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]/Cd(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 doublecomplex c_b1 = {1.,0.};
516 static integer c__1 = 1;
518 /* > \brief \b ZLAHEF_RK computes a partial factorization of a complex Hermitian indefinite matrix using bound
519 ed Bunch-Kaufman (rook) diagonal pivoting method. */
521 /* =========== DOCUMENTATION =========== */
523 /* Online html documentation available at */
524 /* http://www.netlib.org/lapack/explore-html/ */
527 /* > Download ZLAHEF_RK + dependencies */
528 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/zlahef_
531 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/zlahef_
534 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/zlahef_
542 /* SUBROUTINE ZLAHEF_RK( UPLO, N, NB, KB, A, LDA, E, IPIV, W, LDW, */
546 /* INTEGER INFO, KB, LDA, LDW, N, NB */
547 /* INTEGER IPIV( * ) */
548 /* COMPLEX*16 A( LDA, * ), E( * ), W( LDW, * ) */
551 /* > \par Purpose: */
555 /* > ZLAHEF_RK computes a partial factorization of a complex Hermitian */
556 /* > matrix A using the bounded Bunch-Kaufman (rook) diagonal */
557 /* > pivoting method. The 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. */
568 /* > ZLAHEF_RK is an auxiliary routine called by ZHETRF_RK. It uses */
569 /* > blocked code (calling Level 3 BLAS) to update the submatrix */
570 /* > A11 (if UPLO = 'U') or A22 (if UPLO = 'L'). */
576 /* > \param[in] UPLO */
578 /* > UPLO is CHARACTER*1 */
579 /* > Specifies whether the upper or lower triangular part of the */
580 /* > Hermitian matrix A is stored: */
581 /* > = 'U': Upper triangular */
582 /* > = 'L': Lower triangular */
588 /* > The order of the matrix A. N >= 0. */
591 /* > \param[in] NB */
593 /* > NB is INTEGER */
594 /* > The maximum number of columns of the matrix A that should be */
595 /* > factored. NB should be at least 2 to allow for 2-by-2 pivot */
599 /* > \param[out] KB */
601 /* > KB is INTEGER */
602 /* > The number of columns of A that were actually factored. */
603 /* > KB is either NB-1 or NB, or N if N <= NB. */
606 /* > \param[in,out] A */
608 /* > A is COMPLEX*16 array, dimension (LDA,N) */
609 /* > On entry, the Hermitian matrix A. */
610 /* > If UPLO = 'U': the leading N-by-N upper triangular part */
611 /* > of A contains the upper triangular part of the matrix A, */
612 /* > and the strictly lower triangular part of A is not */
615 /* > If UPLO = 'L': the leading N-by-N lower triangular part */
616 /* > of A contains the lower triangular part of the matrix A, */
617 /* > and the strictly upper triangular part of A is not */
620 /* > On exit, contains: */
621 /* > a) ONLY diagonal elements of the Hermitian block diagonal */
622 /* > matrix D on the diagonal of A, i.e. D(k,k) = A(k,k); */
623 /* > (superdiagonal (or subdiagonal) elements of D */
624 /* > are stored on exit in array E), and */
625 /* > b) If UPLO = 'U': factor U in the superdiagonal part of A. */
626 /* > If UPLO = 'L': factor L in the subdiagonal part of A. */
629 /* > \param[in] LDA */
631 /* > LDA is INTEGER */
632 /* > The leading dimension of the array A. LDA >= f2cmax(1,N). */
635 /* > \param[out] E */
637 /* > E is COMPLEX*16 array, dimension (N) */
638 /* > On exit, contains the superdiagonal (or subdiagonal) */
639 /* > elements of the Hermitian block diagonal matrix D */
640 /* > with 1-by-1 or 2-by-2 diagonal blocks, where */
641 /* > If UPLO = 'U': E(i) = D(i-1,i), i=2:N, E(1) is set to 0; */
642 /* > If UPLO = 'L': E(i) = D(i+1,i), i=1:N-1, E(N) is set to 0. */
644 /* > NOTE: For 1-by-1 diagonal block D(k), where */
645 /* > 1 <= k <= N, the element E(k) is set to 0 in both */
646 /* > UPLO = 'U' or UPLO = 'L' cases. */
649 /* > \param[out] IPIV */
651 /* > IPIV is INTEGER array, dimension (N) */
652 /* > IPIV describes the permutation matrix P in the factorization */
653 /* > of matrix A as follows. The absolute value of IPIV(k) */
654 /* > represents the index of row and column that were */
655 /* > interchanged with the k-th row and column. The value of UPLO */
656 /* > describes the order in which the interchanges were applied. */
657 /* > Also, the sign of IPIV represents the block structure of */
658 /* > the Hermitian block diagonal matrix D with 1-by-1 or 2-by-2 */
659 /* > diagonal blocks which correspond to 1 or 2 interchanges */
660 /* > at each factorization step. */
662 /* > If UPLO = 'U', */
663 /* > ( in factorization order, k decreases from N to 1 ): */
664 /* > a) A single positive entry IPIV(k) > 0 means: */
665 /* > D(k,k) is a 1-by-1 diagonal block. */
666 /* > If IPIV(k) != k, rows and columns k and IPIV(k) were */
667 /* > interchanged in the submatrix A(1:N,N-KB+1:N); */
