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

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926
  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. #if defined(_WIN64)
  18. typedef long long BLASLONG;
  19. typedef unsigned long long BLASULONG;
  20. #else
  21. typedef long BLASLONG;
  22. typedef unsigned long BLASULONG;
  23. #endif
  24. #ifdef LAPACK_ILP64
  25. typedef BLASLONG blasint;
  26. #if defined(_WIN64)
  27. #define blasabs(x) llabs(x)
  28. #else
  29. #define blasabs(x) labs(x)
  30. #endif
  31. #else
  32. typedef int blasint;
  33. #define blasabs(x) abs(x)
  34. #endif
  35. typedef blasint integer;
  36. typedef unsigned int uinteger;
  37. typedef char *address;
  38. typedef short int shortint;
  39. typedef float real;
  40. typedef double doublereal;
  41. typedef struct { real r, i; } complex;
  42. typedef struct { doublereal r, i; } doublecomplex;
  43. static inline _Complex float Cf(complex *z) {return z->r + z->i*_Complex_I;}
  44. static inline _Complex double Cd(doublecomplex *z) {return z->r + z->i*_Complex_I;}
  45. static inline _Complex float * _pCf(complex *z) {return (_Complex float*)z;}
  46. static inline _Complex double * _pCd(doublecomplex *z) {return (_Complex double*)z;}
  47. #define pCf(z) (*_pCf(z))
  48. #define pCd(z) (*_pCd(z))
  49. typedef int logical;
  50. typedef short int shortlogical;
  51. typedef char logical1;
  52. typedef char integer1;
  53. #define TRUE_ (1)
  54. #define FALSE_ (0)
  55. /* Extern is for use with -E */
  56. #ifndef Extern
  57. #define Extern extern
  58. #endif
  59. /* I/O stuff */
  60. typedef int flag;
  61. typedef int ftnlen;
  62. typedef int ftnint;
  63. /*external read, write*/
  64. typedef struct
  65. { flag cierr;
  66. ftnint ciunit;
  67. flag ciend;
  68. char *cifmt;
  69. ftnint cirec;
  70. } cilist;
  71. /*internal read, write*/
  72. typedef struct
  73. { flag icierr;
  74. char *iciunit;
  75. flag iciend;
  76. char *icifmt;
  77. ftnint icirlen;
  78. ftnint icirnum;
  79. } icilist;
  80. /*open*/
  81. typedef struct
  82. { flag oerr;
  83. ftnint ounit;
  84. char *ofnm;
  85. ftnlen ofnmlen;
  86. char *osta;
  87. char *oacc;
  88. char *ofm;
  89. ftnint orl;
  90. char *oblnk;
  91. } olist;
  92. /*close*/
  93. typedef struct
  94. { flag cerr;
  95. ftnint cunit;
  96. char *csta;
  97. } cllist;
  98. /*rewind, backspace, endfile*/
  99. typedef struct
  100. { flag aerr;
  101. ftnint aunit;
  102. } alist;
  103. /* inquire */
  104. typedef struct
  105. { flag inerr;
  106. ftnint inunit;
  107. char *infile;
  108. ftnlen infilen;
  109. ftnint *inex; /*parameters in standard's order*/
  110. ftnint *inopen;
  111. ftnint *innum;
  112. ftnint *innamed;
  113. char *inname;
  114. ftnlen innamlen;
  115. char *inacc;
  116. ftnlen inacclen;
  117. char *inseq;
  118. ftnlen inseqlen;
  119. char *indir;
  120. ftnlen indirlen;
  121. char *infmt;
  122. ftnlen infmtlen;
  123. char *inform;
  124. ftnint informlen;
  125. char *inunf;
  126. ftnlen inunflen;
  127. ftnint *inrecl;
  128. ftnint *innrec;
  129. char *inblank;
  130. ftnlen inblanklen;
  131. } inlist;
  132. #define VOID void
  133. union Multitype { /* for multiple entry points */
  134. integer1 g;
  135. shortint h;
  136. integer i;
  137. /* longint j; */
  138. real r;
  139. doublereal d;
  140. complex c;
  141. doublecomplex z;
  142. };
  143. typedef union Multitype Multitype;
  144. struct Vardesc { /* for Namelist */
  145. char *name;
  146. char *addr;
  147. ftnlen *dims;
  148. int type;
  149. };
  150. typedef struct Vardesc Vardesc;
  151. struct Namelist {
  152. char *name;
  153. Vardesc **vars;
  154. int nvars;
  155. };
  156. typedef struct Namelist Namelist;
  157. #define abs(x) ((x) >= 0 ? (x) : -(x))
