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 complex c_b1 = {1.f,0.f};
516 static integer c__1 = 1;
518 /* > \brief \b CLAHEF computes a partial factorization of a complex Hermitian indefinite matrix using the Bunc
519 h-Kaufman diagonal pivoting method (blocked algorithm, calling Level 3 BLAS). */
521 /* =========== DOCUMENTATION =========== */
523 /* Online html documentation available at */
524 /* http://www.netlib.org/lapack/explore-html/ */
527 /* > Download CLAHEF + dependencies */
528 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/clahef.
531 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/clahef.
534 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/clahef.
542 /* SUBROUTINE CLAHEF( UPLO, N, NB, KB, A, LDA, IPIV, W, LDW, INFO ) */
545 /* INTEGER INFO, KB, LDA, LDW, N, NB */
546 /* INTEGER IPIV( * ) */
547 /* COMPLEX A( LDA, * ), W( LDW, * ) */
550 /* > \par Purpose: */
555 /* > CLAHEF computes a partial factorization of a complex Hermitian */
556 /* > matrix A using the Bunch-Kaufman diagonal pivoting method. The */
557 /* > partial factorization has the form: */
559 /* > A = ( I U12 ) ( A11 0 ) ( I 0 ) if UPLO = 'U', or: */
560 /* > ( 0 U22 ) ( 0 D ) ( U12**H U22**H ) */
562 /* > A = ( L11 0 ) ( D 0 ) ( L11**H L21**H ) if UPLO = 'L' */
563 /* > ( L21 I ) ( 0 A22 ) ( 0 I ) */
565 /* > where the order of D is at most NB. The actual order is returned in */
566 /* > the argument KB, and is either NB or NB-1, or N if N <= NB. */
567 /* > Note that U**H denotes the conjugate transpose of U. */
569 /* > CLAHEF is an auxiliary routine called by CHETRF. It uses blocked code */
570 /* > (calling Level 3 BLAS) to update the submatrix A11 (if UPLO = 'U') or */
571 /* > A22 (if UPLO = 'L'). */
577 /* > \param[in] UPLO */
579 /* > UPLO is CHARACTER*1 */
580 /* > Specifies whether the upper or lower triangular part of the */
581 /* > Hermitian matrix A is stored: */
582 /* > = 'U': Upper triangular */
583 /* > = 'L': Lower triangular */
589 /* > The order of the matrix A. N >= 0. */
592 /* > \param[in] NB */
594 /* > NB is INTEGER */
595 /* > The maximum number of columns of the matrix A that should be */
596 /* > factored. NB should be at least 2 to allow for 2-by-2 pivot */
600 /* > \param[out] KB */
602 /* > KB is INTEGER */
603 /* > The number of columns of A that were actually factored. */
604 /* > KB is either NB-1 or NB, or N if N <= NB. */
607 /* > \param[in,out] A */
609 /* > A is COMPLEX array, dimension (LDA,N) */
610 /* > On entry, the Hermitian matrix A. If UPLO = 'U', the leading */
611 /* > n-by-n upper triangular part of A contains the upper */
612 /* > triangular part of the matrix A, and the strictly lower */
613 /* > triangular part of A is not referenced. If UPLO = 'L', the */
614 /* > leading n-by-n lower triangular part of A contains the lower */
615 /* > triangular part of the matrix A, and the strictly upper */
616 /* > triangular part of A is not referenced. */
617 /* > On exit, A contains details of the partial factorization. */
620 /* > \param[in] LDA */
622 /* > LDA is INTEGER */
623 /* > The leading dimension of the array A. LDA >= f2cmax(1,N). */
626 /* > \param[out] IPIV */
628 /* > IPIV is INTEGER array, dimension (N) */
629 /* > Details of the interchanges and the block structure of D. */
631 /* > If UPLO = 'U': */
632 /* > Only the last KB elements of IPIV are set. */
634 /* > If IPIV(k) > 0, then rows and columns k and IPIV(k) were */
635 /* > interchanged and D(k,k) is a 1-by-1 diagonal block. */
637 /* > If IPIV(k) = IPIV(k-1) < 0, then rows and columns */
638 /* > k-1 and -IPIV(k) were interchanged and D(k-1:k,k-1:k) */
639 /* > is a 2-by-2 diagonal block. */
641 /* > If UPLO = 'L': */
642 /* > Only the first KB elements of IPIV are set. */
