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__13 = 13;
516 static integer c__15 = 15;
517 static integer c_n1 = -1;
518 static integer c__12 = 12;
519 static integer c__14 = 14;
520 static integer c__16 = 16;
521 static logical c_false = FALSE_;
522 static integer c__1 = 1;
523 static integer c__3 = 3;
525 /* > \brief \b DLAQR4 computes the eigenvalues of a Hessenberg matrix, and optionally the matrices from the Sc
526 hur decomposition. */
528 /* =========== DOCUMENTATION =========== */
530 /* Online html documentation available at */
531 /* http://www.netlib.org/lapack/explore-html/ */
534 /* > Download DLAQR4 + dependencies */
535 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/dlaqr4.
538 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/dlaqr4.
541 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/dlaqr4.
549 /* SUBROUTINE DLAQR4( WANTT, WANTZ, N, ILO, IHI, H, LDH, WR, WI, */
550 /* ILOZ, IHIZ, Z, LDZ, WORK, LWORK, INFO ) */
552 /* INTEGER IHI, IHIZ, ILO, ILOZ, INFO, LDH, LDZ, LWORK, N */
553 /* LOGICAL WANTT, WANTZ */
554 /* DOUBLE PRECISION H( LDH, * ), WI( * ), WORK( * ), WR( * ), */
558 /* > \par Purpose: */
563 /* > DLAQR4 implements one level of recursion for DLAQR0. */
564 /* > It is a complete implementation of the small bulge multi-shift */
565 /* > QR algorithm. It may be called by DLAQR0 and, for large enough */
566 /* > deflation window size, it may be called by DLAQR3. This */
567 /* > subroutine is identical to DLAQR0 except that it calls DLAQR2 */
568 /* > instead of DLAQR3. */
570 /* > DLAQR4 computes the eigenvalues of a Hessenberg matrix H */
571 /* > and, optionally, the matrices T and Z from the Schur decomposition */
572 /* > H = Z T Z**T, where T is an upper quasi-triangular matrix (the */
573 /* > Schur form), and Z is the orthogonal matrix of Schur vectors. */
575 /* > Optionally Z may be postmultiplied into an input orthogonal */
576 /* > matrix Q so that this routine can give the Schur factorization */
577 /* > of a matrix A which has been reduced to the Hessenberg form H */
578 /* > by the orthogonal matrix Q: A = Q*H*Q**T = (QZ)*T*(QZ)**T. */
584 /* > \param[in] WANTT */
586 /* > WANTT is LOGICAL */
587 /* > = .TRUE. : the full Schur form T is required; */
588 /* > = .FALSE.: only eigenvalues are required. */
591 /* > \param[in] WANTZ */
593 /* > WANTZ is LOGICAL */
594 /* > = .TRUE. : the matrix of Schur vectors Z is required; */
595 /* > = .FALSE.: Schur vectors are not required. */
601 /* > The order of the matrix H. N >= 0. */
604 /* > \param[in] ILO */
606 /* > ILO is INTEGER */
609 /* > \param[in] IHI */
611 /* > IHI is INTEGER */
612 /* > It is assumed that H is already upper triangular in rows */
613 /* > and columns 1:ILO-1 and IHI+1:N and, if ILO > 1, */
614 /* > H(ILO,ILO-1) is zero. ILO and IHI are normally set by a */
615 /* > previous call to DGEBAL, and then passed to DGEHRD when the */
616 /* > matrix output by DGEBAL is reduced to Hessenberg form. */
617 /* > Otherwise, ILO and IHI should be set to 1 and N, */
618 /* > respectively. If N > 0, then 1 <= ILO <= IHI <= N. */
619 /* > If N = 0, then ILO = 1 and IHI = 0. */
622 /* > \param[in,out] H */
624 /* > H is DOUBLE PRECISION array, dimension (LDH,N) */