  158. #define dabs(x) (fabs(x))
  159. #define f2cmin(a,b) ((a) <= (b) ? (a) : (b))
  160. #define f2cmax(a,b) ((a) >= (b) ? (a) : (b))
  161. #define dmin(a,b) (f2cmin(a,b))
  162. #define dmax(a,b) (f2cmax(a,b))
  163. #define bit_test(a,b) ((a) >> (b) & 1)
  164. #define bit_clear(a,b) ((a) & ~((uinteger)1 << (b)))
  165. #define bit_set(a,b) ((a) | ((uinteger)1 << (b)))
  166. #define abort_() { sig_die("Fortran abort routine called", 1); }
  167. #define c_abs(z) (cabsf(Cf(z)))
  168. #define c_cos(R,Z) { pCf(R)=ccos(Cf(Z)); }
  169. #define c_div(c, a, b) {pCf(c) = Cf(a)/Cf(b);}
  170. #define z_div(c, a, b) {pCd(c) = Cd(a)/Cd(b);}
  171. #define c_exp(R, Z) {pCf(R) = cexpf(Cf(Z));}
  172. #define c_log(R, Z) {pCf(R) = clogf(Cf(Z));}
  173. #define c_sin(R, Z) {pCf(R) = csinf(Cf(Z));}
  174. //#define c_sqrt(R, Z) {*(R) = csqrtf(Cf(Z));}
  175. #define c_sqrt(R, Z) {pCf(R) = csqrtf(Cf(Z));}
  176. #define d_abs(x) (fabs(*(x)))
  177. #define d_acos(x) (acos(*(x)))
  178. #define d_asin(x) (asin(*(x)))
  179. #define d_atan(x) (atan(*(x)))
  180. #define d_atn2(x, y) (atan2(*(x),*(y)))
  181. #define d_cnjg(R, Z) { pCd(R) = conj(Cd(Z)); }
  182. #define r_cnjg(R, Z) { pCf(R) = conj(Cf(Z)); }
  183. #define d_cos(x) (cos(*(x)))
  184. #define d_cosh(x) (cosh(*(x)))
  185. #define d_dim(__a, __b) ( *(__a) > *(__b) ? *(__a) - *(__b) : 0.0 )
  186. #define d_exp(x) (exp(*(x)))
  187. #define d_imag(z) (cimag(Cd(z)))
  188. #define r_imag(z) (cimag(Cf(z)))
  189. #define d_int(__x) (*(__x)>0 ? floor(*(__x)) : -floor(- *(__x)))
  190. #define r_int(__x) (*(__x)>0 ? floor(*(__x)) : -floor(- *(__x)))
  191. #define d_lg10(x) ( 0.43429448190325182765 * log(*(x)) )
  192. #define r_lg10(x) ( 0.43429448190325182765 * log(*(x)) )
  193. #define d_log(x) (log(*(x)))
  194. #define d_mod(x, y) (fmod(*(x), *(y)))
  195. #define u_nint(__x) ((__x)>=0 ? floor((__x) + .5) : -floor(.5 - (__x)))
  196. #define d_nint(x) u_nint(*(x))
  197. #define u_sign(__a,__b) ((__b) >= 0 ? ((__a) >= 0 ? (__a) : -(__a)) : -((__a) >= 0 ? (__a) : -(__a)))
  198. #define d_sign(a,b) u_sign(*(a),*(b))
  199. #define r_sign(a,b) u_sign(*(a),*(b))
  200. #define d_sin(x) (sin(*(x)))
  201. #define d_sinh(x) (sinh(*(x)))
  202. #define d_sqrt(x) (sqrt(*(x)))
  203. #define d_tan(x) (tan(*(x)))
  204. #define d_tanh(x) (tanh(*(x)))
  205. #define i_abs(x) abs(*(x))
  206. #define i_dnnt(x) ((integer)u_nint(*(x)))
  207. #define i_len(s, n) (n)
  208. #define i_nint(x) ((integer)u_nint(*(x)))
  209. #define i_sign(a,b) ((integer)u_sign((integer)*(a),(integer)*(b)))
  210. #define pow_dd(ap, bp) ( pow(*(ap), *(bp)))
  211. #define pow_si(B,E) spow_ui(*(B),*(E))
  212. #define pow_ri(B,E) spow_ui(*(B),*(E))
  213. #define pow_di(B,E) dpow_ui(*(B),*(E))
  214. #define pow_zi(p, a, b) {pCd(p) = zpow_ui(Cd(a), *(b));}
  215. #define pow_ci(p, a, b) {pCf(p) = cpow_ui(Cf(a), *(b));}
  216. #define pow_zz(R,A,B) {pCd(R) = cpow(Cd(A),*(B));}
  217. #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++ = ' '; }
  218. #define s_cmp(a,b,c,d) ((integer)strncmp((a),(b),f2cmin((c),(d))))
  219. #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]; }
  220. #define sig_die(s, kill) { exit(1); }
  221. #define s_stop(s, n) {exit(0);}
  222. static char junk[] = "\n@(#)LIBF77 VERSION 19990503\n";
  223. #define z_abs(z) (cabs(Cd(z)))
  224. #define z_exp(R, Z) {pCd(R) = cexp(Cd(Z));}
  225. #define z_sqrt(R, Z) {pCd(R) = csqrt(Cd(Z));}
  226. #define myexit_() break;
  227. #define mycycle() continue;
  228. #define myceiling(w) {ceil(w)}
  229. #define myhuge(w) {HUGE_VAL}
  230. //#define mymaxloc_(w,s,e,n) {if (sizeof(*(w)) == sizeof(double)) dmaxloc_((w),*(s),*(e),n); else dmaxloc_((w),*(s),*(e),n);}
  231. #define mymaxloc(w,s,e,n) {dmaxloc_(w,*(s),*(e),n)}
  232. /* procedure parameter types for -A and -C++ */
