14 typedef long long BLASLONG;
15 typedef unsigned long long BLASULONG;
17 typedef long BLASLONG;
18 typedef unsigned long BLASULONG;
22 typedef BLASLONG blasint;
24 #define blasabs(x) llabs(x)
26 #define blasabs(x) labs(x)
30 #define blasabs(x) abs(x)
33 typedef blasint integer;
35 typedef unsigned int uinteger;
36 typedef char *address;
37 typedef short int shortint;
39 typedef double doublereal;
40 typedef struct { real r, i; } complex;
41 typedef struct { doublereal r, i; } doublecomplex;
43 static inline _Fcomplex Cf(complex *z) {_Fcomplex zz={z->r , z->i}; return zz;}
44 static inline _Dcomplex Cd(doublecomplex *z) {_Dcomplex zz={z->r , z->i};return zz;}
45 static inline _Fcomplex * _pCf(complex *z) {return (_Fcomplex*)z;}
46 static inline _Dcomplex * _pCd(doublecomplex *z) {return (_Dcomplex*)z;}
48 static inline _Complex float Cf(complex *z) {return z->r + z->i*_Complex_I;}
49 static inline _Complex double Cd(doublecomplex *z) {return z->r + z->i*_Complex_I;}
50 static inline _Complex float * _pCf(complex *z) {return (_Complex float*)z;}
51 static inline _Complex double * _pCd(doublecomplex *z) {return (_Complex double*)z;}
53 #define pCf(z) (*_pCf(z))
54 #define pCd(z) (*_pCd(z))
56 typedef short int shortlogical;
57 typedef char logical1;
58 typedef char integer1;
63 /* Extern is for use with -E */
74 /*external read, write*/
83 /*internal read, write*/
113 /*rewind, backspace, endfile*/
125 ftnint *inex; /*parameters in standard's order*/
151 union Multitype { /* for multiple entry points */
162 typedef union Multitype Multitype;
164 struct Vardesc { /* for Namelist */
170 typedef struct Vardesc Vardesc;
177 typedef struct Namelist Namelist;
179 #define abs(x) ((x) >= 0 ? (x) : -(x))
180 #define dabs(x) (fabs(x))
181 #define f2cmin(a,b) ((a) <= (b) ? (a) : (b))
182 #define f2cmax(a,b) ((a) >= (b) ? (a) : (b))
183 #define dmin(a,b) (f2cmin(a,b))
184 #define dmax(a,b) (f2cmax(a,b))
185 #define bit_test(a,b) ((a) >> (b) & 1)
186 #define bit_clear(a,b) ((a) & ~((uinteger)1 << (b)))
187 #define bit_set(a,b) ((a) | ((uinteger)1 << (b)))
189 #define abort_() { sig_die("Fortran abort routine called", 1); }
190 #define c_abs(z) (cabsf(Cf(z)))
191 #define c_cos(R,Z) { pCf(R)=ccos(Cf(Z)); }
193 #define c_div(c, a, b) {Cf(c)._Val[0] = (Cf(a)._Val[0]/Cf(b)._Val[0]); Cf(c)._Val[1]=(Cf(a)._Val[1]/Cf(b)._Val[1]);}
194 #define z_div(c, a, b) {Cd(c)._Val[0] = (Cd(a)._Val[0]/Cd(b)._Val[0]); Cd(c)._Val[1]=(Cd(a)._Val[1]/df(b)._Val[1]);}
196 #define c_div(c, a, b) {pCf(c) = Cf(a)/Cf(b);}
197 #define z_div(c, a, b) {pCd(c) = Cd(a)/Cd(b);}
199 #define c_exp(R, Z) {pCf(R) = cexpf(Cf(Z));}
200 #define c_log(R, Z) {pCf(R) = clogf(Cf(Z));}
201 #define c_sin(R, Z) {pCf(R) = csinf(Cf(Z));}
202 //#define c_sqrt(R, Z) {*(R) = csqrtf(Cf(Z));}
203 #define c_sqrt(R, Z) {pCf(R) = csqrtf(Cf(Z));}
204 #define d_abs(x) (fabs(*(x)))
205 #define d_acos(x) (acos(*(x)))
206 #define d_asin(x) (asin(*(x)))
207 #define d_atan(x) (atan(*(x)))
208 #define d_atn2(x, y) (atan2(*(x),*(y)))
209 #define d_cnjg(R, Z) { pCd(R) = conj(Cd(Z)); }
210 #define r_cnjg(R, Z) { pCf(R) = conjf(Cf(Z)); }
211 #define d_cos(x) (cos(*(x)))
212 #define d_cosh(x) (cosh(*(x)))
213 #define d_dim(__a, __b) ( *(__a) > *(__b) ? *(__a) - *(__b) : 0.0 )
214 #define d_exp(x) (exp(*(x)))
215 #define d_imag(z) (cimag(Cd(z)))
216 #define r_imag(z) (cimagf(Cf(z)))
217 #define d_int(__x) (*(__x)>0 ? floor(*(__x)) : -floor(- *(__x)))
218 #define r_int(__x) (*(__x)>0 ? floor(*(__x)) : -floor(- *(__x)))
219 #define d_lg10(x) ( 0.43429448190325182765 * log(*(x)) )
220 #define r_lg10(x) ( 0.43429448190325182765 * log(*(x)) )
221 #define d_log(x) (log(*(x)))
222 #define d_mod(x, y) (fmod(*(x), *(y)))
223 #define u_nint(__x) ((__x)>=0 ? floor((__x) + .5) : -floor(.5 - (__x)))
224 #define d_nint(x) u_nint(*(x))
225 #define u_sign(__a,__b) ((__b) >= 0 ? ((__a) >= 0 ? (__a) : -(__a)) : -((__a) >= 0 ? (__a) : -(__a)))
226 #define d_sign(a,b) u_sign(*(a),*(b))
227 #define r_sign(a,b) u_sign(*(a),*(b))