644 /* > If IPIV(k) > 0, then rows and columns k and IPIV(k) were */
645 /* > interchanged and D(k,k) is a 1-by-1 diagonal block. */
647 /* > If IPIV(k) = IPIV(k+1) < 0, then rows and columns */
648 /* > k+1 and -IPIV(k) were interchanged and D(k:k+1,k:k+1) */
649 /* > is a 2-by-2 diagonal block. */
652 /* > \param[out] W */
654 /* > W is COMPLEX array, dimension (LDW,NB) */
657 /* > \param[in] LDW */
659 /* > LDW is INTEGER */
660 /* > The leading dimension of the array W. LDW >= f2cmax(1,N). */
663 /* > \param[out] INFO */
665 /* > INFO is INTEGER */
666 /* > = 0: successful exit */
667 /* > > 0: if INFO = k, D(k,k) is exactly zero. The factorization */
668 /* > has been completed, but the block diagonal matrix D is */
669 /* > exactly singular. */
675 /* > \author Univ. of Tennessee */
676 /* > \author Univ. of California Berkeley */
677 /* > \author Univ. of Colorado Denver */
678 /* > \author NAG Ltd. */
680 /* > \date November 2013 */
682 /* > \ingroup complexHEcomputational */
684 /* > \par Contributors: */
685 /* ================== */
689 /* > November 2013, Igor Kozachenko, */
690 /* > Computer Science Division, */
691 /* > University of California, Berkeley */
694 /* ===================================================================== */
695 /* Subroutine */ int clahef_(char *uplo, integer *n, integer *nb, integer *kb,
696 complex *a, integer *lda, integer *ipiv, complex *w, integer *ldw,
699 /* System generated locals */
700 integer a_dim1, a_offset, w_dim1, w_offset, i__1, i__2, i__3, i__4, i__5;
701 real r__1, r__2, r__3, r__4;
702 complex q__1, q__2, q__3, q__4;
704 /* Local variables */
705 integer imax, jmax, j, k;
707 extern /* Subroutine */ int cgemm_(char *, char *, integer *, integer *,
708 integer *, complex *, complex *, integer *, complex *, integer *,
709 complex *, complex *, integer *);
710 extern logical lsame_(char *, char *);
711 extern /* Subroutine */ int cgemv_(char *, integer *, integer *, complex *
712 , complex *, integer *, complex *, integer *, complex *, complex *
713 , integer *), ccopy_(integer *, complex *, integer *,
714 complex *, integer *), cswap_(integer *, complex *, integer *,
715 complex *, integer *);
718 complex d11, d21, d22;
719 integer jb, jj, kk, jp, kp;
721 extern /* Subroutine */ int clacgv_(integer *, complex *, integer *);
723 extern integer icamax_(integer *, complex *, integer *);
724 extern /* Subroutine */ int csscal_(integer *, real *, complex *, integer
730 /* -- LAPACK computational routine (version 3.5.0) -- */
731 /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
732 /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
736 /* ===================================================================== */
739 /* Parameter adjustments */
741 a_offset = 1 + a_dim1 * 1;
745 w_offset = 1 + w_dim1 * 1;
751 /* Initialize ALPHA for use in choosing pivot block size. */
753 alpha = (sqrt(17.f) + 1.f) / 8.f;
755 if (lsame_(uplo, "U")) {
757 /* Factorize the trailing columns of A using the upper triangle */
758 /* of A and working backwards, and compute the matrix W = U12*D */
759 /* for use in updating A11 (note that conjg(W) is actually stored) */
761 /* K is the main loop index, decreasing from N in steps of 1 or 2 */
766 /* KW is the column of W which corresponds to column K of A */
772 if (k <= *n - *nb + 1 && *nb < *n || k < 1) {
778 /* Copy column K of A to column KW of W and update it */
781 ccopy_(&i__1, &a[k * a_dim1 + 1], &c__1, &w[kw * w_dim1 + 1], &c__1);
782 i__1 = k + kw * w_dim1;
783 i__2 = k + k * a_dim1;
785 w[i__1].r = r__1, w[i__1].i = 0.f;
788 q__1.r = -1.f, q__1.i = 0.f;
789 cgemv_("No transpose", &k, &i__1, &q__1, &a[(k + 1) * a_dim1 + 1],
790 lda, &w[k + (kw + 1) * w_dim1], ldw, &c_b1, &w[kw *