  233. #define F2C_proc_par_types 1
  234. #ifdef __cplusplus
  235. typedef logical (*L_fp)(...);
  236. #else
  237. typedef logical (*L_fp)();
  238. #endif
  239. static float spow_ui(float x, integer n) {
  240. float pow=1.0; unsigned long int u;
  241. if(n != 0) {
  242. if(n < 0) n = -n, x = 1/x;
  243. for(u = n; ; ) {
  244. if(u & 01) pow *= x;
  245. if(u >>= 1) x *= x;
  246. else break;
  247. }
  248. }
  249. return pow;
  250. }
  251. static double dpow_ui(double x, integer n) {
  252. double pow=1.0; unsigned long int u;
  253. if(n != 0) {
  254. if(n < 0) n = -n, x = 1/x;
  255. for(u = n; ; ) {
  256. if(u & 01) pow *= x;
  257. if(u >>= 1) x *= x;
  258. else break;
  259. }
  260. }
  261. return pow;
  262. }
  263. static _Complex float cpow_ui(_Complex float x, integer n) {
  264. _Complex float pow=1.0; unsigned long int u;
  265. if(n != 0) {
  266. if(n < 0) n = -n, x = 1/x;
  267. for(u = n; ; ) {
  268. if(u & 01) pow *= x;
  269. if(u >>= 1) x *= x;
  270. else break;
  271. }
  272. }
  273. return pow;
  274. }
  275. static _Complex double zpow_ui(_Complex double x, integer n) {
  276. _Complex double pow=1.0; unsigned long int u;
  277. if(n != 0) {
  278. if(n < 0) n = -n, x = 1/x;
  279. for(u = n; ; ) {
  280. if(u & 01) pow *= x;
  281. if(u >>= 1) x *= x;
  282. else break;
  283. }
  284. }
  285. return pow;
  286. }
  287. static integer pow_ii(integer x, integer n) {
  288. integer pow; unsigned long int u;
  289. if (n <= 0) {
  290. if (n == 0 || x == 1) pow = 1;
  291. else if (x != -1) pow = x == 0 ? 1/x : 0;
  292. else n = -n;
  293. }
  294. if ((n > 0) || !(n == 0 || x == 1 || x != -1)) {
  295. u = n;
  296. for(pow = 1; ; ) {
  297. if(u & 01) pow *= x;
  298. if(u >>= 1) x *= x;
  299. else break;
  300. }
  301. }
  302. return pow;
  303. }
  304. static integer dmaxloc_(double *w, integer s, integer e, integer *n)
  305. {
  306. double m; integer i, mi;
  307. for(m=w[s-1], mi=s, i=s+1; i<=e; i++)
  308. if (w[i-1]>m) mi=i ,m=w[i-1];
  309. return mi-s+1;
  310. }
  311. static integer smaxloc_(float *w, integer s, integer e, integer *n)
  312. {
  313. float m; integer i, mi;
  314. for(m=w[s-1], mi=s, i=s+1; i<=e; i++)
  315. if (w[i-1]>m) mi=i ,m=w[i-1];
  316. return mi-s+1;
  317. }
  318. static inline void cdotc_(complex *z, integer *n_, complex *x, integer *incx_, complex *y, integer *incy_) {
  319. integer n = *n_, incx = *incx_, incy = *incy_, i;
  320. _Complex float zdotc = 0.0;
  321. if (incx == 1 && incy == 1) {
  322. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  323. zdotc += conjf(Cf(&x[i])) * Cf(&y[i]);
  324. }
  325. } else {
  326. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  327. zdotc += conjf(Cf(&x[i*incx])) * Cf(&y[i*incy]);
  328. }
  329. }
  330. pCf(z) = zdotc;
  331. }
  332. static inline void zdotc_(doublecomplex *z, integer *n_, doublecomplex *x, integer *incx_, doublecomplex *y, integer *incy_) {
  333. integer n = *n_, incx = *incx_, incy = *incy_, i;
  334. _Complex double zdotc = 0.0;
  335. if (incx == 1 && incy == 1) {
  336. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  337. zdotc += conj(Cd(&x[i])) * Cd(&y[i]);
  338. }
  339. } else {
  340. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  341. zdotc += conj(Cd(&x[i*incx])) * Cd(&y[i*incy]);
  342. }
  343. }
  344. pCd(z) = zdotc;
  345. }
  346. static inline void cdotu_(complex *z, integer *n_, complex *x, integer *incx_, complex *y, integer *incy_) {
  347. integer n = *n_, incx = *incx_, incy = *incy_, i;
  348. _Complex float zdotc = 0.0;
  349. if (incx == 1 && incy == 1) {
  350. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  351. zdotc += Cf(&x[i]) * Cf(&y[i]);
  352. }
  353. } else {
