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 complex c_b1 = {0.f,0.f};
516 static integer c__2 = 2;
517 static integer c_n1 = -1;
518 static integer c__3 = 3;
519 static integer c__4 = 4;
521 /* > \brief \b CHBTRD_HB2ST reduces a complex Hermitian band matrix A to real symmetric tridiagonal form T */
523 /* =========== DOCUMENTATION =========== */
525 /* Online html documentation available at */
526 /* http://www.netlib.org/lapack/explore-html/ */
529 /* > Download CHBTRD_HB2ST + dependencies */
530 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.tgz?format=tgz&filename=/lapack/lapack_routine/chbtrd_
533 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.zip?format=zip&filename=/lapack/lapack_routine/chbtrd_
536 /* > <a href="http://www.netlib.org/cgi-bin/netlibfiles.txt?format=txt&filename=/lapack/lapack_routine/chbtrd_
544 /* SUBROUTINE CHETRD_HB2ST( STAGE1, VECT, UPLO, N, KD, AB, LDAB, */
545 /* D, E, HOUS, LHOUS, WORK, LWORK, INFO ) */
547 /* #if defined(_OPENMP) */
553 /* CHARACTER STAGE1, UPLO, VECT */
554 /* INTEGER N, KD, IB, LDAB, LHOUS, LWORK, INFO */
555 /* REAL D( * ), E( * ) */
556 /* COMPLEX AB( LDAB, * ), HOUS( * ), WORK( * ) */
559 /* > \par Purpose: */
564 /* > CHETRD_HB2ST reduces a complex Hermitian band matrix A to real symmetric */
565 /* > tridiagonal form T by a unitary similarity transformation: */
566 /* > Q**H * A * Q = T. */
572 /* > \param[in] STAGE1 */
574 /* > STAGE1 is CHARACTER*1 */
575 /* > = 'N': "No": to mention that the stage 1 of the reduction */
576 /* > from dense to band using the chetrd_he2hb routine */
577 /* > was not called before this routine to reproduce AB. */
578 /* > In other term this routine is called as standalone. */
579 /* > = 'Y': "Yes": to mention that the stage 1 of the */
580 /* > reduction from dense to band using the chetrd_he2hb */
581 /* > routine has been called to produce AB (e.g., AB is */
582 /* > the output of chetrd_he2hb. */
585 /* > \param[in] VECT */
587 /* > VECT is CHARACTER*1 */
588 /* > = 'N': No need for the Housholder representation, */
589 /* > and thus LHOUS is of size f2cmax(1, 4*N); */
590 /* > = 'V': the Householder representation is needed to */
591 /* > either generate or to apply Q later on, */
592 /* > then LHOUS is to be queried and computed. */
593 /* > (NOT AVAILABLE IN THIS RELEASE). */
596 /* > \param[in] UPLO */
598 /* > UPLO is CHARACTER*1 */
599 /* > = 'U': Upper triangle of A is stored; */
600 /* > = 'L': Lower triangle of A is stored. */
606 /* > The order of the matrix A. N >= 0. */
609 /* > \param[in] KD */
611 /* > KD is INTEGER */
612 /* > The number of superdiagonals of the matrix A if UPLO = 'U', */
613 /* > or the number of subdiagonals if UPLO = 'L'. KD >= 0. */
616 /* > \param[in,out] AB */
618 /* > AB is COMPLEX array, dimension (LDAB,N) */
619 /* > On entry, the upper or lower triangle of the Hermitian band */
620 /* > matrix A, stored in the first KD+1 rows of the array. The */
621 /* > j-th column of A is stored in the j-th column of the array AB */
623 /* > if UPLO = 'U', AB(kd+1+i-j,j) = A(i,j) for f2cmax(1,j-kd)<=i<=j; */
