You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.

slarrf.c 26 kB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905
  1. /* f2c.h -- Standard Fortran to C header file */
  2. /** barf [ba:rf] 2. "He suggested using FORTRAN, and everybody barfed."
  3. - From The Shogakukan DICTIONARY OF NEW ENGLISH (Second edition) */
  4. #ifndef F2C_INCLUDE
  5. #define F2C_INCLUDE
  6. #include <math.h>
  7. #include <stdlib.h>
  8. #include <string.h>
  9. #include <stdio.h>
  10. #include <complex.h>
  11. #ifdef complex
  12. #undef complex
  13. #endif
  14. #ifdef I
  15. #undef I
  16. #endif
  17. typedef int integer;
  18. typedef unsigned int uinteger;
  19. typedef char *address;
  20. typedef short int shortint;
  21. typedef float real;
  22. typedef double doublereal;
  23. typedef struct { real r, i; } complex;
  24. typedef struct { doublereal r, i; } doublecomplex;
  25. static inline _Complex float Cf(complex *z) {return z->r + z->i*_Complex_I;}
  26. static inline _Complex double Cd(doublecomplex *z) {return z->r + z->i*_Complex_I;}
  27. static inline _Complex float * _pCf(complex *z) {return (_Complex float*)z;}
  28. static inline _Complex double * _pCd(doublecomplex *z) {return (_Complex double*)z;}
  29. #define pCf(z) (*_pCf(z))
  30. #define pCd(z) (*_pCd(z))
  31. typedef int logical;
  32. typedef short int shortlogical;
  33. typedef char logical1;
  34. typedef char integer1;
  35. #define TRUE_ (1)
  36. #define FALSE_ (0)
  37. /* Extern is for use with -E */
  38. #ifndef Extern
  39. #define Extern extern
  40. #endif
  41. /* I/O stuff */
  42. typedef int flag;
  43. typedef int ftnlen;
  44. typedef int ftnint;
  45. /*external read, write*/
  46. typedef struct
  47. { flag cierr;
  48. ftnint ciunit;
  49. flag ciend;
  50. char *cifmt;
  51. ftnint cirec;
  52. } cilist;
  53. /*internal read, write*/
  54. typedef struct
  55. { flag icierr;
  56. char *iciunit;
  57. flag iciend;
  58. char *icifmt;
  59. ftnint icirlen;
  60. ftnint icirnum;
  61. } icilist;
  62. /*open*/
  63. typedef struct
  64. { flag oerr;
  65. ftnint ounit;
  66. char *ofnm;
  67. ftnlen ofnmlen;
  68. char *osta;
  69. char *oacc;
  70. char *ofm;
  71. ftnint orl;
  72. char *oblnk;
  73. } olist;
  74. /*close*/
  75. typedef struct
  76. { flag cerr;
  77. ftnint cunit;
  78. char *csta;
  79. } cllist;
  80. /*rewind, backspace, endfile*/
  81. typedef struct
  82. { flag aerr;
  83. ftnint aunit;
  84. } alist;
  85. /* inquire */
  86. typedef struct
  87. { flag inerr;
  88. ftnint inunit;
  89. char *infile;
  90. ftnlen infilen;
  91. ftnint *inex; /*parameters in standard's order*/
  92. ftnint *inopen;
  93. ftnint *innum;
  94. ftnint *innamed;
  95. char *inname;
  96. ftnlen innamlen;
  97. char *inacc;
  98. ftnlen inacclen;
  99. char *inseq;
  100. ftnlen inseqlen;
  101. char *indir;
  102. ftnlen indirlen;
  103. char *infmt;
  104. ftnlen infmtlen;
  105. char *inform;
  106. ftnint informlen;
  107. char *inunf;
  108. ftnlen inunflen;
  109. ftnint *inrecl;
  110. ftnint *innrec;
  111. char *inblank;
  112. ftnlen inblanklen;
  113. } inlist;
  114. #define VOID void
  115. union Multitype { /* for multiple entry points */
  116. integer1 g;
  117. shortint h;
  118. integer i;
  119. /* longint j; */
  120. real r;
  121. doublereal d;
  122. complex c;
  123. doublecomplex z;
  124. };
  125. typedef union Multitype Multitype;
  126. struct Vardesc { /* for Namelist */
  127. char *name;
  128. char *addr;
  129. ftnlen *dims;
  130. int type;
  131. };
  132. typedef struct Vardesc Vardesc;
  133. struct Namelist {
  134. char *name;
  135. Vardesc **vars;
  136. int nvars;
  137. };
  138. typedef struct Namelist Namelist;
  139. #define abs(x) ((x) >= 0 ? (x) : -(x))
  140. #define dabs(x) (fabs(x))