  354. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  355. zdotc += Cf(&x[i*incx]) * Cf(&y[i*incy]);
  356. }
  357. }
  358. pCf(z) = zdotc;
  359. }
  360. static inline void zdotu_(doublecomplex *z, integer *n_, doublecomplex *x, integer *incx_, doublecomplex *y, integer *incy_) {
  361. integer n = *n_, incx = *incx_, incy = *incy_, i;
  362. _Complex double zdotc = 0.0;
  363. if (incx == 1 && incy == 1) {
  364. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  365. zdotc += Cd(&x[i]) * Cd(&y[i]);
  366. }
  367. } else {
  368. for (i=0;i<n;i++) { /* zdotc = zdotc + dconjg(x(i))* y(i) */
  369. zdotc += Cd(&x[i*incx]) * Cd(&y[i*incy]);
  370. }
  371. }
  372. pCd(z) = zdotc;
  373. }
  374. #endif
  375. /* -- translated by f2c (version 20000121).
  376. You must link the resulting object file with the libraries:
  377. -lf2c -lm (in that order)
  378. */
  379. /* Table of constant values */
  380. static integer c__1 = 1;
  381. /* > \brief \b SLARRF finds a new relatively robust representation such that at least one of the eigenvalues i
  382. s relatively isolated. */
  383. /* =========== DOCUMENTATION =========== */
  384. /* Online html documentation available at */
  385. /* http://www.netlib.org/lapack/explore-html/ */
  386. /* > \htmlonly */
  387. /* > Download SLARRF + dependencies */
  388. /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/slarrf.
  389. f"> */
  390. /* > [TGZ]</a> */
  391. /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/slarrf.
  392. f"> */
  393. /* > [ZIP]</a> */
  394. /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/slarrf.
  395. f"> */
  396. /* > [TXT]</a> */
  397. /* > \endhtmlonly */
  398. /* Definition: */
  399. /* =========== */
  400. /* SUBROUTINE SLARRF( N, D, L, LD, CLSTRT, CLEND, */
  401. /* W, WGAP, WERR, */
  402. /* SPDIAM, CLGAPL, CLGAPR, PIVMIN, SIGMA, */
  403. /* DPLUS, LPLUS, WORK, INFO ) */
  404. /* INTEGER CLSTRT, CLEND, INFO, N */
  405. /* REAL CLGAPL, CLGAPR, PIVMIN, SIGMA, SPDIAM */
  406. /* REAL D( * ), DPLUS( * ), L( * ), LD( * ), */
  407. /* $ LPLUS( * ), W( * ), WGAP( * ), WERR( * ), WORK( * ) */
  408. /* > \par Purpose: */
  409. /* ============= */
  410. /* > */
  411. /* > \verbatim */
  412. /* > */
  413. /* > Given the initial representation L D L^T and its cluster of close */
  414. /* > eigenvalues (in a relative measure), W( CLSTRT ), W( CLSTRT+1 ), ... */
  415. /* > W( CLEND ), SLARRF finds a new relatively robust representation */
  416. /* > L D L^T - SIGMA I = L(+) D(+) L(+)^T such that at least one of the */
  417. /* > eigenvalues of L(+) D(+) L(+)^T is relatively isolated. */
  418. /* > \endverbatim */
  419. /* Arguments: */
  420. /* ========== */
  421. /* > \param[in] N */
  422. /* > \verbatim */
  423. /* > N is INTEGER */
  424. /* > The order of the matrix (subblock, if the matrix split). */
  425. /* > \endverbatim */
  426. /* > */
  427. /* > \param[in] D */
  428. /* > \verbatim */
  429. /* > D is REAL array, dimension (N) */
  430. /* > The N diagonal elements of the diagonal matrix D. */
  431. /* > \endverbatim */
  432. /* > */
  433. /* > \param[in] L */
  434. /* > \verbatim */
  435. /* > L is REAL array, dimension (N-1) */
  436. /* > The (N-1) subdiagonal elements of the unit bidiagonal */
  437. /* > matrix L. */
  438. /* > \endverbatim */
  439. /* > */
  440. /* > \param[in] LD */
  441. /* > \verbatim */
  442. /* > LD is REAL array, dimension (N-1) */
  443. /* > The (N-1) elements L(i)*D(i). */
  444. /* > \endverbatim */
  445. /* > */
  446. /* > \param[in] CLSTRT */
  447. /* > \verbatim */
  448. /* > CLSTRT is INTEGER */
  449. /* > The index of the first eigenvalue in the cluster. */
  450. /* > \endverbatim */
  451. /* > */
  452. /* > \param[in] CLEND */
  453. /* > \verbatim */
  454. /* > CLEND is INTEGER */