668 /* > If IPIV(k) = k, no interchange occurred. */
671 /* > b) A pair of consecutive negative entries */
672 /* > IPIV(k) < 0 and IPIV(k-1) < 0 means: */
673 /* > D(k-1:k,k-1:k) is a 2-by-2 diagonal block. */
674 /* > (NOTE: negative entries in IPIV appear ONLY in pairs). */
675 /* > 1) If -IPIV(k) != k, rows and columns */
676 /* > k and -IPIV(k) were interchanged */
677 /* > in the matrix A(1:N,N-KB+1:N). */
678 /* > If -IPIV(k) = k, no interchange occurred. */
679 /* > 2) If -IPIV(k-1) != k-1, rows and columns */
680 /* > k-1 and -IPIV(k-1) were interchanged */
681 /* > in the submatrix A(1:N,N-KB+1:N). */
682 /* > If -IPIV(k-1) = k-1, no interchange occurred. */
684 /* > c) In both cases a) and b) is always ABS( IPIV(k) ) <= k. */
686 /* > d) NOTE: Any entry IPIV(k) is always NONZERO on output. */
688 /* > If UPLO = 'L', */
689 /* > ( in factorization order, k increases from 1 to N ): */
690 /* > a) A single positive entry IPIV(k) > 0 means: */
691 /* > D(k,k) is a 1-by-1 diagonal block. */
692 /* > If IPIV(k) != k, rows and columns k and IPIV(k) were */
693 /* > interchanged in the submatrix A(1:N,1:KB). */
694 /* > If IPIV(k) = k, no interchange occurred. */
696 /* > b) A pair of consecutive negative entries */
697 /* > IPIV(k) < 0 and IPIV(k+1) < 0 means: */
698 /* > D(k:k+1,k:k+1) is a 2-by-2 diagonal block. */
699 /* > (NOTE: negative entries in IPIV appear ONLY in pairs). */
700 /* > 1) If -IPIV(k) != k, rows and columns */
701 /* > k and -IPIV(k) were interchanged */
702 /* > in the submatrix A(1:N,1:KB). */
703 /* > If -IPIV(k) = k, no interchange occurred. */
704 /* > 2) If -IPIV(k+1) != k+1, rows and columns */
705 /* > k-1 and -IPIV(k-1) were interchanged */
706 /* > in the submatrix A(1:N,1:KB). */
707 /* > If -IPIV(k+1) = k+1, no interchange occurred. */
709 /* > c) In both cases a) and b) is always ABS( IPIV(k) ) >= k. */
711 /* > d) NOTE: Any entry IPIV(k) is always NONZERO on output. */
714 /* > \param[out] W */
716 /* > W is COMPLEX*16 array, dimension (LDW,NB) */
719 /* > \param[in] LDW */
721 /* > LDW is INTEGER */
722 /* > The leading dimension of the array W. LDW >= f2cmax(1,N). */
725 /* > \param[out] INFO */
727 /* > INFO is INTEGER */
728 /* > = 0: successful exit */
730 /* > < 0: If INFO = -k, the k-th argument had an illegal value */
732 /* > > 0: If INFO = k, the matrix A is singular, because: */
733 /* > If UPLO = 'U': column k in the upper */
734 /* > triangular part of A contains all zeros. */
735 /* > If UPLO = 'L': column k in the lower */
736 /* > triangular part of A contains all zeros. */
738 /* > Therefore D(k,k) is exactly zero, and superdiagonal */
739 /* > elements of column k of U (or subdiagonal elements of */
740 /* > column k of L ) are all zeros. The factorization has */
741 /* > been completed, but the block diagonal matrix D is */
742 /* > exactly singular, and division by zero will occur if */
743 /* > it is used to solve a system of equations. */
745 /* > NOTE: INFO only stores the first occurrence of */
746 /* > a singularity, any subsequent occurrence of singularity */
747 /* > is not stored in INFO even though the factorization */
748 /* > always completes. */
754 /* > \author Univ. of Tennessee */
755 /* > \author Univ. of California Berkeley */
756 /* > \author Univ. of Colorado Denver */
757 /* > \author NAG Ltd. */
759 /* > \date December 2016 */
761 /* > \ingroup complex16HEcomputational */
763 /* > \par Contributors: */
764 /* ================== */
768 /* > December 2016, Igor Kozachenko, */
769 /* > Computer Science Division, */
770 /* > University of California, Berkeley */
772 /* > September 2007, Sven Hammarling, Nicholas J. Higham, Craig Lucas, */
773 /* > School of Mathematics, */
774 /* > University of Manchester */
778 /* ===================================================================== */
779 /* Subroutine */ int zlahef_rk_(char *uplo, integer *n, integer *nb, integer
780 *kb, doublecomplex *a, integer *lda, doublecomplex *e, integer *ipiv,
781 doublecomplex *w, integer *ldw, integer *info)
783 /* System generated locals */
784 integer a_dim1, a_offset, w_dim1, w_offset, i__1, i__2, i__3, i__4, i__5;
785 doublereal d__1, d__2;
786 doublecomplex z__1, z__2, z__3, z__4, z__5;
788 /* Local variables */
790 integer imax, jmax, j, k, p;
792 extern logical lsame_(char *, char *);
793 doublereal dtemp, sfmin;
795 extern /* Subroutine */ int zgemm_(char *, char *, integer *, integer *,
796 integer *, doublecomplex *, doublecomplex *, integer *,
797 doublecomplex *, integer *, doublecomplex *, doublecomplex *,
800 extern /* Subroutine */ int zgemv_(char *, integer *, integer *,
801 doublecomplex *, doublecomplex *, integer *, doublecomplex *,