  141. #define f2cmin(a,b) ((a) <= (b) ? (a) : (b))
  142. #define f2cmax(a,b) ((a) >= (b) ? (a) : (b))
  143. #define dmin(a,b) (f2cmin(a,b))
  144. #define dmax(a,b) (f2cmax(a,b))
  145. #define bit_test(a,b) ((a) >> (b) & 1)
  146. #define bit_clear(a,b) ((a) & ~((uinteger)1 << (b)))
  147. #define bit_set(a,b) ((a) | ((uinteger)1 << (b)))
  148. #define abort_() { sig_die("Fortran abort routine called", 1); }
  149. #define c_abs(z) (cabsf(Cf(z)))
  150. #define c_cos(R,Z) { pCf(R)=ccos(Cf(Z)); }
  151. #define c_div(c, a, b) {pCf(c) = Cf(a)/Cf(b);}
  152. #define z_div(c, a, b) {pCd(c) = Cd(a)/Cd(b);}
  153. #define c_exp(R, Z) {pCf(R) = cexpf(Cf(Z));}
  154. #define c_log(R, Z) {pCf(R) = clogf(Cf(Z));}
  155. #define c_sin(R, Z) {pCf(R) = csinf(Cf(Z));}
  156. //#define c_sqrt(R, Z) {*(R) = csqrtf(Cf(Z));}
  157. #define c_sqrt(R, Z) {pCf(R) = csqrtf(Cf(Z));}
  158. #define d_abs(x) (fabs(*(x)))
  159. #define d_acos(x) (acos(*(x)))
  160. #define d_asin(x) (asin(*(x)))
  161. #define d_atan(x) (atan(*(x)))
  162. #define d_atn2(x, y) (atan2(*(x),*(y)))
  163. #define d_cnjg(R, Z) { pCd(R) = conj(Cd(Z)); }
  164. #define r_cnjg(R, Z) { pCf(R) = conj(Cf(Z)); }
  165. #define d_cos(x) (cos(*(x)))
  166. #define d_cosh(x) (cosh(*(x)))
  167. #define d_dim(__a, __b) ( *(__a) > *(__b) ? *(__a) - *(__b) : 0.0 )
  168. #define d_exp(x) (exp(*(x)))
  169. #define d_imag(z) (cimag(Cd(z)))
  170. #define r_imag(z) (cimag(Cf(z)))
  171. #define d_int(__x) (*(__x)>0 ? floor(*(__x)) : -floor(- *(__x)))
  172. #define r_int(__x) (*(__x)>0 ? floor(*(__x)) : -floor(- *(__x)))
  173. #define d_lg10(x) ( 0.43429448190325182765 * log(*(x)) )
  174. #define r_lg10(x) ( 0.43429448190325182765 * log(*(x)) )
  175. #define d_log(x) (log(*(x)))
  176. #define d_mod(x, y) (fmod(*(x), *(y)))
  177. #define u_nint(__x) ((__x)>=0 ? floor((__x) + .5) : -floor(.5 - (__x)))
  178. #define d_nint(x) u_nint(*(x))
  179. #define u_sign(__a,__b) ((__b) >= 0 ? ((__a) >= 0 ? (__a) : -(__a)) : -((__a) >= 0 ? (__a) : -(__a)))
  180. #define d_sign(a,b) u_sign(*(a),*(b))
  181. #define r_sign(a,b) u_sign(*(a),*(b))
  182. #define d_sin(x) (sin(*(x)))
  183. #define d_sinh(x) (sinh(*(x)))
  184. #define d_sqrt(x) (sqrt(*(x)))
  185. #define d_tan(x) (tan(*(x)))
  186. #define d_tanh(x) (tanh(*(x)))
  187. #define i_abs(x) abs(*(x))
  188. #define i_dnnt(x) ((integer)u_nint(*(x)))
  189. #define i_len(s, n) (n)
  190. #define i_nint(x) ((integer)u_nint(*(x)))
  191. #define i_sign(a,b) ((integer)u_sign((integer)*(a),(integer)*(b)))
  192. #define pow_dd(ap, bp) ( pow(*(ap), *(bp)))
  193. #define pow_si(B,E) spow_ui(*(B),*(E))
  194. #define pow_ri(B,E) spow_ui(*(B),*(E))
  195. #define pow_di(B,E) dpow_ui(*(B),*(E))
  196. #define pow_zi(p, a, b) {pCd(p) = zpow_ui(Cd(a), *(b));}
  197. #define pow_ci(p, a, b) {pCf(p) = cpow_ui(Cf(a), *(b));}
  198. #define pow_zz(R,A,B) {pCd(R) = cpow(Cd(A),*(B));}
  199. #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++ = ' '; }
  200. #define s_cmp(a,b,c,d) ((integer)strncmp((a),(b),f2cmin((c),(d))))
  201. #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]; }
  202. #define sig_die(s, kill) { exit(1); }
  203. #define s_stop(s, n) {exit(0);}
  204. static char junk[] = "\n@(#)LIBF77 VERSION 19990503\n";
  205. #define z_abs(z) (cabs(Cd(z)))
  206. #define z_exp(R, Z) {pCd(R) = cexp(Cd(Z));}
  207. #define z_sqrt(R, Z) {pCd(R) = csqrt(Cd(Z));}
  208. #define myexit_() break;
  209. #define mycycle() continue;
  210. #define myceiling(w) {ceil(w)}
  211. #define myhuge(w) {HUGE_VAL}
  212. //#define mymaxloc_(w,s,e,n) {if (sizeof(*(w)) == sizeof(double)) dmaxloc_((w),*(s),*(e),n); else dmaxloc_((w),*(s),*(e),n);}
  213. #define mymaxloc(w,s,e,n) {dmaxloc_(w,*(s),*(e),n)}
  214. /* procedure parameter types for -A and -C++ */