  455. /* > The index of the last eigenvalue in the cluster. */
  456. /* > \endverbatim */
  457. /* > */
  458. /* > \param[in] W */
  459. /* > \verbatim */
  460. /* > W is REAL array, dimension */
  461. /* > dimension is >= (CLEND-CLSTRT+1) */
  462. /* > The eigenvalue APPROXIMATIONS of L D L^T in ascending order. */
  463. /* > W( CLSTRT ) through W( CLEND ) form the cluster of relatively */
  464. /* > close eigenalues. */
  465. /* > \endverbatim */
  466. /* > */
  467. /* > \param[in,out] WGAP */
  468. /* > \verbatim */
  469. /* > WGAP is REAL array, dimension */
  470. /* > dimension is >= (CLEND-CLSTRT+1) */
  471. /* > The separation from the right neighbor eigenvalue in W. */
  472. /* > \endverbatim */
  473. /* > */
  474. /* > \param[in] WERR */
  475. /* > \verbatim */
  476. /* > WERR is REAL array, dimension */
  477. /* > dimension is >= (CLEND-CLSTRT+1) */
  478. /* > WERR contain the semiwidth of the uncertainty */
  479. /* > interval of the corresponding eigenvalue APPROXIMATION in W */
  480. /* > \endverbatim */
  481. /* > */
  482. /* > \param[in] SPDIAM */
  483. /* > \verbatim */
  484. /* > SPDIAM is REAL */
  485. /* > estimate of the spectral diameter obtained from the */
  486. /* > Gerschgorin intervals */
  487. /* > \endverbatim */
  488. /* > */
  489. /* > \param[in] CLGAPL */
  490. /* > \verbatim */
  491. /* > CLGAPL is REAL */
  492. /* > \endverbatim */
  493. /* > */
  494. /* > \param[in] CLGAPR */
  495. /* > \verbatim */
  496. /* > CLGAPR is REAL */
  497. /* > absolute gap on each end of the cluster. */
  498. /* > Set by the calling routine to protect against shifts too close */
  499. /* > to eigenvalues outside the cluster. */
  500. /* > \endverbatim */
  501. /* > */
  502. /* > \param[in] PIVMIN */
  503. /* > \verbatim */
  504. /* > PIVMIN is REAL */
  505. /* > The minimum pivot allowed in the Sturm sequence. */
  506. /* > \endverbatim */
  507. /* > */
  508. /* > \param[out] SIGMA */
  509. /* > \verbatim */
  510. /* > SIGMA is REAL */
  511. /* > The shift used to form L(+) D(+) L(+)^T. */
  512. /* > \endverbatim */
  513. /* > */
  514. /* > \param[out] DPLUS */
  515. /* > \verbatim */
  516. /* > DPLUS is REAL array, dimension (N) */
  517. /* > The N diagonal elements of the diagonal matrix D(+). */
  518. /* > \endverbatim */
  519. /* > */
  520. /* > \param[out] LPLUS */
  521. /* > \verbatim */
  522. /* > LPLUS is REAL array, dimension (N-1) */
  523. /* > The first (N-1) elements of LPLUS contain the subdiagonal */
  524. /* > elements of the unit bidiagonal matrix L(+). */
  525. /* > \endverbatim */
  526. /* > */
  527. /* > \param[out] WORK */
  528. /* > \verbatim */
  529. /* > WORK is REAL array, dimension (2*N) */
  530. /* > Workspace. */
  531. /* > \endverbatim */
  532. /* > */
  533. /* > \param[out] INFO */
  534. /* > \verbatim */
  535. /* > INFO is INTEGER */
  536. /* > Signals processing OK (=0) or failure (=1) */
  537. /* > \endverbatim */
  538. /* Authors: */
  539. /* ======== */
  540. /* > \author Univ. of Tennessee */
  541. /* > \author Univ. of California Berkeley */
  542. /* > \author Univ. of Colorado Denver */
  543. /* > \author NAG Ltd. */
  544. /* > \date June 2016 */
  545. /* > \ingroup OTHERauxiliary */
  546. /* > \par Contributors: */
  547. /* ================== */
  548. /* > */
  549. /* > Beresford Parlett, University of California, Berkeley, USA \n */
  550. /* > Jim Demmel, University of California, Berkeley, USA \n */
  551. /* > Inderjit Dhillon, University of Texas, Austin, USA \n */
  552. /* > Osni Marques, LBNL/NERSC, USA \n */
  553. /* > Christof Voemel, University of California, Berkeley, USA */
  554. /* ===================================================================== */