228 #define d_sin(x) (sin(*(x)))
229 #define d_sinh(x) (sinh(*(x)))
230 #define d_sqrt(x) (sqrt(*(x)))
231 #define d_tan(x) (tan(*(x)))
232 #define d_tanh(x) (tanh(*(x)))
233 #define i_abs(x) abs(*(x))
234 #define i_dnnt(x) ((integer)u_nint(*(x)))
235 #define i_len(s, n) (n)
236 #define i_nint(x) ((integer)u_nint(*(x)))
237 #define i_sign(a,b) ((integer)u_sign((integer)*(a),(integer)*(b)))
238 #define pow_dd(ap, bp) ( pow(*(ap), *(bp)))
239 #define pow_si(B,E) spow_ui(*(B),*(E))
240 #define pow_ri(B,E) spow_ui(*(B),*(E))
241 #define pow_di(B,E) dpow_ui(*(B),*(E))
242 #define pow_zi(p, a, b) {pCd(p) = zpow_ui(Cd(a), *(b));}
243 #define pow_ci(p, a, b) {pCf(p) = cpow_ui(Cf(a), *(b));}
244 #define pow_zz(R,A,B) {pCd(R) = cpow(Cd(A),*(B));}
245 #define s_cat(lpp, rpp, rnp, np, llp) { ftnlen i, nc, ll; char *f__rp, *lp; ll = (llp); lp = (lpp); for(i=0; i < (int)*(np); ++i) { nc = ll; if((rnp)[i] < nc) nc = (rnp)[i]; ll -= nc; f__rp = (rpp)[i]; while(--nc >= 0) *lp++ = *(f__rp)++; } while(--ll >= 0) *lp++ = ' '; }
246 #define s_cmp(a,b,c,d) ((integer)strncmp((a),(b),f2cmin((c),(d))))
247 #define s_copy(A,B,C,D) { int __i,__m; for (__i=0, __m=f2cmin((C),(D)); __i<__m && (B)[__i] != 0; ++__i) (A)[__i] = (B)[__i]; }
248 #define sig_die(s, kill) { exit(1); }
249 #define s_stop(s, n) {exit(0);}
250 static char junk[] = "\n@(#)LIBF77 VERSION 19990503\n";
251 #define z_abs(z) (cabs(Cd(z)))
252 #define z_exp(R, Z) {pCd(R) = cexp(Cd(Z));}
253 #define z_sqrt(R, Z) {pCd(R) = csqrt(Cd(Z));}
254 #define myexit_() break;
255 #define mycycle() continue;
256 #define myceiling(w) {ceil(w)}
257 #define myhuge(w) {HUGE_VAL}
258 //#define mymaxloc_(w,s,e,n) {if (sizeof(*(w)) == sizeof(double)) dmaxloc_((w),*(s),*(e),n); else dmaxloc_((w),*(s),*(e),n);}
259 #define mymaxloc(w,s,e,n) {dmaxloc_(w,*(s),*(e),n)}
261 /* procedure parameter types for -A and -C++ */
263 #define F2C_proc_par_types 1
265 typedef logical (*L_fp)(...);
267 typedef logical (*L_fp)();
270 static float spow_ui(float x, integer n) {
271 float pow=1.0; unsigned long int u;
273 if(n < 0) n = -n, x = 1/x;
282 static double dpow_ui(double x, integer n) {
283 double pow=1.0; unsigned long int u;
285 if(n < 0) n = -n, x = 1/x;
295 static _Fcomplex cpow_ui(complex x, integer n) {
296 complex pow={1.0,0.0}; unsigned long int u;
298 if(n < 0) n = -n, x.r = 1/x.r, x.i=1/x.i;
300 if(u & 01) pow.r *= x.r, pow.i *= x.i;
301 if(u >>= 1) x.r *= x.r, x.i *= x.i;
305 _Fcomplex p={pow.r, pow.i};
309 static _Complex float cpow_ui(_Complex float x, integer n) {
310 _Complex float pow=1.0; unsigned long int u;
312 if(n < 0) n = -n, x = 1/x;
323 static _Dcomplex zpow_ui(_Dcomplex x, integer n) {
324 _Dcomplex pow={1.0,0.0}; unsigned long int u;
326 if(n < 0) n = -n, x._Val[0] = 1/x._Val[0], x._Val[1] =1/x._Val[1];
328 if(u & 01) pow._Val[0] *= x._Val[0], pow._Val[1] *= x._Val[1];
329 if(u >>= 1) x._Val[0] *= x._Val[0], x._Val[1] *= x._Val[1];
333 _Dcomplex p = {pow._Val[0], pow._Val[1]};
337 static _Complex double zpow_ui(_Complex double x, integer n) {
338 _Complex double pow=1.0; unsigned long int u;
340 if(n < 0) n = -n, x = 1/x;
350 static integer pow_ii(integer x, integer n) {
351 integer pow; unsigned long int u;
353 if (n == 0 || x == 1) pow = 1;
354 else if (x != -1) pow = x == 0 ? 1/x : 0;
357 if ((n > 0) || !(n == 0 || x == 1 || x != -1)) {
367 static integer dmaxloc_(double *w, integer s, integer e, integer *n)
369 double m; integer i, mi;
370 for(m=w[s-1], mi=s, i=s+1; i<=e; i++)
371 if (w[i-1]>m) mi=i ,m=w[i-1];
374 static integer smaxloc_(float *w, integer s, integer e, integer *n)
376 float m; integer i, mi;
377 for(m=w[s-1], mi=s, i=s+1; i<=e; i++)
378 if (w[i-1]>m) mi=i ,m=w[i-1];
381 static inline void cdotc_(complex *z, integer *n_, complex *x, integer *incx_, complex *y, integer *incy_) {
382 integer n = *n_, incx = *incx_, incy = *incy_, i;
384 _Fcomplex zdotc = {0.0, 0.0};
385 if (incx == 1 && incy == 1) {
386 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
387 zdotc._Val[0] += conjf(Cf(&x[i]))._Val[0] * Cf(&y[i])._Val[0];