  215. #define F2C_proc_par_types 1
  216. #ifdef __cplusplus
  217. typedef logical (*L_fp)(...);
  218. #else
  219. typedef logical (*L_fp)();
  220. #endif
  221. static float spow_ui(float x, integer n) {
  222. float pow=1.0; unsigned long int u;
  223. if(n != 0) {
  224. if(n < 0) n = -n, x = 1/x;
  225. for(u = n; ; ) {
  226. if(u & 01) pow *= x;
  227. if(u >>= 1) x *= x;
  228. else break;
  229. }
  230. }
  231. return pow;
  232. }
  233. static double dpow_ui(double x, integer n) {
  234. double pow=1.0; unsigned long int u;
  235. if(n != 0) {
  236. if(n < 0) n = -n, x = 1/x;
  237. for(u = n; ; ) {
  238. if(u & 01) pow *= x;
  239. if(u >>= 1) x *= x;
  240. else break;
  241. }
  242. }
  243. return pow;
  244. }
  245. static _Complex float cpow_ui(_Complex float x, integer n) {
  246. _Complex float pow=1.0; unsigned long int u;
  247. if(n != 0) {
  248. if(n < 0) n = -n, x = 1/x;
  249. for(u = n; ; ) {
  250. if(u & 01) pow *= x;
  251. if(u >>= 1) x *= x;
  252. else break;
  253. }
  254. }
  255. return pow;
  256. }
  257. static _Complex double zpow_ui(_Complex double x, integer n) {
  258. _Complex double pow=1.0; unsigned long int u;
  259. if(n != 0) {
  260. if(n < 0) n = -n, x = 1/x;
  261. for(u = n; ; ) {
  262. if(u & 01) pow *= x;
  263. if(u >>= 1) x *= x;
  264. else break;
  265. }
  266. }
  267. return pow;
  268. }
  269. static integer pow_ii(integer x, integer n) {
  270. integer pow; unsigned long int u;
  271. if (n <= 0) {
  272. if (n == 0 || x == 1) pow = 1;
  273. else if (x != -1) pow = x == 0 ? 1/x : 0;
  274. else n = -n;
  275. }
  276. if ((n > 0) || !(n == 0 || x == 1 || x != -1)) {
  277. u = n;
  278. for(pow = 1; ; ) {
  279. if(u & 01) pow *= x;
  280. if(u >>= 1) x *= x;
  281. else break;
  282. }
  283. }
  284. return pow;
  285. }
  286. static integer dmaxloc_(double *w, integer s, integer e, integer *n)
  287. {
  288. double m; integer i, mi;
  289. for(m=w[s-1], mi=s, i=s+1; i<=e; i++)
  290. if (w[i-1]>m) mi=i ,m=w[i-1];
  291. return mi-s+1;
  292. }
  293. static integer smaxloc_(float *w, integer s, integer e, integer *n)
  294. {
  295. float m; integer i, mi;
  296. for(m=w[s-1], mi=s, i=s+1; i<=e; i++)
  297. if (w[i-1]>m) mi=i ,m=w[i-1];
  298. return mi-s+1;
  299. }
  300. static inline void cdotc_(complex *z, integer *n_, complex *x, integer *incx_, complex *y, integer *incy_) {
  301. integer n = *n_, incx = *incx_, incy = *incy_, i;
  302. _Complex float zdotc = 0.0;
  303. if (incx == 1 && incy == 1) {
  304. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  305. zdotc += conjf(Cf(&x[i])) * Cf(&y[i]);
  306. }
  307. } else {
  308. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  309. zdotc += conjf(Cf(&x[i*incx])) * Cf(&y[i*incy]);
  310. }
  311. }
  312. pCf(z) = zdotc;
  313. }
  314. static inline void zdotc_(doublecomplex *z, integer *n_, doublecomplex *x, integer *incx_, doublecomplex *y, integer *incy_) {
  315. integer n = *n_, incx = *incx_, incy = *incy_, i;
  316. _Complex double zdotc = 0.0;
  317. if (incx == 1 && incy == 1) {
  318. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  319. zdotc += conj(Cd(&x[i])) * Cd(&y[i]);
  320. }
  321. } else {
  322. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  323. zdotc += conj(Cd(&x[i*incx])) * Cd(&y[i*incy]);
  324. }
  325. }
  326. pCd(z) = zdotc;
  327. }
  328. static inline void cdotu_(complex *z, integer *n_, complex *x, integer *incx_, complex *y, integer *incy_) {
  329. integer n = *n_, incx = *incx_, incy = *incy_, i;
  330. _Complex float zdotc = 0.0;
  331. if (incx == 1 && incy == 1) {
  332. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  333. zdotc += Cf(&x[i]) * Cf(&y[i]);
  334. }
  335. } else {
