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 doublereal c_b15 = -.125;
516 static integer c__1 = 1;
517 static real c_b49 = 1.f;
518 static real c_b72 = -1.f;
520 /* > \brief \b SBDSQR */
522 /* =========== DOCUMENTATION =========== */
524 /* Online html documentation available at */
525 /* http://www.netlib.org/lapack/explore-html/ */
528 /* > Download SBDSQR + dependencies */
529 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/sbdsqr.
532 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/sbdsqr.
535 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/sbdsqr.
543 /* SUBROUTINE SBDSQR( UPLO, N, NCVT, NRU, NCC, D, E, VT, LDVT, U, */
544 /* LDU, C, LDC, WORK, INFO ) */
547 /* INTEGER INFO, LDC, LDU, LDVT, N, NCC, NCVT, NRU */
548 /* REAL C( LDC, * ), D( * ), E( * ), U( LDU, * ), */
549 /* $ VT( LDVT, * ), WORK( * ) */
552 /* > \par Purpose: */
557 /* > SBDSQR computes the singular values and, optionally, the right and/or */
558 /* > left singular vectors from the singular value decomposition (SVD) of */
559 /* > a real N-by-N (upper or lower) bidiagonal matrix B using the implicit */
560 /* > zero-shift QR algorithm. The SVD of B has the form */
562 /* > B = Q * S * P**T */
564 /* > where S is the diagonal matrix of singular values, Q is an orthogonal */
565 /* > matrix of left singular vectors, and P is an orthogonal matrix of */
566 /* > right singular vectors. If left singular vectors are requested, this */
567 /* > subroutine actually returns U*Q instead of Q, and, if right singular */
568 /* > vectors are requested, this subroutine returns P**T*VT instead of */
569 /* > P**T, for given real input matrices U and VT. When U and VT are the */
570 /* > orthogonal matrices that reduce a general matrix A to bidiagonal */
571 /* > form: A = U*B*VT, as computed by SGEBRD, then */
573 /* > A = (U*Q) * S * (P**T*VT) */
575 /* > is the SVD of A. Optionally, the subroutine may also compute Q**T*C */
576 /* > for a given real input matrix C. */
578 /* > See "Computing Small Singular Values of Bidiagonal Matrices With */
579 /* > Guaranteed High Relative Accuracy," by J. Demmel and W. Kahan, */
580 /* > LAPACK Working Note #3 (or SIAM J. Sci. Statist. Comput. vol. 11, */
581 /* > no. 5, pp. 873-912, Sept 1990) and */
582 /* > "Accurate singular values and differential qd algorithms," by */
583 /* > B. Parlett and V. Fernando, Technical Report CPAM-554, Mathematics */
584 /* > Department, University of California at Berkeley, July 1992 */
585 /* > for a detailed description of the algorithm. */
591 /* > \param[in] UPLO */
593 /* > UPLO is CHARACTER*1 */
594 /* > = 'U': B is upper bidiagonal; */
595 /* > = 'L': B is lower bidiagonal. */
601 /* > The order of the matrix B. N >= 0. */
604 /* > \param[in] NCVT */
606 /* > NCVT is INTEGER */
607 /* > The number of columns of the matrix VT. NCVT >= 0. */
610 /* > \param[in] NRU */
612 /* > NRU is INTEGER */
613 /* > The number of rows of the matrix U. NRU >= 0. */