388 zdotc._Val[1] += conjf(Cf(&x[i]))._Val[1] * Cf(&y[i])._Val[1];
391 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
392 zdotc._Val[0] += conjf(Cf(&x[i*incx]))._Val[0] * Cf(&y[i*incy])._Val[0];
393 zdotc._Val[1] += conjf(Cf(&x[i*incx]))._Val[1] * Cf(&y[i*incy])._Val[1];
399 _Complex float zdotc = 0.0;
400 if (incx == 1 && incy == 1) {
401 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
402 zdotc += conjf(Cf(&x[i])) * Cf(&y[i]);
405 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
406 zdotc += conjf(Cf(&x[i*incx])) * Cf(&y[i*incy]);
412 static inline void zdotc_(doublecomplex *z, integer *n_, doublecomplex *x, integer *incx_, doublecomplex *y, integer *incy_) {
413 integer n = *n_, incx = *incx_, incy = *incy_, i;
415 _Dcomplex zdotc = {0.0, 0.0};
416 if (incx == 1 && incy == 1) {
417 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
418 zdotc._Val[0] += conj(Cd(&x[i]))._Val[0] * Cd(&y[i])._Val[0];
419 zdotc._Val[1] += conj(Cd(&x[i]))._Val[1] * Cd(&y[i])._Val[1];
422 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
423 zdotc._Val[0] += conj(Cd(&x[i*incx]))._Val[0] * Cd(&y[i*incy])._Val[0];
424 zdotc._Val[1] += conj(Cd(&x[i*incx]))._Val[1] * Cd(&y[i*incy])._Val[1];
430 _Complex double zdotc = 0.0;
431 if (incx == 1 && incy == 1) {
432 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
433 zdotc += conj(Cd(&x[i])) * Cd(&y[i]);
436 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
437 zdotc += conj(Cd(&x[i*incx])) * Cd(&y[i*incy]);
443 static inline void cdotu_(complex *z, integer *n_, complex *x, integer *incx_, complex *y, integer *incy_) {
444 integer n = *n_, incx = *incx_, incy = *incy_, i;
446 _Fcomplex zdotc = {0.0, 0.0};
447 if (incx == 1 && incy == 1) {
448 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
449 zdotc._Val[0] += Cf(&x[i])._Val[0] * Cf(&y[i])._Val[0];
450 zdotc._Val[1] += Cf(&x[i])._Val[1] * Cf(&y[i])._Val[1];
453 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
454 zdotc._Val[0] += Cf(&x[i*incx])._Val[0] * Cf(&y[i*incy])._Val[0];
455 zdotc._Val[1] += Cf(&x[i*incx])._Val[1] * Cf(&y[i*incy])._Val[1];
461 _Complex float zdotc = 0.0;
462 if (incx == 1 && incy == 1) {
463 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
464 zdotc += Cf(&x[i]) * Cf(&y[i]);
467 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
468 zdotc += Cf(&x[i*incx]) * Cf(&y[i*incy]);
474 static inline void zdotu_(doublecomplex *z, integer *n_, doublecomplex *x, integer *incx_, doublecomplex *y, integer *incy_) {
475 integer n = *n_, incx = *incx_, incy = *incy_, i;
477 _Dcomplex zdotc = {0.0, 0.0};
478 if (incx == 1 && incy == 1) {
479 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
480 zdotc._Val[0] += Cd(&x[i])._Val[0] * Cd(&y[i])._Val[0];
481 zdotc._Val[1] += Cd(&x[i])._Val[1] * Cd(&y[i])._Val[1];
484 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
485 zdotc._Val[0] += Cd(&x[i*incx])._Val[0] * Cd(&y[i*incy])._Val[0];
486 zdotc._Val[1] += Cd(&x[i*incx])._Val[1] * Cd(&y[i*incy])._Val[1];
492 _Complex double zdotc = 0.0;
493 if (incx == 1 && incy == 1) {
494 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
495 zdotc += Cd(&x[i]) * Cd(&y[i]);
498 for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
499 zdotc += Cd(&x[i*incx]) * Cd(&y[i*incy]);
505 /* -- translated by f2c (version 20000121).
506 You must link the resulting object file with the libraries:
507 -lf2c -lm (in that order)
513 /* Table of constant values */
515 static integer c__1 = 1;
516 static integer c__0 = 0;
517 static real c_b27 = 1.f;
519 /* > \brief \b CGSVJ0 pre-processor for the routine cgesvj. */
521 /* =========== DOCUMENTATION =========== */
523 /* Online html documentation available at */
524 /* http://www.netlib.org/lapack/explore-html/ */
527 /* > Download CGSVJ0 + dependencies */
528 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/cgsvj0.
531 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/cgsvj0.
534 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/cgsvj0.