  555. /* Subroutine */ int slarrf_(integer *n, real *d__, real *l, real *ld,
  556. integer *clstrt, integer *clend, real *w, real *wgap, real *werr,
  557. real *spdiam, real *clgapl, real *clgapr, real *pivmin, real *sigma,
  558. real *dplus, real *lplus, real *work, integer *info)
  559. {
  560. /* System generated locals */
  561. integer i__1;
  562. real r__1, r__2, r__3;
  563. /* Local variables */
  564. real growthbound, fail, fact, oldp;
  565. integer indx;
  566. real prod;
  567. integer ktry;
  568. real fail2;
  569. integer i__;
  570. real s, avgap, ldmax, rdmax;
  571. integer shift;
  572. extern /* Subroutine */ int scopy_(integer *, real *, integer *, real *,
  573. integer *);
  574. real bestshift, smlgrowth;
  575. logical dorrr1;
  576. real ldelta;
  577. extern real slamch_(char *);
  578. logical nofail;
  579. real mingap, lsigma, rdelta;
  580. logical forcer;
  581. real rsigma, clwdth;
  582. extern logical sisnan_(real *);
  583. logical sawnan1, sawnan2;
  584. real eps, tmp;
  585. logical tryrrr1;
  586. real max1, max2, rrr1, rrr2, znm2;
  587. /* -- LAPACK auxiliary routine (version 3.7.1) -- */
  588. /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
  589. /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
  590. /* June 2016 */
  591. /* ===================================================================== */
  592. /* Parameter adjustments */
  593. --work;
  594. --lplus;
  595. --dplus;
  596. --werr;
  597. --wgap;
  598. --w;
  599. --ld;
  600. --l;
  601. --d__;
  602. /* Function Body */
  603. *info = 0;
  604. /* Quick return if possible */
  605. if (*n <= 0) {
  606. return 0;
  607. }
  608. fact = 2.f;
  609. eps = slamch_("Precision");
  610. shift = 0;
  611. forcer = FALSE_;
  612. /* Note that we cannot guarantee that for any of the shifts tried, */
  613. /* the factorization has a small or even moderate element growth. */
  614. /* There could be Ritz values at both ends of the cluster and despite */
  615. /* backing off, there are examples where all factorizations tried */
  616. /* (in IEEE mode, allowing zero pivots & infinities) have INFINITE */
  617. /* element growth. */
  618. /* For this reason, we should use PIVMIN in this subroutine so that at */
  619. /* least the L D L^T factorization exists. It can be checked afterwards */
  620. /* whether the element growth caused bad residuals/orthogonality. */
  621. /* Decide whether the code should accept the best among all */
  622. /* representations despite large element growth or signal INFO=1 */
  623. /* Setting NOFAIL to .FALSE. for quick fix for bug 113 */
  624. nofail = FALSE_;
  625. /* Compute the average gap length of the cluster */
  626. clwdth = (r__1 = w[*clend] - w[*clstrt], abs(r__1)) + werr[*clend] + werr[
  627. *clstrt];
  628. avgap = clwdth / (real) (*clend - *clstrt);
  629. mingap = f2cmin(*clgapl,*clgapr);
  630. /* Initial values for shifts to both ends of cluster */
  631. /* Computing MIN */
  632. r__1 = w[*clstrt], r__2 = w[*clend];
  633. lsigma = f2cmin(r__1,r__2) - werr[*clstrt];
  634. /* Computing MAX */
  635. r__1 = w[*clstrt], r__2 = w[*clend];
  636. rsigma = f2cmax(r__1,r__2) + werr[*clend];
  637. /* Use a small fudge to make sure that we really shift to the outside */
  638. lsigma -= abs(lsigma) * 2.f * eps;
  639. rsigma += abs(rsigma) * 2.f * eps;
  640. /* Compute upper bounds for how much to back off the initial shifts */
  641. ldmax = mingap * .25f + *pivmin * 2.f;
  642. rdmax = mingap * .25f + *pivmin * 2.f;
  643. /* Computing MAX */
  644. r__1 = avgap, r__2 = wgap[*clstrt];
  645. ldelta = f2cmax(r__1,r__2) / fact;
  646. /* Computing MAX */
  647. r__1 = avgap, r__2 = wgap[*clend - 1];
  648. rdelta = f2cmax(r__1,r__2) / fact;
  649. /* Initialize the record of the best representation found */