802 integer *, doublecomplex *, doublecomplex *, integer *);
804 extern /* Subroutine */ int zcopy_(integer *, doublecomplex *, integer *,
805 doublecomplex *, integer *), zswap_(integer *, doublecomplex *,
806 integer *, doublecomplex *, integer *);
807 doublecomplex d11, d21, d22;
808 integer jb, ii, jj, kk;
809 extern doublereal dlamch_(char *);
813 extern /* Subroutine */ int zdscal_(integer *, doublereal *,
814 doublecomplex *, integer *);
816 extern /* Subroutine */ int zlacgv_(integer *, doublecomplex *, integer *)
818 extern integer izamax_(integer *, doublecomplex *, integer *);
823 /* -- LAPACK computational routine (version 3.7.0) -- */
824 /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
825 /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
829 /* ===================================================================== */
832 /* Parameter adjustments */
834 a_offset = 1 + a_dim1 * 1;
839 w_offset = 1 + w_dim1 * 1;
845 /* Initialize ALPHA for use in choosing pivot block size. */
847 alpha = (sqrt(17.) + 1.) / 8.;
849 /* Compute machine safe minimum */
851 sfmin = dlamch_("S");
853 if (lsame_(uplo, "U")) {
855 /* Factorize the trailing columns of A using the upper triangle */
856 /* of A and working backwards, and compute the matrix W = U12*D */
857 /* for use in updating A11 (note that conjg(W) is actually stored) */
858 /* Initialize the first entry of array E, where superdiagonal */
859 /* elements of D are stored */
861 e[1].r = 0., e[1].i = 0.;
863 /* K is the main loop index, decreasing from N in steps of 1 or 2 */
868 /* KW is the column of W which corresponds to column K of A */
874 if (k <= *n - *nb + 1 && *nb < *n || k < 1) {
881 /* Copy column K of A to column KW of W and update it */
885 zcopy_(&i__1, &a[k * a_dim1 + 1], &c__1, &w[kw * w_dim1 + 1], &
888 i__1 = k + kw * w_dim1;
889 i__2 = k + k * a_dim1;
891 w[i__1].r = d__1, w[i__1].i = 0.;
894 z__1.r = -1., z__1.i = 0.;
895 zgemv_("No transpose", &k, &i__1, &z__1, &a[(k + 1) * a_dim1 + 1],
896 lda, &w[k + (kw + 1) * w_dim1], ldw, &c_b1, &w[kw *
898 i__1 = k + kw * w_dim1;
899 i__2 = k + kw * w_dim1;
901 w[i__1].r = d__1, w[i__1].i = 0.;
904 /* Determine rows and columns to be interchanged and whether */
905 /* a 1-by-1 or 2-by-2 pivot block will be used */
907 i__1 = k + kw * w_dim1;
908 absakk = (d__1 = w[i__1].r, abs(d__1));
910 /* IMAX is the row-index of the largest off-diagonal element in */
911 /* column K, and COLMAX is its absolute value. */
912 /* Determine both COLMAX and IMAX. */
916 imax = izamax_(&i__1, &w[kw * w_dim1 + 1], &c__1);
917 i__1 = imax + kw * w_dim1;
918 colmax = (d__1 = w[i__1].r, abs(d__1)) + (d__2 = d_imag(&w[imax +
919 kw * w_dim1]), abs(d__2));
924 if (f2cmax(absakk,colmax) == 0.) {
926 /* Column K is zero or underflow: set INFO and continue */
932 i__1 = k + k * a_dim1;
933 i__2 = k + kw * w_dim1;
935 a[i__1].r = d__1, a[i__1].i = 0.;
938 zcopy_(&i__1, &w[kw * w_dim1 + 1], &c__1, &a[k * a_dim1 + 1],
942 /* Set E( K ) to zero */
946 e[i__1].r = 0., e[i__1].i = 0.;
951 /* ============================================================ */
953 /* BEGIN pivot search */
956 /* Equivalent to testing for ABSAKK.GE.ALPHA*COLMAX */
957 /* (used to handle NaN and Inf) */
958 if (! (absakk < alpha * colmax)) {
960 /* no interchange, use 1-by-1 pivot block */
966 /* Lop until pivot found */
972 /* BEGIN pivot search loop body */
975 /* Copy column IMAX to column KW-1 of W and update it */
979 zcopy_(&i__1, &a[imax * a_dim1 + 1], &c__1, &w[(kw - 1) *
982 i__1 = imax + (kw - 1) * w_dim1;
983 i__2 = imax + imax * a_dim1;
985 w[i__1].r = d__1, w[i__1].i = 0.;
988 zcopy_(&i__1, &a[imax + (imax + 1) * a_dim1], lda, &w[imax +
989 1 + (kw - 1) * w_dim1], &c__1);
991 zlacgv_(&i__1, &w[imax + 1 + (kw - 1) * w_dim1], &c__1);
995 z__1.r = -1., z__1.i = 0.;
996 zgemv_("No transpose", &k, &i__1, &z__1, &a[(k + 1) *
997 a_dim1 + 1], lda, &w[imax + (kw + 1) * w_dim1],
998 ldw, &c_b1, &w[(kw - 1) * w_dim1 + 1], &c__1);
999 i__1 = imax + (kw - 1) * w_dim1;
1000 i__2 = imax + (kw - 1) * w_dim1;
1002 w[i__1].r = d__1, w[i__1].i = 0.;
1005 /* JMAX is the column-index of the largest off-diagonal */
1006 /* element in row IMAX, and ROWMAX is its absolute value. */
1007 /* Determine both ROWMAX and JMAX. */
1011 jmax = imax + izamax_(&i__1, &w[imax + 1 + (kw - 1) *
1013 i__1 = jmax + (kw - 1) * w_dim1;
1014 rowmax = (d__1 = w[i__1].r, abs(d__1)) + (d__2 = d_imag(&
1015 w[jmax + (kw - 1) * w_dim1]), abs(d__2));
1022 itemp = izamax_(&i__1, &w[(kw - 1) * w_dim1 + 1], &c__1);
1023 i__1 = itemp + (kw - 1) * w_dim1;
1024 dtemp = (d__1 = w[i__1].r, abs(d__1)) + (d__2 = d_imag(&w[
1025 itemp + (kw - 1) * w_dim1]), abs(d__2));