542 /* SUBROUTINE CGSVJ0( JOBV, M, N, A, LDA, D, SVA, MV, V, LDV, EPS, */
543 /* SFMIN, TOL, NSWEEP, WORK, LWORK, INFO ) */
545 /* INTEGER INFO, LDA, LDV, LWORK, M, MV, N, NSWEEP */
546 /* REAL EPS, SFMIN, TOL */
547 /* CHARACTER*1 JOBV */
548 /* COMPLEX A( LDA, * ), D( N ), V( LDV, * ), WORK( LWORK ) */
552 /* > \par Purpose: */
557 /* > CGSVJ0 is called from CGESVJ as a pre-processor and that is its main */
558 /* > purpose. It applies Jacobi rotations in the same way as CGESVJ does, but */
559 /* > it does not check convergence (stopping criterion). Few tuning */
560 /* > parameters (marked by [TP]) are available for the implementer. */
566 /* > \param[in] JOBV */
568 /* > JOBV is CHARACTER*1 */
569 /* > Specifies whether the output from this procedure is used */
570 /* > to compute the matrix V: */
571 /* > = 'V': the product of the Jacobi rotations is accumulated */
572 /* > by postmulyiplying the N-by-N array V. */
573 /* > (See the description of V.) */
574 /* > = 'A': the product of the Jacobi rotations is accumulated */
575 /* > by postmulyiplying the MV-by-N array V. */
576 /* > (See the descriptions of MV and V.) */
577 /* > = 'N': the Jacobi rotations are not accumulated. */
583 /* > The number of rows of the input matrix A. M >= 0. */
589 /* > The number of columns of the input matrix A. */
593 /* > \param[in,out] A */
595 /* > A is COMPLEX array, dimension (LDA,N) */
596 /* > On entry, M-by-N matrix A, such that A*diag(D) represents */
597 /* > the input matrix. */
599 /* > A_onexit * diag(D_onexit) represents the input matrix A*diag(D) */
600 /* > post-multiplied by a sequence of Jacobi rotations, where the */
601 /* > rotation threshold and the total number of sweeps are given in */
602 /* > TOL and NSWEEP, respectively. */
603 /* > (See the descriptions of D, TOL and NSWEEP.) */
606 /* > \param[in] LDA */
608 /* > LDA is INTEGER */
609 /* > The leading dimension of the array A. LDA >= f2cmax(1,M). */
612 /* > \param[in,out] D */
614 /* > D is COMPLEX array, dimension (N) */
615 /* > The array D accumulates the scaling factors from the complex scaled */
616 /* > Jacobi rotations. */
617 /* > On entry, A*diag(D) represents the input matrix. */
618 /* > On exit, A_onexit*diag(D_onexit) represents the input matrix */
619 /* > post-multiplied by a sequence of Jacobi rotations, where the */
620 /* > rotation threshold and the total number of sweeps are given in */
621 /* > TOL and NSWEEP, respectively. */
622 /* > (See the descriptions of A, TOL and NSWEEP.) */
625 /* > \param[in,out] SVA */
627 /* > SVA is REAL array, dimension (N) */
628 /* > On entry, SVA contains the Euclidean norms of the columns of */
629 /* > the matrix A*diag(D). */
630 /* > On exit, SVA contains the Euclidean norms of the columns of */
631 /* > the matrix A_onexit*diag(D_onexit). */
634 /* > \param[in] MV */
636 /* > MV is INTEGER */
637 /* > If JOBV = 'A', then MV rows of V are post-multipled by a */
638 /* > sequence of Jacobi rotations. */
639 /* > If JOBV = 'N', then MV is not referenced. */
642 /* > \param[in,out] V */
644 /* > V is COMPLEX array, dimension (LDV,N) */
645 /* > If JOBV = 'V' then N rows of V are post-multipled by a */
646 /* > sequence of Jacobi rotations. */
647 /* > If JOBV = 'A' then MV rows of V are post-multipled by a */
648 /* > sequence of Jacobi rotations. */
649 /* > If JOBV = 'N', then V is not referenced. */
652 /* > \param[in] LDV */
654 /* > LDV is INTEGER */
655 /* > The leading dimension of the array V, LDV >= 1. */
656 /* > If JOBV = 'V', LDV >= N. */
657 /* > If JOBV = 'A', LDV >= MV. */
660 /* > \param[in] EPS */
663 /* > EPS = SLAMCH('Epsilon') */
666 /* > \param[in] SFMIN */
668 /* > SFMIN is REAL */
669 /* > SFMIN = SLAMCH('Safe Minimum') */
672 /* > \param[in] TOL */
675 /* > TOL is the threshold for Jacobi rotations. For a pair */
676 /* > A(:,p), A(:,q) of pivot columns, the Jacobi rotation is */
677 /* > applied only if ABS(COS(angle(A(:,p),A(:,q)))) > TOL. */
680 /* > \param[in] NSWEEP */
682 /* > NSWEEP is INTEGER */
683 /* > NSWEEP is the number of sweeps of Jacobi rotations to be */
687 /* > \param[out] WORK */
689 /* > WORK is COMPLEX array, dimension (LWORK) */
692 /* > \param[in] LWORK */
694 /* > LWORK is INTEGER */
695 /* > LWORK is the dimension of WORK. LWORK >= M. */
698 /* > \param[out] INFO */
700 /* > INFO is INTEGER */
701 /* > = 0: successful exit. */
702 /* > < 0: if INFO = -i, then the i-th argument had an illegal value */
708 /* > \author Univ. of Tennessee */