624 /* > if UPLO = 'L', AB(1+i-j,j) = A(i,j) for j<=i<=f2cmin(n,j+kd). */
625 /* > On exit, the diagonal elements of AB are overwritten by the */
626 /* > diagonal elements of the tridiagonal matrix T; if KD > 0, the */
627 /* > elements on the first superdiagonal (if UPLO = 'U') or the */
628 /* > first subdiagonal (if UPLO = 'L') are overwritten by the */
629 /* > off-diagonal elements of T; the rest of AB is overwritten by */
630 /* > values generated during the reduction. */
633 /* > \param[in] LDAB */
635 /* > LDAB is INTEGER */
636 /* > The leading dimension of the array AB. LDAB >= KD+1. */
639 /* > \param[out] D */
641 /* > D is REAL array, dimension (N) */
642 /* > The diagonal elements of the tridiagonal matrix T. */
645 /* > \param[out] E */
647 /* > E is REAL array, dimension (N-1) */
648 /* > The off-diagonal elements of the tridiagonal matrix T: */
649 /* > E(i) = T(i,i+1) if UPLO = 'U'; E(i) = T(i+1,i) if UPLO = 'L'. */
652 /* > \param[out] HOUS */
654 /* > HOUS is COMPLEX array, dimension LHOUS, that */
655 /* > store the Householder representation. */
658 /* > \param[in] LHOUS */
660 /* > LHOUS is INTEGER */
661 /* > The dimension of the array HOUS. LHOUS = MAX(1, dimension) */
662 /* > If LWORK = -1, or LHOUS=-1, */
663 /* > then a query is assumed; the routine */
664 /* > only calculates the optimal size of the HOUS array, returns */
665 /* > this value as the first entry of the HOUS array, and no error */
666 /* > message related to LHOUS is issued by XERBLA. */
667 /* > LHOUS = MAX(1, dimension) where */
668 /* > dimension = 4*N if VECT='N' */
669 /* > not available now if VECT='H' */
672 /* > \param[out] WORK */
674 /* > WORK is COMPLEX array, dimension LWORK. */
677 /* > \param[in] LWORK */
679 /* > LWORK is INTEGER */
680 /* > The dimension of the array WORK. LWORK = MAX(1, dimension) */
681 /* > If LWORK = -1, or LHOUS=-1, */
682 /* > then a workspace query is assumed; the routine */
683 /* > only calculates the optimal size of the WORK array, returns */
684 /* > this value as the first entry of the WORK array, and no error */
685 /* > message related to LWORK is issued by XERBLA. */
686 /* > LWORK = MAX(1, dimension) where */
687 /* > dimension = (2KD+1)*N + KD*NTHREADS */
688 /* > where KD is the blocking size of the reduction, */
689 /* > FACTOPTNB is the blocking used by the QR or LQ */
690 /* > algorithm, usually FACTOPTNB=128 is a good choice */
691 /* > NTHREADS is the number of threads used when */
692 /* > openMP compilation is enabled, otherwise =1. */
695 /* > \param[out] INFO */
697 /* > INFO is INTEGER */
698 /* > = 0: successful exit */
699 /* > < 0: if INFO = -i, the i-th argument had an illegal value */
705 /* > \author Univ. of Tennessee */
706 /* > \author Univ. of California Berkeley */
707 /* > \author Univ. of Colorado Denver */
708 /* > \author NAG Ltd. */
710 /* > \date November 2017 */
712 /* > \ingroup complexOTHERcomputational */
714 /* > \par Further Details: */
715 /* ===================== */
719 /* > Implemented by Azzam Haidar. */