  650. s = slamch_("S");
  651. smlgrowth = 1.f / s;
  652. fail = (real) (*n - 1) * mingap / (*spdiam * eps);
  653. fail2 = (real) (*n - 1) * mingap / (*spdiam * sqrt(eps));
  654. bestshift = lsigma;
  655. /* while (KTRY <= KTRYMAX) */
  656. ktry = 0;
  657. growthbound = *spdiam * 8.f;
  658. L5:
  659. sawnan1 = FALSE_;
  660. sawnan2 = FALSE_;
  661. /* Ensure that we do not back off too much of the initial shifts */
  662. ldelta = f2cmin(ldmax,ldelta);
  663. rdelta = f2cmin(rdmax,rdelta);
  664. /* Compute the element growth when shifting to both ends of the cluster */
  665. /* accept the shift if there is no element growth at one of the two ends */
  666. /* Left end */
  667. s = -lsigma;
  668. dplus[1] = d__[1] + s;
  669. if (abs(dplus[1]) < *pivmin) {
  670. dplus[1] = -(*pivmin);
  671. /* Need to set SAWNAN1 because refined RRR test should not be used */
  672. /* in this case */
  673. sawnan1 = TRUE_;
  674. }
  675. max1 = abs(dplus[1]);
  676. i__1 = *n - 1;
  677. for (i__ = 1; i__ <= i__1; ++i__) {
  678. lplus[i__] = ld[i__] / dplus[i__];
  679. s = s * lplus[i__] * l[i__] - lsigma;
  680. dplus[i__ + 1] = d__[i__ + 1] + s;
  681. if ((r__1 = dplus[i__ + 1], abs(r__1)) < *pivmin) {
  682. dplus[i__ + 1] = -(*pivmin);
  683. /* Need to set SAWNAN1 because refined RRR test should not be used */
  684. /* in this case */
  685. sawnan1 = TRUE_;
  686. }
  687. /* Computing MAX */
  688. r__2 = max1, r__3 = (r__1 = dplus[i__ + 1], abs(r__1));
  689. max1 = f2cmax(r__2,r__3);
  690. /* L6: */
  691. }
  692. sawnan1 = sawnan1 || sisnan_(&max1);
  693. if (forcer || max1 <= growthbound && ! sawnan1) {
  694. *sigma = lsigma;
  695. shift = 1;
  696. goto L100;
  697. }
  698. /* Right end */
  699. s = -rsigma;
  700. work[1] = d__[1] + s;
  701. if (abs(work[1]) < *pivmin) {
  702. work[1] = -(*pivmin);
  703. /* Need to set SAWNAN2 because refined RRR test should not be used */
  704. /* in this case */
  705. sawnan2 = TRUE_;
  706. }
  707. max2 = abs(work[1]);
  708. i__1 = *n - 1;
  709. for (i__ = 1; i__ <= i__1; ++i__) {
  710. work[*n + i__] = ld[i__] / work[i__];
  711. s = s * work[*n + i__] * l[i__] - rsigma;
  712. work[i__ + 1] = d__[i__ + 1] + s;
  713. if ((r__1 = work[i__ + 1], abs(r__1)) < *pivmin) {
  714. work[i__ + 1] = -(*pivmin);
  715. /* Need to set SAWNAN2 because refined RRR test should not be used */
  716. /* in this case */
  717. sawnan2 = TRUE_;
  718. }
  719. /* Computing MAX */
  720. r__2 = max2, r__3 = (r__1 = work[i__ + 1], abs(r__1));
  721. max2 = f2cmax(r__2,r__3);
  722. /* L7: */
  723. }
  724. sawnan2 = sawnan2 || sisnan_(&max2);
  725. if (forcer || max2 <= growthbound && ! sawnan2) {
  726. *sigma = rsigma;
  727. shift = 2;
  728. goto L100;
  729. }
  730. /* If we are at this point, both shifts led to too much element growth */
  731. /* Record the better of the two shifts (provided it didn't lead to NaN) */
  732. if (sawnan1 && sawnan2) {
  733. /* both MAX1 and MAX2 are NaN */
  734. goto L50;
  735. } else {
  736. if (! sawnan1) {
  737. indx = 1;
  738. if (max1 <= smlgrowth) {
  739. smlgrowth = max1;
  740. bestshift = lsigma;
  741. }
  742. }
  743. if (! sawnan2) {
  744. if (sawnan1 || max2 <= max1) {
  745. indx = 2;
  746. }
  747. if (max2 <= smlgrowth) {
  748. smlgrowth = max2;
  749. bestshift = rsigma;
  750. }
  751. }
  752. }
  753. /* If we are here, both the left and the right shift led to */
  754. /* element growth. If the element growth is moderate, then */
  755. /* we may still accept the representation, if it passes a */
  756. /* refined test for RRR. This test supposes that no NaN occurred. */
  757. /* Moreover, we use the refined RRR test only for isolated clusters. */
  758. if (clwdth < mingap / 128.f && f2cmin(max1,max2) < fail2 && ! sawnan1 && !