1026 if (dtemp > rowmax) {
1033 /* Equivalent to testing for */
1034 /* ABS( REAL( W( IMAX,KW-1 ) ) ).GE.ALPHA*ROWMAX */
1035 /* (used to handle NaN and Inf) */
1037 i__1 = imax + (kw - 1) * w_dim1;
1038 if (! ((d__1 = w[i__1].r, abs(d__1)) < alpha * rowmax)) {
1040 /* interchange rows and columns K and IMAX, */
1041 /* use 1-by-1 pivot block */
1045 /* copy column KW-1 of W to column KW of W */
1047 zcopy_(&k, &w[(kw - 1) * w_dim1 + 1], &c__1, &w[kw *
1048 w_dim1 + 1], &c__1);
1053 /* Equivalent to testing for ROWMAX.EQ.COLMAX, */
1054 /* (used to handle NaN and Inf) */
1056 } else if (p == jmax || rowmax <= colmax) {
1058 /* interchange rows and columns K-1 and IMAX, */
1059 /* use 2-by-2 pivot block */
1068 /* Pivot not found: set params and repeat */
1074 /* Copy updated JMAXth (next IMAXth) column to Kth of W */
1076 zcopy_(&k, &w[(kw - 1) * w_dim1 + 1], &c__1, &w[kw *
1077 w_dim1 + 1], &c__1);
1082 /* END pivot search loop body */
1090 /* END pivot search */
1092 /* ============================================================ */
1094 /* KK is the column of A where pivoting step stopped */
1098 /* KKW is the column of W which corresponds to column KK of A */
1100 kkw = *nb + kk - *n;
1102 /* Interchange rows and columns P and K. */
1103 /* Updated column P is already stored in column KW of W. */
1105 if (kstep == 2 && p != k) {
1107 /* Copy non-updated column K to column P of submatrix A */
1108 /* at step K. No need to copy element into columns */
1109 /* K and K-1 of A for 2-by-2 pivot, since these columns */
1110 /* will be later overwritten. */
1112 i__1 = p + p * a_dim1;
1113 i__2 = k + k * a_dim1;
1115 a[i__1].r = d__1, a[i__1].i = 0.;
1117 zcopy_(&i__1, &a[p + 1 + k * a_dim1], &c__1, &a[p + (p + 1) *
1120 zlacgv_(&i__1, &a[p + (p + 1) * a_dim1], lda);
1123 zcopy_(&i__1, &a[k * a_dim1 + 1], &c__1, &a[p * a_dim1 +
1127 /* Interchange rows K and P in the last K+1 to N columns of A */
1128 /* (columns K and K-1 of A for 2-by-2 pivot will be */
1129 /* later overwritten). Interchange rows K and P */
1130 /* in last KKW to NB columns of W. */
1134 zswap_(&i__1, &a[k + (k + 1) * a_dim1], lda, &a[p + (k +
1138 zswap_(&i__1, &w[k + kkw * w_dim1], ldw, &w[p + kkw * w_dim1],
1142 /* Interchange rows and columns KP and KK. */
1143 /* Updated column KP is already stored in column KKW of W. */
1147 /* Copy non-updated column KK to column KP of submatrix A */
1148 /* at step K. No need to copy element into column K */
1149 /* (or K and K-1 for 2-by-2 pivot) of A, since these columns */
1150 /* will be later overwritten. */
1152 i__1 = kp + kp * a_dim1;
1153 i__2 = kk + kk * a_dim1;
1155 a[i__1].r = d__1, a[i__1].i = 0.;
1157 zcopy_(&i__1, &a[kp + 1 + kk * a_dim1], &c__1, &a[kp + (kp +
1160 zlacgv_(&i__1, &a[kp + (kp + 1) * a_dim1], lda);
1163 zcopy_(&i__1, &a[kk * a_dim1 + 1], &c__1, &a[kp * a_dim1
1167 /* Interchange rows KK and KP in last K+1 to N columns of A */
1168 /* (columns K (or K and K-1 for 2-by-2 pivot) of A will be */
1169 /* later overwritten). Interchange rows KK and KP */
1170 /* in last KKW to NB columns of W. */
1174 zswap_(&i__1, &a[kk + (k + 1) * a_dim1], lda, &a[kp + (k
1175 + 1) * a_dim1], lda);
1178 zswap_(&i__1, &w[kk + kkw * w_dim1], ldw, &w[kp + kkw *
1184 /* 1-by-1 pivot block D(k): column kw of W now holds */
1186 /* W(kw) = U(k)*D(k), */
1188 /* where U(k) is the k-th column of U */
1190 /* (1) Store subdiag. elements of column U(k) */
1191 /* and 1-by-1 block D(k) in column k of A. */
1192 /* (NOTE: Diagonal element U(k,k) is a UNIT element */
1193 /* and not stored) */
1194 /* A(k,k) := D(k,k) = W(k,kw) */
1195 /* A(1:k-1,k) := U(1:k-1,k) = W(1:k-1,kw)/D(k,k) */
1197 /* (NOTE: No need to use for Hermitian matrix */
1198 /* A( K, K ) = REAL( W( K, K) ) to separately copy diagonal */
1199 /* element D(k,k) from W (potentially saves only one load)) */
1200 zcopy_(&k, &w[kw * w_dim1 + 1], &c__1, &a[k * a_dim1 + 1], &
1204 /* (NOTE: No need to check if A(k,k) is NOT ZERO, */
1205 /* since that was ensured earlier in pivot search: */
1206 /* case A(k,k) = 0 falls into 2x2 pivot case(3)) */
1208 /* Handle division by a small number */
1210 i__1 = k + k * a_dim1;
1212 if (abs(t) >= sfmin) {
1215 zdscal_(&i__1, &r1, &a[k * a_dim1 + 1], &c__1);
1218 for (ii = 1; ii <= i__1; ++ii) {
1219 i__2 = ii + k * a_dim1;
1220 i__3 = ii + k * a_dim1;
1221 z__1.r = a[i__3].r / t, z__1.i = a[i__3].i / t;
1222 a[i__2].r = z__1.r, a[i__2].i = z__1.i;
1227 /* (2) Conjugate column W(kw) */
1230 zlacgv_(&i__1, &w[kw * w_dim1 + 1], &c__1);
1232 /* Store the superdiagonal element of D in array E */
1235 e[i__1].r = 0., e[i__1].i = 0.;
1241 /* 2-by-2 pivot block D(k): columns kw and kw-1 of W now hold */