792 i__1 = k + kw * w_dim1;
793 i__2 = k + kw * w_dim1;
795 w[i__1].r = r__1, w[i__1].i = 0.f;
798 /* Determine rows and columns to be interchanged and whether */
799 /* a 1-by-1 or 2-by-2 pivot block will be used */
801 i__1 = k + kw * w_dim1;
802 absakk = (r__1 = w[i__1].r, abs(r__1));
804 /* IMAX is the row-index of the largest off-diagonal element in */
805 /* column K, and COLMAX is its absolute value. */
806 /* Determine both COLMAX and IMAX. */
810 imax = icamax_(&i__1, &w[kw * w_dim1 + 1], &c__1);
811 i__1 = imax + kw * w_dim1;
812 colmax = (r__1 = w[i__1].r, abs(r__1)) + (r__2 = r_imag(&w[imax +
813 kw * w_dim1]), abs(r__2));
818 if (f2cmax(absakk,colmax) == 0.f) {
820 /* Column K is zero or underflow: set INFO and continue */
826 i__1 = k + k * a_dim1;
827 i__2 = k + k * a_dim1;
829 a[i__1].r = r__1, a[i__1].i = 0.f;
832 /* ============================================================ */
834 /* BEGIN pivot search */
837 if (absakk >= alpha * colmax) {
839 /* no interchange, use 1-by-1 pivot block */
844 /* BEGIN pivot search along IMAX row */
847 /* Copy column IMAX to column KW-1 of W and update it */
850 ccopy_(&i__1, &a[imax * a_dim1 + 1], &c__1, &w[(kw - 1) *
852 i__1 = imax + (kw - 1) * w_dim1;
853 i__2 = imax + imax * a_dim1;
855 w[i__1].r = r__1, w[i__1].i = 0.f;
857 ccopy_(&i__1, &a[imax + (imax + 1) * a_dim1], lda, &w[imax +
858 1 + (kw - 1) * w_dim1], &c__1);
860 clacgv_(&i__1, &w[imax + 1 + (kw - 1) * w_dim1], &c__1);
863 q__1.r = -1.f, q__1.i = 0.f;
864 cgemv_("No transpose", &k, &i__1, &q__1, &a[(k + 1) *
865 a_dim1 + 1], lda, &w[imax + (kw + 1) * w_dim1],
866 ldw, &c_b1, &w[(kw - 1) * w_dim1 + 1], &c__1);
867 i__1 = imax + (kw - 1) * w_dim1;
868 i__2 = imax + (kw - 1) * w_dim1;
870 w[i__1].r = r__1, w[i__1].i = 0.f;
873 /* JMAX is the column-index of the largest off-diagonal */
874 /* element in row IMAX, and ROWMAX is its absolute value. */
875 /* Determine only ROWMAX. */
878 jmax = imax + icamax_(&i__1, &w[imax + 1 + (kw - 1) * w_dim1],
880 i__1 = jmax + (kw - 1) * w_dim1;
881 rowmax = (r__1 = w[i__1].r, abs(r__1)) + (r__2 = r_imag(&w[
882 jmax + (kw - 1) * w_dim1]), abs(r__2));
885 jmax = icamax_(&i__1, &w[(kw - 1) * w_dim1 + 1], &c__1);
887 i__1 = jmax + (kw - 1) * w_dim1;
888 r__3 = rowmax, r__4 = (r__1 = w[i__1].r, abs(r__1)) + (
889 r__2 = r_imag(&w[jmax + (kw - 1) * w_dim1]), abs(
891 rowmax = f2cmax(r__3,r__4);
895 if (absakk >= alpha * colmax * (colmax / rowmax)) {
897 /* no interchange, use 1-by-1 pivot block */
902 } else /* if(complicated condition) */ {
903 i__1 = imax + (kw - 1) * w_dim1;
904 if ((r__1 = w[i__1].r, abs(r__1)) >= alpha * rowmax) {
906 /* interchange rows and columns K and IMAX, use 1-by-1 */
911 /* copy column KW-1 of W to column KW of W */
913 ccopy_(&k, &w[(kw - 1) * w_dim1 + 1], &c__1, &w[kw *
919 /* interchange rows and columns K-1 and IMAX, use 2-by-2 */
928 /* END pivot search along IMAX row */
932 /* END pivot search */
934 /* ============================================================ */
936 /* KK is the column of A where pivoting step stopped */
940 /* KKW is the column of W which corresponds to column KK of A */
944 /* Interchange rows and columns KP and KK. */
945 /* Updated column KP is already stored in column KKW of W. */
949 /* Copy non-updated column KK to column KP of submatrix A */
950 /* at step K. No need to copy element into column K */
951 /* (or K and K-1 for 2-by-2 pivot) of A, since these columns */
952 /* will be later overwritten. */
954 i__1 = kp + kp * a_dim1;
955 i__2 = kk + kk * a_dim1;
957 a[i__1].r = r__1, a[i__1].i = 0.f;
959 ccopy_(&i__1, &a[kp + 1 + kk * a_dim1], &c__1, &a[kp + (kp +
962 clacgv_(&i__1, &a[kp + (kp + 1) * a_dim1], lda);
965 ccopy_(&i__1, &a[kk * a_dim1 + 1], &c__1, &a[kp * a_dim1