709 /* > \author Univ. of California Berkeley */
710 /* > \author Univ. of Colorado Denver */
711 /* > \author NAG Ltd. */
713 /* > \date June 2016 */
715 /* > \ingroup complexOTHERcomputational */
717 /* > \par Further Details: */
718 /* ===================== */
720 /* > CGSVJ0 is used just to enable CGESVJ to call a simplified version of */
721 /* > itself to work on a submatrix of the original matrix. */
723 /* > \par Contributor: */
724 /* ================== */
726 /* > Zlatko Drmac (Zagreb, Croatia) */
728 /* > \par Bugs, Examples and Comments: */
729 /* ================================= */
731 /* > Please report all bugs and send interesting test examples and comments to */
732 /* > drmac@math.hr. Thank you. */
734 /* ===================================================================== */
735 /* Subroutine */ int cgsvj0_(char *jobv, integer *m, integer *n, complex *a,
736 integer *lda, complex *d__, real *sva, integer *mv, complex *v,
737 integer *ldv, real *eps, real *sfmin, real *tol, integer *nsweep,
738 complex *work, integer *lwork, integer *info)
740 /* System generated locals */
741 integer a_dim1, a_offset, v_dim1, v_offset, i__1, i__2, i__3, i__4, i__5,
744 complex q__1, q__2, q__3;
746 /* Local variables */
752 extern /* Subroutine */ int crot_(integer *, complex *, integer *,
753 complex *, integer *, real *, complex *);
756 real aapp0, aapq1, temp1;
759 extern /* Complex */ VOID cdotc_(complex *, integer *, complex *, integer
760 *, complex *, integer *);
762 extern logical lsame_(char *, char *);
764 extern /* Subroutine */ int ccopy_(integer *, complex *, integer *,
765 complex *, integer *), cswap_(integer *, complex *, integer *,
766 complex *, integer *);
767 logical applv, rsvec;
768 extern /* Subroutine */ int caxpy_(integer *, complex *, complex *,
769 integer *, complex *, integer *);
772 extern real scnrm2_(integer *, complex *, integer *);
774 extern /* Subroutine */ int clascl_(char *, integer *, integer *, real *,
775 real *, integer *, integer *, complex *, integer *, integer *), xerbla_(char *, integer *, ftnlen);
776 integer ijblsk, swband;
777 extern integer isamax_(integer *, real *, integer *);
779 extern /* Subroutine */ int classq_(integer *, complex *, integer *, real
781 real mxaapq, thsign, mxsinj;
782 integer ir1, emptsw, notrot, iswrot, jbc;
784 integer kbl, lkahead, igl, ibr, jgl, nbl, mvl;
785 real rootbig, rooteps;
790 /* -- LAPACK computational routine (version 3.8.0) -- */
791 /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
792 /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
796 /* ===================================================================== */
801 /* Test the input parameters. */
803 /* Parameter adjustments */
807 a_offset = 1 + a_dim1 * 1;
810 v_offset = 1 + v_dim1 * 1;
815 applv = lsame_(jobv, "A");
816 rsvec = lsame_(jobv, "V");
817 if (! (rsvec || applv || lsame_(jobv, "N"))) {
821 } else if (*n < 0 || *n > *m) {
823 } else if (*lda < *m) {
825 } else if ((rsvec || applv) && *mv < 0) {
827 } else if (rsvec && *ldv < *n || applv && *ldv < *mv) {
829 } else if (*tol <= *eps) {
831 } else if (*nsweep < 0) {
833 } else if (*lwork < *m) {
842 xerbla_("CGSVJ0", &i__1, (ftnlen)6);
851 rsvec = rsvec || applv;
852 rooteps = sqrt(*eps);
853 rootsfmin = sqrt(*sfmin);
854 small = *sfmin / *eps;
856 rootbig = 1.f / rootsfmin;
857 bigtheta = 1.f / rooteps;
858 roottol = sqrt(*tol);
861 emptsw = *n * (*n - 1) / 2;
866 /* [TP] SWBAND is a tuning parameter [TP]. It is meaningful and effective */
867 /* if CGESVJ is used as a computational routine in the preconditioned */
868 /* Jacobi SVD algorithm CGEJSV. For sweeps i=1:SWBAND the procedure */
869 /* works on pivots inside a band-like region around the diagonal. */
870 /* The boundaries are determined dynamically, based on the number of */
871 /* pivots above a threshold. */
874 /* [TP] KBL is a tuning parameter that defines the tile size in the */
875 /* tiling of the p-q loops of pivot pairs. In general, an optimal */
876 /* value of KBL depends on the matrix dimensions and on the */
877 /* parameters of the computer's memory. */
880 if (nbl * kbl != *n) {
884 /* Computing 2nd power */
886 blskip = i__1 * i__1;
887 /* [TP] BLKSKIP is a tuning parameter that depends on SWBAND and KBL. */
889 rowskip = f2cmin(5,kbl);
890 /* [TP] ROWSKIP is a tuning parameter. */
893 /* [TP] LKAHEAD is a tuning parameter. */
895 /* Quasi block transformations, using the lower (upper) triangular */