625 /* > On entry, the upper Hessenberg matrix H. */
626 /* > On exit, if INFO = 0 and WANTT is .TRUE., then H contains */
627 /* > the upper quasi-triangular matrix T from the Schur */
628 /* > decomposition (the Schur form); 2-by-2 diagonal blocks */
629 /* > (corresponding to complex conjugate pairs of eigenvalues) */
630 /* > are returned in standard form, with H(i,i) = H(i+1,i+1) */
631 /* > and H(i+1,i)*H(i,i+1) < 0. If INFO = 0 and WANTT is */
632 /* > .FALSE., then the contents of H are unspecified on exit. */
633 /* > (The output value of H when INFO > 0 is given under the */
634 /* > description of INFO below.) */
636 /* > This subroutine may explicitly set H(i,j) = 0 for i > j and */
637 /* > j = 1, 2, ... ILO-1 or j = IHI+1, IHI+2, ... N. */
640 /* > \param[in] LDH */
642 /* > LDH is INTEGER */
643 /* > The leading dimension of the array H. LDH >= f2cmax(1,N). */
646 /* > \param[out] WR */
648 /* > WR is DOUBLE PRECISION array, dimension (IHI) */
651 /* > \param[out] WI */
653 /* > WI is DOUBLE PRECISION array, dimension (IHI) */
654 /* > The real and imaginary parts, respectively, of the computed */
655 /* > eigenvalues of H(ILO:IHI,ILO:IHI) are stored in WR(ILO:IHI) */
656 /* > and WI(ILO:IHI). If two eigenvalues are computed as a */
657 /* > complex conjugate pair, they are stored in consecutive */
658 /* > elements of WR and WI, say the i-th and (i+1)th, with */
659 /* > WI(i) > 0 and WI(i+1) < 0. If WANTT is .TRUE., then */
660 /* > the eigenvalues are stored in the same order as on the */
661 /* > diagonal of the Schur form returned in H, with */
662 /* > WR(i) = H(i,i) and, if H(i:i+1,i:i+1) is a 2-by-2 diagonal */
663 /* > block, WI(i) = sqrt(-H(i+1,i)*H(i,i+1)) and */
664 /* > WI(i+1) = -WI(i). */
667 /* > \param[in] ILOZ */
669 /* > ILOZ is INTEGER */
672 /* > \param[in] IHIZ */
674 /* > IHIZ is INTEGER */
675 /* > Specify the rows of Z to which transformations must be */
676 /* > applied if WANTZ is .TRUE.. */
677 /* > 1 <= ILOZ <= ILO; IHI <= IHIZ <= N. */
680 /* > \param[in,out] Z */
682 /* > Z is DOUBLE PRECISION array, dimension (LDZ,IHI) */
683 /* > If WANTZ is .FALSE., then Z is not referenced. */
684 /* > If WANTZ is .TRUE., then Z(ILO:IHI,ILOZ:IHIZ) is */
685 /* > replaced by Z(ILO:IHI,ILOZ:IHIZ)*U where U is the */
686 /* > orthogonal Schur factor of H(ILO:IHI,ILO:IHI). */
687 /* > (The output value of Z when INFO > 0 is given under */
688 /* > the description of INFO below.) */
691 /* > \param[in] LDZ */
693 /* > LDZ is INTEGER */
694 /* > The leading dimension of the array Z. if WANTZ is .TRUE. */
695 /* > then LDZ >= MAX(1,IHIZ). Otherwise, LDZ >= 1. */
698 /* > \param[out] WORK */
700 /* > WORK is DOUBLE PRECISION array, dimension LWORK */
701 /* > On exit, if LWORK = -1, WORK(1) returns an estimate of */
702 /* > the optimal value for LWORK. */
705 /* > \param[in] LWORK */
707 /* > LWORK is INTEGER */
708 /* > The dimension of the array WORK. LWORK >= f2cmax(1,N) */
709 /* > is sufficient, but LWORK typically as large as 6*N may */
710 /* > be required for optimal performance. A workspace query */
711 /* > to determine the optimal workspace size is recommended. */
713 /* > If LWORK = -1, then DLAQR4 does a workspace query. */