1243 /* ( W(kw-1) W(kw) ) = ( U(k-1) U(k) )*D(k) */
1245 /* where U(k) and U(k-1) are the k-th and (k-1)-th columns */
1248 /* (1) Store U(1:k-2,k-1) and U(1:k-2,k) and 2-by-2 */
1249 /* block D(k-1:k,k-1:k) in columns k-1 and k of A. */
1250 /* (NOTE: 2-by-2 diagonal block U(k-1:k,k-1:k) is a UNIT */
1251 /* block and not stored) */
1252 /* A(k-1:k,k-1:k) := D(k-1:k,k-1:k) = W(k-1:k,kw-1:kw) */
1253 /* A(1:k-2,k-1:k) := U(1:k-2,k:k-1:k) = */
1254 /* = W(1:k-2,kw-1:kw) * ( D(k-1:k,k-1:k)**(-1) ) */
1258 /* Factor out the columns of the inverse of 2-by-2 pivot */
1259 /* block D, so that each column contains 1, to reduce the */
1260 /* number of FLOPS when we multiply panel */
1261 /* ( W(kw-1) W(kw) ) by this inverse, i.e. by D**(-1). */
1263 /* D**(-1) = ( d11 cj(d21) )**(-1) = */
1266 /* = 1/(d11*d22-|d21|**2) * ( ( d22) (-cj(d21) ) ) = */
1267 /* ( (-d21) ( d11 ) ) */
1269 /* = 1/(|d21|**2) * 1/((d11/cj(d21))*(d22/d21)-1) * */
1271 /* * ( d21*( d22/d21 ) conj(d21)*( - 1 ) ) = */
1272 /* ( ( -1 ) ( d11/conj(d21) ) ) */
1274 /* = 1/(|d21|**2) * 1/(D22*D11-1) * */
1276 /* * ( d21*( D11 ) conj(d21)*( -1 ) ) = */
1277 /* ( ( -1 ) ( D22 ) ) */
1279 /* = (1/|d21|**2) * T * ( d21*( D11 ) conj(d21)*( -1 ) ) = */
1280 /* ( ( -1 ) ( D22 ) ) */
1282 /* = ( (T/conj(d21))*( D11 ) (T/d21)*( -1 ) ) = */
1283 /* ( ( -1 ) ( D22 ) ) */
1285 /* Handle division by a small number. (NOTE: order of */
1286 /* operations is important) */
1288 /* = ( T*(( D11 )/conj(D21)) T*(( -1 )/D21 ) ) */
1289 /* ( (( -1 ) ) (( D22 ) ) ), */
1291 /* where D11 = d22/d21, */
1292 /* D22 = d11/conj(d21), */
1294 /* T = 1/(D22*D11-1). */
1296 /* (NOTE: No need to check for division by ZERO, */
1297 /* since that was ensured earlier in pivot search: */
1298 /* (a) d21 != 0 in 2x2 pivot case(4), */
1299 /* since |d21| should be larger than |d11| and |d22|; */
1300 /* (b) (D22*D11 - 1) != 0, since from (a), */
1301 /* both |D11| < 1, |D22| < 1, hence |D22*D11| << 1.) */
1303 i__1 = k - 1 + kw * w_dim1;
1304 d21.r = w[i__1].r, d21.i = w[i__1].i;
1305 d_cnjg(&z__2, &d21);
1306 z_div(&z__1, &w[k + kw * w_dim1], &z__2);
1307 d11.r = z__1.r, d11.i = z__1.i;
1308 z_div(&z__1, &w[k - 1 + (kw - 1) * w_dim1], &d21);
1309 d22.r = z__1.r, d22.i = z__1.i;
1310 z__1.r = d11.r * d22.r - d11.i * d22.i, z__1.i = d11.r *
1311 d22.i + d11.i * d22.r;
1312 t = 1. / (z__1.r - 1.);
1314 /* Update elements in columns A(k-1) and A(k) as */
1315 /* dot products of rows of ( W(kw-1) W(kw) ) and columns */
1319 for (j = 1; j <= i__1; ++j) {
1320 i__2 = j + (k - 1) * a_dim1;
1321 i__3 = j + (kw - 1) * w_dim1;
1322 z__4.r = d11.r * w[i__3].r - d11.i * w[i__3].i,
1323 z__4.i = d11.r * w[i__3].i + d11.i * w[i__3]
1325 i__4 = j + kw * w_dim1;
1326 z__3.r = z__4.r - w[i__4].r, z__3.i = z__4.i - w[i__4]
1328 z_div(&z__2, &z__3, &d21);
1329 z__1.r = t * z__2.r, z__1.i = t * z__2.i;
1330 a[i__2].r = z__1.r, a[i__2].i = z__1.i;
1331 i__2 = j + k * a_dim1;
1332 i__3 = j + kw * w_dim1;
1333 z__4.r = d22.r * w[i__3].r - d22.i * w[i__3].i,
1334 z__4.i = d22.r * w[i__3].i + d22.i * w[i__3]
1336 i__4 = j + (kw - 1) * w_dim1;
1337 z__3.r = z__4.r - w[i__4].r, z__3.i = z__4.i - w[i__4]
1339 d_cnjg(&z__5, &d21);
1340 z_div(&z__2, &z__3, &z__5);
1341 z__1.r = t * z__2.r, z__1.i = t * z__2.i;
1342 a[i__2].r = z__1.r, a[i__2].i = z__1.i;
1347 /* Copy diagonal elements of D(K) to A, */
1348 /* copy superdiagonal element of D(K) to E(K) and */
1349 /* ZERO out superdiagonal entry of A */
1351 i__1 = k - 1 + (k - 1) * a_dim1;
1352 i__2 = k - 1 + (kw - 1) * w_dim1;
1353 a[i__1].r = w[i__2].r, a[i__1].i = w[i__2].i;
1354 i__1 = k - 1 + k * a_dim1;
1355 a[i__1].r = 0., a[i__1].i = 0.;
1356 i__1 = k + k * a_dim1;
1357 i__2 = k + kw * w_dim1;
1358 a[i__1].r = w[i__2].r, a[i__1].i = w[i__2].i;
1360 i__2 = k - 1 + kw * w_dim1;
1361 e[i__1].r = w[i__2].r, e[i__1].i = w[i__2].i;
1363 e[i__1].r = 0., e[i__1].i = 0.;
1365 /* (2) Conjugate columns W(kw) and W(kw-1) */
1368 zlacgv_(&i__1, &w[kw * w_dim1 + 1], &c__1);
1370 zlacgv_(&i__1, &w[(kw - 1) * w_dim1 + 1], &c__1);
1374 /* End column K is nonsingular */
1378 /* Store details of the interchanges in IPIV */
1387 /* Decrease K and return to the start of the main loop */
1394 /* Update the upper triangle of A11 (= A(1:k,1:k)) as */
1396 /* A11 := A11 - U12*D*U12**H = A11 - U12*W**H */
1398 /* computing blocks of NB columns at a time (note that conjg(W) is */
1399 /* actually stored) */
1402 for (j = (k - 1) / *nb * *nb + 1; i__1 < 0 ? j >= 1 : j <= 1; j +=
1405 i__2 = *nb, i__3 = k - j + 1;
1406 jb = f2cmin(i__2,i__3);
1408 /* Update the upper triangle of the diagonal block */
1411 for (jj = j; jj <= i__2; ++jj) {
1412 i__3 = jj + jj * a_dim1;