721 /* > All details are available on technical report, SC11, SC13 papers. */
723 /* > Azzam Haidar, Hatem Ltaief, and Jack Dongarra. */
724 /* > Parallel reduction to condensed forms for symmetric eigenvalue problems */
725 /* > using aggregated fine-grained and memory-aware kernels. In Proceedings */
726 /* > of 2011 International Conference for High Performance Computing, */
727 /* > Networking, Storage and Analysis (SC '11), New York, NY, USA, */
728 /* > Article 8 , 11 pages. */
729 /* > http://doi.acm.org/10.1145/2063384.2063394 */
731 /* > A. Haidar, J. Kurzak, P. Luszczek, 2013. */
732 /* > An improved parallel singular value algorithm and its implementation */
733 /* > for multicore hardware, In Proceedings of 2013 International Conference */
734 /* > for High Performance Computing, Networking, Storage and Analysis (SC '13). */
735 /* > Denver, Colorado, USA, 2013. */
736 /* > Article 90, 12 pages. */
737 /* > http://doi.acm.org/10.1145/2503210.2503292 */
739 /* > A. Haidar, R. Solca, S. Tomov, T. Schulthess and J. Dongarra. */
740 /* > A novel hybrid CPU-GPU generalized eigensolver for electronic structure */
741 /* > calculations based on fine-grained memory aware tasks. */
742 /* > International Journal of High Performance Computing Applications. */
743 /* > Volume 28 Issue 2, Pages 196-209, May 2014. */
744 /* > http://hpc.sagepub.com/content/28/2/196 */
748 /* ===================================================================== */
749 /* Subroutine */ int chetrd_hb2st_(char *stage1, char *vect, char *uplo,
750 integer *n, integer *kd, complex *ab, integer *ldab, real *d__, real *
751 e, complex *hous, integer *lhous, complex *work, integer *lwork,
754 /* System generated locals */
755 integer ab_dim1, ab_offset, i__1, i__2, i__3, i__4, i__5;
759 /* Local variables */
761 extern integer ilaenv2stage_(integer *, char *, char *, integer *,
762 integer *, integer *, integer *);
763 integer thed, indv, myid, indw, apos, dpos, abofdpos, nthreads, i__, k, m,
765 extern logical lsame_(char *, char *);
766 integer lhmin, sicev, sizea, shift, stind, colpt, lwmin, awpos;
767 logical wantq, upper;
768 integer grsiz, ttype, stepercol, ed, ib;
769 extern /* Subroutine */ int chb2st_kernels_(char *, logical *, integer *,
770 integer *, integer *, integer *, integer *, integer *, integer *,
771 complex *, integer *, complex *, complex *, integer *, complex *);
773 extern /* Subroutine */ int clacpy_(char *, integer *, integer *, complex
774 *, integer *, complex *, integer *), claset_(char *,
775 integer *, integer *, complex *, complex *, complex *, integer *), xerbla_(char *, integer *, ftnlen);
776 integer thgrid, thgrnb, indtau;
778 integer ofdpos, blklastind;
779 extern /* Subroutine */ int mecago_();
780 logical lquery, afters1;
781 integer lda, tid, ldv;
783 integer stt, sweepid, nbtiles, sizetau, thgrsiz;
789 /* -- LAPACK computational routine (version 3.8.0) -- */
790 /* -- LAPACK is a software package provided by Univ. of Tennessee, -- */
791 /* -- Univ. of California Berkeley, Univ. of Colorado Denver and NAG Ltd..-- */