  336. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  337. zdotc += Cf(&x[i*incx]) * Cf(&y[i*incy]);
  338. }
  339. }
  340. pCf(z) = zdotc;
  341. }
  342. static inline void zdotu_(doublecomplex *z, integer *n_, doublecomplex *x, integer *incx_, doublecomplex *y, integer *incy_) {
  343. integer n = *n_, incx = *incx_, incy = *incy_, i;
  344. _Complex double zdotc = 0.0;
  345. if (incx == 1 && incy == 1) {
  346. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  347. zdotc += Cd(&x[i]) * Cd(&y[i]);
  348. }
  349. } else {
  350. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  351. zdotc += Cd(&x[i*incx]) * Cd(&y[i*incy]);
  352. }
  353. }
  354. pCd(z) = zdotc;
  355. }
  356. #endif
  357. /* -- translated by f2c (version 20000121).
  358. You must link the resulting object file with the libraries:
  359. -lf2c -lm (in that order)
  360. */
  361. /* Table of constant values */
  362. static integer c__1 = 1;
  363. /* > \brief \b SLARRF finds a new relatively robust representation such that at least one of the eigenvalues i
  364. s relatively isolated. */
  365. /* =========== DOCUMENTATION =========== */
  366. /* Online html documentation available at */
  367. /* http://www.netlib.org/lapack/explore-html/ */
  368. /* > \htmlonly */
  369. /* > Download SLARRF + dependencies */
  370. /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/slarrf.
  371. f"> */
  372. /* > [TGZ]</a> */
  373. /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/slarrf.
  374. f"> */
  375. /* > [ZIP]</a> */
  376. /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/slarrf.
  377. f"> */
  378. /* > [TXT]</a> */
  379. /* > \endhtmlonly */
  380. /* Definition: */
  381. /* =========== */
  382. /* SUBROUTINE SLARRF( N, D, L, LD, CLSTRT, CLEND, */
  383. /* W, WGAP, WERR, */
  384. /* SPDIAM, CLGAPL, CLGAPR, PIVMIN, SIGMA, */
  385. /* DPLUS, LPLUS, WORK, INFO ) */
  386. /* INTEGER CLSTRT, CLEND, INFO, N */
  387. /* REAL CLGAPL, CLGAPR, PIVMIN, SIGMA, SPDIAM */
  388. /* REAL D( * ), DPLUS( * ), L( * ), LD( * ), */
  389. /* $ LPLUS( * ), W( * ), WGAP( * ), WERR( * ), WORK( * ) */
  390. /* > \par Purpose: */
  391. /* ============= */
  392. /* > */
  393. /* > \verbatim */
  394. /* > */
  395. /* > Given the initial representation L D L^T and its cluster of close */
  396. /* > eigenvalues (in a relative measure), W( CLSTRT ), W( CLSTRT+1 ), ... */
  397. /* > W( CLEND ), SLARRF finds a new relatively robust representation */
  398. /* > L D L^T - SIGMA I = L(+) D(+) L(+)^T such that at least one of the */
  399. /* > eigenvalues of L(+) D(+) L(+)^T is relatively isolated. */
  400. /* > \endverbatim */
  401. /* Arguments: */
  402. /* ========== */
  403. /* > \param[in] N */
  404. /* > \verbatim */
  405. /* > N is INTEGER */
  406. /* > The order of the matrix (subblock, if the matrix split). */
  407. /* > \endverbatim */
  408. /* > */
  409. /* > \param[in] D */
  410. /* > \verbatim */
  411. /* > D is REAL array, dimension (N) */
  412. /* > The N diagonal elements of the diagonal matrix D. */
  413. /* > \endverbatim */
  414. /* > */
  415. /* > \param[in] L */
  416. /* > \verbatim */
  417. /* > L is REAL array, dimension (N-1) */
  418. /* > The (N-1) subdiagonal elements of the unit bidiagonal */
  419. /* > matrix L. */
  420. /* > \endverbatim */
  421. /* > */
  422. /* > \param[in] LD */
  423. /* > \verbatim */
  424. /* > LD is REAL array, dimension (N-1) */
  425. /* > The (N-1) elements L(i)*D(i). */
  426. /* > \endverbatim */
  427. /* > */
  428. /* > \param[in] CLSTRT */
  429. /* > \verbatim */
  430. /* > CLSTRT is INTEGER */
  431. /* > The index of the first eigenvalue in the cluster. */
  432. /* > \endverbatim */
  433. /* > */
  434. /* > \param[in] CLEND */
  435. /* > \verbatim */
  436. /* > CLEND is INTEGER */