1413 i__4 = jj + jj * a_dim1;
1415 a[i__3].r = d__1, a[i__3].i = 0.;
1418 z__1.r = -1., z__1.i = 0.;
1419 zgemv_("No transpose", &i__3, &i__4, &z__1, &a[j + (k + 1) *
1420 a_dim1], lda, &w[jj + (kw + 1) * w_dim1], ldw, &c_b1,
1421 &a[j + jj * a_dim1], &c__1);
1422 i__3 = jj + jj * a_dim1;
1423 i__4 = jj + jj * a_dim1;
1425 a[i__3].r = d__1, a[i__3].i = 0.;
1429 /* Update the rectangular superdiagonal block */
1434 z__1.r = -1., z__1.i = 0.;
1435 zgemm_("No transpose", "Transpose", &i__2, &jb, &i__3, &z__1,
1436 &a[(k + 1) * a_dim1 + 1], lda, &w[j + (kw + 1) *
1437 w_dim1], ldw, &c_b1, &a[j * a_dim1 + 1], lda);
1442 /* Set KB to the number of columns factorized */
1448 /* Factorize the leading columns of A using the lower triangle */
1449 /* of A and working forwards, and compute the matrix W = L21*D */
1450 /* for use in updating A22 (note that conjg(W) is actually stored) */
1452 /* Initialize the unused last entry of the subdiagonal array E. */
1455 e[i__1].r = 0., e[i__1].i = 0.;
1457 /* K is the main loop index, increasing from 1 in steps of 1 or 2 */
1462 /* Exit from loop */
1464 if (k >= *nb && *nb < *n || k > *n) {
1471 /* Copy column K of A to column K of W and update column K of W */
1473 i__1 = k + k * w_dim1;
1474 i__2 = k + k * a_dim1;
1476 w[i__1].r = d__1, w[i__1].i = 0.;
1479 zcopy_(&i__1, &a[k + 1 + k * a_dim1], &c__1, &w[k + 1 + k *
1485 z__1.r = -1., z__1.i = 0.;
1486 zgemv_("No transpose", &i__1, &i__2, &z__1, &a[k + a_dim1], lda, &
1487 w[k + w_dim1], ldw, &c_b1, &w[k + k * w_dim1], &c__1);
1488 i__1 = k + k * w_dim1;
1489 i__2 = k + k * w_dim1;
1491 w[i__1].r = d__1, w[i__1].i = 0.;
1494 /* Determine rows and columns to be interchanged and whether */
1495 /* a 1-by-1 or 2-by-2 pivot block will be used */
1497 i__1 = k + k * w_dim1;
1498 absakk = (d__1 = w[i__1].r, abs(d__1));
1500 /* IMAX is the row-index of the largest off-diagonal element in */
1501 /* column K, and COLMAX is its absolute value. */
1502 /* Determine both COLMAX and IMAX. */
1506 imax = k + izamax_(&i__1, &w[k + 1 + k * w_dim1], &c__1);
1507 i__1 = imax + k * w_dim1;
1508 colmax = (d__1 = w[i__1].r, abs(d__1)) + (d__2 = d_imag(&w[imax +
1509 k * w_dim1]), abs(d__2));
1514 if (f2cmax(absakk,colmax) == 0.) {
1516 /* Column K is zero or underflow: set INFO and continue */
1522 i__1 = k + k * a_dim1;
1523 i__2 = k + k * w_dim1;
1525 a[i__1].r = d__1, a[i__1].i = 0.;
1528 zcopy_(&i__1, &w[k + 1 + k * w_dim1], &c__1, &a[k + 1 + k *
1532 /* Set E( K ) to zero */
1536 e[i__1].r = 0., e[i__1].i = 0.;
1541 /* ============================================================ */
1543 /* BEGIN pivot search */
1546 /* Equivalent to testing for ABSAKK.GE.ALPHA*COLMAX */
1547 /* (used to handle NaN and Inf) */
1549 if (! (absakk < alpha * colmax)) {
1551 /* no interchange, use 1-by-1 pivot block */
1559 /* Loop until pivot found */
1563 /* BEGIN pivot search loop body */
1566 /* Copy column IMAX to column k+1 of W and update it */
1569 zcopy_(&i__1, &a[imax + k * a_dim1], lda, &w[k + (k + 1) *
1572 zlacgv_(&i__1, &w[k + (k + 1) * w_dim1], &c__1);
1573 i__1 = imax + (k + 1) * w_dim1;
1574 i__2 = imax + imax * a_dim1;
1576 w[i__1].r = d__1, w[i__1].i = 0.;
1580 zcopy_(&i__1, &a[imax + 1 + imax * a_dim1], &c__1, &w[
1581 imax + 1 + (k + 1) * w_dim1], &c__1);
1587 z__1.r = -1., z__1.i = 0.;
1588 zgemv_("No transpose", &i__1, &i__2, &z__1, &a[k + a_dim1]
1589 , lda, &w[imax + w_dim1], ldw, &c_b1, &w[k + (k +
1590 1) * w_dim1], &c__1);
1591 i__1 = imax + (k + 1) * w_dim1;
1592 i__2 = imax + (k + 1) * w_dim1;
1594 w[i__1].r = d__1, w[i__1].i = 0.;
1597 /* JMAX is the column-index of the largest off-diagonal */
1598 /* element in row IMAX, and ROWMAX is its absolute value. */
1599 /* Determine both ROWMAX and JMAX. */
1603 jmax = k - 1 + izamax_(&i__1, &w[k + (k + 1) * w_dim1], &
1605 i__1 = jmax + (k + 1) * w_dim1;
1606 rowmax = (d__1 = w[i__1].r, abs(d__1)) + (d__2 = d_imag(&
1607 w[jmax + (k + 1) * w_dim1]), abs(d__2));
1614 itemp = imax + izamax_(&i__1, &w[imax + 1 + (k + 1) *
1616 i__1 = itemp + (k + 1) * w_dim1;
1617 dtemp = (d__1 = w[i__1].r, abs(d__1)) + (d__2 = d_imag(&w[
1618 itemp + (k + 1) * w_dim1]), abs(d__2));
1619 if (dtemp > rowmax) {
1626 /* Equivalent to testing for */
1627 /* ABS( REAL( W( IMAX,K+1 ) ) ).GE.ALPHA*ROWMAX */
1628 /* (used to handle NaN and Inf) */
1630 i__1 = imax + (k + 1) * w_dim1;
1631 if (! ((d__1 = w[i__1].r, abs(d__1)) < alpha * rowmax)) {
1633 /* interchange rows and columns K and IMAX, */
1634 /* use 1-by-1 pivot block */
1638 /* copy column K+1 of W to column K of W */
1641 zcopy_(&i__1, &w[k + (k + 1) * w_dim1], &c__1, &w[k + k *
1647 /* Equivalent to testing for ROWMAX.EQ.COLMAX, */