896 /* structure of the input matrix. The quasi-block-cycling usually */
897 /* invokes cubic convergence. Big part of this cycle is done inside */
898 /* canonical subspaces of dimensions less than M. */
903 for (i__ = 1; i__ <= i__1; ++i__) {
913 /* Each sweep is unrolled using KBL-by-KBL tiles over the pivot pairs */
914 /* 1 <= p < q <= N. This is the first step toward a blocked implementation */
915 /* of the rotations. New implementation, based on block transformations, */
916 /* is under development. */
919 for (ibr = 1; ibr <= i__2; ++ibr) {
921 igl = (ibr - 1) * kbl + 1;
924 i__4 = lkahead, i__5 = nbl - ibr;
925 i__3 = f2cmin(i__4,i__5);
926 for (ir1 = 0; ir1 <= i__3; ++ir1) {
931 i__5 = igl + kbl - 1, i__6 = *n - 1;
932 i__4 = f2cmin(i__5,i__6);
933 for (p = igl; p <= i__4; ++p) {
937 q = isamax_(&i__5, &sva[p], &c__1) + p - 1;
939 cswap_(m, &a[p * a_dim1 + 1], &c__1, &a[q * a_dim1 +
942 cswap_(&mvl, &v[p * v_dim1 + 1], &c__1, &v[q *
949 aapq.r = d__[i__5].r, aapq.i = d__[i__5].i;
952 d__[i__5].r = d__[i__6].r, d__[i__5].i = d__[i__6].i;
954 d__[i__5].r = aapq.r, d__[i__5].i = aapq.i;
959 /* Column norms are periodically updated by explicit */
960 /* norm computation. */
962 /* Unfortunately, some BLAS implementations compute SNCRM2(M,A(1,p),1) */
963 /* as SQRT(S=CDOTC(M,A(1,p),1,A(1,p),1)), which may cause the result to */
964 /* overflow for ||A(:,p)||_2 > SQRT(overflow_threshold), and to */
965 /* underflow for ||A(:,p)||_2 < SQRT(underflow_threshold). */
966 /* Hence, SCNRM2 cannot be trusted, not even in the case when */
967 /* the true norm is far from the under(over)flow boundaries. */
968 /* If properly implemented SCNRM2 is available, the IF-THEN-ELSE-END IF */
969 /* below should be replaced with "AAPP = SCNRM2( M, A(1,p), 1 )". */
971 if (sva[p] < rootbig && sva[p] > rootsfmin) {
972 sva[p] = scnrm2_(m, &a[p * a_dim1 + 1], &c__1);
976 classq_(m, &a[p * a_dim1 + 1], &c__1, &temp1, &
978 sva[p] = temp1 * sqrt(aapp);
990 i__6 = igl + kbl - 1;
991 i__5 = f2cmin(i__6,*n);
992 for (q = p + 1; q <= i__5; ++q) {
1000 rotok = small * aapp <= aaqq;
1001 if (aapp < big / aaqq) {
1002 cdotc_(&q__3, m, &a[p * a_dim1 + 1], &
1003 c__1, &a[q * a_dim1 + 1], &
1005 q__2.r = q__3.r / aaqq, q__2.i =
1007 q__1.r = q__2.r / aapp, q__1.i =
1009 aapq.r = q__1.r, aapq.i = q__1.i;
1011 ccopy_(m, &a[p * a_dim1 + 1], &c__1, &
1013 clascl_("G", &c__0, &c__0, &aapp, &
1014 c_b27, m, &c__1, &work[1],
1016 cdotc_(&q__2, m, &work[1], &c__1, &a[
1017 q * a_dim1 + 1], &c__1);
1018 q__1.r = q__2.r / aaqq, q__1.i =
1020 aapq.r = q__1.r, aapq.i = q__1.i;
1023 rotok = aapp <= aaqq / small;
1024 if (aapp > small / aaqq) {
1025 cdotc_(&q__3, m, &a[p * a_dim1 + 1], &
1026 c__1, &a[q * a_dim1 + 1], &
1028 q__2.r = q__3.r / aapp, q__2.i =
1030 q__1.r = q__2.r / aaqq, q__1.i =
1032 aapq.r = q__1.r, aapq.i = q__1.i;
1034 ccopy_(m, &a[q * a_dim1 + 1], &c__1, &
1036 clascl_("G", &c__0, &c__0, &aaqq, &
1037 c_b27, m, &c__1, &work[1],
1039 cdotc_(&q__2, m, &a[p * a_dim1 + 1], &
1040 c__1, &work[1], &c__1);
1041 q__1.r = q__2.r / aapp, q__1.i =
1043 aapq.r = q__1.r, aapq.i = q__1.i;
1047 /* AAPQ = AAPQ * CONJG( CWORK(p) ) * CWORK(q) */
1048 aapq1 = -c_abs(&aapq);
1050 r__1 = mxaapq, r__2 = -aapq1;
1051 mxaapq = f2cmax(r__1,r__2);
1053 /* TO rotate or NOT to rotate, THAT is the question ... */
1055 if (abs(aapq1) > *tol) {
1056 r__1 = c_abs(&aapq);
1057 q__1.r = aapq.r / r__1, q__1.i = aapq.i /
1059 ompq.r = q__1.r, ompq.i = q__1.i;
1061 /* [RTD] ROTATED = ROTATED + ONE */
1071 aqoap = aaqq / aapp;
1072 apoaq = aapp / aaqq;
1073 theta = (r__1 = aqoap - apoaq, abs(
1074 r__1)) * -.5f / aapq1;
1076 if (abs(theta) > bigtheta) {
1080 r_cnjg(&q__2, &ompq);
1081 q__1.r = t * q__2.r, q__1.i = t *
1083 crot_(m, &a[p * a_dim1 + 1], &
1084 c__1, &a[q * a_dim1 + 1],
1087 r_cnjg(&q__2, &ompq);
1088 q__1.r = t * q__2.r, q__1.i = t * q__2.i;
1089 crot_(&mvl, &v[p * v_dim1 + 1], &c__1, &v[q *
1090 v_dim1 + 1], &c__1, &cs, &q__1);
1093 r__1 = 0.f, r__2 = t * apoaq *
1095 sva[q] = aaqq * sqrt((f2cmax(r__1,
1098 r__1 = 0.f, r__2 = 1.f - t *
1100 aapp *= sqrt((f2cmax(r__1,r__2)));
1102 r__1 = mxsinj, r__2 = abs(t);
1103 mxsinj = f2cmax(r__1,r__2);
1108 thsign = -r_sign(&c_b27, &aapq1);
1109 t = 1.f / (theta + thsign * sqrt(
1110 theta * theta + 1.f));
1111 cs = sqrt(1.f / (t * t + 1.f));
1115 r__1 = mxsinj, r__2 = abs(sn);
1116 mxsinj = f2cmax(r__1,r__2);
1118 r__1 = 0.f, r__2 = t * apoaq *
1120 sva[q] = aaqq * sqrt((f2cmax(r__1,