  437. /* > The index of the last eigenvalue in the cluster. */
  438. /* > \endverbatim */
  439. /* > */
  440. /* > \param[in] W */
  441. /* > \verbatim */
  442. /* > W is REAL array, dimension */
  443. /* > dimension is >= (CLEND-CLSTRT+1) */
  444. /* > The eigenvalue APPROXIMATIONS of L D L^T in ascending order. */
  445. /* > W( CLSTRT ) through W( CLEND ) form the cluster of relatively */
  446. /* > close eigenalues. */
  447. /* > \endverbatim */
  448. /* > */
  449. /* > \param[in,out] WGAP */
  450. /* > \verbatim */
  451. /* > WGAP is REAL array, dimension */
  452. /* > dimension is >= (CLEND-CLSTRT+1) */
  453. /* > The separation from the right neighbor eigenvalue in W. */
  454. /* > \endverbatim */
  455. /* > */
  456. /* > \param[in] WERR */
  457. /* > \verbatim */
  458. /* > WERR is REAL array, dimension */
  459. /* > dimension is >= (CLEND-CLSTRT+1) */
  460. /* > WERR contain the semiwidth of the uncertainty */
  461. /* > interval of the corresponding eigenvalue APPROXIMATION in W */
  462. /* > \endverbatim */
  463. /* > */
  464. /* > \param[in] SPDIAM */
  465. /* > \verbatim */
  466. /* > SPDIAM is REAL */
  467. /* > estimate of the spectral diameter obtained from the */
  468. /* > Gerschgorin intervals */
  469. /* > \endverbatim */
  470. /* > */
  471. /* > \param[in] CLGAPL */
  472. /* > \verbatim */
  473. /* > CLGAPL is REAL */
  474. /* > \endverbatim */
  475. /* > */
  476. /* > \param[in] CLGAPR */
  477. /* > \verbatim */
  478. /* > CLGAPR is REAL */
  479. /* > absolute gap on each end of the cluster. */
  480. /* > Set by the calling routine to protect against shifts too close */
  481. /* > to eigenvalues outside the cluster. */
  482. /* > \endverbatim */
  483. /* > */
  484. /* > \param[in] PIVMIN */
  485. /* > \verbatim */
  486. /* > PIVMIN is REAL */
  487. /* > The minimum pivot allowed in the Sturm sequence. */
  488. /* > \endverbatim */
  489. /* > */
  490. /* > \param[out] SIGMA */
  491. /* > \verbatim */
  492. /* > SIGMA is REAL */
  493. /* > The shift used to form L(+) D(+) L(+)^T. */
  494. /* > \endverbatim */
  495. /* > */
  496. /* > \param[out] DPLUS */
  497. /* > \verbatim */
  498. /* > DPLUS is REAL array, dimension (N) */
  499. /* > The N diagonal elements of the diagonal matrix D(+). */
  500. /* > \endverbatim */
  501. /* > */
  502. /* > \param[out] LPLUS */
  503. /* > \verbatim */
  504. /* > LPLUS is REAL array, dimension (N-1) */
  505. /* > The first (N-1) elements of LPLUS contain the subdiagonal */
  506. /* > elements of the unit bidiagonal matrix L(+). */
  507. /* > \endverbatim */
  508. /* > */
  509. /* > \param[out] WORK */
  510. /* > \verbatim */
  511. /* > WORK is REAL array, dimension (2*N) */
  512. /* > Workspace. */
  513. /* > \endverbatim */
  514. /* > */
  515. /* > \param[out] INFO */
  516. /* > \verbatim */
  517. /* > INFO is INTEGER */
  518. /* > Signals processing OK (=0) or failure (=1) */
  519. /* > \endverbatim */
  520. /* Authors: */
  521. /* ======== */
  522. /* > \author Univ. of Tennessee */
  523. /* > \author Univ. of California Berkeley */
  524. /* > \author Univ. of Colorado Denver */
  525. /* > \author NAG Ltd. */
  526. /* > \date June 2016 */
  527. /* > \ingroup OTHERauxiliary */
  528. /* > \par Contributors: */
  529. /* ================== */
  530. /* > */
  531. /* > Beresford Parlett, University of California, Berkeley, USA \n */
  532. /* > Jim Demmel, University of California, Berkeley, USA \n */
  533. /* > Inderjit Dhillon, University of Texas, Austin, USA \n */
  534. /* > Osni Marques, LBNL/NERSC, USA \n */
  535. /* > Christof Voemel, University of California, Berkeley, USA */
  536. /* ===================================================================== */
  537. /* Subroutine */ int slarrf_(integer *n, real *d__, real *l, real *ld,