616 /* > \param[in] NCC */
618 /* > NCC is INTEGER */
619 /* > The number of columns of the matrix C. NCC >= 0. */
622 /* > \param[in,out] D */
624 /* > D is REAL array, dimension (N) */
625 /* > On entry, the n diagonal elements of the bidiagonal matrix B. */
626 /* > On exit, if INFO=0, the singular values of B in decreasing */
630 /* > \param[in,out] E */
632 /* > E is REAL array, dimension (N-1) */
633 /* > On entry, the N-1 offdiagonal elements of the bidiagonal */
635 /* > On exit, if INFO = 0, E is destroyed; if INFO > 0, D and E */
636 /* > will contain the diagonal and superdiagonal elements of a */
637 /* > bidiagonal matrix orthogonally equivalent to the one given */
641 /* > \param[in,out] VT */
643 /* > VT is REAL array, dimension (LDVT, NCVT) */
644 /* > On entry, an N-by-NCVT matrix VT. */
645 /* > On exit, VT is overwritten by P**T * VT. */
646 /* > Not referenced if NCVT = 0. */
649 /* > \param[in] LDVT */
651 /* > LDVT is INTEGER */
652 /* > The leading dimension of the array VT. */
653 /* > LDVT >= f2cmax(1,N) if NCVT > 0; LDVT >= 1 if NCVT = 0. */
656 /* > \param[in,out] U */
658 /* > U is REAL array, dimension (LDU, N) */
659 /* > On entry, an NRU-by-N matrix U. */
660 /* > On exit, U is overwritten by U * Q. */
661 /* > Not referenced if NRU = 0. */
664 /* > \param[in] LDU */
666 /* > LDU is INTEGER */
667 /* > The leading dimension of the array U. LDU >= f2cmax(1,NRU). */
670 /* > \param[in,out] C */
672 /* > C is REAL array, dimension (LDC, NCC) */
673 /* > On entry, an N-by-NCC matrix C. */
674 /* > On exit, C is overwritten by Q**T * C. */
675 /* > Not referenced if NCC = 0. */
678 /* > \param[in] LDC */
680 /* > LDC is INTEGER */
681 /* > The leading dimension of the array C. */
682 /* > LDC >= f2cmax(1,N) if NCC > 0; LDC >=1 if NCC = 0. */
685 /* > \param[out] WORK */
687 /* > WORK is REAL array, dimension (4*N) */
690 /* > \param[out] INFO */
692 /* > INFO is INTEGER */
693 /* > = 0: successful exit */
694 /* > < 0: If INFO = -i, the i-th argument had an illegal value */
696 /* > if NCVT = NRU = NCC = 0, */
697 /* > = 1, a split was marked by a positive value in E */
698 /* > = 2, current block of Z not diagonalized after 30*N */
699 /* > iterations (in inner while loop) */
700 /* > = 3, termination criterion of outer while loop not met */
701 /* > (program created more than N unreduced blocks) */
702 /* > else NCVT = NRU = NCC = 0, */
703 /* > the algorithm did not converge; D and E contain the */
704 /* > elements of a bidiagonal matrix which is orthogonally */
705 /* > similar to the input matrix B; if INFO = i, i */
706 /* > elements of E have not converged to zero. */
709 /* > \par Internal Parameters: */
710 /* ========================= */
713 /* > TOLMUL REAL, default = f2cmax(10,f2cmin(100,EPS**(-1/8))) */
714 /* > TOLMUL controls the convergence criterion of the QR loop. */
715 /* > If it is positive, TOLMUL*EPS is the desired relative */