714 /* > In this case, DLAQR4 checks the input parameters and */
715 /* > estimates the optimal workspace size for the given */
716 /* > values of N, ILO and IHI. The estimate is returned */
717 /* > in WORK(1). No error message related to LWORK is */
718 /* > issued by XERBLA. Neither H nor Z are accessed. */
721 /* > \param[out] INFO */
723 /* > INFO is INTEGER */
724 /* > = 0: successful exit */
725 /* > > 0: if INFO = i, DLAQR4 failed to compute all of */
726 /* > the eigenvalues. Elements 1:ilo-1 and i+1:n of WR */
727 /* > and WI contain those eigenvalues which have been */
728 /* > successfully computed. (Failures are rare.) */
730 /* > If INFO > 0 and WANT is .FALSE., then on exit, */
731 /* > the remaining unconverged eigenvalues are the eigen- */
732 /* > values of the upper Hessenberg matrix rows and */
733 /* > columns ILO through INFO of the final, output */
736 /* > If INFO > 0 and WANTT is .TRUE., then on exit */
738 /* > (*) (initial value of H)*U = U*(final value of H) */
740 /* > where U is a orthogonal matrix. The final */
741 /* > value of H is upper Hessenberg and triangular in */
742 /* > rows and columns INFO+1 through IHI. */
744 /* > If INFO > 0 and WANTZ is .TRUE., then on exit */
746 /* > (final value of Z(ILO:IHI,ILOZ:IHIZ) */
747 /* > = (initial value of Z(ILO:IHI,ILOZ:IHIZ)*U */
749 /* > where U is the orthogonal matrix in (*) (regard- */
750 /* > less of the value of WANTT.) */
752 /* > If INFO > 0 and WANTZ is .FALSE., then Z is not */
759 /* > \author Univ. of Tennessee */
760 /* > \author Univ. of California Berkeley */
761 /* > \author Univ. of Colorado Denver */
762 /* > \author NAG Ltd. */
764 /* > \date December 2016 */
766 /* > \ingroup doubleOTHERauxiliary */
768 /* > \par Contributors: */
769 /* ================== */
771 /* > Karen Braman and Ralph Byers, Department of Mathematics, */
772 /* > University of Kansas, USA */
774 /* > \par References: */
775 /* ================ */
777 /* > K. Braman, R. Byers and R. Mathias, The Multi-Shift QR */
778 /* > Algorithm Part I: Maintaining Well Focused Shifts, and Level 3 */
779 /* > Performance, SIAM Journal of Matrix Analysis, volume 23, pages */
780 /* > 929--947, 2002. */
782 /* > K. Braman, R. Byers and R. Mathias, The Multi-Shift QR */
783 /* > Algorithm Part II: Aggressive Early Deflation, SIAM Journal */
784 /* > of Matrix Analysis, volume 23, pages 948--973, 2002. */
786 /* ===================================================================== */
787 /* Subroutine */ int dlaqr4_(logical *wantt, logical *wantz, integer *n,
788 integer *ilo, integer *ihi, doublereal *h__, integer *ldh, doublereal
789 *wr, doublereal *wi, integer *iloz, integer *ihiz, doublereal *z__,
790 integer *ldz, doublereal *work, integer *lwork, integer *info)
792 /* System generated locals */
793 integer h_dim1, h_offset, z_dim1, z_offset, i__1, i__2, i__3, i__4, i__5;
794 doublereal d__1, d__2, d__3, d__4;
796 /* Local variables */
797 integer ndec, ndfl, kbot, nmin;
800 doublereal zdum[1] /* was [1][1] */;
801 integer kacc22, i__, k, itmax, nsmax, nwmax, kwtop;
802 extern /* Subroutine */ int dlaqr2_(logical *, logical *, integer *,
803 integer *, integer *, integer *, doublereal *, integer *, integer