  538. integer *clstrt, integer *clend, real *w, real *wgap, real *werr,
  539. real *spdiam, real *clgapl, real *clgapr, real *pivmin, real *sigma,
  540. real *dplus, real *lplus, real *work, integer *info)
  541. {
  542. /* System generated locals */
  543. integer i__1;
  544. real r__1, r__2, r__3;
  545. /* Local variables */
  546. real growthbound, fail, fact, oldp;
  547. integer indx;
  548. real prod;
  549. integer ktry;
  550. real fail2;
  551. integer i__;
  552. real s, avgap, ldmax, rdmax;
  553. integer shift;
  554. extern /* Subroutine */ int scopy_(integer *, real *, integer *, real *,
  555. integer *);
  556. real bestshift, smlgrowth;
  557. logical dorrr1;
  558. real ldelta;
  559. extern real slamch_(char *);
  560. logical nofail;
  561. real mingap, lsigma, rdelta;
  562. logical forcer;
  563. real rsigma, clwdth;
  564. extern logical sisnan_(real *);
  565. logical sawnan1, sawnan2;
  566. real eps, tmp;
  567. logical tryrrr1;
  568. real max1, max2, rrr1, rrr2, znm2;
  569. /* -- LAPACK auxiliary routine (version 3.7.1) -- */
  570. /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
  571. /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
  572. /* June 2016 */
  573. /* ===================================================================== */
  574. /* Parameter adjustments */
  575. --work;
  576. --lplus;
  577. --dplus;
  578. --werr;
  579. --wgap;
  580. --w;
  581. --ld;
  582. --l;
  583. --d__;
  584. /* Function Body */
  585. *info = 0;
  586. /* Quick return if possible */
  587. if (*n <= 0) {
  588. return 0;
  589. }
  590. fact = 2.f;
  591. eps = slamch_("Precision");
  592. shift = 0;
  593. forcer = FALSE_;
  594. /* Note that we cannot guarantee that for any of the shifts tried, */
  595. /* the factorization has a small or even moderate element growth. */
  596. /* There could be Ritz values at both ends of the cluster and despite */
  597. /* backing off, there are examples where all factorizations tried */
  598. /* (in IEEE mode, allowing zero pivots & infinities) have INFINITE */
  599. /* element growth. */
  600. /* For this reason, we should use PIVMIN in this subroutine so that at */
  601. /* least the L D L^T factorization exists. It can be checked afterwards */
  602. /* whether the element growth caused bad residuals/orthogonality. */
  603. /* Decide whether the code should accept the best among all */
  604. /* representations despite large element growth or signal INFO=1 */
  605. /* Setting NOFAIL to .FALSE. for quick fix for bug 113 */
  606. nofail = FALSE_;
  607. /* Compute the average gap length of the cluster */
  608. clwdth = (r__1 = w[*clend] - w[*clstrt], abs(r__1)) + werr[*clend] + werr[
  609. *clstrt];
  610. avgap = clwdth / (real) (*clend - *clstrt);
  611. mingap = f2cmin(*clgapl,*clgapr);
  612. /* Initial values for shifts to both ends of cluster */
  613. /* Computing MIN */
  614. r__1 = w[*clstrt], r__2 = w[*clend];
  615. lsigma = f2cmin(r__1,r__2) - werr[*clstrt];
  616. /* Computing MAX */
  617. r__1 = w[*clstrt], r__2 = w[*clend];
  618. rsigma = f2cmax(r__1,r__2) + werr[*clend];
  619. /* Use a small fudge to make sure that we really shift to the outside */
  620. lsigma -= abs(lsigma) * 2.f * eps;
  621. rsigma += abs(rsigma) * 2.f * eps;
  622. /* Compute upper bounds for how much to back off the initial shifts */
  623. ldmax = mingap * .25f + *pivmin * 2.f;
  624. rdmax = mingap * .25f + *pivmin * 2.f;
  625. /* Computing MAX */
  626. r__1 = avgap, r__2 = wgap[*clstrt];
  627. ldelta = f2cmax(r__1,r__2) / fact;
  628. /* Computing MAX */
  629. r__1 = avgap, r__2 = wgap[*clend - 1];
  630. rdelta = f2cmax(r__1,r__2) / fact;
  631. /* Initialize the record of the best representation found */
  632. s = slamch_("S");