716 /* > precision in the computed singular values. */
717 /* > If it is negative, abs(TOLMUL*EPS*sigma_max) is the */
718 /* > desired absolute accuracy in the computed singular */
719 /* > values (corresponds to relative accuracy */
720 /* > abs(TOLMUL*EPS) in the largest singular value. */
721 /* > abs(TOLMUL) should be between 1 and 1/EPS, and preferably */
722 /* > between 10 (for fast convergence) and .1/EPS */
723 /* > (for there to be some accuracy in the results). */
724 /* > Default is to lose at either one eighth or 2 of the */
725 /* > available decimal digits in each computed singular value */
726 /* > (whichever is smaller). */
728 /* > MAXITR INTEGER, default = 6 */
729 /* > MAXITR controls the maximum number of passes of the */
730 /* > algorithm through its inner loop. The algorithms stops */
731 /* > (and so fails to converge) if the number of passes */
732 /* > through the inner loop exceeds MAXITR*N**2. */
739 /* > Bug report from Cezary Dendek. */
740 /* > On March 23rd 2017, the INTEGER variable MAXIT = MAXITR*N**2 is */
741 /* > removed since it can overflow pretty easily (for N larger or equal */
742 /* > than 18,919). We instead use MAXITDIVN = MAXITR*N. */
748 /* > \author Univ. of Tennessee */
749 /* > \author Univ. of California Berkeley */
750 /* > \author Univ. of Colorado Denver */
751 /* > \author NAG Ltd. */
753 /* > \date June 2017 */
755 /* > \ingroup auxOTHERcomputational */
757 /* ===================================================================== */
758 /* Subroutine */ int sbdsqr_(char *uplo, integer *n, integer *ncvt, integer *
759 nru, integer *ncc, real *d__, real *e, real *vt, integer *ldvt, real *
760 u, integer *ldu, real *c__, integer *ldc, real *work, integer *info)
762 /* System generated locals */
763 integer c_dim1, c_offset, u_dim1, u_offset, vt_dim1, vt_offset, i__1,
765 real r__1, r__2, r__3, r__4;
768 /* Local variables */
775 real unfl, sinl, cosr, smin, smax, sinr;
776 extern /* Subroutine */ int srot_(integer *, real *, integer *, real *,
777 integer *, real *, real *);
779 extern /* Subroutine */ int slas2_(real *, real *, real *, real *, real *)
784 extern logical lsame_(char *, char *);
786 extern /* Subroutine */ int sscal_(integer *, real *, real *, integer *);
788 real shift, sigmn, oldsn, sminl;
789 extern /* Subroutine */ int slasr_(char *, char *, char *, integer *,
790 integer *, real *, real *, real *, integer *);
793 extern /* Subroutine */ int sswap_(integer *, real *, integer *, real *,
796 extern /* Subroutine */ int slasq1_(integer *, real *, real *, real *,
797 integer *), slasv2_(real *, real *, real *, real *, real *, real *
798 , real *, real *, real *);
802 extern real slamch_(char *);
803 extern /* Subroutine */ int xerbla_(char *, integer *, ftnlen);
805 extern /* Subroutine */ int slartg_(real *, real *, real *, real *, real *
811 integer nm12, nm13, lll;
815 /* -- LAPACK computational routine (version 3.7.1) -- */
816 /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