969 /* Interchange rows KK and KP in last K+1 to N columns of A */
970 /* (columns K (or K and K-1 for 2-by-2 pivot) of A will be */
971 /* later overwritten). Interchange rows KK and KP */
972 /* in last KKW to NB columns of W. */
976 cswap_(&i__1, &a[kk + (k + 1) * a_dim1], lda, &a[kp + (k
977 + 1) * a_dim1], lda);
980 cswap_(&i__1, &w[kk + kkw * w_dim1], ldw, &w[kp + kkw *
986 /* 1-by-1 pivot block D(k): column kw of W now holds */
988 /* W(kw) = U(k)*D(k), */
990 /* where U(k) is the k-th column of U */
992 /* (1) Store subdiag. elements of column U(k) */
993 /* and 1-by-1 block D(k) in column k of A. */
994 /* (NOTE: Diagonal element U(k,k) is a UNIT element */
995 /* and not stored) */
996 /* A(k,k) := D(k,k) = W(k,kw) */
997 /* A(1:k-1,k) := U(1:k-1,k) = W(1:k-1,kw)/D(k,k) */
999 /* (NOTE: No need to use for Hermitian matrix */
1000 /* A( K, K ) = DBLE( W( K, K) ) to separately copy diagonal */
1001 /* element D(k,k) from W (potentially saves only one load)) */
1002 ccopy_(&k, &w[kw * w_dim1 + 1], &c__1, &a[k * a_dim1 + 1], &
1006 /* (NOTE: No need to check if A(k,k) is NOT ZERO, */
1007 /* since that was ensured earlier in pivot search: */
1008 /* case A(k,k) = 0 falls into 2x2 pivot case(4)) */
1010 i__1 = k + k * a_dim1;
1011 r1 = 1.f / a[i__1].r;
1013 csscal_(&i__1, &r1, &a[k * a_dim1 + 1], &c__1);
1015 /* (2) Conjugate column W(kw) */
1018 clacgv_(&i__1, &w[kw * w_dim1 + 1], &c__1);
1023 /* 2-by-2 pivot block D(k): columns kw and kw-1 of W now hold */
1025 /* ( W(kw-1) W(kw) ) = ( U(k-1) U(k) )*D(k) */
1027 /* where U(k) and U(k-1) are the k-th and (k-1)-th columns */
1030 /* (1) Store U(1:k-2,k-1) and U(1:k-2,k) and 2-by-2 */
1031 /* block D(k-1:k,k-1:k) in columns k-1 and k of A. */
1032 /* (NOTE: 2-by-2 diagonal block U(k-1:k,k-1:k) is a UNIT */
1033 /* block and not stored) */
1034 /* A(k-1:k,k-1:k) := D(k-1:k,k-1:k) = W(k-1:k,kw-1:kw) */
1035 /* A(1:k-2,k-1:k) := U(1:k-2,k:k-1:k) = */
1036 /* = W(1:k-2,kw-1:kw) * ( D(k-1:k,k-1:k)**(-1) ) */
1040 /* Factor out the columns of the inverse of 2-by-2 pivot */
1041 /* block D, so that each column contains 1, to reduce the */
1042 /* number of FLOPS when we multiply panel */
1043 /* ( W(kw-1) W(kw) ) by this inverse, i.e. by D**(-1). */
1045 /* D**(-1) = ( d11 cj(d21) )**(-1) = */
1048 /* = 1/(d11*d22-|d21|**2) * ( ( d22) (-cj(d21) ) ) = */
1049 /* ( (-d21) ( d11 ) ) */
1051 /* = 1/(|d21|**2) * 1/((d11/cj(d21))*(d22/d21)-1) * */
1053 /* * ( d21*( d22/d21 ) conj(d21)*( - 1 ) ) = */
1054 /* ( ( -1 ) ( d11/conj(d21) ) ) */
1056 /* = 1/(|d21|**2) * 1/(D22*D11-1) * */
1058 /* * ( d21*( D11 ) conj(d21)*( -1 ) ) = */
1059 /* ( ( -1 ) ( D22 ) ) */
1061 /* = (1/|d21|**2) * T * ( d21*( D11 ) conj(d21)*( -1 ) ) = */
1062 /* ( ( -1 ) ( D22 ) ) */
1064 /* = ( (T/conj(d21))*( D11 ) (T/d21)*( -1 ) ) = */
1065 /* ( ( -1 ) ( D22 ) ) */
1067 /* = ( conj(D21)*( D11 ) D21*( -1 ) ) */
1068 /* ( ( -1 ) ( D22 ) ), */
1070 /* where D11 = d22/d21, */
1071 /* D22 = d11/conj(d21), */
1073 /* T = 1/(D22*D11-1). */
1075 /* (NOTE: No need to check for division by ZERO, */
1076 /* since that was ensured earlier in pivot search: */
1077 /* (a) d21 != 0, since in 2x2 pivot case(4) */
1078 /* |d21| should be larger than |d11| and |d22|; */
1079 /* (b) (D22*D11 - 1) != 0, since from (a), */
1080 /* both |D11| < 1, |D22| < 1, hence |D22*D11| << 1.) */
1082 i__1 = k - 1 + kw * w_dim1;
1083 d21.r = w[i__1].r, d21.i = w[i__1].i;
1084 r_cnjg(&q__2, &d21);
1085 c_div(&q__1, &w[k + kw * w_dim1], &q__2);
1086 d11.r = q__1.r, d11.i = q__1.i;
1087 c_div(&q__1, &w[k - 1 + (kw - 1) * w_dim1], &d21);
1088 d22.r = q__1.r, d22.i = q__1.i;
1089 q__1.r = d11.r * d22.r - d11.i * d22.i, q__1.i = d11.r *
1090 d22.i + d11.i * d22.r;
1091 t = 1.f / (q__1.r - 1.f);
1092 q__2.r = t, q__2.i = 0.f;