1123 r__1 = 0.f, r__2 = 1.f - t *
1125 aapp *= sqrt((f2cmax(r__1,r__2)));
1127 r_cnjg(&q__2, &ompq);
1128 q__1.r = sn * q__2.r, q__1.i = sn
1130 crot_(m, &a[p * a_dim1 + 1], &
1131 c__1, &a[q * a_dim1 + 1],
1134 r_cnjg(&q__2, &ompq);
1135 q__1.r = sn * q__2.r, q__1.i = sn * q__2.i;
1136 crot_(&mvl, &v[p * v_dim1 + 1], &c__1, &v[q *
1137 v_dim1 + 1], &c__1, &cs, &q__1);
1142 q__2.r = -d__[i__7].r, q__2.i = -d__[
1144 q__1.r = q__2.r * ompq.r - q__2.i *
1145 ompq.i, q__1.i = q__2.r *
1146 ompq.i + q__2.i * ompq.r;
1147 d__[i__6].r = q__1.r, d__[i__6].i =
1151 ccopy_(m, &a[p * a_dim1 + 1], &c__1, &
1153 clascl_("G", &c__0, &c__0, &aapp, &
1154 c_b27, m, &c__1, &work[1],
1156 clascl_("G", &c__0, &c__0, &aaqq, &
1157 c_b27, m, &c__1, &a[q *
1158 a_dim1 + 1], lda, &ierr);
1159 q__1.r = -aapq.r, q__1.i = -aapq.i;
1160 caxpy_(m, &q__1, &work[1], &c__1, &a[
1161 q * a_dim1 + 1], &c__1);
1162 clascl_("G", &c__0, &c__0, &c_b27, &
1163 aaqq, m, &c__1, &a[q * a_dim1
1166 r__1 = 0.f, r__2 = 1.f - aapq1 *
1168 sva[q] = aaqq * sqrt((f2cmax(r__1,r__2)))
1170 mxsinj = f2cmax(mxsinj,*sfmin);
1172 /* END IF ROTOK THEN ... ELSE */
1174 /* In the case of cancellation in updating SVA(q), SVA(p) */
1175 /* recompute SVA(q), SVA(p). */
1177 /* Computing 2nd power */
1178 r__1 = sva[q] / aaqq;
1179 if (r__1 * r__1 <= rooteps) {
1180 if (aaqq < rootbig && aaqq >
1182 sva[q] = scnrm2_(m, &a[q * a_dim1
1187 classq_(m, &a[q * a_dim1 + 1], &
1189 sva[q] = t * sqrt(aaqq);
1192 if (aapp / aapp0 <= rooteps) {
1193 if (aapp < rootbig && aapp >
1195 aapp = scnrm2_(m, &a[p * a_dim1 +
1200 classq_(m, &a[p * a_dim1 + 1], &
1202 aapp = t * sqrt(aapp);
1208 /* A(:,p) and A(:,q) already numerically orthogonal */
1212 /* [RTD] SKIPPED = SKIPPED + 1 */
1216 /* A(:,q) is zero column */
1223 if (i__ <= swband && pskipped > rowskip) {
1236 /* bailed out of q-loop */
1242 if (ir1 == 0 && aapp == 0.f) {
1244 i__5 = igl + kbl - 1;
1245 notrot = notrot + f2cmin(i__5,*n) - p;
1251 /* end of the p-loop */
1252 /* end of doing the block ( ibr, ibr ) */
1255 /* end of ir1-loop */
1257 /* ... go to the off diagonal blocks */
1259 igl = (ibr - 1) * kbl + 1;
1262 for (jbc = ibr + 1; jbc <= i__3; ++jbc) {
1264 jgl = (jbc - 1) * kbl + 1;
1266 /* doing the block at ( ibr, jbc ) */
1270 i__5 = igl + kbl - 1;
1271 i__4 = f2cmin(i__5,*n);
1272 for (p = igl; p <= i__4; ++p) {
1280 i__6 = jgl + kbl - 1;
1281 i__5 = f2cmin(i__6,*n);
1282 for (q = jgl; q <= i__5; ++q) {
1289 /* Safe Gram matrix computation */
1293 rotok = small * aapp <= aaqq;
1295 rotok = small * aaqq <= aapp;
1297 if (aapp < big / aaqq) {
1298 cdotc_(&q__3, m, &a[p * a_dim1 + 1], &
1299 c__1, &a[q * a_dim1 + 1], &
1301 q__2.r = q__3.r / aaqq, q__2.i =
1303 q__1.r = q__2.r / aapp, q__1.i =
1305 aapq.r = q__1.r, aapq.i = q__1.i;
1307 ccopy_(m, &a[p * a_dim1 + 1], &c__1, &
1309 clascl_("G", &c__0, &c__0, &aapp, &
1310 c_b27, m, &c__1, &work[1],
1312 cdotc_(&q__2, m, &work[1], &c__1, &a[
1313 q * a_dim1 + 1], &c__1);
1314 q__1.r = q__2.r / aaqq, q__1.i =
1316 aapq.r = q__1.r, aapq.i = q__1.i;
1320 rotok = aapp <= aaqq / small;
1322 rotok = aaqq <= aapp / small;
1324 if (aapp > small / aaqq) {
1325 cdotc_(&q__3, m, &a[p * a_dim1 + 1], &
1326 c__1, &a[q * a_dim1 + 1], &
1328 r__1 = f2cmax(aaqq,aapp);
1329 q__2.r = q__3.r / r__1, q__2.i =
1331 r__2 = f2cmin(aaqq,aapp);
1332 q__1.r = q__2.r / r__2, q__1.i =
1334 aapq.r = q__1.r, aapq.i = q__1.i;
1336 ccopy_(m, &a[q * a_dim1 + 1], &c__1, &
1338 clascl_("G", &c__0, &c__0, &aaqq, &
1339 c_b27, m, &c__1, &work[1],
1341 cdotc_(&q__2, m, &a[p * a_dim1 + 1], &
1342 c__1, &work[1], &c__1);
1343 q__1.r = q__2.r / aapp, q__1.i =
1345 aapq.r = q__1.r, aapq.i = q__1.i;
1349 /* AAPQ = AAPQ * CONJG(CWORK(p))*CWORK(q) */
1350 aapq1 = -c_abs(&aapq);
1352 r__1 = mxaapq, r__2 = -aapq1;
1353 mxaapq = f2cmax(r__1,r__2);
1355 /* TO rotate or NOT to rotate, THAT is the question ... */
1357 if (abs(aapq1) > *tol) {
1358 r__1 = c_abs(&aapq);
1359 q__1.r = aapq.r / r__1, q__1.i = aapq.i /
1361 ompq.r = q__1.r, ompq.i = q__1.i;
1363 /* [RTD] ROTATED = ROTATED + 1 */
1369 aqoap = aaqq / aapp;
1370 apoaq = aapp / aaqq;
1371 theta = (r__1 = aqoap - apoaq, abs(
1372 r__1)) * -.5f / aapq1;
1377 if (abs(theta) > bigtheta) {
1380 r_cnjg(&q__2, &ompq);
1381 q__1.r = t * q__2.r, q__1.i = t *
1383 crot_(m, &a[p * a_dim1 + 1], &
1384 c__1, &a[q * a_dim1 + 1],
1387 r_cnjg(&q__2, &ompq);
1388 q__1.r = t * q__2.r, q__1.i = t * q__2.i;
1389 crot_(&mvl, &v[p * v_dim1 + 1], &c__1, &v[q *