804 *, integer *, doublereal *, integer *, integer *, integer *,
805 doublereal *, doublereal *, doublereal *, integer *, integer *,
806 doublereal *, integer *, integer *, doublereal *, integer *,
807 doublereal *, integer *), dlanv2_(doublereal *, doublereal *,
808 doublereal *, doublereal *, doublereal *, doublereal *,
809 doublereal *, doublereal *, doublereal *, doublereal *), dlaqr5_(
810 logical *, logical *, integer *, integer *, integer *, integer *,
811 integer *, doublereal *, doublereal *, doublereal *, integer *,
812 integer *, integer *, doublereal *, integer *, doublereal *,
813 integer *, doublereal *, integer *, integer *, doublereal *,
814 integer *, integer *, doublereal *, integer *);
815 doublereal aa, bb, cc, dd;
818 integer nh, nibble, it, ks, kt;
820 integer ku, kv, ls, ns;
823 extern /* Subroutine */ int dlahqr_(logical *, logical *, integer *,
824 integer *, integer *, doublereal *, integer *, doublereal *,
825 doublereal *, integer *, integer *, doublereal *, integer *,
826 integer *), dlacpy_(char *, integer *, integer *, doublereal *,
827 integer *, doublereal *, integer *);
828 extern integer ilaenv_(integer *, char *, char *, integer *, integer *,
829 integer *, integer *, ftnlen, ftnlen);
833 integer lwkopt, inf, kdu, nho, nve, kwh, nsr, nwr, kwv;
836 /* -- LAPACK auxiliary routine (version 3.7.0) -- */
837 /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
838 /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
842 /* ================================================================ */
844 /* ==== Matrices of order NTINY or smaller must be processed by */
845 /* . DLAHQR because of insufficient subdiagonal scratch space. */
846 /* . (This is a hard limit.) ==== */
848 /* ==== Exceptional deflation windows: try to cure rare */
849 /* . slow convergence by varying the size of the */
850 /* . deflation window after KEXNW iterations. ==== */
852 /* ==== Exceptional shifts: try to cure rare slow convergence */
853 /* . with ad-hoc exceptional shifts every KEXSH iterations. */
856 /* ==== The constants WILK1 and WILK2 are used to form the */
857 /* . exceptional shifts. ==== */
858 /* Parameter adjustments */
860 h_offset = 1 + h_dim1 * 1;
865 z_offset = 1 + z_dim1 * 1;
872 /* ==== Quick return for N = 0: nothing to do. ==== */
881 /* ==== Tiny matrices must use DLAHQR. ==== */
885 dlahqr_(wantt, wantz, n, ilo, ihi, &h__[h_offset], ldh, &wr[1], &
886 wi[1], iloz, ihiz, &z__[z_offset], ldz, info);
890 /* ==== Use small bulge multi-shift QR with aggressive early */
891 /* . deflation on larger-than-tiny matrices. ==== */
893 /* ==== Hope for the best. ==== */
897 /* ==== Set up job flags for ILAENV. ==== */
900 *(unsigned char *)jbcmpz = 'S';
902 *(unsigned char *)jbcmpz = 'E';
905 *(unsigned char *)&jbcmpz[1] = 'V';
907 *(unsigned char *)&jbcmpz[1] = 'N';
910 /* ==== NWR = recommended deflation window size. At this */
911 /* . point, N .GT. NTINY = 15, so there is enough */
912 /* . subdiagonal workspace for NWR.GE.2 as required. */
913 /* . (In fact, there is enough subdiagonal space for */
914 /* . NWR.GE.4.) ==== */
916 nwr = ilaenv_(&c__13, "DLAQR4", jbcmpz, n, ilo, ihi, lwork, (ftnlen)6,
920 i__1 = *ihi - *ilo + 1, i__2 = (*n - 1) / 3, i__1 = f2cmin(i__1,i__2);