817 /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
821 /* ===================================================================== */
824 /* Test the input parameters. */
826 /* Parameter adjustments */
830 vt_offset = 1 + vt_dim1 * 1;
833 u_offset = 1 + u_dim1 * 1;
836 c_offset = 1 + c_dim1 * 1;
842 lower = lsame_(uplo, "L");
843 if (! lsame_(uplo, "U") && ! lower) {
847 } else if (*ncvt < 0) {
849 } else if (*nru < 0) {
851 } else if (*ncc < 0) {
853 } else if (*ncvt == 0 && *ldvt < 1 || *ncvt > 0 && *ldvt < f2cmax(1,*n)) {
855 } else if (*ldu < f2cmax(1,*nru)) {
857 } else if (*ncc == 0 && *ldc < 1 || *ncc > 0 && *ldc < f2cmax(1,*n)) {
862 xerbla_("SBDSQR", &i__1, (ftnlen)6);
872 /* ROTATE is true if any singular vectors desired, false otherwise */
874 rotate = *ncvt > 0 || *nru > 0 || *ncc > 0;
876 /* If no singular vectors desired, use qd algorithm */
879 slasq1_(n, &d__[1], &e[1], &work[1], info);
881 /* If INFO equals 2, dqds didn't finish, try to finish */
894 /* Get machine constants */
896 eps = slamch_("Epsilon");
897 unfl = slamch_("Safe minimum");
899 /* If matrix lower bidiagonal, rotate to be upper bidiagonal */
900 /* by applying Givens rotations on the left */
904 for (i__ = 1; i__ <= i__1; ++i__) {
905 slartg_(&d__[i__], &e[i__], &cs, &sn, &r__);
907 e[i__] = sn * d__[i__ + 1];
908 d__[i__ + 1] = cs * d__[i__ + 1];
910 work[nm1 + i__] = sn;
914 /* Update singular vectors if desired */
917 slasr_("R", "V", "F", nru, n, &work[1], &work[*n], &u[u_offset],
921 slasr_("L", "V", "F", n, ncc, &work[1], &work[*n], &c__[c_offset],
926 /* Compute singular values to relative accuracy TOL */
927 /* (By setting TOL to be negative, algorithm will compute */
928 /* singular values to absolute accuracy ABS(TOL)*norm(input matrix)) */
932 d__1 = (doublereal) eps;
933 r__3 = 100.f, r__4 = pow_dd(&d__1, &c_b15);
934 r__1 = 10.f, r__2 = f2cmin(r__3,r__4);
935 tolmul = f2cmax(r__1,r__2);
938 /* Compute approximate maximum, minimum singular values */
942 for (i__ = 1; i__ <= i__1; ++i__) {
944 r__2 = smax, r__3 = (r__1 = d__[i__], abs(r__1));
945 smax = f2cmax(r__2,r__3);
949 for (i__ = 1; i__ <= i__1; ++i__) {
951 r__2 = smax, r__3 = (r__1 = e[i__], abs(r__1));
952 smax = f2cmax(r__2,r__3);
958 /* Relative accuracy desired */
960 sminoa = abs(d__[1]);
966 for (i__ = 2; i__ <= i__1; ++i__) {
967 mu = (r__2 = d__[i__], abs(r__2)) * (mu / (mu + (r__1 = e[i__ - 1]
969 sminoa = f2cmin(sminoa,mu);
976 sminoa /= sqrt((real) (*n));
978 r__1 = tol * sminoa, r__2 = *n * (*n * unfl) * 6;
979 thresh = f2cmax(r__1,r__2);
982 /* Absolute accuracy desired */
985 r__1 = abs(tol) * smax, r__2 = *n * (*n * unfl) * 6;
986 thresh = f2cmax(r__1,r__2);
989 /* Prepare for main iteration loop for the singular values */
990 /* (MAXIT is the maximum number of passes through the inner */
991 /* loop permitted before nonconvergence signalled.) */
999 /* M points to last element of unconverged part of matrix */
