Workspace to hold the intermediate residual. */
810 /* > \param[in] AYB */
812 /* > AYB is DOUBLE PRECISION array, dimension (N) */
813 /* > Workspace. This can be the same workspace passed for Y_TAIL. */
816 /* > \param[in] DY */
818 /* > DY is DOUBLE PRECISION array, dimension (N) */
819 /* > Workspace to hold the intermediate solution. */
822 /* > \param[in] Y_TAIL */
824 /* > Y_TAIL is DOUBLE PRECISION array, dimension (N) */
825 /* > Workspace to hold the trailing bits of the intermediate solution. */
828 /* > \param[in] RCOND */
830 /* > RCOND is DOUBLE PRECISION */
831 /* > Reciprocal scaled condition number. This is an estimate of the */
832 /* > reciprocal Skeel condition number of the matrix A after */
833 /* > equilibration (if done). If this is less than the machine */
834 /* > precision (in particular, if it is zero), the matrix is singular */
835 /* > to working precision. Note that the error may still be small even */
836 /* > if this number is very small and the matrix appears ill- */
840 /* > \param[in] ITHRESH */
842 /* > ITHRESH is INTEGER */
843 /* > The maximum number of residual computations allowed for */
844 /* > refinement. The default is 10. For 'aggressive' set to 100 to */
845 /* > permit convergence using approximate factorizations or */
846 /* > factorizations other than LU. If the factorization uses a */
847 /* > technique other than Gaussian elimination, the guarantees in */
848 /* > ERR_BNDS_NORM and ERR_BNDS_COMP may no longer be trustworthy. */
851 /* > \param[in] RTHRESH */
853 /* > RTHRESH is DOUBLE PRECISION */
854 /* > Determines when to stop refinement if the error estimate stops */
855 /* > decreasing. Refinement will stop when the next solution no longer */
856 /* > satisfies norm(dx_{i+1}) < RTHRESH * norm(dx_i) where norm(Z) is */
857 /* > the infinity norm of Z. RTHRESH satisfies 0 < RTHRESH <= 1. The */
858 /* > default value is 0.5. For 'aggressive' set to 0.9 to permit */
859 /* > convergence on extremely ill-conditioned matrices. See LAWN 165 */
860 /* > for more details. */
863 /* > \param[in] DZ_UB */
865 /* > DZ_UB is DOUBLE PRECISION */
866 /* > Determines when to start considering componentwise convergence. */
867 /* > Componentwise convergence is only considered after each component */
868 /* > of the solution Y is stable, which we definte as the relative */
869 /* > change in each component being less than DZ_UB. The default value */
870 /* > is 0.25, requiring the first bit to be stable. See LAWN 165 for */
871 /* > more details. */
874 /* > \param[in] IGNORE_CWISE */
876 /* > IGNORE_CWISE is LOGICAL */
877 /* > If .TRUE. then ignore componentwise convergence. Default value */
881 /* > \param[out] INFO */
883 /* > INFO is INTEGER */
884 /* > = 0: Successful exit. */
885 /* > < 0: if INFO = -i, the ith argument to DPOTRS had an illegal */
892 /* > \author Univ. of Tennessee */
893 /* > \author Univ. of California Berkeley */
894 /* > \author Univ. of Colorado Denver */
895 /* > \author NAG Ltd. */
897 /* > \date June 2017 */
899 /* > \ingroup doublePOcomputational */
901 /* ===================================================================== */
902 /* Subroutine */ int dla_porfsx_extended_(integer *prec_type__, char *uplo,
903 integer *n, integer *nrhs, doublereal *a, integer *lda, doublereal *
904 af, integer *ldaf, logical *colequ, doublereal *c__, doublereal *b,
905 integer *ldb, doublereal *y, integer *ldy, doublereal *berr_out__,
906 integer *n_norms__, doublereal *err_bnds_norm__, doublereal *
907 err_bnds_comp__, doublereal *res, doublereal *ayb, doublereal *dy,
908 doublereal *y_tail__, doublereal *rcond, integer *ithresh, doublereal