  759. sawnan2) {
  760. dorrr1 = TRUE_;
  761. } else {
  762. dorrr1 = FALSE_;
  763. }
  764. tryrrr1 = TRUE_;
  765. if (tryrrr1 && dorrr1) {
  766. if (indx == 1) {
  767. tmp = (r__1 = dplus[*n], abs(r__1));
  768. znm2 = 1.f;
  769. prod = 1.f;
  770. oldp = 1.f;
  771. for (i__ = *n - 1; i__ >= 1; --i__) {
  772. if (prod <= eps) {
  773. prod = dplus[i__ + 1] * work[*n + i__ + 1] / (dplus[i__] *
  774. work[*n + i__]) * oldp;
  775. } else {
  776. prod *= (r__1 = work[*n + i__], abs(r__1));
  777. }
  778. oldp = prod;
  779. /* Computing 2nd power */
  780. r__1 = prod;
  781. znm2 += r__1 * r__1;
  782. /* Computing MAX */
  783. r__2 = tmp, r__3 = (r__1 = dplus[i__] * prod, abs(r__1));
  784. tmp = f2cmax(r__2,r__3);
  785. /* L15: */
  786. }
  787. rrr1 = tmp / (*spdiam * sqrt(znm2));
  788. if (rrr1 <= 8.f) {
  789. *sigma = lsigma;
  790. shift = 1;
  791. goto L100;
  792. }
  793. } else if (indx == 2) {
  794. tmp = (r__1 = work[*n], abs(r__1));
  795. znm2 = 1.f;
  796. prod = 1.f;
  797. oldp = 1.f;
  798. for (i__ = *n - 1; i__ >= 1; --i__) {
  799. if (prod <= eps) {
  800. prod = work[i__ + 1] * lplus[i__ + 1] / (work[i__] *
  801. lplus[i__]) * oldp;
  802. } else {
  803. prod *= (r__1 = lplus[i__], abs(r__1));
  804. }
  805. oldp = prod;
  806. /* Computing 2nd power */
  807. r__1 = prod;
  808. znm2 += r__1 * r__1;
  809. /* Computing MAX */
  810. r__2 = tmp, r__3 = (r__1 = work[i__] * prod, abs(r__1));
  811. tmp = f2cmax(r__2,r__3);
  812. /* L16: */
  813. }
  814. rrr2 = tmp / (*spdiam * sqrt(znm2));
  815. if (rrr2 <= 8.f) {
  816. *sigma = rsigma;
  817. shift = 2;
  818. goto L100;
  819. }
  820. }
  821. }
  822. L50:
  823. if (ktry < 1) {
  824. /* If we are here, both shifts failed also the RRR test. */
  825. /* Back off to the outside */
  826. /* Computing MAX */
  827. r__1 = lsigma - ldelta, r__2 = lsigma - ldmax;
  828. lsigma = f2cmax(r__1,r__2);
  829. /* Computing MIN */
  830. r__1 = rsigma + rdelta, r__2 = rsigma + rdmax;
  831. rsigma = f2cmin(r__1,r__2);
  832. ldelta *= 2.f;
  833. rdelta *= 2.f;
  834. ++ktry;
  835. goto L5;
  836. } else {
  837. /* None of the representations investigated satisfied our */
  838. /* criteria. Take the best one we found. */
  839. if (smlgrowth < fail || nofail) {
  840. lsigma = bestshift;
  841. rsigma = bestshift;
  842. forcer = TRUE_;
  843. goto L5;
  844. } else {
  845. *info = 1;
  846. return 0;
  847. }
  848. }
  849. L100:
  850. if (shift == 1) {
  851. } else if (shift == 2) {
  852. /* store new L and D back into DPLUS, LPLUS */
  853. scopy_(n, &work[1], &c__1, &dplus[1], &c__1);
  854. i__1 = *n - 1;
  855. scopy_(&i__1, &work[*n + 1], &c__1, &lplus[1], &c__1);
  856. }
  857. return 0;
  858. /* End of SLARRF */
  859. } /* slarrf_ */