  633. smlgrowth = 1.f / s;
  634. fail = (real) (*n - 1) * mingap / (*spdiam * eps);
  635. fail2 = (real) (*n - 1) * mingap / (*spdiam * sqrt(eps));
  636. bestshift = lsigma;
  637. /* while (KTRY <= KTRYMAX) */
  638. ktry = 0;
  639. growthbound = *spdiam * 8.f;
  640. L5:
  641. sawnan1 = FALSE_;
  642. sawnan2 = FALSE_;
  643. /* Ensure that we do not back off too much of the initial shifts */
  644. ldelta = f2cmin(ldmax,ldelta);
  645. rdelta = f2cmin(rdmax,rdelta);
  646. /* Compute the element growth when shifting to both ends of the cluster */
  647. /* accept the shift if there is no element growth at one of the two ends */
  648. /* Left end */
  649. s = -lsigma;
  650. dplus[1] = d__[1] + s;
  651. if (abs(dplus[1]) < *pivmin) {
  652. dplus[1] = -(*pivmin);
  653. /* Need to set SAWNAN1 because refined RRR test should not be used */
  654. /* in this case */
  655. sawnan1 = TRUE_;
  656. }
  657. max1 = abs(dplus[1]);
  658. i__1 = *n - 1;
  659. for (i__ = 1; i__ <= i__1; ++i__) {
  660. lplus[i__] = ld[i__] / dplus[i__];
  661. s = s * lplus[i__] * l[i__] - lsigma;
  662. dplus[i__ + 1] = d__[i__ + 1] + s;
  663. if ((r__1 = dplus[i__ + 1], abs(r__1)) < *pivmin) {
  664. dplus[i__ + 1] = -(*pivmin);
  665. /* Need to set SAWNAN1 because refined RRR test should not be used */
  666. /* in this case */
  667. sawnan1 = TRUE_;
  668. }
  669. /* Computing MAX */
  670. r__2 = max1, r__3 = (r__1 = dplus[i__ + 1], abs(r__1));
  671. max1 = f2cmax(r__2,r__3);
  672. /* L6: */
  673. }
  674. sawnan1 = sawnan1 || sisnan_(&max1);
  675. if (forcer || max1 <= growthbound && ! sawnan1) {
  676. *sigma = lsigma;
  677. shift = 1;
  678. goto L100;
  679. }
  680. /* Right end */
  681. s = -rsigma;
  682. work[1] = d__[1] + s;
  683. if (abs(work[1]) < *pivmin) {
  684. work[1] = -(*pivmin);
  685. /* Need to set SAWNAN2 because refined RRR test should not be used */
  686. /* in this case */
  687. sawnan2 = TRUE_;
  688. }
  689. max2 = abs(work[1]);
  690. i__1 = *n - 1;
  691. for (i__ = 1; i__ <= i__1; ++i__) {
  692. work[*n + i__] = ld[i__] / work[i__];
  693. s = s * work[*n + i__] * l[i__] - rsigma;
  694. work[i__ + 1] = d__[i__ + 1] + s;
  695. if ((r__1 = work[i__ + 1], abs(r__1)) < *pivmin) {
  696. work[i__ + 1] = -(*pivmin);
  697. /* Need to set SAWNAN2 because refined RRR test should not be used */
  698. /* in this case */
  699. sawnan2 = TRUE_;
  700. }
  701. /* Computing MAX */
  702. r__2 = max2, r__3 = (r__1 = work[i__ + 1], abs(r__1));
  703. max2 = f2cmax(r__2,r__3);
  704. /* L7: */
  705. }
  706. sawnan2 = sawnan2 || sisnan_(&max2);
  707. if (forcer || max2 <= growthbound && ! sawnan2) {
  708. *sigma = rsigma;
  709. shift = 2;
  710. goto L100;
  711. }
  712. /* If we are at this point, both shifts led to too much element growth */
  713. /* Record the better of the two shifts (provided it didn't lead to NaN) */
  714. if (sawnan1 && sawnan2) {
  715. /* both MAX1 and MAX2 are NaN */
  716. goto L50;
  717. } else {
  718. if (! sawnan1) {
  719. indx = 1;
  720. if (max1 <= smlgrowth) {
  721. smlgrowth = max1;
  722. bestshift = lsigma;
  723. }
  724. }
  725. if (! sawnan2) {
  726. if (sawnan1 || max2 <= max1) {
  727. indx = 2;
  728. }
  729. if (max2 <= smlgrowth) {
  730. smlgrowth = max2;
  731. bestshift = rsigma;
  732. }
  733. }
  734. }
  735. /* If we are here, both the left and the right shift led to */
  736. /* element growth. If the element growth is moderate, then */
  737. /* we may still accept the representation, if it passes a */
  738. /* refined test for RRR. This test supposes that no NaN occurred. */
  739. /* Moreover, we use the refined RRR test only for isolated clusters. */
  740. if (clwdth < mingap / 128.f && f2cmin(max1,max2) < fail2 && ! sawnan1 && !