909 *rthresh, doublereal *dz_ub__, logical *ignore_cwise__, integer *info)
911 /* System generated locals */
912 integer a_dim1, a_offset, af_dim1, af_offset, b_dim1, b_offset, y_dim1,
913 y_offset, err_bnds_norm_dim1, err_bnds_norm_offset,
914 err_bnds_comp_dim1, err_bnds_comp_offset, i__1, i__2, i__3;
915 doublereal d__1, d__2;
917 /* Local variables */
918 doublereal dx_x__, dz_z__;
919 extern /* Subroutine */ int dla_lin_berr_(integer *, integer *, integer *
920 , doublereal *, doublereal *, doublereal *);
921 doublereal ymin, dxratmax, dzratmax;
922 integer y_prec_state__;
923 extern /* Subroutine */ int blas_dsymv_x_(integer *, integer *,
924 doublereal *, doublereal *, integer *, doublereal *, integer *,
925 doublereal *, doublereal *, integer *, integer *);
926 integer uplo2, i__, j;
927 extern logical lsame_(char *, char *);
928 extern /* Subroutine */ int blas_dsymv2_x_(integer *, integer *,
929 doublereal *, doublereal *, integer *, doublereal *, doublereal *,
930 integer *, doublereal *, doublereal *, integer *, integer *),
931 dcopy_(integer *, doublereal *, integer *, doublereal *, integer *
936 extern /* Subroutine */ int daxpy_(integer *, doublereal *, doublereal *,
937 integer *, doublereal *, integer *), dla_syamv_(integer *,
938 integer *, doublereal *, doublereal *, integer *, doublereal *,
939 integer *, doublereal *, doublereal *, integer *), dsymv_(char *,
940 integer *, doublereal *, doublereal *, integer *, doublereal *,
941 integer *, doublereal *, doublereal *, integer *);
942 doublereal normx, normy, myhugeval, prev_dz_z__;
943 extern doublereal dlamch_(char *);
944 doublereal yk, final_dx_x__;
945 extern /* Subroutine */ int dla_wwaddw_(integer *, doublereal *,
946 doublereal *, doublereal *);
947 doublereal final_dz_z__, normdx;
948 extern /* Subroutine */ int dpotrs_(char *, integer *, integer *,
949 doublereal *, integer *, doublereal *, integer *, integer *);
950 doublereal prevnormdx;
953 extern integer ilauplo_(char *);
954 integer x_state__, z_state__;
955 doublereal incr_thresh__;
958 /* -- LAPACK computational routine (version 3.7.1) -- */
959 /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
960 /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
964 /* ===================================================================== */
967 /* Parameter adjustments */
968 err_bnds_comp_dim1 = *nrhs;
969 err_bnds_comp_offset = 1 + err_bnds_comp_dim1 * 1;
970 err_bnds_comp__ -= err_bnds_comp_offset;
971 err_bnds_norm_dim1 = *nrhs;
972 err_bnds_norm_offset = 1 + err_bnds_norm_dim1 * 1;
973 err_bnds_norm__ -= err_bnds_norm_offset;
975 a_offset = 1 + a_dim1 * 1;
978 af_offset = 1 + af_dim1 * 1;
982 b_offset = 1 + b_dim1 * 1;
985 y_offset = 1 + y_dim1 * 1;
997 eps = dlamch_("Epsilon");
998 myhugeval = dlamch_("Overflow");
999 /* Force MYHUGEVAL to Inf */
1000 myhugeval *= myhugeval;
1001 /* Using MYHUGEVAL may lead to spurious underflows. */
1002 incr_thresh__ = (doublereal) (*n) * eps;
1003 if (lsame_(uplo, "L")) {
1004 uplo2 = ilauplo_("L");
1006 uplo2 = ilauplo_("U");
1009 for (j = 1; j <= i__1; ++j) {
1011 if (y_prec_state__ == 2) {
1013 for (i__ = 1; i__ <= i__2; ++i__) {
1021 final_dx_x__ = myhugeval;
1022 final_dz_z__ = myhugeval;
1023 prevnormdx = myhugeval;
1024 prev_dz_z__ = myhugeval;
1029 incr_prec__ = FALSE_;
1031 for (cnt = 1; cnt <= i__2; ++cnt) {
1033 /* Compute residual RES = B_s - op(A_s) * Y, */
1034 /* op(A) = A, A**T, or A**H depending on TRANS (and type). */
1036 dcopy_(n, &b[j * b_dim1 + 1], &c__1, &res[1], &c__1);
1037 if (y_prec_state__ == 0) {
1038 dsymv_(uplo, n, &c_b9, &a[a_offset], lda, &y[j * y_dim1 + 1],
1039 &c__1, &c_b11, &res[1], &c__1);