921 nwr = f2cmin(i__1,nwr);
923 /* ==== NSR = recommended number of simultaneous shifts. */
924 /* . At this point N .GT. NTINY = 15, so there is at */
925 /* . enough subdiagonal workspace for NSR to be even */
926 /* . and greater than or equal to two as required. ==== */
928 nsr = ilaenv_(&c__15, "DLAQR4", jbcmpz, n, ilo, ihi, lwork, (ftnlen)6,
931 i__1 = nsr, i__2 = (*n - 3) / 6, i__1 = f2cmin(i__1,i__2), i__2 = *ihi -
933 nsr = f2cmin(i__1,i__2);
935 i__1 = 2, i__2 = nsr - nsr % 2;
936 nsr = f2cmax(i__1,i__2);
938 /* ==== Estimate optimal workspace ==== */
940 /* ==== Workspace query call to DLAQR2 ==== */
943 dlaqr2_(wantt, wantz, n, ilo, ihi, &i__1, &h__[h_offset], ldh, iloz,
944 ihiz, &z__[z_offset], ldz, &ls, &ld, &wr[1], &wi[1], &h__[
945 h_offset], ldh, n, &h__[h_offset], ldh, n, &h__[h_offset],
946 ldh, &work[1], &c_n1);
948 /* ==== Optimal workspace = MAX(DLAQR5, DLAQR2) ==== */
951 i__1 = nsr * 3 / 2, i__2 = (integer) work[1];
952 lwkopt = f2cmax(i__1,i__2);
954 /* ==== Quick return in case of workspace query. ==== */
957 work[1] = (doublereal) lwkopt;
961 /* ==== DLAHQR/DLAQR0 crossover point ==== */
963 nmin = ilaenv_(&c__12, "DLAQR4", jbcmpz, n, ilo, ihi, lwork, (ftnlen)
965 nmin = f2cmax(15,nmin);
967 /* ==== Nibble crossover point ==== */
969 nibble = ilaenv_(&c__14, "DLAQR4", jbcmpz, n, ilo, ihi, lwork, (
970 ftnlen)6, (ftnlen)2);
971 nibble = f2cmax(0,nibble);
973 /* ==== Accumulate reflections during ttswp? Use block */
974 /* . 2-by-2 structure during matrix-matrix multiply? ==== */
976 kacc22 = ilaenv_(&c__16, "DLAQR4", jbcmpz, n, ilo, ihi, lwork, (
977 ftnlen)6, (ftnlen)2);
978 kacc22 = f2cmax(0,kacc22);
979 kacc22 = f2cmin(2,kacc22);
981 /* ==== NWMAX = the largest possible deflation window for */
982 /* . which there is sufficient workspace. ==== */
985 i__1 = (*n - 1) / 3, i__2 = *lwork / 2;
986 nwmax = f2cmin(i__1,i__2);
989 /* ==== NSMAX = the Largest number of simultaneous shifts */
990 /* . for which there is sufficient workspace. ==== */
993 i__1 = (*n - 3) / 6, i__2 = (*lwork << 1) / 3;
994 nsmax = f2cmin(i__1,i__2);
997 /* ==== NDFL: an iteration count restarted at deflation. ==== */
1001 /* ==== ITMAX = iteration limit ==== */
1004 i__1 = 10, i__2 = *ihi - *ilo + 1;
1005 itmax = 30 * f2cmax(i__1,i__2);
1007 /* ==== Last row and column in the active block ==== */
1011 /* ==== Main Loop ==== */
1014 for (it = 1; it <= i__1; ++it) {
1016 /* ==== Done when KBOT falls below ILO ==== */
1022 /* ==== Locate active block ==== */
1025 for (k = kbot; k >= i__2; --k) {
1026 if (h__[k + (k - 1) * h_dim1] == 0.) {
1035 /* ==== Select deflation window size: */
1036 /* . Typical Case: */
1037 /* . If possible and advisable, nibble the entire */
1038 /* . active block. If not, use size MIN(NWR,NWMAX) */
1039 /* . or MIN(NWR+1,NWMAX) depending upon which has */
1040 /* . the smaller corresponding subdiagonal entry */
1041 /* . (a heuristic). */
1043 /* . Exceptional Case: */
1044 /* . If there have been no deflations in KEXNW or */
1045 /* . more iterations, then vary the deflation window */
1046 /* . size. At first, because, larger windows are, */
1047 /* . in general, more powerful than smaller ones, */
1048 /* . rapidly increase the window to the maximum possible. */