1093 c_div(&q__1, &q__2, &d21);
1094 d21.r = q__1.r, d21.i = q__1.i;
1096 /* Update elements in columns A(k-1) and A(k) as */
1097 /* dot products of rows of ( W(kw-1) W(kw) ) and columns */
1101 for (j = 1; j <= i__1; ++j) {
1102 i__2 = j + (k - 1) * a_dim1;
1103 i__3 = j + (kw - 1) * w_dim1;
1104 q__3.r = d11.r * w[i__3].r - d11.i * w[i__3].i,
1105 q__3.i = d11.r * w[i__3].i + d11.i * w[i__3]
1107 i__4 = j + kw * w_dim1;
1108 q__2.r = q__3.r - w[i__4].r, q__2.i = q__3.i - w[i__4]
1110 q__1.r = d21.r * q__2.r - d21.i * q__2.i, q__1.i =
1111 d21.r * q__2.i + d21.i * q__2.r;
1112 a[i__2].r = q__1.r, a[i__2].i = q__1.i;
1113 i__2 = j + k * a_dim1;
1114 r_cnjg(&q__2, &d21);
1115 i__3 = j + kw * w_dim1;
1116 q__4.r = d22.r * w[i__3].r - d22.i * w[i__3].i,
1117 q__4.i = d22.r * w[i__3].i + d22.i * w[i__3]
1119 i__4 = j + (kw - 1) * w_dim1;
1120 q__3.r = q__4.r - w[i__4].r, q__3.i = q__4.i - w[i__4]
1122 q__1.r = q__2.r * q__3.r - q__2.i * q__3.i, q__1.i =
1123 q__2.r * q__3.i + q__2.i * q__3.r;
1124 a[i__2].r = q__1.r, a[i__2].i = q__1.i;
1129 /* Copy D(k) to A */
1131 i__1 = k - 1 + (k - 1) * a_dim1;
1132 i__2 = k - 1 + (kw - 1) * w_dim1;
1133 a[i__1].r = w[i__2].r, a[i__1].i = w[i__2].i;
1134 i__1 = k - 1 + k * a_dim1;
1135 i__2 = k - 1 + kw * w_dim1;
1136 a[i__1].r = w[i__2].r, a[i__1].i = w[i__2].i;
1137 i__1 = k + k * a_dim1;
1138 i__2 = k + kw * w_dim1;
1139 a[i__1].r = w[i__2].r, a[i__1].i = w[i__2].i;
1141 /* (2) Conjugate columns W(kw) and W(kw-1) */
1144 clacgv_(&i__1, &w[kw * w_dim1 + 1], &c__1);
1146 clacgv_(&i__1, &w[(kw - 1) * w_dim1 + 1], &c__1);
1152 /* Store details of the interchanges in IPIV */
1161 /* Decrease K and return to the start of the main loop */
1168 /* Update the upper triangle of A11 (= A(1:k,1:k)) as */
1170 /* A11 := A11 - U12*D*U12**H = A11 - U12*W**H */
1172 /* computing blocks of NB columns at a time (note that conjg(W) is */
1173 /* actually stored) */
1176 for (j = (k - 1) / *nb * *nb + 1; i__1 < 0 ? j >= 1 : j <= 1; j +=
1179 i__2 = *nb, i__3 = k - j + 1;
1180 jb = f2cmin(i__2,i__3);
1182 /* Update the upper triangle of the diagonal block */
1185 for (jj = j; jj <= i__2; ++jj) {
1186 i__3 = jj + jj * a_dim1;
1187 i__4 = jj + jj * a_dim1;
1189 a[i__3].r = r__1, a[i__3].i = 0.f;
1192 q__1.r = -1.f, q__1.i = 0.f;
1193 cgemv_("No transpose", &i__3, &i__4, &q__1, &a[j + (k + 1) *
1194 a_dim1], lda, &w[jj + (kw + 1) * w_dim1], ldw, &c_b1,
1195 &a[j + jj * a_dim1], &c__1);
1196 i__3 = jj + jj * a_dim1;
1197 i__4 = jj + jj * a_dim1;
1199 a[i__3].r = r__1, a[i__3].i = 0.f;
1203 /* Update the rectangular superdiagonal block */
1207 q__1.r = -1.f, q__1.i = 0.f;
1208 cgemm_("No transpose", "Transpose", &i__2, &jb, &i__3, &q__1, &a[(
1209 k + 1) * a_dim1 + 1], lda, &w[j + (kw + 1) * w_dim1], ldw,
1210 &c_b1, &a[j * a_dim1 + 1], lda);
1214 /* Put U12 in standard form by partially undoing the interchanges */
1215 /* in of rows in columns k+1:n looping backwards from k+1 to n */
1220 /* Undo the interchanges (if any) of rows J and JP */
1221 /* at each step J */
1223 /* (Here, J is a diagonal index) */
1228 /* (Here, J is a diagonal index) */
1231 /* (NOTE: Here, J is used to determine row length. Length N-J+1 */
1232 /* of the rows to swap back doesn't include diagonal element) */
1234 if (jp != jj && j <= *n) {
1236 cswap_(&i__1, &a[jp + j * a_dim1], lda, &a[jj + j * a_dim1], lda);
1242 /* Set KB to the number of columns factorized */
1248 /* Factorize the leading columns of A using the lower triangle */
1249 /* of A and working forwards, and compute the matrix W = L21*D */
1250 /* for use in updating A22 (note that conjg(W) is actually stored) */
1252 /* K is the main loop index, increasing from 1 in steps of 1 or 2 */
1257 /* Exit from loop */
1259 if (k >= *nb && *nb < *n || k > *n) {