795 /* ===================================================================== */
798 /* Determine the minimal workspace size required. */
799 /* Test the input parameters */
801 /* Parameter adjustments */
803 ab_offset = 1 + ab_dim1 * 1;
813 afters1 = lsame_(stage1, "Y");
814 wantq = lsame_(vect, "V");
815 upper = lsame_(uplo, "U");
816 lquery = *lwork == -1 || *lhous == -1;
818 /* Determine the block size, the workspace size and the hous size. */
820 ib = ilaenv2stage_(&c__2, "CHETRD_HB2ST", vect, n, kd, &c_n1, &c_n1);
821 lhmin = ilaenv2stage_(&c__3, "CHETRD_HB2ST", vect, n, kd, &ib, &c_n1);
822 lwmin = ilaenv2stage_(&c__4, "CHETRD_HB2ST", vect, n, kd, &ib, &c_n1);
824 if (! afters1 && ! lsame_(stage1, "N")) {
826 } else if (! lsame_(vect, "N")) {
828 } else if (! upper && ! lsame_(uplo, "L")) {
832 } else if (*kd < 0) {
834 } else if (*ldab < *kd + 1) {
836 } else if (*lhous < lhmin && ! lquery) {
838 } else if (*lwork < lwmin && ! lquery) {
843 hous[1].r = (real) lhmin, hous[1].i = 0.f;
844 work[1].r = (real) lwmin, work[1].i = 0.f;
849 xerbla_("CHETRD_HB2ST", &i__1, (ftnlen)12);
855 /* Quick return if possible */
858 hous[1].r = 1.f, hous[1].i = 0.f;
859 work[1].r = 1.f, work[1].i = 0.f;
863 /* Determine pointer position */
869 indv = indtau + sizetau;
870 lda = (*kd << 1) + 1;
886 awpos = inda + *kd + 1;
894 /* The matrix is diagonal. We just copy it (convert to "real" for */
895 /* complex because D is double and the imaginary part should be 0) */
896 /* and store it in D. A sequential code here is better or */
897 /* in a parallel environment it might need two cores for D and E */
901 for (i__ = 1; i__ <= i__1; ++i__) {
902 i__2 = abdpos + i__ * ab_dim1;
903 d__[i__] = ab[i__2].r;
907 for (i__ = 1; i__ <= i__1; ++i__) {
912 hous[1].r = 1.f, hous[1].i = 0.f;
913 work[1].r = 1.f, work[1].i = 0.f;
918 /* The matrix is already Tridiagonal. We have to make diagonal */
919 /* and offdiagonal elements real, and store them in D and E. */
920 /* For that, for real precision just copy the diag and offdiag */
921 /* to D and E while for the COMPLEX case the bulge chasing is */
922 /* performed to convert the hermetian tridiagonal to symmetric */
923 /* tridiagonal. A simpler coversion formula might be used, but then */
924 /* updating the Q matrix will be required and based if Q is generated */
925 /* or not this might complicate the story. */
929 for (i__ = 1; i__ <= i__1; ++i__) {
930 i__2 = abdpos + i__ * ab_dim1;
931 d__[i__] = ab[i__2].r;
935 /* make off-diagonal elements real and copy them to E */
939 for (i__ = 1; i__ <= i__1; ++i__) {
940 i__2 = abofdpos + (i__ + 1) * ab_dim1;
941 tmp.r = ab[i__2].r, tmp.i = ab[i__2].i;
942 abstmp = c_abs(&tmp);
943 i__2 = abofdpos + (i__ + 1) * ab_dim1;
944 ab[i__2].r = abstmp, ab[i__2].i = 0.f;
947 q__1.r = tmp.r / abstmp, q__1.i = tmp.i / abstmp;
948 tmp.r = q__1.r, tmp.i = q__1.i;
950 tmp.r = 1.f, tmp.i = 0.f;
953 i__2 = abofdpos + (i__ + 2) * ab_dim1;
954 i__3 = abofdpos + (i__ + 2) * ab_dim1;
955 q__1.r = ab[i__3].r * tmp.r - ab[i__3].i * tmp.i, q__1.i =
956 ab[i__3].r * tmp.i + ab[i__3].i * tmp.r;