  741. sawnan2) {
  742. dorrr1 = TRUE_;
  743. } else {
  744. dorrr1 = FALSE_;
  745. }
  746. tryrrr1 = TRUE_;
  747. if (tryrrr1 && dorrr1) {
  748. if (indx == 1) {
  749. tmp = (r__1 = dplus[*n], abs(r__1));
  750. znm2 = 1.f;
  751. prod = 1.f;
  752. oldp = 1.f;
  753. for (i__ = *n - 1; i__ >= 1; --i__) {
  754. if (prod <= eps) {
  755. prod = dplus[i__ + 1] * work[*n + i__ + 1] / (dplus[i__] *
  756. work[*n + i__]) * oldp;
  757. } else {
  758. prod *= (r__1 = work[*n + i__], abs(r__1));
  759. }
  760. oldp = prod;
  761. /* Computing 2nd power */
  762. r__1 = prod;
  763. znm2 += r__1 * r__1;
  764. /* Computing MAX */
  765. r__2 = tmp, r__3 = (r__1 = dplus[i__] * prod, abs(r__1));
  766. tmp = f2cmax(r__2,r__3);
  767. /* L15: */
  768. }
  769. rrr1 = tmp / (*spdiam * sqrt(znm2));
  770. if (rrr1 <= 8.f) {
  771. *sigma = lsigma;
  772. shift = 1;
  773. goto L100;
  774. }
  775. } else if (indx == 2) {
  776. tmp = (r__1 = work[*n], abs(r__1));
  777. znm2 = 1.f;
  778. prod = 1.f;
  779. oldp = 1.f;
  780. for (i__ = *n - 1; i__ >= 1; --i__) {
  781. if (prod <= eps) {
  782. prod = work[i__ + 1] * lplus[i__ + 1] / (work[i__] *
  783. lplus[i__]) * oldp;
  784. } else {
  785. prod *= (r__1 = lplus[i__], abs(r__1));
  786. }
  787. oldp = prod;
  788. /* Computing 2nd power */
  789. r__1 = prod;
  790. znm2 += r__1 * r__1;
  791. /* Computing MAX */
  792. r__2 = tmp, r__3 = (r__1 = work[i__] * prod, abs(r__1));
  793. tmp = f2cmax(r__2,r__3);
  794. /* L16: */
  795. }
  796. rrr2 = tmp / (*spdiam * sqrt(znm2));
  797. if (rrr2 <= 8.f) {
  798. *sigma = rsigma;
  799. shift = 2;
  800. goto L100;
  801. }
  802. }
  803. }
  804. L50:
  805. if (ktry < 1) {
  806. /* If we are here, both shifts failed also the RRR test. */
  807. /* Back off to the outside */
  808. /* Computing MAX */
  809. r__1 = lsigma - ldelta, r__2 = lsigma - ldmax;
  810. lsigma = f2cmax(r__1,r__2);
  811. /* Computing MIN */
  812. r__1 = rsigma + rdelta, r__2 = rsigma + rdmax;
  813. rsigma = f2cmin(r__1,r__2);
  814. ldelta *= 2.f;
  815. rdelta *= 2.f;
  816. ++ktry;
  817. goto L5;
  818. } else {
  819. /* None of the representations investigated satisfied our */
  820. /* criteria. Take the best one we found. */
  821. if (smlgrowth < fail || nofail) {
  822. lsigma = bestshift;
  823. rsigma = bestshift;
  824. forcer = TRUE_;
  825. goto L5;
  826. } else {
  827. *info = 1;
  828. return 0;
  829. }
  830. }
  831. L100:
  832. if (shift == 1) {
  833. } else if (shift == 2) {
  834. /* store new L and D back into DPLUS, LPLUS */
  835. scopy_(n, &work[1], &c__1, &dplus[1], &c__1);
  836. i__1 = *n - 1;
  837. scopy_(&i__1, &work[*n + 1], &c__1, &lplus[1], &c__1);
  838. }
  839. return 0;
  840. /* End of SLARRF */
  841. } /* slarrf_ */