1648 /* (used to handle NaN and Inf) */
1650 } else if (p == jmax || rowmax <= colmax) {
1652 /* interchange rows and columns K+1 and IMAX, */
1653 /* use 2-by-2 pivot block */
1662 /* Pivot not found: set params and repeat */
1668 /* Copy updated JMAXth (next IMAXth) column to Kth of W */
1671 zcopy_(&i__1, &w[k + (k + 1) * w_dim1], &c__1, &w[k + k *
1677 /* End pivot search loop body */
1685 /* END pivot search */
1687 /* ============================================================ */
1689 /* KK is the column of A where pivoting step stopped */
1693 /* Interchange rows and columns P and K (only for 2-by-2 pivot). */
1694 /* Updated column P is already stored in column K of W. */
1696 if (kstep == 2 && p != k) {
1698 /* Copy non-updated column KK-1 to column P of submatrix A */
1699 /* at step K. No need to copy element into columns */
1700 /* K and K+1 of A for 2-by-2 pivot, since these columns */
1701 /* will be later overwritten. */
1703 i__1 = p + p * a_dim1;
1704 i__2 = k + k * a_dim1;
1706 a[i__1].r = d__1, a[i__1].i = 0.;
1708 zcopy_(&i__1, &a[k + 1 + k * a_dim1], &c__1, &a[p + (k + 1) *
1711 zlacgv_(&i__1, &a[p + (k + 1) * a_dim1], lda);
1714 zcopy_(&i__1, &a[p + 1 + k * a_dim1], &c__1, &a[p + 1 + p
1718 /* Interchange rows K and P in first K-1 columns of A */
1719 /* (columns K and K+1 of A for 2-by-2 pivot will be */
1720 /* later overwritten). Interchange rows K and P */
1721 /* in first KK columns of W. */
1725 zswap_(&i__1, &a[k + a_dim1], lda, &a[p + a_dim1], lda);
1727 zswap_(&kk, &w[k + w_dim1], ldw, &w[p + w_dim1], ldw);
1730 /* Interchange rows and columns KP and KK. */
1731 /* Updated column KP is already stored in column KK of W. */
1735 /* Copy non-updated column KK to column KP of submatrix A */
1736 /* at step K. No need to copy element into column K */
1737 /* (or K and K+1 for 2-by-2 pivot) of A, since these columns */
1738 /* will be later overwritten. */
1740 i__1 = kp + kp * a_dim1;
1741 i__2 = kk + kk * a_dim1;
1743 a[i__1].r = d__1, a[i__1].i = 0.;
1745 zcopy_(&i__1, &a[kk + 1 + kk * a_dim1], &c__1, &a[kp + (kk +
1748 zlacgv_(&i__1, &a[kp + (kk + 1) * a_dim1], lda);
1751 zcopy_(&i__1, &a[kp + 1 + kk * a_dim1], &c__1, &a[kp + 1
1752 + kp * a_dim1], &c__1);
1755 /* Interchange rows KK and KP in first K-1 columns of A */
1756 /* (column K (or K and K+1 for 2-by-2 pivot) of A will be */
1757 /* later overwritten). Interchange rows KK and KP */
1758 /* in first KK columns of W. */
1762 zswap_(&i__1, &a[kk + a_dim1], lda, &a[kp + a_dim1], lda);
1764 zswap_(&kk, &w[kk + w_dim1], ldw, &w[kp + w_dim1], ldw);
1769 /* 1-by-1 pivot block D(k): column k of W now holds */
1771 /* W(k) = L(k)*D(k), */
1773 /* where L(k) is the k-th column of L */
1775 /* (1) Store subdiag. elements of column L(k) */
1776 /* and 1-by-1 block D(k) in column k of A. */
1777 /* (NOTE: Diagonal element L(k,k) is a UNIT element */
1778 /* and not stored) */
1779 /* A(k,k) := D(k,k) = W(k,k) */
1780 /* A(k+1:N,k) := L(k+1:N,k) = W(k+1:N,k)/D(k,k) */
1782 /* (NOTE: No need to use for Hermitian matrix */
1783 /* A( K, K ) = REAL( W( K, K) ) to separately copy diagonal */
1784 /* element D(k,k) from W (potentially saves only one load)) */
1786 zcopy_(&i__1, &w[k + k * w_dim1], &c__1, &a[k + k * a_dim1], &
1790 /* (NOTE: No need to check if A(k,k) is NOT ZERO, */
1791 /* since that was ensured earlier in pivot search: */
1792 /* case A(k,k) = 0 falls into 2x2 pivot case(3)) */
1794 /* Handle division by a small number */
1796 i__1 = k + k * a_dim1;
1798 if (abs(t) >= sfmin) {
1801 zdscal_(&i__1, &r1, &a[k + 1 + k * a_dim1], &c__1);
1804 for (ii = k + 1; ii <= i__1; ++ii) {
1805 i__2 = ii + k * a_dim1;
1806 i__3 = ii + k * a_dim1;
1807 z__1.r = a[i__3].r / t, z__1.i = a[i__3].i / t;
1808 a[i__2].r = z__1.r, a[i__2].i = z__1.i;
1813 /* (2) Conjugate column W(k) */
1816 zlacgv_(&i__1, &w[k + 1 + k * w_dim1], &c__1);
1818 /* Store the subdiagonal element of D in array E */
1821 e[i__1].r = 0., e[i__1].i = 0.;
1827 /* 2-by-2 pivot block D(k): columns k and k+1 of W now hold */
1829 /* ( W(k) W(k+1) ) = ( L(k) L(k+1) )*D(k) */
1831 /* where L(k) and L(k+1) are the k-th and (k+1)-th columns */
1834 /* (1) Store L(k+2:N,k) and L(k+2:N,k+1) and 2-by-2 */
1835 /* block D(k:k+1,k:k+1) in columns k and k+1 of A. */
1836 /* NOTE: 2-by-2 diagonal block L(k:k+1,k:k+1) is a UNIT */
1837 /* block and not stored. */
1838 /* A(k:k+1,k:k+1) := D(k:k+1,k:k+1) = W(k:k+1,k:k+1) */
1839 /* A(k+2:N,k:k+1) := L(k+2:N,k:k+1) = */
1840 /* = W(k+2:N,k:k+1) * ( D(k:k+1,k:k+1)**(-1) ) */
1844 /* Factor out the columns of the inverse of 2-by-2 pivot */
1845 /* block D, so that each column contains 1, to reduce the */
1846 /* number of FLOPS when we multiply panel */