1003 /* Begin main iteration loop */
1007 /* Check for convergence or exceeding iteration count */
1016 if (iterdivn >= maxitdivn) {
1021 /* Find diagonal block of matrix to work on */
1023 if (tol < 0.f && (r__1 = d__[m], abs(r__1)) <= thresh) {
1026 smax = (r__1 = d__[m], abs(r__1));
1029 for (lll = 1; lll <= i__1; ++lll) {
1031 abss = (r__1 = d__[ll], abs(r__1));
1032 abse = (r__1 = e[ll], abs(r__1));
1033 if (tol < 0.f && abss <= thresh) {
1036 if (abse <= thresh) {
1039 smin = f2cmin(smin,abss);
1041 r__1 = f2cmax(smax,abss);
1042 smax = f2cmax(r__1,abse);
1050 /* Matrix splits since E(LL) = 0 */
1054 /* Convergence of bottom singular value, return to top of loop */
1062 /* E(LL) through E(M-1) are nonzero, E(LL-1) is zero */
1066 /* 2 by 2 block, handle separately */
1068 slasv2_(&d__[m - 1], &e[m - 1], &d__[m], &sigmn, &sigmx, &sinr, &cosr,
1074 /* Compute singular vectors, if desired */
1077 srot_(ncvt, &vt[m - 1 + vt_dim1], ldvt, &vt[m + vt_dim1], ldvt, &
1081 srot_(nru, &u[(m - 1) * u_dim1 + 1], &c__1, &u[m * u_dim1 + 1], &
1082 c__1, &cosl, &sinl);
1085 srot_(ncc, &c__[m - 1 + c_dim1], ldc, &c__[m + c_dim1], ldc, &
1092 /* If working on new submatrix, choose shift direction */
1093 /* (from larger end diagonal element towards smaller) */
1095 if (ll > oldm || m < oldll) {
1096 if ((r__1 = d__[ll], abs(r__1)) >= (r__2 = d__[m], abs(r__2))) {
1098 /* Chase bulge from top (big end) to bottom (small end) */
1103 /* Chase bulge from bottom (big end) to top (small end) */
1109 /* Apply convergence tests */
1113 /* Run convergence test in forward direction */
1114 /* First apply standard test to bottom of matrix */
1116 if ((r__2 = e[m - 1], abs(r__2)) <= abs(tol) * (r__1 = d__[m], abs(
1117 r__1)) || tol < 0.f && (r__3 = e[m - 1], abs(r__3)) <= thresh)
1125 /* If relative accuracy desired, */
1126 /* apply convergence criterion forward */
1128 mu = (r__1 = d__[ll], abs(r__1));
1131 for (lll = ll; lll <= i__1; ++lll) {
1132 if ((r__1 = e[lll], abs(r__1)) <= tol * mu) {
1136 mu = (r__2 = d__[lll + 1], abs(r__2)) * (mu / (mu + (r__1 = e[
1138 sminl = f2cmin(sminl,mu);
1145 /* Run convergence test in backward direction */
1146 /* First apply standard test to top of matrix */
1148 if ((r__2 = e[ll], abs(r__2)) <= abs(tol) * (r__1 = d__[ll], abs(r__1)
1149 ) || tol < 0.f && (r__3 = e[ll], abs(r__3)) <= thresh) {
1156 /* If relative accuracy desired, */
1157 /* apply convergence criterion backward */
1159 mu = (r__1 = d__[m], abs(r__1));
1162 for (lll = m - 1; lll >= i__1; --lll) {
1163 if ((r__1 = e[lll], abs(r__1)) <= tol * mu) {
1167 mu = (r__2 = d__[lll], abs(r__2)) * (mu / (mu + (r__1 = e[lll]
1169 sminl = f2cmin(sminl,mu);
1177 /* Compute shift. First, test if shifting would ruin relative */
1178 /* accuracy, and if so set the shift to zero. */
1181 r__1 = eps, r__2 = tol * .01f;
1182 if (tol >= 0.f && *n * tol * (sminl / smax) <= f2cmax(r__1,r__2)) {
1184 /* Use a zero shift to avoid loss of relative accuracy */