1265 /* Copy column K of A to column K of W and update it */
1267 i__1 = k + k * w_dim1;
1268 i__2 = k + k * a_dim1;
1270 w[i__1].r = r__1, w[i__1].i = 0.f;
1273 ccopy_(&i__1, &a[k + 1 + k * a_dim1], &c__1, &w[k + 1 + k *
1278 q__1.r = -1.f, q__1.i = 0.f;
1279 cgemv_("No transpose", &i__1, &i__2, &q__1, &a[k + a_dim1], lda, &w[k
1280 + w_dim1], ldw, &c_b1, &w[k + k * w_dim1], &c__1);
1281 i__1 = k + k * w_dim1;
1282 i__2 = k + k * w_dim1;
1284 w[i__1].r = r__1, w[i__1].i = 0.f;
1286 /* Determine rows and columns to be interchanged and whether */
1287 /* a 1-by-1 or 2-by-2 pivot block will be used */
1289 i__1 = k + k * w_dim1;
1290 absakk = (r__1 = w[i__1].r, abs(r__1));
1292 /* IMAX is the row-index of the largest off-diagonal element in */
1293 /* column K, and COLMAX is its absolute value. */
1294 /* Determine both COLMAX and IMAX. */
1298 imax = k + icamax_(&i__1, &w[k + 1 + k * w_dim1], &c__1);
1299 i__1 = imax + k * w_dim1;
1300 colmax = (r__1 = w[i__1].r, abs(r__1)) + (r__2 = r_imag(&w[imax +
1301 k * w_dim1]), abs(r__2));
1306 if (f2cmax(absakk,colmax) == 0.f) {
1308 /* Column K is zero or underflow: set INFO and continue */
1314 i__1 = k + k * a_dim1;
1315 i__2 = k + k * a_dim1;
1317 a[i__1].r = r__1, a[i__1].i = 0.f;
1320 /* ============================================================ */
1322 /* BEGIN pivot search */
1325 if (absakk >= alpha * colmax) {
1327 /* no interchange, use 1-by-1 pivot block */
1332 /* BEGIN pivot search along IMAX row */
1335 /* Copy column IMAX to column K+1 of W and update it */
1338 ccopy_(&i__1, &a[imax + k * a_dim1], lda, &w[k + (k + 1) *
1341 clacgv_(&i__1, &w[k + (k + 1) * w_dim1], &c__1);
1342 i__1 = imax + (k + 1) * w_dim1;
1343 i__2 = imax + imax * a_dim1;
1345 w[i__1].r = r__1, w[i__1].i = 0.f;
1348 ccopy_(&i__1, &a[imax + 1 + imax * a_dim1], &c__1, &w[
1349 imax + 1 + (k + 1) * w_dim1], &c__1);
1353 q__1.r = -1.f, q__1.i = 0.f;
1354 cgemv_("No transpose", &i__1, &i__2, &q__1, &a[k + a_dim1],
1355 lda, &w[imax + w_dim1], ldw, &c_b1, &w[k + (k + 1) *
1357 i__1 = imax + (k + 1) * w_dim1;
1358 i__2 = imax + (k + 1) * w_dim1;
1360 w[i__1].r = r__1, w[i__1].i = 0.f;
1362 /* JMAX is the column-index of the largest off-diagonal */
1363 /* element in row IMAX, and ROWMAX is its absolute value. */
1364 /* Determine only ROWMAX. */
1367 jmax = k - 1 + icamax_(&i__1, &w[k + (k + 1) * w_dim1], &c__1)
1369 i__1 = jmax + (k + 1) * w_dim1;
1370 rowmax = (r__1 = w[i__1].r, abs(r__1)) + (r__2 = r_imag(&w[
1371 jmax + (k + 1) * w_dim1]), abs(r__2));
1374 jmax = imax + icamax_(&i__1, &w[imax + 1 + (k + 1) *
1377 i__1 = jmax + (k + 1) * w_dim1;
1378 r__3 = rowmax, r__4 = (r__1 = w[i__1].r, abs(r__1)) + (
1379 r__2 = r_imag(&w[jmax + (k + 1) * w_dim1]), abs(
1381 rowmax = f2cmax(r__3,r__4);
1385 if (absakk >= alpha * colmax * (colmax / rowmax)) {
1387 /* no interchange, use 1-by-1 pivot block */
1392 } else /* if(complicated condition) */ {
1393 i__1 = imax + (k + 1) * w_dim1;
1394 if ((r__1 = w[i__1].r, abs(r__1)) >= alpha * rowmax) {
1396 /* interchange rows and columns K and IMAX, use 1-by-1 */
1401 /* copy column K+1 of W to column K of W */
1404 ccopy_(&i__1, &w[k + (k + 1) * w_dim1], &c__1, &w[k +
1405 k * w_dim1], &c__1);
1410 /* interchange rows and columns K+1 and IMAX, use 2-by-2 */
1419 /* END pivot search along IMAX row */
1423 /* END pivot search */
1425 /* ============================================================ */
1427 /* KK is the column of A where pivoting step stopped */
1431 /* Interchange rows and columns KP and KK. */
1432 /* Updated column KP is already stored in column KK of W. */
1436 /* Copy non-updated column KK to column KP of submatrix A */
1437 /* at step K. No need to copy element into column K */