1390 v_dim1 + 1], &c__1, &cs, &q__1);
1393 r__1 = 0.f, r__2 = t * apoaq *
1395 sva[q] = aaqq * sqrt((f2cmax(r__1,
1398 r__1 = 0.f, r__2 = 1.f - t *
1400 aapp *= sqrt((f2cmax(r__1,r__2)));
1402 r__1 = mxsinj, r__2 = abs(t);
1403 mxsinj = f2cmax(r__1,r__2);
1407 thsign = -r_sign(&c_b27, &aapq1);
1411 t = 1.f / (theta + thsign * sqrt(
1412 theta * theta + 1.f));
1413 cs = sqrt(1.f / (t * t + 1.f));
1416 r__1 = mxsinj, r__2 = abs(sn);
1417 mxsinj = f2cmax(r__1,r__2);
1419 r__1 = 0.f, r__2 = t * apoaq *
1421 sva[q] = aaqq * sqrt((f2cmax(r__1,
1424 r__1 = 0.f, r__2 = 1.f - t *
1426 aapp *= sqrt((f2cmax(r__1,r__2)));
1428 r_cnjg(&q__2, &ompq);
1429 q__1.r = sn * q__2.r, q__1.i = sn
1431 crot_(m, &a[p * a_dim1 + 1], &
1432 c__1, &a[q * a_dim1 + 1],
1435 r_cnjg(&q__2, &ompq);
1436 q__1.r = sn * q__2.r, q__1.i = sn * q__2.i;
1437 crot_(&mvl, &v[p * v_dim1 + 1], &c__1, &v[q *
1438 v_dim1 + 1], &c__1, &cs, &q__1);
1443 q__2.r = -d__[i__7].r, q__2.i = -d__[
1445 q__1.r = q__2.r * ompq.r - q__2.i *
1446 ompq.i, q__1.i = q__2.r *
1447 ompq.i + q__2.i * ompq.r;
1448 d__[i__6].r = q__1.r, d__[i__6].i =
1453 ccopy_(m, &a[p * a_dim1 + 1], &
1454 c__1, &work[1], &c__1);
1455 clascl_("G", &c__0, &c__0, &aapp,
1456 &c_b27, m, &c__1, &work[1]
1458 clascl_("G", &c__0, &c__0, &aaqq,
1459 &c_b27, m, &c__1, &a[q *
1460 a_dim1 + 1], lda, &ierr);
1461 q__1.r = -aapq.r, q__1.i =
1463 caxpy_(m, &q__1, &work[1], &c__1,
1464 &a[q * a_dim1 + 1], &c__1)
1466 clascl_("G", &c__0, &c__0, &c_b27,
1467 &aaqq, m, &c__1, &a[q *
1468 a_dim1 + 1], lda, &ierr);
1470 r__1 = 0.f, r__2 = 1.f - aapq1 *
1472 sva[q] = aaqq * sqrt((f2cmax(r__1,
1474 mxsinj = f2cmax(mxsinj,*sfmin);
1476 ccopy_(m, &a[q * a_dim1 + 1], &
1477 c__1, &work[1], &c__1);
1478 clascl_("G", &c__0, &c__0, &aaqq,
1479 &c_b27, m, &c__1, &work[1]
1481 clascl_("G", &c__0, &c__0, &aapp,
1482 &c_b27, m, &c__1, &a[p *
1483 a_dim1 + 1], lda, &ierr);
1484 r_cnjg(&q__2, &aapq);
1485 q__1.r = -q__2.r, q__1.i =
1487 caxpy_(m, &q__1, &work[1], &c__1,
1488 &a[p * a_dim1 + 1], &c__1)
1490 clascl_("G", &c__0, &c__0, &c_b27,
1491 &aapp, m, &c__1, &a[p *
1492 a_dim1 + 1], lda, &ierr);
1494 r__1 = 0.f, r__2 = 1.f - aapq1 *
1496 sva[p] = aapp * sqrt((f2cmax(r__1,
1498 mxsinj = f2cmax(mxsinj,*sfmin);
1501 /* END IF ROTOK THEN ... ELSE */
1503 /* In the case of cancellation in updating SVA(q), SVA(p) */
1504 /* Computing 2nd power */
1505 r__1 = sva[q] / aaqq;
1506 if (r__1 * r__1 <= rooteps) {
1507 if (aaqq < rootbig && aaqq >
1509 sva[q] = scnrm2_(m, &a[q * a_dim1
1514 classq_(m, &a[q * a_dim1 + 1], &
1516 sva[q] = t * sqrt(aaqq);
1519 /* Computing 2nd power */
1520 r__1 = aapp / aapp0;
1521 if (r__1 * r__1 <= rooteps) {
1522 if (aapp < rootbig && aapp >
1524 aapp = scnrm2_(m, &a[p * a_dim1 +
1529 classq_(m, &a[p * a_dim1 + 1], &
1531 aapp = t * sqrt(aapp);
1535 /* end of OK rotation */
1538 /* [RTD] SKIPPED = SKIPPED + 1 */
1548 if (i__ <= swband && ijblsk >= blskip) {
1553 if (i__ <= swband && pskipped > rowskip) {
1561 /* end of the q-loop */
1570 i__5 = jgl + kbl - 1;
1571 notrot = notrot + f2cmin(i__5,*n) - jgl + 1;
1581 /* end of the p-loop */
1584 /* end of the jbc-loop */
1586 /* 2011 bailed out of the jbc-loop */
1588 i__4 = igl + kbl - 1;
1589 i__3 = f2cmin(i__4,*n);
1590 for (p = igl; p <= i__3; ++p) {
1591 sva[p] = (r__1 = sva[p], abs(r__1));
1597 /* 2000 :: end of the ibr-loop */
1599 if (sva[*n] < rootbig && sva[*n] > rootsfmin) {
1600 sva[*n] = scnrm2_(m, &a[*n * a_dim1 + 1], &c__1);
1604 classq_(m, &a[*n * a_dim1 + 1], &c__1, &t, &aapp);
1605 sva[*n] = t * sqrt(aapp);
1608 /* Additional steering devices */
1610 if (i__ < swband && (mxaapq <= roottol || iswrot <= *n)) {
1614 if (i__ > swband + 1 && mxaapq < sqrt((real) (*n)) * *tol && (real) (*
1615 n) * mxaapq * mxsinj < *tol) {
1619 if (notrot >= emptsw) {
1625 /* end i=1:NSWEEP loop */
1627 /* #:( Reaching this point means that the procedure has not converged. */
1628 *info = *nsweep - 1;
1632 /* #:) Reaching this point means numerical convergence after the i-th */
1636 /* #:) INFO = 0 confirms successful iterations. */
1639 /* Sort the vector SVA() of column norms. */
1641 for (p = 1; p <= i__1; ++p) {
1643 q = isamax_(&i__2, &sva[p], &c__1) + p - 1;
1649 aapq.r = d__[i__2].r, aapq.i = d__[i__2].i;
1652 d__[i__2].r = d__[i__3].r, d__[i__2].i = d__[i__3].i;
1654 d__[i__2].r = aapq.r, d__[i__2].i = aapq.i;
1655 cswap_(m, &a[p * a_dim1 + 1], &c__1, &a[q * a_dim1 + 1], &c__1);
1657 cswap_(&mvl, &v[p * v_dim1 + 1], &c__1, &v[q * v_dim1 + 1], &