1189 /* Compute the shift from 2-by-2 block at end of matrix */
1192 sll = (r__1 = d__[ll], abs(r__1));
1193 slas2_(&d__[m - 1], &e[m - 1], &d__[m], &shift, &r__);
1195 sll = (r__1 = d__[m], abs(r__1));
1196 slas2_(&d__[ll], &e[ll], &d__[ll + 1], &shift, &r__);
1199 /* Test if shift negligible, and if so set to zero */
1202 /* Computing 2nd power */
1204 if (r__1 * r__1 < eps) {
1210 /* Increment iteration count */
1212 iter = iter + m - ll;
1214 /* If SHIFT = 0, do simplified QR iteration */
1219 /* Chase bulge from top to bottom */
1220 /* Save cosines and sines for later singular vector updates */
1225 for (i__ = ll; i__ <= i__1; ++i__) {
1226 r__1 = d__[i__] * cs;
1227 slartg_(&r__1, &e[i__], &cs, &sn, &r__);
1229 e[i__ - 1] = oldsn * r__;
1232 r__2 = d__[i__ + 1] * sn;
1233 slartg_(&r__1, &r__2, &oldcs, &oldsn, &d__[i__]);
1234 work[i__ - ll + 1] = cs;
1235 work[i__ - ll + 1 + nm1] = sn;
1236 work[i__ - ll + 1 + nm12] = oldcs;
1237 work[i__ - ll + 1 + nm13] = oldsn;
1241 d__[m] = h__ * oldcs;
1242 e[m - 1] = h__ * oldsn;
1244 /* Update singular vectors */
1248 slasr_("L", "V", "F", &i__1, ncvt, &work[1], &work[*n], &vt[
1249 ll + vt_dim1], ldvt);
1253 slasr_("R", "V", "F", nru, &i__1, &work[nm12 + 1], &work[nm13
1254 + 1], &u[ll * u_dim1 + 1], ldu);
1258 slasr_("L", "V", "F", &i__1, ncc, &work[nm12 + 1], &work[nm13
1259 + 1], &c__[ll + c_dim1], ldc);
1262 /* Test convergence */
1264 if ((r__1 = e[m - 1], abs(r__1)) <= thresh) {
1270 /* Chase bulge from bottom to top */
1271 /* Save cosines and sines for later singular vector updates */
1276 for (i__ = m; i__ >= i__1; --i__) {
1277 r__1 = d__[i__] * cs;
1278 slartg_(&r__1, &e[i__ - 1], &cs, &sn, &r__);
1280 e[i__] = oldsn * r__;
1283 r__2 = d__[i__ - 1] * sn;
1284 slartg_(&r__1, &r__2, &oldcs, &oldsn, &d__[i__]);
1285 work[i__ - ll] = cs;
1286 work[i__ - ll + nm1] = -sn;
1287 work[i__ - ll + nm12] = oldcs;
1288 work[i__ - ll + nm13] = -oldsn;
1292 d__[ll] = h__ * oldcs;
1293 e[ll] = h__ * oldsn;
1295 /* Update singular vectors */
1299 slasr_("L", "V", "B", &i__1, ncvt, &work[nm12 + 1], &work[
1300 nm13 + 1], &vt[ll + vt_dim1], ldvt);
1304 slasr_("R", "V", "B", nru, &i__1, &work[1], &work[*n], &u[ll *
1309 slasr_("L", "V", "B", &i__1, ncc, &work[1], &work[*n], &c__[
1313 /* Test convergence */
1315 if ((r__1 = e[ll], abs(r__1)) <= thresh) {
1321 /* Use nonzero shift */
1325 /* Chase bulge from top to bottom */
1326 /* Save cosines and sines for later singular vector updates */
1328 f = ((r__1 = d__[ll], abs(r__1)) - shift) * (r_sign(&c_b49, &d__[
1329 ll]) + shift / d__[ll]);
1332 for (i__ = ll; i__ <= i__1; ++i__) {
1333 slartg_(&f, &g, &cosr, &sinr, &r__);
1337 f = cosr * d__[i__] + sinr * e[i__];
1338 e[i__] = cosr * e[i__] - sinr * d__[i__];
1339 g = sinr * d__[i__ + 1];
1340 d__[i__ + 1] = cosr * d__[i__ + 1];
1341 slartg_(&f, &g, &cosl, &sinl, &r__);
1343 f = cosl * e[i__] + sinl * d__[i__ + 1];
1344 d__[i__ + 1] = cosl * d__[i__ + 1] - sinl * e[i__];