1438 /* (or K and K+1 for 2-by-2 pivot) of A, since these columns */
1439 /* will be later overwritten. */
1441 i__1 = kp + kp * a_dim1;
1442 i__2 = kk + kk * a_dim1;
1444 a[i__1].r = r__1, a[i__1].i = 0.f;
1446 ccopy_(&i__1, &a[kk + 1 + kk * a_dim1], &c__1, &a[kp + (kk +
1449 clacgv_(&i__1, &a[kp + (kk + 1) * a_dim1], lda);
1452 ccopy_(&i__1, &a[kp + 1 + kk * a_dim1], &c__1, &a[kp + 1
1453 + kp * a_dim1], &c__1);
1456 /* Interchange rows KK and KP in first K-1 columns of A */
1457 /* (columns K (or K and K+1 for 2-by-2 pivot) of A will be */
1458 /* later overwritten). Interchange rows KK and KP */
1459 /* in first KK columns of W. */
1463 cswap_(&i__1, &a[kk + a_dim1], lda, &a[kp + a_dim1], lda);
1465 cswap_(&kk, &w[kk + w_dim1], ldw, &w[kp + w_dim1], ldw);
1470 /* 1-by-1 pivot block D(k): column k of W now holds */
1472 /* W(k) = L(k)*D(k), */
1474 /* where L(k) is the k-th column of L */
1476 /* (1) Store subdiag. elements of column L(k) */
1477 /* and 1-by-1 block D(k) in column k of A. */
1478 /* (NOTE: Diagonal element L(k,k) is a UNIT element */
1479 /* and not stored) */
1480 /* A(k,k) := D(k,k) = W(k,k) */
1481 /* A(k+1:N,k) := L(k+1:N,k) = W(k+1:N,k)/D(k,k) */
1483 /* (NOTE: No need to use for Hermitian matrix */
1484 /* A( K, K ) = DBLE( W( K, K) ) to separately copy diagonal */
1485 /* element D(k,k) from W (potentially saves only one load)) */
1487 ccopy_(&i__1, &w[k + k * w_dim1], &c__1, &a[k + k * a_dim1], &
1491 /* (NOTE: No need to check if A(k,k) is NOT ZERO, */
1492 /* since that was ensured earlier in pivot search: */
1493 /* case A(k,k) = 0 falls into 2x2 pivot case(4)) */
1495 i__1 = k + k * a_dim1;
1496 r1 = 1.f / a[i__1].r;
1498 csscal_(&i__1, &r1, &a[k + 1 + k * a_dim1], &c__1);
1500 /* (2) Conjugate column W(k) */
1503 clacgv_(&i__1, &w[k + 1 + k * w_dim1], &c__1);
1508 /* 2-by-2 pivot block D(k): columns k and k+1 of W now hold */
1510 /* ( W(k) W(k+1) ) = ( L(k) L(k+1) )*D(k) */
1512 /* where L(k) and L(k+1) are the k-th and (k+1)-th columns */
1515 /* (1) Store L(k+2:N,k) and L(k+2:N,k+1) and 2-by-2 */
1516 /* block D(k:k+1,k:k+1) in columns k and k+1 of A. */
1517 /* (NOTE: 2-by-2 diagonal block L(k:k+1,k:k+1) is a UNIT */
1518 /* block and not stored) */
1519 /* A(k:k+1,k:k+1) := D(k:k+1,k:k+1) = W(k:k+1,k:k+1) */
1520 /* A(k+2:N,k:k+1) := L(k+2:N,k:k+1) = */
1521 /* = W(k+2:N,k:k+1) * ( D(k:k+1,k:k+1)**(-1) ) */
1525 /* Factor out the columns of the inverse of 2-by-2 pivot */
1526 /* block D, so that each column contains 1, to reduce the */
1527 /* number of FLOPS when we multiply panel */
1528 /* ( W(kw-1) W(kw) ) by this inverse, i.e. by D**(-1). */
1530 /* D**(-1) = ( d11 cj(d21) )**(-1) = */
1533 /* = 1/(d11*d22-|d21|**2) * ( ( d22) (-cj(d21) ) ) = */
1534 /* ( (-d21) ( d11 ) ) */
1536 /* = 1/(|d21|**2) * 1/((d11/cj(d21))*(d22/d21)-1) * */
1538 /* * ( d21*( d22/d21 ) conj(d21)*( - 1 ) ) = */
1539 /* ( ( -1 ) ( d11/conj(d21) ) ) */
1541 /* = 1/(|d21|**2) * 1/(D22*D11-1) * */
1543 /* * ( d21*( D11 ) conj(d21)*( -1 ) ) = */
1544 /* ( ( -1 ) ( D22 ) ) */
1546 /* = (1/|d21|**2) * T * ( d21*( D11 ) conj(d21)*( -1 ) ) = */
1547 /* ( ( -1 ) ( D22 ) ) */
1549 /* = ( (T/conj(d21))*( D11 ) (T/d21)*( -1 ) ) = */
1550 /* ( ( -1 ) ( D22 ) ) */
1552 /* = ( conj(D21)*( D11 ) D21*( -1 ) ) */
1553 /* ( ( -1 ) ( D22 ) ) */
1555 /* where D11 = d22/d21, */
1556 /* D22 = d11/conj(d21), */
1558 /* T = 1/(D22*D11-1). */
1560 /* (NOTE: No need to check for division by ZERO, */
1561 /* since that was ensured earlier in pivot search: */
1562 /* (a) d21 != 0, since in 2x2 pivot case(4) */
1563 /* |d21| should be larger than |d11| and |d22|; */
1564 /* (b) (D22*D11 - 1) != 0, since from (a), */
1565 /* both |D11| < 1, |D22| < 1, hence |D22*D11| << 1.) */
1567 i__1 = k + 1 + k * w_dim1;