1049 /* . Then, gradually reduce the window size. ==== */
1051 nh = kbot - ktop + 1;
1052 nwupbd = f2cmin(nh,nwmax);
1054 nw = f2cmin(nwupbd,nwr);
1057 i__2 = nwupbd, i__3 = nw << 1;
1058 nw = f2cmin(i__2,i__3);
1064 kwtop = kbot - nw + 1;
1065 if ((d__1 = h__[kwtop + (kwtop - 1) * h_dim1], abs(d__1))
1066 > (d__2 = h__[kwtop - 1 + (kwtop - 2) * h_dim1],
1074 } else if (ndec >= 0 || nw >= nwupbd) {
1076 if (nw - ndec < 2) {
1082 /* ==== Aggressive early deflation: */
1083 /* . split workspace under the subdiagonal into */
1084 /* . - an nw-by-nw work array V in the lower */
1085 /* . left-hand-corner, */
1086 /* . - an NW-by-at-least-NW-but-more-is-better */
1087 /* . (NW-by-NHO) horizontal work array along */
1088 /* . the bottom edge, */
1089 /* . - an at-least-NW-but-more-is-better (NHV-by-NW) */
1090 /* . vertical work array along the left-hand-edge. */
1095 nho = *n - nw - 1 - kt + 1;
1097 nve = *n - nw - kwv + 1;
1099 /* ==== Aggressive early deflation ==== */
1101 dlaqr2_(wantt, wantz, n, &ktop, &kbot, &nw, &h__[h_offset], ldh,
1102 iloz, ihiz, &z__[z_offset], ldz, &ls, &ld, &wr[1], &wi[1],
1103 &h__[kv + h_dim1], ldh, &nho, &h__[kv + kt * h_dim1],
1104 ldh, &nve, &h__[kwv + h_dim1], ldh, &work[1], lwork);
1106 /* ==== Adjust KBOT accounting for new deflations. ==== */
1110 /* ==== KS points to the shifts. ==== */
1114 /* ==== Skip an expensive QR sweep if there is a (partly */
1115 /* . heuristic) reason to expect that many eigenvalues */
1116 /* . will deflate without it. Here, the QR sweep is */
1117 /* . skipped if many eigenvalues have just been deflated */
1118 /* . or if the remaining active block is small. */
1120 if (ld == 0 || ld * 100 <= nw * nibble && kbot - ktop + 1 > f2cmin(
1123 /* ==== NS = nominal number of simultaneous shifts. */
1124 /* . This may be lowered (slightly) if DLAQR2 */
1125 /* . did not provide that many shifts. ==== */
1129 i__4 = 2, i__5 = kbot - ktop;
1130 i__2 = f2cmin(nsmax,nsr), i__3 = f2cmax(i__4,i__5);
1131 ns = f2cmin(i__2,i__3);
1134 /* ==== If there have been no deflations */
1135 /* . in a multiple of KEXSH iterations, */
1136 /* . then try exceptional shifts. */
1137 /* . Otherwise use shifts provided by */
1138 /* . DLAQR2 above or from the eigenvalues */
1139 /* . of a trailing principal submatrix. ==== */
1141 if (ndfl % 6 == 0) {
1144 i__3 = ks + 1, i__4 = ktop + 2;
1145 i__2 = f2cmax(i__3,i__4);
1146 for (i__ = kbot; i__ >= i__2; i__ += -2) {
1147 ss = (d__1 = h__[i__ + (i__ - 1) * h_dim1], abs(d__1))
1148 + (d__2 = h__[i__ - 1 + (i__ - 2) * h_dim1],
1150 aa = ss * .75 + h__[i__ + i__ * h_dim1];
1154 dlanv2_(&aa, &bb, &cc, &dd, &wr[i__ - 1], &wi[i__ - 1]
1155 , &wr[i__], &wi[i__], &cs, &sn);
1159 wr[ks + 1] = h__[ks + 1 + (ks + 1) * h_dim1];
1161 wr[ks] = wr[ks + 1];
1162 wi[ks] = wi[ks + 1];
1166 /* ==== Got NS/2 or fewer shifts? Use DLAHQR */
1167 /* . on a trailing principal submatrix to */
1168 /* . get more. (Since NS.LE.NSMAX.LE.(N-3)/6, */
1169 /* . there is enough space below the subdiagonal */
1170 /* . to fit an NS-by-NS scratch array.) ==== */
1172 if (kbot - ks + 1 <= ns / 2) {
1175 dlacpy_("A", &ns, &ns, &h__[ks + ks * h_dim1], ldh, &
1176 h__[kt + h_dim1], ldh);
1177 dlahqr_(&c_false, &c_false, &ns, &c__1, &ns, &h__[kt