1847 /* ( W(kw-1) W(kw) ) by this inverse, i.e. by D**(-1). */
1849 /* D**(-1) = ( d11 cj(d21) )**(-1) = */
1852 /* = 1/(d11*d22-|d21|**2) * ( ( d22) (-cj(d21) ) ) = */
1853 /* ( (-d21) ( d11 ) ) */
1855 /* = 1/(|d21|**2) * 1/((d11/cj(d21))*(d22/d21)-1) * */
1857 /* * ( d21*( d22/d21 ) conj(d21)*( - 1 ) ) = */
1858 /* ( ( -1 ) ( d11/conj(d21) ) ) */
1860 /* = 1/(|d21|**2) * 1/(D22*D11-1) * */
1862 /* * ( d21*( D11 ) conj(d21)*( -1 ) ) = */
1863 /* ( ( -1 ) ( D22 ) ) */
1865 /* = (1/|d21|**2) * T * ( d21*( D11 ) conj(d21)*( -1 ) ) = */
1866 /* ( ( -1 ) ( D22 ) ) */
1868 /* = ( (T/conj(d21))*( D11 ) (T/d21)*( -1 ) ) = */
1869 /* ( ( -1 ) ( D22 ) ) */
1871 /* Handle division by a small number. (NOTE: order of */
1872 /* operations is important) */
1874 /* = ( T*(( D11 )/conj(D21)) T*(( -1 )/D21 ) ) */
1875 /* ( (( -1 ) ) (( D22 ) ) ), */
1877 /* where D11 = d22/d21, */
1878 /* D22 = d11/conj(d21), */
1880 /* T = 1/(D22*D11-1). */
1882 /* (NOTE: No need to check for division by ZERO, */
1883 /* since that was ensured earlier in pivot search: */
1884 /* (a) d21 != 0 in 2x2 pivot case(4), */
1885 /* since |d21| should be larger than |d11| and |d22|; */
1886 /* (b) (D22*D11 - 1) != 0, since from (a), */
1887 /* both |D11| < 1, |D22| < 1, hence |D22*D11| << 1.) */
1889 i__1 = k + 1 + k * w_dim1;
1890 d21.r = w[i__1].r, d21.i = w[i__1].i;
1891 z_div(&z__1, &w[k + 1 + (k + 1) * w_dim1], &d21);
1892 d11.r = z__1.r, d11.i = z__1.i;
1893 d_cnjg(&z__2, &d21);
1894 z_div(&z__1, &w[k + k * w_dim1], &z__2);
1895 d22.r = z__1.r, d22.i = z__1.i;
1896 z__1.r = d11.r * d22.r - d11.i * d22.i, z__1.i = d11.r *
1897 d22.i + d11.i * d22.r;
1898 t = 1. / (z__1.r - 1.);
1900 /* Update elements in columns A(k) and A(k+1) as */
1901 /* dot products of rows of ( W(k) W(k+1) ) and columns */
1905 for (j = k + 2; j <= i__1; ++j) {
1906 i__2 = j + k * a_dim1;
1907 i__3 = j + k * w_dim1;
1908 z__4.r = d11.r * w[i__3].r - d11.i * w[i__3].i,
1909 z__4.i = d11.r * w[i__3].i + d11.i * w[i__3]
1911 i__4 = j + (k + 1) * w_dim1;
1912 z__3.r = z__4.r - w[i__4].r, z__3.i = z__4.i - w[i__4]
1914 d_cnjg(&z__5, &d21);
1915 z_div(&z__2, &z__3, &z__5);
1916 z__1.r = t * z__2.r, z__1.i = t * z__2.i;
1917 a[i__2].r = z__1.r, a[i__2].i = z__1.i;
1918 i__2 = j + (k + 1) * a_dim1;
1919 i__3 = j + (k + 1) * w_dim1;
1920 z__4.r = d22.r * w[i__3].r - d22.i * w[i__3].i,
1921 z__4.i = d22.r * w[i__3].i + d22.i * w[i__3]
1923 i__4 = j + k * w_dim1;
1924 z__3.r = z__4.r - w[i__4].r, z__3.i = z__4.i - w[i__4]
1926 z_div(&z__2, &z__3, &d21);
1927 z__1.r = t * z__2.r, z__1.i = t * z__2.i;
1928 a[i__2].r = z__1.r, a[i__2].i = z__1.i;
1933 /* Copy diagonal elements of D(K) to A, */
1934 /* copy subdiagonal element of D(K) to E(K) and */
1935 /* ZERO out subdiagonal entry of A */
1937 i__1 = k + k * a_dim1;
1938 i__2 = k + k * w_dim1;
1939 a[i__1].r = w[i__2].r, a[i__1].i = w[i__2].i;
1940 i__1 = k + 1 + k * a_dim1;
1941 a[i__1].r = 0., a[i__1].i = 0.;
1942 i__1 = k + 1 + (k + 1) * a_dim1;
1943 i__2 = k + 1 + (k + 1) * w_dim1;
1944 a[i__1].r = w[i__2].r, a[i__1].i = w[i__2].i;
1946 i__2 = k + 1 + k * w_dim1;
1947 e[i__1].r = w[i__2].r, e[i__1].i = w[i__2].i;
1949 e[i__1].r = 0., e[i__1].i = 0.;
1951 /* (2) Conjugate columns W(k) and W(k+1) */
1954 zlacgv_(&i__1, &w[k + 1 + k * w_dim1], &c__1);
1956 zlacgv_(&i__1, &w[k + 2 + (k + 1) * w_dim1], &c__1);
1960 /* End column K is nonsingular */
1964 /* Store details of the interchanges in IPIV */
1973 /* Increase K and return to the start of the main loop */
1980 /* Update the lower triangle of A22 (= A(k:n,k:n)) as */
1982 /* A22 := A22 - L21*D*L21**H = A22 - L21*W**H */
1984 /* computing blocks of NB columns at a time (note that conjg(W) is */
1985 /* actually stored) */
1989 for (j = k; i__2 < 0 ? j >= i__1 : j <= i__1; j += i__2) {
1991 i__3 = *nb, i__4 = *n - j + 1;
1992 jb = f2cmin(i__3,i__4);
1994 /* Update the lower triangle of the diagonal block */
1997 for (jj = j; jj <= i__3; ++jj) {
1998 i__4 = jj + jj * a_dim1;
1999 i__5 = jj + jj * a_dim1;
2001 a[i__4].r = d__1, a[i__4].i = 0.;
2004 z__1.r = -1., z__1.i = 0.;
2005 zgemv_("No transpose", &i__4, &i__5, &z__1, &a[jj + a_dim1],
2006 lda, &w[jj + w_dim1], ldw, &c_b1, &a[jj + jj * a_dim1]
2008 i__4 = jj + jj * a_dim1;
2009 i__5 = jj + jj * a_dim1;
2011 a[i__4].r = d__1, a[i__4].i = 0.;
2015 /* Update the rectangular subdiagonal block */
2018 i__3 = *n - j - jb + 1;
2020 z__1.r = -1., z__1.i = 0.;
2021 zgemm_("No transpose", "Transpose", &i__3, &jb, &i__4, &z__1,
2022 &a[j + jb + a_dim1], lda, &w[j + w_dim1], ldw, &c_b1,
2023 &a[j + jb + j * a_dim1], lda);
2028 /* Set KB to the number of columns factorized */
2035 /* End of ZLAHEF_RK */