1568 d21.r = w[i__1].r, d21.i = w[i__1].i;
1569 c_div(&q__1, &w[k + 1 + (k + 1) * w_dim1], &d21);
1570 d11.r = q__1.r, d11.i = q__1.i;
1571 r_cnjg(&q__2, &d21);
1572 c_div(&q__1, &w[k + k * w_dim1], &q__2);
1573 d22.r = q__1.r, d22.i = q__1.i;
1574 q__1.r = d11.r * d22.r - d11.i * d22.i, q__1.i = d11.r *
1575 d22.i + d11.i * d22.r;
1576 t = 1.f / (q__1.r - 1.f);
1577 q__2.r = t, q__2.i = 0.f;
1578 c_div(&q__1, &q__2, &d21);
1579 d21.r = q__1.r, d21.i = q__1.i;
1581 /* Update elements in columns A(k) and A(k+1) as */
1582 /* dot products of rows of ( W(k) W(k+1) ) and columns */
1586 for (j = k + 2; j <= i__1; ++j) {
1587 i__2 = j + k * a_dim1;
1588 r_cnjg(&q__2, &d21);
1589 i__3 = j + k * w_dim1;
1590 q__4.r = d11.r * w[i__3].r - d11.i * w[i__3].i,
1591 q__4.i = d11.r * w[i__3].i + d11.i * w[i__3]
1593 i__4 = j + (k + 1) * w_dim1;
1594 q__3.r = q__4.r - w[i__4].r, q__3.i = q__4.i - w[i__4]
1596 q__1.r = q__2.r * q__3.r - q__2.i * q__3.i, q__1.i =
1597 q__2.r * q__3.i + q__2.i * q__3.r;
1598 a[i__2].r = q__1.r, a[i__2].i = q__1.i;
1599 i__2 = j + (k + 1) * a_dim1;
1600 i__3 = j + (k + 1) * w_dim1;
1601 q__3.r = d22.r * w[i__3].r - d22.i * w[i__3].i,
1602 q__3.i = d22.r * w[i__3].i + d22.i * w[i__3]
1604 i__4 = j + k * w_dim1;
1605 q__2.r = q__3.r - w[i__4].r, q__2.i = q__3.i - w[i__4]
1607 q__1.r = d21.r * q__2.r - d21.i * q__2.i, q__1.i =
1608 d21.r * q__2.i + d21.i * q__2.r;
1609 a[i__2].r = q__1.r, a[i__2].i = q__1.i;
1614 /* Copy D(k) to A */
1616 i__1 = k + k * a_dim1;
1617 i__2 = k + k * w_dim1;
1618 a[i__1].r = w[i__2].r, a[i__1].i = w[i__2].i;
1619 i__1 = k + 1 + k * a_dim1;
1620 i__2 = k + 1 + k * w_dim1;
1621 a[i__1].r = w[i__2].r, a[i__1].i = w[i__2].i;
1622 i__1 = k + 1 + (k + 1) * a_dim1;
1623 i__2 = k + 1 + (k + 1) * w_dim1;
1624 a[i__1].r = w[i__2].r, a[i__1].i = w[i__2].i;
1626 /* (2) Conjugate columns W(k) and W(k+1) */
1629 clacgv_(&i__1, &w[k + 1 + k * w_dim1], &c__1);
1631 clacgv_(&i__1, &w[k + 2 + (k + 1) * w_dim1], &c__1);
1637 /* Store details of the interchanges in IPIV */
1646 /* Increase K and return to the start of the main loop */
1653 /* Update the lower triangle of A22 (= A(k:n,k:n)) as */
1655 /* A22 := A22 - L21*D*L21**H = A22 - L21*W**H */
1657 /* computing blocks of NB columns at a time (note that conjg(W) is */
1658 /* actually stored) */
1662 for (j = k; i__2 < 0 ? j >= i__1 : j <= i__1; j += i__2) {
1664 i__3 = *nb, i__4 = *n - j + 1;
1665 jb = f2cmin(i__3,i__4);
1667 /* Update the lower triangle of the diagonal block */
1670 for (jj = j; jj <= i__3; ++jj) {
1671 i__4 = jj + jj * a_dim1;
1672 i__5 = jj + jj * a_dim1;
1674 a[i__4].r = r__1, a[i__4].i = 0.f;
1677 q__1.r = -1.f, q__1.i = 0.f;
1678 cgemv_("No transpose", &i__4, &i__5, &q__1, &a[jj + a_dim1],
1679 lda, &w[jj + w_dim1], ldw, &c_b1, &a[jj + jj * a_dim1]
1681 i__4 = jj + jj * a_dim1;
1682 i__5 = jj + jj * a_dim1;
1684 a[i__4].r = r__1, a[i__4].i = 0.f;
1688 /* Update the rectangular subdiagonal block */
1691 i__3 = *n - j - jb + 1;
1693 q__1.r = -1.f, q__1.i = 0.f;
1694 cgemm_("No transpose", "Transpose", &i__3, &jb, &i__4, &q__1,
1695 &a[j + jb + a_dim1], lda, &w[j + w_dim1], ldw, &c_b1,
1696 &a[j + jb + j * a_dim1], lda);
1701 /* Put L21 in standard form by partially undoing the interchanges */
1702 /* of rows in columns 1:k-1 looping backwards from k-1 to 1 */
1707 /* Undo the interchanges (if any) of rows J and JP */
1708 /* at each step J */
1710 /* (Here, J is a diagonal index) */
1715 /* (Here, J is a diagonal index) */
1718 /* (NOTE: Here, J is used to determine row length. Length J */
1719 /* of the rows to swap back doesn't include diagonal element) */
1721 if (jp != jj && j >= 1) {
1722 cswap_(&j, &a[jp + a_dim1], lda, &a[jj + a_dim1], lda);
1728 /* Set KB to the number of columns factorized */