1178 + h_dim1], ldh, &wr[ks], &wi[ks], &c__1, &
1179 c__1, zdum, &c__1, &inf);
1182 /* ==== In case of a rare QR failure use */
1183 /* . eigenvalues of the trailing 2-by-2 */
1184 /* . principal submatrix. ==== */
1187 aa = h__[kbot - 1 + (kbot - 1) * h_dim1];
1188 cc = h__[kbot + (kbot - 1) * h_dim1];
1189 bb = h__[kbot - 1 + kbot * h_dim1];
1190 dd = h__[kbot + kbot * h_dim1];
1191 dlanv2_(&aa, &bb, &cc, &dd, &wr[kbot - 1], &wi[
1192 kbot - 1], &wr[kbot], &wi[kbot], &cs, &sn)
1198 if (kbot - ks + 1 > ns) {
1200 /* ==== Sort the shifts (Helps a little) */
1201 /* . Bubble sort keeps complex conjugate */
1202 /* . pairs together. ==== */
1206 for (k = kbot; k >= i__2; --k) {
1212 for (i__ = ks; i__ <= i__3; ++i__) {
1213 if ((d__1 = wr[i__], abs(d__1)) + (d__2 = wi[
1214 i__], abs(d__2)) < (d__3 = wr[i__ + 1]
1215 , abs(d__3)) + (d__4 = wi[i__ + 1],
1220 wr[i__] = wr[i__ + 1];
1224 wi[i__] = wi[i__ + 1];
1235 /* ==== Shuffle shifts into pairs of real shifts */
1236 /* . and pairs of complex conjugate shifts */
1237 /* . assuming complex conjugate shifts are */
1238 /* . already adjacent to one another. (Yes, */
1239 /* . they are.) ==== */
1242 for (i__ = kbot; i__ >= i__2; i__ += -2) {
1243 if (wi[i__] != -wi[i__ - 1]) {
1246 wr[i__] = wr[i__ - 1];
1247 wr[i__ - 1] = wr[i__ - 2];
1251 wi[i__] = wi[i__ - 1];
1252 wi[i__ - 1] = wi[i__ - 2];
1259 /* ==== If there are only two shifts and both are */
1260 /* . real, then use only one. ==== */
1262 if (kbot - ks + 1 == 2) {
1263 if (wi[kbot] == 0.) {
1264 if ((d__1 = wr[kbot] - h__[kbot + kbot * h_dim1], abs(
1265 d__1)) < (d__2 = wr[kbot - 1] - h__[kbot +
1266 kbot * h_dim1], abs(d__2))) {
1267 wr[kbot - 1] = wr[kbot];
1269 wr[kbot] = wr[kbot - 1];
1274 /* ==== Use up to NS of the the smallest magnitude */
1275 /* . shifts. If there aren't NS shifts available, */
1276 /* . then use them all, possibly dropping one to */
1277 /* . make the number of shifts even. ==== */
1280 i__2 = ns, i__3 = kbot - ks + 1;
1281 ns = f2cmin(i__2,i__3);
1285 /* ==== Small-bulge multi-shift QR sweep: */
1286 /* . split workspace under the subdiagonal into */
1287 /* . - a KDU-by-KDU work array U in the lower */
1288 /* . left-hand-corner, */
1289 /* . - a KDU-by-at-least-KDU-but-more-is-better */
1290 /* . (KDU-by-NHo) horizontal work array WH along */
1291 /* . the bottom edge, */
1292 /* . - and an at-least-KDU-but-more-is-better-by-KDU */
1293 /* . (NVE-by-KDU) vertical work WV arrow along */
1294 /* . the left-hand-edge. ==== */
1299 nho = *n - kdu - 3 - (kdu + 1) + 1;
1301 nve = *n - kdu - kwv + 1;
1303 /* ==== Small-bulge multi-shift QR sweep ==== */
1305 dlaqr5_(wantt, wantz, &kacc22, n, &ktop, &kbot, &ns, &wr[ks],
1306 &wi[ks], &h__[h_offset], ldh, iloz, ihiz, &z__[
1307 z_offset], ldz, &work[1], &c__3, &h__[ku + h_dim1],
1308 ldh, &nve, &h__[kwv + h_dim1], ldh, &nho, &h__[ku +
1309 kwh * h_dim1], ldh);
1312 /* ==== Note progress (or the lack of it). ==== */
1320 /* ==== End of main loop ==== */
1324 /* ==== Iteration limit exceeded. Set INFO to show where */
1325 /* . the problem occurred and exit. ==== */
1332 /* ==== Return the optimal value of LWORK. ==== */
1334 work[1] = (doublereal) lwkopt;
1336 /* ==== End of DLAQR4 ==== */