1040 } else if (y_prec_state__ == 1) {
1041 blas_dsymv_x__(&uplo2, n, &c_b9, &a[a_offset], lda, &y[j *
1042 y_dim1 + 1], &c__1, &c_b11, &res[1], &c__1,
1045 blas_dsymv2_x__(&uplo2, n, &c_b9, &a[a_offset], lda, &y[j *
1046 y_dim1 + 1], &y_tail__[1], &c__1, &c_b11, &res[1], &
1049 /* XXX: RES is no longer needed. */
1050 dcopy_(n, &res[1], &c__1, &dy[1], &c__1);
1051 dpotrs_(uplo, n, &c__1, &af[af_offset], ldaf, &dy[1], n, info);
1053 /* Calculate relative changes DX_X, DZ_Z and ratios DXRAT, DZRAT. */
1061 for (i__ = 1; i__ <= i__3; ++i__) {
1062 yk = (d__1 = y[i__ + j * y_dim1], abs(d__1));
1063 dyk = (d__1 = dy[i__], abs(d__1));
1066 d__1 = dz_z__, d__2 = dyk / yk;
1067 dz_z__ = f2cmax(d__1,d__2);
1068 } else if (dyk != 0.) {
1071 ymin = f2cmin(ymin,yk);
1072 normy = f2cmax(normy,yk);
1075 d__1 = normx, d__2 = yk * c__[i__];
1076 normx = f2cmax(d__1,d__2);
1078 d__1 = normdx, d__2 = dyk * c__[i__];
1079 normdx = f2cmax(d__1,d__2);
1082 normdx = f2cmax(normdx,dyk);
1086 dx_x__ = normdx / normx;
1087 } else if (normdx == 0.) {
1092 dxrat = normdx / prevnormdx;
1093 dzrat = dz_z__ / prev_dz_z__;
1095 /* Check termination criteria. */
1097 if (ymin * *rcond < incr_thresh__ * normy && y_prec_state__ < 2) {
1098 incr_prec__ = TRUE_;
1100 if (x_state__ == 3 && dxrat <= *rthresh) {
1103 if (x_state__ == 1) {
1104 if (dx_x__ <= eps) {
1106 } else if (dxrat > *rthresh) {
1107 if (y_prec_state__ != 2) {
1108 incr_prec__ = TRUE_;
1113 if (dxrat > dxratmax) {
1117 if (x_state__ > 1) {
1118 final_dx_x__ = dx_x__;
1121 if (z_state__ == 0 && dz_z__ <= *dz_ub__) {
1124 if (z_state__ == 3 && dzrat <= *rthresh) {
1127 if (z_state__ == 1) {
1128 if (dz_z__ <= eps) {
1130 } else if (dz_z__ > *dz_ub__) {
1133 final_dz_z__ = myhugeval;
1134 } else if (dzrat > *rthresh) {
1135 if (y_prec_state__ != 2) {
1136 incr_prec__ = TRUE_;
1141 if (dzrat > dzratmax) {
1145 if (z_state__ > 1) {
1146 final_dz_z__ = dz_z__;
1149 if (x_state__ != 1 && (*ignore_cwise__ || z_state__ != 1)) {
1153 incr_prec__ = FALSE_;
1156 for (i__ = 1; i__ <= i__3; ++i__) {
1160 prevnormdx = normdx;
1161 prev_dz_z__ = dz_z__;
1163 /* Update soluton. */
1165 if (y_prec_state__ < 2) {
1166 daxpy_(n, &c_b11, &dy[1], &c__1, &y[j * y_dim1 + 1], &c__1);
1168 dla_wwaddw_(n, &y[j * y_dim1 + 1], &y_tail__[1], &dy[1]);
1171 /* Target of "IF (Z_STOP .AND. X_STOP)". Sun's f77 won't CALL MYEXIT. */
1174 /* Set final_* when cnt hits ithresh. */
1176 if (x_state__ == 1) {
1177 final_dx_x__ = dx_x__;
1179 if (z_state__ == 1) {
1180 final_dz_z__ = dz_z__;
1183 /* Compute error bounds. */
1185 if (*n_norms__ >= 1) {
1186 err_bnds_norm__[j + (err_bnds_norm_dim1 << 1)] = final_dx_x__ / (
1189 if (*n_norms__ >= 2) {
1190 err_bnds_comp__[j + (err_bnds_comp_dim1 << 1)] = final_dz_z__ / (
1194 /* Compute componentwise relative backward error from formula */
1195 /* f2cmax(i) ( abs(R(i)) / ( abs(op(A_s))*abs(Y) + abs(B_s) )(i) ) */
1196 /* where abs(Z) is the componentwise absolute value of the matrix */
1199 /* Compute residual RES = B_s - op(A_s) * Y, */
1200 /* op(A) = A, A**T, or A**H depending on TRANS (and type). */
1202 dcopy_(n, &b[j * b_dim1 + 1], &c__1, &res[1], &c__1);
1203 dsymv_(uplo, n, &c_b9, &a[a_offset], lda, &y[j * y_dim1 + 1], &c__1, &
1204 c_b11, &res[1], &c__1);
1206 for (i__ = 1; i__ <= i__2; ++i__) {
1207 ayb[i__] = (d__1 = b[i__ + j * b_dim1], abs(d__1));
1210 /* Compute abs(op(A_s))*abs(Y) + abs(B_s). */
1212 dla_syamv_(&uplo2, n, &c_b11, &a[a_offset], lda, &y[j * y_dim1 + 1],
1213 &c__1, &c_b11, &ayb[1], &c__1);
1214 dla_lin_berr_(n, n, &c__1, &res[1], &ayb[1], &berr_out__[j]);
1216 /* End of loop for each RHS. */
1221 } /* dla_porfsx_extended__ */