1346 g = sinl * e[i__ + 1];
1347 e[i__ + 1] = cosl * e[i__ + 1];
1349 work[i__ - ll + 1] = cosr;
1350 work[i__ - ll + 1 + nm1] = sinr;
1351 work[i__ - ll + 1 + nm12] = cosl;
1352 work[i__ - ll + 1 + nm13] = sinl;
1357 /* Update singular vectors */
1361 slasr_("L", "V", "F", &i__1, ncvt, &work[1], &work[*n], &vt[
1362 ll + vt_dim1], ldvt);
1366 slasr_("R", "V", "F", nru, &i__1, &work[nm12 + 1], &work[nm13
1367 + 1], &u[ll * u_dim1 + 1], ldu);
1371 slasr_("L", "V", "F", &i__1, ncc, &work[nm12 + 1], &work[nm13
1372 + 1], &c__[ll + c_dim1], ldc);
1375 /* Test convergence */
1377 if ((r__1 = e[m - 1], abs(r__1)) <= thresh) {
1383 /* Chase bulge from bottom to top */
1384 /* Save cosines and sines for later singular vector updates */
1386 f = ((r__1 = d__[m], abs(r__1)) - shift) * (r_sign(&c_b49, &d__[m]
1387 ) + shift / d__[m]);
1390 for (i__ = m; i__ >= i__1; --i__) {
1391 slartg_(&f, &g, &cosr, &sinr, &r__);
1395 f = cosr * d__[i__] + sinr * e[i__ - 1];
1396 e[i__ - 1] = cosr * e[i__ - 1] - sinr * d__[i__];
1397 g = sinr * d__[i__ - 1];
1398 d__[i__ - 1] = cosr * d__[i__ - 1];
1399 slartg_(&f, &g, &cosl, &sinl, &r__);
1401 f = cosl * e[i__ - 1] + sinl * d__[i__ - 1];
1402 d__[i__ - 1] = cosl * d__[i__ - 1] - sinl * e[i__ - 1];
1404 g = sinl * e[i__ - 2];
1405 e[i__ - 2] = cosl * e[i__ - 2];
1407 work[i__ - ll] = cosr;
1408 work[i__ - ll + nm1] = -sinr;
1409 work[i__ - ll + nm12] = cosl;
1410 work[i__ - ll + nm13] = -sinl;
1415 /* Test convergence */
1417 if ((r__1 = e[ll], abs(r__1)) <= thresh) {
1421 /* Update singular vectors if desired */
1425 slasr_("L", "V", "B", &i__1, ncvt, &work[nm12 + 1], &work[
1426 nm13 + 1], &vt[ll + vt_dim1], ldvt);
1430 slasr_("R", "V", "B", nru, &i__1, &work[1], &work[*n], &u[ll *
1435 slasr_("L", "V", "B", &i__1, ncc, &work[1], &work[*n], &c__[
1441 /* QR iteration finished, go back and check convergence */
1445 /* All singular values converged, so make them positive */
1449 for (i__ = 1; i__ <= i__1; ++i__) {
1450 if (d__[i__] < 0.f) {
1451 d__[i__] = -d__[i__];
1453 /* Change sign of singular vectors, if desired */
1456 sscal_(ncvt, &c_b72, &vt[i__ + vt_dim1], ldvt);
1462 /* Sort the singular values into decreasing order (insertion sort on */
1463 /* singular values, but only one transposition per singular vector) */
1466 for (i__ = 1; i__ <= i__1; ++i__) {
1468 /* Scan for smallest D(I) */
1472 i__2 = *n + 1 - i__;
1473 for (j = 2; j <= i__2; ++j) {
1474 if (d__[j] <= smin) {
1480 if (isub != *n + 1 - i__) {
1482 /* Swap singular values and vectors */
1484 d__[isub] = d__[*n + 1 - i__];
1485 d__[*n + 1 - i__] = smin;
1487 sswap_(ncvt, &vt[isub + vt_dim1], ldvt, &vt[*n + 1 - i__ +
1491 sswap_(nru, &u[isub * u_dim1 + 1], &c__1, &u[(*n + 1 - i__) *
1492 u_dim1 + 1], &c__1);
1495 sswap_(ncc, &c__[isub + c_dim1], ldc, &c__[*n + 1 - i__ +
1503 /* Maximum number of iterations exceeded, failure to converge */
1508 for (i__ = 1; i__ <= i__1; ++i__) {
1509 if (e[i__] != 0.f) {