957 ab[i__2].r = q__1.r, ab[i__2].i = q__1.i;
959 /* IF( WANTZ ) THEN */
960 /* CALL CSCAL( N, CONJG( TMP ), Q( 1, I+1 ), 1 ) */
966 for (i__ = 1; i__ <= i__1; ++i__) {
967 i__2 = abofdpos + i__ * ab_dim1;
968 tmp.r = ab[i__2].r, tmp.i = ab[i__2].i;
969 abstmp = c_abs(&tmp);
970 i__2 = abofdpos + i__ * ab_dim1;
971 ab[i__2].r = abstmp, ab[i__2].i = 0.f;
974 q__1.r = tmp.r / abstmp, q__1.i = tmp.i / abstmp;
975 tmp.r = q__1.r, tmp.i = q__1.i;
977 tmp.r = 1.f, tmp.i = 0.f;
980 i__2 = abofdpos + (i__ + 1) * ab_dim1;
981 i__3 = abofdpos + (i__ + 1) * ab_dim1;
982 q__1.r = ab[i__3].r * tmp.r - ab[i__3].i * tmp.i, q__1.i =
983 ab[i__3].r * tmp.i + ab[i__3].i * tmp.r;
984 ab[i__2].r = q__1.r, ab[i__2].i = q__1.i;
986 /* IF( WANTQ ) THEN */
987 /* CALL CSCAL( N, TMP, Q( 1, I+1 ), 1 ) */
993 hous[1].r = 1.f, hous[1].i = 0.f;
994 work[1].r = 1.f, work[1].i = 0.f;
998 /* Main code start here. */
999 /* Reduce the hermitian band of A to a tridiagonal matrix. */
1004 r__1 = (real) (*n) / (real) (*kd) + .5f;
1005 nbtiles = r_int(&r__1);
1006 r__1 = (real) shift / (real) grsiz + .5f;
1007 stepercol = r_int(&r__1);
1008 r__1 = (real) (*n - 1) / (real) thgrsiz + .5f;
1009 thgrnb = r_int(&r__1);
1012 clacpy_("A", &i__1, n, &ab[ab_offset], ldab, &work[apos], &lda)
1014 claset_("A", kd, n, &c_b1, &c_b1, &work[awpos], &lda);
1017 /* openMP parallelisation start here */
1020 /* main bulge chasing loop */
1023 for (thgrid = 1; thgrid <= i__1; ++thgrid) {
1024 stt = (thgrid - 1) * thgrsiz + 1;
1026 i__2 = stt + thgrsiz - 1, i__3 = *n - 1;
1027 thed = f2cmin(i__2,i__3);
1029 for (i__ = stt; i__ <= i__2; ++i__) {
1030 ed = f2cmin(i__,thed);
1035 for (m = 1; m <= i__3; ++m) {
1038 for (sweepid = st; sweepid <= i__4; ++sweepid) {
1040 for (k = 1; k <= i__5; ++k) {
1041 myid = (i__ - sweepid) * (stepercol * grsiz) + (m - 1)
1046 ttype = myid % 2 + 2;
1049 colpt = myid / 2 * *kd + sweepid;
1050 stind = colpt - *kd + 1;
1051 edind = f2cmin(colpt,*n);
1054 colpt = (myid + 1) / 2 * *kd + sweepid;
1055 stind = colpt - *kd + 1;
1056 edind = f2cmin(colpt,*n);
1057 if (stind >= edind - 1 && edind == *n) {
1064 /* Call the kernel */
1066 chb2st_kernels_(uplo, &wantq, &ttype, &stind, &edind,
1067 &sweepid, n, kd, &ib, &work[inda], &lda, &
1068 hous[indv], &hous[indtau], &ldv, &work[indw +
1070 if (blklastind >= *n - 1) {
1086 /* Copy the diagonal from A to D. Note that D is REAL thus only */
1087 /* the Real part is needed, the imaginary part should be zero. */
1090 for (i__ = 1; i__ <= i__1; ++i__) {
1091 i__2 = dpos + (i__ - 1) * lda;
1092 d__[i__] = work[i__2].r;
1096 /* Copy the off diagonal from A to E. Note that E is REAL thus only */
1097 /* the Real part is needed, the imaginary part should be zero. */
1101 for (i__ = 1; i__ <= i__1; ++i__) {
1102 i__2 = ofdpos + i__ * lda;
1103 e[i__] = work[i__2].r;
1108 for (i__ = 1; i__ <= i__1; ++i__) {
1109 i__2 = ofdpos + (i__ - 1) * lda;
1110 e[i__] = work[i__2].r;
1115 hous[1].r = (real) lhmin, hous[1].i = 0.f;
1116 work[1].r = (real) lwmin, work[1].i = 0.f;
1119 /* End of CHETRD_HB2ST */
1121 } /* chetrd_hb2st__ */